Dimensionality reduction of cross-spectral density matrices using diffusion map projections

Abstract

Cross-spectral density matrices (CSDMs) constructed for matched-field source localization in an ocean waveguide are both Hermitian and positive semi-definite, allowing them to be interpreted in a non-Euclidean geometric framework as “points” in a Riemannian manifold. This abstract matrix representation can be used to construct distance metrics on the manifold, and distances between pairs of such matrices provide a measure of similarity between the true source CSDM and various replica CSDMs. With this geometric interpretation, the shortest “distance” between pairs of source/replica CSDMs represents an estimate of the source location obtained from acoustic field amplitude and phase acquired on a vertical array [Finette and Mignerey, JASA 143 (2018)]. In this presentation, visualizations of CSDM manifolds obtained from simulated acoustic fields propagating in an ocean waveguide with internal wave-induced variability are illustrated to gain insight into this approach to passive source localization. Since the original manifold resides in a high-dimensional space determined by the number of sensors, manifold learning using diffusion maps is employed to reduce the dimensionality but constrained to retain the relative distance relationships among the CSDMs. The original manifold is projected down to three dimensions for visualization. [Work supported by the Office of Naval Research.]