{
 "artifacts": null,
 "category": "architecture",
 "description": "Replace or augment Fourier features and ordinary coordinate embeddings with a sum of quadratic B-spline grids at explicitly selected resolutions. The finest grid controls the smallest representable feature, while C1 continuity makes spatial derivatives smooth and avoids the high-frequency optimization pathologies of sharp or discontinuous encodings. This is especially suitable for neural fields, implicit scene representations, PDE surrogates, diffusion score fields, and networks trained with derivative losses.",
 "formulas_latex": [
  "$$B_2(t)=\\begin{cases}\\frac{3}{4}-t^2,\u0026|t|\\le\\frac{1}{2},\\\\[2pt]\\frac{1}{2}\\left(\\frac{3}{2}-|t|\\right)^2,\u0026\\frac{1}{2}\u003c|t|\\le\\frac{3}{2},\\\\[2pt]0,\u0026|t|\u003e\\frac{3}{2},\\end{cases}$$",
  "$$S_r(\\mathbf{x})=\\sum_{\\mathbf{k}\\in\\mathbb{Z}^d}A_{r,\\mathbf{k}}\\prod_{j=1}^{d}B_2\\!\\left(\\frac{x_j}{h_r}-k_j\\right),\\qquad S(\\mathbf{x})=\\sum_{r=0}^{R}\\alpha_rS_r(\\mathbf{x}).$$",
  "$$f_{\\theta}(\\mathbf{x})=\\operatorname{MLP}_{\\theta}\\!\\left([\\mathbf{x},S_0(\\mathbf{x}),\\ldots,S_R(\\mathbf{x})]\\right),\\qquad h_R\\le\\text{target feature width}.$$"
 ],
 "id": 124,
 "implementation": "(1) Integration point: modify the coordinate-input stage of a coordinate MLP, neural field, PDE surrogate, or diffusion spatial conditioner. Normalize each coordinate x to [-1,1]^d and concatenate outputs from R+1 trainable quadratic B-spline grids before the first linear layer. Choose grid spacings h_r=h_0/2^r, with the finest spacing selected from the expected minimum feature width rather than relying on training to discover it. (2) Pseudocode: initialize coefficient tables A[r] with small Gaussian values; for each batch of coordinates x, find the three nonzero neighboring knots per dimension at every level; evaluate B2 and, when derivative losses are used, its analytic first and second derivatives; compute S[r]=sum_k A[r,k] product_j B2(x_j/h_r-k_j); form z=concat(x,S[0],...,S[R]); predict y=MLP(z); compute the task loss; backpropagate through A and the MLP; update with AdamW. Optionally learn alpha[r] using softplus-constrained positive parameters. (3) Compute compact-support interpolation and derivatives exactly. Estimate only empirical design choices such as R, h_R, channel count, and whether raw coordinates are retained. Numerically check C1 continuity by finite-differencing gradients on both sides of random knot boundaries. (4) First cheap experiment: fit a 2D synthetic field containing a narrow Gaussian ridge and smooth background, comparing equal-memory Fourier features, a ReLU MLP, and the spline-augmented MLP. Measure target-field error, derivative error, loss versus optimizer steps, and wall-clock time. Success is lower narrow-ridge and derivative error or reaching a fixed error in fewer steps at equal parameter count and memory.",
 "math_summary": "The paper uses C1 quadratic B-spline grids so that the field representation can support both second-order and fourth-order energy densities on the same point discretization. A concrete separable quadratic cardinal B-spline basis is B2(t)=3/4-t^2 for |t|\u003c=1/2, B2(t)=1/2(3/2-|t|)^2 for 1/2\u003c|t|\u003c=3/2, and B2(t)=0 otherwise. For a d-dimensional normalized coordinate x in a bounded box, grid level r has spacing h_r and coefficient tensor A_r; its scalar field is S_r(x)=sum over k in Z^d of A_{r,k} times the product over j of B2(x_j/h_r-k_j). The multiresolution representation is S(x)=sum from r=0 to R of alpha_r S_r(x), where alpha_r are learned or fixed level weights and h_R is the explicitly chosen finest scale. For a neural network, concatenate S_0(x),...,S_R(x) or use them to modulate an MLP: f_theta(x)=MLP_theta([x,S_0(x),...,S_R(x)]). The compact-support basis gives local evaluation, and matching first derivatives at knot boundaries give a C1 representation. The transferable property is that increasing R changes approximation bandwidth directly instead of requiring gradient descent to synthesize fine-scale oscillations from a coarse parameterization.",
 "math_tags": [
  "approximation-theory",
  "numerical-analysis"
 ],
 "ml_areas": [
  "mlp",
  "embedding",
  "training"
 ],
 "paper": {
  "arxiv_id": "2608.24126",
  "arxiv_url": "https://arxiv.org/abs/2608.24126",
  "summary_what_math_gives_to_ml": "The paper combines two concrete numerical ideas with direct relevance to neural-network training: a globally smooth, multiresolution coordinate representation and unbiased, continually resampled integration. The spline representation exposes a selectable finest spatial scale, avoiding the optimization burden of discovering narrow features through ordinary MLP weights, while its C1 continuity makes spatial derivatives smooth and supports derivative-based losses. The Monte Carlo construction gives exact importance weights for concentrating computation near localized structures without changing the target objective, and redrawing points prevents a model from overfitting a fixed quadrature set. The strongest ML transfer is to use spline-grid features for coordinate-conditioned networks and adaptive stratified resampling for expensive spatial or structured loss estimation.",
  "title": "A mesh-free multiresolution deep energy method with phase-field modeling of brittle fracture",
  "year": "2026"
 },
 "ratings": {
  "difficulty": 4,
  "novelty": 5,
  "usefulness": 7
 },
 "solves": [
  "accuracy",
  "stability",
  "speedup"
 ],
 "title": "C1 multiresolution spline features",
 "url": "https://synthcore.org/idea/124/c1-multiresolution-spline-features",
 "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
   }
  }
 }
}