arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 1506.01082cs.DScs.DM
收藏

基于矩形快速矩阵乘法的最大团列表上界的改进

An Improved Upper Bound on Maximal Clique Listing via Rectangular Fast Matrix Multiplication

Carlo Comin, Romeo Rizzi

摘要:

最大团列表问题的第一个输出敏感算法是由Tsukiyama等人在1977年给出的。作为落入逆向搜萦范例的任何算法,它执行对具有要列出的对象(即,最大团)作为其节点的有向树(RS树)的DFS访问。在递归实现中,RS树对应于算法的递归树。时间延迟由生成节点的下一个子节点的成本给出,Tsukiyama表明它是$O(mn)$。2004年,Makino和Uno通过计算\emph{square}快速矩阵乘法在一个Single Shot中生成一个节点的所有子节点,将时间延迟锐化为$O(n^ω)$。在本文中,我们进一步提高了渐近探索相同的RS-树的分组后代的计算更进一步。我们的想法是依靠矩形快速矩阵乘法,以计算所有的孩子$n^2$节点在一个镜头。根据快速矩阵乘法的当前上界,由此时间延迟从$O(n^{2.3728639})$改善到$O(n^{2.093362})$。

英文摘要:

The first output-sensitive algorithm for the Maximal Clique Listing problem was given by Tsukiyama et.al. in 1977. As any algorithm falling within the Reverse Search paradigm, it performs a DFS visit of a directed tree (the RS-tree) having the objects to be listed (i.e. maximal cliques) as its nodes. In a recursive implementation, the RS-tree corresponds to the recursion tree of the algorithm. The time delay is given by the cost of generating the next child of a node, and Tsukiyama showed it is $O(mn)$. In 2004, Makino and Uno sharpened the time delay to $O(n^ω)$ by generating all the children of a node in one single shot performed by computing a \emph{square} fast matrix multiplication. In this paper, we further improve the asymptotics for the exploration of the same RS-tree by grouping the offsprings' computation even further. Our idea is to rely on rectangular fast matrix multiplication in order to compute all children of $n^2$ nodes in one shot. According to the current upper bounds on fast matrix multiplication, with this the time delay improves from $O(n^{2.3728639})$ to $O(n^{2.093362})$.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