Stochasticity in high dimensions has properties that make it appealing for use in algorithms, particularly for surfaces with non-differentiable points. Empirically, stochastic gradient has been found to be effective and theory has followed suit.