Comparing ordinary probability distributions is a well-studied problem, but some modern learning tasks go one level up: each observation is itself a distribution. The resulting objects, known as meta-measures, live in the space of probability measures over probability measures, denoted P(P(Ω)). A new preprint on arXiv proposes a way to compute distances between them using vertical sliced Wasserstein distances.
The abstract motivates the setting by noting that many learning problems require comparing such hierarchical probabilistic objects. However, the standard toolbox for comparing distributions does not directly apply, which motivates the new construction. The paper's title points to a sliced approach, a technique that typically reduces the computational cost of optimal transport distances by projecting measures onto one-dimensional slices.
Because only this single source was provided, this digest reflects the paper's own framing rather than a comparison with related work. The full details of the proposed distance and its properties are not included in the announcement abstract.