comp.lang.idl-pvwave archive: archive

Home » Public Forums » archive » Re: randomn problem

Show: Today's Messages :: Show Polls :: Message Navigator
E-mail to friend

Re: randomn problem [message #53030 is a reply to message #52948]

Tue, 13 March 2007 04:03

Paolo Grigis
Messages: 171
Registered: December 2003

Senior Member

Ben Panter wrote:
> Nigel Wade wrote:
>
>> Something also changed between IDL 6.1 and IDL 6.2:
>> Maybe the algorithm has been changed to one which propogates more
>> round-off
>> error?
>
> It would seem that it didn't get better in 6.3:
>
> IDL> print, !version
> { x86 linux unix linux 6.3 Mar 23 2006 32 64}
>
> IDL> print, stddev(randomn(seed, 1e8))
> % Compiled module: STDDEV.
> 0.852648
>
> IDL> print, stddev(randomn(seed, 1e8, /double) )
> 1.0000419
>
> Ben
>

Ok, let's try to understand what's going on. stddev calls
moment which uses the built-in function total. The moment
algorithm did not change in recent versions. We actually
expect total to fail to sum 1d8 small floats accurately
in single precision, whereas this should be no problem
in double precision. But it looks like in some versions of
IDL double precision is used even when the input is float
and no /double keyword is set. To test this, we can use
the following commands:

n=10L^7
a=replicate(0.1,n)
print,'TOTAL A:(SINGLE PREC.):',total(a)
print,'TOTAL A (DOUBLE PREC.): ',total(a,/double)

The output is (the exact value will depend on the hardware)

TOTAL A:(SINGLE PREC.): 1.08794e+06
TOTAL A (DOUBLE PREC.): 1000000.0

in 5.4,5.5, 6.2 and 6.3 and

TOTAL A:(SINGLE PREC.): 1.00000e+06
TOTAL A (DOUBLE PREC.): 1000000.0

in 5.6 and 6.0.

So, I would guess that in these two versions, the total
is internally computed in double precision even when
/double is not set. Or may this be due to a different
way in threading the "total" operation?

Ciao,
Paolo

Report message to a moderator

[Message index]

		Re: randomn problem By: Ben Panter on Mon, 12 March 2007 09:17
		Re: randomn problem By: Nigel Wade on Mon, 12 March 2007 09:07
		Re: randomn problem By: Paolo Grigis on Mon, 12 March 2007 02:07
		Re: randomn problem By: Ingo von Borstel on Mon, 12 March 2007 01:03
		Re: randomn problem By: David Fanning on Sun, 11 March 2007 18:10
		Re: randomn problem By: askemer on Sun, 11 March 2007 16:04
		Re: randomn problem By: David Fanning on Sun, 11 March 2007 12:41
		Re: randomn problem By: JD Smith on Tue, 13 March 2007 09:25
		Re: randomn problem By: Paolo Grigis on Tue, 13 March 2007 04:03

Previous Topic:	Re: New IDLDE
Next Topic:	Delete bad data and interplate the new data

-=] Back to Top [=-

[ Syndicate this forum (XML) ] [

] [

]

Current Time: Wed Oct 08 18:11:29 PDT 2025

Total time taken to generate the page: 0.00201 seconds