A Note on Graphon-Signal Analysis of Graph Neural Networks
This work analyzes message passing graph neural networks (MPNNs) by translating their input into a space of attributed graphons, known as 'graph-signals.' The authors establish generalization bounds and sampling lemmas for these models. This framework extends analysis to multidimensional signals and proves key properties like Lipschitz continuity and commutativity, providing a rigorous theoretical foundation for understanding MPNN performance.