Random Number Generators: Professor Karl Sigman Columbia University Department of IEOR New York City USA

Random Number Generators
Professor Karl Sigman

Columbia University
Department of IEOR
New York City
USA
1/17
Introduction
Your computer “generates" numbers U1 , U2 , U3 , . . . that are

considered independent and uniformly randomly distributed on the
continuous interval (0, 1).
Recall that the probability distribution (cumulative distribution

function) of such a uniformly distributed random variable U is given by
F(x) = P(U ≤ x) = x, x ∈ (0, 1),
and more generally, for 0 ≤ y < x < 1,
P(y < U ≤ x) = x − y.
In other words, the probability that U falls in an interval (y, x] is just

the length of the interval, x − y.
2/17
Introduction
Independence means that regardless of the values of the first n

random numbers, U1 , . . . , Un , the value of the next one, Un+1 , still has
the same uniform distribution over (0, 1); it is not in any way effected
by those previous values.
This is analogous to sequentially flipping a (fair) coin: regardless of

the first n flips, the next one will still land heads (H) or tails (T) with
probability 1/2.
The sequence of random variables (rvs) U1 , U2 , . . . is an example of

an independent and identically distributed(iid) sequence. Here, the
identical distribution is the uniform over (0, 1), which is a continuous
analog of “equally likely" probabilities.
3/17
Introduction
In Python, for example, you can obtain such U as follows:
import random
U = random.random()
Once random is imported, then each time you use the command
U = random.random()
you receive a new uniform number within [0, 1).
4/17
Introduction
It turns out once we have access to such uniform numbers U, we can

use them to construct (“simulate/generate") random variables of any
desired distribution, construct stochastic processes such as random
walks, Markov chains, Poisson processes, renewal processes,
Brownian motion and many other processes.
Suppose for example, that we want a random variable X that has an
exponential distribution at rate λ: The cumulative distribution function
(CDF) is given by
F(x) = P(X ≤ x) = 1 − e −λx , x ≥ 0.
5/17
Introduction
Then simply define

1
X =− ln (U),
λ
where ln (y) denotes the natural logarithm of y > 0.
Proof:
1
P(X ≤ x) = P(− ln (U) ≤ x)
λ
= P(ln (U) ≥ −λx)
= P(U ≥ e −λx )
= 1 − e −λx .
(Recall that P(U ≥ y) = 1 − y, y ∈ (0, 1).)
6/17
Pseudorandom numbers
It turns out that the numbers generated by a computer are not really
random nor independent as we said, but what are called
Pseudorandom numbers.
This means that they appear, for all practical purposes, to be random
(and independent) in the sense that they would pass various
statistical tests for checking the random/independent property. We
thus can use them in our simulations as if they were truly
random—and we do.
Next we will discuss how the computer generates these numbers.

This is deeply related to "cryptography" in which one wants to hide
important information (from an opponent/enemy) in data that
“appears" to be random.
7/17
As an example to get you to think: Suppose I hand you a sequence of

zeros and ones:
(0, 1, 1, 0, 1, 0, 1, 0, 0, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 0, 1, 0, 0, 1).
I tell you that I flipped a fair coin 25 times where 1 = Head (H), and
0 = Tails (T).
How can you check with certainty that I am telling the truth?
ANSWER: You can’t
8/17
But if I keep handing you such sequences of various lengths, then

you can perform statistical tests that would help you decide if the
sequences are consistent with coin flips.
9/17
Linear Congruential Generators
The most common and easy to understand and implement random
number generator is called a Linear Congruential Generator (LCG)
and is defined by a recursion as follows:
Zn+1 = (aZn + c) mod m, n ≥ 0,
Un = Zn /m,
where 0 < a < m, 0 ≤ c < m are constant integers, and mod m
means modulo m which means you divide by m and leave the
remainder. For example 6 mod 4 = 2, 2 mod 4 = 2, 7 mod 4 = 3,
12 mod 4 = 0. Thus all the Zn fall between 0 and m − 1; the Un are
thus between 0 and 1.
0 ≤ Z0 < c is called the seed. m is chosen to be very large, usually of

the form m = 232 or m = 264 because your computer architecture is
based on 32 or 64 bits per word; the modulo computation merely
involves truncation by the computer, hence is immediate.
10/17
For example if m = 23 = 8, then in binary, there are 8 numbers

representing {0, 1, 2, 3, 4, 5, 6, 7} given by a 3-tuple of 0s and 1s,
denoted by
(i0 , i1 , i2 ) = i0 20 + i1 21 + i2 22 , ij ∈ {0, 1}, j = 0, 1, 2.
Thus, (0, 0, 0) = 0, (1, 0, 0) = 20 = 1, (1, 1, 0) = 20 + 21 = 3,

(0, 0, 1) = 22 = 4 and (1, 1, 1) = 20 + 21 + 22 = 7, and so on.
Note how 7 + 1 = 8 = 0 mod 8 is computed by truncation:
(1, 1, 1) + (1, 0, 0) = (0, 0, 0, 1) = 23 . The last component gets
truncated yielding (0, 0, 0) = 0.
10 = (0, 1, 0, 1) gets truncated to (0, 1, 0) = 2, and so on.
11/17
Here is a more typical example:
Zn+1 = (1664525 × Zn + 1013904223) mod 232 .

