Georgia Institute of Technology College of Sciences

In 1665, Huygens discovered that, when two pendulum clocks hanged from a same wooden beam supported by two chairs, they synchronize in anti-phase mode. On the other hand, metronomes synchronize in-phase when oscillating on top of the same movable surface. In this talk, I will describe and analyze a model to help understand the conditions that lead to anti-phase synchronization vs. the conditions that lead to in-phase synchronization.

In current convex optimization literature, there are significant gaps between algorithms that produce high accuracy (1+1/poly(n))-approximate solutions vs. algorithms that produce O(1)-approximate solutions for symmetrized special cases. This gap is reflected in the differences between interior point methods vs. (accelerated) gradient descent for regression problems, and between exact vs. approximate undirected max-flow. In this talk, I will discuss generalizations of a fundamental building block in numerical analysis, preconditioned iterative methods, to convex functions that include p-norms. This leads to algorithms that converge to high accuracy solutions by crudely solving a sequence of symmetric residual problems. I will then briefly describe several recent and ongoing projects, including p-norm regression using m^{1/3} linear system solves, p-norm flow in undirected unweighted graphs in almost-linear time, and further improvements to the dependence on p in the runtime.

Logistic regression is arguably the most widely used and studied non-linear model in statistics. Classical maximum-likelihood theory based statistical inference is ubiquitous in this context. This theory hinges on well-known fundamental results—(1) the maximum-likelihood-estimate (MLE) is asymptotically unbiased and normally distributed, (2) its variability can be quantified via the inverse Fisher information, and (3) the likelihood-ratio-test (LRT) is asymptotically a Chi-Squared. In this talk, I will show that in the common modern setting where the number of features and the sample size are both large and comparable, classical results are far from accurate. In fact, (1) the MLE is biased, (2) its variability is far greater than classical results, and (3) the LRT is not distributed as a Chi-Square. Consequently, p-values obtained based on classical theory are completely invalid in high dimensions. In turn, I will propose a new theory that characterizes the asymptotic behavior of both the MLE and the LRT under some assumptions on the covariate distribution, in a high-dimensional setting. Empirical evidence demonstrates that this asymptotic theory provides accurate inference in finite samples. Practical implementation of these results necessitates the estimation of a single scalar, the overall signal strength, and I will propose a procedure for estimating this parameter precisely. This is based on joint work with Emmanuel Candes and Yuxin Chen.

In the infinite-dimensional KAM theory, solving the homological equations is the one of the main parts. Generally, the coefficients of the homological equations are constants, by comparing the coefficients of the functions, it is easy to solve these equations. If the coefficients of homological equations depend on the angle variables, we call these equations as the variable coefficients homological equations. In this talk we will talk about how to solve these equations.