Georgia Institute of Technology College of Sciences

Abstract: We consider the hidden Markov model, where the dynamic of theprocess is modelled by a latent Markov chain Y and the observations X aresuch that: 1) given the realization of Y, the observations areindependent; 2) the distribution of the i-th observations (X_i) depends onthe i-th element of the Y (Y_i), only.The segmentation problem consists of estimating the underlying realization(path) of Y given the n observation. Usually the realization with maximumlikelihood, the so called Viterbi alignment is used. On the other hand, itis easy to see that the Viterbi alignment does not minimize the expectednumber of misclassification errors.We consider the segmentation problem in the framework of statisticallearning. This unified risk-based approach helps to analyse many existingalignments as well as defining many new ones. We also study theasymptotics of the risks and infinite alignments.

Finding ground states of spin glasses, a model of disordered materials, has a deep connection to many hard combinatorial optimization problems, such as satisfiability, maxcut, graph-bipartitioning, and coloring. Much insight has been gained for the combinatorial problems from the intuitive approaches developed in physics (such as replica theory and the cavity method), some of which have been proven rigorously recently. I present a treasure trove of numerical data obtained with heuristic methods that suggest a number conjectures, such as an equivalence between maxcut and bipartitioning for r-regular graphs, a simple relation for their optimal configurations as a function of degree r, and anomalous extreme-value fluctuations in a variety of models, hotly debated in physics currently. For some, such as those related to finite-size effects, not even a physics theory exists, for others theory exists that calls for rigorous methods.

We consider two problems: (1) estimate a normal mean under a general divergence loss introduced in Amari (1982) and Cressie and Read (1984) and (2) find a predictive density of a new observation drawn independently of the sampled observations from a normal distribution with the same mean but possibly with a different variance under the same loss. The general divergence loss includes as special cases both the Kullback-Leibler and Bhattacharyya-Hellinger losses. The sample mean, which is a Bayes estimator of the population mean under this loss and the improper uniform prior, is shown to be minimax in any arbitrary dimension. A counterpart of this result for predictive density is also proved in any arbitrary dimension. The admissibility of these rules holds in one dimension, and we conjecture that the result is true in two dimensions as well. However, the general Baranchik (1970) class of estimators, which includes the James-Stein estimator and the Strawderman (1971) class of estimators, dominates the sample mean in three or higher dimensions for the estimation problem. An analogous class of predictive densities is defined and any member of this class is shown to dominate the predictive density corresponding to a uniform prior in three or higher dimensions. For the prediction problem, in the special case of Kullback-Leibler loss, our results complement to a certain extent some of the recent important work of Komaki (2001) and George, Liang and Xu (2006). While our proposed approach produces a general class of predictive densities (not necessarily Bayes) dominating the predictive density under a uniform prior, George et al. (2006) produced a class of Bayes predictors achieving a similar dominance. We show also that various modifications of the James-Stein estimator continue to dominate the sample mean, and by the duality of the estimation and predictive density results which we will show, similar results continue to hold for the prediction problem as well. This is a joint research with Professor Malay Ghosh and Dr. Victor Mergel.

My aim is to explain how to prove multi-dimensional central limit theorems for the spectral moments (of arbitrary degrees) associated with random matrices with real-valued i.i.d. entries, satisfying some appropriate moment conditions. The techniques I will use rely on a universality principle for the Gaussian Wiener chaos as well as some combinatorial estimates. Unlike other related results in the probabilistic literature, I will not require that the law of the entries has a density with respect to the Lebesgue measure. The talk is based on a joint work with Giovanni Peccati, and use an invariance principle obtained in a joint work with G. P. and Gesine Reinert