Filtering non-periodic functions (Q1892452)

A polynomial approximation \(p_ N\) of degree \(N\) for a function \(f\) defined on the interval \(I = [-1, 1]\) is said to have (global) spectral accuracy if the approximation error at \(x \in I\) decreases faster than any finite power of \(1/N\) for all \(x \in I\). If \(f\) is regular, the best \(L^ 2\) polynomial approximant has this property. Then the function value or the derivative has a discontinuity at \(s \in I\), the Gibbs phenomenon plays, but local spectral accuracy can be obtained, i.e., we have the behaviour of the error as above for any \(x\) staying away from \(s\). In this paper, such an approximant is constructed for non-periodic functions. It uses a filtering technique on the best \(L^ 2\) approximant (based on Legendre polynomials) which is the analog of a filtering technique used in the periodic case (based on Chebyshev polynomials). This adaptation is not trivial since the Chebyshev coefficients correspond to Fourier coefficients for periodic signals but the Legendre coefficients allow to recover only a part of the Fourier coefficients in the non-periodic case. Numerical examples show that this technique can be extended to the case of several discontinuities in \(I\) or to functions in two variables with discontinuities.