{
 "artifacts": null,
 "category": "training",
 "description": "Use the paper's quotient-space and Euler-characteristic logic as a runtime invariant for adaptive neural discretizations. A refinement controller should add capacity only when the residual warrants it and should reject refinements that create new nullspace modes, disconnected constant modes, or closed non-exact feature modes.",
 "formulas_latex": [
  "$$\\sum_{k=0}^{2}(-1)^k\\dim\\bigl(H^k(\\mathbb{P}^{\\bullet})\\bigr)=\\sum_{k=0}^{2}(-1)^k\\dim(\\mathbb{P}^k).$$",
  "$$\\dim H^1=1-n^0+n^1-n^2.$$",
  "$$\\ker(D_1)=\\operatorname{im}(D_0),\\qquad D_1D_0=0.$$"
 ],
 "id": 3109,
 "implementation": "Use this as a controller around any adaptive neural field discretization, rather than as a conventional differentiable regularizer. Maintain sparse matrices D0 and D1 and basis counts n0, n1, and n2 for the currently active hierarchy. The network predicts coefficient vectors or residual corrections in these spaces. Generate a cellwise refinement score e_K = ||r_K||_2^2 plus an optional feature-gradient score ||D0 x0|| on scalar channels. Mark the top 10 percent of cells, apply a precomputed local B-spline refinement or prolongation operator P, update coefficients xk \u003c- Pk xk, and construct new derivative matrices using the compatible relation Dk_new Pk = P{k+1} Dk_old. Before accepting refinement, check three invariants: the Frobenius norm ||D1_new D0_new||, the Euler count 1 - n0_new + n1_new - n2_new, and numerical rank consistency rank(D1_new) + rank(D0_new) = n1_new - dim H1_target, with target dim H1 = 0 for a disk. Estimate ranks with sparse QR or SVD using tolerance 1e-7. If any check fails, undo the refinement or refine the coupled 0-, 1-, and 2-form neighborhoods together. The paper supplies the algebraic acceptance criteria; residual scores, tolerances, and maximum hierarchy depth are empirical. The first experiment should use a small PINN or neural operator for a Poisson equation with a polar source, comparing residual-driven refinement with parameter-count-matched uniform refinement. Measure validation error versus active basis count, rejected refinements, and forbidden cohomological residual. The expected signal is improved error per parameter and no increase in pole-localized oscillations or spurious nullspace modes.",
 "math_summary": "For a finite complex P^0 -\u003e P^1 -\u003e P^2, the paper uses the Euler characteristic identity sum_{k=0}^2 (-1)^k dim H^k = sum_{k=0}^2 (-1)^k dim P^k. Since H^0 is one-dimensional and H^2 is zero on the disk, dim H^1 = 1 - n^0 + n^1 - n^2. In the intended polar complex this equals zero, so refinement should preserve 1 - n^0 + n^1 - n^2 = 0 and preserve rank(D0), rank(D1) with ker(D1) = im(D0). Here H^k denotes kth cohomology, n^k = dim P^k, and D0 and D1 are consecutive discrete derivative matrices. The ML adaptation uses these as computable algebraic checks rather than as a training-loss approximation.",
 "math_tags": [
  "topology",
  "linear-algebra",
  "pde",
  "approximation-theory"
 ],
 "ml_areas": [
  "training",
  "regularization",
  "world-model"
 ],
 "paper": {
  "arxiv_id": "2609.03461",
  "arxiv_url": "https://arxiv.org/abs/2609.03461",
  "summary_what_math_gives_to_ml": "The paper provides a constructive hierarchical spline complex for disk-like domains whose parameterization collapses an entire boundary edge to a pole. Its transferable asset is not ordinary B-spline approximation, but the combination of local adaptive refinement with pole-compatible basis constraints and an exact discrete de Rham sequence, so differentiation, curl, and divergence remain structurally consistent after refinement. This suggests a neural operator or physics-informed network whose spatial features are hierarchical polar-spline coefficients, with refinement driven by residuals while preserving the algebraic identity that the discrete divergence of a discrete curl is exactly zero.",
  "title": "Adaptively-refinable polar-spline discrete differential forms: hierarchical construction, exactness, and applications",
  "year": "2026"
 },
 "ratings": {
  "difficulty": 5,
  "novelty": 8,
  "usefulness": 5
 },
 "solves": [
  "accuracy",
  "stability",
  "scalability"
 ],
 "title": "Cohomology-safe adaptive residual refinement",
 "url": "https://synthcore.org/idea/3109/cohomology-safe-adaptive-residual-refinement",
 "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
   }
  }
 }
}