New Probabilistic Methods for Observational Cosmology

New York University, New York NY

Investigators

Abstract

Future astronomical surveys will be producing such enormous quantities of data that new analysis methods are becoming imperative. Especially when large quantities of information must be distilled into theoretical insights, and the individual items are themselves actually of less certainty, a probability-based approach offers impressive advantages. This project will generate toolkits to implement such novel approaches to the extraction of knowledge from overwhelmingly large data sets. With the potential to transform the way science is done both in cosmology and in other areas under threat of being swamped with data, the impact of this work cannot be overestimated. The methods will both use and propagate collaborations with applied mathematics that make the impossible possible. New cosmological surveys will require measurement of smaller signals using larger numbers of galaxies which are individually observed at lower confidence. This will require data analyses that are as information-preserving as possible. This project will create new methods for cosmological data analysis that permit inferences using not lossy, derived data products, such as galaxy catalogs, best-fit redshifts, or correlation function point estimates, but something much closer to the original imaging and spectroscopic data. These techniques will be informed by the principles of probabilistic inference and applied-mathematics technology. The work creates three related toolsets. Toolset 1 is for reconstruction and marginalization of cosmological density fields, which could be the mass, galaxy, or neutral-gas density field, or the two-dimensional projected mass density. Toolset 2 is for cosmological inference that makes proper use of probabilistic information about galaxy and quasar redshifts, improving probabilistic redshift information, and providing informative imputation of missing redshifts, thereby producing predictions and tools for smaller-scale scientific questions. Toolset 3 is for propagation of probabilistic image-level quantities such as galaxy shapes and the point-spread function, into weak-lensing studies of large-scale structure. This will permit inference from survey data conditional priors over galaxy shapes, and use them in a justified forward-modeling measurement of the shear field and cosmological parameters. The toolsets from this project will be the first practical methods for cosmological inference and large-scale structure measurement that can make full and proper, justified, use of probabilistic outputs. For the first time, it will be possible to perform simultaneous inference or refinement of catalog-level properties along with large-scale structure and cosmological inferences. Simultaneous inference will significantly reduce statistical biases in cosmological measurements, and also reduce variance in catalog-level quantities. The toolsets will be papers and methods but also open-source codebases, with benefits beyond cosmology. These developments will help create standards for generating and delivering probabilistic outputs. The research will reach populations inside and outside academia by producing pedagogical papers on inference, data analysis, and computational statistics in the physical sciences.