R devel - Oct 2009 - Accuracy (PR#13999)

Full_Name: Viktor Witkovsky
Version: 2.9.2
OS: Windows XP
Submission from: (NULL) (78.98.89.227)


Hello,

I have found strange behavior of the function qchisq (the non-central qchisq is
based on inversion of pchisq, which is further based on pgamma). The function
gives wrong results without any warning. For example:

qchisq(1e-12,1,8.94^2,lower.tail=FALSE) gives 255.1840972465858 (notice that
here the correct value should be 255.1841334848075),
but
qchisq(1e-12,1,8.95^2,lower.tail=FALSE) gives 1249.681320136836
Here, the correct value should be 255.5037231613135.

So, it seems that qchisq is inaccurate for small probability values and larger
non/centrality parameter.


I am using the precompiled binary version of R, under Windows XP.

               _                           
platform       i386-pc-mingw32             
arch           i386                        
os             mingw32                     
system         i386, mingw32               
status                                     
major          2                           
minor          9.2                         
year           2009                        
month          08                          
day            24                          
svn rev        49384                       
language       R                           
version.string R version 2.9.2 (2009-08-24)

witkovsky wrote:
> 
> Full_Name: Viktor Witkovsky
> Version: 2.9.2
> OS: Windows XP
> Submission from: (NULL) (78.98.89.227)
> 
> 
> Hello,
> 
> I have found strange behavior of the function qchisq (the non-central
> qchisq is
> based on inversion of pchisq, which is further based on pgamma). The
> function
> gives wrong results without any warning. For example:
> 
> qchisq(1e-12,1,8.94^2,lower.tail=FALSE) gives 255.1840972465858 (notice
> that
> here the correct value should be 255.1841334848075),
> but
> qchisq(1e-12,1,8.95^2,lower.tail=FALSE) gives 1249.681320136836
> Here, the correct value should be 255.5037231613135.
> 
> So, it seems that qchisq is inaccurate for small probability values and
> larger
> non/centrality parameter.
> 
> 
> I am using the precompiled binary version of R, under Windows XP.
> 
>                _                           
> platform       i386-pc-mingw32             
> arch           i386                        
> os             mingw32                     
> system         i386, mingw32               
> status                                     
> major          2                           
> minor          9.2                         
> year           2009                        
> month          08                          
> day            24                          
> svn rev        49384                       
> language       R                           
> version.string R version 2.9.2 (2009-08-24)
> 
> ______________________________________________
> R-devel at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-devel
> 
> 
More entertainingly (and perhaps enlighteningly?)
> curve(qchisq(1e-12,1,x^2,lower.tail=FALSE),from=0.1,to=50)
 Of course, I haven't taken the time to dig through and see
what's happening :-(
> sessionInfo()
R version 2.9.2 (2009-08-24) 
i486-pc-linux-gnu 

locale:
LC_CTYPE=en_US.UTF-8;LC_NUMERIC=C;LC_TIME=en_US.UTF-8;LC_COLLATE=en_US.UTF-8;LC_MONETARY=C;LC_MESSAGES=en_US.UTF-8;LC_PAPER=en_US.UTF-8;LC_NAME=C;LC_ADDRESS=C;LC_TELEPHONE=C;LC_MEASUREMENT=en_US.UTF-8;LC_IDENTIFICATION=C

attached base packages:
[1] stats     graphics  grDevices utils     datasets  grid      methods  
[8] base     

other attached packages:
[1] reshape_0.8.3 plyr_0.1.9    proto_0.3-8  

loaded via a namespace (and not attached):
[1] ggplot2_0.8.4

-- 
View this message in context:
http://www.nabble.com/Accuracy-%28PR-13999%29-tp25843737p25845807.html
Sent from the R devel mailing list archive at Nabble.com.

You crossed the value ncp = 80: the help page warned you of 
cancellation there, and at p = 1 - 1e-12, that is what you got 
(badly).

On Sun, 11 Oct 2009, witkovsky at savba.sk wrote:
> Full_Name: Viktor Witkovsky
> Version: 2.9.2
> OS: Windows XP
> Submission from: (NULL) (78.98.89.227)
>
>
> Hello,
>
> I have found strange behavior of the function qchisq (the non-central
qchisq is
> based on inversion of pchisq, which is further based on pgamma). The
function
> gives wrong results without any warning. For example:
>
> qchisq(1e-12,1,8.94^2,lower.tail=FALSE) gives 255.1840972465858 (notice
that
> here the correct value should be 255.1841334848075),
> but
> qchisq(1e-12,1,8.95^2,lower.tail=FALSE) gives 1249.681320136836
> Here, the correct value should be 255.5037231613135.
>
> So, it seems that qchisq is inaccurate for small probability values and
larger
> non/centrality parameter.
Actually, for probability values near 1, and it does say so in the 
documentation.  But see
> qchisq(1e-12, 1, 81, lower.tail=FALSE)
[1] 1258.412
> qchisq(1-1e-12, 1, 81)
[1] 257.1488

which I suggest gives you a workaround -- we'll look into giving a 
warning from the code.
>
> I am using the precompiled binary version of R, under Windows XP.
>
>               _
> platform       i386-pc-mingw32
> arch           i386
> os             mingw32
> system         i386, mingw32
> status
> major          2
> minor          9.2
> year           2009
> month          08
> day            24
> svn rev        49384
> language       R
> version.string R version 2.9.2 (2009-08-24)
>
> ______________________________________________
> R-devel at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-devel
>
-- 
Brian D. Ripley,                  ripley at stats.ox.ac.uk
Professor of Applied Statistics,  http://www.stats.ox.ac.uk/~ripley/
University of Oxford,             Tel:  +44 1865 272861 (self)
1 South Parks Road,                     +44 1865 272866 (PA)
Oxford OX1 3TG, UK                Fax:  +44 1865 272595