← Research overview

How Wide and How Deep? Mitigating Over-squashing of GNNs via Channel Capacity Constrained Estimation

Zinuo You, Jin Zheng, John Cartlidge
AAAI 2026

Choose GNN width and depth by treating representation learning as an information channel.

First author and lead implementation author · 2023–2025

How much information survives propagation?

Hidden width and depth determine how a graph neural network represents and compresses signals. Choosing them only by convention can leave the architecture poorly matched to the graph.

A capacity-based view

C3E models spectral graph propagation as a communication channel and formulates architecture selection as a constrained optimization problem. It connects graph structure, propagation, and representation size.

Reading the demonstration

A small path graph carries eight orthogonal signals. Lazy graph diffusion attenuates its spectral modes; width limits how many channels are kept. The displayed Gaussian-channel capacity proxy is calculated for this toy system. It is not the complete C3E estimator, and it does not equate over-squashing with over-smoothing or claim that every graph behaves this way.

Interactive toy experiment

Width, depth, and information capacity

Increase propagation depth to attenuate graph modes, or increase width to retain more channels.

Width, depth, and information capacityFixed eight-node path1234567812345678Spectral amplitude after propagationGraph mode

Blue: retained spectral channels · Gray outlines: discarded channels

4 of 8 modes retained · 3 propagation steps · Gaussian-channel capacity proxy: 4.982 bits.

How this toy works

On an eight-node path, symmetric lazy diffusion has λₖ = 0.5 + 0.5 cos(πk/8). Mode gain after L steps is λₖᴸ. Capacity proxy C = ½ Σ log₂(1 + 10 λₖ²ᴸ), summed over retained channels.

This Gaussian-channel toy is not the complete C3E estimator. It shows attenuation and dimensional truncation; it does not equate over-squashing with over-smoothing or predict task accuracy.

Synthetic illustration · Not a trained-model output or a published benchmark.