Section III: Linear Congruential Generator II   < < Index > >

 

Potency:

As discussed in last section, a long period is of course necessary for a sequence to look random, but it's not sufficient. As an easy example, we let a = 1, c = 1, m = 231 , X0 = 0, then we get the sequence 0, 1, 2, 3,... . It has a very long period, but nobody would ever think it random. So there are more to consider. In this section, we will introduce some concept called Potency, which will take care of situations like this.

The potency of a linear conguential sequence with maximum period is defined to be the least positive integer s such that

bs = 0 mod m, where b = a-1.

When the multiplier satisfies the conditions of Theorem A, such s always exists since b is a multiple of every prime dividing m. For example, if b = 22*3*5, m = 29*32*5, then the potency is 5.

Now that we always assume the conditions of Theorem A is satisfied, we may analyze the randomness of the sequence by taking X0 = 0, since 0 occurs somewhere in the sequence. With this assumption, we have Xn = ((an - 1)c/b) mod m. If we expand an - 1 = (b+1) n - 1 by the binomial theorem, we find that

Xn = c(n + (n!/(2!(n-2)!))*b + ... + (n!/(s!(n-s)!))*bs-1) mod m.

All terms in bs, bs+1, etc., may be igored, since they are multiples of m.

Let's consider some special cases. If a = 1, the potency is 1; and Xn = cn mod m, this sequence can be observed to be not random. The case we discussed at the beginning is a =1, c = 1. If the potency is 2, we have Xn = cn + cbn(n-1)/2, and again the sequence is not very random; indeed,

Xn+1 - Xn = c + cbn mod m.

so the differences between one generated number and the next are very simply related from one value of n to the next.

The situation gets better when the potency becomes bigger. A potency of at least 5 would seem to be required for sufficiently random values.

At the beginning , people didn't realise the importance of big potency. In history, the following form of generators:

Xn+1 = (2k + 1)Xn + 1 mod 2e. (3.1)

have been widely discussed in the literature and recommended by many authors. It satisfies the condition in Theorem A if k ≥ 2, so has period as long as 2e. However, some five years of experimentation with this method show that this kind of generators should be avoided. Potency is one of the most important factors.

Exercise 3.1: Discuss the potency of generators in the form of formula (3.1).
Exercise 3.2: From previous exercise, we can see generator (3.1) can get relatively large potency when k is much smaller than e. But for some reason, small multipliers are also undesirable. Think about why.

 

Other Methods:

Besides the simple linear congruential method mentioned above, there are many other kinds of generators. Many of them are based on this simple method or are modification of this method. We give a brief introduction to some of them in the following:

Multiple Recursive Generators:

A simple extension of simple linear congruential method is to use multiples of previous k values to generate the next one:

Xi = (a1Xi-1 + a2Xi-2 + ... +akXi-k) mod m, k > 1.  

The number of previous number used, k, is called the "order" of the generator.

One advantage of this method is the the period can be much longer than the simple linear conguential method. For m a prime, Knuth has shown that the maximum period is mk - 1 with properly chosen ai's.

Lagged Fibonacci congruential generator:

This is a special case of the multiple recursive generators.

A simple Fibonacci sequence has Xi+2 = Xi+1 + Xi. Reducing the numbers mod m produces a sequence that looks somewhat random but actually does not have satisfactory randomness properties.

Instead of combining successive terms, we combine terms at some greater distance apart, so the lagged Fibonacci congruential generator is

Xi = (Xi-j + Xi-k) mod m.

If j, k, and m are chosen properly, the lagged Fibonacci generator can perform well.

Inversive Congrutial Generators:

All the generators mentioned before are linear generators, i.e. any number in the sequence is linear function of its predecessors. Many nonlinear generators have been proposed. But since the output of nonlinear generators are more complicated, more study about their properties is still needed. Following is one of them:

Xi = (aXi-1- + c) mod m, with 0 ≤ Xi < m,

where X- denotes the multiplicative inverse of X modulo m, if it exists, or else it denotes 0.

The multiplicative inverse, X-, of X modulo m is defined for all nonzero x relatively prime to m by

X*X- = 1 mod m.

Get random integers in any interval:

Using linear congruential generator, we get random number sequence of integers in [0, m-1]. Devide the sequence by m, we get random number sequence in [0,1). Denote the sequence in [0,1) by Un. To get a random integer sequence Yn between 0 and k - 1, we can multiply Un by k, and let Yn = [kUn], where [U] denotes the biggest integer that is no larger than U. If Un is uniformly distributed in [0,1), then Yn is uniformly distributed in the set {0, 1, 2,..., k-1}.

 

Exercise 3.3: What if we want to get integer sequence between 10 and 99(include 10 and 99) but not eaqual to 50?
Exercise 3.4: If we want to get random interger sequence between 0 and 9(include 0 and 9), can we just let m = 10 in the linear congruential sequence?
Exercise 3.5: The method given above gives each integer with equal probability. What if we want to give different weights to different integers? e.g. We want a 0-1 sequence(a sequence consists of 0 and 1), but the probability Xn = 0 is 70%, while the probability Xn = 1 is 30%.

 

Back to MEC Homepage Previous Section Back to Index Next Section