Spectrophotometric parallaxes with linear models: Accurate distances for luminous red-giant stars

Abstract

With contemporary infrared spectroscopic surveys like APOGEE, red-giant stars can be observed to distances and extinctions at which Gaia parallaxes are not highly informative. Yet the combination of effective temperature, surface gravity, composition, and age - all accessible through spectroscopy - determines a giant's luminosity. Therefore spectroscopy plus photometry should enable precise spectrophotometric distance estimates. Here we use the APOGEE-Gaia-2MASS-WISE overlap to train a data-driven model to predict parallaxes for red-giant branch stars with $0<\log g\leq2.2$ (more luminous than the red clump). We employ (the exponentiation of) a linear function of APOGEE spectral pixel intensities and multi-band photometry to predict parallax spectrophotometrically. The model training involves no logarithms or inverses of the Gaia parallaxes, and needs no cut on the Gaia parallax signal-to-noise ratio. It includes an L1 regularization to zero out the contributions of uninformative pixels. The training is performed with leave-out subsamples such that no star's astrometry is used even indirectly in its spectrophotometric parallax estimate. The model implicitly performs a reddening and extinction correction in its parallax prediction, without any explicit dust model. We assign to each star in the sample a new spectrophotometric parallax estimate; these parallaxes have uncertainties of a few to 15 percent, depending on data quality, which is more precise than the Gaia parallax for the vast majority of targets, and certainly any stars more than a few kpc distance. We obtain 10-percent distance estimates out to heliocentric distances of $20\,$kpc, and make global maps of the Milky Way's disk.