Georgia Institute of Technology College of Sciences

We improve on some recent results of Sagiv and Steinerberger that quantify the following uncertainty principle: for a function f with mean zero, then either the size of the zero set of the function or the cost of transporting the mass of the positive part of f to its negative part must be big. We also provide a sharp upper estimate of the transport cost of the positive part of an eigenfunction of the Laplacian.

This proves a conjecture of Steinerberger and provides a lower bound of the size of a nodal set of the eigenfunction. Finally, we use a similar technique to provide a measure of how well the points in a design in a manifold are equidistributed. This is a joint work with Tom Carroll and Xavier Massaneda.

In many optimization problems, a feasible solution induces a multi-dimensional cost vector. For example, in load-balancing a schedule induces a load vector across the machines. In k-clustering, opening k facilities induces an assignment cost vector across the clients. In this paper we consider the following minimum norm optimization problem : given an arbitrary monotone, symmetric norm, find a solution which minimizes the norm of the induced cost-vector. This generalizes many fundamental NP-hard problems. We give a general framework to tackle the minimum norm problem, and illustrate its efficacy in load balancing and, time permitting, in the clustering setting.

(The speaker is an ACO alum; after the lecture, the speaker will engage with the ACO students for 30-45 minutes.)

Learning latent structures in noisy data has been a central task in statistical and computational sciences. For applications such as ranking, matching and clustering, the structure of interest is non-convex and, furthermore, of combinatorial nature. This talk will be a gentle introduction to selected models and methods for statistical inference of such combinatorial structures. I will particularly discuss some of my recent research interests.

This talk will introduce a precise high-dimensional asymptotic theory for Boosting (AdaBoost) on separable data, taking both statistical and computational perspectives. We will consider the common modern setting where the number of features p and the sample size n are both large and comparable, and in particular, look at scenarios where the data is asymptotically separable. Under a class of statistical models, we will provide an (asymptotically) exact analysis of the generalization error of AdaBoost, when the algorithm interpolates the training data and maximizes an empirical L1 margin. On the computational front, we will provide a sharp analysis of the stopping time when boosting approximately maximizes the empirical L1 margin. Our theory provides several insights into properties of Boosting; for instance, the larger the dimensionality ratio p/n, the faster the optimization reaches interpolation. At the heart of our theory lies an in-depth study of the maximum L1-margin, which can be accurately described by a new system of non-linear equations; we analyze this margin and the properties of this system, using Gaussian comparison techniques and a novel uniform deviation argument. Time permitting, I will present analogous results for a new class of boosting algorithms that correspond to Lq geometry, for q>1. This is based on joint work with Tengyuan Liang.