(a) Wall-clock timing of LGC, for different number of observations along the x-axis and varying dimensions up to 20D. (b) Wall-clock time measurement of LGC on a 10K (20D) data set, using different learning rates α. (c) Wall-clock time measurements of RP, LMDS, and t-SNE applied to nD LGC data. We note that the wall-clock time of LGC increases with increasing α and depends heavily on the number of dimensions. Moreover, RP and LMDS (landmark ratio=0.05) take less than a minute to run, while t-SNE takes longer and the speed heavily depends on the number of samples. More results of speed experiments using different numbers of dimensions in RP, LMDS, and t-SNE and landmark ratios in LMDS are found in the supplemental materials.