[Rd] Kinderman-Ramage (PR#2846)

[Prof Brian Ripley]

Unfortunately it was the default prior to 1.7.0, so for reproducibility I
think we can't change this.  We could add a new option, a corrected 
version, but is it worth it?  I would definitely add a warning if the 
buggy version was selected, and perhaps make it an error to select it via 
RNGkind (rather than RNGversion or a saved .Random.seed).

On 25 Apr 2003, Peter Dalgaard BSA wrote:

> leydold at statistik.wu-wien.ac.at writes:
> 
> > Hi,
> > 
> > Our department has detected a bug in the implementation of the
> > Kinderman-Ramage generator for normal random variates in version
> > 1.7.0, which can be seen from the below R session.
> > (Consecutive calls for chisq.test(...) always gives p-values very
> > close to 0.)
> > We have already encountered this bug in version 1.6.2
> > 
> > The error is in file 
> > R-1.7.0/src/nmath/snorm.c
> > 
> > Here is a patch for this file to correct the error
> > (the added two lines are crucial):
> 
> Thanks. Both look like transcription errors. What was the source of
> the corrected constants? The JASA paper cited in snorm.c? 
> 
> (Another way of seeing the effect is
> 
>  m <- sapply(1:50,function(i)table(trunc(100*pnorm(rnorm(1e6)))))
>  matplot(m,type="l")
> 
> Two small, but clear, spikes at around 30 and 70 can be seen. The
> effect is not seen with the Inversion method, which is the default in
> R-1.7.0.) 
> 
> 
> 

-- 
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