The polynomial Fourier transform with minimum mean square error for noisy data (Q972771): Difference between revisions
From MaRDI portal
Changed an Item |
Set profile property. |
||
Property / MaRDI profile type | |||
Property / MaRDI profile type: MaRDI publication profile / rank | |||
Normal rank |
Revision as of 01:47, 5 March 2024
scientific article
Language | Label | Description | Also known as |
---|---|---|---|
English | The polynomial Fourier transform with minimum mean square error for noisy data |
scientific article |
Statements
The polynomial Fourier transform with minimum mean square error for noisy data (English)
0 references
21 May 2010
0 references
Building on the earlier work of \textit{N. Saito} and \textit{J.-F. Remy} [Appl. Comput. Harmon. Anal. 20, No. 1, 41--73 (2006; Zbl 1089.94003)], the authors consider various least squares approximation schemes involving polyharmonic local Fourier transforms (PHLFTs). The essential problem in each case is to minimize \(\|f-u_p-v_q\|\) in \(L^2[-1,1]\) where \(f\) is a noise-corrupted signal, \(u\) is a polynomial of degree at most \(p\), and \(v\) is a trigonometric polynomial of degree at most \(q\). One refers to a corresponding approximant \(u_p+v_q\) as a {\textit{one-step least squares approximant}}. It is shown that a best approximant exists when \(p\) and \(q\) are specified, and then that \(v_q\) has to be the partial Fourier sum of \(f-u_p\). In particular, such an approximation can be made with arbitrary precision by allowing \(q\) to be large enough. If one specifies that \(u\) has the same boundary values as \(f\), that the degrees \(p,q\) are fixed, and that \(v_q\) is the partial Fourier series of \(f-u\), the approximant \(u_p+v_q\) is called the {\textit{near least squares approximant}}. The authors also consider a {\textit{two-step}} approximant in which a minimizer \(u_\ast\) of \(\|f-u_p\|\) is found first, and then \(\|(f-u_\ast)-v_q\|\) is minimized. Polyharmonic approximations are constructed using Hermite interpolants. For example, Theorem 2.11 states that if \(f\in C^{2m}[-1,1]\) has an \((m-1)\)st derivative that is Hölder continuous of order \(\alpha\in (0,1]\), then its PHLFT representation of order \(m\) has the form \(f(t)=\phi(t)+{1\over \sqrt 2} \sum_{} c_j e^{i j\pi t}\), where \(\phi\) is a Hermite interpolating polynomial of degree at most \(2m-1\) satisfying the polyharmonic equation \(\phi^{(2m)}=0\) and \(f^{(\ell)}(\pm 1) =\phi{(\ell)}(\pm 1)\), \(\ell\leq m\), so that the Fourier coefficients \(c_j\) decay like \(O(j^{-m-1})\). One then obtains a near-least squares approximant by truncating the Fourier series to degree \(q\), and these approximants converge uniformly of order \(O(\log{q}/q^{m-1-\alpha})\) as \(q\to\infty\). Approximate numerical implementations of the one-step least squares, two-step least squares, near least squares, and PHLFT approximations are all considered and experimental results are compared. The one-step approximation scheme is the most accurate, although the near least squares and PHLFT approximations are nearly as good in \(\ell^2\), and better in \(\ell^\infty\) for uncorrupted data with high SNR. The one-step and two-step methods perform better, though, both in \(\ell^2\) and \(\ell^\infty\), when noise levels are high.
0 references
algebraic polynomials
0 references
trigonometric polynomials
0 references
least squares approximation
0 references
mean square error
0 references
noisy data
0 references