University of Glasgow
Variational inference and normalising flows for gravitational wave parameter estimation
Abstract
dc:description.abstractThis thesis explores the possibility of the application of Variational Inference with Normalising Flows to the gravitational wave parameter estimation problem, providing a potential machine-learning alternative to the currently used stochastic sampling methods. This approach translates Bayesian inference into an optimisation problem, where a variational distribution parametrised by a neural network is matched to the target posterior by minimising a similarity measure between the two. If successful, it offers fast training utilising parallel GPU likelihood evaluations, low-cost post-training posterior sampling and lower bound evidence estimation. The variational inference algorithm is tested on a binary black hole gravitational wave signal injected into the ground-based interferometer network. Successful results are achieved only on constrained subspaces of the full posterior, with the main obstacle identified to be mode collapse fuelled by the reverse Kullback-Leibler divergence training objective. Different strategies of mitigating the mode collapse problem are investigated, with best improvements seen from posterior temperature annealing. The thesis further investigates Bernstein polynomial normalising flows as an alternative to spline-based transforms, for use both within the variational inference algorithm, as well as other normalising flow applications across gravitational wave data analysis. While showing promising results on simple targets, the Bernstein flow architecture performs comparably to spline transforms when used within the variational inference algorithm applied to the gravitational wave problem. In the course of the investigation several useful properties of the Bernstein transform flows are observed — a circular transform configuration allowing for defining densities on periodic domains, as well as a density upper bound providing a lower limit on the Bernstein transform degree necessary to match a target with a given maximum density.
Degree
thesis:*- Level dc:type.qualificationlevel
- MSc(R)
- Grantor dc:publisher.institution
- University of Glasgow
- Year dc:date.issued
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Zięba, Dominika Zuzanna
Subjects
dc:subject × 2Rights
- Language dc:language
- en
Identifiers
dc:identifier.*- DOI dc:identifier.doi
- https://doi.org/10.5525/gla.thesis.86304