1: \begin{abstract}

2: %	\todo{

3: %	In this paper, we propose a novel algorithm, namely the normal map-based prox-SGD, for composite stochastic optimization problems. First, we show the stationary measure converges to zero in almost sure and expectation sense. This is by far the first asymptotic convergence analysis for proximal type algorithms in the nonconvex setting under the traditional assumptions of stochastic programming.

4: %	Furthermore, the event ``every accumulation point of the algorithm iterates is a stationary point'' occurs with probability 1. When the objective function is subanalytic or semialgebrabic, the iterates generated by our proposed algorithm is proved to be convergent to a $\crit \psi$-valued random variable almost surely relying on the Kurdyka {\L}ojasiewicz (KL) inequality. In particular, under proper KL exponent and step sizes, we obtain the asymptotic convergence rate $\tilde \cO(1/\sqrt{k})$ for the iterates where $\tilde{\cO}$ hides logarithmic terms, and the asymptotic rate matches that of prox-SGD in the strongly convex case. %This is the first paper that establishes the iterates convergence of the proximal type algorithms in the nonconvex setting to the best of our knowledge. 

5: %	More importantly, it provides a new and original perspective for analyzing the behavior of stochastic algorithms.}

6: %	\end{abstract}

7: