Collaborative Research: Photometric redshifts via Bayesian functional data analysis

Johns Hopkins University, Baltimore MD

Investigators

Abstract

Many projects mapping the sky require precise estimates of the speed at which galaxies are receding. A wide-spread technique derives this number, the redshift, from carefully measuring galaxy colors, using an analysis called "photo-zs". Unfortunately, current estimation methods are not precise enough to achieve major survey science goals. The present team of astronomers and statisticians will improve both the precision and the reliability of photo-zs. This work will provide key enabling technology for large surveys in progress and in development, which represent a considerable investment in astronomy. This project will increase the return on that investment. The research requires innovation in both astronomy and statistics and the development of new statistical methods likely to have significant additional impact. The research will help to train a diverse population of students and postdocs in advanced statistics via summer schools and other special sessions, and historically, many such trainees have gone on to pursue careers in data science. Current and forthcoming automated digital sky surveys aim to push further into "precision cosmology" territory by meticulously mapping the distribution and properties of hundreds of millions of galaxies, and measuring the details of thousands of supernovae. This requires accurate and precise estimation of the redshifts of galaxies using broad-band photometric data, by the technique known as photometric redshifts, or photo-zs for short. Unfortunately, current estimation methods do not enable the most complete science return from these surveys. The present project unites astronomers and statisticians to improve the precision as well as the reliability of photo-zs, by modeling spectral energy distributions of galaxies, using Bayesian functional data analysis (FDA), an approach that emphasizes predictive modeling and thorough propagation of information and uncertainty across hierarchical, multi-stage discovery chains. The project will use a modular, hierarchical modeling framework and account for similarity and diversity, with both conventional parameterizations, and new data-driven parameterizations. Because this framework will produce probabilistic photo-z estimates, with possibly complex uncertainties, the team will also study how optimally to provide such estimates and to use them for cosmological science. Open-source implementations of their algorithms will be accelerated by graphics processing units where appropriate. The research will provide valuable inter-disciplinary training to a graduate student, while developing new statistical methods by innovatively combining FDA, machine learning, and high-performance computing. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.