Paper detail

Estimating distances from parallaxes. II. Performance of Bayesian distance estimators on a Gaia-like catalogue

Estimating a distance by inverting a parallax is only valid in the absence of noise. As most stars in the Gaia catalogue will have non-negligible fractional parallax errors, we must treat distance estimation as a constrained inference problem. Here we investigate the performance of various priors for estimating distances, using a simulated Gaia catalogue of one billion stars. We use three minimalist, isotropic priors, as well an anisotropic prior derived from the observability of stars in a Milky Way model. The two priors that assume a uniform distribution of stars--either in distance or in space density---give poor results: The root mean square fractional distance error, f_RMS, grows far in excess of 100% once the fractional parallax error, f_true, is larger than 0.1. A prior assuming an exponentially decreasing space density with increasing distance performs well once its single scale length parameter has been set to an appropriate value: f_RMS is roughly equal to f_true for f_true < 0.4, yet does not increase further as f_true increases up to to 1.0. The Milky Way prior performs well except towards the Galactic centre, due to a mismatch with the (simulated) data. Such mismatches will be inevitable (and remain unknown) in real applications, and can produce large errors. We therefore suggest to adopt the simpler exponentially decreasing space density prior, which is also less time-consuming to compute. Including Gaia photometry improves the distance estimation significantly for both the Milky Way and exponentially decreasing space density prior, yet doing so requires additional assumptions about the physical nature of stars.