Integer overflow in R

Question

I’m working on a population dynamics simulation that involves generating whole numbers. Due to the hypotheses of my model, I am generating random numbers through the function rmultinom R. However, I am having overflow problems in my simulation.

The largest integer R can generate is defined by .Machine$integer.max. In my PC, this number is equal to 2147483648. In scientific notation, this value is equivalent to approximately 2,147 x 10 9.

But many of the simulations I run exceed this limit. The parameter size of rmultinom will be greater than 10 (10) or even 10 (12). And if so, I cannot generate random numbers with the distribution I desire.

What could I do to solve this problem? Any suggestions?

Answer 1

From what I understand, what the rmultinom do, it’s like this:

• You own lenght(prob) different types of balls, where prob is the paretero prob of function.

• Then you will withdraw size balls independently, according to these probabilities. (size is the parameter size of function)

• This procedure will be repeated n, times. (n is p parameter n of function)

Given this scenario I imagine you can use sizes larger than what fits in the R, as follows:

A possible function that takes samples with the maximum size of the R

library(bit64)
rmultinom2 <- function(size, prob){

  n <- size %/% 1000000000L
  resto <- size %% 1000000000L

  amostra <- rmultinom(n = as.integer(n), size = 1000000000L, prob = prob)
  amostra_resto <- rmultinom(n = 1, size = resto, prob = prob)

  return(rowSums(cbind(amostra, amostra_resto)))
}

Repeating experiment 100x

amostra <- plyr::ldply(1:100, function(x, size, prob) {
  rmultinom2(size, prob)
}, size = as.integer64("10000000000"), prob = c(1,2,4,5)
)

I think the big balcony here is to use the package bit64 which supports larger integers and take several samples independently to then add up. It is possible that it is also necessary to transform the lines of the amostra (inside the function) in large integers so that the sum also does not explode.

Now, if lenght(prob) is bigger than the whole largest, I don’t know.