Georgia Institute of Technology College of Sciences

Motivated by the investigation on the dependence on ``epsilon" in the Dvoretzky's theorem, I will show some refinements of the classical concentration of measure for convex functions. Applications to convexity will be presented if time permits. The talk will be based on joint works with Peter Pivovarov and Petros Valettas.

We propose a new convex relaxation for the problem of solving (random) quadratic equations known as phase retrieval. The main advantage of the proposed method is that it operates in the natural domain of the signal. Therefore, it has significantly lower computational cost than the existing convex methods that rely on semidefinite programming and competes with the recent non-convex methods. In the proposed formulation the quadratic equations are relaxed to inequalities describing a "complex polytope". Then, using an *anchor vector* that itself can be constructed from the observations, a simple convex program estimates the ground truth as an (approximate) extreme point of the polytope. We show, using classic results in statistical learning theory, that with random measurements this convex program produces accurate estimates. I will also discuss some preliminary results on a more general class of regression problems where we construct accurate and computationally efficient estimators using anchor vectors.

Demand forecasting plays an important role in many inventory control problems. To mitigate the potential harms of model misspecification, various forms of distributionally robust optimization have been applied. Although many of these methodologies suffer from the problem of time-inconsistency, the work of Klabjan et al. established a general time-consistent framework for such problems by connecting to the literature on robust Markov decision processes. Motivated by the fact that many forecasting models exhibit very special structure, as well as a desire to understand the impact of positing different dependency structures, in this talk we formulate and solve a time-consistent distributionally robust multi-stage newsvendor model which naturally unifies and robustifies several inventory models with demand forecasting. In particular, many simple models of demand forecasting have the feature that demand evolves as a martingale (i.e. expected demand tomorrow equals realized demand today). We consider a robust variant of such models, in which the sequence of future demands may be any martingale with given mean and support. Under such a model, past realizations of demand are naturally incorporated into the structure of the uncertainty set going forwards. We explicitly compute the minimax optimal policy (and worst-case distribution) in closed form, by combining ideas from convex analysis, probability, and dynamic programming. We prove that at optimality the worst-case demand distribution corresponds to the setting in which inventory may become obsolete at a random time, a scenario of practical interest. To gain further insight, we prove weak convergence (as the time horizon grows large) to a simple and intuitive process. We also compare to the analogous setting in which demand is independent across periods (analyzed previously by Shapiro), and identify interesting differences between these models, in the spirit of the price of correlations studied by Agrawal et al. This is joint work with Linwei Xin, and the paper is available on arxiv at https://arxiv.org/abs/1511.09437v1

Exact oracle inequalities for regression have been extensively studied in statistics and learning theory over the past decade. In the case of a misspecified model, the focus has been on either parametric or convex classes. We present a new estimator that steps outside of the model in the non-convex case and reduces to least squares in the convex case. To analyze the estimator for general non-parametric classes, we prove a generalized Pythagorean theorem and study the supremum of a negative-mean stochastic process (which we term the offset Rademacher complexity) via the chaining technique.(joint work with T. Liang and K. Sridharan)