{
 "artifacts": null,
 "category": "architecture",
 "description": "Replace an SVD-based pseudoinverse of a learned rectangular matrix with a low-degree polynomial initialization followed by a few Newton–Schulz iterations. The polynomial approximates the inverse Gram operator, while a cheap residual test accepts it only when the iteration is contractive and otherwise selects a conservative transpose-scaled initialization.",
 "formulas_latex": [
  "$$X_0=A^{\\mathsf T}p_d(AA^{\\mathsf T}),\\qquad p_d(t)=\\sum_{r=0}^{d}c_rt^r,\\qquad c=\\arg\\min_c\\sum_{\\ell=1}^{L}w_\\ell\\left(p_d(t_\\ell)-\\frac{1}{t_\\ell}\\right)^2.$$",
  "$$X_{k+1}=X_k\\left(2I_m-AX_k\\right),\\qquad R_k=I_m-AX_k.$$",
  "$$e_{k,i}=1-\\sigma_i x_{k,i}\\quad\\Longrightarrow\\quad e_{k+1,i}=e_{k,i}^{\\,2},\\qquad \\rho_0=\\left\\|I_m-AX_0\\right\\|_2\u003c1.$$",
  "$$X_0^{\\mathrm{safe}}=\\alpha A^{\\mathsf T},\\qquad \\alpha=\\frac{1}{\\widehat{\\|A\\|}_2^{\\,2}+\\varepsilon}.$$"
 ],
 "id": 1588,
 "implementation": "(1) Integration point: replace a batched SVD or pseudoinverse call for matrices A with shape [batch,m,n], such as the solve inside a low-rank adapter, whitening layer, implicit block, or CUR middle-factor module. The output X should have shape [batch,n,m]. (2) Pseudocode: estimate b, the largest eigenvalue of AA^T, by power iteration; choose a positive lower bound a from calibration data or a conservative fraction of b; build L nodes t_l in [a,b]; solve the small weighted least-squares problem for polynomial coefficients c_r; evaluate P=sum_{r=0}^d c_r(AA^T)^r using Horner evaluation; set X=A^T P; estimate rho0=||I-AX||_2 with one or two power iterations. If rho0\u003c0.9, retain X; otherwise set alpha=1/(b+epsilon) and replace X with alpha A^T. Run K updates X \u003c- X @ (2I - A @ X), stopping early when ||I-AX||_F divided by sqrt(m) is below tolerance. (3) The paper-derived computations are the polynomial inverse approximation, Newton–Schulz update, and quadratic residual recursion. Estimate b, rho0, and the useful spectral interval empirically; detach these estimates from autograd if their gradients are unstable. Use float32 or float64 for the matrix iteration even when the surrounding model uses lower precision. (4) First experiment: compare SVD pseudoinverse, safe transpose-scaled Newton–Schulz, and polynomial-screened Newton–Schulz on 64 by 32 matrices with prescribed condition numbers and rank deficiencies. Then insert each method into a LoRA linear regression layer on CIFAR-10. Record pseudoinverse residual, task loss, wall-clock time, peak memory, number of iterations, and gradient norm. The method succeeds if it reaches the SVD residual and task accuracy with lower cost, ideally at least 2x faster, while avoiding divergence on ill-conditioned batches.",
 "math_summary": "Let A be an m by n real matrix and let A^+ be its Moore–Penrose inverse. Define a polynomial p_d(t)=sum from r=0 to d of c_r t^r that approximates 1/t on an interval [a,b] containing the nonzero eigenvalues of AA^T. The coefficients minimize the weighted least-squares error over nodes t_l: c=argmin sum_l w_l (p_d(t_l)-1/t_l)^2. Initialize X_0=A^T p_d(AA^T), where X_0 has shape n by m. Newton–Schulz updates X_{k+1}=X_k(2I_m-AX_k). If A=U Sigma V^T and x_{k,i} is the scalar coefficient of X_k on a singular direction with singular value sigma_i, then e_{k,i}=1-sigma_i x_{k,i} satisfies e_{k+1,i}=e_{k,i}^2. Thus the start is contractive when the residual R_0=I_m-AX_0 has spectral norm rho_0 less than 1. If the polynomial start is unsafe, use X_0_safe=alpha A^T with alpha=1/(estimated spectral norm of A squared plus epsilon). Rank-deficient directions are preserved because the iteration acts independently on singular subspaces and zero singular values remain zero.",
 "math_tags": [
  "linear-algebra",
  "numerical-analysis",
  "approximation-theory",
  "spectral-theory"
 ],
 "ml_areas": [
  "architecture",
  "training",
  "inference-speedup"
 ],
 "paper": {
  "arxiv_id": "2607.29270",
  "arxiv_url": "https://arxiv.org/abs/2607.29270",
  "summary_what_math_gives_to_ml": "The transferable contribution is a representation-based pseudoinverse method with a polynomial inverse initialization, an explicit Newton–Schulz residual recursion, and a spectral acceptance test. This can replace repeated SVD pseudoinverses inside low-rank adapters, whitening layers, implicit modules, or matrix-compression blocks. The most practical adaptation is to compute a cheap polynomial start, verify contraction using an estimated residual norm, and fall back to a safe transpose-scaled start when the test fails.",
  "title": "Iterative Methods for Computing the Moore--Penrose Inverse of Split-Quaternion Matrices with Applications",
  "year": "2026"
 },
 "ratings": {
  "difficulty": 5,
  "novelty": 6,
  "usefulness": 6
 },
 "solves": [
  "speedup",
  "stability",
  "memory"
 ],
 "title": "Spectrally screened polynomial pseudoinverse layer",
 "url": "https://synthcore.org/idea/1588/spectrally-screened-polynomial-pseudoinverse-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
   }
  }
 }
}