Re: Again an FFT question [message #80625] |
Thu, 28 June 2012 06:07 |
Kenneth P. Bowman
Messages: 585 Registered: May 2000
|
Senior Member |
|
|
In article <27bbd728-a65d-4925-bd54-d003cfad6493@googlegroups.com>,
Helder <helder@marchetto.de> wrote:
> Thank you all for the answers. I guess the difference between changing sign
> of every second element and shift the array is due to a round-off error. This
> becomes clear when analyzing the differences between the two results using
> double precision or not.
> Thanks also for the explanation of how the frequencies in an FFT are
> displayed.
> I guess I'll have to update old code and remove the shift and use instead the
> center option.
I didn't plug it yesterday, but you might want to look at the chapter
on FFTs in my book.
http://www.amazon.com/An-Introduction-Programming-IDL-Intera ctive/dp/012088559X/r
ef=sr_1_1?ie=UTF8&qid=1340888241&sr=8-1&keywords =idl+bowman
I used to center FFTs, but have gotten away from that. For real input
data, the positive and negative frequencies are complex conjugates, so
the negative frequencies are redundant and you can get all of the information
by plotting only the positive frequencies.
For calculations in the spectral domain, such as filtering, I usually
compute the frequencies (or periods) in dimensional units as a function
of the transform array indices and use that. Then you don't have
to remember to uncenter when transforming back.
Ken Bowman
http://idl.tamu.edu/
|
|
|
Re: Again an FFT question [message #80626 is a reply to message #80625] |
Thu, 28 June 2012 02:00  |
Helder Marchetto
Messages: 520 Registered: November 2011
|
Senior Member |
|
|
On Wednesday, June 27, 2012 6:45:46 PM UTC+2, alx wrote:
> On 27 juin, 18:45, alx <lecacheux.al...@wanadoo.fr> wrote:
>> On 27 juin, 18:33, David Fanning <n...@idlcoyote.com> wrote:
>>
>>
>>
>>
>>
>>> alx writes:
>>>> You can use the /CENTER keyword in FFT function since IDL7.1
>>
>>> Well, it was *present* in IDL 7.1. I don't think
>>> it actually worked until later.
>>
>>> Cheers,
>>
>>> David
>>
>>> --
>>> David Fanning, Ph.D.
>>> Fanning Software Consulting, Inc.
>>> Coyote's Guide to IDL Programming:http://www.idlcoyote.com/
>>> Sepore ma de ni thui. ("Perhaps thou speakest truth.")
>>
>> ?
>> IDL> print,fft(cos(2*!pi*findgen(16)/16*3))
>> ( 5.06635e-008, 0.000000)( 8.06580e-008, 3.52675e-008)
>> (-9.04610e-009, 2.63418e-008)( 0.500000, 1.28007e-007)
>> (-9.08976e-008, 1.49012e-008)
>> (-2.65443e-008,-1.35536e-008)( 1.20273e-008, 2.63418e-008)
>> (-7.60014e-008,-9.51177e-008)( 1.25169e-007, -0.000000)
>> (-7.60014e-008, 9.51177e-008)
>> ( 1.20273e-008,-2.63418e-008)(-2.65443e-008, 1.35536e-008)
>> (-9.08976e-008,-1.49012e-008)( 0.500000,-1.28007e-007)
>> (-9.04610e-009,-2.63418e-008)
>> ( 8.06580e-008,-3.52675e-008)
>> IDL> print,fft(cos(2*!pi*findgen(16)/16*3),/CENTER)
>> ( 1.25169e-007, -0.000000)(-7.60014e-008, 9.51177e-008)
>> ( 1.20273e-008,-2.63418e-008)(-2.65443e-008, 1.35536e-008)
>> (-9.08976e-008,-1.49012e-008)
>> ( 0.500000,-1.28007e-007)(-9.04610e-009,-2.63418e-008)
>> ( 8.06580e-008,-3.52675e-008)( 5.06635e-008, 0.000000)
>> ( 8.06580e-008, 3.52675e-008)
>> (-9.04610e-009, 2.63418e-008)( 0.500000, 1.28007e-007)
>> (-9.08976e-008, 1.49012e-008)(-2.65443e-008,-1.35536e-008)
>> ( 1.20273e-008, 2.63418e-008)
>> (-7.60014e-008,-9.51177e-008)
>> Both statements give expected results: i.e. non zero values at
>> positions 3 and 13 (CENTER=0) or 5 and 11 (CENTER=1).
