Competitive guarantee for the depth-averaged LSA prior

Establish whether the depth-averaged layered simplex architecture prior admits a Good–Turing-type competitive guarantee, such as regret against the natural oracle vanishing uniformly over all distributions in the d-dimensional simplex.

Background

The paper shows empirically that the depth-averaged layered simplex architecture (LSA) prior performs competitively with Good–Turing estimation on many heavy-tailed targets, while retaining a coherent Bayesian mixture interpretation and a non-uniform prior-mass regret guarantee. However, Good–Turing-type estimators also possess worst-case competitive guarantees relative to a natural oracle that knows the target distribution but must assign equal probabilities to symbols with equal observed counts.

The authors identify as unresolved whether the depth-averaged LSA prior can achieve an analogous uniform guarantee over the entire probability simplex. Such a result would connect the Bayesian, minimum-description-length interpretation of the LSA predictor with worst-case competitive distribution estimation, while addressing its observed weakness on flat targets.