arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 1807.09982cs.CG
收藏

斯帕里普斯

Sparips

Bernhard Brehm, Hanne Hardering

摘要:

Rips过滤的持久同源性允许在不同尺度上跟踪点云的拓扑特征,并且是拓扑数据分析的基础工具。不幸的是,当被视为滤波单纯复形时,Rips滤波是指数大小的。因此,除了最小的数据集之外,完全持久化模块的计算是不可能的;当截断拓扑特征的维度时,情况变得稍微不那么棘手,但对于中型数据集仍然令人生畏。理论上可以通过更小和更稀疏的线性大小的单纯复形来近似Rips滤波,然而,可能由于现有方法的复杂性,我们不知道任何现有的实现。我们提出了一个不同的稀疏化方案,覆盖树的基础上,这是很容易实现的,同时提供类似的保证计算缩放。我们进一步提出了一个可视化,是适应近似持久性图,通过纳入一个变量的误差条和跟踪所有的近似保证,明确。

英文摘要:

Persistent homology of the Rips filtration allows to track topological features of a point cloud over scales, and is a foundational tool of topological data analysis. Unfortunately, the Rips-filtration is exponentially sized, when considered as a filtered simplicial complex. Hence, the computation of full persistence modules is impossible for all but the tiniest of datasets; when truncating the dimension of topological features, the situation becomes slightly less intractable, but still daunting for medium-sized datasets. It is theoretically possible to approximate the Rips-filtration by a much smaller and sparser, linear-sized simplicial complexs, however, possibly due to the complexity of existing approaches, we are not aware of any existing implementation. We propose a different sparsification scheme, based on cover-trees, that is easy to implement, while giving similar guarantees on the computational scaling. We further propose a visualization that is adapted to approximate persistence diagrams, by incorporating a variant of error bars and keeping track of all approximation guarantees, explicitly.

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。

↑