Pith. sign in

REVIEW 4 major objections 5 minor 2 cited by

Embedding the singular DVCS integral as a differentiable layer in a neural network gives a stable, model-independent extraction of generalized parton distributions from Compton form factors.

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

2026-08-03 14:06 UTC pith:73EPNWB7

load-bearing objection Novel differentiable PV-inversion scheme, but Eq. (5) has a sign error that flips the extracted GPD, and the closure test is blind to it. the 4 major comments →

arxiv 2512.21761 v2 pith:73EPNWB7 submitted 2025-12-25 hep-ph

Differentiable Principal-Value Inversion for Neural-Network Extraction of Generalized Parton Distributions

Dima Watkins , Ishara Fernando , Dustin Keller This is my paper
classification hep-ph
keywords generalized parton distributionsCompton form factorsdeeply virtual Compton scatteringprincipal-value integral inversiondifferentiable physics layerneural networksuncertainty quantificationinverse problem
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.

This paper claims that the ill-posed inverse problem of extracting the C-even quark GPD H^(+)(x,ξ,t,Q²) from the real part of the DVCS Compton form factor can be solved by building the principal-value integral transform directly into a neural network as a fixed, differentiable layer. The network outputs the GPD itself, the transform produces Re H, and training happens end-to-end, avoiding restrictive functional parametrizations. Replica ensembles propagate experimental uncertainties into a probabilistic GPD with credible bands. A closure test with synthetic data reports reconstruction errors near 10⁻⁶, and the method is demonstrated on Jefferson Lab CFF inputs. If correct, this turns a longstanding ill-posed inversion into a tractable, uncertainty-quantified learning problem.

Core claim

The paper establishes that the singular PV transform relating H^(+) to Re H can be discretized with an excised trapezoidal kernel and embedded as a differentiable operator inside a neural network, enabling direct optimization over the GPD itself. The architecture enforces exact oddness in x, endpoint suppression via (1−|x|)^β, and curvature regularization in x, which together stabilize the inversion. A Monte-Carlo replica ensemble then converts the global CFF fit uncertainties into a full probability distribution over H^(+)(x,ξ,t,Q²). The authors report closure-test agreement at the 10⁻⁶ level and present one-dimensional slices and three-dimensional surfaces of the extracted GPD.

What carries the argument

The central object is the discrete principal-value operator K_i(ξ)=[1/(ξ−x_i)−1/(ξ+x_i)] Θ(|x_i−ξ|−ε_PV) Θ(|x_i+ξ|−ε_PV) with trapezoidal weights, implemented as a fixed differentiable layer. It carries the physics of the singular Hilbert-transform-type integral, while the oddness operator O[f](x)=½[f(x)−f(−x)] and endpoint envelope (1−|x|)^β impose the GPD's known analytic structure. A second-derivative Tikhonov penalty in x suppresses the spurious oscillations amplified by the ill-conditioned kernel, and the replica ensemble provides uncertainty propagation.

Load-bearing premise

The load-bearing premise is that the discrete excised-principal-value operator (trapezoid rule with hard exclusion of the singular points, no analytic pole correction) is an accurate proxy for the exact continuum PV transform — and that closure tests, which use this same operator for pseudo-data, loss, and truth, can detect any bias introduced by that discretization.

What would settle it

Generate pseudo-Re-H data using an independent, high-accuracy PV evaluation (e.g., analytic pole subtraction or a much finer grid with local correction) from a known analytic GPD, feed these data into the network, and compare the extracted GPD to the known input; if the difference exceeds the quoted replica uncertainty bands — especially near x=ξ or for a GPD with a sharp ridge at the crossing point — the discrete operator's bias is the cause and the central claim fails.

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

If this is right

  • A direct, nonparametric extraction of H^(+) becomes feasible from Re H data alone, without double-distribution or Regge-style functional ansätze.
  • Uncertainties from experimental CFF fits propagate, through the replica ensemble, into pointwise credible bands on the GPD across (x, ξ, t, Q²).
  • The same differentiable-layer architecture extends immediately to the other CFFs (E, H̃, Ẽ) and to joint fits of Re H and Im H, with Im H anchoring the GPD on the cross-over line x=ξ.
  • The learned Q² dependence emerges from data and the smoothness priors, providing a no-evolution baseline against which QCD-evolution prescriptions can be compared.
  • Closure tests with tuned hyperparameters give a quantitative lower bound on the inversion's algorithmic error and a way to calibrate the smoothness prior strength.

