{
 "artifacts": null,
 "category": "architecture",
 "description": "For a network processing sparsely sampled functions, evaluate a shared encoder on several cyclic transformations of the domain and average the resulting predictions or latent covariance features. Cap the orbit size at the bandwidth-limited effective count, because increasing the nominal number of transformations beyond this point cannot reduce the global variance floor.",
 "formulas_latex": [
  "$$m^{*}(q)\\asymp n^{1/(2\\beta)}q^{-1/2}.$$",
  "$$r_q=\\min\\!\\left(q,\\frac{c_K}{h}\\right),\\qquad m^{*}(q)\\asymp n^{1/(2\\beta)}r_q^{-1/2}.$$",
  "$$\\operatorname{Var}_{\\mathrm{local}}\\asymp\\frac{1}{nm^2h^2},\\qquad \\operatorname{Var}_{\\mathrm{global}}\\asymp\\frac{1}{n}.$$",
  "$$\\bar z_q(x)=\\frac{1}{q}\\sum_{j=0}^{q-1}f_\\theta(g_j\\!\\cdot x),\\qquad q\\le q^\\sharp=\\left\\lfloor\\frac{c_K}{h}\\right\\rfloor.$$"
 ],
 "id": 221,
 "implementation": "Integrate the method at the output of a sparse-functional encoder or at a covariance-prediction head. Represent each input curve as observations {(t_i, y_i)} for i = 1,...,m on a periodic domain [0,1). Define g_j(t) = (t + j/q) mod 1 for j = 0,...,q-1, transform the observation locations while preserving their values, and run the same network f_theta on every transformed set. Average the outputs z_bar = (1/q) sum_j f_theta(g_j applied to x) before computing the supervised loss. Use shared weights and vectorize the q orbit evaluations; for an MVP, choose q from {1, 2, 4, 8}. Estimate h from the median nearest-neighbor spacing of the observed domain locations or from a fixed kernel bandwidth used by a local interpolation front-end. Set q_sharp = floor(c_K/h), initially taking c_K = 1, and do not expect gains beyond that cap. If the output is a covariance matrix, average positive-semidefinite matrices or Cholesky factors rather than unconstrained entries. The formulas determine the square-root scaling and the saturation rule; the constants and suitable h must be measured empirically. First test on a synthetic periodic functional-regression task and a small sparse-FDA benchmark, comparing q = 1 with q = 2, 4, and 8 at fixed model width and fixed numbers of observations per curve. Record validation error, prediction variance across repeated observation subsamples, training FLOPs, and inference latency. Success is lower error or lower prediction variance at fixed m, with diminishing returns after q_sharp rather than a claimed linear improvement in q.",
 "math_summary": "The paper establishes the threshold law m*(q) approximately proportional to n^(1/(2 beta)) q^(-1/2), where n is the number of independent curves, m is observations per curve, beta is the smoothness exponent, and q is the cyclic symmetry order. The effective orbit count is r_q = min(q, c_K/h), where h is the smoothing bandwidth and c_K is a kernel-dependent constant. The square-root gain follows from averaging the local variance term of order 1/(n m^2 h^2), while the global whole-curve variance remains order 1/n. For a network f_theta, let g_j be cyclic domain transformations and define the orbit average z_bar_q(x) = q^(-1) sum_{j=0}^{q-1} f_theta(g_j applied to x). The useful range is q \u003c= q_sharp = floor(c_K/h).",
 "math_tags": [
  "statistics",
  "probability",
  "harmonic-analysis"
 ],
 "ml_areas": [
  "architecture",
  "data-augmentation",
  "regularization"
 ],
 "paper": {
  "arxiv_id": "2608.27055",
  "arxiv_url": "https://arxiv.org/abs/2608.27055",
  "summary_what_math_gives_to_ml": "The paper gives a quantitative account of how symmetry averaging changes the sparse-functional-data sampling threshold. Averaging over a cyclic orbit reduces the local variance contribution only as the square root of the effective orbit count, with saturation at a bandwidth-dependent orbit size, while the global parametric variance floor of order 1/n is unchanged. For neural networks, this supports controllable orbit-averaged encoders for sparse functional inputs, combined with validation-based selection of the symmetry strength so that approximation bias does not exceed the statistical benefit.",
  "title": "How far can symmetry help? Phase transitions and symmetry selection in sparse functional data analysis",
  "year": "2026"
 },
 "ratings": {
  "difficulty": 5,
  "novelty": 5,
  "usefulness": 6
 },
 "solves": [
  "sample-efficiency",
  "stability",
  "accuracy"
 ],
 "title": "Saturated cyclic orbit averaging",
 "url": "https://synthcore.org/idea/221/saturated-cyclic-orbit-averaging",
 "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
   }
  }
 }
}