Mirror Likelihood for Modeling Manifold Data with Application on Correlation Time Series

Abstract

Manifold-valued data arise in a wide range of scientific applications. Existing statistical methods typically follow either an intrinsic approach, which characterizes the manifold itself via quantities such as geodesic distance or the Fréchet mean, or an extrinsic approach, which embeds the manifold into an ambient Euclidean space, performs unconstrained statistical estimation, and projects the result back to the manifold. While the latter is conceptually and computationally simple, likelihood-based extrinsic solutions remain scarce, due to the lack of a general recipe for choosing an embedding map and for characterizing its impact on the transformed distribution. In this article, we tackle the problem using an augmented likelihood approach: we view the manifold-valued data as the Fenchel mirror map of a latent variable, which is nearly unconstrained, and model the latent variable with a canonical distribution, such as Gaussian. The key transform, the Fenchel mirror map, is a generalization of the Euclidean projection and can effectively map into the interior of the manifold. Taking a Bayesian approach, the proposed framework enjoys a generative model interpretation and an augmented likelihood free from intractable normalization constants, and it supports principled uncertainty quantification and flexible model-based extensions. We study the theoretical and computational properties of the framework and demonstrate its utility in analyzing neuroimaging data of high-dimensional correlation time series, where each correlation matrix is rank deficient.