Where Pith is reading between the lines

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

  • The 10⁻⁶ closure error is measured only against the same discrete PV operator used inside the network, so it validates the network's in-class fitting capability, not the fidelity of the excised-trapezoid discretization to the true continuum PV integral; an independent high-accuracy PV evaluation is needed to bound that systematic.
  • Because ξ generally lies off the x-grid and the mask removes a band of width O(Δx), the discretization error is O(Δx) and could be largest precisely where the GPD varies steeply near x=ξ; joint use of Im H, or a local refinement near the crossing line, would directly test this.
  • The endpoint envelope (1−|x|)^β functions as a modest structural prior; making β trainable with a soft positivity penalty would let the replica ensemble quantify how much of the large-|x| behavior is data-driven versus imposed.
  • The extracted GPD inherits all systematics from the input global CFF DNN fit, so the method's overall reliability is bounded by that fit's coverage and validity, not just by the inversion machinery itself.

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

4 major / 5 minor

Summary. The manuscript presents a neural-network method for extracting the C-even GPD H^{(+)}(x,\xi,t,Q^2) from the real part of the DVCS CFF H. The QCD principal-value integral transform is implemented as a fixed differentiable layer inside a PyTorch network; oddness in x and endpoint suppression are imposed analytically, curvature regularization stabilizes the inversion, and a Monte Carlo replica ensemble propagates CFF uncertainties. A closure test with a GPD surrogate is used to tune regularization, and the method is applied to a global DNN CFF fit to Jefferson Lab DVCS data, yielding one-dimensional slices and three-dimensional GPD surfaces with credible bands. The central claim is that this provides a model-independent, uncertainty-quantified inversion of the singular PV transform.

Significance. The differentiable-PV-layer idea is a sensible and potentially useful methodological advance: it avoids a two-step CFF-to-GPD post-processing, makes the singular transform part of the training graph, and the replica ensemble provides a clear uncertainty-propagation scheme. The exact oddness and endpoint factors are appropriate structural priors, and the paper is explicit about several limitations, including the role of the endpoint prior and the null space of the inversion. However, the current manuscript contains a sign error in the central integral relation that, if reflected in the code, flips the physical sign of the extracted GPD. The closure test is constructed from the pipeline's own first inversion and from the same discrete operator used for pseudo-data and loss, so it cannot detect this or other systematic errors. The framework is promising, but the physical results as presented are not yet supported.

