{
 "artifacts": null,
 "category": "dynamics",
 "description": "Partition a network's parameters into M ordered blocks and represent blockwise normalized update activity by a nonnegative density n_i. Instead of assigning independent learning rates, evolve this density through a discrete conservative current whose diffusivity depends on local activity, while adding calibrated multiplicative noise from the corresponding mobility. This couples learning-rate adaptation across depth or layer order and prevents isolated blocks from becoming arbitrarily overactive.",
 "formulas_latex": [
  "$$\\partial_\\tau n(x,\\tau)=-\\partial_x J(x,\\tau),\\qquad J(x,\\tau)=-D(n)\\,\\partial_x n+\\sqrt{\\sigma(n)}\\,\\xi(x,\\tau),\\qquad \\langle\\xi(x,\\tau)\\xi(x',\\tau')\\rangle=\\delta(x-x')\\delta(\\tau-\\tau').$$",
  "$$J_{i+1/2}=-D(\\bar n_{i+1/2})(n_{i+1}-n_i)+\\sqrt{\\sigma(\\bar n_{i+1/2})}\\,z_{i+1/2},\\qquad n_i^{t+1}=n_i^t-\\alpha(J_{i+1/2}-J_{i-1/2}),\\qquad \\bar n_{i+1/2}=\\frac{n_i+n_{i+1}}{2}.$$",
  "$$\\delta n_k^{t+1}\\approx\\left(1-\\alpha D(\\bar n)\\lambda_k\\right)\\delta n_k^t,\\qquad 0\u003c\\alpha D(\\bar n)\\lambda_{\\max}\u003c2.$$",
  "$$\\eta(\\ell)\\sim\\Delta(\\ell)\\sim e^{(2-d)\\ell};\\qquad d=1\\Rightarrow\\eta(\\ell),\\Delta(\\ell)\\text{ grow under coarse-graining}.$$"
 ],
 "id": 2964,
 "implementation": "1. Integration point: insert the module between gradient computation and the base parameter update. Divide parameters into ordered blocks, preferably one block per Transformer layer or recurrent-state group. Compute each block's raw update u_i = −g_i, and define activity a_i = ||u_i||₂/(||θ_i||₂+ε). Set n_i = M(a_i+ε_n)/(Σ_j(a_j+ε_n)), so total density is M. Use the transported density to scale the actual update: θ_i ← θ_i + η₀(n_i/n̄)u_i. 2. Pseudocode: estimate D_i = D₀(1+βn_i), or fit D(n) = softplus(d₀+d₁n); estimate mobility from the exponential moving variance of a_i, using σ_i = c₀+c₁ EMA[(a_i−EMA(a_i))²]. Form each interface current using the stated formula, set boundary currents to zero, update n, project to n_i ≥ 0, and renormalize its sum to M. Add Gaussian interface noise only after multiplying by the mobility and scale it by a separately tunable noise coefficient. 3. Computed quantities are block norms, moving variances, currents, and the path-Laplacian eigenvalue λ_max = 4 sin²(π/(2M)); D and σ are initially hyperparameters, then optionally fitted by regressing observed activity drift and variance against local density. 4. First experiment: train a small Transformer or MLP on a fixed data order with AdamW, cosine decay, and equal wall-clock and update budgets. Compare constant D, nonlinear D(n) without noise, and nonlinear D(n) plus mobility noise. Use M = 12 layer blocks and sweep the hydrodynamic step α. 5. Preregistered signature: after perturbing one block's n_i, the discrete Fourier amplitude at mode k must decay exponentially with slope D(n̄)λ_k before nonlinear saturation, and the divergence boundary must occur near α_c = 2/[D(n̄)λ_max]. Reject the transfer if fitted decay slopes differ by more than 30 percent from prediction, if no finite stability boundary is observed, or if the nonlinear-plus-noise variant fails to reduce block-activity variance by at least 10 percent at matched training loss.",
 "math_summary": "The paper's effective mechanism is a conserved fluctuating hydrodynamic field: ∂τ n(x,τ) = −∂x J(x,τ), with current J = −D(n)∂x n + √σ(n) ξ. Here n is density, D(n) \u003e 0 is density-dependent diffusivity, σ(n) ≥ 0 is mobility, and ξ is unit space-time white noise. For a neural optimizer, let n_i ≥ 0 be normalized update allocation in parameter block i, with Σ_i n_i = M. Define the interface current J_{i+1/2} = −D((n_i+n_{i+1})/2)(n_{i+1}−n_i) + √σ((n_i+n_{i+1})/2) z_{i+1/2}, where z is Gaussian noise. The conservative update n_i^{t+1} = n_i^t − α(J_{i+1/2}−J_{i−1/2}) preserves total allocation. Linearization around a uniform state n̄ gives Fourier-mode decay rate D(n̄)k²; explicit Euler is stable when αD(n̄)λ_max(L) \u003c 2, where L is the path-graph Laplacian. The paper's one-dimensional RG result η(ℓ) ∼ Δ(ℓ) ∼ exp((2−d)ℓ) motivates making nonlinear transport and noise active rather than asymptotically irrelevant in a one-dimensional block chain.",
 "math_tags": [
  "dynamical-systems",
  "stochastic-processes",
  "statistical-mechanics",
  "renormalization",
  "pde",
  "numerical-analysis"
 ],
 "ml_areas": [
  "optimizer",
  "training-dynamics",
  "scheduler"
 ],
 "paper": {
  "arxiv_id": "2609.00159",
  "arxiv_url": "https://arxiv.org/abs/2609.00159",
  "summary_what_math_gives_to_ml": "The paper derives a coarse-grained fluctuating hydrodynamics in which a conserved density evolves through a current with density-dependent diffusivity and multiplicative noise, rather than through a linear Gaussian diffusion equation. Its transferable asset is the separation between deterministic transport D(n), fluctuation mobility σ(n), and a conservation law, together with a renormalization-group prediction that interaction and noise couplings become relevant in one dimension. A concrete neural-network transfer is a parameter-block optimizer that treats normalized update activity as a conserved density, transports it across blocks with nonlinear diffusion, and injects noise proportional to the estimated mobility. The scheme makes a sharp prediction: long-wavelength activity modes should decay at rate D(n̄)k², with an explicit Euler stability ceiling determined by the largest discrete hydrodynamic eigenvalue.",
  "title": "Nonlinear Fluctuating Hydrodynamics from Interacting Noisy Quantum Matter",
  "year": "2026"
 },
 "ratings": {
  "difficulty": 5,
  "novelty": 8,
  "usefulness": 7
 },
 "solves": [
  "stability",
  "generalization",
  "speedup"
 ],
 "title": "Nonlinear Hydrodynamic Optimizer",
 "url": "https://synthcore.org/idea/2964/nonlinear-hydrodynamic-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
   }
  }
 }
}