>> alx.
>
> I forgot to say, by using 8.2.
> alx.
Thank you all for the answers. I guess the difference between changing sign of every second element and shift the array is due to a round-off error. This becomes clear when analyzing the differences between the two results using double precision or not.
Thanks also for the explanation of how the frequencies in an FFT are displayed.
I guess I'll have to update old code and remove the shift and use instead the center option.
Regards,
Helder
|
|
|
Re: Again an FFT question [message #80647 is a reply to message #80626] |
Wed, 27 June 2012 09:45  |
lecacheux.alain
Messages: 325 Registered: January 2008
|
Senior Member |
|
|
On 27 juin, 18:45, alx <lecacheux.al...@wanadoo.fr> wrote:
> On 27 juin, 18:33, David Fanning <n...@idlcoyote.com> wrote:
>
>
>
>
>
>> alx writes:
>>> You can use the /CENTER keyword in FFT function since IDL7.1
>
>> Well, it was *present* in IDL 7.1. I don't think
>> it actually worked until later.
>
>> Cheers,
>
>> David
>
>> --
>> David Fanning, Ph.D.
>> Fanning Software Consulting, Inc.
>> Coyote's Guide to IDL Programming:http://www.idlcoyote.com/
>> Sepore ma de ni thui. ("Perhaps thou speakest truth.")
>
> ?
> IDL> print,fft(cos(2*!pi*findgen(16)/16*3))
> ( 5.06635e-008, 0.000000)( 8.06580e-008, 3.52675e-008)
> (-9.04610e-009, 2.63418e-008)( 0.500000, 1.28007e-007)
> (-9.08976e-008, 1.49012e-008)
> (-2.65443e-008,-1.35536e-008)( 1.20273e-008, 2.63418e-008)
> (-7.60014e-008,-9.51177e-008)( 1.25169e-007, -0.000000)
> (-7.60014e-008, 9.51177e-008)
> ( 1.20273e-008,-2.63418e-008)(-2.65443e-008, 1.35536e-008)
> (-9.08976e-008,-1.49012e-008)( 0.500000,-1.28007e-007)
> (-9.04610e-009,-2.63418e-008)
> ( 8.06580e-008,-3.52675e-008)
> IDL> print,fft(cos(2*!pi*findgen(16)/16*3),/CENTER)
> ( 1.25169e-007, -0.000000)(-7.60014e-008, 9.51177e-008)
> ( 1.20273e-008,-2.63418e-008)(-2.65443e-008, 1.35536e-008)
> (-9.08976e-008,-1.49012e-008)
> ( 0.500000,-1.28007e-007)(-9.04610e-009,-2.63418e-008)
> ( 8.06580e-008,-3.52675e-008)( 5.06635e-008, 0.000000)
> ( 8.06580e-008, 3.52675e-008)
> (-9.04610e-009, 2.63418e-008)( 0.500000, 1.28007e-007)
> (-9.08976e-008, 1.49012e-008)(-2.65443e-008,-1.35536e-008)
> ( 1.20273e-008, 2.63418e-008)
> (-7.60014e-008,-9.51177e-008)
> Both statements give expected results: i.e. non zero values at
> positions 3 and 13 (CENTER=0) or 5 and 11 (CENTER=1).
> alx.
I forgot to say, by using 8.2.
alx.
|
|
|
Re: Again an FFT question [message #80648 is a reply to message #80647] |
Wed, 27 June 2012 09:45  |
lecacheux.alain
Messages: 325 Registered: January 2008
|
Senior Member |
|
|
On 27 juin, 18:33, David Fanning <n...@idlcoyote.com> wrote:
> alx writes:
>> You can use the /CENTER keyword in FFT function since IDL7.1
>
> Well, it was *present* in IDL 7.1. I don't think
> it actually worked until later.
