Skip to content

README.md

Latest commit

 

History

History
90 lines (67 loc) · 2.74 KB

README.md

File metadata and controls

90 lines (67 loc) · 2.74 KB

Faul–Goodsell–Powell (FGP) Algorithm

Fast NumPy implementation to generate an RBF interpolant in high dimensions

✨ Features

  • Core solver returns interpolation coefficients (lambdas) and constant (alpha).

  • Supports multiple radial basis function (RBF) kernels: multiquadric, Gaussian, inverse multiquadric, inverse quadratic.

  • Reproducible runs with seed control (np.random.default_rng).

  • Helper functions to generate sample points from balls, cubes, grids, or Gaussian distributions.

  • Helper function interp to evaluate the interpolant.

  • Returns per-iteration diagnostics (helpful_stats) for analysis.

  • The interpolant generated, s(x), is of the form

    $s(x) = \sum_i^n \lambda_i \phi(|x-x_i|) + \alpha$

    where $\phi$ is your chosen kernel function.

📦 Requirements

  • Python 3.9+
  • numpy, pandas

🚀 Quick Start

import numpy as np
from helper_functions import points_in_unit_ball, mq
from fgp_general import FGP

# Sample data
centers = points_in_unit_ball(100, 2, seed=42)
values = np.random.default_rng(42).uniform(0, 1, 100)

# Run FGP
num_iterations, lambdas, alpha, err, stats = FGP(
    data=centers, values=values,
    c=0.1, q=20, error=1e-5,
    max_iterations=1000, rbf_function=mq
)

🧮 API

FGP(data, values, c=0.1, q=30, error=1e-5, seed=42,
    max_iterations=1000, rbf_function=mq)

Args

  • data: (n, d) array – interpolation centres
  • values: (n,) array – function values
  • c: float – RBF shape parameter
  • q: int – neighborhood size (20–50 typical)
  • error: float – stopping tolerance
  • seed: int – RNG seed
  • max_iterations: int – iteration cap
  • rbf_function: callable – one of mq, gaussian, inv_mq, inv_quadratic

Returns

  • iterations: int – number of iterations
  • lambdas: ndarray – interpolation coefficients
  • alpha: float – interpolation constant
  • error: float – final residual
  • stats: pandas.DataFrame – per-iteration diagnostics

⚙️ Notes

  • q too small → unstable; too large → slower. Stick to 20–50, independent of dataset size.
  • NaNs/divergence usually mean poor kernel/shape parameter choice.

📖 Citation

If you use this repo, please cite this repo and the original paper:

A.C. Faul, G. Goodsell, M.J.D. Powell.
A Krylov subspace algorithm for multiquadric interpolation in many dimensions.
IMA Journal of Numerical Analysis, 25 (2005), 1–24.
doi:10.1093/imanum/drh021

Tags: radial-basis-functions · rbf-interpolation · fgp-algorithm · scattered-data · high-dimensional-data · numerical-analysis