1: \begin{abstract}

2: This paper derives statistical guarantees for the performance of Graph Neural Networks (GNNs) in link prediction tasks on graphs generated by a graphon. We propose a linear GNN architecture (LG-GNN) that produces consistent estimators for the underlying edge probabilities. We establish a bound on the mean squared error and give guarantees on the ability of LG-GNN to detect high-probability edges. Our guarantees hold for both sparse and dense graphs. Finally, we demonstrate some of the shortcomings of the classical GCN architecture, as well as verify our results on real and synthetic datasets. 

3: 

4: % and demonstrate how LG-GNN circumvents these problems. 

5: 

6: 

7: % We establish rates of convergence for estimators of these edges probabilities.  Then, we prove statistical guarantees that bound the convergence rate of these estimators, and compare the performance to that of the traditional GCN architecture. 

8: \end{abstract}

9: