[SciPy-Dev] Accuracy of single-precision FFT

[Pauli Virtanen]

Thu, 24 Jun 2010 13:41:01 -0400, Anne Archibald wrote:
[clip]
> We can't use FFTW, unfortunately, but it's clear that prime-size FFTs
> can do much better than we are. (Though what exactly is the definition
> of the errors, below?) 

>From Ralf's test program:

x = np.random.randn(size).astype(np.float32)
err = abs(x - ifft(fft(x))).max()

It's anyway in the order of magnitude of the rms error in this case.

Since 2**n FFTs have a much better accuracy, the problem is probably 
indeed in FFTPACK.

-- 
Pauli Virtanen