IVA using complex multivariate GGD: application to fMRI analysis

Abstract

Examples of complex-valued random phenomena in science and engineering are abound, and joint blind source separation (JBSS) provides an effective way to analyze multiset data. Thus there is a need for flexible JBSS algorithms for efficient data-driven feature extraction in the complex domain. Independent vector analysis (IVA) is a prominent recent extension of independent component analysis to multivariate sources, i.e., to perform JBSS, but its effectiveness is determined by how well the source models used match the true latent distributions and the optimization algorithm employed. The complex multivariate generalized Gaussian distribution (CMGGD) is a simple, yet effective parameterized family of distributions that account for full second-and higher-order statistics including noncircularity, a property that has been often omitted for convenience. In this paper, we marry IVA and CMGGD to derive, IVA-CMGGD, with a number of numerical optimization implementations including steepest descent, the quasi-Newton method Broyden-Fletcher-Goldfarb-Shanno (BFGS), and its limited-memory sibling limited-memory BFGS (L-BFGS) all in the complex-domain. We demonstrate the performance of our algorithm on simulated data as well as a 14-subject real-world complex-valued functional magnetic resonance imaging (fMRI) dataset against a number of competing algorithms.