Digital Signal Processing Seminar

A major problem in time series analysis is choosing an algorithm to estimate the spectrum from a finite observation of a process in such a way that the estimation is not dominated by bias, is consistent and statistically meaningful, and maintains these properties in the presence of minor variations of assumptions. Classic nonparametric methods (e.g., periodogram and its windowed versions) often provide both inconsistent and severely biased estimates. Modern methods typically assume a model (usually AR, MA, or ARMA) for the data and then estimate the model parameters from the data. These parametric methods work well only when the underlying process can be accurately described by such a model. In his classic 1982 paper, David Thompson proposed the powerful multiple-window method, which has been widely regarded as the most elegant and successful technique in the field of spectrum estimation. Based on the Cramer representation, Thompson's method is nonparametric, consistent, efficient, and optimally suited for finite data samples, has good bias control and stability, provides an analysis of variance test for line components, and finally, works very well in many real-world applications. Unfortunately, such an important work has been inappropriately neglected in most textbooks and graduate courses on statistical signal processing.

A list of digital signal processing seminars is available at from the ECE department Web pages under "Seminars". The Web address for the digital signal processing seminars is http://www.ece.utexas.edu/~bevans/dsp_seminars.html