A low frequency stochastic background of gravitational waves may be detected by pulsar timing experiments in the next five to ten years. Using methods developed to analyze interferometric gravitational wave data, in this paper we lay out the optimal techniques to detect a background of gravitational waves using a pulsar timing array. We show that for pulsar distances and gravitational wave frequencies typical of pulsar timing experiments, neglecting the effect of the metric perturbation at the pulsar does not result in a significant deviation from optimality. We discuss methods for setting upper limits using the optimal statistic, show how to construct skymaps using the pulsar timing array, and consider several issues associated with realistic analysis of pulsar timing data.