{
 "artifacts": null,
 "category": "dynamics",
 "description": "Replace continuous stochastic-gradient updates by a flashing schedule with alternating ON phases, where gradients act normally, and OFF phases, where gradients are suppressed or weakened and controlled noise allows escape from local traps. Estimate directional asymmetry of the local loss basin from forward and backward probe distances, then set the flashing frequency using the ratchet resonance law so that noise-assisted transitions preferentially produce net progress toward lower loss.",
 "formulas_latex": [
  "$$\\dot{x}=-\\mu U_0\\sigma(t)u'(x)+\\sqrt{2D_0}\\,\\xi(t),\\qquad \\sigma(t)\\in\\{0,1\\},$$",
  "$$\\delta=\\frac{s_2-s_1}{s_1+s_2},\\qquad s_1+s_2=\\lambda,$$",
  "$$\\nu_\\star(\\delta,\\Delta)=\\frac{\\nu_0(\\Delta)}{1-b_0\\delta^2},\\qquad J_{\\max}(\\delta,\\Delta)\\approx C(\\Delta)\\delta\\quad (|\\delta|\u003c0.5),$$"
 ],
 "id": 2860,
 "implementation": "1. Integration point: implement this as an optimizer wrapper around a standard minibatch gradient update. For each parameter block or for the full parameter vector, alternate an ON phase of ordinary AdamW or SGD descent with an OFF phase in which the gradient coefficient is reduced to zero or to a small value \\(\\alpha_{\\rm off}\\), while isotropic or gradient-preconditioned Gaussian noise is applied. The cycle frequency is the inverse number of minibatches per ON+OFF cycle. 2. Pseudocode: initialize \\(\\theta\\), cycle length \\(L\\), and fitted \\(b_0\\). Every \\(K\\) steps, choose a normalized probe direction \\(v\\), evaluate losses at \\(\\theta\\pm rv\\), and increase \\(r\\) until the loss rises by a prescribed barrier threshold \\(\\Delta_{\\rm probe}\\). Let \\(s_1\\) and \\(s_2\\) be the positive distances to the two threshold crossings, assign their ordering so that \\(s_1+s_2\\) is fixed, compute \\(\\delta=(s_2-s_1)/(s_1+s_2)\\), and set \\(\\nu_\\star=\\nu_0/(1-b_0\\delta^2)\\). Use \\(L=\\operatorname{round}(B/\\nu_\\star)\\), where \\(B\\) is a calibration window, and perform ON updates \\(\\theta\\leftarrow\\theta-\\eta g\\) for \\(L_{\\rm on}\\) steps and OFF updates \\(\\theta\\leftarrow\\theta-\\eta\\alpha_{\\rm off}g+\\sqrt{2\\eta T_{\\rm off}}\\,\\epsilon\\) for \\(L_{\\rm off}\\) steps. 3. Computed from the paper: the functional form \\((1-b_0\\delta^2)^{-1}\\) and the prediction of linear signed current. Estimated empirically: \\(\\nu_0(\\Delta)\\), \\(b_0\\), probe distances, and the best ON/OFF duty cycle. Fit \\(\\nu_0\\) and \\(b_0\\) from a short logarithmic sweep of cycle frequencies on a small training subset, then freeze them. 4. First cheap experiment: train a two-layer MLP on MNIST or CIFAR-10 with SGD, AdamW, and flashing SGD at 8--12 cycle frequencies. Measure loss decrease per 100 updates, escape rate from deliberately initialized sharp minima, and signed displacement along the probe direction. The required signature is a nonmonotonic progress curve versus frequency with a peak within 20% of \\(\\nu_0/(1-b_0\\delta^2)\\); across runs with imposed probe asymmetries \\(|\\delta|\u003c0.5\\), the peak displacement should change approximately linearly with \\(\\delta\\), and reversing the probe direction should reverse the measured current.",
 "math_summary": "The source system alternates an active asymmetric potential with a flat potential: \\(\\dot{x}=-\\mu U_0\\sigma(t)u'(x)+\\sqrt{2D_0}\\xi(t)\\), where \\(\\sigma(t)\\in\\{0,1\\}\\), \\(u\\) is periodic, \\(\\mu\\) is mobility, and \\(D_0\\) is diffusivity. Its branch lengths \\(s_1,s_2\\) define asymmetry \\(\\delta=(s_2-s_1)/(s_1+s_2)\\), and the measured resonant switching frequency obeys \\(\\nu(\\delta,\\Delta)=\\nu_0(\\Delta)/(1-b_0\\delta^2)\\), with barrier height \\(\\Delta\\), baseline resonance \\(\\nu_0\\), and fitted constant \\(b_0\\). The adapted optimizer treats a one-dimensional local training coordinate \\(z\\) as the ratchet coordinate, estimates distances \\(s_1,s_2\\) to loss barriers along a probe direction, and uses \\(\\nu\\) as the ON/OFF cycle frequency. The falsifiable prediction is that progress per update is nonmonotonic in switching frequency and peaks near the predicted \\(\\nu\\), while the signed progress amplitude is approximately linear in \\(\\delta\\) for \\(|\\delta|\u003c0.5\\).",
 "math_tags": [
  "statistical-mechanics",
  "stochastic-processes",
  "dynamical-systems",
  "bifurcations",
  "optimization"
 ],
 "ml_areas": [
  "optimizer",
  "scheduler",
  "training-dynamics"
 ],
 "paper": {
  "arxiv_id": "2608.29991",
  "arxiv_url": "https://arxiv.org/abs/2608.29991",
  "summary_what_math_gives_to_ml": "The paper provides a nonstandard resonance mechanism: randomly flashing an asymmetric periodic potential produces a directed stationary current that is maximal at an intermediate switching frequency. Its transferable asset is the quantitative dependence of the resonant frequency on geometric asymmetry, \\(\\nu(\\delta,\\Delta)=\\nu_0(\\Delta)/(1-b_0\\delta^2)\\), together with approximately linear current amplitude in \\(\\delta\\). A neural-network analogue is a curvature-aware optimizer that alternates between deterministic loss descent and stochastic exploration, estimates local asymmetry of the loss landscape, and selects the switching frequency predicted to maximize net directional progress rather than injecting noise continuously.",
  "title": "Asymmetry-controlled resonant transport in a Brownian flashing ratchet",
  "year": "2026"
 },
 "ratings": {
  "difficulty": 6,
  "novelty": 8,
  "usefulness": 5
 },
 "solves": [
  "stability",
  "speed",
  "sample-efficiency"
 ],
 "title": "Asymmetry-Tuned Flashing Optimizer",
 "url": "https://synthcore.org/idea/2860/asymmetry-tuned-flashing-optimizer",
 "verification": {
  "peer_reviewed": false,
  "status": "unverified",
  "status_label": "Unverified",
  "verdict_source": "deterministic test code (paired-seed permutation statistics)"
 }
}