New Abilities and Limitations of Spectral Graph Bisection

Spectral based heuristics belong to well-known commonly used methods for finding a minimum-size bisection in a graph. The heuristics are usually easy to implement and they work well for several practice-relevant classes of graphs. However, only a few research efforts are focused on providing rigorous analysis of such heuristics and often they lack of proven optimality or approximation quality. This paper focuses on the spectral heuristic proposed by Boppana almost three decades ago, which still belongs to one of the most important bisection methods. It is well known that Boppana's algorithm finds and certifies an optimal bisection with high probability in the random planted bisection model -- the standard model which captures many real-world instances. In this model the vertex set is partitioned randomly into two equal sized sets, and then each edge inside the same part of the partition is chosen with probability $p$ and each edge crossing the partition is chosen with probability $q$, with $p \ge q$. In our paper we investigate the problem if Boppana's algorithm works well in the semirandom model introduced by Feige and Kilian. The model generates initially an instance by random selection within the planted bisection model, followed by adversarial decisions. Feige and Kilian posed the question if Boppana's algorithm works well in the semirandom model and it has remained open so far. In our paper we answer the question affirmatively. We show also that the algorithm achieves similar performance on graph models which generalize the semirandom model. On the other hand we prove some limitations: we show that if the density difference $p-q \le o(\sqrt{p\cdot \ln n}/\sqrt{n})$ then the algorithm fails with high probability in the planted bisection model. This bound is sharp, since under assumption $p-q \ge \Omega(\sqrt{p\cdot \ln n}/\sqrt{n})$ Boppana's algorithm works well in the model.