Neural-Network extraction of TMDs with SIDIS data

* · 2026

Read Neural-Network extraction of TMDs with SIDIS data by * (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2606-28218.txt.

Introduction

Matteo Cerutti reports the first global extraction of transverse-momentum-dependent distributions, maps of how quarks move sideways inside protons, that uses a neural network to describe their unknown shape. Earlier maps relied on fixed formulas, which can bias the result toward the formula chosen.

Cerutti fits two kinds of experiment together: Drell-Yan, where quark pairs produce leptons in collisions, and semi-inclusive deep inelastic scattering, where a struck quark is also observed. Adding the scattering data made the extracted distributions broader, meaning quarks carry more sideways motion than collision data alone suggested. Uncertainties shrank compared with a collisions-only network fit, though they stayed larger than those from traditional formulas, which Cerutti reads as a more honest accounting.

Catalog record

CREATOR*
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE11 KB
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2606.28218

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Neural-Network extraction of TMDs with SIDIS data * Abstract A first global analysis of unpolarized Transverse-Momentum-Dependent (TMD) distributions based on a neural-network (NN) parametrization is presented. Drell-Yan (DY) and semi-inclusive deep inelastic scattering (SIDIS) data are simultaneously included at next-to-next-to-next-to-leading logarithmic (N 3 LL) accuracy. The results indicate that the inclusion of SIDIS data leads to broader unpolarized TMD PDFs compared to a DY-only NN extraction. The associated uncertainties are reduced with respect to the DY-only case, while remaining larger than the ones obtained using traditional models. These results demonstrate the potential of flexible NN parametrizations in reducing model dependence and provide guidance for future high-precision measurements at Jefferson Lab and the Electron-Ion Collider. [a]Matteo Cerutti 1 Introduction Transverse-Momentum-Dependent (TMD) distributions provide a three-dimensional map of hadron structure in momentum space. Drell-Yan (DY) lepton-pair production in hadron-hadron collisions provides information on quark TMD parton distribution functions (TMD PDFs) [ 1 ] , while semi-inclusive Deep Inelastic Scattering (SIDIS) measurements provide simultaneous access to TMD PDFs and fragmentation functions (TMD FFs) [ 2 ] . Gluon TMDs can be accessed through other processes, like Higgs and/or quarkonia production [ 3 ] . Despite significant progress in improving the perturbative accuracy and increasing the global dataset, there are tensions between current extractions of unpolarized quark TMD PDFs from different global fits [ 4 , 5 , 6 ] . This is in part due to the limited flexibility of standard parametrizations and motivates the exploration of more flexible approaches in modeling the nonperturbative part of TMDs. A proof-of-concept study on DY data has recently demonstrated the feasibility of applying Neural Networks (NNs) to reduce this parameterization bias [ 7 ] . In this contribution, I present an extension of Ref. [ 7 ] to a global TMD analysis. I focus in particular on the impact of SIDIS measurements, assessing their potential to constrain TMD distributions in a global framework. These results are particularly relevant for maximizing the physics potential of forthcoming experimental programs at future facilities, such as SoLID [ 8 ] at Jefferson Lab and the Electron-Ion Collider [ 9 ] . 