Research Horizons Seminar
Wednesday, April 18, 2012 - 12:05
1 hour (actually 50 minutes)
Recently, there has been a lot of interest in estimation of sparse vectors in high-dimensional spaces and large low rank matrices based on a finite number of measurements of randomly picked linear functionals of these vectors/matrices. Such problems are very basic in several areas (high-dimensional statistics, compressed sensing, quantum state tomography, etc). The existing methods are based on fitting the vectors (or the matrices) to the data using least squares with carefully designed complexity penalties based on the $\ell_1$-norm in the case of vectors and on the nuclear norm in the case of matrices. Proving error bounds for such methods that hold with a guaranteed probability is based on several tools from high-dimensional probability that will be also discussed.