LIGO Seminar

https://caltech.zoom.us/j/82072539167

Title: "Exploiting Gaussian Processes for data analysis: a Gravitational-wave Cosmology example" with Virginia d'Emilio, Cardiff University

Abstract: Gaussian Processes (GP) can be simply described as a regression method with a probabilistic (Bayesian) interpretation. They have been employed for various tasks in Gravitational-wave (GW) data analysis, from modelling calibration uncertainties in the LIGO-VIRGO detectors, to generating a non-parametric fit for equations of state [2], to more recently being proposed for modelling long-duration glitches [2]. GW astrophysical analysis is largely based on Bayesian inference results of individual observations. These are what we call parameter estimation (PE) products. In this talk, I will focus on two applications of GPs for improving the post-processing of PE results. In particular, I will give a brief overview of a GP-based density estimation technique [3] that can offer improved accuracy over other Kernel Density Estimators (KDEs). I will present an example application of this technique for GW cosmology with dark (and bright [4]) sirens. This example allows me to highlight how density estimation can be difficult for non-smooth, multi-dimensional surfaces and why alternative interpolation methods like GPs can strengthen our confidence in the results. I will then outline how GPs could be used to model the full PE likelihood-surface and generally enhance our data analysis techniques.