Regression on high-dimensional phase space

Abstract

Machine learning tasks in particle physics are typically expressed in terms of regression, or relaxation toward some optimal solution as defined by an objective or loss function. In this paper, we introduce a large multiplicity expansion for analyzing regression tasks, which enables significant analytic understanding. For concreteness, we study regression with energy flow polynomials (EFPs), a complete linear basis of infrared and collinear safe observables on Lorentz invariant phase space. While the expectation value of any EFP scales like a constant in the large multiplicity limit, we prove that the covariance on uniform phase space of any two EFPs is highly suppressed and scales like the inverse of the squared multiplicity. This very unintuitive result relies on detailed cancellation guaranteed by the interplay of permutation invariance, momentum conservation, and rotational invariance. This scaling of the covariance with multiplicity further implies that the coefficients in an expansion of EFPs must scale quadratically in multiplicity. We present an example of regression with EFPs as applied to $e^+e^-\to$ hadrons events, and show how the regression can be solved analytically through the complementary expansions of the large multiplicity limit and of familiar fixed-order perturbation theory of QCD.