[Rd] Precision of function mean,bug?

brodie gaslam brod|e@g@@|@m @end|ng |rom y@hoo@com
Wed May 20 20:09:12 CEST 2020


 > On Wednesday, May 20, 2020, 7:00:09 AM EDT, peter dalgaard <pdalgd using gmail.com> wrote:
>
> Expected, see FAQ 7.31.
>
> You just can't trust == on FP operations. Notice also

Additionally, since you're implementing a "mean" function you are testing
against R's mean, you might want to consider that R uses a two-pass
calculation[1] to reduce floating point precision error.

Best,

Brodie.

[1] https://github.com/wch/r-source/blob/tags/R-4-0-0/src/main/summary.c#L482

> > a2=(z[idx]+x[idx]+y[idx])/3
> > a2==a
> [1] FALSE
> > a2==b
> [1] TRUE
>
> -pd
>
> > On 20 May 2020, at 12:40 , Morgan Morgan <morgan.emailbox using gmail.com> wrote:
> >
> > Hello R-dev,
> >
> > Yesterday, while I was testing the newly implemented function pmean in
> > package kit, I noticed a mismatch in the output of the below R expressions.
> >
> > set.seed(123)
> > n=1e3L
> > idx=5
> > x=rnorm(n)
> > y=rnorm(n)
> > z=rnorm(n)
> > a=(x[idx]+y[idx]+z[idx])/3
> > b=mean(c(x[idx],y[idx],z[idx]))
> > a==b
> > # [1] FALSE
> >
> > For idx= 1, 2, 3, 4 the last line is equal to TRUE. For 5, 6 and many
> > others the difference is small but still.
> > Is that expected or is it a bug?



More information about the R-devel mailing list