new experimental result
1 New experimental results (R2)
We thank all the reviewers for their insightful comments which help us to improve the article. We will include the pointed references on IPW, thank you. A reference to Joonyoung et al., in which they propose a heuristic to debiase It can be actually relaxed into a bound "in average": more precisely only bounds The noise structure in the SGD iterates (Lemma 2(3.)) becomes even more technical to control. An efficient code is also provided.