Neural network determination of parton distributions: the nonsinglet case

TL;DR

The paper demonstrates a novel neural-network Monte Carlo framework to determine the nonsinglet quark distribution from deep-inelastic scattering data with unbiased uncertainty estimates. By coupling Monte Carlo data replicas to neural-network parametrizations and an efficient N-space/x-space evolution strategy, it delivers a statistically faithful parton distribution up to NNLO, including rigorous data-driven validation and stability checks. The work shows good agreement with data while revealing larger small-x uncertainties relative to traditional fits, and it emphasizes the method's potential to remove parametrization bias in global PDF determinations. These results establish the viability of applying neural networks to full PDF determinations and outline a clear path toward a comprehensive global neural PDF fit.

Abstract

We provide a determination of the isotriplet quark distribution from available deep--inelastic data using neural networks. We give a general introduction to the neural network approach to parton distributions, which provides a solution to the problem of constructing a faithful and unbiased probability distribution of parton densities based on available experimental information. We discuss in detail the techniques which are necessary in order to construct a Monte Carlo representation of the data, to construct and evolve neural parton distributions, and to train them in such a way that the correct statistical features of the data are reproduced. We present the results of the application of this method to the determination of the nonsinglet quark distribution up to next--to--next--to--leading order, and compare them with those obtained using other approaches.