{
 "artifacts": null,
 "category": "architecture",
 "description": "Represent a neural density or feature field by two positive reciprocal branches whose product is the modeled density, analogous to the forward and backward fields in the paper. Add stochastic branching perturbations to the two branches and train their cross-covariance so that the diagonal paired density matches the target while off-diagonal correlations remain finite-range. This creates a structured alternative to an unconstrained single-field uncertainty representation.",
 "formulas_latex": [
  "$$\\rho_\\theta(x)=\\phi_F(x)\\phi_B(x),\\qquad \\phi_F(x)\u003e0,\\ \\phi_B(x)\u003e0$$",
  "$$\\partial_t\\phi_F=\\mathcal F[\\rho]\\phi_F,\\qquad \\partial_t\\phi_B=-\\mathcal F[\\rho]\\phi_B,\\qquad \\mathcal F[\\rho]=D\\nabla^2+\\frac{V+2Q[\\rho]}{\\hbar}$$",
  "$$Q[\\rho]= -\\frac{\\hbar^2}{2m}\\frac{\\nabla^2\\sqrt{\\rho}}{\\sqrt{\\rho}},\\qquad D=\\frac{\\hbar}{2m}$$",
  "$$C_{FB}(x,y)=\\mathbb E_r[\\psi_F^r(x)\\psi_B^r(y)],\\qquad \\rho_{\\mathrm{pair}}(x)=-C_{FB}(x,x)$$"
 ],
 "id": 2850,
 "implementation": "1. Integration point: use this as a two-branch output head for a diffusion model, energy-based model, or latent-density estimator on a low-dimensional grid or learned token coordinate. Replace a single positive score or intensity head by two heads \\(f_\\theta\\) and \\(b_\\theta\\), set \\(\\phi_F=\\operatorname{softplus}(f_\\theta)+10^{-5}\\) and \\(\\phi_B=\\operatorname{softplus}(b_\\theta)+10^{-5}\\), and normalize their product over spatial or token positions. 2. Pseudocode: sample \\(K\\) dropout, Gaussian, or bootstrap replicas; compute \\(\\rho^r=\\phi_F^r\\phi_B^r\\); compute \\(\\bar\\rho=K^{-1}\\sum_r\\rho^r\\), \\(\\psi_F^r=\\phi_F^r-\\bar\\phi_F\\), and \\(\\psi_B^r=\\phi_B^r-\\bar\\phi_B\\). Minimize task loss plus \\(\\lambda_\\rho\\|\\bar\\rho-\\rho_{target}\\|_2^2+\\lambda_c\\sum_x[\\max(0,C_{FB}(x,x))]^2\\), where the second term encourages anticorrelation, and add \\(\\lambda_{prod}\\|\\log\\bar\\rho-\\log(\\bar\\phi_F\\bar\\phi_B)\\|_2^2\\) for reciprocal consistency. For continuous grids, optionally add \\(\\lambda_{smooth}(\\|\\nabla\\log\\phi_F\\|^2+\\|\\nabla\\log\\phi_B\\|^2)\\). 3. Computed from the paper: the positive reciprocal product, centered cross-kernel, and negative diagonal paired-density construction. Estimated empirically: replica expectations, target density, and all loss weights. Do not initially implement the Bohm term, because estimating second derivatives of \\(\\sqrt\\rho\\) may destabilize training. 4. First cheap experiment: train a two-head MLP density estimator on two-dimensional eight-Gaussian data and compare it with a single-head positive MLP and a standard ensemble. Measure held-out NLL, calibration, and \\(C_{FB}(x,y)\\). The quantitative prediction is that the paired model can retain a broad mean density while its cross-correlation becomes negative near the diagonal; fitting \\(|C_{FB}(x,y)|\\) versus distance should reveal a finite correlation length rather than the domain-scale correlations of an unconstrained ensemble. On a controllable synthetic target, increasing \\(\\lambda_c\\) should monotonically reduce diagonal cross-covariance without substantially changing \\(\\bar\\rho\\).",
 "math_summary": "Let \\(\\phi_F(x,t)\u003e0\\) and \\(\\phi_B(x,t)\u003e0\\) be forward and backward neural fields, with modeled density \\(\\rho_\\theta(x,t)=\\phi_F(x,t)\\phi_B(x,t)\\). The paper's reciprocal dynamics are \\(\\partial_t\\phi_F=\\mathcal F[\\rho]\\phi_F\\) and \\(\\partial_t\\phi_B=-\\mathcal F[\\rho]\\phi_B\\), where \\(\\mathcal F[\\rho]=D\\nabla^2+(V+2Q[\\rho])/\\hbar\\), \\(D=\\hbar/(2m)\\), and \\(Q[\\rho]=-(\\hbar^2/(2m))\\nabla^2\\sqrt{\\rho}/\\sqrt{\\rho}\\). In the neural adaptation, \\(\\phi_F=\\operatorname{softplus}(f_\\theta)+\\epsilon\\) and \\(\\phi_B=\\operatorname{softplus}(b_\\theta)+\\epsilon\\), with \\(f_\\theta,b_\\theta\\) outputs of two coupled networks. For stochastic replicas \\(r\\), define centered fluctuations \\(\\psi_F^r(x)=\\phi_F^r(x)-\\bar\\phi_F(x)\\), \\(\\psi_B^r(x)=\\phi_B^r(x)-\\bar\\phi_B(x)\\), and cross-kernel \\(C_{FB}(x,y)=\\mathbb E_r[\\psi_F^r(x)\\psi_B^r(y)]\\). Use the anticorrelated paired density \\(\\rho_{\\mathrm{pair}}(x)=-C_{FB}(x,x)\\), after parameterizing the cross-covariance with a negative sign. The working property is that the mean/product sector can stay broad while the connected sector is independently controlled.",
 "math_tags": [
  "pde",
  "stochastic-processes",
  "probability",
  "dynamical-systems"
 ],
 "ml_areas": [
  "diffusion",
  "regularization",
  "architecture"
 ],
 "paper": {
  "arxiv_id": "2608.29807",
  "arxiv_url": "https://arxiv.org/abs/2608.29807",
  "summary_what_math_gives_to_ml": "The paper offers a nonstandard reciprocal-field construction in which two positive diffusion-reaction fields have a product equal to a target density, while branching fluctuations generate a connected correlation kernel. Its most transferable mechanism is the separation between an extended mean profile and a finite-range fluctuation structure, quantified by the screening length \\(\\xi=\\sqrt{D_{\\mathrm{eff}}/\\mu}\\). A neural implementation could represent uncertainty or latent density with forward and backward positive fields, and regulate their cross-covariance rather than only matching marginal activations. The strongest falsifiable prediction is that the learned off-diagonal correlation decays exponentially over a controllable length, while the diagonal paired density remains matched to the target distribution.",
  "title": "Branching stochastic mechanics. I. Clustering and connected correlations within a branching-process representation of the Schrödinger equation",
  "year": "2026"
 },
 "ratings": {
  "difficulty": 6,
  "novelty": 7,
  "usefulness": 6
 },
 "solves": [
  "stability",
  "generalization",
  "accuracy"
 ],
 "title": "Reciprocal Paired-Density Network",
 "url": "https://synthcore.org/idea/2850/reciprocal-paired-density-network",
 "verification": {
  "peer_reviewed": false,
  "status": "unverified",
  "status_label": "Unverified",
  "verdict_source": "deterministic test code (paired-seed permutation statistics)"
 }
}