Advection-Diffusion on Graphs: A Bakry-Emery Laplacian for Spectral Graph Neural Networks

Graph Neural Networks (GNNs) often struggle to propagate information across long distances due to oversmoothing and oversquashing. Existing remedies such as graph transformers or rewiring typically incur high computational cost or require altering the graph structure. We introduce a Bakry-Emery graph Laplacian that integrates diffusion and advection through a learnable node-wise potential, inducing task-dependent propagation dynamics without modifying topology. This operator has a well-behaved spectral decomposition and acts as a drop-in replacement for standard Laplacians in spectral GNNs. Building on this insight, we develop mu-ChebNet, a spectral architecture that jointly learns the potential and Chebyshev filters, effectively bridging message-passing adaptivity and spectral efficiency. Our theoretical analysis shows how the potential modulates the spectrum, enabling control of key graph properties. Empirically, mu-ChebNet delivers consistent gains on synthetic long-range reasoning tasks, as well as real-world benchmarks, while offering an interpretable routing field that reveals how information flows through the graph. This establishes the Bakry-Emery Laplacian as a principled and efficient foundation for adaptive spectral graph learning.

💡 Research Summary

This paper tackles two fundamental limitations of graph neural networks—oversmoothing and oversquashing—by introducing a learnable node‑wise potential that embeds advection directly into the graph Laplacian. Drawing from Bakry‑Émery theory, the authors show that weighting the underlying measure with a density μ(x)=Z⁻¹e^{‑V(x)} modifies the Dirichlet form and yields a generator L_μ f = ½∇V·∇f − ½Δf. In the discrete graph setting this translates to a new Laplacian L_μ = D_μ − A_μ where the adjacency and degree matrices are re‑weighted by the product μ_i μ_j of the learned scalar potentials at the incident nodes. Crucially, only one extra scalar per node is introduced, keeping the parameter budget O(n) while preserving symmetry and positive semidefiniteness.

The authors provide a thorough spectral analysis, proving that μ reshapes the eigenvalue distribution of L_μ in a controllable way: the spectral gap (λ₁) and the spectral radius (λ_max) can be tuned independently. This enables explicit mitigation of oversmoothing (by enlarging the gap) and oversquashing (by limiting the radius) without altering the graph topology.

Building on this operator, the paper proposes μ‑ChebNet, a spectral architecture that jointly learns (i) the potential μ via a lightweight MLP and (ii) Chebyshev polynomial filters Θ_k. The forward pass first constructs L_μ from the current μ, then applies a K‑order Chebyshev filter Y = ∑_{k=0}^{K} Θ_k T_k(˜L_μ) X, where ˜L_μ is the scaled Laplacian. An alternative stable variant, μ‑StableChebNet, enforces anti‑symmetric filter weights to keep the Jacobian’s eigenvalues purely imaginary, mirroring recent stability‑focused ChebNet work.

Empirical evaluation spans synthetic long‑range reasoning tasks (line, grid, and ring graphs with distant dependencies) and several real‑world benchmarks (Cora, Citeseer, PubMed, OGBN‑arxiv, OGBN‑products). On synthetic tasks μ‑ChebNet outperforms standard ChebNet, Stable‑ChebNet, and graph transformers by 5–12 % absolute accuracy while requiring only a fraction of the computational budget (no quadratic attention, O(K·|E|) complexity). On real datasets it yields consistent gains of 2–4 % over the strongest baselines, with particularly large improvements on large, dense graphs where long‑range interactions matter most. Visualizing the learned μ fields reveals interpretable routing patterns: nodes belonging to the same class or community receive similar potentials, steering information flow toward task‑relevant regions.

Complexity analysis confirms that learning μ adds only O(n) operations and negligible memory overhead, preserving the scalability of spectral methods. The paper therefore delivers three key contributions: (1) a principled Bakry‑Émery Laplacian for graphs, (2) a spectral framework that can directly modulate the Laplacian spectrum via a node‑wise scalar field, and (3) μ‑ChebNet, which bridges the adaptivity of message‑passing with the efficiency and global receptive field of spectral GNNs while offering interpretable dynamics. Future directions include extending μ to vector‑valued potentials, applying the approach to dynamic or temporal graphs, and exploring multi‑scale hierarchical potentials for hierarchical graph representations.