Implement generalized "linear minimization oracles"

[GeoffNN]

Lu and Freund, 2020 use a generalization of the LMO found in Frank-Wolfe optimization. It generalizes minimizing a sum of a linear function and an indicator function by minimizing a sum of a linear function and a bounded domain function (potentially strongly convex / smooth).

It seems from their paper that they obtain linear convergence in the case of a strongly convex regularizer thus defined. It may work also for other stochastic variants.

Also see Bach, 2012 and Yu et al. 2017