Variance Reduction for Matrix Games
We present a randomized primal-dual algorithm that solves the problem _x_y y^ A x to additive error ϵ in time nnz(A) + √(nnz(A)n)/ϵ, for matrix A with larger dimension n and nnz(A) nonzero entries. This improves on Nemirovski's mirror-prox method by a factor of √(nnz(A)/n) and is faster than stochastic gradient methods in the accurate and/or sparse regime ϵ<√(n/nnz(A)). Our results hold for x,y in the simplex (matrix games, linear programming) and for x in an ℓ_2 ball and y in the simplex (perceptron / SVM, minimum enclosing ball). Our algorithm combines the mirror-prox method and a novel variance-reduced gradient estimator based on "sampling from the difference" between the current iterate and a reference point.
READ FULL TEXT