approximation-rates

This paper proves that graph neural networks with random node features can universally approximate permutation-invariant or equivariant functions on directed graphs, and provides approximation rate bounds for differentiable functions.