Mixed Membership Model of Low-rank Matrices with Multimodal Extension

Abstract

Matrix-valued observations arise in multiplex networks, neuroimaging, and other domains where population-level patterns are often low-rank and subjects may express several latent patterns simultaneously. Existing tensor PCA methods provide continuous subject scores but their loading matrices can be difficult to interpret as population prototypes, while low-rank clustering yields interpretable prototypes with hard labels. We introduce a low-rank mixed membership model for matrix-valued data in which the expected value of each subject's matrix is a convex combination of latent low-rank basis matrices. The model yields both interpretable population-level extreme profiles and continuous subject-level memberships. Our multimodal extension shares memberships across modalities with modality-specific basis matrices and can restore identifiability when one modality is insufficient. We establish identifiability under a pure-subject condition, propose a constrained least-squares estimator and scalable algorithm with spectral initialization and low-rank refinement, and derive nonasymptotic error bounds. The estimator achieves a minimax-optimal reconstruction rate up to a logarithmic factor, with separate basis and membership convergence rates under a geometric condition. Simulations corroborate the theoretical rates and show strong performance. In an analysis of Human Connectome Project functional connectivity data, the proposed method identifies interpretable brain connectivity profiles whose estimated memberships are strongly associated with cognitive phenotypes.