△ Mechanism confirmed, baseline not beaten 2026

Self-Correcting Euler Horizon Rule

Implementation & benchmark of arXiv:2607.14291 — Wasserstein Stability of Contracting Flows: Effective Rates, Euler Self-Correction, and Noise Tightening

Usefulness7/10

Difficulty4/10

Novelty6/10

Dynamics Stability Speedup Accuracy Diffusion sampling Inference speedup Ssm World model Dynamical systems Numerical analysis Control theory Spectral theory

Source paper: Wasserstein Stability of Contracting Flows: Effective Rates, Euler Self-Correction, and Noise Tightening arXiv:2607.14291 ⓘ · analyzed Aug 30, 2026

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 Not peer-reviewed.

Idea description

Use contraction-aware integration rather than assuming that Euler discretization error grows monotonically with sampling time. For a contracting neural ODE, permit a transient error peak but choose the step size and terminal horizon using the predicted peak time and subsequent exponential decay.

Formulas

$$x_{k+1}=x_k+h v(x_k),\qquad \dot{x}=v(x),$$

$$\dot{x}=-\lambda x:\quad x(t)=e^{-\lambda t}x_0,\qquad x_k=(1-\lambda h)^kx_0,$$

$$0<h<\frac{2}{\lambda}\quad\Longleftrightarrow\quad |1-\lambda h|<1,$$

$$e(t)\ \text{is predicted to grow initially, peak at }t_{\mathrm{peak}}=O(\lambda^{-1}),\text{ then decay as }e(t)\propto e^{-\lambda(t-t_{\mathrm{peak}})}.$

Mathematical statement

Consider the exact contracting flow $\dot{x}=v(x)$ and Euler iterates $x_{k+1}=x_k+h v(x_k)$, where $h$ is the step size and $\lambda>0$ is a contraction rate. Local truncation errors injected at early steps are propagated by the contraction, so the global error has a forcing-versus-decay profile rather than monotone accumulation. In the linear surrogate $\dot{x}=-\lambda x$, the exact multiplier is $e^{-\lambda h}$ and Euler's multiplier is $1-\lambda h$, requiring $0<h<2/\lambda$ for discrete stability. For nonlinear contracting fields, the paper predicts a universal transient peak whose time is controlled by $\lambda$, followed by exponential decay; this creates a testable criterion for safe coarse integration and for distinguishing numerical error from genuine model instability.

Implementation notes

Add a contraction-aware solver controller to neural ODE, continuous normalizing flow, or diffusion-ODE sampling. Estimate a conservative local contraction rate $\hat\lambda_t$ from minibatch secant pairs or from the largest eigenvalue of the symmetric Jacobian part, $\lambda_t=-\lambda_{\max}((J_v+J_v^\top)/2)$. Before each Euler or Heun step, enforce $h_t\leq\alpha/\max(\hat\lambda_t,\epsilon)$ with $\alpha<2$, and log the discrepancy between one full step and two half steps as an error proxy. Do not reject a trajectory merely because this proxy increases: maintain a running error curve, identify its peak near $t_{\mathrm{peak}}\approx c/\hat\lambda$, and permit larger later steps when the proxy begins decaying at the predicted rate. Pseudocode: estimate $\hat\lambda_t$; clip $h_t$; compute full and half-step solutions; record $e_t$; update the peak detector; after the peak, increase $h_t$ only while $e_{t+\Delta}/e_t$ is below $\exp(-\hat\lambda_t\Delta)$ up to tolerance. The paper supplies the self-correction mechanism and contraction dependence; Jacobian rates and constants must be estimated empirically for the neural field. First cheap experiment: use a 2D contracting MLP neural ODE and a CIFAR-10 latent diffusion ODE with fixed Euler versus this controller. Plot error against a high-accuracy RK reference. The prediction is a non-monotone error curve with a peak at a time proportional to $1/\lambda$, followed by a log-linear decay slope near $-\lambda$; violating $h<2/\lambda$ should remove this decay and produce persistent or divergent error.

