On the solution of a rational matrix equation arising in G-networks (Q1675433)

First the definition and properties of a G-network (also called random neural network) are recalled. Finding its steady state corresponds to solving a system of rational equations, expressed in matrix form as \(z=T(z)\) with \(T(z)=\Lambda^+(D_z-P^+)^{-1}P^-D_\mu^{-1}+\alpha D_\mu^{-1}\) where \(z\) is a vector of unknowns in \(\mathcal{D}=\{z\in\mathbb{R}^N: z\geq\mathbf{1}\}\), \(D_w\) denotes a diagonal matrix with diagonal given by the vector \(w\), \(P^+\) and \(P^-\) are matrices with positive elements and zero diagonal while \(P^+ + P^-\) is row sub-stochastic, \(\Lambda^+,\alpha,\mu\) are given vectors. The main contribution of the paper is the analysis of two methods to solve this system. A fixed point iteration \(z^{(k+1)}=T(z^{(k)})\) which is linearly convergent enclosing alternatingly the solution from below and above. A Newton-Raphson method on the other hand solving \(T(z)-z=0\) is locally quadratic convergent. The performance of the two methods is tested numerically and compared with the method proposed in [\textit{J.M. Fourneau}, Computing the steady-state distribution of networks with positive and negative customers. In: 13th IMACS World Congress on Computation and Applied Mathematics, Dublin].