Random Number Generator Bugs
Name: Wayne Landsman and others
E-mail Address: landsman@stars.gsfc.nasa.gov
IDL version: IDL 5.1
Platform and OS: All?
Description of Behavior:
[This subject generated a lot of interest on the IDL
newsgroup. I've just collected the discussion and include
it here for your information. DWF]
Subject: RANDOMU bug (and HTML help)
From: landsman@stars.gsfc.nasa.gov (Wayne Landsman)
Date: 21 Feb 1999 00:21 EST
Versions 5.1.1 and 5.2 of IDL have a bug in the RANDOMU (and RANDOMN) function
such that the SEED variable is initialized to the same value at the start of
each session, rather than being intialized by the system clock (see example
below). I believe that so long as one as one stays within one IDL session
that this causes no problems, but if one is, say, combining Monte Carlo
simulations from different IDL sessions, then the results will be decidedly
unrandom.
RSI knows about this problem and say that they are giving it very high
priority for a fix.
On another subject, does anyone know if there is a HTML help distribution for
IDL V5.2, like there was for IDL V5.1? I couldn't find one on the CD-Rom
distribution.
--Wayne Landsman landsman@mpb.gsfc.nasa.gov
mpb{landsman}102: idl
IDL> print,!VERSION
{ sparc sunos unix 5.2 Oct 30 1998}
IDL> print,randomu(seed)
0.415999
IDL> print,randomu(seed)
0.0919649
IDL> exit
mpb{landsman}102: idl
IDL Version 5.2 (sunos sparc). Research Systems, Inc.
IDL> print,randomu(seed)
0.415999
IDL> print,randomu(seed)
0.0919649
Subject: Re: RANDOMU bug (and HTML help)
From: ajschmitt@my-dejanews.com
Date: Mon, 22 Feb 1999 19:55:13 GMT
This is not the worst of it. For some of us, it is important to be able
to have the same seed so that the same "random" sequence is produced.
However, RSI quietly changed the manner in which the SEED variable
interacts with the RANDOM functions between version 5.0 & 5.1.
Consider the following behavior from IDL v 5..0.3:
IDL> seed = 2 & print, randomu(seed, 3)
0.342299 0.402381 0.307838
...doing this multiple times will always give the same result.
However, in IDL v.5..1 & later,. using this several times in a row
produces different results each time.:
IDL> seed = 2 & print, randomu(seed, 3)
0.0594004 0.982075 0.358593
IDL> seed = 2 & print, randomu(seed, 3)
0.831999 0.303037 0.506712
...etcetera. It turns out that you now have to specify a NEGATIVE seed
in order for it to have any influence on the generated sequence:
seed = -2 & print, randomu(seed, 3)
0.342299 0.402381 0.307838
seed = -2 & print, randomu(seed, 3)
0.342299 0.402381 0.307838
I assume that RSI knows about this (since it's documented in the new online
help file), but they didn't see fit to mention it in the release notes
accompanying version 5.1. Whatever reason they may have had for changing
this behavior (which has been working the previous way since IDL version 1.0
at least 14 years ago) I cannot fathom. (Is there some reason someone would
bother to set the seed only to have it be ignored by the random functions???)
This new "feature" caused one of my coworkers to waste more than a day
of his time tracking down the poorly documented change. Needless to say,
we are not very satisfied with RSI about this.
-Andy Schmitt
-----------== Posted via Deja News, The Discussion Network ==----------
http://www.dejanews.com/ Search, Read, Discuss, or Start Your Own
:56 EST
In article <7ascms$kin$1@nnrp1.dejanews.com>, ajschmitt@my-dejanews.com writes...
>This is not the worst of it. For some of us, it is important to be able
>to have the same seed so that the same "random" sequence is produced.
>However, RSI quietly changed the manner in which the SEED variable
>interacts with the RANDOM functions between version 5.0 & 5.1.
>
>Consider the following behavior from IDL v 5..0.3:
>IDL> seed = 2 & print, randomu(seed, 3)
> 0.342299 0.402381 0.307838
>....doing this multiple times will always give the same result.
>
>However, in IDL v.5..1 & later,. using this several times in a row
>produces different results each time.:
>IDL> seed = 2 & print, randomu(seed, 3)
> 0.0594004 0.982075 0.358593
>IDL> seed = 2 & print, randomu(seed, 3)
> 0.831999 0.303037 0.506712
What exact version and machine are you using? On Solaris under IDL V5.2,
RANDOMU works as described above for V5.0.3. I suppose that those of us who
use the RANDOMU function will have explicitly watch for specific IDL versions
and machines (like we used to have to do for BYTEORDER). Sigh...
IDL> print,!VERSION
{ sparc sunos unix 5.2 Oct 30 1998}
IDL> seed = 2 & print,randomu(seed,3)
0.342299 0.402381 0.307838
IDL> seed = 2 & print,randomu(seed,3)
0.342299 0.402381 0.307838
--Wayne Landsman landsman@mpb.gsfc.nasa.gov
Subject: Re: RANDOMU bug (and HTML help)
From: thompson@orpheus.nascom.nasa.gov (William Thompson)
Date: 23 Feb 1999 23:43:01 GMT
ajschmitt@my-dejanews.com writes:
>In article <21FEB199900213357@stars.gsfc.nasa.gov>,
> landsman@stars.gsfc.nasa.gov (Wayne Landsman) wrote:
>> Versions 5.1.1 and 5.2 of IDL have a bug in the RANDOMU (and RANDOMN) function
>> such that the SEED variable is initialized to the same value at the start of
>> each session, rather than being intialized by the system clock (see example
>> below). I believe that so long as one as one stays within one IDL session
>> that this causes no problems, but if one is, say, combining Monte Carlo
>> simulations from different IDL sessions, then the results will be decidedly
>> unrandom.
>>
>> RSI knows about this problem and say that they are giving it very high
>> priority for a fix.
>>
>This is not the worst of it. For some of us, it is important to be able
>to have the same seed so that the same "random" sequence is produced.
>However, RSI quietly changed the manner in which the SEED variable
>interacts with the RANDOM functions between version 5.0 & 5.1.
>Consider the following behavior from IDL v 5..0.3:
>IDL> seed = 2 & print, randomu(seed, 3)
> 0.342299 0.402381 0.307838
>...doing this multiple times will always give the same result.
>However, in IDL v.5..1 & later,. using this several times in a row
>produces different results each time.:
>IDL> seed = 2 & print, randomu(seed, 3)
> 0.0594004 0.982075 0.358593
>IDL> seed = 2 & print, randomu(seed, 3)
> 0.831999 0.303037 0.506712
>...etcetera. It turns out that you now have to specify a NEGATIVE seed
>in order for it to have any influence on the generated sequence:
>seed = -2 & print, randomu(seed, 3)
> 0.342299 0.402381 0.307838
>seed = -2 & print, randomu(seed, 3)
> 0.342299 0.402381 0.307838
(rest deleted)
That's interesting. When I try this in IDL/v5.1.1, I get the same result over
and over again, even with positive seeds.
IDL> seed = 2 & print,randomu(seed,3)
0.342299 0.402381 0.307838
IDL> seed = 2 & print,randomu(seed,3)
0.342299 0.402381 0.307838
IDL> print,!version
{ alpha OSF unix 5.1.1 Jul 20 1998}
The online help, though, still states that one is supposed to put in a negative
number to re-use a seed. In fact, the online help in version 5.1.1 is rather
misleading in that it implies that the seed is a scalar value, whereas it's
actually returned as a 36-element array just like in 5.2. The 5.2 online help
does away with any mention of needing to put in negative values.
It looks like the business of needing to put in a negative number to force seed
initialization is restricted to version 5.1.0. Maybe fixing that problem is
what introduced the other problem?
Here's another weird behavior of RANDOMU. If you call IDL with an undefined
seed, it's supposed to initialize the seed for you. Thus, when you type in
IDL> print,randomu(seed,3)
0.653919 0.0668422 0.722660
before seed is defined to anything, it still works. Naively, I expected that
if one deleted the seed variable, that would force IDL to re-initialize the
seed, based on the system time or whatever. However, if my next command is
IDL> delvar,seed & print,randomu(seed,3)
I don't get any random numbers at all. To start getting random numbers again,
I have to use a completely new variable name for the seed, e.g.
