Permutation-Invariant Log-Euclidean Geometries on Full-Rank Correlation Matrices
Yann Thanwerdas
DOI: https://doi.org/10.1137/22m1538144
IF: 1.908
2024-04-26
SIAM Journal on Matrix Analysis and Applications
Abstract:SIAM Journal on Matrix Analysis and Applications, Volume 45, Issue 2, Page 930-953, June 2024. There is a growing interest in defining specific tools on correlation matrices which depart from those suited to SPD matrices. Several geometries have been defined on the open elliptope of full-rank correlation matrices: some are permutation-invariant, some others are log-Euclidean, i.e., diffeomorphic to a Euclidean space. In this work, we prove the existence of permutation-invariant log-Euclidean metrics by defining the families of off-log metrics and log-scaled metrics. First, we prove that the recently introduced off-log bijection is a smooth diffeomorphism, allowing pullback of (permutation-invariant) inner products. We introduce the "cor-inverse" involution on the open elliptope, which can be seen as analogous to the inversion of SPD matrices. We show that off-log metrics are not inverse-consistent. That is why, second, we define the log-scaling smooth diffeomorphism between the open elliptope and the vector space of symmetric matrices with null row sums. This map is based on the congruence action of positive diagonal matrices on SPD matrices, more precisely on the existence and uniqueness of a "scaling," i.e., an SPD matrix with unit row sums within an orbit. Thanks to this multiplicative approach, log-scaled metrics are inverse-consistent. We provide the main Riemannian operations in closed form for the two families modulo the computation of the respective bijections.
mathematics, applied