[Paper Review] Behaviors of $\phi$-exponential distributions in Wasserstein geometry and an evolution equation

Asuka Takatsu|arXiv (Cornell University)|Sep 30, 2011

Geometric Analysis and Curvature Flows9 references4 citations

TL;DR

This paper establishes that the space of ϕ-exponential distributions on ℝᵈ is convex in Wasserstein geometry and stable under a nonlinear evolution equation if and only if ϕ(s) = s^q for some q ∈ (0, (d+4)/(d+2)). The key contribution is the characterization of Gaussian and q-Gaussian measures as the only such distributions with these geometric and dynamical properties, using Wasserstein geometry and evolution equations on measure spaces with dually flat structures.

ABSTRACT

A $\phi$-exponential distribution is a generalization of an exponential distribution associated to functions $\phi$ in an appropriate class, and the space of $\phi$-exponential distributions has a dually flat structure. We study features of the space of $\phi$-exponential distributions, such as the convexity in Wasserstein geometry and the stability under an evolution equation. From this study, we provide the new characterizations to the space of Gaussian measures and the space of $q$-Gaussian measures.

Motivation & Objective

• To characterize the space of ϕ-exponential distributions on ℝᵈ with respect to Wasserstein geometry and evolution equations.
• To determine under what conditions the space of ϕ-exponential distributions is convex in Wasserstein geometry.
• To identify the necessary and sufficient conditions under which the space is stable under a nonlinear evolution equation of porous medium type.
• To show that only Gaussian and q-Gaussian measures satisfy both convexity and stability under the evolution equation.
• To extend the dually flat structure and Cramér–Rao lower bound to generalized exponential families via ϕ-exponential distributions.

Proposed method

• Utilizes Wasserstein geometry on the space of probability measures with finite second moments, defined via optimal couplings and geodesics.
• Applies the Brenier–Knott–Smith theorem to characterize optimal couplings and geodesics via gradient maps of convex functions.
• Defines ϕ-exponential distributions using the inverse of the ϕ-logarithmic function ∫₁ᵗ ds/ϕ(s), with ϕ(s) = s^q yielding q-Gaussian measures.
• Analyzes the behavior of the space of ϕ-exponential distributions under the evolution equation ∂ₜρ = div(ρ∇(lnϕ(ρ) + Ψϕ)) with Ψϕ(x) = cϕ(Id)|x|².
• Employs differential geometry on the manifold of covariance matrices to derive evolution equations for mean and covariance parameters.
• Uses asymptotic analysis and integration of ODEs along geodesics to derive functional identities that constrain ϕ(s) to power laws.

Experimental results

Research questions

• RQ1 Under what conditions is the space of ϕ-exponential distributions convex in Wasserstein geometry?
• RQ2 When is the space of ϕ-exponential distributions stable under the evolution equation ∂ₜρ = div(ρ∇(lnϕ(ρ) + Ψϕ))?
• RQ3 Which ϕ-exponential families satisfy both convexity and stability under the evolution equation?
• RQ4 Is the class of Gaussian and q-Gaussian measures the only one with both properties?
• RQ5 What functional form of ϕ(s) ensures that the space of ϕ-exponential distributions with mean and covariance parameters is dually flat and convex in Wasserstein geometry?

Key findings

• The space of ϕ-exponential distributions is convex in Wasserstein geometry if and only if ϕ(s) = s^q for some q ∈ (0, (d+4)/(d+2)) and α > 0.
• The space of ϕ-exponential distributions is stable under the evolution equation ∂ₜρ = div(ρ∇(lnϕ(ρ) + Ψϕ)) if and only if ϕ(s) = s^q with q ∈ (0, (d+4)/(d+2)) and α > 0.
• The only ϕ-exponential families that are both convex in Wasserstein geometry and stable under the evolution equation are the Gaussian and q-Gaussian measures.
• For ϕ(s) = s^q, the space of ϕ-exponential distributions with mean and covariance matrix parameters is dually flat and coincides with the space of q-Gaussian measures.
• The functional form αϕ(s) = s^q is necessary and sufficient for both convexity and stability, as shown via asymptotic analysis and integration of the evolution equation along geodesics.
• The result holds for both the space Gϕ (with mean and covariance) and Nϕ (with mean and covariance), with identical necessary and sufficient conditions on ϕ.