major comments (4)
  1. [II, Eq. (5); IV.C, Eq. (25)] Taking the real part of Eq. (2) with 1/(x±i0) identities and using H^(+)=H(x)-H(-x) gives Re H = PV∫_0^1 dx [1/(x-ξ)+1/(x+ξ)] H^(+)(x) = -PV∫_0^1 dx [1/(ξ-x)-1/(ξ+x)] H^(+)(x). Equation (5) prints the second form without the minus sign, i.e., the negative of the kernel derived from Eq. (2). The same wrong-sign kernel appears in Eq. (25) and Eq. (37). Because pseudo-data, loss, and closure truth all use this kernel, the closure test cannot detect the sign: a network output H^(+)_θ ≈ -H^(+)_phys would satisfy the loss. The sign convention in Eq. (4) is consistent with Eq. (2), so this is not a convention ambiguity. Please correct the sign and add a physical consistency check such as the forward limit H^(+)(x,0,0)>0.
  2. [V.A and V.C] The closure 'truth' is not independent. Section V.A states that H_true is obtained by first running the PV-DNN inversion on the CFF mean and then fitting the result with Eq. (31); Section V.C repeats this. Therefore the closure test verifies only that the network can fit the function class spanned by its own first inversion when the same discrete operator is used to generate pseudo-data and to compute the loss. Any systematic error common to the operator (sign, normalization, excision bias) is invisible. To support the claim of ~10^-6 methodological error, use an externally specified H_true (e.g., a double-distribution model) and generate pseudo-ReH with an independent PV evaluation.
  3. [IV.A-C, Eqs. (10), (25), (36)] The normalization of the discrete operator is inconsistent as printed. Eq. (5) integrates over x∈[0,1]. On the symmetric grid x_i∈[-1,1], both K_i(ξ) and H^(+)(x_i) are odd, so the product is even and the full sum in Eq. (25) equals 2∫_0^1 dx K H^(+) (up to quadrature). Eq. (36) contains the needed factor 1/2, but Eq. (10) and Eq. (25) do not. If the code follows Eq. (25), the network's predicted ReH is a factor of two larger than the pseudo-data projection, so the trained GPD would be half of H_true; if the code follows Eq. (36), Eqs. (10)/(25) are misprinted. The manuscript must state the exact normalization used and demonstrate it in the closure test.
  4. [IV.A-B, Eqs. (11), (24)] The hard-excision trapezoid rule with no analytic pole term has an O(Δx) bias when ξ is not on the grid. The stability scan over n_PV only varies the excision within the same approximation; it does not compare with the continuum PV integral. Since Eq. (36) generates the pseudo-data with the same discrete operator, the closure residual cannot bound this bias. Add a direct validation of Eq. (25) against an analytic PV integral (or a high-order quadrature with pole subtraction) for a known H^(+), reporting the error as a function of N_x and n_PV.
minor comments (5)
  1. [Abstract and I] The phrase 'model-independent' is too strong given the explicit endpoint envelope E(x)=(1-|x|)^β and the smoothness priors. The text later acknowledges this; please soften the earlier wording.
  2. [II, Eqs. (3)-(5)] The notation H^(+) is overloaded: Eq. (3) defines the flavor-specific H^(+)_q, while Eq. (5) introduces the charge-weighted H^(+). Please clarify the change of notation explicitly.
  3. [V.A, Eqs. (31), (36)] The function is called H_fit in Eq. (31) but H_true in Eq. (36). Use a single name or define the mapping to improve readability.
  4. [IV.E, Eq. (30)] The replica noise in Eq. (30) assumes independent σ_n and neglects correlations from the global CFF fit. A sentence justifying this or discussing its effect would be useful.
  5. [VI, Figs. 2-4] The figures would benefit from stating N_rep (the number of replicas used) and from colorbars with units for the GPD surfaces; currently the scale is only implicit.

Circularity Check

3 steps flagged

Validation loop is partially circular: the closure 'truth' is produced by the method's own first inversion and the pseudo-data use the same discrete PV operator as the network's forward layer, so the ~1e-6 closure error certifies internal consistency only; the printed kernel also has the opposite sign of the continuum real part derived from Eq. (2), an error the closure cannot expose.

specific steps
  1. fitted input called prediction [Sec. V.A (Eqs. 31-36) and Sec. V.C]
    "To parameterize this function to use as the closure test generator we first perform an initial GPD extraction on the CFF mean, and then perform a fit to the GPD to obtain H_true. ... The corresponding true CFF ReH_true is obtained by applying exactly the same discrete PV operator as used in the network architecture."

    The 'analytic truth' H_true is not an independent test function: it is the output of the PV-DNN pipeline, smoothed by the analytic form Eq. (31), and then re-fit. The pseudo-data are generated by projecting this pipeline-derived H_true through the same discrete operator that appears as the network's forward layer. The closure test therefore measures only how well a network can fit a smoothed version of its own first inversion; any systematic error in the operator or the first inversion is imprinted on H_true and is invisible to the ~1e-6 agreement.

  2. self definitional [Sec. IV.C Eq. (25) and Sec. V.A Eqs. (36)-(37)]
    "ReH_theta(xi,t,Q^2) = sum_{i in M(xi)} w_i [1/(xi-x_i) - 1/(xi+x_i)] H_theta(x_i,xi,t,Q^2) ... ReH_true(xi,t,Q^2) = 1/2 sum H_true(x_n,xi,t,Q^2) K_n(xi) w_n ... K_n ... implements the numerical principal value with an exclusion half-width epsilon_PV = n_PV Delta x, as in Eq. (11)."

    The same discretized excised-PV kernel (trapezoid rule with the same n_PV excision mask) is used in three places: to define the network's forward prediction, to generate the synthetic CFF pseudo-data, and to compute the closure 'truth.' Thus the ~1e-6 closure residual bounds only the network's in-class fitting error for this fixed operator. It cannot test the fidelity of the discretized/excised operator to the continuum PV transform, the O(Delta x) excision error, or any global sign error, because every stage of the validation uses the identical operator.

