Wasserstein gradient flows of Maximum Mean Discrepancy with energy kernels

What the math gives to ML

The paper gives a concrete particle gradient flow for the nonsmooth negative-distance kernel K(z)=-|z|^q, together with the essential prescription of removing self-interactions. Its transferable asset is a repulsive-attractive dynamics: particles are attracted toward a target distribution while pairwise interactions spread them apart and can prevent prototype collapse. A practical neural-network use is to maintain trainable latent prototypes or synthetic embeddings and update them with this flow against minibatch feature distributions. The theory supports testing collision avoidance and coverage, but it also warns that global convergence rates and Polyak–Łojasiewicz guarantees should not be assumed.