In particular, we establish the surprising result that: F or any constant learning rate η > 0, the stochastic gradient bandit algorithm is guaranteed to converge to the globally optimal policy almost surely.