# Correlation-Irreversibility Learning-Rate Controller

- ID: 1753
- Canonical URL: https://synthcore.org/idea/1753/correlation-irreversibility-learning-rate-controller
- API JSON: https://synthcore.org/api/idea/1753.json
- API Markdown: https://synthcore.org/api/idea/1753.md
- Verification status: unverified
- Source: [arXiv:2608.03619](https://arxiv.org/abs/2608.03619)
- Category: dynamics
- Solves: stability, speedup
- ML areas: optimizer, training-dynamics, scheduler
- Math tags: stochastic-processes, statistical-mechanics, dynamical-systems, probability, control-theory
- Ratings: usefulness 6/10; difficulty 5/10; novelty 7/10

## Idea description

Measure time-reversal asymmetry in coarse-grained parameter or update trajectories and convert it into a lower bound on the irreversibility of training dynamics. Use this bound as a feedback signal: when irreversible circulation increases sharply, reduce the learning rate or momentum; when it remains low and the loss decreases, permit larger steps.

## Mathematical statement

For a continuous-time Markov state X_t with stationary probabilities p_i and transition rates k_{ij}, define the stationary probability current J_{ij}=p_i k_{ij}-p_j k_{ji} for each unordered edge i<j. The entropy-production rate is sigma=sum_{i<j} J_{ij} log[(p_i k_{ij})/(p_j k_{ji})]. The inequality (a-b)log(a/b)>=2(a-b)^2/(a+b) gives a quadratic lower bound on sigma. For state observables f and g, the two-time correlation is C_fg(tau)=E[f(X_t)g(X_{t+tau})], and its antisymmetric part A_fg(tau)=C_fg(tau)-C_gf(tau) measures time-reversal breaking. For indicator observables of states i and j, A_ij(tau)=tau J_ij+O(tau^2), so at short lag the current bound can be estimated from correlation asymmetry. In the neural-network application, X_t is a coarse optimizer state containing gradient, update, momentum, and loss statistics.

## Key formulas

- $$\sigma=\sum_{i<j}J_{ij}\log\frac{p_i k_{ij}}{p_j k_{ji}},\qquad J_{ij}=p_i k_{ij}-p_j k_{ji}.$$
- $$\sigma\ge 2\sum_{i<j}\frac{J_{ij}^{2}}{p_i k_{ij}+p_j k_{ji}},\qquad (a-b)\log(a/b)\ge\frac{2(a-b)^2}{a+b}.$$
- $$C_{fg}(\tau)=\mathbb E[f(X_t)g(X_{t+\tau})],\qquad A_{fg}(\tau)=C_{fg}(\tau)-C_{gf}(\tau).$$
- $$A_{ij}(\tau)=\tau J_{ij}+O(\tau^2),\qquad \widehat{\sigma}_{\rm corr}(\tau)=2\sum_{i<j}\frac{A_{ij}(\tau)^2}{\tau^2[\widehat p_i\widehat k_{ij}+\widehat p_j\widehat k_{ji}]}.$$

## Implementation notes

Integrate the controller between optimizer updates without changing the model architecture. Every K minibatches, construct a low-dimensional coarse state z_t from the current training dynamics: concatenate log gradient norm, parameter-update norm, cosine similarity between successive updates, momentum norm, minibatch loss, and optionally 8-32 random projections of the parameter vector. Normalize coordinates with exponential moving means and variances, then discretize z_t into B bins or cluster it into B states. Store a rolling window of states and estimate transition counts n_ij at a lag of one or several optimizer steps. Set p_i=n_i divided by the total outgoing count and estimate k_ij=n_ij divided by the product of the timestep and n_i, using symmetric pseudocounts to avoid zero denominators. Compute J_ij=p_i k_ij-p_j k_ji and the quadratic lower bound. Smooth it with an EMA. Pseudocode is: observe z_t; assign state i; update n_ij; estimate p and k; compute sigma_lb=2 sum((p_i k_ij-p_j k_ji)^2/(p_i k_ij+p_j k_ji+epsilon)); if sigma_lb exceeds a threshold or its slope is positive, set learning_rate to 0.5 times its current value and momentum to 0.8 times its current value; if sigma_lb is low and loss decreases, multiply learning_rate by 1.02. The paper supplies the current-based lower-bound mechanism; coarse states, transition estimates, lag, and thresholds are empirical. First test a 2-layer MLP on MNIST and a small Transformer on CIFAR-10, comparing SGD, momentum SGD, and Adam across learning rates. Measure divergence step, loss oscillation, gradient-norm growth, and sigma_lb. The quantitative prediction is that correlation asymmetry rises before instability: its median value should cross a reproducible threshold 5-20 optimizer steps before gradient-norm divergence, and feedback control should extend the stable learning-rate ceiling. Also verify A_ij(tau) proportional to tau at short lags; failure indicates non-Markovian coarse dynamics and invalidates the simple estimator.

## Disclaimer

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