A Riemannian viewpoint on the Amari-Cencov $\alpha$-connections and Proudman-Johnson equations

Abstract

We give a new geometric interpretation of the Amari-Cencov $\alpha$-connections $\nabla^{(\alpha)}$ from information geometry: On the space of densities $\operatorname{Dens}_+(M)$, we show that there exist Riemannian metrics $G^\alpha$, which we call $\alpha$-Fisher-Rao metrics, whose Levi-Civita connections are $\nabla^{(\alpha)}$. With the exception of $\alpha=0$ (the Fisher-Rao metric), these metrics are non-invariant to the action of the diffeomorphism group $\operatorname{Diff}(M)$, even though the connections are invariant. This gives a new way of interpreting the geodesics of the $\nabla^{(\alpha)}$ as energy-minimizing curves. On the space of probability densities $\operatorname{Prob}(M)$, we show that the same phenomenon holds for $\alpha\in \{-1,0,1\}$ and that the $\alpha$-connections are not metric otherwise. We show that $\nabla^{(\alpha)}$-geodesics on this space can be interpreted as radial projections of straight lines on appropriate hyper-surfaces, and use this geometric picture to obtain geodesic convexity for any $\alpha\in \mathbb{R}$. In addition, we prove analogous results for appropriate metrics and connections on $\operatorname{Diff}(M)$, which, for the case $M=\mathbb{R}$, imply that the generalized Proudman-Johnson equations on the real line are the Euler-Arnold equations of non-right invariant metrics. Finally, in the finite-dimensional case, we show that $\nabla^{(\alpha)}$ can be metric or non-metric depending on the considered statistical model.