[R] Variance for Vector of Constants is STILL Not Zero

[Thomas Lumley]

On Fri, 17 Feb 2006, Barry Zajdlik wrote:
> Hello all,
>
> Thanks for the responses but I am still annoyed by this seemingly simple
> problem; I recorded sessionInfo() as below.
>
> x<-rep(0.02,10)
>> var(x)
> [1] 1.337451e-35

Well, yes, it is accurate only to 35 digits and this is a bit unfortunate 
since it could be more accurate in this case. The exact result is probably 
system-dependent, but there has been a bug report for a long time 
(PR#1228) that var() is not as accurate as it could possibly be, and 
patches would be welcome.  They obviously aren't a very high priority for 
anyone.

> I Changed .Machine$double.eps to make the calculations LESS accurate.
> My thought was that if I reduced the precision, 1-eps would return 1
> instead of some number less than 1.  My thought was that if eps were
> sufficiently large my sample problem would return a zero.  This didn't
> happen though.
>

.Machine is a set of constants that describe your hardware and C 
implementation.  They are right when you start. If you change them all 
that happens is that they become wrong -- R doesn't go in and re-engineer 
the floating point hardware.

 	-thomas