Low-Complexity Nonparametric Bayesian Online Prediction with Universal Guarantees

Weprove its pointwise universality, i.e., it achieves a normalized log loss performance asymptotically as good as the true conditional entropy of the labels given the features.