On 20 mai, 19:56, Christian Gollwitzer <aurio...@gmx.de> wrote: > Oops, I thought we were posting to comp.dsp. Nevertheless, I think > numpy.fft does mixed-radix (can't check it now) > > Am 20.05.13 19:50, schrieb Christian Gollwitzer: > > > > > > > > > Am 20.05.13 19:23, schrieb jmfauth: > >> Non sense. > > > Dito. > > >> The discrete fft algorithm is valid only if the number of data > >> points you transform does correspond to a power of 2 (2**n). > > > Where did you get this? The DFT is defined for any integer point number > > the same way. > > > Just if you want to get it fast, you need to worry about the length. For > > powers of two, there is the classic Cooley-Tukey. But there do exist FFT > > algorithms for any other length. For example, there is the Winograd > > transform for a set of small numbers, there is "mixed-radix" to reduce > > any length which can be factored, and there is finally Bluestein which > > works for any size, even for a prime. All of the aforementioned > > algorithms are O(log n) and are implemented in typical FFT packages. All > > of them should result (up to rounding differences) in the same thing as > > the naive DFT sum. Therefore, today > > >> Keywords to the problem: apodization, zero filling, convolution > >> product, ... > > > Not for a periodic signal of integer length. > > >> eg.http://en.wikipedia.org/wiki/Convolution > > > How long do you read this group? > > > Christian
------ Forget what I wrote. I'm understanding what I wanted to say, it is badly formulated. jmf -- http://mail.python.org/mailman/listinfo/python-list
