| Correlate function (bug?) [message #11103] |
Fri, 13 March 1998 00:00  |
Theo Brauers
Messages: 58
Registered: November 1997
|
Member |
|
|
Hi
I was facing a problem with the correlate function
which is demonstrated by the following IDL procedure:
pro testc
x=[1.D-5, 1.D-6, 0.0]
y=[1.D-5, 1.1D-6, 0.0]
f=[1.D0, 1.D1, 1.D2, 1.D3, 1.D4, 1.D5, 1.D6]
for i=0,6 do begin
print, i, correlate(x,y, /double) $
, correlate(x,y*f[i], /double) $
, correlate(x*f[i],y, /double) $
, correlate(x*f[i],y*f[i], /double)
endfor
END
giving the following result:
0 1.#QNAN00 1.#QNAN00 1.#QNAN00 1.#QNAN00
1 1.#QNAN00 1.#QNAN00 1.#QNAN00 1.#QNAN00
2 1.#QNAN00 1.#QNAN00 1.#QNAN00 1.#QNAN00
3 1.#QNAN00 1.#QNAN00 1.#QNAN00 0.99995432
4 1.#QNAN00 1.#QNAN00 1.#QNAN00 0.99995432
5 1.#QNAN00 0.99995432 0.99995432 0.99995432
6 1.#QNAN00 0.99995432 0.99995432 0.99995432
In my opinion there is no reason for an underflow or whatever causes
correlate to return NAN.
Best Theo
----------------------------------------------
Dr. Theo Brauers
Institut fuer Atmosphaerische Chemie (ICG-3)
Forschungszentrum Juelich
52425 JUELICH, Germany
Tel. +49-2461-61-6646 Fax. +49-2461-61-5346
|
|
|
|
| Re: Correlate function (bug?) [message #11269 is a reply to message #11103] |
Tue, 17 March 1998 00:00  |
John Smith
Messages: 4
Registered: March 1998
|
Junior Member |
|
|
William Connolley wrote:
> In article 9AA03CE1@fz-juelich.de, Theo Brauers <Th.Brauers@fz-juelich.de> writes:
>> I was facing a problem with the correlate function
>> which is demonstrated by the following IDL procedure:
>
>> pro testc
>> x=[1.D-5, 1.D-6, 0.0]
>> y=[1.D-5, 1.1D-6, 0.0]
>> print, correlate(x,y, /double)
>
>> giving the following result:
>>
>> NAN
>
> The problem is due to stupid programming in the correlate function.
> You can look at the source in $IDL_DIR/lib/correlate.pro.
> There is a tolerance (set to 1e-12 at /double, or 1e-6 otherwise) and if
> ||x'||^2*||y'||^2 is less than this, NAN is returned. Double precision should be
> good for a lot better than 1e-12, and anyway its more the ratio between x and y
> that would matter not the absolute value. Ho hum.
>
> Just take the code and insert tol=0 in the appropriate place...
>
> ps - the documentation doesn't mention this tolerance factor, presumably to avoid
> worrying our pretty little heads...
>
> - William
This bug was reported long time ago to RSI and they said they will fix it in one of
their
next releases. Meanwhile I use the old version.
Samuel Haimov
haimov@uwyo.edu
|
|
|
|
| Re: Correlate function (bug?) [message #11277 is a reply to message #11103] |
Mon, 16 March 1998 00:00 |
thompson
Messages: 584
Registered: August 1991
|
Senior Member |
|
|
Theo Brauers <Th.Brauers@fz-juelich.de> writes:
> Hi
> I was facing a problem with the correlate function
> which is demonstrated by the following IDL procedure:
> pro testc
> x=[1.D-5, 1.D-6, 0.0]
> y=[1.D-5, 1.1D-6, 0.0]
> f=[1.D0, 1.D1, 1.D2, 1.D3, 1.D4, 1.D5, 1.D6]
> for i=0,6 do begin
> print, i, correlate(x,y, /double) $
> , correlate(x,y*f[i], /double) $
> , correlate(x*f[i],y, /double) $
> , correlate(x*f[i],y*f[i], /double)
> endfor
> END
> giving the following result:
> 0 1.#QNAN00 1.#QNAN00 1.#QNAN00 1.#QNAN00
> 1 1.#QNAN00 1.#QNAN00 1.#QNAN00 1.#QNAN00
> 2 1.#QNAN00 1.#QNAN00 1.#QNAN00 1.#QNAN00
> 3 1.#QNAN00 1.#QNAN00 1.#QNAN00 0.99995432
> 4 1.#QNAN00 1.#QNAN00 1.#QNAN00 0.99995432
> 5 1.#QNAN00 0.99995432 0.99995432 0.99995432
> 6 1.#QNAN00 0.99995432 0.99995432 0.99995432
> In my opinion there is no reason for an underflow or whatever causes
> correlate to return NAN.
> Best Theo
This behavior seems to have been introduced in IDL version 5. IDL/v4 doesn't
have any problem, and returns the result
0 0.99995432 0.99995432 0.99995432 0.99995432
1 0.99995432 0.99995432 0.99995432 0.99995432
2 0.99995432 0.99995432 0.99995432 0.99995432
3 0.99995432 0.99995432 0.99995432 0.99995432
4 0.99995432 0.99995432 0.99995432 0.99995432
5 0.99995432 0.99995432 0.99995432 0.99995432
6 0.99995432 0.99995432 0.99995432 0.99995432
William Thompson
|
|
|
|
comp.lang.idl-pvwave archive
Members
Search
Help
Login
Home