Verification

Mechanism confirmed, baseline not beaten

Stage 1 · Toy mechanism gate: Passed ✓

Stage 2 · Mechanism transferred to benchmark: Confirmed ✓

Stage 2 · Practical benchmark result: No measurable effect

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 scalar contracting-ODE MVP with Euler error sweeps, stability-boundary verification, peak-time scaling, decay-rate fitting, and a contraction-aware controller. The three mechanism predictions manifested: the stability transition was observed at lambda*h=2.00, peak time scaled as 1/lambda with 9.7% deviation from the analytic estimate, and post-peak log-error slopes matched -lambda within 0.91%. The controller used 33 versus 40 steps but had worse terminal error (3.26e-4 versus 2.03e-4), so this is evidence for the mechanism, not an end-to-end accuracy or speed win.

Agent confidence: 9/10

Baseline: Fixed Euler, lambda=1, T=8, h=0.2: 40 steps, terminal absolute error 2.0254e-4.
Idea: Contraction-capped two-half-step controller, alpha=1.8: 33 steps, terminal absolute error 3.2625e-4; stability boundary observed z=2.00, peak lambda*t=0.9278 versus predicted 0.8456, decay slope relative error 0.57%-0.91%.

Limitations: Only the scalar linear contracting surrogate was tested; no nonlinear MLP neural ODE, CIFAR-10 latent diffusion, Jacobian/secant-rate estimation, wall-clock benchmark, adaptive horizon extension, or GPU experiment was performed.

How to run: python3 experiment.py

Stage 2 — Benchmark vs. tuned baseline (arXiv:2607.14291) Sep 2, 2026

No significant difference vs. the tuned baseline (−-0.0%, p=1, wins 0 / 8 paired seeds).

Benchmark: Dynamics forecasting
Model: rnn_small
Paired seeds: 8
Baseline mean: 0
Idea mean: 0
Effect (Δ): 0 (−-0.0%; negative = idea better)
Wins: 0 / 8 paired seeds
p-value: 1 (permutation test, 20 000 shuffles)
Smallest detectable effect: ±0.0%
Mechanism: Confirmed ✓
Practical effect: No effect
Baseline tuning: swept over 3 configs

Limitations:

Only the built-in damped controlled-pendulum dynamics track was tested. CIFAR-10 latent diffusion and nonlinear neural-ODE inference were not tested. The local full-vs-two-half perturbation used a constant secant field, making the proxy effectively inactive; genuine time-varying neural vector-field integration and wall-clock speedup remain untested.

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 Aug 31, 2026 latest Failed ✗

Agent confidence: 9/10

Implemented the contraction-aware Euler consistency intervention on the matched dynamics track using the shared rnn_small architecture, with 8 paired seeds and parity-preserving baseline/idea sweeps. Baseline and best idea both achieved mean test MSE 1.3684802866009704e-05; paired delta was 0.0 with permutation p=1.0, so there was no benchmark win. The Euler stability sanity check passed and the trained-model signature observed mean lambda*h=0.2277172260841973 below alpha=1.8, but this did not improve task performance.

Baseline: Best lr=0.006; 8-seed test MSE mean=1.3684802866009704e-05, std=2.4297331089709535e-06.
Idea: Best lr=0.006, alpha=1.8, penalty=0.1; 8-seed test MSE mean=1.3684802866009704e-05, std=2.4297331089709535e-06.
Math sanity: Predicted Euler stability boundary lambda*h=2.0; observed grid boundary=1.9999999999999998.
Mechanism signature: Observed mean lambda*h=0.2277172260841973 versus predicted cap alpha=1.8; confirmed=true, but no task-metric gain.
Paired comparison: delta_mean=0.0, idea_wins=0/8, p_value=1.0, verdict=no measurable effect.

Benchmark result