vix.ing · top · new · best · stats · spec

Information Geometry via the Q-Root Transform

2026/03/20 by Levin Maier · 1 voice
Computer Science · Mathematics · #cs.IT #math.DG #math.SG

paper · pdf · doi:10.48550/arxiv.2603.20081

Abstract

In this paper, we introduce p-information geometry, an infinite-dimensional framework that shares key features with the geometry of the space of probability densities \( Dens(M) \) on a closed manifold, while also incorporating aspects of measure-valued information geometry. We define the 2-probability simplex with a noncanonical differentiable structure induced via the q-root transform from an open subset of the \( ℓq \)-sphere. This choice makes the \(q\)-root transform an isometry and allows us to construct the \(ℓ2\)- and \(ℓq\)-Fisher--Rao geometries, including Amari--Čencov \(α\)-connections and a Chern connection in the \(ℓq\)-setting. We then apply this framework to an infinite-dimensional linear optimization problem. We show that the corresponding gradient flow with respect to the \(ℓ2\)--Fisher--Rao metric can be solved explicitly, converges to a maximizer under a natural monotonicity assumption, and admits an interpretation as the geodesic flow of an exponential connection. In particular, we prove that this \(e\)-connection is geodesically complete. We further relate these flows to a completely integrable Hamiltonian system through a momentum map associated with a Hamiltonian torus action on infinite-dimensional complex projective space. Finally, inspired by the \(ℓ2\)-theory, we outline an analogous Fisher--Rao geometry for \( Dens(M) \) on possibly noncompact Riemannian manifolds, showing that, with a suitable spherical differentiable structure, the square-root transform remains an isometry.

Citations

Discussions

Related