{
 "artifacts": null,
 "category": "architecture",
 "description": "Replace a dense weak pairwise interaction over tokens or graph nodes by an analytically resummed Gaussian quadratic term plus explicitly evaluated corrections containing only subsets of size at most R. Choose R from the target approximation error and interaction scale instead of fixing the interaction order heuristically. This is intended for dense global attention, graph interactions, or latent-variable energy models with individually small couplings.",
 "formulas_latex": [
  "$$Z(J)=2^{n}\\det(I-J)^{-1/2}\\left(1+O_{\\beta,\\kappa}(1/n)\\right),\\qquad \\max_{i,j}|J_{ij}|\\leq\\frac{\\beta}{n},\\quad \\lambda_{\\max}(J)\\leq1-\\kappa.$$",
  "$$\\left|\\mathrm{Tail}_{\u003eR}\\right|\\leq\\left(\\frac{CR}{n}\\right)^{R/2}+e^{-cn}.$$",
  "$$R\\log\\!\\left(\\frac{en}{R}\\right)=O\\!\\left(\\log n+\\log\\frac{1}{\\varepsilon}\\right).$$",
  "$$\\widehat{\\log Z}(X)=-\\frac12\\log\\det(I-J(X))+\\sum_{\\substack{S\\subseteq[n]\\\\1\\leq |S|\\leq R}}c_S(X,J),\\qquad \\left|\\log Z-\\widehat{\\log Z}\\right|\\lesssim\\left(\\frac{CR}{n}\\right)^{R/2}+e^{-cn}.$$"
 ],
 "id": 3163,
 "implementation": "(1) Integration point: insert the block after token or graph embeddings and before the next residual layer. Given X in R^{n x d}, compute q_i=W_q x_i and define J_ij=(beta/n)tanh(q_i^T q_j/sqrt(d)) for i!=j and J_ii=0; symmetrize J as (J+J^T)/2. Enforce lambda_max(J)\u003c=1-kappa with 3-5 power-iteration steps and the rescaling J \u003c- J*min(1,(1-kappa)/(lambda_max(J)+1e-6)). (2) Pseudocode: compute Q=I-J; obtain base=-0.5*logdet(Q) using Cholesky for n\u003c=256 or stochastic Lanczos for larger n. For every subset S with 1\u003c=|S|\u003c=R, evaluate c_S using Gauss-Hermite quadrature over z_S under the Gaussian covariance Q^{-1}; for each scalar residual potential r_i(z_i), subtract its degree-0, degree-1, and degree-2 Taylor terms before multiplying residual factors, ensuring the correction starts at degree three. Sum base and all c_S, then use the result as a global residual bias or interaction feature. (3) Use the paper's tail expression to choose the smallest R satisfying (C*R/n)^(R/2)\u003cepsilon/2. Estimate C from the maximum normalized degree-three residual on a calibration batch; estimate the e^{-cn} contribution by comparing outputs at R and R+1. For large n, replace exact subset enumeration by uniform or magnitude-weighted Monte Carlo subset sampling with unbiased reweighting. (4) First experiment: train a small transformer with sequence lengths 32, 64, and 128 on a synthetic task whose target is a non-Gaussian Gaussian-coupled partition function, where brute-force quadrature is available for n\u003c=12. Compare dense softmax attention, linear attention, and this block at matched hidden size and wall-clock budget. Measure partition-function relative error, training loss, gradient variance, peak memory, and throughput. A successful result is below 1% approximation error at n=32, lower gradient variance near the coupling boundary, and faster loss reduction or lower memory than dense attention at comparable accuracy.",
 "math_summary": "The paper considers Gaussian integrals with a possibly indefinite moment matrix J whose entries satisfy |J_ij|=O(1/n), performs linear cancellation and exact quadratic resummation, and makes the remaining coordinate factor vanish to order at least three. Its truncation guarantee is |Tail_{\u003eR}| \u003c= (CR/n)^{R/2}+e^{-cn}, where n is the number of interacting coordinates, R is the largest retained support size, C\u003e0 depends on bounded scalar factors and the weak-coupling regime, and c\u003e0 is a large-deviation rate. It also proves Z(J)=2^n det(I-J)^(-1/2)(1+O_{beta,kappa}(1/n)) for zero-diagonal symmetric J with max_{i,j}|J_ij|\u003c=beta/n and lambda_max(J)\u003c=1-kappa. In the neural adaptation, X is the input feature matrix, J(X) is a learned symmetric zero-diagonal coupling matrix, z is a scalar latent vector, and the quadratic Gaussian base is -1/2 log det(I-J). The non-Gaussian residual corrections c_S are indexed by subsets S of coordinates and are retained only for 1\u003c=|S|\u003c=R. The order-three cancellation is what yields the faster-than-geometric tail rather than a generic first-order truncation bound.",
 "math_tags": [
  "probability",
  "combinatorics",
  "approximation-theory",
  "numerical-analysis"
 ],
 "ml_areas": [
  "attention",
  "graph-nn",
  "world-model",
  "loss"
 ],
 "paper": {
  "arxiv_id": "2609.04079",
  "arxiv_url": "https://arxiv.org/abs/2609.04079",
  "summary_what_math_gives_to_ml": "The paper's transferable asset is a deterministic truncation principle for weak, dense interactions: exact quadratic resummation removes all second-order effects, leaving a remainder whose first nonzero terms have degree at least three. The resulting tail is bounded by a rapidly decaying term, $(CR/n)^{R/2}$, plus an exponentially small large-support term, while the number of retained subsets is controlled by $R\\log(en/R)$. This suggests a weak-interaction neural layer in which a dense Gaussian or energy-based interaction is analytically resummed and only low-order residual interaction subsets are explicitly evaluated.",
  "title": "Diffuse Gaussian Truncation For Deterministic Approximate Counting",
  "year": "2026"
 },
 "ratings": {
  "difficulty": 7,
  "novelty": 8,
  "usefulness": 5
 },
 "solves": [
  "scalability",
  "speedup",
  "stability"
 ],
 "title": "Gaussian-resummed sparse interaction layer",
 "url": "https://synthcore.org/idea/3163/gaussian-resummed-sparse-interaction-layer",
 "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
   }
  }
 }
}