Hoffmann-Jørgensen inequalities for random walks on the cone of positive definite matrices (Q6161608)

Let \(\mathcal{P}_m\) denote the cone of \(m\times m\) positive definite symmetric matrices, and \(\{X_n:n\in\mathbb{N}\}\) be a sequence of independent and orthogonally invariant random matrices in \(\mathcal{P}_m\). Define the random walk \(S_1=X_1\) and \(S_{j+1}=S_j^{1/2}X_{j+1}S_j^{1/2}\) for \(j=1,2,\dots\) Proximity of elements of \(\mathcal{P}_m\) is measured by the Riemannian metric defined by \[ d_R(A,B)=\left(\sum_{j=1}^m\left[\log\lambda_j(A^{-1/2}BA^{-1/2})\right]^2\right)^{1/2}, \] for \(A,B\in\mathcal{P}_m\), with \(\lambda_1(A),\dots,\lambda_m(A)\) denoting the eigenvalues of \(A\). The authors establish Hoffmann-Jørgensen inequalities for this random walk \(\{S_n:n\in\mathbb{N}\}\). That is, they give upper bounds on the right tail probabilities of \(U_n\), where \[ U_n=\max\{d_R(I_m,S_1),\dots,d_R(I_m,S_n)\}\,, \] and \(I_m\) is the \(m\times m\) identity matrix. Particular attention is paid to the situation where the \(X_n\) have the Wishart distribution \(W_m(a,I_m)\), i.e., in the case where the underlying random matrices have the following density with respect to Lebesgue measure on \(\mathcal{P}_m\): \[ w(x)=\frac{1}{2^{ma}\Gamma_m(a)}\left(\det x\right)^{a-\frac{1}{2}(m+1)}\exp\left\{-\frac{1}{2}\text{tr }x\right\}\,, \] where \(\Gamma_m(a)=\pi^{m(m-1)/4}\prod_{j=1}^m\Gamma\left(a-\frac{1}{2}(j-1)\right)\) for \(a>\frac{1}{2}(m-1)\), and where \(\Gamma(\cdot)\) is the gamma function. In this example, explicit and computable upper bounds are given for each of the terms appearing in the corresponding Hoffmann-Jørgensen inequality.