  3. other [Sec. II Eq. (2) vs Eq. (5), and Sec. IV.C Eq. (25)/Sec. V.A Eq. (36)]
    "H(xi,t,Q^2)|_LO = sum_q e_q^2 int_{-1}^{1} dx [1/(x-xi+i0) + 1/(x+xi-i0)] H^q(x,xi,t,Q^2) ... ReH(xi,t,Q^2) = PV int_0^1 dx [1/(xi-x) - 1/(xi+x)] H^(+)(x,xi,t,Q^2)."

    Taking the real part of Eq. (2) with the standard distribution identity gives ReH = PV int_0^1 [1/(x-xi) + 1/(x+xi)] H^(+)(x) dx, which is the negative of the printed kernel in Eq. (5). The discrete operator Eq. (25) and the pseudo-data generator Eq. (36) both use the printed sign. If the implementation follows the printed kernel, the trained H^(+) is approximately -H^(+)_phys (up to regularized null-space contributions), and the forward-limit check H^(+)(x,0,0)=q(x)-qbar(x)>0 would fail. Because pseudo-data and closure truth are built with the same sign-flipped operator, the ~1e-6 closure test cannot detect this: it only certifies that the network inverts the network's own forward operator, not the physical kernel.

full rationale

The central inversion itself is a legitimate inverse problem: the network H_theta is trained by minimizing chi^2 between the discrete PV projection of H_theta and the input CFF values, so the extracted GPD is not a mere re-labeling of the input. The oddness projection, endpoint envelope, and curvature regularization are explicit priors rather than hidden inputs. The circularity is concentrated in the validation loop. In Sec. V.A/V.C the 'analytic truth' is obtained by first running the PV-DNN inversion on the mean CFF and then fitting the result with Eq. (31); the pseudo-data and closure target are generated with exactly the same discrete PV operator used as the network's forward layer (Eqs. 25 and 36). Hence the reported ~1e-6 closure error bounds only the network's ability to reconstruct a smoothed version of its own first inversion, and any systematic error in the operator—including the sign inconsistency between Eq. (2) and Eq. (5)/(25)—cancels in the closure test. The paper itself acknowledges that this tuning procedure 'does not remove methodological bias.' Because the extraction from the experimental CFF ensemble is still a genuine fit to input data, the circularity is partial rather than total. Self-citations to the global CFF fit and replica methodology are not treated as load-bearing circularity here: they are external inputs or standard techniques, and the GPD inversion does not feed back into them.

Axiom & Free-Parameter Ledger

6 free parameters · 6 axioms · 0 invented entities

No new physical entities are postulated - no new particles, forces, mediators, or conserved quantities. The differentiable PV layer is a computational construct, not a physical entity. The ledger is dominated by the regularization/envelope priors (lambda_x, beta), the numerical PV implementation (n_PV, N_x), and the pipeline-derived closure truth (a_ijk), plus the standard LO factorization and distribution-identity axioms. The sign inconsistency of Eq. (5) is a derivation error, not an axiom.

