# ISS-Certified Sampled Optimizer Wrapper

- ID: 1113
- Canonical URL: https://synthcore.org/idea/1113/iss-certified-sampled-optimizer-wrapper
- API JSON: https://synthcore.org/api/idea/1113.json
- API Markdown: https://synthcore.org/api/idea/1113.md
- Verification status: failed_benchmark
- Source: [arXiv:2607.18500](https://arxiv.org/abs/2607.18500)
- Category: dynamics
- Solves: stability, accuracy
- ML areas: optimizer, rnn, ssm, training-dynamics
- Math tags: dynamical-systems, control-theory, optimization, numerical-analysis
- Ratings: usefulness 7/10; difficulty 6/10; novelty 7/10

## Idea description

Wrap a recurrent or state-space neural network in a sampled-data feedback loop: latent states evolve continuously or at every fine solver step, while a constrained optimizer updates the control, adapter, or residual-gating vector only every M steps. Between optimizer updates, use zero-order hold or linear interpolation and reject updates that violate a learned Lyapunov decrease condition. This should prevent large transient latent explosions caused by aggressive optimizer updates while retaining larger update steps than a globally contractive architecture.

## Mathematical statement

Let x_k be the neural latent state at optimizer sample k, u_k the constrained control or residual-gating vector, d_k an external perturbation, and T the sampling interval. The sampled closed loop is x_{k+1}=F_T(x_k,u_k,d_k), with optimizer update u_{k+1}=Pi_U[u_k-eta g_k], where Pi_U projects onto feasible set U and g_k is a gradient or policy direction. Between samples, the applied control is u(t)=(1-tau)u_k+tau u_{k+1}, with tau=(t-kT)/T, or u(t)=u_k under zero-order hold. The transferable stability condition is an input-to-state inequality: V(x_{k+1})-V(x_k)<=-alpha(||x_k||)+gamma(||d_k||), where V is positive definite, alpha is a positive dissipation function, and gamma measures disturbance amplification. For V(x)=x^T P x with P positive definite, use the empirical certificate Delta V_k+lambda V(x_k)-c||d_k||^2<=0, where lambda>0 and c>=0. Accept an update only when this inequality holds; otherwise reduce eta or shrink the trust region. The predicted ultimate bound is limsup V(x_k)<=(c/lambda)sup_k||d_k||^2.

## Key formulas

- $$x_{k+1}=F_T(x_k,u_k,d_k),\qquad u_{k+1}=\Pi_{\mathcal U}\left[u_k-\eta g_k\right]$$
- $$u(t)=(1-\tau)u_k+\tau u_{k+1},\qquad \tau=\frac{t-kT}{T},\quad kT\leq t<(k+1)T$$
- $$V(x_{k+1})-V(x_k)\leq-\alpha(\|x_k\|)+\gamma(\|d_k\|)$$
- $$\Delta V_k+\lambda V(x_k)-c\|d_k\|^2\leq0,\qquad V(x)=x^\top P x,\quad P\succ0$$

## Implementation notes

1. Integration point: use a GRU, residual RNN, or neural state-space model whose latent state x is updated at every fine step. Add a low-dimensional vector u controlling residual strength, damping, normalization gain, or an input-feedback correction. Every M fine steps, compute a gradient direction g_k from the task loss, a prediction-error objective, or a one-step MPC surrogate; constrain u to a box or norm ball. 2. Pseudocode: initialize P as diagonal positive, u in U, and sampling period T=M*dt. At sample k, save x_k and V_k=x_k^T P x_k; propose u_prop=project_U(u-eta*g). Roll the network forward for M steps using linear interpolation or hold between u and u_prop. Compute x_next, DeltaV=x_next^T P x_next-V_k, and a disturbance estimate d_k from input residuals or teacher-forcing error. Accept u_prop if DeltaV+lambda*V_k-c*||d_k||^2<=0; otherwise halve eta and retry, up to a fixed number of attempts. 3. Computed from the paper's mechanism: sampled update, constrained projection, interpolation, Lyapunov decrease, and ISS residual. Estimated empirically: P, lambda, c, and the worst-case local dissipation margin. Fit P by minimizing certificate violations over rollout data, then validate it on held-out initial states. 4. First cheap experiment: compare a 2-layer tanh RNN on sequential MNIST or a synthetic nonlinear oscillator-prediction task against the same RNN with Adam and gradient clipping. Sweep eta and M with bounded impulse disturbances. The quantitative prediction is a sharp degradation boundary when the accepted dissipation margin crosses zero; the maximum stable eta should decrease as M increases, and the empirical ultimate bound should scale approximately linearly with disturbance energy, with V plateau near (c/lambda)*sup||d_k||^2.

## Verification

- Status: failed_benchmark
- Mechanism evidence: yes
- Mechanism confirmed: no
- Practical verdict: no_effect
- Verdict: Built an ISS-certified sampled optimizer wrapper with box projection, linear control interpolation, Lyapunov-certificate rejection/backtracking, and safe fallback control. All three toy predictions were confirmed: the fixed-control stability boundary matched |a-u|<1 within grid resolution, the certificate sampling transition occurred at M=2 as predicted, and disturbance plateau energy scaled exactly as d^2. The wrapper reduced the baseline's state explosion from 2.22e22 maximum magnitude to a maximum of 1.0 and final magnitude 2.09e-3, demonstrating a real stabilization effect in this toy setting.

### Mechanism check

- Verdict: Built an ISS-certified sampled optimizer wrapper with box projection, linear control interpolation, Lyapunov-certificate rejection/backtracking, and safe fallback control. All three toy predictions were confirmed: the fixed-control stability boundary matched |a-u|<1 within grid resolution, the certificate sampling transition occurred at M=2 as predicted, and disturbance plateau energy scaled exactly as d^2. The wrapper reduced the baseline's state explosion from 2.22e22 maximum magnitude to a maximum of 1.0 and final magnitude 2.09e-3, demonstrating a real stabilization effect in this toy setting.
- Confidence: 8/10
- Limitations: Only a scalar linear state-space toy was tested, not a GRU/SSM trained on a task. The optimizer direction is synthetic, P is fixed to the scalar identity rather than learned, disturbances are constant, and no FLOP/accuracy or larger-scale neural-network comparison was performed.

### Practical benchmark

- Paired seeds: 8
- Baseline mean: 0
- Idea mean: 0
- p-value: 1
- Paired wins: 0/8
- Benchmark verdict: no measurable effect

## Artifacts

- [bench_report.json](https://synthcore.org/code/443/bench_report.json)
- [iss_sampled_optimizer.py](https://synthcore.org/code/443/iss_sampled_optimizer.py)
- [report.md](https://synthcore.org/code/443/report.md)
- [report_bench_2026-08-31T121904.md](https://synthcore.org/code/443/report_bench_2026-08-31T121904.md)
- [results.json](https://synthcore.org/code/443/results.json)
- [stage2_iss_bench.py](https://synthcore.org/code/443/stage2_iss_bench.py)
- [Download all files as ZIP](https://synthcore.org/download/443)

## Disclaimer

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