1: \begin{abstract}

2: We extend the Frank-Wolfe (FW) optimization algorithm to solve constrained smooth convex-concave saddle point (SP) problems.

3: Remarkably, the method only requires access to linear minimization oracles. 

4: Leveraging recent advances in FW optimization, we provide the first proof of convergence of a FW-type saddle point solver over polytopes, thereby partially answering a 30 year-old conjecture.

5: We also survey other convergence results and highlight gaps in the theoretical underpinnings of FW-style algorithms. 

6: Motivating applications without known efficient alternatives are explored through structured prediction with combinatorial penalties as well as games over matching polytopes involving an exponential number of constraints. 

7: %

8: \end{abstract}

9: