FRG: Topological methods in data analysis

点击翻译按钮获取中文摘要

DMS-0101364Gunnar CarlssonThe overall goal of this project is to develop flexible topological methods which will allow the analysis of data which is difficult to analyze using classical linear methods. Data obtained by sampling from highly curved manifolds or singular algebraic varieties in Euclidean space are typical examples where our methods will be useful. We intend to develop and refine two pieces of software which have been written by members of our research group, ISOMAP (Tenenbaum) and PLEX (de Silva-Carlsson). ISOMAP is a tool for dimension reduction and parameterization of high dimensional data sets, and PLEX is a homology computing tool which we will use in locating and analyzing singular points in data sets, as well as estimating dimension in situations where standard methods do not work well. We plan to extend the range of applicability of both tools, in the case of ISOMAP by studying embeddings into spaces with non-Euclidean metrics, and in the case of PLEX by building in the Mayer-Vietoris spectral sequence as a tool Both ISOMAP and PLEX will be adapted for parallel computing. We will also begin the theoretical study of statistical questions relating to topology. For instance, we will initiate the study of higher dimensional homology of subsets sampled from Euclidean space under various sampling hypotheses. The key object of study will be the family of Cech complexes constructed using the distance function in Euclidean space together with a randomly chosen finite set of points in Euclidean space. The goal of this project is to develop tools for understanding data sets which are not easy to understand using standard methods. This kind of data might include singular points, or might be strongly curved. The data is also high dimensional, in the sense that each data point has many coordinates. For instance, we might have a data set whose points each of which is an image, which has one coordinate for each pixel. Many standard tools rely on linear approximations, which do not work well in strongly curved or singular problems. The kind of tools we have in mind are in part topological, in the sense that they measure more qualitative properties of the spaces involved, such as connectedness, or the number of holes in a space, and so on. This group of methods has the capability of recognizing the number of parameters required to describe a space, without actually parameterizing it. These methods also have the capability of recognizing singular points (like points where two non-parallel planes or non-parallel lines intersect), without actually having to construct coordinates on the space. We will also be further developing and refining methods we have already constructed which can actually find good parameterizations for many high dimensional data sets. Both projects will involve the adaptation for the computer of many methods which have heretofore been used in by-hand calculations for solving theoretical problems. We will also initiate the theoretical development of topological tools in a setting which includes errors and sampling.