{
 "artifacts": [
  {
   "name": "analyze_results.py",
   "url": "https://synthcore.org/code/1028/analyze_results.py"
  },
  {
   "name": "report.md",
   "url": "https://synthcore.org/code/1028/report.md"
  },
  {
   "name": "results.json",
   "url": "https://synthcore.org/code/1028/results.json"
  },
  {
   "name": "run_experiment.py",
   "url": "https://synthcore.org/code/1028/run_experiment.py"
  }
 ],
 "category": "dynamics",
 "description": "Replace or augment a deterministic recurrent hidden state with a stochastic Markov transition, then explicitly measure its entropy production and output memory time. Penalize operating points where the target changes faster than the hidden state can track at the available dissipation, while allowing the model to satisfy the bound either by increasing transition activity or by developing a longer-lived memory mode.",
 "download_zip": "https://synthcore.org/download/1028",
 "formulas_latex": [
  "$$\\sigma(t)=\\frac{1}{2}\\sum_{i,j}\\left[p_i(t)q_{ij}(t)-p_j(t)q_{ji}(t)\\right]\\log\\left(\\frac{p_i(t)q_{ij}(t)}{p_j(t)q_{ji}(t)}\\right)=\\sum_r\\left(j_r^+-j_r^-\\right)\\log\\frac{j_r^+}{j_r^-}.$$",
  "$$\\tau_y=\\int_0^\\infty C_y(s)\\,ds,\\qquad C_y(s)=\\frac{\\mathbb E_{\\rm eq}[(y(t)-\\bar y)(y(t+s)-\\bar y)]}{\\operatorname{Var}_{\\rm eq}(y)}.$$",
  "$$\\epsilon_{\\rm lag}\\,\\sigma\\;\\gtrsim\\;\\kappa\\,\\tau_y\\left\\|\\frac{d y_\\star}{dt}\\right\\|^2,\\qquad\\epsilon_{\\rm lag}\\gtrsim\\frac{\\kappa\\tau_yv^2}{\\sigma}.$$",
  "$$z_{k+1}\\sim P_\\theta(\\cdot\\mid z_k,x_k),\\qquad\\mathcal L_{\\rm total}=\\mathcal L_{\\rm task}+\\lambda_{\\rm budget}\\left[\\frac{\\kappa\\widehat\\tau_y\\widehat v_k^2}{\\widehat\\sigma_k+\\epsilon}-\\widehat\\epsilon_{\\rm lag,k}\\right]_+^2+\\lambda_\\sigma\\widehat\\sigma_k.$$"
 ],
 "id": 2710,
 "implementation": "1. Integration point: implement a stochastic continuous-time-inspired RNN or SSM. For the cheapest test, use a finite-state hidden variable with positive transition rates $q_{ij}=\\operatorname{softplus}(a_{ij})$ and simulate transitions by uniformization; decode $y_k=g_\\theta(z_k)$. 2. Pseudocode: simulate trajectories; accumulate transition counts $N_{ij}$ and state occupancies $n_i$; estimate $\\widehat p_i=n_i/\\sum_jn_j$ and $\\widehat q_{ij}=N_{ij}/(n_i\\Delta t)$; compute $\\widehat\\sigma=\\frac12\\sum_{ij}(\\widehat p_i\\widehat q_{ij}-\\widehat p_j\\widehat q_{ji})\\log[(\\widehat p_i\\widehat q_{ij}+\\epsilon)/(\\widehat p_j\\widehat q_{ji}+\\epsilon)]$. Estimate $\\widehat C_y(s)$ from constant-input rollouts and integrate it to obtain $\\widehat\\tau_y$. Estimate $\\widehat v_k=|y_{\\star,k+1}-y_{\\star,k}|/\\Delta t$ and subtract the constant-input representation MSE from total MSE to obtain $\\widehat\\epsilon_{\\rm lag}$. Add the displayed hinge penalty, or use it as a controller that increases the global rate multiplier when the bound is violated. 3. Computed versus estimated: the entropy-production expression and speed–accuracy relation come from the paper; probabilities, rates, correlation time, representation error, and prefactor $\\kappa$ are estimated empirically. Initially stop gradients through rate and correlation estimates. 4. First experiment: train a 16-state stochastic RNN on sine-wave tracking and randomly pulsed input sequences, comparing a GRU, an unconstrained stochastic RNN, and the proposed regularized model. Sweep a global rate multiplier $a$ with $q_{ij}\\leftarrow aq_{ij}$ and sweep input frequency. The prediction is that at fixed frequency, lag MSE versus measured entropy production has log-log slope near $-1$ until representation error dominates; at fixed dissipation and memory, lag MSE versus squared target velocity has slope near $+1$. The acceptable-tracking boundary should occur near $\\widehat\\sigma\\approx\\kappa\\widehat\\tau_y\\widehat v^2/\\widehat\\epsilon_{\\rm target}$, within a factor of two. Increasing a slow mode should increase $\\widehat\\tau_y$, reduce required dissipation at low frequencies, and worsen high-frequency tracking.",
