Limit theorems for eigenvectors of the normalized Laplacian for random graphs (Q1800805)

Distributional limit results are established for the eigenvectors corresponding to the largest eigenvalues of the adjacency or Laplacian matrix. It is shown that the Chernoff information between the multivariate normals approximation of the embedding characterizes the minimum error rate achievable by any clustering procedure that operates only on the spectral embedding. The contribution reinforces results concerning the prediction accuracy's improvements using the normalization from [\textit{P. Sarkar} and \textit{P. J. Bickel}, Ann. Stat. 43, No. 3, 962--990 (2015; Zbl 1320.62150)], ref. also [\textit{U. von Luxburg}, ``A tutorial on spectral clustering'', Stat. Comput. 17, No. 4, 395--416 (2007; \url{doi:10.1007/s11222-007-9033-z})].