Hyperspherical Dirac Mixture Reapproximation

We propose a novel scheme for efficient Dirac mixture modeling of distributions on unit hyperspheres. A so-called hyperspherical localized cumulative distribution (HLCD) is introduced as a local and smooth characterization of the underlying continuous density in hyperspherical domains. Based on HLCD, a manifold-adapted modification of the Cram\'er-von Mises distance (HCvMD) is established to measure the statistical divergence between two Dirac mixtures of arbitrary dimensions. Given a (source) Dirac mixture with many components representing an unknown hyperspherical distribution, a (target) Dirac mixture with fewer components is obtained via matching the source in the sense of least HCvMD. As the number of target Dirac components is configurable, the underlying distributions is represented in a more efficient and informative way. Based upon this hyperspherical Dirac mixture reapproximation (HDMR), we derive a density estimation method and a recursive filter. For density estimation, a maximum likelihood method is provided to reconstruct the underlying continuous distribution in the form of a von Mises-Fisher mixture. For recursive filtering, we introduce the hyperspherical reapproximation discrete filter (HRDF) for nonlinear hyperspherical estimation of dynamic systems under unknown system noise of arbitrary form. Simulations show that the HRDF delivers superior tracking performance over filters using sequential Monte Carlo and parametric modeling.