# Affinity-Controlled Three-Phase Optimizer

- ID: 2738
- Canonical URL: https://synthcore.org/idea/2738/affinity-controlled-three-phase-optimizer
- API JSON: https://synthcore.org/api/idea/2738.json
- API Markdown: https://synthcore.org/api/idea/2738.md
- Verification status: unverified
- Source: [arXiv:2608.25638](https://arxiv.org/abs/2608.25638)
- Category: dynamics
- Solves: stability, speedup, sample-efficiency
- ML areas: optimizer, training-dynamics, scheduler
- Math tags: statistical-mechanics, stochastic-processes, dynamical-systems, optimization
- Ratings: usefulness 6/10; difficulty 5/10; novelty 8/10

## Idea description

Replace a conventional optimizer step by a three-phase cyclic update in which successive parameter blocks or gradient components are exposed to two low-noise phases and one high-noise, chemically driven phase. Treat the loss decrease as mechanical work, phase-dependent gradient-noise scales as reservoir temperatures, and an auxiliary drive as chemical free energy. Adapt the drive toward a target positive cycle affinity rather than increasing the learning rate indefinitely, creating a measurable stall boundary between backward drift, zero progress, and forward optimization.

## Mathematical statement

The source model has a clockwise three-link cycle with energy levels 0, E, and 2E, two cold links at temperature T_c, one hot link at temperature T_h, mechanical load f, and chemical drive Delta_mu on the hot link. Its cycle affinity is A = E times (1/T_c minus 1/T_h) plus Delta_mu divided by T_h minus f times (2/T_c plus 1/T_h). Map E to a normalized gradient-barrier scale, T_c and T_h to noise temperatures or inverse gradient signal-to-noise ratios, f to the normalized loss-improvement requirement per cycle, and Delta_mu to an auxiliary drive injected only in the second phase. The predicted stall load is f_stall = [E(T_h - T_c) + T_c Delta_mu] divided by (2T_h + T_c). Positive affinity should produce positive average cycle current, zero affinity should produce zero current, and negative affinity should reverse the net update direction. Increasing Delta_mu should increase current but eventually reach a kinetic ceiling because only one of three bottlenecks is accelerated.

## Key formulas

- $$A=E\left(\frac{1}{T_c}-\frac{1}{T_h}\right)+\frac{\Delta\mu}{T_h}-f\left(\frac{2}{T_c}+\frac{1}{T_h}\right)$$
- $$f_{stall}=\frac{E(T_h-T_c)+T_c\Delta\mu}{2T_h+T_c}$$
- $$\Delta\mu_* = \frac{3E}{2}\quad\Longrightarrow\quad f_{stall}=\frac{E}{2}$$
- $$\dot{S}_{cycle}=J A\geq 0$$

## Implementation notes

Integrate the method as a wrapper around SGD, Adam, or Lion. Partition each minibatch update into three ordered phases, either by assigning three parameter groups or by applying three successive transformations to the same gradient. In phases one and three, inject isotropic Gaussian noise with variance T_c. In phase two, use variance T_h greater than T_c and add a bounded auxiliary impulse Delta_mu in the normalized negative-gradient direction, optionally through a separate momentum accumulator. Estimate E during a calibration period as the median loss barrier or median squared gradient norm after normalization. Define f as the required normalized loss decrease per cycle, for example f = max(0, negative target loss change) divided by three times the update scale. Before each cycle calculate A from the displayed formula. If A is below a target A_target, increase Delta_mu; if it is too large, decrease Delta_mu. Pseudocode: sample a minibatch; compute its gradient; apply phase-one update with cold noise; apply phase-two update with hot noise and the bounded chemical impulse; apply phase-three update with cold noise; record the signed cycle displacement q as the displacement projected onto the initial negative-gradient direction; update the empirical current J as the window average of q. The paper supplies the affinity and stall equations; effective temperatures, E, and kinetic rates must be estimated from training traces. First test a two-layer MLP on MNIST against SGD with matched average step norm, sweeping Delta_mu and f. The predicted signature is a current reversal near A = 0, with measured stall f within 20 percent of the formula, followed by saturation of J as Delta_mu increases.

## Disclaimer

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