Exact multi-parameter persistent homology of time-series data: Fast and variable topological inferences

TL;DR

This paper introduces an exact multi-parameter persistent homology method for time-series analysis using the Liouville torus, offering faster, more flexible topological inferences with reduced computational costs.

Contribution

It develops a novel multi-parameter filtration based on Fourier decomposition and provides an exact barcode formula, improving efficiency over traditional sliding window TDA methods.

Findings

Exact barcode formula derived for Liouville torus-based analysis

Significant reduction in computational complexity

Enhanced flexibility in topological inferences from time-series data

Abstract

We propose a novel exact multi-parameter persistent homology method for analyzing time-series data utilizing the Liouville torus. In the field of topological data analysis (TDA), the conventional approach to analyzing time-series data often involves sliding window embedding. From the perspective of Takens' embedding theorem, we justify the analysis of the Liouville torus in TDA and discuss the similarities and differences between the Liouville torus and sliding window embedding approaches. We develop a multi-parameter filtration method based on Fourier decomposition and provide an exact formula of persistent homology with its one-parameter reduction of the multi-parameter filtration. The conventional TDA of time-series data via sliding window is known to be computationally expensive, but the proposed method yields the exact barcode formula with the symmetry of the Liouville torus promptly, which significantly reduces computational complexity while demonstrating comparable or superior performance compared to the existing TDA methods. Furthermore, the proposed method provides a way of obtaining various topological inferences by exploring different filtration rays within the multi-parameter filtration space, utilizing the nearly real-time computational capabilities of the proposed method. The advantages of the proposed method significantly improve the efficiency and flexibility of TDA when handling extensive time-series data within machine learning workflows.