χ² and Beyond: Robust Statistics for Hadronisation Tuning

by James Whitehead

Friday 13 Nov 2026, 13:00 → 14:00 Europe/London

Bohr Conference Room

Monte Carlo Event Generators in use for LHC phenomenology combine perturbative modelling of hard processes with semi-empirical models of non-perturbative physics, including hadronisation and the underlying event. These models introduce parameters that must be calibrated, or 'tuned', to data. Standard tuning approaches minimise a weighted χ² objective, so large residuals can exert disproportionate influence. Yet such discrepancies are expected to arise from imperfect phenomenological models, and may reflect structural limitations that cannot be resolved by adjusting the parameters. Consequently, a tune can be driven by precisely the measurements which the model is least able to describe.

The concept of ‘robustness’, developed in the statistical framework of Huber and others, offers a principled way to limit this sensitivity. Robust statistics concerns the stability of inference under departures from assumed statistical distributions, including contamination by observations drawn from a different distribution. We compare χ² with robust alternatives, both through loss-function minimisation and Bayesian inference with MCMC sampling. We apply these approaches to the final stage of a recent Herwig 7 tune using the Pythia Lund string model, and assess how these statistical choices affect the inferred parameters, their uncertainties, and sensitivity to tuning weights.