Institute for Advanced Study Astrophysics Seminar

In applications as varied as the measurement of stellar proper motions, the determination of the Milky Way mass with maser kinematics, and the selection of quasar targets for SDSS-III BOSS, precise---and, more important, accurate---data analysis requires a model that generates the data. A generative model produces a probability distribution function in the space of the noisy data, after convolution by observational uncertainty distribution functions. I show that proper modeling of the data-generating process performs better than other data analysis and classification methods, in scientific applications in which measurements come with relatively reliable uncertainty estimates. I make also some comments on the theoretical basis for and ideal outputs from any principled program of data analysis. These results have implications for almost all ongoing and future astrophysics projects.