 "math_summary": "Let $x_t$ be the input, $z_t$ the stochastic hidden state, and $y_t=g(z_t)$ the output observable. For a continuous-time Markov process with probability $p_i(t)$ and transition rates $q_{ij}(t)$, the entropy-production rate is $\\sigma(t)=\\frac{1}{2}\\sum_{i,j}[p_iq_{ij}-p_jq_{ji}]\\log[(p_iq_{ij})/(p_jq_{ji})]$, equivalently $\\sum_r(j_r^+-j_r^-)\\log(j_r^+/j_r^-)$ for reaction channels. Define the normalized equilibrium output autocorrelation $C_y(s)=\\operatorname{Cov}_{\\mathrm{eq}}[y(t),y(t+s)]/\\operatorname{Var}_{\\mathrm{eq}}[y]$ and memory time $\\tau_y=\\int_0^\\infty C_y(s)\\,ds$. If the desired output has local velocity $v(t)=d y_\\star(t)/dt$, the paper's speed–accuracy mechanism gives a lag-error lower bound of the dimensionally normalized form $\\epsilon_{\\mathrm{lag}}\\sigma\\gtrsim\\kappa v^2\\tau_y$, where $\\epsilon_{\\mathrm{lag}}$ is mean-square error after subtracting static representation error and $\\kappa$ is an order-one convention-dependent constant. Thus faster input variation requires more dissipation or a different memory spectrum.",
 "math_tags": [
  "statistical-mechanics",
  "stochastic-processes",
  "dynamical-systems",
  "information-theory",
  "control-theory"
 ],
 "ml_areas": [
  "rnn",
  "ssm",
  "world-model",
  "training-dynamics",
  "regularization"
 ],
 "paper": {
  "arxiv_id": "2608.23764",
  "arxiv_url": "https://arxiv.org/abs/2608.23764",
  "summary_what_math_gives_to_ml": "The paper provides a transferable thermodynamic speed–accuracy mechanism: in a stochastic Markov computation, lag error cannot be reduced arbitrarily unless the network either dissipates more entropy or retains information in slowly relaxing modes. The relevant quantities are entropy-production rate, integrated output autocorrelation time, and the rate at which the desired output changes. A neural analogue is a stochastic recurrent or state-space model with an explicit dissipation estimate and memory-time monitor. The resulting regularizer predicts a measurable accuracy–energy–memory tradeoff rather than merely encouraging generic smoothness.",
  "title": "Entropy Production Bounds the Accuracy of Computation in Markov Networks",
  "year": "2026"
 },
 "ratings": {
  "difficulty": 6,
  "novelty": 8,
  "usefulness": 8
 },
 "solves": [
  "stability",
  "accuracy",
  "memory"
 ],
 "title": "Dissipation–Memory Budget for Stochastic RNNs",
 "url": "https://synthcore.org/idea/2710/dissipation-memory-budget-for-stochastic-rnns",
 "verification": {
  "peer_reviewed": false,
  "stage1_mechanism_check": {
   "worked": false,
   "confidence": 9,
   "verdict": "Built a three-state nonequilibrium stochastic Markov hidden-state MVP with entropy-production and autocorrelation-time estimators, plus activity/frequency sweeps and a budget-based rate controller. Two quantitative predictions were confirmed: entropy production scales as activity^1.000 (predicted +1) and memory time as activity^-1.000 (predicted -1); Monte Carlo estimates matched exact sigma=0.5493 and tau=0.6667 within about 5%. However, the central kappa=1 budget inequality was not supported: epsilon*sigma/(tau*v^2) ranged from 0.054 to 0.819, and the controller helped only at high frequency while worsening low-frequency tracking, so the full proposed mechanism is not validated.",
   "metrics": {
    "baseline": "Fixed rate=1 tracking MSE at frequencies 0.03, 0.1, 0.3: 0.01575, 0.14104, 0.38923",
    "idea": "Budget controller with predeclared kappa=0.5 selected rates 0.5, 1, 4 and produced MSE 0.05818, 0.14104, 0.08954; budget ratio min/median/max was 0.054/0.768/0.819; observed activity exponents were sigma +1.000 and tau -1.000"
   },
   "how_to_run": "python3 run_experiment.py \u0026\u0026 /home/maxwelhelp/main/bin/python3 analyze_results.py",
   "files": [
    "run_experiment.py",
    "analyze_results.py",
    "results.json",
    "run_output.txt",
    "diagnostics.txt"
   ],
   "limitations": "No trained GRU or neural stochastic RNN was tested; the tracking experiment uses an analytically simulated first-order decoder rather than learned finite-state transition parameters. The budget prefactor kappa was not fitted, and only a small three-state cycle, sinusoidal targets, finite-duration Monte Carlo estimates, and a limited rate/frequency grid were evaluated."
  },
  "status": "mechanism_failed",
  "status_label": "Mechanism failed",
  "updated_at": "2026-09-01T18:30:42",
  "verdict_source": "deterministic test code (paired-seed permutation statistics)"
 }
}