All the three reviewers rated this paper favorably, well above the acceptance threshold. They agreed that the proposed method, unconstrained mirror descent in continuous time with the mirror map (4) and the initialization (5), is interesting and significant in several respects: it is almost parameter-free, the recovery guarantee has been given (Theorem 2), theoretical prediction of the three-stage behavior was confirmed empirically, and an interpretation of the Hadamard Wirtinger flow algorithm as a discrete-time first-order approximation of the proposal was provided. They also expect that this study, being inspiring, would stimulate follow-up work. I am thus glad to recommend acceptance of this paper for presentation at the NeurIPS conference.