Harmonic Extension on the Point Cloud
Zuoqiang Shi,Jian Sun,Minghao Tian
DOI: https://doi.org/10.1137/16m1098747
2018-01-01
Multiscale Modeling and Simulation
Abstract:In this paper, we consider the harmonic extension problem, which is widely used in many applications of machine learning. We formulate the harmonic extension as solving a Laplace-Beltrami equation with Dirichlet boundary condition. We use the point integral method (PIM) proposed in [Z. Li, Z. Shi, and J. Sun, Commun. Comput. Phys., 22 (2017), pp. 228-258; Z. Shi and J. Sun, Res. Math. Sci., to appear; Z. Li and Z. Shi, Multiscale Model. Simul., 14 (2016), pp. 874-905] to solve the Laplace-Beltrami equation. The basic idea of the PIM method is to approximate the Laplace equation using an integral equation, which is easy to discretize from points. Based on the integral equation, we found that the traditional graph Laplacian method (GLM) fails to approximate the harmonic functions near the boundary. One important application of the harmonic extension in machine learning is semisupervised learning. We run a popular semisupervised learning algorithm by Zhu, Ghahramani, and Lafferty [Machine Learning, Proceedings of the Twentieth International Conference (ICML 2003), 2003, Washington, DC, 2003, ARAI Press, Menlo Park, CA pp. 912-919] over a couple of well-known datasets and compare the performance of the aforementioned approaches. Our experiments show the PIM performs the best. We also apply PIM to an image recovery problem and show it outperforms GLM. Finally, on the model problem of the Laplace Beltrami equation with Dirichlet boundary, we prove the convergence of the point integral method.