How to Shrink Confidence Sets for Many Equivalent Discrete Distributions?

We consider the situation when a learner faces a set of unknown discrete distributions (pk)k∈ mathcal K defined over a common alphabet mathcal X, and can build for each distribution pk an individual high-probability confidence set thanks to nk observations sampled from pk. The set (pk)k∈ mathcal K is structured: each distribution pk is obtained from the same common, but unknown, distribution q via applying an unknown permutation to mathcal X. We call this permutation-equivalence. The goal is to build refined confidence sets exploiting this structural property. Like other popular notions of structure (Lipschitz smoothness, Linearity, etc.) permutation-equivalence naturally appears in machine learning problems, and to benefit from its potential gain calls for a specific approach. We present a strategy to effectively exploit permutation-equivalence, and provide a finite-time high-probability bound on the size of the refined confidence sets output by the strategy. Since a refinement is not possible for too few observations in general, under mild technical assumptions, our finite-time analysis establish when the number of observations (nk)k∈ mathcal K are large enough so that the output confidence sets improve over initial individual sets. We carefully characterize this event and the corresponding improvement. Further, our result implies that the size of confidence sets shrink at asymptotic rates of O(1/√sumk∈ mathcal K nk) and O(1/maxk∈ K nk), respectively for elements inside and outside the support of q, when the size of each individual confidence set shrinks at respective rates of O(1/√nk) and O(1/nk). We illustrate the practical benefit of exploiting permutation equivalence on a reinforcement learning task.