Novel Matrix Problems in Modern Applications

点击翻译按钮获取中文摘要

Matrix computations are ubiquitous in scientific and engineering applications.The existence of high-quality algorithms and resulting software packages, such asLAPACK and MATLAB, greatly alleviates the burden on scientists and engineers, allowingthem to devote their energy to the particular application at hand.This proposal is addressed towards identifying and solving some novel and fundamental matrixproblems that arise in modern applications. The last decade hasseen significant advances in communications technology and infrastructure;prime examples are the world-wide web, wireless computing and sensor networks.These modern applications lead to a vast variety of interesting problems --- these rangefrom complex design constructions, data warehousing and retrieval problems, to the need for intelligent analysis of huge amounts of data. As a response to these needs,many new matrix problems arise that need to be solved robustly and efficiently.Some of the most challenging problems in the analyses of large data matrices are therelated problems of clustering, co-clustering (simultaneous clusteringof rows and columns), and matrix approximations. In many applications, the datais not Gaussian, and hence traditional techniques, such as singular valuedecomposition (SVD) and classical Euclidean k-means are often inadequate.The proposed project formulates new matrix approximation problems related toco-clustering: these formulations appeal to a generalized maximum entropy principleto search for structured matrix approximations that arise naturally for an importantclass of statistical distributions, known as exponential families. The aim of theproject is to develop the theory as well as robust, numerical software for thesematrix computations.Another important class of problems contains varied inverse problems: givena set of desirable properties, construct a data matrix that satisfies these properties.Such inverse problems arise in diverse applications, such as the design ofkernel similarity matrices for statistical learning and the design of Gram matriceswith low cross-correlations for wireless communications. The plan is touse alternate projection schemes, and numerical optimization to solve such designproblems. The outcome of the project would be a unified framework for these inverseproblems, including modeling, algorithms and software.It is anticipated that there will be two broad research impacts of the proposed project.The first contribution would be to applied mathematics and computer science--- the formulation and proposed solutions of the novel matrix problems thatarise naturally in applications. It is hoped that the proposed work will vitalize thesearch for better algorithms, and further deepen the connections between applied mathematicsand engineering. The second contribution would be to the applications themselves. Posing particularapplication tasks in terms of well-defined matrix problems should lead to fasterand more reliable methods.The project will support the graduate education of 2 or 3 students. These students will be trained in themultidisciplinary areas of computational linear algebra, data mining, statistical pattern recognition and wireless computing. Specialized courses in all these areas are available in the Computer Science,Applied Mathematics and Electrical & Computer Engineering departments at UT Austin.Such a broad education will fill the need for substantial academic and industrial demand in thesemultidisciplinary areas for educated personnel.The project will also support research experience for one undergraduate, and a conscious effortwill be made to recruit women and minority students.Results of the project will be published in the linear algebra, data-mining, and information theoryliteratures, contributing to each of these areas as well as their cross-fertilization.Data and software developed under the project will be shared with the scientific communityvia a public web site.