1: \begin{abstract}

2: Techniques for reducing the variance of gradient estimates used in

3: stochastic programming algorithms for convex finite-sum problems have

4: received a great deal of attention in recent years. By leveraging

5: dissipativity theory from control, we provide a new perspective on two

6: important variance-reduction algorithms: SVRG and its direct

7: accelerated variant Katyusha.  Our perspective provides a physically

8: intuitive understanding of the behavior of SVRG-like methods via

9: a principle of energy conservation.  The tools discussed

10: here allow us to automate the convergence analysis of SVRG-like

11: methods by capturing their essential properties in small semidefinite

12: programs amenable to standard analysis and computational techniques.  Our approach recovers existing convergence results

13: for SVRG and Katyusha and generalizes the theory to alternative parameter choices.  We also discuss how our approach complements the linear

14: coupling technique. Our combination of perspectives leads to

15: a better understanding of accelerated variance-reduced stochastic

16: methods for finite-sum problems.

17: \end{abstract}

18: