On the normalization of the step size in nonsymmetric stochastic gradient algorithms