A Stochastic View of Optimal Regret through Minimax Duality