# Delay-Budget Controller for Coupled Training

- ID: 302
- Canonical URL: https://synthcore.org/idea/302/delay-budget-controller-for-coupled-training
- API JSON: https://synthcore.org/api/idea/302.json
- API Markdown: https://synthcore.org/api/idea/302.md
- Verification status: failed_benchmark
- Source: [arXiv:2607.00227](https://arxiv.org/abs/2607.00227)
- Category: dynamics
- Solves: stability, speedup
- ML areas: optimizer, training-dynamics, gan, distributed
- Math tags: dynamical-systems, control-theory, numerical-analysis
- Ratings: usefulness 6/10; difficulty 5/10; novelty 6/10

## Idea description

Treat a coupled neural training loop as a delayed feedback system with two hard delays and two first-order implementation filters. Estimate the dominant coupled Jacobian mode and use the characteristic equation to distinguish a recoverable delay-induced oscillation from a filter-induced instability; then reduce stale-gradient delay only in the former case, and slow or retune EMA or relaxation filters in the latter.

## Mathematical statement

The paper gives the linearized characteristic factor for tangent spectral value \(\gamma_\rho\): \((z+\mu_X)(z+\mu_Y)(z+\kappa_X)(z+\kappa_Y)=\frac{\lambda_X\lambda_Y\kappa_X\kappa_Y}{m^2}\gamma_\rho e^{-z\tau_\Sigma}\), with total hard delay \(\tau_\Sigma=\sigma_X+\sigma_Y+\theta_X+\theta_Y\). Here \(z\) is a continuous-time growth rate, \(\mu_X,\mu_Y>0\) are effective response rates of the two coupled learners, \(\kappa_X,\kappa_Y>0\) are first-order implementation-filter rates, \(\lambda_X,\lambda_Y\) are coupling gains, \(m\) is a normalization scale, and \(\gamma_\rho\) is an eigenvalue of the projected coupled interaction operator. For an antagonistic mode set \(\gamma_\rho=-|\gamma_\rho|\) and define \(G=(\lambda_X\lambda_Y\kappa_X\kappa_Y/m^2)|\gamma_\rho|\). A Hopf crossing \(z=i\omega\), \(\omega>0\), satisfies \((\mu_X^2+\omega^2)(\mu_Y^2+\omega^2)(\kappa_X^2+\omega^2)=G\) after taking magnitudes, and \(\omega\tau_\Sigma+\arctan(\omega/\mu_X)+\arctan(\omega/\mu_Y)+\arctan(\omega/\kappa_X)+\arctan(\omega/\kappa_Y)=(2q+1)\pi\) from the phase. The zero-delay filter margin is \(G<\mu_X\mu_Y\kappa_X\kappa_Y\); if it fails, reducing hard delay alone cannot restore stability.

## Key formulas

- $$(z+\mu_X)(z+\mu_Y)(z+\kappa_X)(z+\kappa_Y)=\frac{\lambda_X\lambda_Y\kappa_X\kappa_Y}{m^{2}}\,\gamma_{\rho}e^{-z\tau_{\Sigma}},\qquad \tau_{\Sigma}=\sigma_X+\sigma_Y+\theta_X+\theta_Y.$$
- $$G=\frac{\lambda_X\lambda_Y\kappa_X\kappa_Y}{m^2}|\gamma_\rho|,\qquad (\mu_X^2+\omega^2)(\mu_Y^2+\omega^2)(\kappa_X^2+\omega^2)=G.$$
- $$\omega\tau_\Sigma+\tan^{-1}\!\frac{\omega}{\mu_X}+\tan^{-1}\!\frac{\omega}{\mu_Y}+\tan^{-1}\!\frac{\omega}{\kappa_X}+\tan^{-1}\!\frac{\omega}{\kappa_Y}=(2q+1)\pi.$$
- $$\tau_{\mathrm{crit}}(\omega,q)=\frac{(2q+1)\pi-\tan^{-1}(\omega/\mu_X)-\tan^{-1}(\omega/\mu_Y)-\tan^{-1}(\omega/\kappa_X)-\tan^{-1}(\omega/\kappa_Y)}{\omega}.$$

