Stochastic Bregman Proximal Gradient Method Revisited: Kernel Conditioning and Painless Variance Reduction

Abstract: We investigate Bregman proximal gradient (BPG) methods for solving nonconvex composite stochastic optimization problems. Instead of the standard gradient Lipschitz continuity (GLC) assumption, the objective function only satisfies a smooth-adaptability assumption w.r.t. some kernel function. An in-depth analysis of the stationarity measure is made in this paper, where we reveal an interesting fact that the widely adopted Bregman proximal gradient mapping in the existing works may not correctly depict the near stationarity of the solutions. To resolve this issue, a new Bregman proximal gradient mapping is proposed and analyzed in this paper. Second, a thorough analysis is made on the sample complexities of the stochastic Bregman proximal gradient methods under both the old and the newly proposed gradient mappings are analyzed. Note that the absence of GLC disables the standard analysis of the stochastic variance reduction techniques, existing stochastic BPG methods only obtain an $O(\epsilon^{-2})$ sample complexity under the old gradient mapping, we show that such a limitation in the existing analyses mainly comes from the insufficient exploitation of the kernel's properties. By proposing a new kernel-conditioning regularity assumption on the kernel, we show that a simple epoch bound mechanism is enough to enable all the existing variance reduction techniques for stochastic BPG methods. Combined with a novel probabilistic argument, we show that there is a high probability event conditioning on which $O(\sqrt{n}\epsilon^{-1})$ sample complexity for the finite-sum stochastic setting can be derived for the old gradient mapping. Moreover, with a novel adaptive step size control mechanism, we also show an $\tilde{O}(\sqrt{n}L_h(\mathcal{X}_\epsilon)\epsilon^{-1})$ complexity under the new gradient mapping.