Measured performance of our draws of u_∥ in terms of CPU time per draw vs. input frequency x_in. The methods of Semelin et al. (2007) and Smith et al. (2015) are shown in blue and red, while our approach is shown in green. The solid lines are obtained with a = 10⁻³, while the dotted and dashed lines correspond to a = 10⁻² and a = 10⁻⁴. The shaded areas indicate the area between these two curves. The non-monotonic behaviour of CPU cost with a for our method and that of Smith et al. (2015) can be seen: at low x_in, low a values are the cheapest, while at high x_in, low a values are the most expensive draws.