A new arXiv preprint, "Sliced Wasserstein Sketches with Error Certificates," addresses the problem of representing probability distributions in compact forms that still allow meaningful geometric comparisons. The abstract argues that such representations are valuable for learning from distributions.
The authors study optimal, free-weight M-atom quantization of one-dimensional probability measures. In this setting, a distribution is approximated by M atoms whose weights are not fixed in advance, and the quality of the approximation is judged by a Wasserstein-style geometry.
The title promises error certificates, but the abstract excerpt does not specify the form of those certificates or the algorithm. As a result, the main contribution is best understood as a theoretical study of optimal quantization for sliced Wasserstein sketches.
This digest draws on a single source, so there are no independent results to compare.