Speaker
Marco Zaccheddu
(Jefferson Lab)
Description
We present a new nonparametric framework for the Bayesian inference of TMD parton distributions. By discretizing the impact parameter space into "pixels", we eliminate the biases of traditional functional forms. To sample the complex high-dimensional posterior, we employ a hybrid Normalizing Flow-driven Metropolis-Hastings algorithm. Using Singular Value Decomposition (SVD), we formally identify the resolution limits of the integral transform, introducing the concept of unobservable "null TMDs". Through multi-scale closure tests within the CSS formalism, we demonstrate how scale evolution combined with generative AI can break these degeneracies, enabling rigorous and unbiased 3D imaging of the nucleon.
Author
Marco Zaccheddu
(Jefferson Lab)