[Paper Review] Strong Consistency, Graph Laplacians, and the Stochastic Block Model

Shaofeng Deng, Shuyang Ling+1 more|arXiv (Cornell University)|Apr 21, 2020

Complex Network Analysis Techniques42 references4 citations

TL;DR

This paper establishes conditions under which classical two-step spectral clustering using the graph Laplacian achieves strong consistency—exact recovery of hidden communities—in the stochastic block model. By deriving entrywise $∞$-norm perturbation bounds for the Fiedler eigenvector of both unnormalized and normalized Laplacians, the authors prove exact recovery is possible when the model parameters $(n,p,q)$ match the information-theoretic limits.

ABSTRACT

Spectral clustering has become one of the most popular algorithms in data clustering and community detection. We study the performance of classical two-step spectral clustering via the graph Laplacian to learn the stochastic block model. Our aim is to answer the following question: when is spectral clustering via the graph Laplacian able to achieve strong consistency, i.e., the exact recovery of the underlying hidden communities? Our work provides an entrywise analysis (an $\ell_{\infty}$-norm perturbation bound) of the Fielder eigenvector of both the unnormalized and the normalized Laplacian associated with the adjacency matrix sampled from the stochastic block model. We prove that spectral clustering is able to achieve exact recovery of the planted community structure under conditions that match the information-theoretic limits.

Motivation & Objective

• To determine the conditions under which spectral clustering via the graph Laplacian achieves strong consistency in community detection.
• To address the theoretical gap in understanding entrywise perturbation of the Fiedler eigenvector in spectral clustering.
• To close the gap between algorithmic performance and information-theoretic limits in the stochastic block model.
• To provide a rigorous entrywise ($\ell_\infty$-norm) analysis of the Fiedler eigenvector for both unnormalized and normalized Laplacians.
• To establish that spectral clustering can achieve exact recovery even when parameters approach the fundamental limits of detectability.

Proposed method

• Derives an entrywise $\ell_\infty$-norm perturbation bound for the Fiedler eigenvector of the unnormalized and normalized graph Laplacian under the stochastic block model.
• Uses a perturbation framework based on the matrix inverse and spectral gap analysis to control the deviation of the eigenvector from its population counterpart.
• Applies concentration inequalities and high-probability bounds on the adjacency matrix deviation $\|A - A^*\|$ to control eigenvector error.
• Introduces a localized perturbation analysis by conditioning on the removal of a single node, enabling control of the $\ell_\infty$-norm of the eigenvector difference.
• Combines bounds on the eigenvector norm $\|u_2\|_\infty = O(1/\sqrt{n})$ with concentration of the Laplacian action to derive tight error bounds.
• Employs a recursive decomposition of the eigenvector difference using the resolvent identity and matrix perturbation theory to achieve $O(\|u_2\|_\infty)$ error scaling.

Experimental results

Research questions

• RQ1 Under what conditions on $(n,p,q)$ does spectral clustering via the graph Laplacian achieve exact recovery of the hidden community structure in the stochastic block model?
• RQ2 Can spectral clustering achieve strong consistency when the model parameters are near the information-theoretic threshold for community detection?
• RQ3 What is the entrywise ($\ell_\infty$-norm) perturbation behavior of the Fiedler eigenvector of the graph Laplacian under the stochastic block model?
• RQ4 How does the normalized Laplacian compare to the unnormalized Laplacian in terms of eigenvector stability and recovery performance?
• RQ5 What is the role of the spectral gap and minimum degree in ensuring robustness of the eigenvector to adjacency matrix noise?

Key findings

• The Fiedler eigenvector of both unnormalized and normalized Laplacians satisfies $\|u_2\|_\infty = O(1/\sqrt{n})$ with high probability.
• The entrywise perturbation of the Fiedler eigenvector is bounded by $\|v\|_\infty = O(\|u_2\|_\infty)$, which is $O(1/\sqrt{n})$.
• The spectral clustering algorithm achieves exact recovery of the planted community structure when $p$ and $q$ are such that the model is within the information-theoretic limits.
• The perturbation bound for the normalized Laplacian is derived via a localized analysis, conditioning on node removal, to control $\ell_\infty$-norm deviations.
• The method establishes that $\|A(u_2 - u_2^*)\|_\infty = O(\log n / \sqrt{n} \log \log n)$ with high probability, enabling tight control on eigenvector deviation.
• The final result confirms strong consistency: spectral clustering recovers the true partition with probability $1 - o(1)$ when $p$ and $q$ are above the information-theoretic threshold.