What is the best angular density of multipath in MIMO channels?

While various models of multipath angular density are used for modeling MIMO fading channels, it is not known which one is the best, i.e. gives the highest channel capacity. We answer this question in two steps. First, we show that the asymptotic capacity of a broad class of MIMO channels (not necessarily Rayleigh-fading) with an arbitrary correlation structure does not depend on particular channel distribution, but only on the correlation between antennas. Second, the best multipath angular density, which eliminates the correlation and thus maximizes the asymptotic capacity of uniform linear antenna arrays, is shown to be non-uniform, which implies that the popular Clarke's model does not represent the best-case scenario. We identify the best angular density for different values of the antenna spacing. The Gaussian angular density (adopted in IEEE 802.11n) is shown to be near optimum when the spacing equals to the wavelength. Based on these results, practical guidelines for the antenna array design are given.