Dirichlet Energy Preservation in Deep Graph Neural Networks

32 concepts21 papers1 proposal

Imagine a drop of ink falling into a glass of water. At first, the ink forms intricate, highly distinct patterns. But as time passes, the ink diffuses outward. Eventually, the water turns a uniform, featureless gray.

This is exactly what happens inside a deep Graph Neural Network (GNN). Standard message passing acts like a global diffusion process. As representations pass through layer after layer, nodes aggregate information from their neighbors, then their neighbors' neighbors, until every node's embedding looks identical to every other node's embedding. This phenomenon is known as oversmoothing, and it prevents traditional GNNs from scaling effectively to deeper architectures.

To diagnose and combat this collapse, researchers use a mathematical metric called Dirichlet energy. Intuitively, Dirichlet energy measures the variance or difference between connected node features across a graph's topology. If the energy is high, the graph signal is crisp and heterogeneous; if the energy collapses to zero, the signal has completely smoothed out into a uniform consensus. As demonstrated by Dai Shi et al. in Revisiting Generalized p-Laplacian Regularized Framelet GCNs: Convergence, Energy Dynamic and as Non-Linear Diffusion, maintaining non-zero Dirichlet energy during graph signal propagation is the key to preventing oversmoothing and keeping node features distinguishable at extreme depths.


Structural Interventions: Pruning and Rewiring

If a graph's topology forces too much mixing, one intuitive solution is to change the graph itself. By strategically removing or adding edges, we can slow down the rate at which energy decays.

For instance, Valentina Giunchiglia et al. developed a framework in Towards Training GNNs using Explanation Directed Message Passing that restricts message passing to high-importance edges. By pruning irrelevant edges that contribute to global smoothing, the architecture maintains higher Dirichlet energy across layers.

Taking a spectral perspective, Adarsh Jamadandi et al. show in Spectral Graph Pruning Against Over-Squashing and Over-Smoothing that optimizing graph structure specifically to slow the decay of Dirichlet energy allows deep networks to sustain high-frequency feature components that standard message passing typically washes out.

Topology can also be adjusted using geometric principles. Lukas Fesser et al. in Mitigating Over-smoothing and Over-squashing using Augmentations of Forman-Ricci Curvature leverage local curvature to guide graph rewiring. This strategic adjustment preserves feature diversity by maintaining higher Dirichlet energy across deep layers, delaying the convergence of node representations.


Beyond Pure Diffusion: Waves, Reactions, and Phase Transitions

Instead of just fixing the graph structure, what if we change the underlying physics of how information propagates? Standard GNNs mimic pure heat diffusion, which inherently minimizes Dirichlet energy. A powerful alternative is introducing counter-forces that actively pump energy back into the system.

One approach is to introduce repulsive forces alongside attractive ones. Yuelin Wang et al. design a system in ACMP: Allen-Cahn Message Passing with Attractive and Repulsive Forces for Graph Neural Networks inspired by the Allen-Cahn phase transition equation. By maintaining a strictly positive lower bound on Dirichlet energy, their model prevents node features from converging to a single average value, enabling GNNs to scale to over 100 layers without losing discriminative power.

Another method is to couple diffusion with a reaction term. Moshe Eliasof et al. introduce this in Graph Neural Reaction Diffusion Models. While standard diffusion causes Dirichlet energy to collapse toward zero, their learnable reaction function actively increases this energy to sharpen node features, allowing the network to capture high-frequency signals.

Alternatively, we can abandon diffusion entirely in favor of wave-like propagation. Christian Nauck et al. propose this in Dirac-Bianconi Graph Neural Networks - Enabling Non-Diffusive Long-Range Graph Predictions. By basing their architecture on the topological Dirac equation, they preserve Dirichlet energy over many propagation steps, maintaining feature heterogeneity across deep architectures.


Architectural Safeguards and Energy Bounding

Finally, we can build safeguards directly into the neural network layers via gating, normalization, and geometric constraints to ensure that Dirichlet energy remains stable.

Gating and Residuals

T. Konstantin Rusch et al. demonstrate in Gradient Gating for Deep Multi-Rate Learning on Graphs that preserving Dirichlet energy via gradient gating prevents the exponential collapse of feature diversity. Similarly, AJAY KUMAR JAISWAL et al. show in Old can be Gold: Better Gradient Flow can Make Vanilla-GCNs Great Again that maintaining Dirichlet energy through isometric transformations and topology-aware initialization allows vanilla GCNs to scale to 64 layers. To extend this to hierarchical data, Yuwei Wu et al. utilize Dirichlet energy in Residual Hyperbolic Graph Convolution Networks as a formal metric to ensure node signals remain distinguishable across deep hyperbolic layers.

Bilateral Bounding

Energy preservation does not mean letting energy explode into noise. Jialong Chen et al. argue in Graph Neural Ricci Flow: Evolving Feature from a Curvature Perspective that discrete Ricci flow provides bilateral bounds on Dirichlet energy. This keeps the energy within a fixed range, preventing the representation from becoming either excessively smooth or overly noisy. In a similar vein, Anna Bison in A Spectral Interpretation of Redundancy in a Graph Reservoir manages graph signal evolution by navigating the gradient of Dirichlet energy, alternating between energy-decreasing (smoothing) and energy-increasing (sharpening) steps to prevent signal collapse.

Normalization and Boundary Conditions

Li Sun et al. shift the energy decay rate from exponential to polynomial in Deeper with Riemannian Geometry: Overcoming Oversmoothing and Oversquashing for Graph Foundation Models by introducing nonhomogeneous boundary conditions. To prevent energy from vanishing entirely, Arseny Skryagin et al. introduce localized normalization and activation in Graph Neural Networks Need Cluster-Normalize-Activate Modules, providing a theoretical safeguard against the collapse of node distinctiveness.

Regularization and Adaptive Depth

Xiaojun Ma et al. propose local energy regularization in LEReg: Empower Graph Neural Networks with Local Energy Regularization to explicitly minimize feature variance within predicted classes while maximizing it between them. Yuankai Luo et al. reveal in Beyond Random Masking: When Dropout meets Graph Convolution Networks that dropout prevents oversmoothing precisely by increasing Dirichlet energy and maintaining average pairwise distances. Finally, Federico Errica et al. demonstrate in Adaptive Message Passing: A General Framework to Mitigate Oversmoothing, Oversquashing, and Underreaching that maintaining high Dirichlet energy at significant depths serves as a reliable indicator that a model can successfully preserve distinct node features.

Key Foundational Work

Towards Training GNNs using Explanation Directed Message Passing (2022) — contributed 2 concept elements to this direction.
Old can be Gold: Better Gradient Flow can Make Vanilla-GCNs Great Again (2022) — contributed 2 concept elements to this direction.
Dirichlet Energy Constrained Learning for Deep Graph Neural Networks (2021) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_dirichlet_energy_preservation_87c4b2e9,
  title = {Dirichlet energy preservation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/dirichlet-energy-preservation/rdr_87c4b2e945dd49fe9623a41cf7f242a2},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_dirichlet_energy_preservation_87c4b2e9_v1,
  title = {Dirichlet energy preservation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/dirichlet-energy-preservation/rdr_87c4b2e945dd49fe9623a41cf7f242a2/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake