A characterization of the alpha-connections on the statistical manifold of multivariate normal distributions (Q6955296)

A statistical manifold \((M,g,\nabla)\) is given by a manifold, a Riemannian matric and a torsion-free affine connection satisfying that the tensor field \(C(X,Y,Z)=(\nabla_Xg)(Y,Z)\) is symmetric. The statistical manifold of normal distributions is given by \(\mathcal{N}={\mathbb R}^n\times \mathrm{Sym}^+(n,\mathbb{R})\) with the Fisher metric \(g^F\) and the Amari-Chentsov connection \(\nabla^A\). \N\NFor \(n=1\), \textit{H. Furuhata} et al. [Inf. Geom. 4, No. 1, 177--188 (2021; Zbl 1473.53023)] have shown that this statistical manifold \(\left( \mathcal{N},g^{F},\nabla^{A}\right) \)\ admits a solvable Lie group structure, with respect to which \(g^{F}\)\ is a left-invariant metric and \(\nabla^{A}\)\ is a left-invariant connection. As a generalization of that result, this paper shows first (Theorem 2.3) that the statistical manifold \(\left( \mathcal{N},g^{F},\nabla^{A\left( \alpha\right) }\right) \)\ admits a statistical Lie group structure with \(\mathcal{N}\)\ a solvable Lie group and the Fisher metric \(g^{F} \)\ left-invariant. It is further shown (Theorem 3.4) that the Amari-Chentsov \(\alpha\)-connection is characterized by the conjugate symmetry of \(\left( \mathcal{N},g^{F},\nabla\right) \). The covariance matrix \(\Sigma\in \mathrm{Sym}^+(n,\mathbb{R})\) is a positive definite symmetric matrix of order \(n\)\ and the conjugate symmetry conditions give a large number of linear equations between the coefficients. To deal with this situation, the authors introduce a canonical orthonormal basis on the Lie algebra \(\mathcal{N}\)\ and then solve these equations, showing that there exists a \(1\)-parameter family of solutions which is no other than the Amari-Chentsov \(\alpha\)-connection.