free parameters (6)
  • lambda_x = final value not quoted; scanned in [1,5]e-3
    Curvature-regularization weight in the loss (Eqs. 19/26/38). Tuned in closure tests (Sec. V.C); controls the trade-off between CFF fidelity and smoothness, and is a load-bearing prior of the inversion.
  • beta = not stated
    Endpoint-envelope exponent in E(x) = (1-|x|)^beta (Eq. 14). Described as calibrated in closure studies (Sec. VII.C) but the chosen value is never quoted.
  • n_PV = 2
    Principal-value excision width in units of Delta x (Eqs. 11, 23-24). Set by hand; all reported results use n_PV = 2 (Sec. IV.B).
  • N_x = 181
    Uniform x-grid size (Eq. 21). Numerical discretization choice that controls resolution and conditioning of the PV operator.
  • a_ijk (Eq. 31 surrogate coefficients) = fit to the pipeline's own first inversion
    Coefficients of the analytic closure truth H_true, obtained from a fit to the method's own initial GPD extraction (Sec. V.C). They define the benchmark, so the closure test validates the pipeline against itself.
  • DNN width/depth, learning rate = 64x2 units, tanh, Adam eta0 = 2e-2
    Architecture choices stated in Sec. III.B; standard ML settings, not physically constrained.
axioms (6)
  • domain assumption LO leading-twist QCD factorization of the DVCS amplitude, Eq. (2): H = sum_q e_q^2 integral dx [1/(x-xi+i0) + 1/(x+xi-i0)] H_q, neglecting gluon GPDs, higher twist, and the D-term.
    Stated in Sec. II. Load-bearing: the entire PV operator and hence the extraction rests on this factorization being adequate at the kinematics used.
  • standard math Distribution identity (x +/- i0)^-1 = PV(1/x) -/+ i pi delta(x) and the C-even decomposition H^(+) = H(x) - H(-x), Eqs. (3)-(4).
    Used to obtain Eqs. (4)-(5) from Eq. (2). This is exactly the step where the sign inconsistency of Eq. (5) enters.
  • domain assumption The true GPD is smooth enough to be represented by a 2-layer 64-unit tanh network times (1-|x|)^beta, and the minimal-curvature solution is close to the true GPD.
    Network ansatz Eq. (12) and Tikhonov framing Eq. (38). Sec. VII.C admits the envelope and smoothness priors are a controlled but nontrivial source of modeling.
  • ad hoc to paper The closure surrogate H_true (Eq. 31), fit to the pipeline's own first inversion, spans the physically relevant function class for tuning lambda_x.
    Sec. V.C: 'a preliminary GPD model is obtained using the PV-DNN inversion ... fitted using the smooth analytic functional form (31)'. The benchmark truth is pipeline-derived, so the closure test cannot detect systematic pipeline errors.
  • domain assumption Experimental CFF uncertainties are i.i.d. Gaussian with sigma_n from the global CFF model, so replicas via Eq. (20)/(30) represent the posterior.
    Correlations of the Ref [40] CFF fit are not propagated; the paper only partially acknowledges this limitation.
  • domain assumption Trapezoid quadrature with hard excision |x_i +/- xi| < n_PV Delta x (Eqs. 10-11, 21-25) accurately represents the continuum PV integral, including off-grid xi, without an analytic pole correction.
    Never validated against an exact PV evaluation; the closure test reuses the same discrete operator on both sides (Eq. 36), so discretization bias and the O(Delta x) excision error are invisible.

pith-pipeline@v1.3.0-alltime-deepseek · 5735 in / 5997 out tokens · 503557 ms · 2026-08-03T14:06:24.069097+00:00 · methodology

0 comments
read the original abstract

We present a machine-learning method for the nonparametric extraction of generalized parton distributions (GPDs) from Compton form factors (CFFs) constrained by experimental data. The method addresses the longstanding inverse problem posed by the principal-value (PV) linear integral transform with a singular kernel that relates the charge-conjugation-even (C-even) quark GPD $H^{(+)}$ to the real part of the deeply virtual Compton scattering (DVCS) amplitude. Our approach constructs a differentiable representation of the Quantum Chromodynamics (QCD) PV kernel and embeds it as a fixed, physics-preserving layer inside a neural network that parameterizes the GPD $H^{(+)}(x,\xi,t,Q^{2})$ itself. The model enforces exact oddness in $x$, implements endpoint suppression, and includes curvature-based regularization that stabilizes the inversion in kinematically ill-conditioned regions. A Monte Carlo ensemble of CFFs, obtained from a global neural-network fit to unpolarized DVCS measurements with propagated experimental uncertainties, serves as input to a replica ensemble of GPD networks, yielding a fully probabilistic extraction of $H^{(+)}$ over the phase space. We demonstrate the method using a global determination of $\mathrm{Re}\,\mathcal{H}$ for Jefferson Lab measurements, and present a direct neural-network reconstruction of three-dimensional GPD surfaces $H^{(+)}(x_{0},\xi,t,Q_{0}^{2})$ obtained from experimental CFF inputs. This work establishes a flexible, scalable, and model-independent strategy for extracting multidimensional hadronic structure from current and future DVCS data and other GPD-related processes.

Figures

Figures reproduced from arXiv: 2512.21761 by Dima Watkins, Dustin Keller, Ishara Fernando.

