Revisit last-iterate convergence of mSGD under milder requirement on step size

Understanding convergence of SGD-based optimization algorithms can help deal with enormous machine learning problems.