IDL> print,randomu(seed2,3)
0.671149 0.383416 0.631635
I don't know if this is properly speaking a bug or not, but it's certainly
weird. As far as I can tell, this behavior is seen in all versions of IDL.
Bill Thompson
Subject: Re: RANDOMU bug (and HTML help)
From: ajschmitt@my-dejanews.com
Date: Wed, 24 Feb 1999 17:57:16 GMT
In article <7avee5$ct8@post.gsfc.nasa.gov>,
thompson@orpheus.nascom.nasa.gov (William Thompson) wrote:
> ajschmitt@my-dejanews.com writes:
(snip)
.
>
> >Consider the following behavior from IDL v 5..0.3:
> >IDL> seed = 2 & print, randomu(seed, 3)
> > 0.342299 0.402381 0.307838
> >...doing this multiple times will always give the same result.
>
> >However, in IDL v.5..1 & later,. using this several times in a row
> >produces different results each time.:
> >IDL> seed = 2 & print, randomu(seed, 3)
> > 0.0594004 0.982075 0.358593
> >IDL> seed = 2 & print, randomu(seed, 3)
> > 0.831999 0.303037 0.506712
>
> >...etcetera. It turns out that you now have to specify a NEGATIVE seed
> >in order for it to have any influence on the generated sequence:
>
> >seed = -2 & print, randomu(seed, 3)
> > 0.342299 0.402381 0.307838
> >seed = -2 & print, randomu(seed, 3)
> > 0.342299 0.402381 0.307838
>
> (rest deleted)
>
> That's interesting. When I try this in IDL/v5.1.1, I get the same result over
> and over again, even with positive seeds.
>
> IDL> seed = 2 & print,randomu(seed,3)
> 0.342299 0.402381 0.307838
> IDL> seed = 2 & print,randomu(seed,3)
> 0.342299 0.402381 0.307838
> IDL> print,!version
> { alpha OSF unix 5.1.1 Jul 20 1998}
>
> The online help, though, still states that one is supposed to put in a negative
> number to re-use a seed. In fact, the online help in version 5.1.1 is rather
> misleading in that it implies that the seed is a scalar value, whereas it's
> actually returned as a 36-element array just like in 5.2. The 5.2 online help
> does away with any mention of needing to put in negative values.
>
> It looks like the business of needing to put in a negative number to force seed
> initialization is restricted to version 5.1.0. Maybe fixing that problem is
> what introduced the other problem?
I agree, this behavior DOES appear to be restricted to version 5.1.0; here's
the version string from the IDL version I was running when I encountere the
problem:
IDL>print,!version
{ sparc sunos unix 5.1 Apr 13 1998}
I had made the statement that the bug occurred in idl v.5.1 and later, but
hadn't actually tested it in a later version. Silly me, I had assumed that if
RSI had bothered to change the documentation (however quietly), that they had
really changed the behavior of the SEED for good.
>
> Here's another weird behavior of RANDOMU. If you call IDL with an undefined
> seed, it's supposed to initialize the seed for you. Thus, when you type in
>
> IDL> print,randomu(seed,3)
> 0.653919 0.0668422 0.722660
>
> before seed is defined to anything, it still works. Naively, I expected that
> if one deleted the seed variable, that would force IDL to re-initialize the
> seed, based on the system time or whatever. However, if my next command is
>
> IDL> delvar,seed & print,randomu(seed,3)
>
> I don't get any random numbers at all. To start getting random numbers again,
> I have to use a completely new variable name for the seed, e.g.
>
Strictly speaking, this isn't a bug, but a documented side effect of DELVAR
(unless they changed that TOO since my last copy of printed documentation):
"Each time DELVAR is called, the main program is erased. Variables that are
not deleted remain unchanged."
Thus, after the use of delvar in the first statement, the second statement is
"erased". Calling the two in sequence SHOULD give the expected result:
IDL> delvar, seed
IDL> print, randomu(seed,3)
0.860873 0.661931 0.766831
-Andy Schmitt
-----------== Posted via Deja News, The Discussion Network ==----------
http://www.dejanews.com/ Search, Read, Discuss, or Start Your Own
[Return to Table of Contents]
Copyright © 1999 David W. Fanning
Last Updated 3 March 1999