# The missing radial velocities of Gaia: a catalogue of Bayesian estimates for DR3

**Authors:** Aneesh P. Naik, Axel Widmark  
**arXiv:** [2307.13398](https://arxiv.org/abs/2307.13398)  
**Date:** 2023-07-25  
**Summary page:** https://www.summarizepaper.com/en/paper/2307.13398v1/

## Key points

- Bayesian neural networks were effective in estimating missing line-of-sight velocities of Gaia stars
- A catalogue of blind predictions for line-of-sight velocities in Gaia DR3 was published
- The predictions were probability distributions reflecting knowledge about each star
- The accuracy of the predictions was validated by comparing them to DR3 measurements
- The measurements were statistically consistent with the prediction distributions, with an error rate of 1.5%
- A publicly available catalogue of predictive probability distributions for 185 million stars up to G-band magnitude 17.5 was created
- Validation tests showed that the predictions were accurate for stars within approximately 7 kpc from the Sun and with distance precisions better than around 20%
- Typical prediction uncertainty for such stars is around 25-30 km/s
- The catalogue can be used by the scientific community for analyses of stellar kinematics and dynamics
- Preliminary explorations using the catalogue reveal rotation structure, vertical bending, and breathing mode disturbances in the Galactic disc plane at larger distances than previously possible with Gaia data alone
- Prediction distributions have a broad width due to intrinsic conditional distribution function width and epistemic uncertainty arising from limited training data
- Actual measurements are still preferable when line-of-sight velocity is the primary focus of analysis, but the catalogue is useful in areas where proper motions govern interesting dynamics

## Summary

In a previous study, we demonstrated the effectiveness of Bayesian neural networks in estimating missing line-of-sight velocities of Gaia stars and published a catalogue of blind predictions for the line-of-sight velocities in Gaia DR3. These predictions were not just point estimates, but probability distributions that reflected our knowledge about each star. In this follow-up work, we validate the accuracy of these predictions by comparing them to the DR3 measurements. We find that the measurements are statistically consistent with our prediction distributions, with an approximate error rate of 1.5%. Building on this success, we use the same technique to create a publicly available catalogue of predictive probability distributions for the 185 million stars up to a G-band magnitude of 17.5 that still have missing line-of-sight velocities in Gaia DR3. To ensure reliability, we perform validation tests and find that the predictions are accurate for stars within approximately 7 kpc from the Sun and with distance precisions better than around 20%. The typical prediction uncertainty for such stars is around 25-30 km/s. We invite the scientific community to utilize these radial velocities in analyses of stellar kinematics and dynamics. To illustrate its capabilities, we provide some preliminary explorations using this catalogue. We plot the median line-of-sight velocity in the Galactic disc plane, revealing the rotation structure of the disc at larger distances than previously possible with Gaia data alone. Additionally, we plot the mean vertical velocity and north-south asymmetry in vertical velocity to large distances in the disc plane, uncovering vertical bending and breathing mode disturbances in the disc at large Galactocentric radii (12–14 kpc). It is important to note that while our prediction distributions provide valuable information about each star's line-of-sight velocity, they have a broad width due to intrinsic conditional distribution function width and epistemic uncertainty arising from limited training data. Therefore, when line-of sight velocity is primary focus of analysis actual measurements are still preferable; however our catalogue proves particularly useful in areas where proper motions govern interesting dynamics such as vertical motions in distant regions of disc by retaining stars with missing line–of–sight velocity measurements in analysis and marginalizing over this missing dimension valuable information is preserved. The aims of this work are twofold: Firstly we aim to follow up on our previous study by comparing our blind predictions with published measurements from ground–based surveys and confirming their accuracy; secondly we train new model using DR3 measurements and generate catalogue predictive probability distributions remaining missing line–of–sight velocities DR3.