R devel - Dec 2001 - behavior of r* and d* functions at boundaries (PR#1218)

(Sent this to r-help back in October, got no comments, forgot
to re-submit it as a bug report.)

There are a few inconsistencies, at least, in some of the functions that 
generate random deviates from particular distributions (I think they're 
bugs because they're inconvenient, but maybe someone can make an argument 
for the current behavior). If people think these are really bugs I can 
submit them, together or separately. 

1. rlnorm(n,mean,sd) gives NaN for sd=0, rather than always returning 
the mean. This is inconsistent with rnorm(), which returns the mean. 

2. rnbinom(n,prob,size) gives NaN when prob=1; always returning 0 would be 
reasonable, and consistent with the output of dnbinom(). [This is 
equivalent to rnbinom(n,mu=0,size) in the new parameterization.] 

3. rgamma(n,shape,scale) happily returns 0 when shape=0 or scale=0. However, 
dgamma(x,shape=0) [and pgamma(x,shape=0)] gives NaN. This behavior is 
actually convenient for me, but I could make an argument for returning a 
warning in this case. 

  In general, I don't know what (e.g.) dnorm(m,mean=m,sd=0)
should return: NaN (as at present) or Inf?  I guess it depends
how fussy you're being and how you're taking limits.

  The fixes seem straightforward. 

  Ben Bolker 

Version:
 platform = i686-pc-linux-gnu
 arch = i686
 os = linux-gnu
 system = i686, linux-gnu
 status = 
 major = 1
 minor = 3.1
 year = 2001
 month = 08
 day = 31
 language = R

Search Path:
 .GlobalEnv, package:ctest, Autoloads, package:base

-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-
r-devel mailing list -- Read http://www.ci.tuwien.ac.at/~hornik/R/R-FAQ.html
Send "info", "help", or "[un]subscribe"
(in the "body", not the subject !)  To:
r-devel-request@stat.math.ethz.ch
_._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._

>>>>> "Ben" == Ben Bolker <bolker@zoo.ufl.edu>
writes:
    Ben>   (Sent this to r-help back in October, got no
    Ben> comments, forgot to re-submit it as a bug report.)

    Ben> There are a few inconsistencies, at least, in some of
    Ben> the functions that generate random deviates from
    Ben> particular distributions (I think they're bugs because
    Ben> they're inconvenient, but maybe someone can make an
    Ben> argument for the current behavior). If people think
    Ben> these are really bugs I can submit them, together or
    Ben> separately.

    Ben> 1. rlnorm(n,mean,sd) gives NaN for sd=0, rather than
    Ben> always returning the mean. This is inconsistent with
    Ben> rnorm(), which returns the mean.

    Ben> 2. rnbinom(n,prob,size) gives NaN when prob=1; always
    Ben> returning 0 would be reasonable, and consistent with
    Ben> the output of dnbinom(). [This is equivalent to
    Ben> rnbinom(n,mu=0,size) in the new parameterization.]

these two above are fixed for 1.4

    Ben> 3. rgamma(n,shape,scale) happily returns 0 when shape=0
    Ben> or scale=0. However, dgamma(x,shape=0) [and
    Ben> pgamma(x,shape=0)] gives NaN. This behavior is actually
    Ben> convenient for me, but I could make an argument for
    Ben> returning a warning in this case.

    Ben>   In general, I don't know what (e.g.)
    Ben> dnorm(m,mean=m,sd=0) should return: NaN (as at present)
    Ben> or Inf?

    Ben> I guess it depends how fussy you're being and how
    Ben> you're taking limits.

My approach would be to always return the limit if it's well
defined (as here).  
For any random variate X, when Var(X) = 0, r<dist>()  should
return E(X) (if that is defined) I think.  
This would mean that rgamma() is fine and dnorm() could be
improved.  
dgamma(0,shape=0) should give Inf as well and pgamma(0,shape=0) |-> 1.
The fact that pnorm(mu, mu, sd = 0) currently gives NaN instead
of 1 is a bug IMO.

Another line of reasoning could be that "shape=0" (or less
clearly "sd = 0") are not allowed (as per ?rgamma), and giving
NaN would reflect this.  I'm not share this view
but leave all this for after release.

    Ben>   The fixes seem straightforward.
(yes, at least for 1) and 2))

Thank you, Ben, for reporting these!

Martin Maechler <maechler@stat.math.ethz.ch>
http://stat.ethz.ch/~maechler/
Seminar fuer Statistik, ETH-Zentrum  LEO C16	Leonhardstr. 27
ETH (Federal Inst. Technology)	8092 Zurich	SWITZERLAND
phone: x-41-1-632-3408		fax: ...-1228			<><
-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-
r-devel mailing list -- Read http://www.ci.tuwien.ac.at/~hornik/R/R-FAQ.html
Send "info", "help", or "[un]subscribe"
(in the "body", not the subject !)  To:
r-devel-request@stat.math.ethz.ch
_._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._