“…In the unconstrained setting, when the gradient is Lipschitz continuous, the standard gradient descent [52] achieves the lower iteration complexity bound O(ε −2 ) [19,20] to find a first-order ε-stationary point x such that ∇f (x) 2 ε. In the composite optimization setting which includes problems with simple, projectionfriendly, constraints similar iteration complexity is achieved by the mirror descent algorithm [14,34,36,48]. Various acceleration strategies of mirror and gradient descent methods have been derived in the literature, attaining the same bound as of gradient descent in the unconstrained case [35,39,40,53] or improving upon it under additional assumptions [2,18].…”