GPU correlations • Likelihood-free inference • Tools
Here's some public code that I have written and contributed to.
A GPU pair-counting code written by Arnaud de Mattia. I added auto- and cross-correlations of arbitrary spin quantities, which is what it takes to correlate galaxy shapes, which are spin 2, with galaxy positions, which are spin 0. Validated against treecorr, and about a hundred times faster than the CPU codes for shape–position correlations, which is what made the environmental splits in my intrinsic-alignment analysis possible at all.
Written with Constantin Payerne. Capish forward models a cluster survey end to end: a halo catalogue drawn from a set of cosmological parameters, with the mass function, halo bias and super-sample covariance; then the things a survey actually observes — weak-lensing mass, richness, photometric redshift — with completeness, purity and selection applied to them; then the summary statistics you would measure. The point is to constrain cosmology from cluster counts without having to write down a likelihood for them, which is the situation any model with a realistic selection function leaves you in.
Ask an assistant for a talk and it writes a reveal.js HTML file, usually with the content right and the layout wrong. Lectern is a drag-and-drop canvas over that same file: you fix the layout by hand while the assistant keeps working on the content, and each of you sees the other's changes within a second. Where you want something different, you leave a note on the slide and the assistant works through it. The .html file is the deck — no proprietary format, no import, no export, and it stays readable in git diff.