Hyperfast Second-Order Local Solvers for Efficient Statistically Preconditioned Distributed Optimization

Statistical preconditioning can be used to design fast methods for distributed large-scale empirical risk minimization problems, for strongly convex and smooth loss functions, allowing fewer communication rounds. Multiple worker nodes compute gradients in parallel, which are then used by the central node to update the parameter by solving an auxiliary (preconditioned) smaller-scale optimization problem. However, previous works require an exact solution of an auxiliary optimization problem by the central node at every iteration, which may be impractical. This paper proposes a method that allows the inexact solution of the auxiliary problem, reducing the total computation time. Moreover, for loss functions with high-order smoothness, we exploit the structure of the auxiliary problem and propose a hyperfast second-order method with complexity $\tilde{O}(\kappa)$, where $\kappa$ is the local condition number. Combining these two building blocks (inexactness and hyperfast methods), we show complexity estimates for the proposed algorithm, which is provably better than classical variance reduction methods and has the same convergence rate as statistical preconditioning with exact solutions. Finally, we illustrate the proposed method's practical efficiency by performing large-scale numerical experiments on logistic regression models.