Orateur
Description
Generalized parton distributions (GPDs) encode the three-dimensional partonic structure of the nucleon, but exclusive processes such as deeply virtual Compton scattering constrain them only in part of their kinematic domain — the so-called DGLAP region — while leaving the complementary ERBL region essentially undetermined by data. A standard way to relate the two is through the double distribution: a single underlying function whose Radon transform generates the GPD, tying the measured and unmeasured regions together through Lorentz covariance. We represent this double distribution as a neural field — a coordinate network encoding it continuously — and enforce the physical support, symmetry, and polynomiality by construction. Trained on Compton form factor pseudodata with realistic experimental coverage, this approach reproduces the input GPD across both regions, while a direct parametrization of the GPD fails in the unmeasured region — quantifying the information carried from measured to unmeasured kinematics by the double-distribution structure. We then ask how much of the reconstruction is fixed by data and how much by the modelling assumptions. Using a Jacobian analysis, we identify the directions in model space that deform the GPD while leaving every predicted observable unchanged — deformations the data are blind to by construction — and calibrate their size to the tolerance of the data. The same framework quantifies how much future measurements, in particular at the Electron-Ion Collider, would shrink this residual freedom, providing a data-driven, assumption-aware uncertainty for GPD extractions.
| Working Group | WG1. Nucleons and Hadrons |
|---|---|
| Type of oral contribution | Theory |