1: \begin{abstract}

2: In this paper, we study decentralized online stochastic non-convex optimization over a network of nodes. Integrating a technique called gradient tracking in decentralized stochastic gradient descent, we show that the resulting algorithm, \textbf{\texttt{GT-DSGD}}, enjoys certain desirable characteristics towards minimizing a sum of smooth non-convex functions. In particular, for general smooth non-convex functions, we establish non-asymptotic characterizations of \textbf{\texttt{GT-DSGD}} and derive the conditions under which it achieves network-independent performances that match the centralized minibatch \textbf{\texttt{SGD}}. In contrast, the existing results suggest that \textbf{\texttt{GT-DSGD}} is always network-dependent and is therefore strictly worse than the centralized minibatch \textbf{\texttt{SGD}}.

3: When the global non-convex function additionally satisfies the Polyak-\L ojasiewics (PL) condition, we establish the linear convergence of \textbf{\texttt{GT-DSGD}} up to a steady-state error with appropriate constant step-sizes. Moreover, under stochastic approximation step-sizes, we establish, for the first time, the optimal global sublinear convergence rate on almost every sample path, in addition to the asymptotically optimal sublinear rate in expectation. Since strongly convex functions are a special case of the functions satisfying the PL condition, our results are not only immediately applicable but also improve the currently known best convergence rates and their dependence on problem parameters.

4: 

5: \begin{IEEEkeywords}

6: Decentralized optimization, stochastic gradient methods, non-convex problems, multi-agent systems.

7: \end{IEEEkeywords}

8: \end{abstract}

9: