[Paper Review] Nonparametric graphon estimation

Patrick J. Wolfe, Sofia C. Olhede|arXiv (Cornell University)|Sep 23, 2013

Stochastic processes and statistical mechanicsMathematics27 references137 citations

TL;DR

This paper introduces a nonparametric framework for estimating graphons—limit objects representing network structures—using profile likelihood methods. It establishes consistency and convergence rates for graphon estimation under general conditions, including sparse networks and model misspecification, linking the theory to approximation theory and graph limits.

ABSTRACT

We propose a nonparametric framework for the analysis of networks, based on a natural limit object termed a graphon. We prove consistency of graphon estimation under general conditions, giving rates which include the important practical setting of sparse networks. Our results cover dense and sparse stochastic blockmodels with a growing number of classes, under model misspecification. We use profile likelihood methods, and connect our results to approximation theory, nonparametric function estimation, and the theory of graph limits.

Motivation & Objective

• Address the lack of flexible, theoretically grounded tools for statistical network analysis, especially in sparse network regimes.
• Develop a nonparametric framework for network analysis based on the graphon, a limit object from graph limit theory.
• Establish theoretical consistency and convergence rates for graphon estimation under general conditions, including model misspecification.
• Connect network estimation to approximation theory, nonparametric function estimation, and the theory of graph limits.
• Provide a foundation for coherent statistical inference in large-scale network data using exchangeable random graph models.

Proposed method

• Model networks via an adjacency matrix $ A $ with entries $ A_{ij} \sim \text{Bernoulli}(\rho_n f(\xi_i, \xi_j)) $, where $ \xi_i \sim \text{Uniform}(0,1) $, and $ f $ is the graphon.
• Use profile likelihood estimation to infer the graphon $ f $, treating the latent variables $ \xi_i $ as nuisance parameters.
• Apply a blockmodel approximation to $ f $, using a partition of $[0,1]^2$ into $ k \times k $ blocks with constant values, enabling nonparametric smoothing.
• Establish convergence rates by relating estimation error to the $ L^2 $-norm of the difference between true and estimated graphons.
• Leverage tools from approximation theory and empirical process theory to control bias and variance in graphon estimation.
• Use the invariance of the graphon under measure-preserving transformations to ensure identifiability up to equivalence classes.

Experimental results

Research questions

• RQ1 Can a nonparametric framework for graphon estimation achieve consistency and convergence rates in both dense and sparse network regimes?
• RQ2 How does the estimation performance of graphons behave under model misspecification, particularly for stochastic blockmodels with growing numbers of blocks?
• RQ3 What is the theoretical relationship between graphon estimation and approximation theory, especially in terms of bias-variance trade-offs?
• RQ4 How can profile likelihood methods be adapted to handle latent variables $ \xi_i $ in exchangeable random networks?
• RQ5 What are the convergence rates of graphon estimators when the network sparsity $ \rho_n \to 0 $ as $ n \to \infty $?

Key findings

• The proposed nonparametric graphon estimator is consistent under general conditions, including for sparse networks where $ \rho_n \to 0 $.
• Convergence rates are established for both dense and sparse network regimes, with explicit dependence on the smoothness of the graphon and the number of blocks in the approximation.
• The framework covers stochastic blockmodels with a growing number of blocks, even under model misspecification.
• The estimation error is bounded via a comparison to the $ L^2 $-norm of the difference between the true and estimated graphon, with rates derived using Taylor expansions of the Kullback-Leibler divergence.
• The method achieves consistency by leveraging the invariance of the graphon under measure-preserving transformations, ensuring identifiability up to equivalence classes.
• Theoretical results connect graphon estimation to nonparametric function estimation and approximation theory, providing a unified foundation for statistical network analysis.