Agile Inexact Methods for Spectral Projector-Based Graph Fourier Transforms

We propose an inexact method for the graph Fourier transform of a graph signal, as defined by the signal decomposition over the Jordan subspaces of the graph adjacency matrix. This method projects the signal over the generalized eigenspaces of the adjacency matrix, which accelerates the transform computation over large, sparse, and directed adjacency matrices. The trade-off between execution time and fidelity to the original graph structure is discussed. In addition, properties such as a generalized Parseval’s identity and total variation ordering of the generalized eigenspaces are discussed. The method is applied to 2010-2013 NYC taxi trip data to identify traffic hotspots on the Manhattan grid. Our results show that identical highly expressed geolocations can be identified with the inexact method and the method based on eigenvector projections, while reducing computation time by a factor of 26,000 and reducing energy dispersal among the spectral components corresponding to the multiple zero eigenvalue.