Figure 1
Figure 1. Figure 1: FIG. 1. Reconstructed one-dimensional slice of [PITH_FULL_IMAGE:figures/full_fig_p008_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: FIG. 2. Reconstructed one-dimensional slice of [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: FIG. 3. Three-dimensional GPD surface [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: FIG. 4. Comparison of reconstructed surfaces [PITH_FULL_IMAGE:figures/full_fig_p010_4.png] view at source ↗

discussion (0)

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

Forward citations

Cited by 2 Pith papers

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

  1. Constraining DVCS Compton Form Factors Using Lattice QCD informed Neural Network

    hep-ph 2026-06 unverdicted novelty 5.0

    A neural network framework informed by lattice QCD uses all-order dispersion relations to significantly constrain both real and imaginary parts of Compton Form Factors extracted from DVCS proton data.

  2. Generalized parton distributions of a deuteron in an AdS/QCD hard-wall model

    hep-ph 2026-06 unverdicted novelty 3.0

    Deuteron GFFs and GPDs calculated in hard-wall AdS/QCD agree in momentum dependence with soft-wall results and match experimental gravitational mean square radius.

Reference graph

Works this paper leans on

58 extracted references · 17 linked inside Pith · cited by 2 Pith papers

  1. [1]

    The solid curve denotes the ensemble mean and the shaded band indicates the1σ uncertainty inferred from the replica variance

    = ( median(ξ),median(t),median(Q 2) ) . The solid curve denotes the ensemble mean and the shaded band indicates the1σ uncertainty inferred from the replica variance. Several features are immediately visible: (i) the exact oddness inx enforced by the architecture, (ii) the sup- pression near|x|→ 1due to the endpoint envelope, and (iii) a smooth behavior ne...

  2. [2]

    Unpolarized cross sections, beam–spin and beam–charge asymmetries, target–spin (lon- gitudinal and transverse) asymmetries, and double–spin asymmetries, together with polarized and longitudinally separated cross sections

  3. [3]

    Next-generation CFF extractions that use DNNs to build global models with low uncertainty across phase space

  4. [4]

    Future high-precision measurements with multiple observables taken simultaneously

  5. [5]

    the fullH

    Access to raw, unbinned data for optimal sampling of experimental features and covariances. An important practical extension of the present method is to treat the real and imaginary parts of a given CFF on the same footing. At leading order, the imaginary part ofH fixes H (+) on the cross-over linex =ξ, while the real part is obtained from the PV integral...

  6. [6]

    Müller, D

    D. Müller, D. Robaschik, B. Geyer, F. M. Dittes, and J. Horejsi, Fortsch. Phys.42, 101 (1994), arXiv:hep- ph/9812448 [hep-ph]

    arXiv 1994

  7. [7]

    A. V. Radyushkin, Phys. Rev. D56, 5524 (1997), hep- ph/9704207

    arXiv 1997

  8. [8]

    X. D. Ji, Phys. Rev. D55, 7114 (1997), hep-ph/9609381

    Pith/arXiv arXiv 1997

  9. [9]

    Ji, Phys

    X. Ji, Phys. Rev. Lett.78, 610 (1997)

    1997

  10. [10]

    J. C. Collins and A. Freund, Phys. Rev. D59, 074009 (1999), hep-ph/9801262

    Pith/arXiv arXiv 1999

  11. [11]

    Ji and J

    X. Ji and J. Osborne, Phys. Rev. D58, 094018 (1998), hep-ph/9801260

    Pith/arXiv arXiv 1998

  12. [12]

    A. V. Belitsky and A. V. Radyushkin, Phys. Rept.418, 1 (2005), hep-ph/0504030

    Pith/arXiv arXiv 2005

  13. [13]

    P. C. Hansen,Rank-Deficient and Discrete Ill-Posed Prob- lems: Numerical Aspects of Linear Inversion(Society for Industrial and Applied Mathematics, Philadelphia, PA, 1998)

    1998

  14. [14]

    A. N. Tikhonov and V. Y. Arsenin,Solutions of Ill-posed Problems(Winston & Sons, Washington, DC, 1977) origi- nal Russian edition: Nauka, Moscow (1974)

    1977

  15. [15]

    Müller, Few Body Syst.55, 317 (2014), arXiv:1405.2817 [hep-ph]

    D. Müller, Few Body Syst.55, 317 (2014), arXiv:1405.2817 [hep-ph]

    Pith/arXiv arXiv 2014

  16. [16]