Thus a = 1664525 and c = 1013904223 and
m = 232 = 4, 294, 967, 296;
more than 4.2 billion.
12/17
The numbers a, c, m must be carefully chosen to get a “good"

random number generator, in particular we would want all c values
0, 1, . . . c − 1 to be generated in which case we say that the LCG has
full period of length c. Such generators will cyclically run thru the
numbers over and over again.
To illustrate, consider
Zn+1 = (5Zn + 1) mod 8, n ≥ 0,
with Z0 = 0. Then
(Z0 , Z1 , . . . , Z7 ) = (0, 1, 6, 7, 4, 5, 2, 3),
and Z8 = 16 mod 8 = 0, hence causing the sequence to repeat.
13/17
If we increase c to c = 16,
Zn+1 = (5Zn + 1) mod 16, n ≥ 0,

with Z0 = 0, then
(Z0 , Z1 , . . . , Z15 ) = (0, 1, 6, 15, 12, 13, 2, 11, 8, 9, 14, 7, 4, 5, 10, 3),
and Z16 = 16 mod 16 = 0, hence causing the sequence to repeat.
14/17
Choosing good numbers a, c, m involves the sophisticated use of

number theory; prime numbers and such, and has been extensively
researched/studied by computer scientists and mathematicians for
many years.
Note that the numbers generated are entirely deterministic: If you

know the values a, c, m, then once you know one value (Z0 , say) you
know them all. But if you are handed a long sequence of the Un , they
certainly appear random, and that is the point.
15/17
The advantages of using a LCG:

1. Very fast to implement
2. Requires no storage of the numbers, only the most recent value.
3. Replication: Using the same seed, you can generate exactly the
same sequence again and again which is extremely useful when
comparing alternative systems/models: By using the same
numbers you are reducing the variability of differences that
would be caused by using different numbers; any difference in
the comparisons are thus due to the inherent difference in the
models themselves.
16/17
More sophisticated generators
Python currently uses the Mersenne Twister as its core random

number generator; U = random.random(). It produces at double
precision (64 bit), 53-bit precision (floating), and has a period of
219937 − 1 (a Mersenne prime number). The Mersenne Twister is one
of the most extensively tested random number generators in
existence. (There is both a 32-bit and a 64-bit implementation.) It is
not a LCG, it is far more complex, but yet again is deterministic and
recursive. The basic Mersenne Twister algorithm continues to be
refined and modified over the years to make it faster to implement,
etc.
17/17

Random Number Generators: Professor Karl Sigman Columbia University Department of IEOR New York City USA

Uploaded by

Copyright:

Available Formats

Random Number Generators: Professor Karl Sigman Columbia University Department of IEOR New York City USA

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Random Number Generators: Professor Karl Sigman Columbia University Department of IEOR New York City USA

Uploaded by

Copyright:

Available Formats

Random Number Generators

Professor Karl Sigman

Your computer “generates" numbers U1 , U2 , U3 , . . . that are

Recall that the probability distribution (cumulative distribution

F(x) = P(U ≤ x) = x, x ∈ (0, 1),

and more generally, for 0 ≤ y < x < 1,

In other words, the probability that U falls in an interval (y, x] is just

Independence means that regardless of the values of the first n

This is analogous to sequentially flipping a (fair) coin: regardless of

The sequence of random variables (rvs) U1 , U2 , . . . is an example of

In Python, for example, you can obtain such U as follows:

It turns out once we have access to such uniform numbers U, we can

F(x) = P(X ≤ x) = 1 − e −λx , x ≥ 0.

Then simply define

(Recall that P(U ≥ y) = 1 − y, y ∈ (0, 1).)

Next we will discuss how the computer generates these numbers.

As an example to get you to think: Suppose I hand you a sequence of

ANSWER: You can’t

But if I keep handing you such sequences of various lengths, then

0 ≤ Z0 < c is called the seed. m is chosen to be very large, usually of

For example if m = 23 = 8, then in binary, there are 8 numbers

(i0 , i1 , i2 ) = i0 20 + i1 21 + i2 22 , ij ∈ {0, 1}, j = 0, 1, 2.

Thus, (0, 0, 0) = 0, (1, 0, 0) = 20 = 1, (1, 1, 0) = 20 + 21 = 3,

Here is a more typical example:

Zn+1 = (1664525 × Zn + 1013904223) mod 232 .

m = 232 = 4, 294, 967, 296;

more than 4.2 billion.

The numbers a, c, m must be carefully chosen to get a “good"

Zn+1 = (5Zn + 1) mod 8, n ≥ 0,

(Z0 , Z1 , . . . , Z7 ) = (0, 1, 6, 7, 4, 5, 2, 3),

and Z8 = 16 mod 8 = 0, hence causing the sequence to repeat.

Zn+1 = (5Zn + 1) mod 16, n ≥ 0,

and Z16 = 16 mod 16 = 0, hence causing the sequence to repeat.

Choosing good numbers a, c, m involves the sophisticated use of

Note that the numbers generated are entirely deterministic: If you

The advantages of using a LCG:

Python currently uses the Mersenne Twister as its core random

You might also like