Sub-sampled Newton Methods with Non-uniform Sampling

Abstract: We consider the problem of finding the minimizer of a convex function $F: \mathbb R^d \rightarrow \mathbb R$ of the form $F(w) := \sum_{i=1}ⁿ f_i(w) + R(w)$ where a low-rank factorization of $\nabla² f_i(w)$ is readily available. We consider the regime where $n \gg d$. As second-order methods prove to be effective in finding the minimizer to a high-precision, in this work, we propose randomized Newton-type algorithms that exploit \textit{non-uniform} sub-sampling of ${\nabla² f_i(w)}<em>{i=1}<sup>{n}$, as well as inexact updates, as means to reduce the computational complexity. Two non-uniform sampling distributions based on {\it block norm squares} and {\it block partial leverage scores} are considered in order to capture important terms among ${\nabla² f_i(w)}</em>{i=1}<sup>{n}$. We show that at each iteration non-uniformly sampling at most $\mathcal O(d \log d)$ terms from ${\nabla² f_i(w)}_{i=1}<sup>{n}$ is sufficient to achieve a linear-quadratic convergence rate in $w$ when a suitable initial point is provided. In addition, we show that our algorithms achieve a lower computational complexity and exhibit more robustness and better dependence on problem specific quantities, such as the condition number, compared to similar existing methods, especially the ones based on uniform sampling. Finally, we empirically demonstrate that our methods are at least twice as fast as Newton's methods with ridge logistic regression on several real datasets.