PULLBACK FISHER–RAO GEOMETRIES OF PARAMETRIC PROBABILITY FAMILIES: AN EXPOSITORY ATLAS
Description
We study the Fisher-Rao geometry induced on finite-dimensional
statistical manifolds by pulling back the Fisher-Rao metric from the
space of positive probability densities to the parameter space of a para-
metric probability model. Starting from this common construction, we
derive and interpret the resulting Riemannian metrics for univariate and
multivariate Gaussian distributions, elliptical Laplace distributions, lin-
ear and nonlinear Gibbs families, and probability families obtained as
pushforwards of a fixed reference distribution. For each family, we com-
pute the differential of the parameter-to-density map, derive the pull-
back metric tensor, and give a geometric interpretation of its components
in terms of location, scale, anisotropy, orientation, energy fluctuations,
or local mass compression. We then study the associated geodesics, us-
ing closed-form expressions when available and discrete Riemannian en-
ergy minimization otherwise. The purpose of this article is expository:
to provide a unified geometric reading of several classical parametric
probability models and to show how the same Fisher-Rao construction
leads to distinct geometries on their parameter spaces. We finally con-
trast this viewpoint with Wasserstein-Otto geometry at the infinitesimal
level.
Files
fisher_rao_hal_v1.pdf
Additional details
Software
- Repository URL
- https://pierreguilbertf.github.io/pullback-fisher-rao-atlas/
- Programming language
- Python