Brittany Terese Fasy

Stochastic convergence of persistence landscapes and silhouettes

Persistent homology is a widely used tool in Topological Data Analysis that encodes multi-scale topological information as a multiset of points in the plane called a persistence diagram. It is difficult to apply statistical theory directly to a random sample of diagrams. Instead, we summarize persistent homology with a persistence landscape, introduced by Bubenik, which converts a diagram into a well-behaved real-valued function. We investigate the statistical properties of landscapes, such as weak convergence of the average landscapes and convergence of the bootstrap. In addition, we introduce an alternate functional summary of persistent homology, which we call the silhouette, and derive an analogous statistical theory.

On the Bootstrap for Persistence Diagrams and Landscapes

Persistent homology probes topological properties from point clouds and functions. By looking at multiple scales simultaneously, one can record the births and deaths of topological features as the scale varies. In this paper we use a statistical technique, the empirical bootstrap, to separate topological signal from topological noise. In particular, we derive confidence sets for persistence diagrams and confidence bands for persistence landscapes.

Local persistent homology based distance between maps

We define a topology-based distance metric between road networks embedded in the plane. This distance measure is based on local persistent homology, and employs a local distance signature that enables identification and visualization of local differences between the road networks. This paper is motivated by the need to recognize changes in road networks over time and to assess the quality of different map construction algorithms. One particular challenge is evaluating the results when no ground truth is known. However, we demonstrate that we can overcome this hurdle by using a statistical technique known as the bootstrap.

Add isotropic Gaussian kernels at own risk: more and more resilient modes in higher dimensions

It has been an open question whether the sum of finitely many isotropic Gaussian kernels in n ≥ 2 dimensions can have more modes than kernels, until in 2003 Carreira-Perpinan and Williams exhibited n+1 isotropic Gaussian kernels in Rn with n+2 modes. We give a detailed analysis of this example, showing that it has exponentially many critical points and that the resilience of the extra mode grows like √n. In addition, we exhibit finite configurations of isotropic Gaussian kernels with superlinearly many modes.

A Path-Based Distance for Street Map Comparison

Comparing two geometric graphs embedded in space is important in the field of transportation network analysis. Given street maps of the same city collected from different sources, researchers often need to know how and where they differ. However, the majority of current graph comparison algorithms are based on structural properties of graphs, such as their degree distribution or their local connectivity properties, and do not consider their spatial embedding. This ignores a key property of road networks since the similarity of travel over two road networks is intimately tied to the specific spatial embedding. Likewise, many current algorithms specific to street map comparison either do not provide quality guarantees or focus on spatial embeddings only. Motivated by road network comparison, we propose a new path-based distance measure between two planar geometric graphs that is based on comparing sets of travel paths generated over the graphs. Surprisingly, we are able to show that using paths of bounded link-length, we can capture global structural and spatial differences between the graphs. We show how to utilize our distance measure as a local signature in order to identify and visualize portions of high similarity in the maps. Finally, we present an experimental evaluation of our distance measure and its local signature on street map data from Berlin, Germany and Athens, Greece.

Stochastic Convergence of Persistence Landscapes and Silhouettes

Persistent homology is a widely used tool in Topological Data Analysis that encodes multiscale topological information as a multi-set of points in the plane called a persistence diagram. It is difficult to apply statistical theory directly to a random sample of diagrams. Instead, we can summarize the persistent homology with the persistence landscape, introduced by Bubenik, which converts a diagram into a well-behaved real-valued function. We investigate the statistical properties of landscapes, such as weak convergence of the average landscapes and convergence of the bootstrap. In addition, we introduce an alternate functional summary of persistent homology, which we call the silhouette, and derive an analogous statistical theory.

Choosing thresholds for density-based map construction algorithms

Due to the ubiquitous use of various positioning technologies in smart phones and other devices, geospatial tracking data has become a routine data source. One of its uses that has gained recent popularity is the construction of street maps from vehicular tracking data. Due to the inherent noise in the data, many map construction algorithms are based on thresholding a density function. While kernel density estimation provides a firm theoretical foundation for computing the density from the measurements, the thresholds are generally picked in a heuristic, and often brute-force way, which results in slow algorithms with no guarantees on the map construction quality. In this paper, we formalize the selection of thresholds in a density-based street map construction algorithm. We propose a new thresholding technique that uses persistent homology combined with statistical analysis to determine a small set of thresholds that captures all relevant topological features. We formally prove that when the samples are drawn uniformly from the street map, a constant number of thresholds suffices to recover the street map. We also provide algorithms to compute the thresholds for different sampling assumptions. Finally, we show the effectiveness of our algorithms in several experiments on artificially generated data and on real GPS trajectory data.

A simplicial complex-based approach to unmixing tumor progression data

BackgroundTumorigenesis is an evolutionary process by which tumor cells acquire mutations through successive diversification and differentiation. There is much interest in reconstructing this process of evolution due to its relevance to identifying drivers of mutation and predicting future prognosis and drug response. Efforts are challenged by high tumor heterogeneity, though, both within and among patients. In prior work, we showed that this heterogeneity could be turned into an advantage by computationally reconstructing models of cell populations mixed to different degrees in distinct tumors. Such mixed membership model approaches, however, are still limited in their ability to dissect more than a few well-conserved cell populations across a tumor data set.ResultsWe present a method to improve on current mixed membership model approaches by better accounting for conserved progression pathways between subsets of cancers, which imply a structure to the data that has not previously been exploited. We extend our prior methods, which use an interpretation of the mixture problem as that of reconstructing simple geometric objects called simplices, to instead search for structured unions of simplices called simplicial complexes that one would expect to emerge from mixture processes describing branches along an evolutionary tree. We further improve on the prior work with a novel objective function to better identify mixtures corresponding to parsimonious evolutionary tree models. We demonstrate that this approach improves on our ability to accurately resolve mixtures on simulated data sets and demonstrate its practical applicability on a large RNASeq tumor data set.ConclusionsBetter exploiting the expected geometric structure for mixed membership models produced from common evolutionary trees allows us to quickly and accurately reconstruct models of cell populations sampled from those trees. In the process, we hope to develop a better understanding of tumor evolution as well as other biological problems that involve interpreting genomic data gathered from heterogeneous populations of cells.

Add Isotropic Gaussian Kernels at Own Risk: More and More Resilient Modes in Higher Dimensions

The fact that a sum of isotropic Gaussian kernels can have more modes than kernels is surprising. Extra (ghost) modes do not exist in