I noticed the following when running gmp from R :
Rgames> log2(Inf)
[1] Inf
Rgames> log2(as.bigz(Inf))
[1] 8000
What's happening here? Or does the gmp language (or gmp::log2.bigz method) just not support the concept of "Inf" ?
EDIT: Josh's comment is correct: it's not the log2 function per se but rather the fact that as.bigz appears to convert Inf to 2^8000 as a bigz integer. Changed title to match.
BTW, there's also this:
Rgames> log2(-Inf)
[1] NaN
Warning message:
NaNs produced
Rgames> log2(as.bigz(-Inf))
[1] NaN
EDIT: that was stupid. log2(negative_anything) is NaN
But,
Rgames> log2(-as.bigz(-Inf))
[1] 8000
Fits with Josh's answer.
Posting Josh's detective work for him, per request:
If you download the sources for gmp, and search for "Inf" in $gmp-HOME$/src/bigintegerR.cc, you'll find the following lines:
/// New: numeric '+- Inf' give +- "Large" instead of NA
and a bit later
else { // dj is +- Inf : use LARGE ( = +- 2 ^ 80000 -- arbitrarily )
My guess is that since the GMP library doesn't have a concept/representation of Inf, the R package authors had to make some decision about how to pass on user-supplied Inf values to the external library. It looks like they used to convert Inf to NA, and now they convert it to 2^80000.