>
> Cheers,
>
> David
>
> --
> David Fanning, Ph.D.
> Fanning Software Consulting, Inc.
> Coyote's Guide to IDL Programming:http://www.idlcoyote.com/
> Sepore ma de ni thui. ("Perhaps thou speakest truth.")
>
>
?
IDL> print,fft(cos(2*!pi*findgen(16)/16*3))
( 5.06635e-008, 0.000000)( 8.06580e-008, 3.52675e-008)
(-9.04610e-009, 2.63418e-008)( 0.500000, 1.28007e-007)
(-9.08976e-008, 1.49012e-008)
(-2.65443e-008,-1.35536e-008)( 1.20273e-008, 2.63418e-008)
(-7.60014e-008,-9.51177e-008)( 1.25169e-007, -0.000000)
(-7.60014e-008, 9.51177e-008)
( 1.20273e-008,-2.63418e-008)(-2.65443e-008, 1.35536e-008)
(-9.08976e-008,-1.49012e-008)( 0.500000,-1.28007e-007)
(-9.04610e-009,-2.63418e-008)
( 8.06580e-008,-3.52675e-008)
IDL> print,fft(cos(2*!pi*findgen(16)/16*3),/CENTER)
( 1.25169e-007, -0.000000)(-7.60014e-008, 9.51177e-008)
( 1.20273e-008,-2.63418e-008)(-2.65443e-008, 1.35536e-008)
(-9.08976e-008,-1.49012e-008)
( 0.500000,-1.28007e-007)(-9.04610e-009,-2.63418e-008)
( 8.06580e-008,-3.52675e-008)( 5.06635e-008, 0.000000)
( 8.06580e-008, 3.52675e-008)
(-9.04610e-009, 2.63418e-008)( 0.500000, 1.28007e-007)
(-9.08976e-008, 1.49012e-008)(-2.65443e-008,-1.35536e-008)
( 1.20273e-008, 2.63418e-008)
(-7.60014e-008,-9.51177e-008)
Both statements give expected results: i.e. non zero values at
positions 3 and 13 (CENTER=0) or 5 and 11 (CENTER=1).
alx.
|
|
|
Re: Again an FFT question [message #80650 is a reply to message #80648] |
Wed, 27 June 2012 09:33  |
David Fanning
Messages: 11724 Registered: August 2001
|
Senior Member |
|
|
alx writes:
> You can use the /CENTER keyword in FFT function since IDL7.1
Well, it was *present* in IDL 7.1. I don't think
it actually worked until later.
Cheers,
David
--
David Fanning, Ph.D.
Fanning Software Consulting, Inc.
Coyote's Guide to IDL Programming: http://www.idlcoyote.com/
Sepore ma de ni thui. ("Perhaps thou speakest truth.")
|
|
|
Re: Again an FFT question [message #80652 is a reply to message #80650] |
Wed, 27 June 2012 09:22  |
lecacheux.alain
Messages: 325 Registered: January 2008
|
Senior Member |
|
|
On 26 juin, 23:11, Helder <hel...@marchetto.de> wrote:
> Dear FFTers,
> I was just wondering about the translation of the FFT so that the what is located at (0,0) goes in the middle of the image (N/2,M/2).
> Until now I did this doing simply a shift of the FFT image, that is:
> FFT_Img = SHIFT(FFT(Img),N/2,M/2)
> Now I have seen that some people use a FFT "trick" to shift the image. They switch every second pixel of an image to its negative value. This is justified by the translation properties of the DFTs and results in a translation of half the image size (for those seeking to understand the math, try to multiply the function (image) by exp(i*2*Pi(u0*x/N)) and after integration you will get a translation of the Fourier image (signal for 1D) of F(u-u0).
