The authors establish logical characterizations of the expressiveness of certain classes of graph neural networks (GNNs), and use these characterizations to determine the decidability of key verification problems for GNNs.
