Section 6.

2: Generating Discrete Random Variates

Discrete-Event Simulation: A First Course
c 2006 Pearson Ed., Inc. 0-13-142917-5

Discrete-Event Simulation: A First Course

Section 6.2: Generating Discrete Random Variates

1/ 24

Section 6.2: Generating Discrete Random Variates

The inverse distribution function (idf) of X is the function F : (0, 1) X for all u (0, 1) as F (u ) = min{x : u < F (x )}
x

F () is the cdf of X That is, if F (u ) = x , x is the smallest possible value of X for which F (x ) is greater than u

Discrete-Event Simulation: A First Course

Section 6.2: Generating Discrete Random Variates

2/ 24

Example 6.2.1
Two common ways of plotting a cdf with X = {a, a + 1, . . . , b }:
1.0 F (u) = x

1.0 F (u) = x F () u

F ()

0.0

0.0

Theorem (6.2.1) Let X = {a, a + 1, . . . , b } where b may be and F () be the cdf of X For any u (0, 1), if u < F (a), F (u ) = a else F (u ) = x where x X is the unique possible value of X for which F (x 1) u < F (x )
Discrete-Event Simulation: A First Course Section 6.2: Generating Discrete Random Variates 3/ 24

Algorithm 6.2.1
For X = {a, a + 1, . . . , b }, the following linear search algorithm denes F (u ) Algorithm 6.2.1
x = a; while (F(x) <= u ) x++; return x; /*x is F (u )*/

Average case analysis:

Let Y be the number of while loop passes Y =X a E [Y ] = E [X a ] = E [X ] a = a

Discrete-Event Simulation: A First Course

Section 6.2: Generating Discrete Random Variates

4/ 24

Algorithm 6.2.2
Idea: start at a more likely point For X = {a, a + 1, . . . , b }, a more ecient linear search algorithm denes F (u ) Algorithm 6.2.2
x = mode; /*initialize with the mode of X */ if (F(x) <= u) while (F(x) <= u) x++; else if (F(a) <= u) while (F(x-1) > u) x--; else x = a; return x; /* x is F (u )*/

For large X , consider binary search

Discrete-Event Simulation: A First Course Section 6.2: Generating Discrete Random Variates 5/ 24

Idf Examples

In some cases F (u ) can be determined explicitly If X is Bernoulli(p ) and F (x ) = u , then x = 0 i 0 < u < 1 p: F (u ) = 0 0 < u < 1p 1 1p u <1

Discrete-Event Simulation: A First Course

Section 6.2: Generating Discrete Random Variates

6/ 24

Example 6.2.3: Idf for Equilikely

If X is Equilikely(a, b ), F (x ) = x a+1 ba+1 x = a, a + 1, . . . , b

For 0 < u < F (a), For F (a) u < 1,

F (u ) = a

F (x 1) u < F (x )

(x 1) a + 1 x a+1 u< ba+1 ba+1 x a + (b a + 1)u < x + 1

Therefore, for all u (0, 1) F (u ) = a + (b a + 1)u

Discrete-Event Simulation: A First Course

Section 6.2: Generating Discrete Random Variates

7/ 24

Example 6.2.4: Idf for Geometric

If X is Geometric(p ), F (x ) = 1 p x +1 x = 0, 1, 2, . . .

For 0 < u < F (0), For F (0) u < 1,

F (u ) = 0

F (x 1) u < F (x ) 1 p x u < 1 p x +1 . . . ln(1 u ) <x +1 x ln(p ) For all u (0, 1) F (u ) =

Discrete-Event Simulation: A First Course

ln(1 u ) ln(p )
8/ 24

Section 6.2: Generating Discrete Random Variates

Random Variate Generation By Inversion

X is a discrete random variable with idf F () Continuous random variable U is Uniform(0, 1) Z is the discrete random variable dened by Z = F (U ) Theorem (6.2.2) Z and X are identically distributed Theorem 6.2.2 allows any discrete random variable (with known idf) to be generated with one call to Random() Algorithm 6.2.3
If X is a discrete random variable with idf F (), a random variate x can be generated as u = Random(); return F (u);
Discrete-Event Simulation: A First Course Section 6.2: Generating Discrete Random Variates 9/ 24

Proof for Theorem 6.2.2

Prove that X = Z
F : (0, 1) X , so u (0, 1) such that F (u ) = x Z = F (U ) It follows that x Z so X Z From denition of Z , if z Z then u (0, 1) such that F (u ) = z F : (0, 1) X It follows that z X so Z X

