We provide a precise description on how the test risk scales with a suitable norm-based capacity when compared to a classical metric of model size.

We show that kernel ridge regression under covariate shift can be made computationally efficient using Nyström random projections, achieving optimal statistical rates without sacrificing accuracy.