Hello,

In preparation for a talk I'm going to give, I wanted to estimate how  
good 128 bits MD5 hashes were: how many hashes must be taken before  
the probability for a collision become non negligible? (I'm assuming  
equi-probability of every hash.)

The brute force approach is simply to enumerate the non-collision  
probability for k different hashes, and compute until it becomes lower  
than 1. This probability is (writing N for 2 ^ 128):
N * (N-1) * (N - 2) * ... * (N - k)
---------------------------------------
N^k

I tried computing this using the bignum library that comes with OCaml,  
and it slows down to a crawl very fast (for k ~ 1000).

So I tried to be more subtle and approximate the result (using  
Stirling's approximation of factorials), but OCaml's floats are not  
precise enough to yield anything significant. (I'm trying to compute  
the log of the approximation of N! / (N^k * (N-k)!), which is N (ln N)  
- N - (k (ln N) + (N - k)(ln (N - k)) - (N - k)).)

Is there a library with better floating point precision than the OCaml  
one?

Thanks,

Alan