Skip to content

Dr. Will Faithfull

Linear Congruential Generators and Uniqueness

rng, data-science2 min read

I am not an expert in number theory, or PRNGs for that matter. I am sometimes forced to do a little academic digging to solve a particular problem. So it was with Linear Congruential Generators or LCGs. I thought I would write up what I found, so that it might be easier for the next person to find what they need.

LCGs

LCGs are a particular class of psuedorandom number generating algorithms, which are cryptographically insecure, but fast and simple. The latter two points continue to make them attractive for a number of use cases.

An LCG takes the following general form:

$$X_{n+1} = (aX_n + c) \mod m$$

Now let's say you have a PRNG of this form, and you are concerned about it's uniqueness, it's cycle or it's period to throw out a few terms people use to talk about the same thing.

Specifically, we would like to know whether we will get any repeated integers or not. And if we do, when will it happen?

Fortunately, these properties are well understood for LCGs. If your PRNG looks like the above equation (remember $a$ might be 1), then the following is applicable.

Full cycle

What we are interested in is called the full cycle of a PRNG. A PRNG has a full cycle if for any seed, it will traverse every valid state before returning to the seed state. In this particular case, state refers to the integers.

My PRNG might have the following valid states:

10, 1, 2, 3, 4, 5, 6, 7, 8, 9

So, if it had a full cycle, whatever number I seed it with, over 10 iterations it will generate the numbers 0-9 in a psuedorandom order, with no repetitions.

Conditions for full cycle in LCGs

Now to the crux. If we want an LCG to have a full cycle of size $m$, we can follow a proven theorem (the Hull-Dobell theorem[1]) to guarantee this. Your LCG parameters must pass the following tests:

  1. $m$ and $c$ must be coprime.
  2. $a-1$ must be divisible by all prime factors of $m$
  3. $a-1$ must be divisible by 4 if $m$ is divisible by 4.

So for example, for parameters m = 1000, a = 1, c = 1049

  1. $gcd(1000,1049) = 1$, so $m$ and $c$ are coprime.
  2. $a-1 = 0$, $0$ is divisible by everything.
  3. $mod(1000,4) = 0$, but $a-1 = 0$ and $0$ is divisible by everything.

In this instance, we're all good. The generator will cycle through the numbers 1-1000 pseudorandomly for any given seed value.

[1] Hull, Thomas E., and Alan R. Dobell. "Random number generators." SIAM review 4.3 (1962): 230-254.

© 2020 by Dr. Will Faithfull. All rights reserved.
Theme by LekoArts