IDL hilbert() function [message #63183] |
Sat, 01 November 2008 02:14  |
lecacheux.alain
Messages: 325 Registered: January 2008
|
Senior Member |
|
|
Is this function actually computing the Hilbert transform ?
The Hilbert transform is known to be idempotent, i.e. H(H(x)) = -x.
However, by applying the IDL function, one get for instance :
IDL> print, hilbert (hilbert (indgen(8)))
( 6.00000, 0.000000)( 7.00000, 0.000000)
( 4.00000, 0.000000)( 5.00000, 0.000000)
( 2.00000, 0.000000)( 3.00000, 0.000000)
( 0.000000, 0.000000)( 1.00000, 0.000000)
|
|
|
Re: IDL hilbert() function [message #63294 is a reply to message #63183] |
Tue, 04 November 2008 05:45  |
lecacheux.alain
Messages: 325 Registered: January 2008
|
Senior Member |
|
|
On 3 nov, 20:54, "R.G. Stockwell" <notha...@noemail.com> wrote:
> <ed.schm...@gmail.com> wrote in message
>
> news:f43242be-734b-4098-8bc1-3ab0691270d9@c22g2000prc.google groups.com...
> On Nov 3, 2:16 am, Wox <s...@nomail.com> wrote:
>
>
>
>
>
>> On Sat, 1 Nov 2008 02:14:54 -0700 (PDT), lecacheux.al...@wanadoo.fr
>> wrote:
>
>>> Is this function actually computing the Hilbert transform ?
>>> The Hilbert transform is known to be idempotent, i.e. H(H(x)) = -x.
>>> However, by applying the IDL function, one get for instance :
>>> IDL> print, hilbert (hilbert (indgen(8)))
>>> ( 6.00000, 0.000000)( 7.00000, 0.000000)
>>> ( 4.00000, 0.000000)( 5.00000, 0.000000)
>>> ( 2.00000, 0.000000)( 3.00000, 0.000000)
>>> ( 0.000000, 0.000000)( 1.00000, 0.000000)
>
>> It works for this:
>> x=findgen(180)/90.*!pi
>> plot,x,hilbert(hilbert(sin(x))),/xs
>> oplot,x,-sin(x),psym=1
>
>> Not for this (flips):
>> plot,hilbert(hilbert((indgen(1000))))
>> oplot,indgen(1000),psym=1
>
>> I'm not sure why, but check the hilbert.pro in the IDL-lib directory
>> to see how it's implemented.
>
> Hello Hilbert fans,
>
> Hilbert.pro is not a perfect implementation of the Hilbert function.
> One of its failures is an inability to treat odd numbered arrays
> properly. Try this:
>
> x1=findgen(9)-4 ; Evenly distributed around the origin
> y1=abs(x1) ; An even function of x1
> h1=hilbert(y1) ; This should be an ODD function of x1, i.e. H(-
> x)=-H(x)
> plot,x1,y1
> oplot,x1,h1,psym=-6 ; But it's clearly not exactly anti-symmetric
> about the origin!
>
> The problem goes away if you use 10 values instead of 9:
>
> x2=findgen(10)-4.5 ; Again evenly distributed around the origin
> y2=abs(x2) ; An even function of x2
> h2=hilbert(y2) ; This should be an ODD function, i.e., H(-x)=-
> H(x)
> plot,x2,y2
> oplot,x2,h2,psym=-6 ; ... and it is: h2(x2)=-h2(-x2) plus a constant
>
> Looking at the source code suggests that an easy fix is possible.
>
> Ed Schmahl
> ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
>
> A very long time ago, I modified the code (it is below), check it out.
> I did run into a number of problems with hilbert in both IDL and MATLAB.
> Matlab was just flat out wrong (it expanded the time series to be a power of
> 2,
> but did not expand the convolving kernel properly). After many arguements
> in email
> with a matlab developer, they finally acquiesced (but simply making the
> wrong behaviour
> a keyword option (rolls eyes)).
>
> Cheers,
> bob
>
> Here is the code:
>
> FUNCTION HILBERT,X,D, ANALYTIC = a ; performs the Hilbert transform of
> some data.
> ON_ERROR,2 ; Return to caller if an error occurs
> Y=FFT(X,-1) ; go to freq. domain.
> N=N_ELEMENTS(Y)
>
> I=COMPLEX(0.0,-1.0)
>
> IF N_PARAMS(X) EQ 2 THEN I=I*D
> N2=ceil(N/2.)-1 ; effect of odd and even # of elements
> ; considered here.
> y(0) = complex(0,0) ; zero the DC value (required for hilbert
> trans.)
> Y(1)=Y(1:N2)*I ; multiplying by I rotates counter c.w. 90 deg.
> if (n mod 2) eq 0 then Y(N2+1) = complex(0,0) ; don't need this
> N2=N-N2
> Y(N2)=Y(N2:N-1)/I
> Y=float(FFT(Y,1)) ; go back to time domain
> if keyword_set(a) then y = complex(x,y)
> RETURN,y
>
> END- Masquer le texte des messages précédents -
>
> - Afficher le texte des messages précédents -
Thank you all for answering: I was a little bit confused in finding
both IDL (6.4)
and Matlab (5.4) in error. But you confirmed that !
