Title:Discrete Signal Processing on Graphs: Frequency Analysis

Abstract:It is becoming increasingly important to work with datasets that arise not only from physical and engineering applications but from social, genetics, biomolecular, or many other domains. We recently introduced discrete signal processing on graphs that extends concepts and techniques from traditional signal processing and from algebraic signal processing to data indexed by arbitrary graphs. With time and image signals, low- and high-pass frequency components are naturally defined because frequencies can be ordered. But with signals defined on graphs there is no natural ordering of frequencies. This paper considers these concepts and defines low-, high-, and band-pass graph signals and graph filters by introducing a measure of the total variation of graph frequency components. We show how to design graph filters with specified frequency response and apply our methods to problems of sensor malfunctioning, regularization of graph signals, and classification of partially labeled data.