2 Framework In the DY process h_{A}+h_{B}\rightarrow\ell^{+}+\ell^{-}+X two hadrons with mass M and four-momenta P_{A} and P_{B} collide with center-of-mass energy squared s=(P_{A}+P_{B})^{2} , and produce a lepton pair with four-momentum q and invariant mass Q . If Q\gg M and the transverse momentum \bm{q}_{T} relative to the collision axis is such that |\bm{q}_{T}|\equiv q_{T}\ll Q , the differential cross section reads \begin{split}&\frac{d\sigma^{\text{DY}}}{dq_{T}\,dy\,dQ}=\frac{8\pi\alpha^{2}q_{T}}{9Q^{3}}\,{\cal P}\,x_{A}\,x_{B}\,{\cal H}^{\text{DY}}(Q,\mu)\,\sum_{a}c_{a}(Q^{2})\\ &\times\int_{0}^{\infty}\!\!db_{T}\,b_{T}J_{0}\big(b_{T}q_{T}\big)\hat{f}_{1}^{a}(x_{A},b_{T}^{2};\mu,\zeta_{A})\,\hat{f}_{1}^{\bar{a}}(x_{B},b_{T}^{2};\mu,\zeta_{B}),\end{split} (1) where \alpha is the electromagnetic coupling, y=\ln\sqrt{(q_{0}+q_{z})/(q_{0}-q_{z})} is the lepton-pair rapidity, {\cal P} is the phase-space-reduction factor due to possible fiducial cuts on the final-state leptons. x_{A,B}=Qe^{\pm y}/\sqrt{s} are the longitudinal momentum fractions of the quarks in the initial state, {\cal H}^{\text{DY}} is a perturbative hard factor, and the sum runs over all active quark flavors a with c_{a} the quark electroweak charges. In Eq. ( 1 ), \hat{f}^{a}_{1} denotes the Fourier transform of the unpolarized TMD PDF of quark flavor a . It depends on the longitudinal momentum fraction x and on b_{T}=|\bm{b}_{T}| , Fourier conjugate to the quark intrinsic transverse momentum \bm{k}_{\perp} . Also, it depends on the UV renormalization scale \mu and on the rapidity scale \zeta (with \zeta_{A}\zeta_{B}=Q^{4} ) [ 10 ] . We also consider the SIDIS process \ell+h\rightarrow\ell^{\prime}+h^{\prime}+X where a lepton and a hadron with mass M and four-momenta l and P collide, respectively, and inclusively produce a hadron h^{\prime} via the exchange of a vector boson with total four-momentum q and invariant mass q^{2}=-Q^{2} . When Q^{2}\gg M^{2} and q_{T}\ll Q , the differential cross section can be written as \begin{split}&\frac{d\sigma^{\text{SIDIS}}}{dq_{T}\,dx\,dz\,dQ}=\frac{16\pi^{2}\,\alpha^{2}\,z^{2}\,q_{T}}{x\,Q^{3}}\,\bigg(1+\bigg(1-\frac{Q^{2}}{xs}\bigg)^{2}\bigg)\,\frac{x}{2\pi}\,{\cal H}^{\text{SIDIS}}(Q,\mu)\sum_{a}e_{a}^{2}\\ &\times\int_{0}^{\infty}\!\!db_{T}\,b_{T}J_{0}\big(b_{T}q_{T}\big)\hat{f}_{1}^{a/h}(x,b_{T}^{2};\mu,\zeta_{A})\,\hat{D}_{1}^{a\rightarrow h^{\prime}}(z,b_{T}^{2};\mu,\zeta_{B}),\end{split} (2) where x and z are the longitudinal momentum fractions in the incoming parton (w.r.t the parent hadron) and the outgoing hadron (w.r.t. the struck quark), respectively, and \hat{D}_{1}^{a\rightarrow h^{\prime}} is the TMD FF of a parton a fragmenting into a hadron h^{\prime} . The structure of a TMD PDF is the following: \hat{f}_{1}^{a/h}(x,b_{T}^{2};\mu,\zeta)=[C\otimes f_{1}]^{a}(x,b_{*}(b_{T}),\mu_{b_{*}})\exp\bigg\{\int_{\mu_{b_{*}}}^{\mu}\frac{d\mu^{\prime}}{\mu^{\prime}}\gamma(\mu^{\prime},\zeta)\bigg\}\bigg(\frac{\zeta}{\mu^{2}_{b_{*}}}\bigg)^{K(\mu_{b_{*}})}\,f_{\rm NP}(x,b_{T},\zeta)\, (3) where C is the perturbatively calculable coefficient function providing the matching onto collinear PDFs f_{1} at small b_{T}\ll 1/\Lambda_{QCD} , \mu_{b}=2e^{-\gamma_{E}}/b_{T} is