Mass classification of dark matter perturbers of stellar tidal streams

Stellar streams formed by tidal stripping of progenitors orbiting around the Milky Way are expected to be perturbed by encounters with dark matter subhalos. Recent studies have shown that they are an excellent proxy to infer properties of the perturbers, such as their mass. Here we present two different methodologies that make use of the fully non-Gaussian density distribution of stellar streams: a Bayesian model selection based on the probability density function (PDF) of stellar density, and a likelihood-free gradient boosting classifier. While the schemes do not assume a specific dark matter model, we are mainly interested in discerning the primordial black holes cold dark matter (PBH CDM) hypothesis form the standard particle dark matter one. Therefore, as an application we forecast model selection strength of evidence for cold dark matter clusters of masses 10(3)-10(5)M(circle dot) and 10(5)-10(9)M(circle dot), based on a GD-1-like stellar stream and including realistic observational errors. Evidence for the smaller mass range, so far under-explored, is particularly interesting for PBH CDM. We expect weak to strong evidence for model selection based on the PDF analysis, depending on the fiducial model. Instead, the gradient boosting model is a highly efficient classifier (99% accuracy) for all mass ranges here considered. As a further test of the robustness of the method, we reach similar conclusions when performing forecasts further dividing the largest mass range into 10(5)-10(7)M(circle dot) and 10(7)-10(9)M(circle dot) ranges. (C) 2022 Elsevier B.V. All rights reserved.

Automated summary

This study presents two methodologies for inferring properties of perturbers from stellar streams formed by tidal stripping, using the non-Gaussian density distribution of stellar streams. The Bayesian model selection and likelihood-free gradient boosting classifier aim to discern the PBH CDM hypothesis from standard particle dark matter. Model selection based on the PDF analysis shows varying degrees of evidence, while the gradient boosting model is highly efficient with 99% accuracy for all mass ranges considered. Further tests dividing the largest mass range into smaller intervals provide similar conclusions.