Idea description

Replace ordinary graph message passing by diffusion over a simplicial complex or hypergraph, using incidence matrices to propagate information through nodes, edges, and higher-order faces. Mix the local higher-order walk with a teleportation operator so that the layer remains globally connected and avoids the slow mixing or oversmoothing caused by poorly connected complexes.

Formulas

$$A_d(\sigma_i^{(d-1)},\sigma_j^{(d)})=\mathbf{1}\{\sigma_i^{(d-1)}\subset\sigma_j^{(d)}\},\qquad k_{i,up}^{(d-1)}=\sum_{j=1}^{N_d}A_d(i,j),\qquad k_{j,down}^{(d)}=\sum_{i=1}^{N_{d-1}}A_d(i,j)=d+1.$$

$$P_{loc}=D_{d-1}^{-1}A_dD_d^{-1}A_d^{\mathsf T},\qquad P_\alpha=(1-\alpha)P_{loc}+\alpha R,\qquad H_{\ell+1}=\phi\!\left(P_\alpha H_\ell W_\ell+b_\ell\right),$$

$$\rho\!\left(P_\alpha-\mathbf{1}\pi_\alpha^{\mathsf T}\right)\leq 1-\alpha\quad\text{for uniform teleportation},$$

$$t^{(g)}=(I-Q_\alpha)^{-1}\mathbf{1},\qquad Q_\alpha=P_\alpha[\mathcal{V}\setminus\{g\},\mathcal{V}\setminus\{g\}],$$

Mathematical statement

Let $\mathcal{X}$ be a $D$-dimensional simplicial complex with $N_d$ $d$-simplices $\sigma_i^{(d)}$. The incidence matrix $A_d\in\{0,1\}^{N_{d-1}\times N_d}$ is defined by $A_d(i,j)=1$ iff $\sigma_i^{(d-1)}\subset\sigma_j^{(d)}$. Its upper degree is $k_{i,up}^{(d-1)}=\sum_j A_d(i,j)$, while the lower degree of a $d$-simplex is $k_{j,down}^{(d)}=\sum_i A_d(i,j)=d+1$. Construct a row-stochastic local diffusion matrix $P_{loc}$ from normalized incidence products, for example $P_{loc}=D_{d-1}^{-1}A_dD_d^{-1}A_d^{\mathsf T}$, with self-loops added if needed for aperiodicity. Let $R$ be a row-stochastic teleportation matrix, such as uniform global mixing $R=\mathbf{1}\mathbf{1}^{\mathsf T}/N_0$ or a learned feature-dependent kernel. The local-to-long-range interpolation is $P_\alpha=(1-\alpha)P_{loc}+\alpha R$, where $0\leq\alpha\leq1$. Because both terms are stochastic, $P_\alpha$ is stable in the probability norm and, for uniform teleportation, all nonprincipal eigenvalues are contracted by approximately $1-\alpha$.

Implementation notes

Integrate this as a message-passing layer for hypergraphs or simplicial complexes. Start with a dataset containing explicit hyperedges, or convert a graph into a clique complex and retain nodes, edges, and triangles. Precompute sparse incidence matrices $A_d$, degree vectors, and the normalized local operator $P_{loc}$. At each layer, choose $R$ as either uniform global mixing, a sparse random landmark matrix, or a learned low-rank stochastic kernel. Compute $H_{next}=\phi(((1-\alpha)P_{loc}+\alpha R)HW+b)$; use residual addition if the nonlinear layer is unstable. Pseudocode is: build incidence matrices; compute degrees with an epsilon floor; normalize $P_{loc}$; set $P=(1-alpha)P_{loc}+alpha R$; propagate $H=P@H@W+b$; apply normalization and activation; optionally estimate the second eigenvalue by ten power iterations. The incidence construction and interpolation rule come from the paper's mechanism; the exact neural weights, residual path, and choice of $R$ are engineering adaptations. First run a cheap node-classification experiment on Cora, a small hypergraph benchmark, and a synthetic two-community simplicial complex. Compare GCN/GAT against the same architecture with $\alpha\in\{0,.01,.03,.1,.3\}$. The quantitative prediction is that the measured nonprincipal spectral radius decreases approximately linearly as $1-\alpha$, while feature mixing time and oversmoothing decrease sharply once $\alpha$ exceeds the inverse graph mixing time. On the synthetic complex, mean first-passage time to a target community should decrease monotonically with $\alpha$; test whether validation accuracy peaks before the fully global limit $\alpha=1$, indicating a locality-globality tradeoff rather than a generic parameter-count gain.

