We propose a new approach for justifying complexity bounds for dual optimization methods. Dual problems often have very big or unbounded size of the optimal solutions. This makes impossible to apply to the complexity analysis of corresponding schemes the standard framework. In this talk we propose new methods, which can work with unbounded feasible sets. All these methods are primal-dual: they generate both primal and dual solutions with required accuracy/feasibility guarantees.
This is a joint work with A. Gasnikov (IITP, Moscow)