In a graph convolutional network, we assume that the graph $G$ is generated
wrt some observation noise. During learning, we make small random perturbations
$\Delta{}G$ of the graph and try to improve generalization. Based on quantum
information geometry, $\Delta{}G$ can be characterize