Prove that Z and X have the same pdf

Let X = Z = {a, a + 1, . . . , b }, from denition of Z and F () and theorem 6.2.1: if z = a, Pr(Z = a) = Pr(U < F (a)) = F (a) = f (a) if z Z , z = a, Pr(Z = z ) = Pr(F (z 1) U < F (z )) = F (z )F (z 1) = f (z )
Discrete-Event Simulation: A First Course Section 6.2: Generating Discrete Random Variates 10/ 24

Inversion Examples
Example 6.2.5 Consider X with pdf 0.1 x=2 f (x ) = 0.3 x=3 0.6 x=6
1.0 u F () F ()

The cdf for X is plotted using two formats

1.0 u

0.4 0.1 0.0

0.4 0.1 0.0

2 3 4 5 6 x

2 3 4 5 6 x

Discrete-Event Simulation: A First Course

Section 6.2: Generating Discrete Random Variates

11/ 24

Algorithm for Example 6.2.5

Example 6.2.5
if (u < 0.1) return 2; else if (u < 0.4) return 3; else return 6;

returns 2 with probability 0.1, 3 with probability 0.3 and 6 with probability 0.6 which corresponds to the pdf of X This example can be made more ecient: check the ranges for u associated with x = 6 rst (the mode), then x = 3, then x =2 Problems may arise when |X | is large or innite

Discrete-Event Simulation: A First Course

Section 6.2: Generating Discrete Random Variates

12/ 24

More Inversion Examples

Example 6.2.6: Generating a Bernoulli(p ) Random Variate
u = Random(); if (u < 1-p) return 0; else return 1;

Example 6.2.7: Generating an Equilikely(a, b ) Random Variate

u = Random(); return a + (long) (u * (b - a + 1));

Example 6.2.8: Generating a Geometric(p ) Random Variate

u = Random() return (long) (log(1.0 - u) / log(p));

Discrete-Event Simulation: A First Course

Section 6.2: Generating Discrete Random Variates

13/ 24

Example 6.2.9
X is a Binomial(n, p ) random variate
x

F (x ) =
t =0

n x

p x (1 p )nx

x = 0, 1, 2, . . . , n

Incomplete beta function F (x ) = 1 I (x + 1, n x , p ) x = 0, 1, . . . , n 1 1 x =n

Except for special cases, an incomplete beta function cannot be inverted to form a closed form expression for the idf Inversion is not easily applied to generation of Binomial(n, p ) random variates

Discrete-Event Simulation: A First Course

Section 6.2: Generating Discrete Random Variates

14/ 24

Algorithm Design Criteria

The design of a correct, exact and ecient algorithm to generate corresponding random variates is often complex
Portability - implementable in high-level languages Exactness - histogram of variates should converge to pdf Robustness - performance should be insensitive to small changes in parameters and should work properly for all reasonable parameter values Eciency - it should be time ecient (set-up time and marginal execution time) and memory ecient Clarity - it is easy to understand and implement Synchronization - exactly one call to Random() is required Monotonicity - it is synchronized and the transformation from u to x is monotone increasing (or decreasing)

Inversion satises some criteria, but not necessarily all

Discrete-Event Simulation: A First Course

Section 6.2: Generating Discrete Random Variates

15/ 24

Example 6.2.10
To generate Binomial(10, 0.4), the pdf is (to 0.ddd precision)
x: 0 1 2 3 4 5 6 7 8 9 10 f (x ) : 0.006 0.040 0.121 0.215 0.251 0.201 0.111 0.042 0.011 0.002 0.000

Random variates can be generated by lling a 1000-element integer-valued array a[] with 6 0s, 40 1s, 121 2s, etc. Binomial(10, 0.4) Random Variate
j = Equilikely(0,999); return a[j];

This algorithm is portable, robust, clear, synchronized and monotone, with small marginal execution time The algorithm is not exact: f (10) = 1/9765625 Set-up time and memory eciency could be problematic: for 0.ddddd precision, need 100 000-element array
Discrete-Event Simulation: A First Course Section 6.2: Generating Discrete Random Variates 16/ 24

Example 6.2.11: Exact Algorithm for Binomial(10, 0.4)

An exact algorithm is based on

lling an 11-element oating-point array with cdf values then using Alg. 6.2.2 with x = 4 to initialize the search

