runif limited precision

runif appears to give 31 bits of precision, but this isn't mentioned
in the documentation page. The R numeric type supports 53 digits of
precision, and other numeric functions (sin, etc.) give full-precision
results.  So I'd think that either runif should give full precision or
its documentation should mention this limitation.

#integers
table(runif(10000,-2^30,2^30) %% 1)
   0  0.5
4972 5028

#natural numbers

table(runif(10000,0,2^31) %% 1)

   0  0.5
4956 5044
#fractions
table((r<-runif(10000,0,1)*2^31)-floor(r))
   0  0.5
5021 4979

On 02/01/2009 2:45 PM, Stavros Macrakis wrote:

...So I'd think that either runif should give full precision or
its documentation should mention this limitation.

It refers to the .Random.seed page for details, and that page mentions the
limitation to at most 2^32 different values for most of the generators.

On Fri, Jan 2, 2009 at 4:03 PM, Duncan Murdoch <murdoch at stats.uwo.ca> wrote:

On 02/01/2009 2:45 PM, Stavros Macrakis wrote:

...So I'd think that either runif should give full precision or
its documentation should mention this limitation.

It refers to the .Random.seed page for details, and that page mentions the
limitation to at most 2^32 different values for most of the generators.

Indeed it does -- thanks for the pointer!  I still think it would be
helpful to the user to mention the limitation in the individual
r<dist> doc pages, as not many users are likely to work through the
very complete RNG doc page.

-----Original Message-----
From: r-help-bounces at r-project.org 
[mailto:r-help-bounces at r-project.org] On Behalf Of Duncan Murdoch
Sent: Friday, January 02, 2009 2:57 PM
To: Stavros Macrakis
Cc: r-help at r-project.org
Subject: Re: [R] runif limited precision

On 02/01/2009 4:16 PM, Stavros Macrakis wrote:

On Fri, Jan 2, 2009 at 4:03 PM, Duncan Murdoch 

<murdoch at stats.uwo.ca> wrote:

On 02/01/2009 2:45 PM, Stavros Macrakis wrote:

...So I'd think that either runif should give full precision or
its documentation should mention this limitation.

It refers to the .Random.seed page for details, and that 

page mentions the

limitation to at most 2^32 different values for most of 

the generators.

Indeed it does -- thanks for the pointer!  I still think it would be
helpful to the user to mention the limitation in the individual
r<dist> doc pages, as not many users are likely to work through the
very complete RNG doc page.

I don't agree.  If you add too much technical detail to a topic, then 
people don't "work through it".  I'd say the r<dist> pages generally 
give enough detail now, but not too much.  If you add every 
detail that 
might interest someone somewhere, then they'd all be hard to read.

Duncan Murdoch

I don't agree.  If you add too much technical detail to a topic, then people
don't "work through it".  I'd say the r<dist> pages generally give enough
detail now, but not too much.  If you add every detail that might interest
someone somewhere, then they'd all be hard to read.

On Fri, Jan 2, 2009 at 5:56 PM, Duncan Murdoch <murdoch at stats.uwo.ca> wrote:

I don't agree.  If you add too much technical detail to a topic, then people
don't "work through it".  I'd say the r<dist> pages generally give enough
detail now, but not too much.  If you add every detail that might interest
someone somewhere, then they'd all be hard to read.

Duncan, Daniel,

Thanks for your responses.

I agree that it's a delicate balance between too little and too much
information on individual documentation pages -- especially when, as
in this case, a number of different functions (r<dist>) share some
content (RNG).  That said, I find in general that R documentation
pages are somewhat on the "too little" side. In particular, I'd find
it useful if there were more references to synthetic overviews: a link
from the connection page to an overview of I/O; a link from the runif
page to an overview of distributions; etc.