Gradient accelerated cosmology

Thesis

Cosmology is currently sitting on a ticking time bomb that will result in an unprecedented explosion in the quantity and quality of data. In preparation, physicists are starting to incorporate into their theoretical predictions more of the physical, observational and instrumental effects which, until now, could be overlooked. In practice, this translates into a dramatic increase in the number of parameters that future analyses will have to consider. This combination of large data sets with complex models will (and in many cases already does) overwhelm the inference methods we currently use to constrain the values of these parameters. In this thesis, we propose two solutions to this problem. First, we show how gradient-based inference algorithms can dramatically speed up the numerical marginalisation of high dimensional parameter spaces. Second, we show how analytical marginalisation schemes, such as the Laplace ap- proximation, can achieve similar speed increases. Crucially, both these methods rely on having access to computationally affordable gradients of the cosmological models, stressing the importance of developing differen- tiable analyses pipelines for future cosmological surveys.

Authors: Jaime Ruiz-Zapatero

• DOI: 10.5287/ora-nbadoa8my
• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford
• Language: English
• Keywords: statistical inference; gaussian processes; auto-differentiation; cosmology; angular power spectra; Hamiltonian Monte Carlo
• Subjects: Statistics; Cosmology