It seems fair that the usual sample correlation coefficient should allow improvement when additional samples from the marginals are available. However, some intuitive attempts fail badly. Using control variates, a simple method is presented which asymptotically achieves the optimal improvement.