Parallel MR imaging methods like SMASH, SENSE etc. use multiple receiver coils to accelerate the imaging process by reducing the Fourier space sampling requirement. In this paper we show how one can optimally select the sampling locations based upon (1) knowledge of the statistics of the object being imaged and, (2) a statistical criterion for optimality determined by the application. In particular we show that the optimal uniform sample spacing is not necessarily an integer multiple of the Nyquist interval and depends upon the specific coil sensitivities and configuration.