Comparing conditional distributions—not just marginal ones—is a recurring problem in statistics and machine learning. Kernel embeddings in a reproducing kernel Hilbert space (RKHS) offer powerful non-parametric tools, but the literature has been fragmented. This paper from Moskvichev, Chau, and Sejdinovic builds a coherent framework around what they call conditional maximum mean discrepancy (CMMD), a family of metrics defined by different embedding choices.
The authors distinguish three special cases: CMMD₀ using conditional mean operators, CMMD₁ using conditional mean embeddings, and CMMD₂ using joint mean embeddings. They then introduce a general level-s CMMD, spell out the assumptions needed, and show how the levels relate through operator-based smoothing. This gives practitioners a clearer map of when each variant is appropriate.
On the estimation side, the paper reviews existing estimators and introduces a novel doubly robust estimator for CMMD. It remains consistent provided the product of errors from the underlying models decreases sufficiently fast. Numerical experiments show the approach captures complex conditional dependencies in statistical testing, suggesting the framework could be useful across domains where conditional relationships matter.