Approximate Differentiable Likelihoods for Astroparticle Physics Experiments

TL;DR

This paper introduces a continuous approximation of event simulators within a probabilistic framework for astroparticle physics, enabling efficient gradient-based inference and improved measurement precision.

Contribution

It presents a novel method to approximate likelihoods in dark matter experiments, facilitating high-dimensional inference and integration of multiple measurement channels.

Findings

Percent-level decrease in measurement uncertainties

Enhanced inference in high-dimensional parameter spaces

Seamless incorporation of multi-channel observables

Abstract

Traditionally, inference in liquid xenon direct detection dark matter experiments has used estimators of event energy or density estimation of simulated data. Such methods have drawbacks compared to the computation of explicit likelihoods, such as an inability to conduct statistical inference in high-dimensional parameter spaces, or a failure to make use of all available information. In this work, we implement a continuous approximation of an event simulator model within a probabilistic programming framework, allowing for the application of high performance gradient-based inference methods such as the No-U-Turn Sampler. We demonstrate an improvement in inference results, with percent-level decreases in measurement uncertainties. Finally, in the case where some observables can be measured using multiple independent channels, such a method also enables the incorporation of additional information seamlessly, allowing for full use of the available information to be made.