Path: csiph.com!x330-a1.tempe.blueboxinc.net!newsfeed.hal-mli.net!feeder1.hal-mli.net!weretis.net!feeder4.news.weretis.net!news.musoftware.de!wum.musoftware.de!fu-berlin.de!uni-berlin.de!news.dfncis.de!not-for-mail
From: =?ISO-8859-1?Q?Hans-Bernhard_Br=F6ker?= <HBBroeker@t-online.de>
Newsgroups: comp.graphics.apps.gnuplot
Subject: Re: trying to get a fit to work.
Date: Thu, 09 Jun 2011 22:26:25 +0200
Lines: 33
Message-ID: <95coieFepcU1@mid.dfncis.de>
References: <62b76acf-a50c-49f2-a02a-908398651e5b@q12g2000prb.googlegroups.com> <ispja4$fmf$1@dont-email.me>
Mime-Version: 1.0
Content-Type: text/plain; charset=ISO-8859-1; format=flowed
Content-Transfer-Encoding: 7bit
X-Trace: news.dfncis.de rkLtCEIC1Rh2FFih1FZhJgJ2AT7Szwrlo+menfqe79f89wOx66WbGvqxz3
Cancel-Lock: sha1:6zwvKE4lTFgM7EBf2pYuXIigFEY=
User-Agent: Mozilla/5.0 (Windows; U; Windows NT 6.1; de; rv:1.9.2.15) Gecko/20110303 Thunderbird/3.1.9
In-Reply-To: <ispja4$fmf$1@dont-email.me>
Xref: x330-a1.tempe.blueboxinc.net comp.graphics.apps.gnuplot:366

On 09.06.2011 06:48, sfeam wrote:

> Neither OS underflows with exp(-x) - both go to zero.

Actually the OS would overflow.  We just don't let it.  See gp_exp() and 
its treatment of E_MINEXP.

> I think this means there's something wrong with the way erfc works in
> windows, other than basic underflow handling. So an even simpler fix uses
> the definition of erfc:
>
> myerfc(x) = 1 - erf(x)

Unfortunately that's really to simple to be a valid fix.  This loses 
accuracy _way_ earlier than the actual erfc() does:

gnuplot> print erfc(27)
5.23704643935263e-319 

gnuplot> print 1-erf(7)
0.0

[Using our Windows binary of 4.4 pl2]

> I am guessing that you are using a machine that has configured in the
> internal version, and there is an overflow bug in it somewhere.

That's no bug, it's a deliberate, reasonably error reaction.  It's a 
bona fide floating point exception being raised by the implementation, 
which under the usual roles of standard C library routines is apparently 
_supposed_ to send out a floating point exception.  That's the same 
reason implementations started to return "undefined" for 
exp(-largenumber), one after the other, a while ago.