Graph topology invariant gradient and sampling complexity for decentralized and stochastic optimization (2101.00143v2)

Abstract: One fundamental problem in decentralized multi-agent optimization is the trade-off between gradient/sampling complexity and communication complexity. We propose new algorithms whose gradient and sampling complexities are graph topology invariant while their communication complexities remain optimal. For convex smooth deterministic problems, we propose a primal dual sliding (PDS) algorithm that computes an $\epsilon$-solution with $O((\tilde{L}/\epsilon)^{1/2})$ gradient and $O((\tilde{L}/\epsilon)^{1/2}+|\mathcal{A}|/\epsilon)$ communication complexities, where $\tilde{L}$ is the smoothness parameter of the objective and $\mathcal{A}$ is related to either the graph Laplacian or the transpose of the oriented incidence matrix of the communication network. The results can be improved to $O((\tilde{L}/\mu)^{1/2}\log(1/\epsilon))$ and $O((\tilde{L}/\mu)^{1/2}\log(1/\epsilon) + |\mathcal{A}|/\epsilon^{1/2})$ respectively with $\mu$-strong convexity. We also propose a stochastic variant, the primal dual sliding (SPDS) algorithm for problems with stochastic gradients. The SPDS algorithm utilizes the mini-batch technique and enables the agents to perform sampling and communication simultaneously. It computes a stochastic $\epsilon$-solution with $O((\tilde{L}/\epsilon)^{1/2} + (\sigma/\epsilon)^2)$ sampling complexity, which can be improved to $O((\tilde{L}/\mu)^{1/2}\log(1/\epsilon) + \sigma^2/\epsilon)$ with strong convexity. Here $\sigma^2$ is the variance. The communication complexities of SPDS remain the same as that of the deterministic case. All the aforementioned gradient and sampling complexities match the lower complexity bounds for centralized convex smooth optimization and are independent of the network structure. To the best of our knowledge, these gradient and sampling complexities have not been obtained before for decentralized optimization over a constraint feasible set.