✗ Failed on benchmark 2026

Discriminant-Gated Positive Edge Adaptation

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

Usefulness7/10

Difficulty7/10

Novelty8/10

Regularization Stability Accuracy Graph nn Optimizer Regularization Training dynamics Spectral theory Linear algebra Graph theory Optimization 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 Not peer-reviewed.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

Make directed edge weights trainable while constraining optimization to remain away from eigenvalue collisions of the graph Laplacian. The network can learn task-specific interaction strengths while preserving a measurable diagonalizability margin and avoiding ill-conditioned modal dynamics.

Formulas

$$\omega_e(\theta_e)=\omega_{\min}+\operatorname{softplus}(\theta_e),\qquad L(\theta)=D(\theta)-A(\theta).$$

$$q_\theta(s)=\frac{\det(sI-L(\theta))}{s^r},\qquad \operatorname{Disc}(q_\theta)=0\Longleftrightarrow\exists s:\ q_\theta(s)=q_\theta'(s)=0.$$

$$g(\theta)=\min_{i\ne j,\ \lambda_i,\lambda_j\ne0}\frac{|\lambda_i-\lambda_j|}{1+|\lambda_i|+|\lambda_j|},\qquad \mathcal{L}_{\mathrm{gap}}=\beta[\max(0,g_{\min}-g(\theta))]^2.$$

$$\theta_{k+1}=\theta_k-\eta\nabla_\theta\left(\mathcal{L}_{\mathrm{task}}+\mathcal{L}_{\mathrm{gap}}+\gamma\log(1+\kappa(V(\theta)))\right).$$

Mathematical statement

For trainable edge parameters theta_e, use strictly positive weights omega_e=omega_min+softplus(theta_e), and construct L(theta). Let r be the algebraic multiplicity of the zero eigenvalue and q_theta(s)=det(sI-L(theta))/s^r. A repeated nonzero eigenvalue occurs when q_theta and its derivative share a root, equivalently Disc(q_theta)=0. In practice define the normalized nonzero eigenvalue gap g(theta) as the minimum pairwise distance divided by the eigenvalue magnitudes. Penalize violations of g(theta)>=g_min and optionally penalize the eigenvector condition number kappa(V), because near-collisions produce unstable modal coordinates even before exact defectiveness.

Implementation notes

Integration point: attach trainable positive edge gates to a directed GCN, graph transformer bias, graph neural ODE, or graph state-space transition. Use omega_e=omega_min+softplus(theta_e), with omega_min small but nonzero. Recompute the Laplacian spectrum every K optimizer steps, initially K=10, rather than for every minibatch. Pseudocode: omega=softplus(theta)+omega_min; L=build_in_laplacian(omega); eig,V=eig(L); nonzero=remove_zero_modes(eig); gap=min_normalized_pairwise_gap(nonzero); barrier=relu(g_min-gap)^2; cond=log(1+condition_number(V)); loss=task_loss+beta*barrier+gamma*cond; backward(loss); optimizer.step(). For graphs with at most 64 nodes, additionally compute the characteristic polynomial and its discriminant; for larger graphs use Arnoldi or stochastic subspace iteration to estimate the smallest eigenvalue gap. The paper supplies positivity and the repeated-root discriminant boundary; numerical eigensolvers estimate gaps, condition numbers, and gradients. First experiment: train a two-layer directed GCN on synthetic directed stochastic block graphs and Cora, comparing unconstrained positive gates, fixed paper-designed weights, and gap-gated gates. Sweep g_min over 1e-4, 1e-3, and 1e-2. Prediction: as g(theta) approaches zero, eigenvector conditioning and modal reconstruction error rise sharply; the gated model maintains g(theta)>=g_min. On small graphs, the first exact loss of diagonalizability must coincide with Disc(q_theta)=0 within numerical tolerance. A useful falsification is failure to observe this spectral transition or absence of improved long-horizon stability despite preserved gaps.

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 positive-weight directed-Laplacian toy implementing softplus edge weights, normalized spectral-gap measurement, discriminant evaluation, conditioning diagnostics, and gap-gated optimization. The sweeps matched the predictions: gap scales as epsilon^1 (observed 1.000000), the nonzero characteristic-polynomial discriminant scales as epsilon^2 (observed 2.000000), and condition(V)*epsilon is approximately constant (1.568 to 1.558). At exact collision, Disc(q)=0 and the repeated eigenvalue has algebraic multiplicity 2 but eigenspace nullity 1, confirming defectiveness. In optimization, the unconstrained positive model reached task loss 1.23e-32 but gap approximately 3.7e-17 and condition number 2.66e16; the g_min=0.03 gated model reached gap 0.0414, condition 24.4, and task loss 0.00772. Thus the proposed mechanism manifested clearly, though this is not evidence of a downstream accuracy win.

Agent confidence: 9/10

Baseline: Positive unconstrained: task loss 1.23e-32, normalized gap 3.70e-17, cond(V)=2.66e16.
Idea: Gap-gated (g_min=0.03, beta=5000): task loss 7.72e-3, normalized gap 4.14e-2, cond(V)=24.37. Sweep final gaps: g_min=1e-4 -> 1.64e-3; 1e-3 -> 4.49e-2; 1e-2 -> 2.49e-2; 3e-2 -> 4.14e-2. Math slopes: gap 1.000000, discriminant 2.000000.

Limitations: Only a 3-node directed path and a synthetic quadratic task were tested; no directed GCN, stochastic minibatches, Cora, long-horizon dynamics, FLOP/speed comparison, condition-number penalty, characteristic-polynomial numerical stability study, or multi-seed statistical evaluation was performed.

How to run: python3 experiment.py

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

No significant difference vs. the tuned baseline (+13.6%, p=0.3076, wins 3 / 8 paired seeds).

Benchmark: Dynamics forecasting
Model: directed_rnn_laplacian
Paired seeds: 8
Baseline mean: 0.0081
Idea mean: 0.0092
Effect (Δ): 0.0011 (+13.6%; negative = idea better)
Wins: 3 / 8 paired seeds
p-value: 0.3076 (permutation test, 20 000 shuffles)
Smallest detectable effect: ±29.0%
Mechanism: Not confirmed ✗
Practical effect: Inconclusive
Baseline tuning: swept over 3 configs

Limitations:

Only the built-in dynamics track was tested; no vision, sequence, tabular, Cora, long-horizon rollout stability, condition-number penalty, or corrected complex-spectrum gap implementation was tested. The current signature computation filters eigenvalues by positive real part and therefore cannot validate the intended normalized nonzero gap for directed matrices with complex eigenvalues.

How to run: python3 bench_experiment.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

Built a matched directed recurrent dynamics model with positive softplus edge gates and a normalized Laplacian-gap penalty, plus a numerical collision/discriminant sanity check. The eight-seed benchmark found baseline MSE 0.00808 versus idea MSE 0.00918, paired delta +0.00110 and permutation p=0.30765; therefore there was no significant task-metric improvement. The trained-model mechanism signature was not confirmed because the implementation incorrectly discarded complex/nonzero eigenvalues, producing gap=0 for all models.

Baseline: dynamics directed_rnn_laplacian, best lr=0.006: mean test MSE 0.0080788132, std 0.0022047114
Idea: gmin=0.01, beta=20, best lr=0.006: mean test MSE 0.0091805133, std 0.0043079289
Math check: gap log slope 1.0, discriminant log slope 2.0, exact collision defective=true
Paired delta: idea - baseline = +0.0011017001; idea wins 3/8; permutation p=0.30765

Benchmark result