6d0c932802f6953f70eb20931645fa40-AuthorFeedback.pdf

We will remove the term "for simplicity" as15 recommended. Bydoing optimization20 in a learned representation space instead of the problem's original space, we enable substantial benefits against the21 Wethink thetheoretical result foralinear objectiveservesasanimportant steptowardtheconvexcase. Thesample38 complexity of the linear case depends on the slope of the linear objective function (constantC in Equation 4).