Poisson sample() hangs when lambda is close to max of the float type.

[WarrenWeckesser]

The follow programs appear to hang indefinitely in the call poisson.sample(&mut rng):

f32

use rand_distr::{Distribution, Poisson};

fn main() {
    let lambda = 1.0e37f32;
    let poisson = Poisson::<f32>::new(lambda).unwrap();
    let mut rng = rand::thread_rng();
    let k = poisson.sample(&mut rng);
    println!("{}", k);
}

f64

use rand_distr::{Distribution, Poisson};

fn main() {
    let lambda = 1.0e306f64;
    let poisson = Poisson::<f64>::new(lambda).unwrap();
    let mut rng = rand::thread_rng();
    let k = poisson.sample(&mut rng);
    println!("{}", k);
}

I ran into this issue while testing gh-1296.

[dhardy]

Related to #1290?

[WarrenWeckesser]

The problem is that when $\lambda$ is sufficiently large, the value of magic_val computed here:

rand/rand_distr/src/poisson.rs

Line 84 in b593db6

magic_val: lambda * log_lambda - crate::utils::log_gamma(F::one() + lambda),

will be nan, because both terms in that expression will overflow to inf, and inf - inf gives nan.

Then down at

rand/rand_distr/src/poisson.rs

Lines 141 to 144 in b593db6

	// check with uniform random value - if below the threshold, we are within the target distribution
	if rng.gen::<F>() <= check {
	break;
	}

check will be nan, so the comparison in the if statement will be false, and the loop will never break.

When $\lambda$ is that large, terms result * self.log_lambda and crate::utils::log_gamma(F::one() + result) from

rand/rand_distr/src/poisson.rs

Lines 134 to 139 in b593db6

	let check = F::from(0.9).unwrap()
	* (F::one() + comp_dev * comp_dev)
	* (result * self.log_lambda
	- crate::utils::log_gamma(F::one() + result)
	- self.magic_val)
	.exp();

can also overflow, because result will be $\lambda + c\sqrt{\lambda}$, where $c$ is not large. ($\sqrt{\lambda}$ is the standard deviation of the Poisson distribution.)

There is another problem with large $\lambda$, even if it isn't large enough to cause overflow: the terms being subtracted in the expressions shown above will be very close, resulting in significant loss of precision.

I've looked into modifying the calculation when $\lambda$ is large to use an asymptotic version of the check expression, but I don't have anything useful yet.

A different approach to avoiding the problem would be switch to a normal distribution when $\lambda$ is sufficiently large. Convential wisdom is that the normal distribution is a good approximation to the Poisson distribution when $\lambda$ is large, but it would be nice to have a more quantitative result on the quality of the approximation. (John D. Cook has a short blog post about it.)