We extend the concept of distribution compression to joint and conditional distributions, introducing algorithms to produce compressed sets that approximate the joint or conditional distribution of a target dataset.

We frame dynamic regret minimization as a static regret problem in an RKHS