>
> I have tested the difference between the two shifting methods with the following code (in double precision just to be sure I wouldn't get rounding errors or so):
> ;**************************
> n = 256
> img = RANDOMU(S, n, n) ; Eventually put your image here, with the *right* size
> xx = LINDGEN(n) # (LONARR(n)+ 1)
> yy = (LONARR(n)+ 1) # LINDGEN(n)
> xxyy = xx+yy
> OddCenteredImg = Img
> EvenCenteredImg = Img
> Odd = WHERE((xxyy MOD 2), COMPLEMENT=Even)
> OddCenteredImg[Odd] = -Img[Odd]
> EvenCenteredImg[Even] = -Img[Even]
> WINDOW, XSIZE=2*n,YSIZE=3*n
> TVSCL, Img, 0 ;Transform by standard shifting
> TVSCL, OddCenteredImg, 1 ;Transform by standard shifting
> TVSCL, ALOG(ABS(SHIFT(FFT(Img,/DOUBLE),n/2,n/2))), 2 ;Transform by standard shifting
> TVSCL, ALOG(ABS(FFT(OddCenteredImg, /DOUBLE))), 3 ;Transform by pixel sign inversion
> TVSCL, ALOG(ABS(SHIFT(FFT(Img,/DOUBLE),n/2,n/2)) - ABS(FFT(OddCenteredImg, /DOUBLE))), 4 ;Power Spectrum Image Difference
> TVSCL, ALOG(ABS(SHIFT(FFT(Img,/DOUBLE),n/2,n/2)) - ABS(FFT(EvenCenteredImg, /DOUBLE))), 5 ;Power Spectrum Image Difference
> ;**************************
>
> The result is that the two are not exactly the same. Very similar, but not the same.
> I have tried varying the size of the image or switching even instead of odd numbers index numbers , but could not get any improvement (with odd image sizes, the difference is even higher).
>
> Does anybody have a reason to use one way (shift(fft(Img...))) rather than the other (switch pixels with index ((x+y) MOD 2 EQ 1)?
>
> I'm more confident using SHIFT, but I would just like to understand why the other method gives different values.
>
> Thanks,
> Helder
>
>
You can use the /CENTER keyword in FFT function since IDL7.1
alx.
|
|
|
Re: Again an FFT question [message #80657 is a reply to message #80652] |
Wed, 27 June 2012 08:53  |
Kenneth P. Bowman
Messages: 585 Registered: May 2000
|
Senior Member |
|
|
In article <614d4d33-39a8-4059-bc5f-f99f08fbc966@googlegroups.com>,
Helder <helder@marchetto.de> wrote:
> Dear FFTers,
> I was just wondering about the translation of the FFT so that the what is
> located at (0,0) goes in the middle of the image (N/2,M/2).
> Until now I did this doing simply a shift of the FFT image, that is:
> FFT_Img = SHIFT(FFT(Img),N/2,M/2)
> Now I have seen that some people use a FFT "trick" to shift the image. They
> switch every second pixel of an image to its negative value. This is
> justified by the translation properties of the DFTs and results in a
> translation of half the image size (for those seeking to understand the math,
> try to multiply the function (image) by exp(i*2*Pi(u0*x/N)) and after
> integration you will get a translation of the Fourier image (signal for 1D)
> of F(u-u0).
>
> The result is that the two are not exactly the same. Very similar, but not
> the same.
> I have tried varying the size of the image or switching even instead of odd
> numbers index numbers , but could not get any improvement (with odd image
> sizes, the difference is even higher).
>
> Does anybody have a reason to use one way (shift(fft(Img...))) rather than
> the other (switch pixels with index ((x+y) MOD 2 EQ 1)?
>
> I'm more confident using SHIFT, but I would just like to understand why the
> other method gives different values.
>
> Thanks,
> Helder
If you do an FFT of an 8-point array, the output frequencies are
stored in this order
f = [0, 1, 2, 3, 4, -3, -2, -1]
The IDL FFT always does a complex FFT. If the input data are real,
then the real part of f=0 is the mean, the imaginary part is 0 to
within roundoff error.
The positive and negative frequencies are complex conjugates of
each other - again within roundoff error.
I think the differences that you are seeing are due to round-off
error. Check the magnitudes of the differences.
Ken Bowman
|
|
|