An Improved Cutting Plane Method for Convex Optimization, Convex-Concave Games and its Applications

Given a separation oracle for a convex set \(K\subset R_n\) that is contained in a box of radius \(R\), the goal is to either compute a point in \(K\) or prove that \(K\) does not contain a ball of radius \(\epsilon\). We propose a new cutting plane algorithm that uses an optimal \(O(nlog(\kappa ))\) evaluations of the oracle and an additional \(O(n_2)\) time per evaluation, where \(\kappa =nR/\epsilon\).

• This improves upon Vaidya’s \(O(\text{SO}\cdot nlog(\kappa )+n_{\omega +1}log(\kappa ))\) time algorithm [Vaidya, FOCS 1989a] in terms of polynomial dependence on \(n\), where \(\omega <2.373\) is the exponent of matrix multiplication and \(\text{SO}\) is the time for oracle evaluation.
• This improves upon Lee-Sidford-Wong’s \(O(\text{SO}\cdot nlog(\kappa )+n_3{log}_{O(1)}(\kappa ))\) time algorithm [Lee, Sidford and Wong, FOCS 2015] in terms of dependence on \(\kappa\).

For many important applications in economics, \(\kappa =\Omega (exp(n))\) and this leads to a significant difference between \(log(\kappa )\) and \(poly(log(\kappa ))\). We also provide evidence that the \(n_2\) time per evaluation cannot be improved and thus our running time is optimal. A bottleneck of previous cutting plane methods is to compute leverage scores, a measure of the relative importance of past constraints. Our result is achieved by a novel multi-layered data structure for leverage score maintenance, which is a sophisticated combination of diverse techniques such as random projection, batched low-rank update, inverse maintenance, polynomial interpolation, and fast rectangular matrix multiplication. Interestingly, our method requires a combination of different fast rectangular matrix multiplication algorithms.