Estimating information loss in LHC simulations: how to tackle the curse of dimensionality

Abstract

In this project, we study the computationally challenging task of estimating the Kullback-Leibler divergence for high-dimensional probability distributions from particle physics. Our approach is based on using a trained classifier (a boosted decision tree) as a tool for dimensional reduction. As an interesting and challenging test case, we study simulated kinematic distributions for the production of supersymmetric particles at the Large Hadron Collider. We estimate the Kullback-Leibler divergence between kinematic distributions simulated at leading order and at next-to-leading order in perturbation theory, and find divergences of the order 10^-2 bits for the studied examples.