A billion-star search maps the Milky Way’s nested structures

Finding groups within groups

The Milky Way’s stellar distribution reflects both its accumulation of smaller systems and its internal evolution. Understanding that history requires finding structures at different scales, rather than treating every stellar group as a separate, self-contained object.

In a preprint, William H. Oliver and Tobias Buck apply the AstroLink clustering algorithm to more than a billion stars from Gaia Data Release 3. Their question is whether a single, model-agnostic approach can recover the Galaxy’s nested structure and relate it consistently to existing catalogues. The supplied source does not establish peer-reviewed publication.

Here, “model-agnostic” means the method does not prescribe a particular shape for the groups it seeks. It still requires choices about data preparation, measurement scaling and statistical significance.

Preparing a billion-star sample

The researchers selected approximately 1.03 billion stars with five-dimensional information: two coordinates on the sky, two components of proper motion and a distance estimate. Proper motion describes a star’s apparent movement across the sky. The distances came from estimates combining Gaia measurements with information about stellar brightness.

The sample excludes sources failing the authors’ astrometric quality requirements. Rather than applying one brightness limit everywhere, they used a limit that varies across the sky, retaining more stars while accounting for differences in crowding and Gaia’s observations.

Selection nevertheless changes the apparent distribution of stars. The team estimated the probability that stars would enter their sample and weighted the density calculation inversely by that probability, aiming to reduce artificial patterns introduced by uneven sampling.

They also transformed distances so that uncertain estimates for distant stars contributed less strongly to the clustering. Positions and motion-related coordinates were rescaled to balance their influence. These operations help reveal patterns, but also affect which patterns the method identifies.

What AstroLink found

AstroLink searches for overdensities: collections of points more concentrated than their surroundings in the chosen data space. It connects these into a hierarchy, allowing smaller groups to sit inside larger ones. A statistical noise model helps distinguish prominent overdensities from random fluctuations.

At the authors’ chosen significance threshold of 3.8 sigma, the algorithm identified 7,410 groups. That threshold refers to AstroLink’s assessment of density structure; it does not independently establish that every group is a distinct physical stellar system.

About 42.3 million stars, or 4.1% of the selected sample, belonged to identified groups. The hierarchy extended through six levels. The authors report that the clustering calculation took approximately 80 minutes on 48 CPU cores.

Comparing maps with different boundaries

The team compared its hierarchy with catalogues of open clusters, globular clusters, dwarf galaxies and stellar streams. Their “availability-aware” comparisons account for the fact that different searches may not include the same stars. Where the original search samples were unknown, they calculated bounds on the comparison scores.

They also asked whether each association was reciprocal: did both the AstroLink group and the catalogue object identify the other as their best match? Other relationships indicate fragmentation, where one catalogue system corresponds to several detected groups, or merging, where one detected group encompasses several catalogue systems.

Compact, high-contrast systems often produced reciprocal matches. Some globular clusters also showed nested substructure. Extended systems were more often split or recovered only in their denser regions. A substantial population had no counterpart in the comparison catalogues, making these groups candidates for investigation—not confirmed discoveries.

What remains uncertain

The analysis lacks radial velocities, which measure motion toward or away from the observer, so it cannot capture the Galaxy’s full dynamical structure. Residual selection effects and the adopted transformations also influence the results; distance contraction can combine distant stars into composite groups.

Catalogue comparisons introduce further ambiguity. Stream tracks describe expected locations and motions rather than securely established membership lists, so unrelated compact groups can overlap their footprints. Low coverage also weakens interpretations of dwarf-galaxy substructure.

The study therefore offers a scalable framework for locating and comparing stellar patterns, not a definitive physical classification. Establishing what unmatched groups represent—and whether apparent fragmentation is physical—requires additional checks.