I’m excited to share my latest research on one of astronomy’s biggest mysteries: dark matter. Using data from the European Space Agency’s Gaia telescope, we inferred how much dark matter exists in our local neighborhood of the galaxy. Dark matter makes up about 85% of all matter in the universe, but it’s invisible and can (as of 2025) only be detected through its gravitational effects. By studying how stars move vertically through the Milky Way’s disk, we can infer the presence and amount of this elusive substance.
I want to highlight three novel aspects of this particular analysis:
- Flexible modeling with Gaussian processes: Instead of assuming simple mathematical formulas for the expressions appearing in our equations, we let the data tell the story. Gaussian processes allow our models to adapt and capture complex variations that simpler approaches would miss. Imagine letting a flexible curve fit the data rather than forcing it into a rigid straight line.
- Full Bayesian analysis without data binning: Traditional methods often group sparse velocity measurements into bins, which can lose important information. Our approach analyzes every individual star’s velocity while properly accounting for uncertainties and correlations between different velocity components. This gives us a more complete and accurate picture of stellar motions.
- Understanding systematic biases from the “tilt term”: We discovered that making overly restrictive assumptions about how stellar velocities are correlated and distributed can significantly bias dark matter estimates. By comparing different modeling approaches, we show how important it is to let the data guide these correlations rather than imposing rigid constraints that might lead to incorrect conclusions.
Title: Local dark matter density from Gaia DR3 K-dwarfs using Gaussian processes
Abstract:
Modern astrophysical surveys have produced a wealth of data on the positions and velocities of stars in the Milky Way with varying accuracies. An ever-increasing detail in observational data calls for more complex physical models and in turn more sophisticated statistical methods to extract information. We perform a vertical Jeans analysis, including a local approximation of the tilt term, using a sample of \(200\,000\) K-dwarf stars from the Gaia DR3 catalogue. After combination with the Survey-of-Surveys (SoS) catalogue, \(160\,888\) of those have radial velocity measurements. We use Gaussian processes as priors for the covariance matrix of radial and vertical velocities. Joint inference of the posterior distribution of the local dark matter density and the velocity moments is performed using geometric variational inference. We find a local dark matter density of \({\rho_\mathrm{dm} = 0.0131 \pm 0.0041\, \mathrm{M}_\odot\,\mathrm{pc}^{-3} = 0.50 \pm 0.15\, \mathrm{GeV}\,\mathrm{cm}^{-3}}\) at the Sun’s position, which is in agreement with most other recent analyses. By comparing a (\(z\)-dependent) Gaussian process prior with a (\(z\)-independent) scalar prior for the tilt term, we quantify its impact on estimates of the local dark matter density and argue that careful modelling is required to mitigate systematic biases.
Authors: Laurin Söding, Ruben Bartel, and Philipp Mertsch
Link to arXiv: arXiv:2506.02956
Link to publisher: Coming soon
