{
 "artifacts": null,
 "category": "dynamics",
 "description": "Treat discrete training events such as gradient-norm spikes, curvature changes, rejected steps, or minibatch outliers as jump channels and apply an event-specific parameter update map. The optimizer should be evaluated using both progress and the information cost of selecting the feedback map, because feedback may reduce loss fluctuations or improve adaptation without changing the average update magnitude.",
 "formulas_latex": [
  "$$\\dot{\\rho}=i[\\rho,H]+\\sum_{k\\geq1}\\left(\\mathcal{F}_k[L_k\\rho L_k^\\dagger]-\\frac{1}{2}\\{L_k^\\dagger L_k,\\rho\\}\\right),$$",
  "$$\\theta_{t+1}=F_{k_t}(\\theta_t,g_t),\\qquad k_t\\sim p_t(k\\mid z_t),$$",
  "$$F_k(\\theta,g)=\\theta-\\eta A_k g,\\qquad I_{\\mathrm{fb}}(t)=\\sum_kp_t(k)\\log\\frac{p_t(k)}{q(k)},$$",
  "$$\\mathcal{R}_T=\\frac{\\operatorname{Var}(J_T)}{\\mathbb{E}[J_T]^2}\\left(\\Sigma_T+\\sum_{t=1}^{T}I_{\\mathrm{fb}}(t)\\right),\\qquad J_T=\\sum_{t=1}^{T}[\\ell(\\theta_{t-1})-\\ell(\\theta_t)],$$"
 ],
 "id": 2982,
 "implementation": "Integrate this as a wrapper around SGD or Adam, leaving the neural-network architecture unchanged. At each minibatch compute the gradient g_t, loss \\ell_t, gradient norm, cosine similarity c_t=g_t^\\top g_{t-1}/(\\|g_t\\|\\|g_{t-1}\\|), and a cheap curvature proxy such as \\|g_t-g_{t-1}\\|/(\\|\\theta_t-\\theta_{t-1}\\|+\\epsilon). Discretize these features into event labels k: normal step, gradient spike, sign reversal, plateau, or high-loss minibatch. Maintain one small diagonal preconditioner A_k for each label and apply \\theta\\leftarrow\\theta-\\eta A_k g_t. Estimate p_t(k) with an exponential-moving histogram and calculate the information charge relative to a uniform prior q. Pseudocode: compute g_t and z_t; assign k_t=event_classifier(z_t); update p_t; compute I_t=\\sum_kp_t(k)\\log((p_t(k)+\\epsilon)/(q(k)+\\epsilon)); select A_{k_t}; update parameters; record loss decrease, step norm, and J_t. The paper supplies the event-conditioned feedback mechanism; event bins, A_k, and the estimator are empirical adaptations. First test on a two-layer MLP on MNIST or CIFAR-10, comparing vanilla SGD, fixed per-event preconditioners, and feedback with an information penalty that discourages rare modes. Match mean step norm and mean loss decrease across methods. The predicted signature is a reduction in \\operatorname{Var}(J_T)/\\mathbb{E}[J_T]^2 at fixed mean progress, with the reduction increasing with cumulative feedback information and disappearing when event labels are randomly shuffled. Sweep the event threshold and identify the stability boundary in learning rate; feedback should not reduce variance beyond its measured information-cost budget.",
 "math_summary": "The paper models an event channel k with jump operator L_k and an immediately applied completely positive trace-preserving map F_k, giving the feedback generator \\(\\dot{\\rho}=i[\\rho,H]+\\sum_k(F_k[L_k\\rho L_k^\\dagger]-\\frac12\\{L_k^\\dagger L_k,\\rho\\})\\). For optimization, replace the state \\(\\rho\\) by parameters \\(\\theta\\), the observed jump channel by an event label \\(k_t\\), and F_k by an update map \\(F_k\\). Let p_t(k) be the empirical probability of selecting mode k and q(k) a fixed no-feedback prior. The information cost of feedback is estimated by \\(I_{\\mathrm{fb}}(t)=\\sum_kp_t(k)\\log[p_t(k)/q(k)]\\). The transferred mechanism predicts lower progress-current variance at fixed mean progress when event-conditioned maps exploit observed training events, with the gain charged to cumulative information cost.",
 "math_tags": [
  "control-theory",
  "stochastic-processes",
  "statistics",
  "information-theory",
  "dynamical-systems"
 ],
 "ml_areas": [
  "optimizer",
  "training-dynamics",
  "scheduler"
 ],
 "paper": {
  "arxiv_id": "2609.00622",
  "arxiv_url": "https://arxiv.org/abs/2609.00622",
  "summary_what_math_gives_to_ml": "The paper's transferable mechanism is jump-conditioned feedback: an event-dependent completely positive map changes the post-event state while preserving explicit accounting of fluctuations, Fisher information, and information-processing cost. For neural optimization, this suggests an event-conditioned optimizer that applies different update maps after identifiable training events such as gradient spikes, sign reversals, curvature changes, or plateaus. The central falsifiable prediction is that feedback can reduce progress-current fluctuations at fixed mean progress, but the improvement should correlate with the information cost required to select feedback modes.",
  "title": "Feedback-Enhanced Quantum Metrology and Clock Precision under Thermodynamic Uncertainty",
  "year": "2026"
 },
 "ratings": {
  "difficulty": 5,
  "novelty": 6,
  "usefulness": 6
 },
 "solves": [
  "stability",
  "speedup",
  "sample-efficiency"
 ],
 "title": "Information-Charged Event-Feedback Optimizer",
 "url": "https://synthcore.org/idea/2982/information-charged-event-feedback-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
   }
  }
 }
}