In various papers, different window functions were considered, e.g. Gaussian pulse tapered with a Hanning window in [4], Gaussian kernels combined with sinc kernels in [19], and special optimised windows in [15,4]. Furthermore, special approaches based on scaling vectors [17], based on minimising the Frobenius norm of certain error matrices [18] or based on min-max interpolation [8] are proposed. However, the numerical results in [18,8] show that these approaches are not superior to the approach based on Kaiser-Bessel functions.
An even more general FFT, where both the knots in frequency and time/spatial domain are arbitrary, was developed in [6]. Up to now, the library does not support this extension.
Our algorithms are based on the approach in [20]. Here, one can change the window functions in a very simple way. See Section 3 and Section 4 for a suitable choice of the window function with respect to accuracy, speed and memory usage.