[Rd] rounding errors in max.col()

[ripley@stats.ox.ac.uk]

On Fri, 3 Jan 2003, Philippe Grosjean wrote:

> Prof. Brian Ripley wrote:
>
> >Do read the help page:
>
> >     Ties are broken at random.  The determination of ``tie'' assumes
> >     that the entries are probabilities.
>
> >and then don't blame your tools when you misuse them ....
>
> Yes, I did read this. So, I know why items are randomly selected in case of
> `ties'.
> But then, I reformulate the question: what are the criteria for deciding two
> values are `ties' in max.col?
> I know, the answer is certainly somewhere (at least in the code), however,
> if someone knows it, it would help save time.

In maxcol.c, of course.

#define RELTOL 1e-5
            tol = RELTOL * fmax2(fabs(a), fabs(b));


> Thanks,
>
> Philippe Grosjean
>
>
> On Thu, 2 Jan 2003, Philippe Grosjean wrote:
>
> > I suppose this is a general behavior with external function calls, so I do
> > not post (yet) a specific bug report. Could someone explain this?
> >
> > a <- rep(1, 20) + rnorm(20, mean=0.00001, sd=0.0001)
> > b <- embed(a, 3)
> > # I want to know where the item in column 2 is greated than both col 1 and
> 3
> > (peak)
> > test1 <- max.col(b) == 2
> > # ... or I could use a less optimal code
> > test2 <- apply(b, 1, max) == b[, 2]
> > any(test1 != test2)	# both are equivalent
> >
> > # but when numbers are very close
> > a <- rep(1, 20) + rnorm(20, mean=0.0000001, sd=0.0000001)
> > b <- embed(a, 3)
> > test1 <- max.col(b) == 2
> > test2 <- apply(b, 1, max) == b[, 2]
> > any(test1 != test2)	# tests are now DIFFERENT!
> > # Indeed, test2 is correct, and test1 suffers from wrong calculations due
> > probably to rounding errors in max.col()
>
> No, to max.col working as documented.
>
> [Large waste of bandwidth deleted]
>
> --
> Brian D. Ripley,                  ripley@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
>
> ______________________________________________
> R-devel@stat.math.ethz.ch mailing list
> http://www.stat.math.ethz.ch/mailman/listinfo/r-devel
>
>
>
>

-- 
Brian D. Ripley,                  ripley@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