number point under-flow

Hi,

yes, I did compile it with gcc 2.96 ... . Do you've an estimate on how "bad"
this error is, e.g. how much it effects the calculations in R?

	kind regards,

	Arne
-----Original Message-----
From: Roger D. Peng [mailto:rpeng at jhsph.edu]
Sent: 04 February 2004 14:49
To: Muller, Arne PH/FR
Cc: r-help at stat.math.ethz.ch
Subject: Re: [R] number point under-flow

Did you compile with gcc-2.96?  I think there were some 
problems with the floating point arithmetic with that 
compiler (at least for the earlier versions released by Red 
Hat).

-roger

Arne.Muller at aventis.com wrote:
Hello,

I've come across the following situation in R-1.8.1 
(compile + running under
RedHat 7.1):

phyper(24, 514, 5961-514, 53, lower.tail=T)
[1] 1

phyper(24, 514, 5961-514, 53, lower.tail=F)
[1] -1.037310e-11

I'd expect the later to be 0 or some very small positive 
number. Is this a
number under-flow of the calculation? Do you think I'm safe 
if I just set the
result to 0 in these cases?

	kind regards,

	Arne

______________________________________________
R-help at stat.math.ethz.ch mailing list
https://www.stat.math.ethz.ch/mailman/listinfo/r-help
PLEASE do read the posting guide! 
http://www.R-project.org/posting-guide.html

"ArneM" ==   <Arne.Muller at aventis.com>
    on Wed, 4 Feb 2004 14:53:11 +0100 writes:
ArneM> Hi, yes, I did compile it with gcc 2.96 ... . Do
    ArneM> you've an estimate on how "bad" this error is,
    ArneM> e.g. how much it effects the calculations in R?

In my memory, the effects are bad enough to very quickly drop
such a version of R and get one which is correctly compiled.

Even if only 1 in 10000 computations go wrong; you will hardly
notice and you have chance of knowing which of your final
results will be inaccurate by how much.

This has been on this mailing list and in many installation
instructions. "gcc 2.96" has never been an official release of
gcc and the fact that Redhat had bundled it with one of their
distributions was probably the biggest mistake they've ever
made.

Google for "gcc 2.96" and you will find
       http://gcc.gnu.org/gcc-2.96.html
and much more variations on this theme.

Current is gcc 3.3.2, released last October, but gcc 3.2.x might
be acceptable too (but 3.2.1 and 3.2.2 are broken on Solaris).
Martin Maechler <maechler at stat.math.ethz.ch> writes:
"ArneM" ==   <Arne.Muller at aventis.com>
    on Wed, 4 Feb 2004 14:53:11 +0100 writes:
    ArneM> Hi, yes, I did compile it with gcc 2.96 ... . Do
    ArneM> you've an estimate on how "bad" this error is,
    ArneM> e.g. how much it effects the calculations in R?

In my memory, the effects are bad enough to very quickly drop
such a version of R and get one which is correctly compiled.

Even if only 1 in 10000 computations go wrong; you will hardly
notice and you have chance of knowing which of your final
results will be inaccurate by how much.
Correct me if I'm wrong, but I don't think the compiler is the only
problem here. I got negative phyper values with Martyn's RPM on RH8.0,
and that appears to have been done with GCC 3.2. The devel version, on
the same machine, gets it right (well, gets a positive result...) and
I do have a vague recollection of seeing a variation of this before, so
I suspect there was a bug and someone fixed it.

Possibly, this records the fix:

    o   [l]choose() use a more accurate formula which also slightly
        improves p- and qhyper(); choose(n, k) now returns 0 instead
        of NaN for k < 0 or > n.

If so, then r-patched should fix it too.
O__  ---- Peter Dalgaard             Blegdamsvej 3  
  c/ /'_ --- Dept. of Biostatistics     2200 Cph. N   
 (*) \(*) -- University of Copenhagen   Denmark      Ph: (+45) 35327918
~~~~~~~~~~ - (p.dalgaard at biostat.ku.dk)             FAX: (+45) 35327907