a4d92e2cd541fca87e4620aba658316d-AuthorFeedback.pdf

The example suggested by Reviewer 1, where the outlier distribution is a point-mass at the true mean of8 the inlier distribution, does not fit this spectral outliers setting. It is possible to formally state and prove such a result characterizing spectral outliers and our algorithm's success12 atsuch outlier detection tasks. Simple generativemodels (such as those we use tocreate synthetic data sets used in13 our experiments) lead to data sets with spectral outliers where our algorithm provably finds more outliers than e.g.14 naivespectralmethods,outlierdetectionbasedon `2 norms,etc. See plots (g),(h),(i) and section 10 of supplementary25 material. Further Remarks We thank the reviewers for pointing out several typos -42 we will fixthem.