Verification

Failed on benchmark

Stage 1 · Toy mechanism gate: Passed ✓

Stage 2 · Mechanism transferred to benchmark: Not reproduced ✗

Stage 2 · Practical benchmark result: Inconclusive

Methodology: 8 paired seeds, permutation test (p<0.05) against a learning-rate-tuned baseline. How verification works

Stage 1 — Mechanism check agent confidence 9/10

Built a simplicial diffusion layer using triangle-node incidence matrices, normalized local walks, uniform teleportation, MFPT computation, and a small classification probe. The mechanism manifested: spectral radius matched rho_alpha=rho_0(1-alpha) with maximum absolute error 1.9e-15, slow-mode mixing at 12 steps matched prediction within 1.4e-14, and mean MFPT decreased monotonically from 528.45 at alpha=0 to 71.66 at alpha=0.6. In the toy classifier, accuracy improved from 0.617 for local diffusion to 0.817 at alpha=0.6, although this was only a synthetic experiment.

Agent confidence: 9/10

Baseline: alpha=0 local diffusion: spectral radius 0.993438; mean MFPT 528.451; classification accuracy 0.617; feature std after 8 steps 0.8761
Idea: alpha=0.6 teleporting diffusion: spectral radius 0.397375, exactly predicted by 0.993438*(1-0.6); mean MFPT 71.661; classification accuracy 0.817; feature std 0.000575

Limitations: Only a regular synthetic cyclic triangle complex and a tiny fixed-seed ridge classifier were tested. No Cora, hypergraph benchmark, learned or sparse teleportation kernel, GPU implementation, runtime/scalability benchmark, multi-seed statistics, or comparison with GCN/GAT was evaluated.

How to run: python3 verify_and_report.py

Stage 2 — Benchmark vs. tuned baseline (arXiv:2607.29450) Sep 2, 2026

Ahead of the baseline (−4.8%) but not statistically significant (p=0.1439, wins 6 / 8 paired seeds).

Benchmark: Geometric_graph_diffusion
Model: graph_diffusion_net
Paired seeds: 8
Baseline mean: 0.0177
Idea mean: 0.0169
Effect (Δ): -0.0009 (−4.8%; negative = idea better)
Wins: 6 / 8 paired seeds
p-value: 0.1439 (permutation test, 20 000 shuffles)
Smallest detectable effect: n/a
Mechanism: Not confirmed ✗
Practical effect: Inconclusive
Baseline tuning: swept over 0 configs

Limitations:

The registered benchmark uses a synthetic geometric kNN graph track rather than a real-world simplicial-complex dataset. The intervention uses node-level weighted graph diffusion reconstructed from the registered track; explicit higher-order triangle incidence was not available in that track. The complete baseline sweep and exact baseline per-seed values are retained in bench_report.json.

How to run: python3 registered_bench.py

Verdict computed by deterministic test code from paired-seed statistics — not by the language model.

Stage 2 — Benchmark Aug 31, 2026 latest Failed ✗

Agent confidence: 9/10

On the registered geometric_graph_diffusion benchmark, teleporting diffusion at alpha=0.1 reduced mean MSE from approximately 0.01771914 to 0.01686609, but the paired permutation test gave p=0.14395, so the improvement is not significant. The mechanism signature was also not confirmed: trained-model contraction differed from the nominal 1-alpha prediction by up to 0.39057. Therefore the system correctly reports no significant win.

Baseline: Registered track geometric_graph_diffusion; swept alpha=0, lr in {0.003,0.01,0.03}, weight decay in {0,0.0001}, 20 epochs; selected lr=0.03, wd=0.0001. Full 8-seed mean MSE approximately 0.01771914.
Idea: Registered track, same architecture and selected training hyperparameters, alpha sweep {0.03,0.1,0.3}; best alpha=0.1, mean MSE 0.01686609, paired delta -0.00085305, 6/8 wins, p=0.14395.

Benchmark result