Reverse Telescoping Coordinates Tame SPD Matrices for Generative AI

A new paper from Purdue University researcher Anindya Bhadra proposes the Reverse Telescoping (RT) coordinate system for symmetric positive definite (SPD) matrices—a foundational object in machine learning, statistics, and neuroscience. Instead of dealing with the restrictive SPD constraint directly, RT maps any p×p SPD matrix Θ to an unconstrained three-part vector x = (v, d, r). Here v encodes the log-determinant (volume), d captures log relative diagonal scales (shape), and r stores partial covariances. This representation is lossless: both Θ and its inverse are symbolically encoded in x, and key computations like determinant, trace, and quadratic forms can be performed in O(p²) time, with O(p³) only required when one explicitly reconstructs the matrix.

The paper further leverages RT for generative modeling on the SPD manifold. Because any two unit-determinant matrices in RT space can be joined by a straight line that stays on the unit-determinant manifold, Bhadra designs a split volume-shape flow model: a one-dimensional flow for volume (determinant) and a separate conditional flow matching model for shape. This surprisingly avoids the normalisation challenges of unconstrained spaces. Experiments show the method works up to p=200 on a difficult synthetic bimodal target and on real fMRI data to generate brain connectivity networks. The RT system also supports intrinsic diffusion on the SPD manifold. Overall, RT transforms the “forbidding” SPD constraint into a structured guide for scalable and geometry-aware generative AI.