Speaker
Description
For every source, gaia_source publishes the five astrometric uncertainties and all ten correlations between them: the joint law of one observation event. Queries typically read the point values, so derived distances, magnitudes and velocities inherit the familiar biases, which are then repaired for that one use.
We describe a reference implementation of a measure-relational algebra: a cell is a probability measure, dependence among cells is carried as an explicit coupling, and operations act on the measures themselves, every derived quantity being the pushforward of the joint law. Here the archive's own $5\times5$ covariance is the within-source coupling; the parallax zero point is one cell per catalogue, shared by every derivation; a population prior is a relation keyed by direction. A point value requires an explicit collapse; what the collapse changes in the answer is its price, which we measure directly and compare with a second-order estimate from the law's covariance.
On the Hyades (466 sources), M67 (2,774) and a simulated cone with truth (2,859) we run the common ad hoc queries both ways: collapsed first, as the archive's SQL does, and with the law carried. Collapsing each M67 member's parallax before inverting it shifts its distance by 2 pc on average — the measured price of the collapse, 3% above its second-order estimate. The published correlations matter for tangential velocities: the uncorrelated default understates their uncertainties for 953 of 960 members. Where the truth is known, the carried law's central 68% interval covers it 66% of the time — the widths are right — but the law is centred on the observation, so the count within 500 pc is inflated by both routes when no population prior is applied: 433 by the cut, 512 expected, 245 true. Conditioning the carried law on a prior reweights its own representation (checked against exact integration to 0.004 per star); the population's own distance law, given the star's parallax error, recovers the count (250) — an operation a point value cannot admit. With the published per-direction prior (Bailer-Jones et al. 2021) as the population relation, the same operation reproduces their geometric distances across all 2,774 M67 sources (median ratio 1.019, 16–84% 0.987–1.037); the 2% is their per-star zero point against our catalogue constant.
About two thousand lines of C run each subset in seconds on a laptop; the specification is published (doi:10.5281/zenodo.22999911), the code and queries are available.