1: \begin{abstract}\vspace{-2mm}

2: The Frank-Wolfe (FW) optimization algorithm has lately re-gained popularity

3: thanks in particular to its ability to nicely handle the structured

4: constraints appearing in machine learning applications. However, its

5: convergence rate is known to be slow (sublinear) when the solution lies at

6: the boundary. A simple less-known fix is to add the possibility to take `away

7: steps' during optimization, an operation that importantly \emph{does not}

8: require a feasibility oracle. %

9: %

10: In this paper, we highlight and clarify several variants of the Frank-Wolfe

11: optimization algorithm that have been successfully applied in practice: 

12: away-steps FW, pairwise FW, fully-corrective FW and Wolfe's minimum norm

13: point algorithm, and prove for the first time that they all enjoy global

14: linear convergence, under a weaker condition than strong convexity of the objective.

15: The constant in the convergence rate has an elegant interpretation as the product

16: of the (classical) condition number of the function with a novel geometric

17: quantity that plays the role of a `condition number' of the constraint set. 

18: We provide pointers to where these algorithms have made a difference in

19: practice, in particular with the flow polytope, the

20: marginal polytope and the base polytope for submodular optimization.

21: 

22: %

23: %

24: %

25: %

26: %

27: %

28: %

29: \end{abstract}

30: