{
 "artifacts": null,
 "category": "dynamics",
 "description": "Replace constant friction and optimizer noise with a velocity-dependent friction gamma(u) and noise amplitude tied by a fluctuation-dissipation relation. High-speed momentum states can be damped and randomized differently from low-speed states, creating controlled transient exploration while preserving a known equilibrium momentum distribution.",
 "formulas_latex": [
  "$$\\partial_t p(u,t)=\\partial_u\\left[\\gamma(u)u\\,p(u,t)+T\\gamma(u)\\partial_u p(u,t)\\right],\\qquad \\gamma(u)\u003e0.$$",
  "$$B(v)=\\lambda(v)k_B T,\\qquad \\lambda(v)=m\\gamma(v).$$",
  "$$du=\\left[-\\gamma(u)u+T\\gamma'(u)\\right]dt+\\sqrt{2T\\gamma(u)}\\,dW_t,\\qquad p_*(u)=\\frac{1}{\\sqrt{2\\pi T}}\\exp\\left(-\\frac{u^2}{2T}\\right).$$",
  "$$\\theta_{k+1}=\\theta_k+h u_k,\\quad u_{k+1}=u_k+h\\left[-\\gamma(u_k)u_k-\\nabla_\\theta L(\\theta_k)+T\\gamma'(u_k)\\right]+\\sqrt{2T\\gamma(u_k)h}\\,\\xi_k,\\quad \\xi_k\\sim\\mathcal N(0,I).$$"
 ],
 "id": 2752,
 "implementation": "Integrate this as a replacement for the momentum state in SGD, initially without Adam's adaptive preconditioner. Maintain parameters theta and a momentum vector u. Choose a positive componentwise friction such as gamma(u) = gamma_0(1 + a u^2), with gamma_0 \u003e 0 and a \u003e= 0. At each minibatch, compute g = grad_theta L(theta), evaluate gamma(u), evaluate gamma'(u) = 2 gamma_0 a u, draw independent standard-normal noise xi, and apply the displayed Euler-Maruyama update before theta \u003c- theta + h u. Use the same temperature T and step size h for all comparisons. The paper supplies the Fokker-Planck equation and friction-diffusion matching rule; gamma_0, a, T, and h are engineering hyperparameters. The first cheap experiment should use a two-layer MLP on MNIST, comparing ordinary momentum SGD, constant-friction Langevin momentum, and the nonlinear matched optimizer at equal minibatch size and compute. First run with gradients set to zero and measure the momentum histogram, variance, autocorrelation, displacement mean-squared growth, and excess kurtosis. The quantitative prediction is that matched dynamics converge to variance T and approximately Gaussian stationary momentum, while a control with the same nonlinear friction but constant noise has a variance and kurtosis that depend on a. For parameter trajectories, the prediction is finite long-time diffusion with non-Gaussian intermediate displacement and excess kurtosis tending toward zero at long times. Sweep h and locate the divergence boundary; the measured stable region should shrink when the maximum effective friction or noise amplitude makes h gamma(u) large. Training should then test whether the transient heavy-tailed displacement improves escape from sharp minibatch-loss basins without violating the equilibrium diagnostics.",
 "math_summary": "After setting the mass to one and defining T as the velocity-temperature scale, the paper uses the Fokker-Planck equation partial_t p = partial_u[gamma(u) u p + T gamma(u) partial_u p], where u is velocity, gamma(u) \u003e 0 is friction, T \u003e 0 is temperature, and p(u,t) is the velocity density. The paper's relation B(v) = lambda(v) k_B T requires diffusion strength to be proportional to friction. The zero-flux stationary solution is p_star(u) proportional to exp(-u^2/(2T)), independent of the nonlinear friction law. The equivalent Ito stochastic differential equation for this divergence-form operator is du = [-gamma(u)u + T gamma'(u)]dt + sqrt(2T gamma(u))dW, where W is standard Brownian motion and gamma' is the derivative of gamma. The derivative term is required for the stated equilibrium law.",
 "math_tags": [
  "stochastic-processes",
  "statistics",
  "dynamical-systems",
  "control-theory"
 ],
 "ml_areas": [
  "optimizer",
  "training-dynamics",
  "regularization"
 ],
 "paper": {
  "arxiv_id": "2608.26773",
  "arxiv_url": "https://arxiv.org/abs/2608.26773",
  "summary_what_math_gives_to_ml": "The paper provides a constructive fluctuation-dissipation mechanism for Brownian yet non-Gaussian motion: velocity-dependent friction must be paired with velocity-dependent diffusion according to B(v) = lambda(v) k_B T. This preserves a Gaussian equilibrium velocity distribution while producing non-Gaussian displacement statistics at intermediate times and ordinary diffusion asymptotically. A transferable neural-network construction is a momentum optimizer with nonlinear, state-dependent friction and matched noise, including the drift correction required by the Fokker-Planck operator. The key falsifiable prediction is that matched dynamics preserve the target stationary momentum law, while mismatched friction and noise produce a measurable variance and kurtosis error.",
  "title": "Brownian yet non-Gaussian diffusion through equilibrium nonlinear friction",
  "year": "2026"
 },
 "ratings": {
  "difficulty": 5,
  "novelty": 7,
  "usefulness": 6
 },
 "solves": [
  "stability",
  "generalization"
 ],
 "title": "Equilibrium-Matched Nonlinear Momentum Optimizer",
 "url": "https://synthcore.org/idea/2752/equilibrium-matched-nonlinear-momentum-optimizer",
 "verification": {
  "peer_reviewed": false,
  "status": "unverified",
  "status_label": "Unverified",
  "verdict_source": "deterministic test code (paired-seed permutation statistics)"
 }
}