# Saturation-Persistence Trust Region

- ID: 1806
- Canonical URL: https://synthcore.org/idea/1806/saturation-persistence-trust-region
- API JSON: https://synthcore.org/api/idea/1806.json
- API Markdown: https://synthcore.org/api/idea/1806.md
- Verification status: failed_benchmark
- Source: [arXiv:2608.04740](https://arxiv.org/abs/2608.04740)
- Category: training
- Solves: stability, generalization
- ML areas: optimizer, scheduler, training-dynamics
- Math tags: control-theory, dynamical-systems, bifurcations, optimization
- Ratings: usefulness 7/10; difficulty 6/10; novelty 8/10

## Idea description

Use the distinction between persistent saturated equilibria and immediate equilibrium loss to adapt the clipping threshold or learning rate. Increase the allowable update only when saturation is locally persistent and attracting; reduce it when saturation produces a nonpositive branch slope, a shrinking stability margin, or a sharp increase in clipped residual variance.

## Mathematical statement

For a parameter update \(\dot\theta=-\eta\tilde g(\theta)\), define the saturation residual \(e(\theta)=\|g(\theta)\|_2-C\). The clipped region is \(e>0\). Let \(\theta_C^*(\eta,C)\) denote a locally averaged stationary state under clipping, obtained from \(\mathbb E[\Delta\theta]=0\). Estimate the persistence score \(s_\eta=\partial e(\theta_C^*)/\partial\eta\) or \(s_C=\partial e(\theta_C^*)/\partial C\) using paired perturbations. A positive slope with a positive margin \(m=e(\theta_C^*)\) predicts that the saturated branch persists. A nonpositive slope predicts an impending boundary fold, so the controller must reduce the learning rate or increase the clipping margin. The criterion is applied as an empirical local continuation test around the current training state.

## Key formulas

- $$e(\theta;C)=\|\nabla_\theta L(\theta)\|_2-C,\qquad \theta_{t+1}=\theta_t-\eta\,\tilde g(\theta_t;C).$$
- $$\hat s_\eta=\frac{e(\bar\theta_{+};C)-e(\bar\theta_{-};C)}{2\delta_\eta},\qquad \hat s_C=\frac{e(\bar\theta;C+\delta_C)-e(\bar\theta;C-\delta_C)}{2\delta_C},$$
- $$\eta_{k+1}=\begin{cases}\min\{\eta_k(1+\alpha),\eta_{\max}\},&m_k>m_{\min}\ \wedge\ \hat s_\eta>0,\\\eta_k(1-\beta),&\hat s_\eta\le0\ \vee\ \operatorname{Var}(L)_{k-W:k}>V_{\max},\\\eta_k,&\text{otherwise}.\end{cases}$$

## Implementation notes

Add a slow supervisory controller around an existing optimizer; do not modify backpropagation. Maintain a primary run at threshold \(C_k\), and create two cheap shadow probes every K optimizer steps. The probes reuse the same minibatch sequence and optimizer state but apply \(\eta_k+\delta_\eta\) and \(\eta_k-\delta_\eta\) for H short steps, with H between 5 and 20. Compute the averaged endpoint \(\bar\theta_+\) and \(\bar\theta_-\), their unclipped gradient norms, loss variance, fraction of clipped steps, and update-direction cosine. Estimate the branch slope \(\hat s_\eta\) from the finite difference in \(e=\|g\|-C\). If the primary run is clipped and \(e\) has a positive margin while \(\hat s_\eta>0\), increase the learning rate by \(1+\alpha\), for example \(\alpha=0.02\). If \(\hat s_\eta\le0\), loss variance rises, or update cosine falls below zero, reduce the learning rate by \(1-\beta\), for example \(\beta=0.1\). Optionally increase \(C\) only when the issue is persistent clipping without instability. The exact equilibrium continuation and slope criterion come from the paper; branch states, slopes, and thresholds are estimated empirically. First experiment: compare this controller with fixed clipping and cosine-decay SGD on CIFAR-10 using a small ResNet-18, restricting probes to 10 steps every 200 training steps. Sweep initial learning rates around the fixed-clipping divergence boundary. The predicted signature is a sign change in \(\hat s_\eta\) near the fixed-run instability boundary, with the controller reducing catastrophic loss spikes by at least 50 percent while retaining a learning rate within 20 percent of the largest stable fixed rate.

## Verification

- Status: failed_benchmark
- Mechanism evidence: yes
- Mechanism confirmed: no
- Practical verdict: no_effect
- Verdict: Built a readable NumPy MVP with the paper-derived saturation branch and a supervisory clipped-SGD controller using paired shadow probes. The mechanism manifested: the predicted branch-sign transition was P*=0.600 versus observed 0.600, persistence classification matched across the sweep, and the collapse-voltage formula matched numerical inversion to 1.1e-16 with a mean boundary-to-collapse gap of 0.403 on persistent branches. The optimizer controller did not reduce peak loss in this toy setup and generally worsened final loss at moderate/high learning rates, so there is no evidence of a practical training win yet.

### Mechanism check

- Verdict: Built a readable NumPy MVP with the paper-derived saturation branch and a supervisory clipped-SGD controller using paired shadow probes. The mechanism manifested: the predicted branch-sign transition was P*=0.600 versus observed 0.600, persistence classification matched across the sweep, and the collapse-voltage formula matched numerical inversion to 1.1e-16 with a mean boundary-to-collapse gap of 0.403 on persistent branches. The optimizer controller did not reduce peak loss in this toy setup and generally worsened final loss at moderate/high learning rates, so there is no evidence of a practical training win yet.
- Confidence: 8/10
- Limitations: Only the required toy mechanism and a two-dimensional noisy quadratic were tested; no CIFAR-10, ResNet, CUDA run, wall-clock/FLOP accounting, multiple random seeds, or comparison with cosine-decay SGD was performed. The exact paper-equation sweep is an analytical reduced model and does not establish that the empirical optimizer controller estimates branch persistence reliably in neural-network training.

### Practical benchmark

- Paired seeds: 8
- Baseline mean: 0
- Idea mean: 0
- p-value: 0.61655
- Paired wins: 4/8
- Benchmark verdict: no significant win

## Artifacts

- [bench_experiment.py](https://synthcore.org/code/675/bench_experiment.py)
- [bench_report.json](https://synthcore.org/code/675/bench_report.json)
- [experiment.py](https://synthcore.org/code/675/experiment.py)
- [report.md](https://synthcore.org/code/675/report.md)
- [report_bench_2026-08-31T213342.md](https://synthcore.org/code/675/report_bench_2026-08-31T213342.md)
- [results.json](https://synthcore.org/code/675/results.json)
- [Download all files as ZIP](https://synthcore.org/download/675)

## Disclaimer

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