On generation of elliptical distributions with Gaussian form covariance matrix (Q1338607)

This paper shows methods for generating random vectors \(Z= (z_ 1,\dots, z_ N)\) having multivariate Gaussian, exponential and Bessel distributions with Gaussian form variance-covariance matrix \(R\). All three distributions belong to the class of elliptical distributions for which the probability density function (PDF) of \(Z\) takes the form \(P_ Z(Z)= f(r)/| R|^{1/2}\), where \(r= [(Z- \mu)^ T R^{-1}(Z- \mu)]^{1/2}= (W^ T W)^{1/2}\), \(f(r)= P_ W(W)\) (\(W\) is spherically distributed) and \(\mu\) is the mean vector of \(Z\). The main results include a method to generate \(r\) for the more general Bessel case in which the PDF of \(r\) has the form \(r^ a K_ b(r)\), where \(K_ b(r)\) is the modified Bessel function of the second kind of order \(b\), as well as a method to generate exponentially and Bessel- distributed vectors \(W\) from the Gaussian ones. The PDF of the form \(r^ a K_ b(r)\), \(r> 0\) can be generated from the product of two Chi random variables. The authors discover that the matrix \(R\) is representable by the Kronecker product \(\sigma^ 2 B\otimes A\) of two Toeplitz matrices \(B\) and \(A\) of similar elements of dimensions smaller than \(R\) which have analytical triangular (LDU) decomposition. The analytical forms of \(L\) and \(L^{-1}\) in the LDU decomposition of \(R\) are obtained. These two results are applied to reduce the computations in correlating vectors for random vector generation and in the evaluation of the quadratic form of \(R^{-1}\). The set of random vectors \(Z\) can be used to simulate the surface height sampled from a rough terrain, which can be applied toward analysis using the Monte Carlo method for the quantities related to the electromagnetic wave scattering from a rough terrain.