Measured variance (mean variance over the range r = 80 − 120 h⁻¹ Mpc) vs. computational cost (mean computation time) for the different methods (markers with error bars) and our theoretical prediction (solid lines). The solid lines (blue for the split method, red for dilution, and black for standard LS with M_r ≤ 50) are our theoretical predictions for the increase in variance and computation time ratio when compared to the standard LS, M_r = 50, case, and the dots on the curves correspond to the measured cases (except for LS they are, from right to left, M_r = 25, 12.5, and (50/7); only the first of which was measured). The curve for split ends at M_s = 2500; the optimal case, M_s = M_r, is the circled dot. The error bars for the variance measurement are naive estimates that do not account for error correlations between bins. The theoretical predictions overestimate the cost savings (data points are to the right of the dots on curves; except for the smaller split factors, where the additional speed-up compared to theory is related to some other performance differences between our split and standard LS implementations). This plot would have a different appearance for other r ranges.