A new preprint on arXiv tackles the problem of learning directed acyclic graph (DAG) structures from observational data when the noise terms are heavy-tailed and differ from node to node. The authors build on the established approach of enforcing acyclicity through smooth constraints, specifically citing DAGMA's log-determinant function as an example. This places the work within a well-known family of structure-learning methods.
The abstract notes that observations and the underlying DAG are typically related through linear structural equations, though the full details of the proposed model are cut off in the available text. The title and framing make clear that the contribution is aimed at relaxing common assumptions about disturbance distributions, which are often taken to be Gaussian or otherwise light-tailed.
Because the source is a single preprint with a truncated abstract, the article cannot offer independent verification or compare the method against alternatives. The significance lies in the explicit focus on node-varying heavy-tailed disturbances, a scenario that is realistic in many applications but underexplored in the DAG learning literature.