I have an old PR [1] to fix #2142 [2]. The idea is to have a new keyword for all fft functions to define the normalisation of the fft: - if 'norm' is None (the default), the normalisation is the current one: fft() is not normalized ans ifft is normalized by 1/n. - if norm is "ortho", the direct and inverse transforms are both normalized by 1/sqrt(n). The results are then unitary. The keyword name and value is consistent with scipy.fftpack.dct. Do you feel that it should be merged ? Joseph [1] https://github.com/numpy/numpy/pull/3883 [2] https://github.com/numpy/numpy/issues/2142