# Hermitian Contraction Regularizer for Complex Neural Dynamics

- ID: 1949
- Canonical URL: https://synthcore.org/idea/1949/hermitian-contraction-regularizer-for-complex-neural-dynamic
- API JSON: https://synthcore.org/api/idea/1949.json
- API Markdown: https://synthcore.org/api/idea/1949.md
- Verification status: mech_ok_no_baseline
- Source: [arXiv:2608.07551](https://arxiv.org/abs/2608.07551)
- Category: dynamics
- Solves: stability, accuracy
- ML areas: rnn, ssm, optimizer, training-dynamics
- Math tags: dynamical-systems, control-theory, geometry, linear-algebra, optimization
- Ratings: usefulness 8/10; difficulty 6/10; novelty 7/10

## Idea description

Constrain the Jacobian of a complex-valued neural ODE or recurrent state update so that it is contracting in a state-dependent Hermitian metric. The resulting model should forget perturbations and initialization differences exponentially, improving long-horizon rollout stability while retaining coordinate-invariant stability information.

## Mathematical statement

Let $z\in\mathbb{C}^{n}$ be the hidden state, $\dot z=f_\theta(z,t)$ a holomorphic vector field, and $J(z,t)=\partial f_\theta/\partial z$ its complex Jacobian. Let $H(z,t)=H(z,t)^*\succ0$ be a Hermitian metric and define $V(\delta z,z)=\delta z^*H(z,t)\delta z$ for an infinitesimal displacement $\delta z$. Along $\dot{\delta z}=J\delta z$, the sufficient contraction condition is $\dot H+J^*H+HJ\preceq-2\alpha H$ for rate $\alpha>0$, where $\dot H$ is the directional derivative of $H$ along the state trajectory and $J^*$ is the conjugate transpose. If $mI\preceq H\preceq MI$ on a forward-invariant set, the induced distance contracts at rate $\alpha$ up to the factor $\sqrt{M/m}$.

## Key formulas

- $$\dot H(z,t)+J(z,t)^*H(z,t)+H(z,t)J(z,t)\preceq-2\alpha H(z,t),\qquad H(z,t)=H(z,t)^*\succ0.$$
- $$V(t)=\delta z(t)^*H(z(t),t)\delta z(t),\qquad \dot V(t)\le-2\alpha V(t),\qquad V(t)\le e^{-2\alpha t}V(0).$$
- $$d_H(z_1(t),z_2(t))\le\sqrt{\frac{M}{m}}e^{-\alpha t}d_H(z_1(0),z_2(0)),\qquad mI\preceq H\preceq MI.$$

## Implementation notes

Use a complex-valued neural ODE, continuous-time state-space model, or complex RNN with hidden dynamics $\dot z=f_\theta(z,x,t)$. Parameterize the metric as $H_\phi(z)=R_\phi(z)^*R_\phi(z)+\epsilon I$, with $\epsilon>0$, or start with a constant diagonal positive metric for the minimum viable experiment. For each minibatch of states, compute $f$, the complex Jacobian $J$ using automatic differentiation, and the directional derivative $\dot H$; form $C=\dot H+J^*H+HJ+2\alpha H$ and add $\lambda_H\max(0,\lambda_{\max}(H^{-1/2}CH^{-1/2}))^2$ to the task loss. Alternatively reduce the ODE solver step when the largest generalized eigenvalue violates the target bound. Jacobians, matrix products, and eigenvalues are computed directly; the invariant region and global metric bounds $m,M$ are estimated from sampled states or interval bounds. The first cheap experiment should train a two-layer complex neural ODE on damped-oscillator trajectory prediction, comparing ordinary MSE training, spectral-Jacobian regularization, and the Hermitian penalty. Perturb initial states and fit the semilog slope of pairwise distances. The prediction is that the observed slope is no greater than $-\alpha$ after transients, within 20% of the sampled certificate. Sweeping penalty strength should reveal a transition when the maximum sampled generalized eigenvalue crosses zero: beyond that boundary perturbations decay, while below it they can persist or amplify.

## Verification

- Status: mech_ok_no_baseline
- Mechanism evidence: yes
- Mechanism confirmed: yes
- Practical verdict: inconclusive
- Verdict: Built a complex linear-dynamics MVP with a Hermitian contraction certificate and a CUDA mini-training comparison. The certificate matched 2(alpha-gamma), the contraction boundary matched gamma=alpha across five alpha values, perturbation slopes matched -gamma, and the result was invariant under a fixed nonidentity metric. The mechanism is real, but this experiment did not demonstrate a practical improvement over baseline because both models already learned strongly contracting dynamics.

### Mechanism check

- Verdict: Built a complex linear-dynamics MVP with a Hermitian contraction certificate and a CUDA mini-training comparison. The certificate matched 2(alpha-gamma), the contraction boundary matched gamma=alpha across five alpha values, perturbation slopes matched -gamma, and the result was invariant under a fixed nonidentity metric. The mechanism is real, but this experiment did not demonstrate a practical improvement over baseline because both models already learned strongly contracting dynamics.
- Confidence: 9/10
- Limitations: The MVP uses constant metrics rather than a learned state-dependent metric, and a 2D complex linear system rather than a nonlinear neural ODE. It does not test nonlinear long-horizon forecasting, adaptive solvers, global metric bounds, or whether regularization rescues an unstable baseline.

### Practical benchmark

- Paired seeds: 8
- Baseline mean: 0
- Idea mean: 0
- p-value: 0.05725
- Paired wins: 2/8
- Benchmark verdict: no significant win

## Artifacts

- [bench_experiment.py](https://synthcore.org/code/734/bench_experiment.py)
- [bench_report.json](https://synthcore.org/code/734/bench_report.json)
- [experiment.py](https://synthcore.org/code/734/experiment.py)
- [report.md](https://synthcore.org/code/734/report.md)
- [report_bench_2026-09-01T022029.md](https://synthcore.org/code/734/report_bench_2026-09-01T022029.md)
- [results.json](https://synthcore.org/code/734/results.json)
- [Download all files as ZIP](https://synthcore.org/download/734)

## Disclaimer

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