# Cycle-Aware Q-Order Scheduler

- ID: 2331
- Canonical URL: https://synthcore.org/idea/2331/cycle-aware-q-order-scheduler
- API JSON: https://synthcore.org/api/idea/2331.json
- API Markdown: https://synthcore.org/api/idea/2331.md
- Verification status: unverified
- Source: [arXiv:2608.15202](https://arxiv.org/abs/2608.15202)
- Category: dynamics
- Solves: stability, speedup
- ML areas: optimizer, training-dynamics, scheduler
- Math tags: dynamical-systems, optimization, numerical-analysis
- Ratings: usefulness 5/10; difficulty 4/10; novelty 7/10

## Idea description

Monitor optimizer convergence over a cycle of p updates instead of judging every update independently. Estimate the p-step contraction factor and effective convergence order from parameter or loss errors, then reduce learning rate only when the cycle-level contraction worsens, avoiding false alarms caused by alternating or oscillatory iterates.

## Mathematical statement

Classical Q-order convergence for a sequence x_k converging to x_* is defined by lim_{k to infinity} |x_{k+1}-x_*| divided by |x_k-x_*| raised to alpha, equal to c, where alpha is the convergence order and c is the asymptotic error constant. The paper generalizes this by comparing errors p iterations apart: lim_{k to infinity} e_{k+p}/e_k^alpha = c, where p is a positive integer, e_k is the error norm ||x_k-x_*||, x_k is the iterate, x_* is the limit, alpha is at least 1, and c is positive. For finite neural-network training, x_* is unavailable, so use a moving terminal reference bar_z_k and define e_k = ||z_k-bar_z_k||_2 + epsilon, where z_k is a low-dimensional projection of the parameter vector or a scalar loss and epsilon prevents division by zero. Estimate the p-step linear factor by c_hat_{p,k}=e_{k+p}/e_k. Estimate effective order from three p-spaced errors using alpha_hat_k = log(e_{k+p}/e_k) / log(e_k/e_{k-p}), when the logarithms are numerically safe. The transferable property is that one-step errors may oscillate while p-step errors contract consistently.

## Key formulas

- $$\lim_{k\to\infty}\frac{|x_{k+1}-x_*|}{|x_k-x_*|^{\alpha}}=c$$
- $$\lim_{k\to\infty}\frac{e_{k+p}}{e_k^{\alpha}}=c,\qquad e_k=\lVert x_k-x_*\rVert,\quad p\in\mathbb{N}$$
- $$\widehat{c}_{p,k}=\frac{e_{k+p}}{e_k},\qquad \widehat{\alpha}_k=\frac{\log(e_{k+p}/e_k)}{\log(e_k/e_{k-p})}$$
- $$\eta_{k+1}=\begin{cases}\gamma_{\downarrow}\eta_k,&\operatorname{median}_{j\in W}\widehat{c}_{p,j}>\tau_{\mathrm{bad}},\\\min(\gamma_{\uparrow}\eta_k,\eta_{\max}),&\operatorname{median}_{j\in W}\widehat{c}_{p,j}<\tau_{\mathrm{good}},\\\eta_k,&\text{otherwise.}\end{cases}$$

## Implementation notes

Integrate this into the optimizer scheduler, not into the forward network. The primary target is AdamW or SGD with momentum applied to a model with a known update period p, such as p=2 for alternating minimization or p=4 for a four-phase schedule. At every optimizer step, save a detached low-dimensional state z_k. To avoid storing all parameters, choose a fixed random projection R with r=16 to 64 columns and compute z_k=R transpose theta_k, or use the scalar validation loss. Maintain an exponential reference bar_z_k = beta bar_z_{k-1} + (1-beta) z_k, then compute e_k = norm(z_k-bar_z_k)_2 + epsilon. Store the most recent 2p errors. Once they exist, calculate c_hat_{p,k}=e_{k+p}/e_k and, when both ratios are positive and sufficiently far from one, calculate alpha_hat_k from the three p-spaced errors. Smooth c_hat with a median window W to reject minibatch noise. At cycle boundaries, reduce the learning rate by gamma_down=0.5 if the median factor exceeds tau_bad, increase it mildly by gamma_up=1.05 if it is below tau_good, and otherwise leave it unchanged; impose a cooldown of several cycles. The reference, projection dimension, thresholds, and smoothing window are empirical choices, while the p-spaced ratios are the mathematical quantities being transferred. First test a two-layer MLP and a small Transformer on MNIST or CIFAR-10 with SGD momentum and AdamW, deliberately choosing a learning rate that produces visible two-step oscillations. Compare fixed learning rates, ReduceLROnPlateau, and this scheduler for p=1, 2, and 4. Success means lower validation loss at equal optimizer steps, fewer instability spikes, and no premature learning-rate reduction when one-step loss oscillates but the p-step envelope decreases.

## Disclaimer

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