Note that the R function rpois() always returns an integer vector and thus
commonly runs into problems when lambda is near or above 2^31 (the
smallest positive integral double that cannot represented as a 32-bit signed
integer).
> set.seed(1)
> rpois(10, 2^31)
[1] 2147454617 NA NA NA 2147412285 NA
NA
[8] NA 2147469496 2147446621
Warning message:
In rpois(10, 2^31) : NAs produced
> storage.mode(.Last.value)
[1] "integer"
Bill Dunlap
TIBCO Software
wdunlap tibco.com
On Sat, Apr 25, 2015 at 9:33 AM, Michael Weylandt <
michael.weyla...@gmail.com> wrote:
>
> On Apr 24, 2015, at 23:58, Christian Gunning <x...@unm.edu> wrote:
>
> Quick question, mainly out of curiousity. I get that Rcpp uses the R api
> on the backend, and in the R api, R::rpois returns a double (and
> Rcpp::rpois returns NumericVector). The R C code doesn't offer much in the
> way of explanation of why double is returned, but always returns a floor() (
> https://svn.r-project.org/R/trunk/src/nmath/rpois.c).
>
> It seems to me that, in this case, upstream could (should?) cast to int.
> Since it doesn't, downstream users like Rcpp and me must choose whether to
> follow upstream convention or to cast. In this case, is the "best
> practice" to cast-to-int right away, or to follow along and just use
> doubles for everything?
>
> I've run into this recently with other RNG code that returns doubles for
> everything, and I'm wondering if there's an obvious tradeoff that I'm
> missing, like speed versus type-correctness? If anyone has any insights,
> I'd love to hear them.
>
>
> The set of integers exactly representable by a double is a superset of
> those which can represented by a 32bit int.
>
> http://stackoverflow.com/questions/759201/representing-integers-in-doubles
>
> _______________________________________________
> Rcpp-devel mailing list
> Rcpp-devel@lists.r-forge.r-project.org
> https://lists.r-forge.r-project.org/cgi-bin/mailman/listinfo/rcpp-devel
>
_______________________________________________
Rcpp-devel mailing list
Rcpp-devel@lists.r-forge.r-project.org
https://lists.r-forge.r-project.org/cgi-bin/mailman/listinfo/rcpp-devel
