[Paper Review] Progress in the NNPDF global analyses of proton structure

Juan Rojo|arXiv (Cornell University)|Apr 19, 2021

Particle physics theoretical and experimental studies8 references4 citations

TL;DR

This paper presents NNPDF4.0, a next-generation global fit of proton parton distribution functions (PDFs) that integrates LHC Run II data at 13 TeV, employs advanced machine learning techniques including stochastic gradient descent and automated hyperparameter optimization, and rigorously validates PDF uncertainties via closure and future tests. The result is a highly precise, basis-independent PDF determination with reduced uncertainties and improved constraints on gluon, anti-quark, and charm PDFs.

ABSTRACT

I review recent progress in the NNPDF global analyses of parton distributions (PDFs) focusing on developments contributing to its new upcoming release: NNPDF4.0. The NNPDF4.0 determination represents unprecedented progress in three main directions: i) the systematic inclusion of LHC Run II data at 13 TeV and of new processes from dijets to single top distributions, ii) the deployment of state-of-the-art machine learning algorithms, from automated hyperparameter optimisation to stochastic gradient descent training; and iii) the complete statistical validation of PDF uncertainties, both in the data and extrapolation regions, by means of closure and future tests. Other methodological improvements in NNPDF4.0 include strict PDF positivity, integrability constraints at small-$x$, and deuteron and heavy nuclear corrections. I present representative results from NNPDF4.0 and assess its impact on open issues such as the light anti-quark asymmetry and the charm content of protons.

Motivation & Objective

• To develop a next-generation global PDF fit, NNPDF4.0, incorporating LHC Run II data and novel processes.
• To implement state-of-the-art machine learning algorithms to improve fitting efficiency and accuracy.
• To ensure robust statistical validation of PDF uncertainties through closure and future tests.
• To enforce theoretical constraints such as PDF positivity and integrability at small-x.
• To assess open questions in proton structure, including light anti-quark asymmetry and intrinsic charm.

Proposed method

• Utilizes neural networks to parametrize PDFs in both evolution and flavor bases, ensuring basis independence.
• Applies stochastic gradient descent with automated hyperparameter optimization to accelerate training and reduce minimization inefficiencies.
• Incorporates LHC Run II data including dijet, single top, direct photon, and W+jets production cross-sections.
• Implements strict positivity constraints on MS PDFs and integrability conditions for non-singlet quark combinations at small-x.
• Corrects for deuteron and heavy nuclear effects in neutrino-nucleus data using nNNPDF2.0 as input.
• Validates uncertainties via closure tests (fitting to pseudo-data) and future tests (forecasting novel datasets).

Experimental results

Research questions

• RQ1 How do LHC Run II dijet and single top distributions constrain the gluon PDF compared to previous fits?
• RQ2 To what extent do machine learning techniques like stochastic gradient descent improve the convergence and precision of PDF fits?
• RQ3 Can the statistical interpretation of PDF uncertainties be rigorously validated using closure and future tests?
• RQ4 What is the impact of new data on the light anti-quark asymmetry $\bar{d}/\bar{u}$ and the charm PDF at low scales?
• RQ5 Does the data support a valence-like or intrinsic charm component in the proton wave function?

Key findings

• The inclusion of dijet cross-sections at 13 TeV leads to a consistent and improved description of jet data without tailored decorrelation models.
• NNPDF4.0 achieves a 10-fold speed-up in per-replica training time and reduces PDF uncertainties by eliminating inefficiencies from genetic algorithms.
• The gluon-gluon luminosity $\mathcal{L}_{gg}$ at $\sqrt{s}=14$ TeV shows a suppression at $m_X \sim 100$ GeV, enhancement above 1 TeV, and suppression beyond 4 TeV, with reduced errors above 500 GeV.
• The light anti-quark ratio $\bar{d}/\bar{u}$ in NNPDF4.0 is in good agreement with SeaQuest measurements, confirming the observed sea quark asymmetry.
• The charm PDF at $Q=1.65$ GeV exhibits a valence-like structure, consistent with an intrinsic charm component, and is largely insensitive to the inclusion of EMC charm data.
• The analysis demonstrates full consistency between fits performed in the evolution basis and the flavor basis, confirming basis independence and methodological robustness.