Rate-Optimal Spectral Likelihood-Ratio Clustering in Stochastic Block Models
Abstract
Spectral clustering commonly assigns communities by clustering a low-dimensional representation, whereas likelihood refinement can improve an initial assignment using the observed edges. We ask whether sharp community-recovery accuracy can be retained when the refinement no longer accesses individual edges. We develop spectral likelihood-ratio clustering for stochastic block models using one regularized adjacency eigensystem and the original vertex degrees. The method estimates connection probabilities and updates each vertex's label by a likelihood-ratio classifier on reconstructed rows. The refinement accepts any sufficiently accurate initialization, including one computed from the same graph. Its expected misclassification proportion is bounded by a constant times the error of a raw-edge oracle classifier, plus the probability that initialization misses the required accuracy. Repeated k-means++ on the retained spectral coordinates supplies one complete procedure with a constant-factor oracle guarantee. The oracle knows the model parameters and the other labels. The guarantee holds for a fixed number of communities, comparable positive block probabilities and community sizes, and a nondegenerate normalized block matrix. It allows arbitrarily slowly diverging expected degrees under the stated upper sparsity condition. We explain the likelihood-ratio construction first, then the regularization and exact-degree correction required by the general result. The proof combines sharp local testing, truncated spectral-row estimates and leave-one-out comparisons.
Related
- Hierarchical clustering of X-Ray Fluorescence images using a dissimilarity based on the likelihood ratio statistic
- A Likelihood-Ratio Test for Verifying Weak Stochastic Transitivity
- Generalized Stochastic Approximation of the Log-Likelihood Ratio for Robust Sequential Change-Point Detection
- Is Spectral Self-Supervised Learning a Minimax Optimal Method of Density Ratio Estimation?
- Maximum Likelihood Estimation and Likelihood Ratio Tests for the Power Spherical Distribution
BibTeX
@misc{zhou2026rateoptimal,
title = {{Rate-Optimal Spectral Likelihood-Ratio Clustering in Stochastic Block Models}},
author = {Zhixin Zhou},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2610.40531},
eprint = {2610.40531},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.40531},
}
RIS
TY - UNPB
TI - Rate-Optimal Spectral Likelihood-Ratio Clustering in Stochastic Block Models
AU - Zhixin Zhou
PY - 2026
DA - 2026/10/29
PB - arXiv
DO - 10.48550/arxiv.2610.40531
UR - https://arxiv.org/abs/2610.40531
AB - Spectral clustering commonly assigns communities by clustering a low-dimensional representation, whereas likelihood refinement can improve an initial assignment using the observed edges. We ask whether sharp community-recovery accuracy can be retained when the refinement no longer accesses individual edges. We develop spectral likelihood-ratio clustering for stochastic block models using one regularized adjacency eigensystem and the original vertex degrees. The method estimates connection probabilities and updates each vertex's label by a likelihood-ratio classifier on reconstructed rows. The refinement accepts any sufficiently accurate initialization, including one computed from the same graph. Its expected misclassification proportion is bounded by a constant times the error of a raw-edge oracle classifier, plus the probability that initialization misses the required accuracy. Repeated k-means++ on the retained spectral coordinates supplies one complete procedure with a constant-factor oracle guarantee. The oracle knows the model parameters and the other labels. The guarantee holds for a fixed number of communities, comparable positive block probabilities and community sizes, and a nondegenerate normalized block matrix. It allows arbitrarily slowly diverging expected degrees under the stated upper sparsity condition. We explain the likelihood-ratio construction first, then the regularization and exact-degree correction required by the general result. The proof combines sharp local testing, truncated spectral-row estimates and leave-one-out comparisons.
ER -