Convergence Analysis of the Graph Allen-Cahn Scheme
Abstract
Graph partitioning problems have a wide range of applications in machine learning. This work analyzes convergence conditions for a class of diffuse interface algorithm [A.L. Bertozziand A. Flenner, Multiscale Modeling and Simulation, 10(3):1090 1118, 2012] for binary and multi-class partitioning. Using techniques from numerical PDE and convex optimization, convergence and monotonicity are shown for a class of schemes under a graph-independent timestep restriction. We also analyze the effects of spectral truncation, a common technique used to save computational cost. Convergence of the scheme with spectral truncation is also proved under a timestep restriction inversely proportional to the size of the graph. Moreover, this restriction is shown to be sharp in a worst case. Various numerical experiments are done to compare theoretical results with practical performance.
Document Details
- Document Type
- Technical Report
- Publication Date
- Feb 01, 2016
- Accession Number
- AD1014925
Entities
People
- Andrea Bertozzi
- Xiyang Luo
Organizations
- University of California