[R] floor of n observations in number generators

[Peter Dalgaard]

"Marcus Davy" <MDavy at hortresearch.co.nz> writes:

> I couldnt find a previous posting on this in the archives, maybe it has
> already been mentioned.
> 
> If you use a calculation to generate n observations in random number
> generators and you don't round to the nearest integer you may be
> generating n-1 numbers not n numbers as you thought depending on the
> storage precision of the calculation. 
> 
> e.g.
> > m <- 1000
> > pi0 <- 0.9
> > length(rnorm(m * (1-pi0)))
> [1] 99  # Should be 100
> > options(digits=16)
> > m * (1-pi0)
> [1] 99.99999999999997
> > identical(m*(1-pi0), 100)
> [1] FALSE
> 
> Random number generation generates the floor of n observations, this
> feature occurs on R-1.8.1 on linux Redhat8, and winXP (also on Unix
> SPlus 3.4) 
> for probably all of the random number generators.

Nothing to do with random number generation, everything to do with
coercion of floats to integers. Hence, e.g.

> as.integer(100*(1-0.9))
[1] 9
> numeric(100*(1-0.9))
[1] 0 0 0 0 0 0 0 0 0

There are a couple of places where we do have hidden fuzz factors
because people were getting bitten by imprecision effects a bit too
easily, like in

> 1:(100*(1-0.9))
 [1]  1  2  3  4  5  6  7  8  9 10

but in general we use the standard C rules and simply truncate. 

-- 
   O__  ---- Peter Dalgaard             Blegdamsvej 3  
  c/ /'_ --- Dept. of Biostatistics     2200 Cph. N   
 (*) \(*) -- University of Copenhagen   Denmark      Ph: (+45) 35327918
~~~~~~~~~~ - (p.dalgaard at biostat.ku.dk)             FAX: (+45) 35327907