Date of Award

Spring 2022

Document Type


Degree Name

Doctor of Philosophy (PhD)


Statistics and Data Science

First Advisor

Wu, Yihong


Learning distributions that govern generation of data and estimation of related functionals are the foundations of many classical statistical problems. In the following dissertation we intend to investigate such topics when either the hypothesized model is non-parametric or the number of free parameters in the model grows along with the sample size. Especially, we study the above scenarios for the following class of problems with the goal of obtaining minimax rate-optimal methods for learning the target distributions when the sample size is finite. Our techniques are based on information-theoretic divergences and related mutual-information based methods. (i) Estimation in compound decision and empirical Bayes settings: To estimate the data-generating distribution, one often takes the following two-step approach. In the first step the statistician estimates the distribution of the parameters, either the empirical distribution or the postulated prior, and then in the second step plugs in the estimate to approximate the target of interest. In the literature, the estimation of empirical distribution is known as the compound decision problem and the estimation of prior is known as the problem of empirical Bayes. In our work we use the method of minimum-distance estimation for approximating these distributions. Considering certain discrete data setups, we show that the minimum-distance based method provides theoretically and practically sound choices for estimation. The computational and algorithmic aspects of the estimators are also analyzed. (ii) Prediction with Markov chains: Given observations from an unknown Markov chain, we study the problem of predicting the next entry in the trajectory. Existing analysis for such a dependent setup usually centers around concentration inequalities that uses various extraneous conditions on the mixing properties. This makes it difficult to achieve results independent of such restrictions. We introduce information-theoretic techniques to bypass such issues and obtain fundamental limits for the related minimax problems. We also analyze conditions on the mixing properties that produce a parametric rate of prediction errors.