# Entry-Exit Curvature Scheduler

- ID: 1798
- Canonical URL: https://synthcore.org/idea/1798/entry-exit-curvature-scheduler
- API JSON: https://synthcore.org/api/idea/1798.json
- API Markdown: https://synthcore.org/api/idea/1798.md
- Verification status: failed_benchmark
- Source: [arXiv:2608.04594](https://arxiv.org/abs/2608.04594)
- Category: dynamics
- Solves: stability, speedup
- ML areas: optimizer, scheduler, training-dynamics
- Math tags: dynamical-systems, control-theory, bifurcations, numerical-analysis
- Ratings: usefulness 7/10; difficulty 5/10; novelty 7/10

## Idea description

Replace pointwise curvature-based learning-rate decisions with a slow-fast entry-exit scheduler. The optimizer maintains a slowly varying state representing effective curvature or gradient-noise level, accumulates the weak transverse growth rate along that slow trajectory, and changes learning regime only when the accumulated rate returns to zero. This permits controlled passage through locally unstable or poorly conditioned regions while preventing indefinite residence in a regime with net positive perturbation growth.

## Mathematical statement

In the paper's traveling-wave reduction, the weak eigenvalue transverse to the slow manifold is \(\lambda_-(S)=c-\sqrt{c^2-4D\mathcal{R}_0S+4D}\), where \(c\) is front speed, \(D\) is infected diffusion, \(\mathcal{R}_0\) is the reproduction parameter, and \(S\) is the slow variable. Along the slow flow, \(S(\zeta)=1+(S_\infty-1)e^{-\zeta/c}\). Defining \(a=c^2-4D(\mathcal{R}_0-1)\) and \(B=4D\mathcal{R}_0(S_\infty-1)<0\), the eigenvalue becomes \(\lambda_-(\zeta)=c-\sqrt{a-Be^{-\zeta/c}}\). The entry-exit time \(T_E\) is the nontrivial positive solution of \(\int_0^{T_E}\lambda_-(\zeta)d\zeta=0\). The neural analogue uses a slow state \(s_t\), an estimated transverse growth rate \(\hat\lambda_t\), and an accumulated budget \(G_{t+1}=G_t+\hat\lambda_t\Delta t\). The optimizer exits its exploratory regime when \(G_t\) crosses zero from below after previously becoming negative. The transfer assumes that local optimizer perturbations admit a fast linear approximation whose dominant transverse exponent varies slowly through \(s_t\).

## Key formulas

- $$\lambda_-(S)=c-\sqrt{c^2-4D\mathcal{R}_0S+4D}.$$
- $$S(\zeta)=1+(S_\infty-1)e^{-\zeta/c},\qquad \lambda_-(\zeta)=c-\sqrt{a-Be^{-\zeta/c}},$$
- $$\int_0^{T_E}\lambda_-(\zeta)\,d\zeta=0,\qquad a=c^2-4D(\mathcal{R}_0-1),\quad B=4D\mathcal{R}_0(S_\infty-1)<0.$$
- $$G_{t+1}=G_t+\hat\lambda_t\Delta t,\qquad \eta_t=\begin{cases}\eta_{\mathrm{explore}},&G_t<0\ \text{and no exit has occurred},\\\eta_{\mathrm{exploit}},&G_t\geq 0\ \text{after the first crossing}.\end{cases}$$

