△ Mechanism confirmed, baseline not beaten 2026

Diagonalizable Directed Message Passing

Implementation & benchmark of arXiv:2608.14439 — Positive Arc-Weight Design Makes Every Directed Laplacian Diagonalizable

Usefulness8/10

Difficulty6/10

Novelty7/10

Architecture Stability Accuracy Scalability Graph nn Ssm Training dynamics Spectral theory Linear algebra Graph theory Dynamical systems

Source paper: Positive Arc-Weight Design Makes Every Directed Laplacian Diagonalizable arXiv:2608.14439 ⓘ · analyzed Sep 1, 2026

AI-generated research hypothesis, automatically tested. Not peer-reviewed.

Idea description

Replace arbitrary directed-edge weights in a graph neural ODE or recurrent message-passing layer by weights constructed to make the directed Laplacian diagonalizable. This removes Jordan-block coupling, allowing the linearized graph dynamics to be represented as independent eigenmodes rather than modes with polynomial transients such as t^k exp(lambda t).

Formulas

$$A_{ij}=\begin{cases}\omega_{ji},&(v_j,v_i)\in E,\\0,&\text{otherwise},\end{cases}\qquad L=D-A,\qquad D_{ii}=\sum_{j:(v_j,v_i)\in E}\omega_{ji}.$$

$$L(\varepsilon)=L_D+\varepsilon R,\qquad q_\varepsilon(s)=\frac{\det(sI-L(\varepsilon))}{s^r}.$$

$$\operatorname{Disc}(q_\varepsilon)=0\Longleftrightarrow q_\varepsilon\text{ has a repeated nonzero eigenvalue},\qquad 0<\varepsilon<\varepsilon_c:=\inf\{\varepsilon>0:\operatorname{Disc}(q_\varepsilon)=0\}.$$

$$e^{-\tau L(\varepsilon)}=V\operatorname{diag}(e^{-\tau\lambda_1},\ldots,e^{-\tau\lambda_n})V^{-1}\quad\text{when }L(\varepsilon)=V\Lambda V^{-1}.$$

Mathematical statement

Let G=(V,E) be a weakly connected loopless digraph with positive arc weights omega_ij. The in-degree Laplacian is L=D-A, where A_ij=omega_ji when arc (v_j,v_i) exists and D_ii is the weighted indegree of vertex v_i. Select a spanning directed acyclic subgraph with one source per root strongly connected component. Assign its edges so all non-source weighted indegrees are distinct; its Laplacian L_D is triangular with distinct nonzero diagonal entries and semisimple zero eigenvalues. Restore every omitted edge with common weight epsilon, giving L(epsilon)=L_D+epsilon R. If q_epsilon(s)=det(sI-L(epsilon))/s^r is the reduced characteristic polynomial, then sufficiently small positive epsilon preserves simple nonzero roots. The first collision is detected by Disc(q_epsilon)=0.

Implementation notes

Integration point: use the construction before training a directed graph neural ODE, graph state-space model, or recurrent layer with propagation dh/dt=-tau Lh+f_theta(h,x), or residual update h_(k+1)=h_k-eta Lh_k+f_theta(h_k,x_k). Compute strongly connected components, choose a directed spanning forest with one source per root SCC, and topologically order its edges. Assign selected-edge weights greedily so every non-source weighted indegree is distinct, for example target indegrees d_i=i delta. Set every omitted edge to epsilon. Pseudocode: SCC=tarjan(G); T=choose_rooted_forest(G,SCC); omega=assign_positive_distinct_indegrees(T); Ld=laplacian(T,omega); R=laplacian(G\T,unit_weights); scan eps_grid; L=Ld+eps*R; eig=eigvals(L); gap=min_pairwise_nonzero_gap(eig); choose largest safe eps; run model. The paper supplies the forest construction and diagonalizability guarantee; code estimates eigenvalues, eigenvectors, condition number, and the discriminant numerically. First experiment: compare this design with unit and random weights on a directed GCN or graph neural ODE using Cora and synthetic directed stochastic block graphs. Sweep epsilon across the predicted collision boundary. Measure eigenvalue gaps, eigenvector condition number, modal reconstruction error, and hidden-state norm over long horizons. Prediction: below epsilon_c there are no nonzero eigenvalue collisions and reconstruction error is numerical precision; at the first discriminant root, the gap vanishes and conditioning sharply worsens. A defective baseline should show polynomial Jordan transients, while the designed model should remove them.

