The question of which estimator of the two-point correlation function has the smallest sampling error is discussed. An expression is given for this smallest sampling error in terms of the correlation functions, although it is found to be of rather limited use. A more useful approach is considered, in which the sampling error is shown to be made up of two components: one is due to the finiteness of the volume of the sample and may approximately be assessed with a simple expression; for the other, which is due to the finite number of points in the sample, an expression is derived which only makes use of the information required to estimate the correlation function itself. The results given by this last expression are in good agreement with existing bootstrap estimates.