Background modeling and signal estimation using Gaussian Processes in the H→yy channel (2019)

Abstract
The signal to background ratio for Higgs decaying to 2 photons is small, therefore, powerful estimation of the background is needed to accurately measure the signal. Since the underlying physical function is unknown, various functional parameterizations are considered for the background estimation.  When the number of events increases, the relative uncertainty decreases.  This might give rise to some previously hidden features of the distribution, leading to the need for re-estimation of the background to avoid creating spurious signal. The current process is lengthy and awkward, and therefore, this study has focused on investigating Gaussian Processes (GP) as a new method for estimating the background and signal distributions. GP is a machine learning method that does not depend on a specific parametric function and could therefore be employed in numerous scientific areas. Herein, it has been shown that GP manages to find the underlying function for a large number of toy data set, and it proved to be independent of the integrated luminosity. From bias tests performed, it was shown that bias generated during fitting, does not scale with the luminosity relative to the expected signal for a standard model Higgs boson. It was further shown that the background component of the GP does not model a signal when this is injected into the testing distributions. Last, it was found that a signal estimation using the GP's internal parameters was unsuccessful. However, using the linearity of the GP, it proved possible to estimate the number of signal events in a
distribution.