One of the fundamental challenges with multivariate data is that a complete statistical or deterministic model involves a state space that grows exponentially with the number of dimensions (i.e. neurons, in this case). Therefore, in the case of a probabilistic model, it becomes computationally unfeasible to estimate the complete joint probability distribution of the firing of a large group of neurons because there are not enough data points (i.e. trials) to adequately fill the state space. One way to overcome this problem is to assume a parametric model such as a Gaussian distribution which has only mean and covariance parameters to estimate. This is, in fact, what we did with our data collected in the center-out task (Maynard et al., submitted for publication). Using a population of 16 neurons, we estimated the joint probability of firing of all 16 neurons by estimating the mean firing rates of the neurons and their covariances (i.e. how the firing rates of two neurons co-varied on a trial by trial basis) measured from the start to the end movement for each movement direction separately. Eight different joint distributions were estimated, one for each movement direction. Using these models and Bayes' rule, the probability of a movement direction given the firing rates of the neurons on a single trial could be estimated. By selecting the most probable direction given the firing rates (i.e. maximum likelihood estimation), we found that the actual movement direction was correctly chosen 63% of the time which was 8% higher than what was estimated if we assumed the neurons fired independently and 24% higher than the performance of the population vector method (Georgopoulos, Schwartz & Kettner, 1986). We plan on using a similar approach to show that a multivariate model will reveal more saliently a neural representation underlying the binding of movement segments.