Reviews: A Primal-Dual link between GANs and Autoencoders

Summary: It is the purpose of this article to establish links between the two most popular generative models nowadays, namely GANs and (W)AEs. In Theorem 8, an (in)equality linking both criteria is stated, tending to explain the performance similarities between the models. An introduced f-WAE model is also thoroughly analyzed, linked to an f-divergence or a Wasserstein distance depending on the weighting operated. Finally, Authors use findings of Theorem 8 to derive generalization bounds on WAEs. Major Comments: - The paper is well written, related works are correctly discussed and notions well introduced, making the submission self-contained although quite dense. The fact that Theorem 8 can be viewed as a generalization of Lemma 4 is particularly striking, and attest of the soundness of the approach.