Title:Statistical Bergman geometry

Abstract:We introduce an embedding $\Phi: \Omega \rightarrow \mathcal{P}(\Omega)$, where $\Omega \subset \mathbb{C}^n$ is a bounded domain, and $\mathcal{P}(\Omega)$ denotes the space of all probability distributions defined on $\Omega$. We establish that the pullback of the Fisher information metric, the fundamental Riemannian metric in Information geometry, from $\mathcal{P}(\Omega)$ through $\Phi$ is exactly the Bergman metric of $\Omega$. By employing the embedding $\Phi$, we consider $\Omega$ as a statistical manifold in $\mathcal{P}(\Omega)$ and present several results in this framework. Firstly, we express the holomorphic sectional curvature of the Bergman metric in terms of the expectation and provide a reconfirmation of Kobayashi's result that the holomorphic sectional curvature lies within the range $(-\infty, 2]$. Secondly, regarding two bounded domains $\Omega_1$ and $\Omega_2$ as statistical manifolds, we show that if a proper holomorphic map $f: \Omega_1 \rightarrow \Omega_2$ preserves the Fisher information metrics, then $f$ must be a biholomorphism. Finally, we prove the asymptotic normality with the Bergman metric, subject to mild conditions imposed on the Bergman kernels that are satisfied by bounded Hermitian symmetric domains.