# IQC-Certified Training Dynamics

- ID: 2888
- Canonical URL: https://synthcore.org/idea/2888/iqc-certified-training-dynamics
- API JSON: https://synthcore.org/api/idea/2888.json
- API Markdown: https://synthcore.org/api/idea/2888.md
- Verification status: failed_benchmark
- Source: [arXiv:2608.30431](https://arxiv.org/abs/2608.30431)
- Category: dynamics
- Solves: stability, generalization
- ML areas: optimizer, training-dynamics, regularization, rnn
- Math tags: control-theory, dynamical-systems, optimization, linear-algebra, convex-analysis
- Ratings: usefulness 8/10; difficulty 6/10; novelty 8/10

## Idea description

Represent a learned optimizer or recurrent training controller as a discrete-time feedback system and certify its sensitivity to one-sample dataset replacement using an IQC dissipativity inequality. Penalize the smallest certified disturbance-to-state gain during meta-training or use it as a post-training acceptance test, favoring update dynamics that do not amplify microscopic data perturbations over many iterations.

## Mathematical statement

Let x_k in R^n be the optimizer or recurrent-module state, y_k = C x_k + D u_k its feedback signal, and u_k = Phi(y_k,S) the data-dependent neural update. For neighboring datasets S and S', decompose the operator difference as u_k - u'_k = w_k + e_k, where w_k is the incremental response to the same feedback signal and e_k is the one-sample replacement disturbance. The sensitivity dynamics are delta_x_(k+1) = A delta_x_k + B w_k + B e_k and delta_y_k = C delta_x_k + D w_k + D e_k. Assume the incremental operator satisfies the IQC q_k^T Q q_k >= 0, where q_k = [delta_y_k; w_k], Q is a symmetric matrix, and delta_y_k = y_k - y'_k. A sufficient dissipativity certificate is an LMI in a storage matrix P, decay factor rho, disturbance gain gamma, and IQC multiplier lambda. It implies V_(k+1) - rho V_k + ||delta_x_k||^2 - gamma^2 ||e_k||^2 + lambda q_k^T Q q_k <= 0, with V_k = delta_x_k^T P delta_x_k. Summation gives a finite-horizon sensitivity bound, so if the one-sample operator sensitivity is at most epsilon_N, cumulative algorithmic stability scales as O(gamma epsilon_N).

## Key formulas

- $$\delta x_{k+1}=A\delta x_k+B w_k+B e_k,\qquad \delta y_k=C\delta x_k+D w_k+D e_k,\qquad q_k=\begin{bmatrix}\delta y_k\\w_k\end{bmatrix},\qquad q_k^T Qq_k\ge 0.$$
- $$T=\begin{bmatrix}A&B&B\end{bmatrix},\qquad H=\begin{bmatrix}C&D&D\\0&I_m&0\end{bmatrix},\qquad P\succ0,\quad \lambda\ge0.$$
- $$V_{k+1}-\rho V_k+\|\delta x_k\|^2-\gamma^2\|e_k\|^2+\lambda q_k^TQq_k\le0,\qquad V_k=\delta x_k^TP\delta x_k.$$
- $$\sum_{k=0}^{K-1}\|\delta x_k\|^2\le\frac{V_0}{1-\rho}+\gamma^2\sum_{k=0}^{K-1}\|e_k\|^2,\qquad 0<\rho<1.$$

## Implementation notes

First, integrate the mechanism into a recurrent optimizer or learned training controller whose state contains parameters, momentum, adaptive statistics, or a compact controller state. A cheap baseline is heavy-ball training with state x_k = [theta_k; theta_(k-1)] and a data-dependent gradient operator Phi; the same construction can later be applied to an RNN or state-space model used for long-horizon prediction. Second, use the following MVP procedure: sample a task and two datasets differing by one example; run the controller on both; estimate e_k = Phi(y_k,S) - Phi(y_k,S') after aligning the feedback state; estimate an incremental slope bound L from pairs of outputs and updates; choose an IQC such as Q = diag(L^2 I_p, -I_m) for the condition ||w_k|| <= L ||delta_y_k||; construct the lifted matrices T and H; solve an SDP for P, lambda, rho, and the smallest gamma; and add alpha log(gamma + 1e-8) or alpha max(0, lambda_max(M))^2 to the controller objective. Initially stop gradients through the SDP and alternate controller updates with certificate updates. The matrices A, B, C, and D come from the optimizer realization; L and one-sample disturbance magnitudes are estimated empirically, while the SDP is the formal certificate conditional on those estimates. Test on a d = 20 finite-sum quadratic regression problem with N = 100, comparing vanilla gradient descent, heavy-ball, Nesterov, and a learned two-state optimizer. Sweep learning rate and momentum, measure leave-one-sample-out sensitivity, loss divergence, and certified gamma. The prediction is that the empirical divergence or sensitivity boundary appears close to the first hyperparameter value where the LMI becomes infeasible, and cumulative sensitivity grows approximately linearly with gamma, with empirical sensitivity within a factor of two of the certificate.

## Verification

- Status: failed_benchmark
- Mechanism evidence: yes
- Mechanism confirmed: no
- Practical verdict: inconclusive
- Verdict: Built a scalar IQC-certified feedback-dynamics MVP with lifted LMI construction, numerical certification, and reproducible parameter sweeps. The mechanism manifested: the predicted stability boundary L*=1-A=0.8 was observed between L=0.79 (stable) and L=0.81 (unstable), while disturbance sensitivity scaled exactly linearly with epsilon. The certificate was feasible on stable grid points and infeasible on unstable points, although conservative near the boundary.

### Mechanism check

- Verdict: Built a scalar IQC-certified feedback-dynamics MVP with lifted LMI construction, numerical certification, and reproducible parameter sweeps. The mechanism manifested: the predicted stability boundary L*=1-A=0.8 was observed between L=0.79 (stable) and L=0.81 (unstable), while disturbance sensitivity scaled exactly linearly with epsilon. The certificate was feasible on stable grid points and infeasible on unstable points, although conservative near the boundary.
- Confidence: 8/10
- Limitations: This is a scalar toy realization, not a learned optimizer or d=20 finite-sum regression meta-training experiment. The certificate uses fixed rho=0.99 and a bounded SciPy SLSQP fallback because CVXPY was unavailable; its gamma is conservative and should be independently checked with a production SDP solver. No generalization-error or wall-clock comparison was tested.

### Practical benchmark

- Paired seeds: 8
- Baseline mean: 0
- Idea mean: 0
- p-value: 0.0522
- Paired wins: 6/8
- Benchmark verdict: no significant win

## Artifacts

- [bench_iqc.py](https://synthcore.org/code/1072/bench_iqc.py)
- [bench_report.json](https://synthcore.org/code/1072/bench_report.json)
- [iqc_experiment.py](https://synthcore.org/code/1072/iqc_experiment.py)
- [report.md](https://synthcore.org/code/1072/report.md)
- [report_bench_2026-09-01T234656.md](https://synthcore.org/code/1072/report_bench_2026-09-01T234656.md)
- [results.json](https://synthcore.org/code/1072/results.json)
- [Download all files as ZIP](https://synthcore.org/download/1072)

## Disclaimer

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