    Diehl, Eur

    M. Diehl, Eur. Phys. J. C25, 223 (2002), arXiv:hep- ph/0205208 [hep-ph]

    arXiv 2002

  17. [17]

    Diehl, Phys

    M. Diehl, Phys. Rept.388, 41 (2003), arXiv:hep- ph/0307382 [hep-ph]

    arXiv 2003

  18. [18]

    Kumerički and D

    K. Kumerički and D. Müller, EPJ A52, 157 (2016), 1602.02763

    Pith/arXiv arXiv 2016

  19. [19]

    Guidal, EPJ A37, 319 (2008), 0711.3743

    M. Guidal, EPJ A37, 319 (2008), 0711.3743

    Pith/arXiv arXiv 2008

  20. [20]

    Moutardeet al., EPJ C78, 890 (2018), 1807.07620

    H. Moutardeet al., EPJ C78, 890 (2018), 1807.07620

    Pith/arXiv arXiv 2018

  21. [21]

    Kumerički, D

    K. Kumerički, D. Müller, and A. Schäfer, Journal of High Energy Physics2011, 73 (2011), arXiv:1106.2808 [hep- ph]

    Pith/arXiv arXiv 2011

  22. [22]

    M. Čuić, K. Kumerički, and A. Sch"afer, Physical Review Letters125, 232005 (2020)

    2020

  23. [23]

    Moutarde, B

    H. Moutarde, B. Pire, L. Szymanowski, and J. Wagner, EPJ C79, 614 (2019), 1905.02089

    Pith/arXiv arXiv 2019

  24. [24]

    Calero Díaz and D

    L. Calero Díaz and D. Keller, Physical Review D 10.1103/PRD/sb63-sfdt (2025), arXiv:2509.18331 [nucl- ex]

    arXiv 2025

  25. [25]

    R. D. Ball, L. Del Debbio, S. Forte, A. Guffanti, J. I. Latorre, A. Piccione, J. Rojo, and M. Ubiali, Nuclear Physics B809, 1 (2009)

    2009

  26. [26]

    R. D. Ball, L. Del Debbio, S. Forte, A. Guffanti, J. I. Latorre, A. Piccione, J. Rojo, and M. Ubiali, Nuclear Physics B816, 293 (2009)

    2009

  27. [27]

    R. D. Ball, V. Bertone, S. Carrazza, C. S. Deans, L. Del Debbio, S. Forte, A. Guffanti, N. P. Hartland, J. I. Latorre, J. Rojo, and M. Ubiali, Nuclear Physics B867, 244 (2013)

    2013

  28. [28]

    T. N. Collaboration, R. D. Ball, V. Bertone, S. Carrazza, C. S. Deans, L. Del Debbio, S. Forte, A. Guffanti, N. P. Hartland, J. I. Latorre, J. Rojo, and M. Ubiali, Journal of High Energy Physics2015, 1 (2015), arXiv:1410.8849 [hep-ph]

    Pith/arXiv arXiv 2015

  29. [29]

    E. R. Nocera, R. D. Ball, S. Forte, G. Ridolfi, and J. Rojo, Nuclear Physics B887, 276 (2014)

    2014

  30. [30]

    Del Debbio, T

    L. Del Debbio, T. Giani, J. Karpie, K. Orginos, A. Radyushkin, and S. Zafeiropoulos, J. High Energ. Phys. 02, 138, arXiv:2010.03996 [hep-ph]

    Pith/arXiv arXiv 2010

  31. [31]

    Dutrieux, H

    H. Dutrieux, H. Moutarde, and P. Sznajder, Eur. Phys. J. C82, 300 (2022), arXiv:2202.06888 [hep-ph]

    Pith/arXiv arXiv 2022

  32. [32]

    Ji, Journal of Physics G: Nuclear and Particle Physics 24, 1181 (1998)

    X. Ji, Journal of Physics G: Nuclear and Particle Physics 24, 1181 (1998)

    1998

  33. [33]

    Radyushkin, Physics Letters B449, 81 (1999)

    A. Radyushkin, Physics Letters B449, 81 (1999)

    1999

  34. [34]

    M. V. Polyakov and C. Weiss, Phys. Rev. D60, 114017 (1999)

    1999

  35. [35]

