Random reshuffling stochastic gradient descent (RR-SGD) is a standard workhorse for finite-sum optimization, but its theoretical guarantees have often lagged behind practical performance. A new paper on arXiv takes a step toward closing that gap by proving convergence for the last iterate of RR-SGD, rather than only for an averaged iterate.

The setting is convex optimization with L-smooth component functions. According to the abstract, recent work had already established convergence guarantees for an averaged iterate in this regime. The new contribution, as reflected in the title, extends the analysis to the last iterate while allowing for large step sizes.

Because the abstract is truncated, the precise conditions and convergence rates are not fully visible in the source. Still, the emphasis on last-iterate behavior is notable: practitioners typically use the final model produced by SGD, so guarantees on that iterate are more directly relevant than those on a theoretical average.