Supplementary Information for: Fast Matrix Square Roots with Applications to Gaussian Processes and Bayesian Optimization

Fig. S1 and Fig. S2 are continuations of Figure 1. This work was conducted while David Eriksson was at Uber AI. Figure S2: Randomized SVD relative error at computing In all cases, randomized SVD is unable to achieve a relative error better than about 0. 25. Fig. S3 further demonstrates the effect of preconditioning on msMINRES-CIQ. To further compare msMINRES-CIQ to randomized methods, Fig. S4 plots the empirical covariance Cholesky-based sampling tend to have very similar empirical covariance error. This additional error is due to the randomness in the RFF approximation.