Abstract: Graph Neural Networks (graph NNs) are a promising deep learning approach for analyzing graph-structured data. However, it is known that they do not improve (or sometimes worsen) their predictive performance as we pile up many layers and add non-lineality. To tackle this problem, we investigate the expressive power of graph NNs via their asymptotic behaviors as the layer size tends to infinity. Our strategy is to generalize the forward propagation of a Graph Convolutional Network (GCN), which is a popular graph NN variant, as a specific dynamical system. In the case of a GCN, we show that when its weights satisfy the conditions determined by the spectra of the (augmented) normalized Laplacian, its output exponentially approaches the set of signals that carry information of the connected components and node degrees only for distinguishing nodes. Our theory enables us to relate the expressive power of GCNs with the topological information of the underlying graphs inherent in the graph spectra. To demonstrate this, we characterize the asymptotic behavior of GCNs on the Erdős–Rényi graph. We show that when the Erdős–Rényi graph is sufficiently dense and large, a broad range of GCNs on it suffers from the "information loss" in the limit of infinite layers with high probability. Based on the theory, we provide a principled guideline for weight normalization of graph NNs. We experimentally confirm that the proposed weight scaling enhances the predictive performance of GCNs in real data. Code is available at https://github.com/delta2323/gnn-asymptotics.
Introduction
Graph Neural Networks (GNNs) have revolutionized the analysis of graph-structured data. However, unlike standard architectures in computer vision or natural language processing, GNNs often suffer from performance degradation as depth increases. This phenomenon, known as over-smoothing, causes node features to become increasingly similar until they are indistinguishable. While empirically well-documented, a formal mathematical characterization of this collapse—especially for non-linear networks—has remained elusive. This work provides a rigorous theoretical explanation by modeling Graph Convolutional Networks (GCNs) as dynamical systems.
GCNs as Dynamical Systems
The core of our approach is to view the forward propagation of a GCN as a discrete dynamical system. In this framework, node features evolve over "time" (layers). We identify a specific invariant subspace, $\mathcal{M}$, which contains signals carrying only basic topological information, such as connected components and node degrees. For tasks like node classification, signals falling into $\mathcal{M}$ are effectively "information-less" as they cannot distinguish between nodes within the same component.
The Information Loss Theorem
By combining spectral graph theory with the properties of the ReLU activation function, we prove that node representations exponentially approach the subspace $\mathcal{M}$. Specifically, the distance between the network's output and this "dead zone" shrinks at a rate of $O((s\lambda)^L)$, where $L$ is the number of layers, $s$ is the maximum singular value of the weights, and $\lambda$ is determined by the graph's Laplacian spectrum.
This theorem reveals that "information loss" is an exponential certainty when $s\lambda < 1$. Furthermore, the rate of loss is tied to the "spectral gap" of the graph; dense graphs or expander graphs, which mix information quickly, cause GCNs to over-smooth much faster than sparse networks.
Weight Scaling Guideline and Results
To maintain expressive power in deep GCNs, we propose a principled weight-scaling strategy: weights must be initialized and maintained such that $s > \lambda^{-1}$.
Empirical validation on both synthetic Erdős–Rényi graphs and real-world citation networks (Cora, CiteSeer, PubMed) confirms these findings. In "Noisy Cora" experiments—where random edges increase graph density and accelerate over-smoothing—GCNs using our proposed scaling maintained significantly higher accuracy at depth compared to standard unnormalized versions. By applying this principled weight scaling, we can "rescue" deep GCNs that would otherwise fail, achieving comparable or better performance than shallow networks.
Conclusion and Limitations
Our work provides a mathematical foundation for understanding the depth limitations of GCNs. While our current theory focuses on ReLU activations and node-wise updates, it offers a clear path forward for designing more expressive deep GNNs. Future work will explore the impact of other activation functions like Leaky ReLU or ELU, graph-level pooling operations, and the complex interaction between residual connections and weight scaling.
Summary updated