In general, to generate Binomial(n, p ) by inversion

compute a oating-point array of n + 1 cdf values use Alg. 6.2.2 with x = np to initialize the search

The library rvms can be used to compute the cdf array by calling cdfBinomial(n,p,x) for x = 0, 1, . . . , n Only drawback is some ineciency (setup time and memory)

Discrete-Event Simulation: A First Course

Section 6.2: Generating Discrete Random Variates

17/ 24

Example 6.2.12
The cdf array from Example 6.2.11 can be eliminated
cdf values computed as needed by Alg. 6.2.2 Reduces set-up time and memory Increases marginal execution time

Function idfBinomial(n,p,u) in library rvms does this Binomial(n, p ) random variates can be generated by inversion Generating a Binomial Random Variate
u = Random(); return idfBinomial(n, p, u); /* in library rvms*/

Inversion can be used for the six models:

Ideal for Equilikely(a, b ), Bernoulli(p ) and Geometric(p ) For Binomial(n, p ), Pascal(n, p ) and Poisson(), time and memory eciency can be a problem if inversion is used
Discrete-Event Simulation: A First Course Section 6.2: Generating Discrete Random Variates 18/ 24

Alternative Random Variate Generation Algorithms

Example 6.2.13 Binomial Random Variates A Binomial(n, p ) random variate can be generated by summing an iid Bernoulli(p ) sequence Generating a Binomial Random Variate
x = 0; for (i = 0; i < n; i++) x += Bernoulli(p); return x;

The algorithm is: portable, exact, robust, clear The algorithm is not: synchronized or monotone Marginal execution: O(n) complexity

Discrete-Event Simulation: A First Course

Section 6.2: Generating Discrete Random Variates

19/ 24

Poisson Random Variates

A Poisson() random variable is the n limiting case of a Binomial(n, /n) random variable For large n, Poisson() Binomial(n, /n) The previous O(n) algorithm for Binomial(n, p ) should not be used when n is large The Poisson() cdf F () is equal to an incomplete gamma function F (x ) = 1 P (x + 1, ) x = 0, 1, 2, . . .

An incomplete gamma function cannot be inverted to form an idf Inversion to generate a Poisson() requires searching the cdf as in Examples 6.2.11 and 6.2.12
Discrete-Event Simulation: A First Course Section 6.2: Generating Discrete Random Variates 20/ 24

Example 6.2.14
Generating a Poisson Random Variate
a = 0.0; x = 0; while (a < ) { a += Exponential(1.0); x++; } return x - 1;

The algorithm does not rely on inversion or the large n version of Binomial(n, p ) The algorithm is: portable, exact, robust; not synchronized or monotone; marginal execution time can be long for large It is obscure. Clarity will be provided in Section 7.3

Discrete-Event Simulation: A First Course

Section 6.2: Generating Discrete Random Variates

21/ 24

Pascal Random Variates

A Pascal(n, p ) cdf is equal to an incomplete beta function: F (x ) = 1 I (x + 1, n, p ) x = 0, 1, 2, . . . X is Pascal(n, p ) i X = X1 + X2 + + Xn where X1 , X2 , . . . , Xn is an iid Geometric(p ) sequence Example 6.2.15 Summing Geometric(p ) random variates to generate a Pascal(n, p ) random variate Generating a Pascal Random Variate
x = 0; for(i = 0; i < n; i++) x += Geometric(p); return x;

The algorithm is: portable, exact, robust, clear; not synchronized or monotone; marginal execution complexity is O(n)
Discrete-Event Simulation: A First Course Section 6.2: Generating Discrete Random Variates 22/ 24

Library rvgs

Includes 6 discrete random variate generators (as below) and 7 continuous random variate generators
long long long long long long Bernoulli(double p ) Binomial(long n, double p ) Equilikely(long a, long b ) Geometric(double p ) Pascal(long n, double p ) Poisson(double )

Functions Bernoulli, Equilikely, Geometric use inversion; essentially ideal Functions Binomial, Pascal, Poisson do not use inversion

Discrete-Event Simulation: A First Course

Section 6.2: Generating Discrete Random Variates

23/ 24

Library rvms

Provides accurate pdf, cdf, idf functions for many random variates Idfs can be used to generate random variates by inversion Functions idfBinomial, idfPascal, idfPoisson may have high marginal execution times Not recommended when many observations are needed due to time ineciency Array of cdf values with inversion may be preferred

Discrete-Event Simulation: A First Course

Section 6.2: Generating Discrete Random Variates

24/ 24