## Implementation notes

(1) Integration point: apply this controller to a two-player training loop, initially a small GAN or actor-critic. The two players are X and Y; sigma_X and sigma_Y are measured stale-gradient or stale-parameter ages, theta_X and theta_Y are explicit communication or update delays, and kappa_X and kappa_Y are EMA target rates or relaxation rates. Instrument the online-to-opponent feedback path rather than changing the model architecture. Use delayed opponent snapshots and EMA deployment copies so the experiment contains the four relevant poles. (2) Pseudocode: every control interval, collect the last K parameter and gradient pairs; estimate local cross-Jacobian blocks J_XY and J_YX with two or three Hutchinson or finite-difference probes; set |gamma_rho| approximately equal to rho(J_XY J_YX) using 5-step power iteration; estimate lambda_X and lambda_Y from the norms of the corresponding update blocks; set mu_X approximately eta_X and mu_Y approximately eta_Y in normalized time; and set kappa_i equal to -log(beta_i), where beta_i is the EMA coefficient. Compute G, solve the scalar magnitude equation for omega by bisection, and compute tau_crit from the phase formula. If measured tau_Sigma exceeds 0.8 tau_crit while G is below the zero-delay margin, reduce stale-update age or communication interval. If G reaches or exceeds the margin, do not merely reduce delay: increase filter rates, reduce EMA inertia, lower coupling learning rates, or add gradient damping. (3) The characteristic equation, magnitude condition, phase condition, and filter margin come from the paper; Jacobian spectral radius, effective gains, continuous-time rates, and actual delays are estimated from traces. Smooth estimates with EMAs and require warnings to persist for several windows. (4) First experiment: compare AdamW GAN training on CIFAR-10 with synchronous updates, artificially delayed opponent snapshots of 0 to 8 steps, and EMA coefficients beta in {0, 0.9, 0.99}. Compare fixed baseline settings against the controller at equal wall-clock budget. Record loss spectral peaks, gradient spikes, FID, mode coverage, and divergence rate. The expected signal is fewer oscillatory loss bursts and lower divergence probability at the same delay; in the filter-induced regime, reducing delay alone should fail while reducing EMA inertia should recover stability.

## Verification

- Status: failed_benchmark
- Mechanism evidence: yes
- Mechanism confirmed: no
- Practical verdict: inconclusive
- Verdict: Built a readable four-pole delayed-feedback MVP with the stated characteristic equation, controller branch logic, numerical root estimation, and Euler DDE simulation. The magnitude and phase equations matched numerically to approximately 5e-13 relative magnitude error and zero phase error; the dominant root crossed from -0.0061 at 0.8 tau_crit to +0.0033 at 1.2 tau_crit, and reducing delay lowered tail RMS from 0.01357 to 0.00749. In the filter-margin case, increasing filter rates reduced tail RMS from 1.84e-9 to 2.10e-22 at zero delay, while delay-only intervention did nothing; this is a promising toy-level confirmation, not evidence of a neural-training win.

### Mechanism check

- Verdict: Built a readable four-pole delayed-feedback MVP with the stated characteristic equation, controller branch logic, numerical root estimation, and Euler DDE simulation. The magnitude and phase equations matched numerically to approximately 5e-13 relative magnitude error and zero phase error; the dominant root crossed from -0.0061 at 0.8 tau_crit to +0.0033 at 1.2 tau_crit, and reducing delay lowered tail RMS from 0.01357 to 0.00749. In the filter-margin case, increasing filter rates reduced tail RMS from 1.84e-9 to 2.10e-22 at zero delay, while delay-only intervention did nothing; this is a promising toy-level confirmation, not evidence of a neural-training win.
- Confidence: 7/10
- Limitations: Only a scalar linear toy DDE was tested; no GAN, neural network, parameter or gradient trace estimator, Hutchinson Jacobian estimation, noisy measurements, wall-clock comparison, FID, mode coverage, or divergence-rate study was performed. The implementation uses a four-factor magnitude equation consistent with the characteristic polynomial; the idea text's displayed magnitude formula appears to omit one filter factor. Filter-rate changes hold G fixed, so they do not model every physical dependence of G on kappa.

### Practical benchmark

- Paired seeds: 8
- Baseline mean: 0
- Idea mean: 0
- p-value: 0.11805
- Paired wins: 6/8
- Benchmark verdict: no significant win

## Artifacts

- [bench_delay_controller.py](https://synthcore.org/code/1261/bench_delay_controller.py)
- [bench_report.json](https://synthcore.org/code/1261/bench_report.json)
- [delay_controller.py](https://synthcore.org/code/1261/delay_controller.py)
- [report.md](https://synthcore.org/code/1261/report.md)
- [report_bench_2026-09-03T232010.md](https://synthcore.org/code/1261/report_bench_2026-09-03T232010.md)
- [results.json](https://synthcore.org/code/1261/results.json)
- [Download all files as ZIP](https://synthcore.org/download/1261)

## Disclaimer

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