✗ Failed on benchmark 2026

Wedge-Positive Tangent Dynamics

Implementation & benchmark of arXiv:2607.27176 — Some intuition for why cooperative systems "look 1-dimensional" and 2-cooperative systems "look 2-dimensional"

Usefulness8/10

Difficulty7/10

Novelty8/10

Dynamics Stability Generalization Accuracy Rnn Ssm Regularization Training dynamics Dynamical systems Linear algebra Geometry Control theory

Source paper: Some intuition for why cooperative systems "look 1-dimensional" and 2-cooperative systems "look 2-dimensional" arXiv:2607.27176 ⓘ · analyzed Aug 31, 2026

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 Not peer-reviewed.

Idea description

Constrain the Jacobian of a neural ODE or recurrent transition so that its second additive compound is Metzler and irreducible, then regularize the resulting finite-window wedge transition toward strict positivity. This should contract projective distances between admissible tangent 2-planes, causing perturbation planes to align and making long-horizon representations effectively two-dimensional rather than allowing uncontrolled orientation growth.

Formulas

wedge2 Dphi_(t+T,t)(v wedge w) = Dphi_(t+T,t)v wedge Dphi_(t+T,t)w

A^[2] = derivative at h=0 of wedge2 exp(hA), and wedge2 exp(hA) = I + h A^[2] + O(h^2)

d_H(Mu,Mv) <= q_T d_H(u,v), with q_T = tanh(Delta(M)/4) < 1 and Delta(M) = sup_(u,v>0) d_H(Mu,Mv)

L_wedge = lambda_Metzler sum_(i != j) max(0,-A^[2]_ij)^2 + lambda_align max(0,q_hat_T - q_*)^2

Mathematical statement

Let the neural dynamics be dot x = f_theta(x,t), with Jacobian A(t) = D_x f_theta(x(t),t). The second additive compound A^[2] is defined by wedge2(exp(hA)) = I + h A^[2] + O(h^2), and its diagonal entries are A^[2]_(i,j),(i,j) = a_ii + a_jj. Require A^[2](t) to be Metzler, meaning all off-diagonal entries are nonnegative after an optional fixed sign gauge, with a fixed strongly connected support graph and nonzero entries bounded below by alpha > 0. Then the finite-window wedge transition M = wedge2 D phi_(t+T,t) becomes entrywise positive. For positive vectors u and v, Hilbert distance is d_H(u,v) = log((max_i u_i/v_i)/(min_i u_i/v_i)). Birkhoff contraction gives d_H(Mu,Mv) <= q_T d_H(u,v), where q_T = tanh(Delta(M)/4) < 1 and Delta(M) = sup_(u,v>0) d_H(Mu,Mv). The neural regularizer penalizes negative compound off-diagonal entries and empirical projective ratios above a target q_* < 1.

Implementation notes

1. Integration point: begin with a continuous-time neural ODE dot x = f_theta(x,t), then test the same construction in a residual RNN update x_(k+1) = x_k + h f_theta(x_k). Use state dimension n = 8 or 16, making the compound dimension m = n choose 2 computationally manageable. At randomly selected training states, obtain A = D_x f_theta with automatic differentiation, construct A^[2], and add the wedge loss to the prediction loss. 2. Pseudocode: sample states; compute A; construct its additive compound; penalize every negative off-diagonal compound entry; integrate dot V = A(t)V for two pairs of tangent vectors; form positive wedge-coordinate vectors z1 = vec(v1 wedge w1) and z2 = vec(v2 wedge w2); estimate q_hat_T = d_H(z1(T),z2(T)) / d_H(z1(0),z2(0)); optimize task loss plus the wedge penalty. 3. Jacobians, compounds, wedge vectors, and Hilbert distances are computed directly. The exact Birkhoff diameter is expensive, so estimate it using 16 to 64 random positive tangent planes and retain the largest observed contraction ratio. The paper supplies the cone mechanism; alpha, window T, and target q_* are empirical hyperparameters. 4. First experiment: train a neural ODE on Lorenz-63 one-step prediction and a damped-pendulum trajectory, comparing no regularizer, spectral-norm regularization, and wedge positivity. Measure rollout error and tangent-plane alignment. The falsifiable prediction is that once compound off-diagonal violations disappear and the finite-window wedge matrix becomes positive, q_hat_T < 1 and distances decay geometrically as d_H(kT) approximately equals q_hat_T^k d_H(0). Test whether the observed decay rate matches log(q_hat_T)/T within 20 percent; models violating the cone condition should not exhibit uniform decay.

Verification

Failed on benchmark

Stage 1 · Toy mechanism gate: Passed ✓

Stage 2 · Mechanism transferred to benchmark: Not reproduced ✗

Stage 2 · Practical benchmark result: No measurable effect

Methodology: 0 paired seeds, permutation test (p<0.05) against a learning-rate-tuned baseline. How verification works

Stage 1 — Mechanism check agent confidence 8/10

Built and ran a NumPy/SciPy MVP for second additive compounds, wedge transitions, positivity, and Hilbert projective contraction. The mechanism manifested: the compound identity had observed O(h^2) error slope 1.992, the six-cycle worst-entry scaling had observed alpha^3 slope 3.000 as predicted, and stronger irreducible coupling reduced the Birkhoff bound from 0.99996 at alpha=0.05 to 0.87525 at alpha=0.8. This validates the toy mathematical mechanism, but does not establish a neural-network accuracy win.

Agent confidence: 8/10

Baseline: Signed/non-Metzler cycle: minimum finite-window transition entry -0.1463; among positive-output samples, maximum Hilbert ratio 2.492, so uniform positive-cone contraction was absent.
Idea: Metzler irreducible cycle: all finite-window entries were positive; sampled maximum Hilbert ratios decreased from 0.9904 to 0.6799 as alpha increased from 0.05 to 0.8, while exact Birkhoff bounds decreased from 0.99996 to 0.87525. Compound identity slope was 1.992 versus predicted 2.0; path scaling slope was 3.000 versus predicted 3.0.

Limitations: This is a controlled linear compound-dynamics toy experiment, not a trained neural ODE or residual RNN. It does not test prediction loss, rollout accuracy, FLOPs, learned regularization, time-varying generators, sign-gauge learning, or downstream task performance. Hilbert ratios are random-sample estimates and the Birkhoff bound can be loose.

How to run: python3 wedge_experiment.py

Stage 2 — Benchmark vs. tuned baseline (arXiv:2607.27176) Sep 2, 2026

No significant difference vs. the tuned baseline (0 metric points, p=1, wins 0 / 0 paired seeds).

Benchmark: Dynamics forecasting
Model: rnn_small
Paired seeds: 0
Baseline mean: 0
Idea mean: 0
Effect (Δ): 0 (n/a; negative = idea better)
Wins: 0 / 0 paired seeds
p-value: 1 (permutation test, 20 000 shuffles)
Smallest detectable effect: n/a
Mechanism: Not confirmed ✗
Practical effect: No effect
Baseline tuning: swept over 0 configs

Limitations:

The full 8-seed baseline sweep, three-setting idea sweep, paired permutation test, and trained-model mechanism signature were not completed. No bench_report.json was generated.

How to run: python3 wedge_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: 10/10

Implemented Wedge-Positive Tangent Dynamics for the dynamics/rnn_small track, including the additive-compound penalty and numerical compound identity check. The required benchmark did not complete within the 40-minute budget, even after reducing the dataset and epochs, so there is no valid baseline-versus-idea metric or evidence of a win.

Baseline: No valid benchmark result; baseline sweep timed out.
Idea: No valid benchmark result; wedge-regularized training timed out.

Benchmark result