My research focuses on studying models that depict complex dependencies between random variables. Such models include directed graphical models with hidden variables, discrete mixture models (which give rise to low rank tensors), and models that impose strong positive dependence called total positivity. In this talk I will give a brief overview of my work in these areas, and will particularly focus on the problem of density estimation under total positivity.

 Nonparametric density estimation is a challenging statistical problem -- in general the maximum likelihood estimate (MLE) does not even exist! Introducing shape constraints such as total positivity allows a path forward. Though they possess very special structure, totally positive random variables are quite common in real world data and exhibit appealing mathematical properties. Given i.i.d. samples from a totally positive distribution, we prove that the MLE exists with probability one if there are at least 3 samples. We characterize the domain of the MLE, and give algorithms to compute it. If the observations are 2-dimensional or binary, we show that the logarithm of the MLE is a piecewise linear function and can be computed via a certain convex program. Finally, I will discuss statistical guarantees for the convergence of the MLE, and will conclude with a variety of further research directions.

Classical continuous optimization starts with linear and nonlinear programming. In the past two decades, convex optimization (e.g., sparse linear regressions with convex regularizers) has been a very effective computational tool in applications to statistical estimations and machine learning. However, many modern data-science problems involve some basic ``non’’-properties that are ignored by the convex approach for the sake of the computation convenience. These non-properties include the coupling of the non-convexity, non-differentiability and non-(Clarke) regularity. In this talk, we present a rigorous computational treatment for solving two non-problems: the piecewise affine regression and the feed-forward deep neural network. The algorithmic framework is an integration of the first order non-convex majorization-minimization method and the second order non-smooth Newton methods. Numerical experiments demonstrate the effectiveness of our proposed approach. Contrary to existing methods for solving non-problems which provide at best very weak guarantees on the computed solutions obtained in practical implementation, our rigorous mathematical treatment aims to understand properties of these computed solutions with reference to both the empirical and the population risk minimizations. This is based on joint work with Jong-Shi Pang, Bodhisattva Sen and Ziyu He.

In this talk, I will present on my experiences and ideas related to teaching across three different institutions (USU, Colby College, and UCSB). First, I will discuss my approach to lower division math classes that emphasizes interdisciplinary applications of mathematics to physics and engineering. Second, I will talk about my experiences with engaging undergraduate students in my research program, in particular as it relates to Kummer surfaces and string dualities in mathematical physics. I will showcase the results of three of my recent undergraduate researchers with respect to 1) period integrals on Kummer surfaces and their relation to special function identities 2) explicit relations between theta functions of genus two and 1 from geometry, and 3) new normal forms of non-principally polarized Kummer surfaces. I will also discuss plans for future involvement of undergraduates in hands-on research.