My personal favorite from ICLR 2020. The paper shows on which conditions GNN can compute any function and that the product of depth*width of GNN should be of size ~n in order to compute popular statistics on graphs (e.g. diameter, vertex cover, coloring, etc.).