Code covered by the BSD License

4.25

4.2 | 4 ratings Rate this file 35 Downloads (last 30 days) File Size: 102 KB File ID: #25473

# Why use fftshift(fft(fftshift(x))) in Matlab instead of fft(x)?

by

Explain why we use fftshift(fft(fftshift(x))) in Matlab instead of fft(x)

File Information
Description

Explain why we use fftshift(fft(fftshift(x))) in Matlab instead of fft(x). An example is given. The example and Matlab codes are partially copied from Daniele Disco ‘s work in "A guide to the Fast Fourier Transform, 2nd Edition".

Acknowledgements

A Guide To The Fft 2nd Edition Plus inspired this file.

MATLAB release MATLAB 7.6 (R2008a)
Tags for This File   Please login to tag files.
 Please login to add a comment or rating.
Comments and Ratings (8)
03 Jan 2014

Hi Yuji,

ifftshift is to correct the bug in fftshift when dealing with the sequence with odd length (that is the number of elements in the sequence is odd). So it is safe to always use: fftshift(fft(ifftshift(sig))) or fftshift(ifft(ifftshift(spectrum))).

13 May 2013

hi Kan,Shalin,

Thanks for the code and the discussion - nice and helpful.

A book I'm reading says it should be
fftshift(fft(fftshift(sig)))
and
ifftshift(ifft(ifftshift(spectrum)))

I'm confused. Could you explain where to use ifftshift when N=odd?

Thank you!
(The book is "Numerical Simulation of Optical Wave Propagation with examples in Matlab")

23 May 2010

thanl you very much !

18 May 2010

You're welcome. ifftshift is the same as fftshift for even length sequence, but different for odd length. So if one uses fftshift(fft(ifftshift(...))) things work well. By the way, scaling by dt and df to correct for scaling introduced by FFT algorithm is neat trick.

16 May 2010

To Mehta, yes you r right. ifftshift is used for sequence with odd length. Thank you for the correction!

14 May 2010

There is a problem with above recipe. It fails when you have sequence of odd length. Correct recipe is:
fftshift(fft(ifftshift(sig))) or fftshift(ifft(ifftshift(spectrum))).

A description of this can be found on my submission on fftshift, ifftshift.

To observe that above is true, run the following code with fftshift and ifftshift on inner call for computing Xfinal.
-------------------------
Bx = 50;
A = sqrt(log(2))/(2*pi*Bx);
fs = 500; %sampling frequency
dt = 1/fs; %time step
T=1; %total time window
t = -T/2:dt:T/2; %time grids
df = 1/T; %freq step
Fmax = 1/2/dt; %freq window
f=-Fmax:df:Fmax; %freq grids, not used in our examples, could be used by plot(f, X)

x = exp(-t.^2/(2*A^2));
Xan = A*sqrt(2*pi)*exp(-2*pi^2*f.^2*A^2); %X(f), analytical Fourier transform of x(t), real
Xfft = dt * fft(x); %directly using fft()
Xfftshift = dt * fft(fftshift(x)); %using fftshift() before fft()
Xfinal = dt * fftshift(fft(ifftshift(x))); %identical with analytical X(f), also note dt
subplot(211); plot(f,Xan,f,real(Xfinal),'--');
subplot(212); plot(f,imag(Xfinal));

18 Dec 2009

Very interesting!

21 Oct 2009

Contact us