{
 "artifacts": null,
 "category": "dynamics",
 "description": "Replace a purely memoryless optimizer step by a partially observed feedback controller for parameters evolving under colored, active gradient fluctuations. Estimate the hidden persistent component of the gradient from parameter displacement and observed minibatch gradients, then use that estimate to cancel predictable activity or adapt the effective update target without directly observing the latent disturbance.",
 "formulas_latex": [
  "$$V(\\theta,\\lambda)=\\frac{1}{2}(\\theta-\\lambda)^\\top K(\\theta-\\lambda),\\qquad W=\\mathbb{E}\\left[\\int_0^T \\nabla_\\lambda V(\\theta,\\lambda)^\\top\\circ d\\lambda\\right].$$",
  "$$\\dot\\theta=-K(\\theta-\\lambda)+a+\\xi,\\qquad \\dot a=-\\frac{1}{\\tau}a+\\sqrt{\\frac{2\\Sigma_a}{\\tau}}\\,\\zeta,$$",
  "$$\\dot\\lambda=v_0-G(\\theta-\\lambda)-G_a\\hat a,$$",
  "$$A_{\\mathrm{cl}}=I+\\Delta t\\,A_{\\mathrm{cont}},\\qquad \\text{stable if }\\rho(A_{\\mathrm{cl}})\u003c1.$$$"
 ],
 "id": 3077,
 "implementation": "1. Integration point: begin with a one-dimensional and then low-dimensional quadratic-loss problem, followed by a two-layer MLP on MNIST. Treat the parameter error e_t = theta_t - lambda_t as the observed position. Maintain a hidden-activity estimate a_hat and covariance P for each parameter block or for a low-rank gradient subspace. 2. Pseudocode: initialize a_hat = 0 and P. At each minibatch, compute gradient g_t and predict a_hat \u003c- exp(-Delta/tau) a_hat; propagate P using the OU process covariance. Predict the gradient residual r_t from the local curvature model, observe r_t from the minibatch gradient, and perform a Kalman measurement update to obtain a_hat. Form u_t = v_0 - G e_t - G_a a_hat, update lambda \u003c- lambda + Delta u_t, and update parameters with theta \u003c- theta - eta (g_t - c a_hat), where c is a tunable cancellation gain. Estimate tau from the empirical autocorrelation of minibatch-gradient residuals, Sigma_a from residual covariance, and K from diagonal Hessian estimates or Hessian-vector products. Compute the linearized augmented closed-loop matrix periodically; reduce eta or gains whenever rho(A_cl) exceeds 0.98. 3. Taken from the paper are the OU latent-disturbance model, inference from position-only observations, and feedback objective; curvature, noise covariance, observation covariance, and gains must be estimated or tuned in the neural-network experiment. 4. First cheap experiment: compare SGD, Adam, momentum, and this controller on a quadratic objective with synthetic OU gradient noise, then on an MLP for MNIST. Sweep tau over 0.1, 1, 10, and 100 optimizer steps. Prediction: filtering gives little benefit for tau much smaller than one step, strongest improvement when tau matches the optimizer relaxation time, and a residual benefit for tau much larger than the training horizon. On the quadratic system, divergence should occur at rho(A_cl) = 1, with the measured learning-rate boundary within 20 percent of the spectral prediction.",
 "math_summary": "Let theta be the parameter vector, lambda an effective moving target, K positive definite local curvature, a a hidden active gradient disturbance, and xi white noise. The local model is dot(theta) = -K(theta-lambda) + a + xi, while a follows an Ornstein–Uhlenbeck process with persistence time tau and stationary covariance Sigma_a. Only theta and noisy gradients are observed, so a Kalman filter produces a_hat. Feedback uses dot(lambda) = v_0 - G(theta-lambda) - G_a a_hat. In discrete time, stability requires the spectral radius of the augmented closed-loop transition matrix A_cl to satisfy rho(A_cl) \u003c 1. The work-like quantity being minimized is the expected parameter-space work generated by moving the quadratic potential center.",
 "math_tags": [
  "control-theory",
  "stochastic-processes",
  "dynamical-systems",
  "optimization"
 ],
 "ml_areas": [
  "optimizer",
  "training-dynamics",
  "scheduler"
 ],
 "paper": {
  "arxiv_id": "2609.02720",
  "arxiv_url": "https://arxiv.org/abs/2609.02720",
  "summary_what_math_gives_to_ml": "The paper provides a nontrivial partial-observation control mechanism: an active Ornstein–Uhlenbeck degree of freedom is hidden, but its correlation with the measured particle position allows work extraction and near-optimal feedback using position alone. The transferable asset is a controller that infers persistent latent activity from observable state histories rather than measuring the disturbance directly. In neural-network optimization, parameters or activations can be modeled as trapped states driven by colored gradient noise, and a Kalman-like estimator can separate predictable activity from irreducible stochastic noise before applying a feedback update. The key falsifiable prediction is that the benefit depends sharply on gradient-noise persistence time and remains nonzero for very large persistence when latent activity is inferred from parameter or gradient history.",
  "title": "Optimal-work feedback on particles with activity --- gliding on active fluctuations using positional information",
  "year": "2026"
 },
 "ratings": {
  "difficulty": 6,
  "novelty": 7,
  "usefulness": 7
 },
 "solves": [
  "stability",
  "speedup",
  "accuracy"
 ],
 "title": "Position-only active-noise optimizer",
 "url": "https://synthcore.org/idea/3077/position-only-active-noise-optimizer",
 "verification": {
  "peer_reviewed": false,
  "status": "unverified",
  "status_label": "Unverified",
  "verdict_source": "deterministic test code (paired-seed permutation statistics)",
  "verification_axes": {
   "benchmark_mechanism": {
    "confirmed": null,
    "tested": false
   },
   "practical_benchmark": {
    "beats_baseline": null,
    "tested": false
   },
   "toy_mechanism_gate": {
    "confirmed": null,
    "tested": false
   }
  }
 }
}