Simulation-based inference: a new paradigm for cosmoparticle data analysis
by
Justin Alsing(SU / OKC)
→
Europe/Stockholm
FA32
FA32
Description
The data analysis paradigm that we all know and love as cosmologists is a path well-trodden: (1) collect data, (2) write down a likelihood-function for those data given your model for the underlying physics and measurement process, (3) interrogate that likelihood with MCMC sampling (or ratio tests, etc). However, as the era of precision cosmology matures, we find ourselves needing to analyze increasingly complex data — governed by non-linear physics, nasty systematics and selection effects, etc — to make progress. In the face of these complexities, writing down a nice closed-form likelihood becomes hard, or even intractable, and our standard inference methods begin to fail. Meanwhile, making forward simulations of your experiment (with all the nastiness included) is much easier; wouldn’t it be nice if we could just use forward simulations directly for inference? In this talk I’ll show how recent developments in simulation-based inference have opened up a new paradigm for cosmoparticle data analysis, requiring only the ability to forward simulate your dataset. I’ll showcase some examples of this new framework in action: analysis of tSZ maps, dark matter detection with XENON, analysis of strong lenses, and constraining the post-EoR ionizing background from high-z quasars. I’ll argue that simulation-based inference has the potential to greatly simplify the chain of scientific reasoning, and gets physicists back to doing what they are good at — physics.