Is there a point at which the statistical properties of the random number generator will start to influence the results of Monte Carlo simulation?

I have a scenario where I need to calculate the distribution functions of the $i^{th}$ order statistic with non-identically distributed random variables. They all follow the exponential distribution, but they have different rate parameters. Rather than trying to compute these probabilities analytically, I think it would be better to approximate them through simulation. Eventually the goal is to maximize an expected value for a certain set of parameters. I will be doing this through some numerical optimization approximation method. This could potentially result in hundreds of iterations each with a thousand or so random variables to approximate the probabilities.

Should I be concerned that the entropy generated through standard language libraries (like python's numpy.random which uses the Mersenne Twister PRG) is not enough for the quantity of random variables I may need to sample? More generally, at what point might it make sense to call the operating system's /dev/random or /dev/urandom for random number generation?



Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.