Finite-Time Node Separation in Recurrent GNNs with Gaussian Perturbations

A new arXiv paper examines how persistent Gaussian noise affects recurrent graph neural networks. While such perturbations are known to keep a positive stationary Dirichlet energy and thus avoid asymptotic oversmoothing, the authors show that this global bound alone does not ensure individual nodes stay distinguishable. The work analyzes finite-time separation between node representations under these random perturbations.