Characterizing Datapoints via Second-Split Forgetting Supplementary Material A Theoretical Results A.1 Preliminaries Let w 2 R

We assume the sample complexity required to estimate the distribution as a proxy for the complexity of the distribution. We make these assumptions to simplify the theoretical exposition. However, our results can be observed even after relaxing them at the expense of more book-keeping. Based on Chatterji and Long [ 9 ], we make the following assumptions about the problem setup: (A.1) The labels are reversed for mislabeled examples.