the natural choice for the TMD initial scale, \gamma and K are the anomalous dimensions of the evolution in \mu and \zeta , respectively. In order to avoid hitting the Landau pole when b_{T} becomes large, the variable b_{T} is replaced with b_{*}(b_{T}) , known as “ b_{*} prescription”, which provides the transition between perturbative and nonperturbative physics [ 11 , 12 ] . The information on the genuine large- b_{T} region is then encoded in a nonperturbative function f_{\rm NP} , which parametrizes long-distance QCD dynamics. Perturbative ingredients are included consistently up to next-to-next-to-next-to-leading logarithmic accuracy (N 3 LL). Collinear PDFs and FFs are provided through the LHAPDF library, using MSHT and DSS sets, respectively. Their uncertainties are incorporated consistently. SIDIS predictions include normalization factors computed prior to the fit according to Ref. [ 6 ] . Nonperturbative components of TMD PDFs and FFs are parametrized as \begin{split}&f_{NP}(x,b_{T};\zeta)=\frac{\mathbb{NN}(x,b_{T})}{\mathbb{NN}(x,0)}\exp\left[-g_{2}^{2}b_{T}^{2}\log\left(\frac{\zeta}{Q_{0}^{2}}\right)\right]\,,\\ &D_{NP}(z,b_{T};\zeta)=\frac{\mathbb{NN}(z,b_{T})}{\mathbb{NN}(z,0)}\exp\left[-g_{2}^{2}b_{T}^{2}\log\left(\frac{\zeta}{Q_{0}^{2}}\right)\right]\,,\end{split} (4) where f_{\rm NP} is split into an “intrinsic” nonperturbative part, entirely parametrized by a NN, and the nonperturbative contribution to evolution, encoded in the exponential function (with Q_{0}=1 GeV). The two different NNs have architecture [2,10,1] , i.e. with two inputs corresponding to x (or z for D_{NP} ) and b_{T} , 10 hidden nodes, and one output node. The activation function is a sigmoid. Note that f_{NP}\to 1 for b_{T}\to 0 and f_{\rm NP}\ll 1 for large b_{T} . In this setup, the number of free parameters is 83, 41 for f_{\rm NP} , 41 for D_{NP} , and one ( g_{2} ) for the evolution. To mitigate the risk of overfitting, the dataset is divided into training and validation samples of the same size. 3 Results We discuss here the results of the simultaneous extraction of TMD PDFs and FFs obtained with the NangaParbat fitting framework [ 13 ] . The global dataset includes low-invariant-mass DY measurements from Fermilab, high-energy Z -boson q_{T} -differential cross sections from RHIC and LHC, and SIDIS multiplicities from HERMES and COMPASS collaborations. Kinematic cuts are imposed to fulfill the requirements of TMD factorization. For DY data, the analysis is restricted to data at q_{T}/Q<0.2 , while the cut on SIDIS data is more involved (see Refs. [ 5 , 6 ] ). The total number of data points in the fit is 2029 (482 for DY, 1547 for SIDIS). The propagation of experimental errors in the fit is performed via the bootstrap method, namely by fitting 100 Monte Carlo replicas of the experimental data. The values of the best-fit parameters are obtained by minimizing a \chi^{2} function. The quality of the fit is represented by the \chi^{2} value normalized by the number of data points N_{\rm dat} . The resulting value is \chi^{2}/N_{\rm dat}=1.01 , indicating that the analysis is able to simultaneously describe DY and SIDIS datasets. In Fig. 1 , we compare the extracted quantities with previous analyses. In the left panel, we compare the results of this NN fit (blue band), the NNDY [ 7 ] fit (red band) and the MAP22 [ 6 ] fit (green band) for the

Read the complete text