Bessel convolutions on matrix cones: Algebraic properties and random walks (Q842399)

Let \(\mathbb{F}\) denote the field of real or complex numbers or the quaternions respectively. Let for \(q\in \mathbb{N}\), \(\Pi_q\) denote the cone of positive semidefinite \(q\times q\)-matrices over \(\mathbb{F}\). \textit{M. Rösler} [Compos. Math. 143, No.~3, 749--779 (2007; Zbl 1115.33011)] introduced positivity preserving convolution structures on \(\Pi_q\) defined by Bessel functions (with matrix arguments); in fact, hypergroup structures generalizing Bessel Kingman structures (in the case \(q=1\)). In analogy to these one-dimensional examples, the convolutions are motivated by the \textit{group case}: Let, for \(p\geq q\), \(M_{p,q}\) denote the real vector space of \(p\times q-\)matrices over \(\mathbb{F}\) with dimension \(d p q\), \(d\) denoting the real dimension of \(\mathbb{F}\). The unitary group \(U_p\) acts canonically on the vector space and the orbits of this action are parametrized by \(\Pi_q\). The quotient map \(M_{p,q}\to \Pi_q\) transfers the convolution structure of \textit{radial measures}, i.e. \(U_q-\)invariant measures, into a hypergroup convolution structure on \(\Pi_q\). Hence for fixed \(q\in\mathbb{N}\) there exists a sequence of hypergroup structures indexed by a discrete real parameter \(\mu\), and this convolution corresponds to product formulas for Bessel functions \(\mathcal{J}_\mu\) (with matrix arguments). As in the case \(q=1\), these convolutions are extended, interpolating the discrete parameters, to the parameter set \(\{\mu\geq d(q-1/2)\}\), providing new examples of hypergroups \(X_{q,\mu}\) on \(\Pi_q\). The author continues and extends these investigations with emphasis on algebraic and probabilistic features: Section 2 is concerned with properties of these hypergroups and their duals, and with important properties of matrix-Bessel functions, while Section 3 is concerned with general properties of commutative hypergroups which are needed in the sequel. In Section 4 the author investigates hypergroup-automorphisms, -homomorphisms, sub-hypergroups and quotients for the hypergroups \(X_{q,\mu}\) on \(\Pi_q\). There exists an injective homomorphism from \(GL(\mathbb{F},q)\) into \(\mathrm{Aut}(X_{q,\mu})\), \(a\mapsto T_a\), where \(T_a(r):= \left(ar^2a^*\right)^{1/2}\), \(r\in\Pi_q\). For the field of real numbers, \(\mathrm{Aut}(X_{q,\mu})\) is given by the group \(\left\{T_a: \;a\in GL(\mathbb{R},q)\right\}\), for \(\mathbb{F } =\mathbb{C}\) transposition induces an additional automorphism (Theorems 4.11 and 4.12). Furthermore, it follows that for \(\mu\neq \mu'\) the convolution structures \(X_{q,\mu}\) and \(X_{q,\mu'}\) are not equivalent. Sub-hypergroups and quotient hypergroups are kernels and images of homomorphisms \(T_a : a\in\mathrm{End}(\mathbb{F},q)\) and isomorphic to \(X_{k,\mu}\) for some \(k\leq q\). Thus these \textit{group-like} hypergroups possess rich structures of automorphisms and endomorphisms, an important tool for investigations in normalized random walks and limit theorems. Note that for \(q=1\), on \(\mathbb{R}_+\) roughly spoken, the only examples of hypergroups with non-trivial automorphisms are of Bessel-Kingman type. Sections 5 and 6 are devoted to probability, in particular to Gaussian laws: Gaussian continuous convolution semigroups are defined as usual as convolution semigroups of local type. Important examples are (squared) Wishart distributions. For the \textit{group case} Wishart distributions are represented as homomorphic images of Gaussian laws on \(M_{p,q}\), beyond the group case, it is not known if all Gaussian laws are Wishart distributions. (Note that the 'classical' and the 'squared' Wishart distributions correspond in a 1-1 way by the bijective mapping \(r\mapsto r^2\): \(\Pi_q\to \Pi_q\).) The paper closes with a central limit theorem for random walks normalized by automorphisms \(T_{n^{-1/2} I}\) -- the limits laws are as expected Wishart distributions -- and a strong law of large numbers.