Color refinement is a crucial subroutine in symmetry detection in theory as well as practice. It has further applications in machine learning and in computational problems from linear algebra. While tight lower bounds for the worst case complexity are known [Berkholz, Bonsma, Grohe, ESA2013] no comparative analysis of design choices for color refinement algorithms is available. We devise two models within which we can compare color refinement algorithms using formal methods, an online model and an approximation model. We use these to show that no online algorithm is competitive beyond a logarithmic factor and no algorithm can approximate the optimal color refinement splitting scheme beyond a logarithmic factor. We also directly compare strategies used in practice showing that, on some graphs, queue based strategies outperform stack based ones by a logarithmic factor and vice versa. Similar results hold for strategies based on priority queues.
翻译:彩色精炼是理论和实践对称检测中的关键次常规。 它在机器学习和线性代数的计算问题中还有进一步的应用。 虽然已知最坏情况复杂程度的下限很紧[Berkholz、Bonsma、Grohe、ESA2013], 但没有关于颜色精炼算法设计选择的比较分析。 我们设计了两种模型, 我们可以在其中比较颜色精细算法, 使用正式方法、 在线模型和近似模型。 我们用这些模型来显示, 任何在线算法除了对数系数之外, 没有竞争性, 也没有算法可以比对数性系数以外的最佳色精细分割方案。 我们还直接比较实践中使用的战略, 显示在某些图表中, 以队列为基础的战略比以对数系数为基础的堆数, 反之, 以优先队列为基础的战略也有相似的结果。