We have a generalization guarantee in Section 4.5. It's not especially tight though; in practice, the errors from different codebooks tend to be mostly independent, and you get nice Gaussian-like concentration. I would look at the empirical results in Section 5 to get a better feel for how it performs in practice.