K-Tensors: Clustering Positive Semi-Definite Matrices

This paper introduces a novel self-consistency clustering algorithm ($K$-Tensors) designed for {partitioning a distribution of} positive-semidefinite matrices based on their eigenstructures. As positive semi-definite matrices can be represented as ellipsoids in $\mathbb R^p$, $p \ge 2$, it is critical to maintain their structural information to perform effective clustering. However, traditional clustering algorithms {applied to matrices} often {involve vectorization of} the matrices, resulting in a loss of essential structural information. To address this issue, we propose a distance metric {for clustering} that is specifically based on the structural information of positive semi-definite matrices. This distance metric enables the clustering algorithm to consider the differences between positive semi-definite matrices and their projections onto {a} common space spanned by \thadJulyTen{orthonormal vectors defined from a set of} positive semi-definite matrices. This innovative approach to clustering positive semi-definite matrices has broad applications in several domains including financial and biomedical research, such as analyzing functional connectivity data. By maintaining the structural information of positive semi-definite matrices, our proposed algorithm promises to cluster the positive semi-definite matrices in a more meaningful way, thereby facilitating deeper insights into the underlying data in various applications.