Learning to discover: expressive Gaussian mixture models for multi-dimensional simulation and parameter inference in the physical sciences

Abstract We show that density models describing multiple observables with (1) hard boundaries and (2) dependence on external parameters may be created using an auto-regressive Gaussian mixture model. The model is designed to capture how observable spectra are deformed by hypothesis variations, made more expressive projecting data onto a configurable latent space. It used as statistical for scientific discovery in interpreting experimental observations, example when constraining the of physical or tuning simulation according calibration data. also sampled use within Monte Carlo chain, estimate likelihood ratios event classification. method demonstrated simulated high-energy particle physics considering anomalous electroweak production Z boson association dijet system at Large Hadron Collider, accuracy inference tested realistic toy example. developed methods domain agnostic; they any field perform where dataset consisting many real-valued has conditional parameters.