Grassmann Averages for Scalable Robust PCA — Supplementary Material —

• pseudo-code for the different algorithms (Sec. B); • a proof of convergence (Sec. C); • a more detailed look at the impact of the trimming parameter in TGA (Sec. D); • some thoughts on statistical efficiency in the context of Fig. 6 in the paper (Sec. E); • an empirical verification of Theorem 2 stating that the Grassmann Average conincides with PCA for Gaussian data (Sec. F); • an empirical study of the influence of outliers when the inliers are Gaussian (Sec. G); • an investigation on the impact of robustness in projections (Sec. H); • some notes on the comparison with Inexact ALM [2, 10] (Sec. I); • a brief discussion of the use of extrinsic averages (Sec. J); and • the images of the 20 leading components of Star Wars IV as estimated by EM PCA [9] and TGA (Sec. K).