Speaker
Description
Weak gravitational lensing is now one of the most powerful probes of cosmological parameters and the nature of dark matter and dark energy. Stage-IV analyses will require rapid evaluation of summary statistics across wide angular scales and many tomographic bins, to sample efficiently the high-dimensional parameter spaces of modern likelihoods.
I will present a new mathematical framework that accelerates cosmological inference by recasting nested line-of-sight integrals as tensor operations. Approximating the matter power spectrum, kernel functions, and scale factors as piecewise linear functions of comoving distance allows each segment of the integral to be solved analytically, yielding a compact expression for the lensing angular power spectrum as a contraction between a cosmology-dependent coefficient tensor and the survey-specific tomographic redshift distributions, on which the dependence is quadratic. Validation against the Core Cosmology Library (CCL; Chisari et al. 2019) confirms sub-percent accuracy across all relevant scales and tomographic configurations.
This factorisation is naturally suited to emulation: I train neural-network emulators that map cosmological parameters directly to the coefficient tensors, bypassing intermediate Boltzmann-code evaluations. On a realistic LSST 3×2pt data vector, the combined framework delivers an order-of-magnitude speed-up over standard pipelines under representative Stage-IV conditions, with further acceleration from a JAX implementation that enables automatic differentiation and seamless CPU/GPU parallelisation.
The quadratic dependence on n(z) further enables fully analytic marginalisation over redshift-distribution uncertainties — directly addressing one of the dominant nuisance-parameter bottlenecks in Stage-IV inference. More broadly, this work establishes a foundation for fast, scalable, differentiable cosmological inference tools, with natural extensions to curved-spacetime modelling, beyond-Limber computation, and higher-order lensing statistics.