Verification

Mechanism confirmed, baseline not beaten

Stage 1 · Toy mechanism gate: Passed ✓

Stage 2 · Mechanism transferred to benchmark: Confirmed ✓

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 reproducible directed-Laplacian weighting experiment for a 3-node chain with a restored edge. The math predictions manifested: the reduced-polynomial discriminant predicted the first collision at epsilon_c=0.1715728753, and the numerical collision matched it with gap 5.0e-8 and discriminant 1.8e-15. The measured eigenvalue gap matched the predicted square-root discriminant within 4.8e-15; conditioning worsened from 8.68 at epsilon=0.1 to 1279.8 near the collision, while the unit-weight chain was defective with condition number 2.66e16.

Agent confidence: 9/10

Baseline: Unit-weight directed chain: repeated nonzero eigenvalues [1,1], discriminant 0, eigenvector condition 2.66e16, modal reconstruction relative error at t=10 = 3.77e-4.
Idea: Designed chain with restored epsilon=0.1: nonzero eigenvalues [1.2298438,1.8701562], gap 0.6403124, discriminant 0.4100000, eigenvector condition 8.68, modal reconstruction relative error at t=10 = 7.07e-16. Predicted and observed collision epsilon were both 0.1715728753.

Limitations: Only a 3-node synthetic graph was tested; no GNN training, graph neural ODE benchmark, Cora data, FLOP/speed comparison, larger SCC construction, or general discriminant root solver was evaluated. The transient norm remains affected by nonnormality even for diagonalizable matrices, so the primary evidence is from eigenvalue, discriminant, conditioning, and modal-reconstruction checks.

How to run: python3 experiment.py

Stage 2 — Benchmark vs. tuned baseline (arXiv:2608.14439) Sep 2, 2026

Ahead of the baseline (−7.8%) but not statistically significant (p=0.6236, wins 4 / 8 paired seeds).

Benchmark: Dynamics forecasting
Model: rnn_small
Paired seeds: 8
Baseline mean: 0.0007
Idea mean: 0.0006
Effect (Δ): -0.0001 (−7.8%; negative = idea better)
Wins: 4 / 8 paired seeds
p-value: 0.6236 (permutation test, 20 000 shuffles)
Smallest detectable effect: ±31.4%
Mechanism: Confirmed ✓
Practical effect: Inconclusive
Baseline tuning: swept over 3 configs

Limitations:

The experiment used the built-in dynamics pendulum track with 15 epochs and 400 training/200 test examples. The intervention applies a fixed Laplacian to the three input state channels before the GRU rather than implementing a full graph neural ODE or learned multi-node message-passing layer; larger SCC/forest construction, Cora-scale graphs, long-horizon rollout stability, and FLOP/speed measurements were not tested.

How to run: python3 stage2_bench.py

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

Stage 2 — Benchmark Sep 1, 2026 latest Failed ✗

Agent confidence: 9/10

The dynamics track is structurally matched to stability/control and was evaluated with the canonical rnn_small architecture. The diagonalizable directed-Laplacian intervention produced a small lower MSE, but not a significant win: paired delta_mean=-5.085e-05 and permutation p_value=0.6236. The trained-model mechanism signature was measurable and marked confirmed, while the mathematical collision/gap prediction also passed numerically.

Baseline: Tuned baseline lr=0.01: test MSE 0.000652883 ± 0.000169289 over 8 seeds.
Idea: Best idea epsilon=0.10, lr=0.01: test MSE 0.000602033 ± 0.000178326 over 8 seeds; 4/8 paired wins, delta=-0.0000508505, p=0.6236.

Benchmark result