    P. V. Pobylitsa, Phys. Rev. D67, 034009 (2003)

    2003

  36. [36]

    A. V. Radyushkin, Phys. Rev. D59, 014030 (1998)

    1998

  37. [37]

    B. Pire, J. Soffer, and O. Teryaev, Eur. Phys. J. C8, 103 (1999)

    1999

  38. [38]

    Diehl, T

    M. Diehl, T. Feldmann, R. Jakob, and P. Kroll, Nuclear Physics B596, 33 (2001)

    2001

  39. [39]

    Diehl, T

    M. Diehl, T. Feldmann, R. Jakob, and P. Kroll, Nuclear Physics B605, 647 (2001)

    2001

  40. [40]

    P. V. Pobylitsa, Phys. Rev. D65, 077504 (2002). 14

    2002

  41. [41]

    P. V. Pobylitsa, Phys. Rev. D65, 114015 (2002)

    2002

  42. [42]

    P. V. Pobylitsa, Phys. Rev. D66, 094002 (2002)

    2002

  43. [43]

    P. V. Pobylitsa, Phys. Rev. D67, 094012 (2003)

    2003

  44. [44]

    P. V. Pobylitsa, Phys. Rev. D70, 034004 (2004)

    2004

  45. [45]

    B. B. Le and D. Keller, arXiv preprint arXiv:2504.15458 10.48550/arXiv.2504.15458 (2025), arXiv:2504.15458 [cs.LG]

    Pith/arXiv arXiv doi:10.48550/arxiv.2504.15458 2025

  46. [46]

    A. V. Belitsky and D. Müller, Phys. Rev. D82, 074010 (2010)

    2010

  47. [47]

    Anselet al., in29th ACM International Conference on Architectural Support for Programming Languages and Op- erating Systems, Volume 2 (ASPLOS ’24)(ACM, 2024)

    J. Anselet al., in29th ACM International Conference on Architectural Support for Programming Languages and Op- erating Systems, Volume 2 (ASPLOS ’24)(ACM, 2024)

    2024

  48. [48]

    U. D. of Energy (USDOE),A New Era of Discovery: The 2023 Long Range Plan for Nuclear Science, Tech. Rep. (US Department of Energy (USDOE), Washington, DC (United States). Office of Science, 2023)

    2023

  49. [49]

    R. A. K. et al., Snowmass 2021 white paper: Electron ion collider for high energy physics (2022), arXiv:2203.13199 [hep-ph]

    Pith/arXiv arXiv 2021

  50. [50]

    E. C. Aschenauer, S. Fazio, K. Kumerički, and et al., Journal of High Energy Physics2013, 93 (2013)

    2013

  51. [51]

    A. A. et al., Electron ion collider: The next qcd fron- tier - understanding the glue that binds us all (2014), arXiv:1212.1701 [nucl-ex]

    Pith/arXiv arXiv 2014

  52. [52]

    Kumerički, D

    K. Kumerički, D. Müller, and K. Passek-Kumerički, Nu- clear Physics B794, 244 (2008)

    2008

  53. [53]

    Diehl and D

    M. Diehl and D. Ivanov, Eur. Phys. J. C52, 919 (2007)

    2007

  54. [54]

    Kumerički, D

    K. Kumerički, D. Müller, and K. Passek-Kumerički, Eur. Phys. J. C58, 193 (2008)

    2008

  55. [55]

    H. W. Engl, M. Hanke, and A. Neubauer,Regularization of Inverse Problems, Mathematics and Its Applications, Vol. 375 (Kluwer Academic Publishers, Dordrecht, 1996)

    1996

  56. [56]

    Keller, arXiv preprint arXiv:2509.11456 (2025), arXiv:2509.11456 [hep-ph]

    D. Keller, arXiv preprint arXiv:2509.11456 (2025), arXiv:2509.11456 [hep-ph]

    arXiv 2025

  57. [57]

    I. P. Fernando and D. Keller, Phys. Rev. D108, 054007 (2023)

    2023

  58. [58]

    I. P. Fernando and D. Keller, arXiv preprint arXiv:2510.17243 10.48550/arXiv.2510.17243 (2025), arXiv:2510.17243 [hep-ph]

    Pith/arXiv arXiv doi:10.48550/arxiv.2510.17243 2025