## Implementation notes

Integrate this as a wrapper around SGD or AdamW. Maintain a scalar slow state \(s_t\), an accumulated budget \(G_t\), and two learning rates \(\eta_{\mathrm{explore}}>\eta_{\mathrm{exploit}}\). Every \(K\) optimizer steps, estimate a dominant curvature direction with one or two Hessian-vector products. Draw a normalized vector \(v_t\), compute \(H_tv_t\) by automatic differentiation, set \(q_t=v_t^TH_tv_t\), update \(v_t\leftarrow H_tv_t/\|H_tv_t\|\), and smooth curvature using \(s_{t+1}=(1-\epsilon)s_t+\epsilon q_t\), with small \(\epsilon\). Define \(\hat\lambda_t=\alpha(s_t-\mu)\), where \(\mu\) is a target stability curvature and \(\alpha\) is calibrated from short probe trajectories. Update \(G\leftarrow G+K\hat\lambda_t\). Use the exploratory learning rate while \(G<0\); after the first crossing from negative to nonnegative, permanently switch to the exploitative rate. The paper contributes the integral-crossing rule; curvature trajectories, calibration constants, and Hessian directions are estimated empirically. First test a two-layer MLP on MNIST and a small ResNet on CIFAR-10 against SGD, AdamW, cosine decay, and a pointwise curvature-sign scheduler with matched average learning-rate budget. Log \(s_t\), \(\hat\lambda_t\), \(G_t\), loss, gradient norms, and nearby-trajectory separation. The prediction is that switching occurs at the first discrete root of \(\sum_{j=1}^{m}K\hat\lambda_j=0\), after the first positive local \(\hat\lambda_t\) whenever the later integrated rate is negative. Predicted switch times should match observed perturbation-contraction onset within 20 percent; pointwise sign switching should occur earlier.

## Verification

- Status: failed_benchmark
- Mechanism evidence: yes
- Mechanism confirmed: no
- Practical verdict: inconclusive
- Verdict: Built an entry-exit curvature scheduler with slow curvature state, accumulated growth budget, permanent crossing-triggered regime switching, and a pointwise-sign comparator. The paper reduction confirmed three mechanism predictions: integral exit occurs after the pointwise zero (T_E/lambda-zero = 2.01–2.28), exit time decreases monotonically with S_inf, and T_E scales approximately linearly with c (T_E/c relative spread 1.6%); root residuals were below 5e-15. In the toy optimizer, entry-exit switched at step 145 versus pointwise step 50 and achieved final loss 7.86e-5 versus 1.47e-4 for pointwise, but it did not beat constant high-rate SGD (1.79e-9), so the demonstrated gain is specifically delayed integrated switching rather than general optimization superiority.

### Mechanism check

- Verdict: Built an entry-exit curvature scheduler with slow curvature state, accumulated growth budget, permanent crossing-triggered regime switching, and a pointwise-sign comparator. The paper reduction confirmed three mechanism predictions: integral exit occurs after the pointwise zero (T_E/lambda-zero = 2.01–2.28), exit time decreases monotonically with S_inf, and T_E scales approximately linearly with c (T_E/c relative spread 1.6%); root residuals were below 5e-15. In the toy optimizer, entry-exit switched at step 145 versus pointwise step 50 and achieved final loss 7.86e-5 versus 1.47e-4 for pointwise, but it did not beat constant high-rate SGD (1.79e-9), so the demonstrated gain is specifically delayed integrated switching rather than general optimization superiority.
- Confidence: 8/10
- Limitations: Only a deterministic nonstationary 2D quadratic optimizer toy was tested; no MNIST, CIFAR-10, AdamW, neural-network HVP implementation, FLOP matching, or nearby-trajectory separation experiment was run. The curvature signal in the optimizer demo is supplied analytically rather than estimated from automatic-differentiation HVPs.

### Practical benchmark

- Paired seeds: 8
- Baseline mean: 0
- Idea mean: 0
- p-value: 0.7926
- Paired wins: 4/8
- Benchmark verdict: no significant win

## Artifacts

- [bench_entry_exit.py](https://synthcore.org/code/671/bench_entry_exit.py)
- [bench_report.json](https://synthcore.org/code/671/bench_report.json)
- [entry_exit_experiment.py](https://synthcore.org/code/671/entry_exit_experiment.py)
- [report.md](https://synthcore.org/code/671/report.md)
- [report_bench_2026-08-31T213617.md](https://synthcore.org/code/671/report_bench_2026-08-31T213617.md)
- [results.json](https://synthcore.org/code/671/results.json)
- [Download all files as ZIP](https://synthcore.org/download/671)

## Disclaimer

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