The paper was overall well-received, and R4 in particular liked the combination of randomized lin algebra with IPM and the solid technical analysis. R3 brought up some major points and thought of this as a borderline paper, in part because of a narrow scope of applicability. However, overall, the AC and SAC agree this is an interesting paper (as well as well-written and technically solid), and is enough to be over the bar for NeurIPS. R3 presents a concern that some of the presentation relative to past methods is a bit misleading, and this should be addressed in the minor revisions. Please see R3s review for full details.