Gaussian Process Methods for Very Large Astrometric Data Sets

Abstract We present a novel nonparametric method for inferring smooth models of the mean velocity field and velocity dispersion tensor of the Milky Way from astrometric data. Our approach is based on stochastic variational Gaussian process regression (SVGPR) and provides an attractive alternative to binning procedures. SVGPR is an approximation to standard Gaussian process regression (GPR), the latter of which suffers severe computational scaling with N and assumes independently distributed Gaussian noise. In the Galaxy, however, velocity measurements exhibit scatter from both observational uncertainty and the intrinsic velocity dispersion of the distribution function. We exploit the factorization property of the objective function in SVGPR to simultaneously model both the mean velocity field and velocity dispersion tensor as separate Gaussian processes. This achieves a computational complexity of O ( M 3 ) versus GPR’s O ( N 3 ), where M < < N is a subset of points chosen in a principled way to summarize the data. Applied to a sample of ∼8 × 10 5 stars from the Gaia DR3 Radial Velocity Survey catalog, we construct differentiable profiles of the mean velocity and velocity dispersion as functions of height above the Galactic midplane. We find asymmetric features in all three diagonal components of the velocity dispersion tensor, providing evidence that the vertical dynamics of the Milky Way are in a state of disequilibrium. Furthermore, our dispersion profiles exhibit correlated structures at several locations in ∣ z ∣, which we interpret as signatures of the Gaia phase spiral. Our method also has interdisciplinary applicability for any data set requiring simultaneous modeling of both latent functions and input-dependent noise.