Nonperturbative TMD distributions leave the neural-network black box: symbolic regression yields a closed form with 9 constants
Synopsis
Nonperturbative transverse-momentum-dependent distributions have mostly existed as neural-network fits that cannot be written as analytic expressions. This work first trains a factorized neural network directly against Drell–Yan cross-section data from fixed-target, Tevatron, RHIC and LHC experiments at next-to-next-to-next-to-leading logarithmic accuracy, then applies symbolic regression to each network component and selects, on a Pareto front of expression complexity and experimental fit, a closed-form function with 9 free numerical constants covering more than 482 data points. A compact formula requires a non-trivial cross term between the longitudinal momentum fraction and the transverse-impact-parameter dependence, which encodes a mild correlation between them.
Interpretation
The nonperturbative transverse-momentum-dependent parton distribution function now has a closed-form analytical parametrization instead of existing only as a numerical neural-network fit. The prior route let a neural network fit Drell–Yan data directly, giving a flexible but unreadable distribution that could not be written as a compact analytic expression. The closed form is supported by Drell–Yan cross-section data from fixed-target, Tevatron, RHIC and LHC experiments, fitted at next-to-next-to-next-to-leading logarithmic accuracy and covering more than 482 data points.
The approach splits the fit into two steps: the network learns the data, then symbolic regression translates the network components into formulas. A factorized neural network decomposes the overall fit into separate network components, and symbolic regression is then applied to each component to search for compact analytical expressions. The final formula is chosen from a Pareto front in the space of expression complexity and experimental fit, contains 9 free numerical constants, and keeps a non-trivial cross term to stay compact.
The retained cross term carries physical meaning: it describes a mild correlation between the longitudinal momentum fraction and the transverse-impact-parameter dependence. Without that cross term the formula cannot remain both compact and descriptive of the data. The cross term encodes a mild correlation between the longitudinal momentum fraction and the transverse-impact-parameter dependence, a non-trivial structural ingredient of the fit.
Perspective
The demonstrated scope is the nonperturbative TMD parton distribution function of unpolarized quarks, using Drell–Yan cross-section data from fixed-target, Tevatron, RHIC and LHC experiments at next-to-next-to-next-to-leading logarithmic accuracy.
The final formula is selected from a Pareto front in expression complexity and experimental fit and contains 9 free numerical constants; a non-trivial cross term is required for compactness, indicating a trade-off between compactness and descriptive power.