Thanks for the code. Using IDL, a simpler one might be :
function Hilbert, x, DOUBLE=double
n = N_elements (x)
m = ((n mod 2) eq 0) ? [0, dblarr(n/2-1) + 1, 0, dblarr(n/2-1) -
1] : [0, dblarr((n-1)/2) + 1, dblarr((n-1)/2) - 1]
return, fft (fft (x, -1, DOUBLE=double)*m*complex(0,1), 1,
DOUBLE=double)
end
function Analytic, x, DOUBLE=double
n = N_elements (x)
m = ((n mod 2) eq 0) ? [1, dblarr(n/2-1) + 2, 1, dblarr(n/2-1)] :
[1, dblarr((n-1)/2) + 2, dblarr((n-1)/2)]
return, fft (fft (x, -1, DOUBLE=double)*m, 1, DOUBLE=double)
end
alx.
|
|
|
Re: IDL hilbert() function [message #63312 is a reply to message #63183] |
Mon, 03 November 2008 11:54  |
R.G. Stockwell
Messages: 363 Registered: July 1999
|
Senior Member |
|
|
<ed.schmahl@gmail.com> wrote in message
news:f43242be-734b-4098-8bc1-3ab0691270d9@c22g2000prc.google groups.com...
On Nov 3, 2:16 am, Wox <s...@nomail.com> wrote:
> On Sat, 1 Nov 2008 02:14:54 -0700 (PDT), lecacheux.al...@wanadoo.fr
> wrote:
>
>> Is this function actually computing the Hilbert transform ?
>> The Hilbert transform is known to be idempotent, i.e. H(H(x)) = -x.
>> However, by applying the IDL function, one get for instance :
>> IDL> print, hilbert (hilbert (indgen(8)))
>> ( 6.00000, 0.000000)( 7.00000, 0.000000)
>> ( 4.00000, 0.000000)( 5.00000, 0.000000)
>> ( 2.00000, 0.000000)( 3.00000, 0.000000)
>> ( 0.000000, 0.000000)( 1.00000, 0.000000)
>
> It works for this:
> x=findgen(180)/90.*!pi
> plot,x,hilbert(hilbert(sin(x))),/xs
> oplot,x,-sin(x),psym=1
>
> Not for this (flips):
> plot,hilbert(hilbert((indgen(1000))))
> oplot,indgen(1000),psym=1
>
> I'm not sure why, but check the hilbert.pro in the IDL-lib directory
> to see how it's implemented.
Hello Hilbert fans,
Hilbert.pro is not a perfect implementation of the Hilbert function.
One of its failures is an inability to treat odd numbered arrays
properly. Try this:
x1=findgen(9)-4 ; Evenly distributed around the origin
y1=abs(x1) ; An even function of x1
h1=hilbert(y1) ; This should be an ODD function of x1, i.e. H(-
x)=-H(x)
plot,x1,y1
oplot,x1,h1,psym=-6 ; But it's clearly not exactly anti-symmetric
about the origin!
The problem goes away if you use 10 values instead of 9:
x2=findgen(10)-4.5 ; Again evenly distributed around the origin
y2=abs(x2) ; An even function of x2
h2=hilbert(y2) ; This should be an ODD function, i.e., H(-x)=-
H(x)
plot,x2,y2
oplot,x2,h2,psym=-6 ; ... and it is: h2(x2)=-h2(-x2) plus a constant
Looking at the source code suggests that an easy fix is possible.
Ed Schmahl
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
A very long time ago, I modified the code (it is below), check it out.
I did run into a number of problems with hilbert in both IDL and MATLAB.
Matlab was just flat out wrong (it expanded the time series to be a power of
2,
but did not expand the convolving kernel properly). After many arguements
in email
with a matlab developer, they finally acquiesced (but simply making the
wrong behaviour
a keyword option (rolls eyes)).
Cheers,
bob
Here is the code:
FUNCTION HILBERT,X,D, ANALYTIC = a ; performs the Hilbert transform of
some data.
ON_ERROR,2 ; Return to caller if an error occurs
Y=FFT(X,-1) ; go to freq. domain.
N=N_ELEMENTS(Y)
I=COMPLEX(0.0,-1.0)
IF N_PARAMS(X) EQ 2 THEN I=I*D
N2=ceil(N/2.)-1 ; effect of odd and even # of elements
; considered here.
y(0) = complex(0,0) ; zero the DC value (required for hilbert
trans.)
Y(1)=Y(1:N2)*I ; multiplying by I rotates counter c.w. 90 deg.
if (n mod 2) eq 0 then Y(N2+1) = complex(0,0) ; don't need this
N2=N-N2
Y(N2)=Y(N2:N-1)/I
Y=float(FFT(Y,1)) ; go back to time domain
if keyword_set(a) then y = complex(x,y)
RETURN,y
END
|
|
|