N-fold integer programming via LP rounding
Jana Cslovjecsek, Friedrich Eisenbrand, Robert Weismantel · arXiv (Cornell University) · 2020
We consider N-fold integer programming problems. After a decade of continuous progress, the currently fastest algorithm for N-fold integer programming by Jansen et al. (2019) has a running time of $(rs\Delta)^{O(r^2s + s^2)} {\phi}^2 \cdot nt \log^{O(1)}(nt)$. Here ${\phi}$ is the largest binary encoding length of a number in the input. This algorithm, like its predecessors are based on the augmentation framework, a tailored integer programming variant of local search. In this paper we propose a different approach that is not based on augmentation. Our algorithm relies on a stronger LP-relaxation of the N-fold integer program instead. This relaxation can be solved in polynomial time with parameter dependence $(s{\Delta})^{O(s^2)}$ by resorting to standard techniques from convex optimization. We show that, for any given optimal vertex solution $x^*$ of this relaxation, there exists an optimal integer solution $z^*$ that is within short $\ell_1$-distance, namely $\|x^* - z^*\|_{1} \leq (rs\Delta)^{O(rs)}$. With dynamic programming one can then find an optimal integer solution of the N-fold IP in time $(rs\Delta)^{O(r^2s + s^2)} \,nt $. This, together with an off-the-shelf-method from convex optimization, results in the currently fastest algorithm for N-fold integer programming.