# Miller-Rabin Primality Test

Aug 15 2016

Back when I started messing around with code, I was told to try out Project Euler to improve my coding skills (something I suggest everyone do). I was able to progress fairly well along but I always struggled with the questions involving prime numbers because I could never get my code to run efficiently on them.

I looked for a solution and repeatedly only found mention to the Sieve of Eratosthenes, which is fairly intuitive and appeared to be the go-to primality test. However, in my early days of programming, I still struggled to implement the algorithm so I ultimately pushed those prime-based questions off to the side.

Years later, while reading Brian Christian’s and Tom Griffiths’ Algorithms to Live By, I happened across the **Miller-Rabin Primality Test**, an alternative to the Sieve of Eratosthenes. The algorithm works as follows: say you wanted to test if (an odd integer) is a prime number. To do so, we can first find and such that . Next, we can pick a random integer such that . With all defined, we next test if or for some . If either of these equalities are true, then we can say that passes the test and any prime number will pass the test. It is fairly trivial to test this for pretty much any . However, the trick with this algorithm is that non-prime numbers will also pass this test (return a false-positive) of the time.

To correct for this, we can try the test again with a new randomly-selected . If again passes the test, we know that there is a chance that is not actually a prime. Therefore, if we repeated the test times successfully, we can say that there is a chance that is not actually a prime number. Using this method, we can determine if a particular number is prime with fairly high certainty after only a few iterations. For example, if we repeat the test times, we can declare to be prime with 99.9% certainty. If passes after iterations, we know with 99.9999% certainty.

Using this, I took another crack at those pesky prime-based Project Euler problems and what do you know, it worked like a charm! While there is a very, very, very slim chance that this still returns a false-positive (I’m using so a 0.00153% chance), given the non-critical context of the application, I’ll definitely take the speed increase over the non-100% certainty.

For anyone interested, I’ve included a sample implementation in Python of the algorithm below.