Table of Contents
Fetching ...

Exponential families and affine Grassmannians

Danuzia Nascimento Figueirêdo, Hale Aytaç, Mathieu Molitor

TL;DR

This work addresses the classification of discrete exponential families on a finite set by establishing a one-to-one correspondence between minimal $n$-dimensional exponential families and the affine Grassmannian $\\text{Graff}_{n}(C(\\Omega)/\\mathbb{R})$. Using a Lie group $G_{n}$ acting on pairs $(C,F)$ that define $p_{C,F}$, the authors prove the action is free and proper and that the orbit space is diffeomorphic to the Graff variety, enabling a geometric parametrization of all minimal exponential families. A key consequence is that minimal exponential families are in bijection with affine $n$-dimensional subspaces of $C(\\Omega)/\\mathbb{R}$, yielding a corollary of uniqueness when $|\\Omega|=n+1$. The approach combines reduction by stages with a group-action framework to connect exponential-family representations to affine Grassmannians, providing a canonical classification and opening routes for geometric analyses of statistical models on finite spaces.

Abstract

We establish a one-to-one correspondence between the set of minimal exponential families of dimension n defined on a finite sample space Ω and the affine Grassmannian associated to an appropriate vector space of functions.

Exponential families and affine Grassmannians

TL;DR

This work addresses the classification of discrete exponential families on a finite set by establishing a one-to-one correspondence between minimal -dimensional exponential families and the affine Grassmannian . Using a Lie group acting on pairs that define , the authors prove the action is free and proper and that the orbit space is diffeomorphic to the Graff variety, enabling a geometric parametrization of all minimal exponential families. A key consequence is that minimal exponential families are in bijection with affine -dimensional subspaces of , yielding a corollary of uniqueness when . The approach combines reduction by stages with a group-action framework to connect exponential-family representations to affine Grassmannians, providing a canonical classification and opening routes for geometric analyses of statistical models on finite spaces.

Abstract

We establish a one-to-one correspondence between the set of minimal exponential families of dimension n defined on a finite sample space Ω and the affine Grassmannian associated to an appropriate vector space of functions.

Paper Structure

This paper contains 7 sections, 11 theorems, 45 equations.

Key Result

Proposition 2.1

Let $G$ be a Lie group that acts freely and properly on the left on a manifold $M$, and let $H$ be a closed Lie subgroup of $G$ that is normal in $G$.

Theorems & Definitions (22)

  • Proposition 2.1: Reduction by stages
  • Lemma 2.2
  • proof
  • Proposition 2.3
  • Remark 2.4
  • Definition 3.1
  • Proposition 3.2
  • proof
  • Lemma 3.3
  • proof
  • ...and 12 more