2WB05 Simulation

Lecture 8: Generating
random variables
Marko Boon
http://www.win.tue.nl/courses/2WB05

January 7, 2013

Outline
1. How do we generate random variables?
2. Fitting distributions

Department of Mathematics and Computer Science

2/36

Generating random variables

How do we generate random variables?
Sampling from continuous distributions
Sampling from discrete distributions

Department of Mathematics and Computer Science

3/36

Continuous distributions

4/36

Inverse Transform Method

Let the random variable X have a continuous and increasing distribution function F. Denote the inverse of F
by F 1 . Then X can be generated as follows:
Generate U from U (0, 1);
Return X = F 1 (U ).
If F is not continuous or increasing, then we have to use the generalized inverse function
F 1 (u) = min{x : F(x) u}.

Department of Mathematics and Computer Science

Continuous distributions

5/36

Examples
X = a + (b a)U is uniform on (a, b);
X = ln(U )/ is exponential with parameter ;
X = ( ln(U ))1/a / is Weibull, parameters a and .
Unfortunately, for many distribution functions we do not have an easy-to-use (closed-form) expression for the
inverse of F.

Department of Mathematics and Computer Science

Continuous distributions

6/36

Composition method
This method applies when the distribution function F can be expressed as a mixture of other distribution functions F1 , F2 , . . .,

X
F(x) =
pi Fi (x),
i=1

where

pi 0,

pi = 1

i=1

The method is useful if it is easier to sample from the Fi s than from F.

First generate an index I such that P(I = i) = pi ,

i = 1, 2, . . .

Generate a random variable X with distribution function FI .

Department of Mathematics and Computer Science

Continuous distributions

7/36

Examples
Hyper-exponential distribution:
F(x) = p1 F1 (x) + p2 F2 (x) + + pk Fk (x),

x 0,

where Fi (x) is the exponential distribution with parameter i , i = 1, . . . , k.

Double-exponential (or Laplace) distribution:
f (x) =

1 x
2e ,

where f denotes the density of F.

Department of Mathematics and Computer Science

1 x
2e ,

x < 0;
x 0,

Continuous distributions
Convolution method
In some case X can be expressed as a sum of independent random variables Y1 , . . . , Yn , so
X = Y1 + Y2 + + Yn .
where the Yi s can be generated more easily than X .
Algorithm:
Generate independent Y1 , . . . , Yn , each with distribution function G;
Return X = Y1 + + Yn .

Department of Mathematics and Computer Science

8/36

Continuous distributions

9/36

Example
If X is Erlang distributed with parameters n and , then X can be expressed as a sum of n independent
exponentials Yi , each with mean 1/.
Algorithm:
Generate n exponentials Y1 , . . . , Yn , each with
mean ;
Set X = Y1 + + Yn .
More efficient algorithm:
Generate n uniform (0, 1) random variables
U1 , . . . , Un ;
Set X = ln(U1U2 Un )/.

Department of Mathematics and Computer Science

Continuous distributions

10/36

Acceptance-Rejection method
Denote the density of X by f . This method requires a function g that majorizes f ,
g(x) f (x)
for all x. Now g will not be a density, since
Z

g(x)d x 1.

c=

Assume that c < . Then h(x) = g(x)/c is a density. Algorithm:

1. Generate Y having density h;
2. Generate U from U (0, 1), independent of Y ;
3. If U f (Y )/g(Y ), then set X = Y ; else go back to step 1.
The random variable X generated by this algorithm has density f .

Department of Mathematics and Computer Science

Continuous distributions

11/36

Validity of the Acceptance-Rejection method

Note
P(X x) = P(Y x|Y accepted).
Now,
Z

P(Y x, Y accepted) =

f (y)
1
h(y)dy =
g(y)
c

f (y)dy,

and thus, letting x gives

1
P(Y accepted) = .
c
Hence,
P(Y x, Y accepted)
P(X x) =
=
P(Y accepted)

Department of Mathematics and Computer Science

f (y)dy.

Continuous distributions

12/36

Note that the number of iterations is geometrically distributed with mean c.

How to choose g?
Try to choose g such that the random variable Y can be generated rapidly;
The probability of rejection in step 3 should be small; so try to bring c close to 1, which mean that g should
be close to f .

Department of Mathematics and Computer Science

Continuous distributions

13/36

Example
The Beta(4,3) distribution has density
f (x) = 60x 3 (1 x)2 ,

0 x 1.

The maximal value of f occurs at x = 0.6, where f (0.6) = 2.0736. Thus, if we define
g(x) = 2.0736,

0 x 1,

then g majorizes f . Algorithm:

1. Generate Y and U from U (0, 1);
60Y 3 (1 Y )2
2. If U
, then set X = Y ; else reject Y and return to step 1.
2.0736

Department of Mathematics and Computer Science

Continuous distributions
Normal distribution
Methods:
Acceptance-Rejection method
Box-Muller method

Department of Mathematics and Computer Science

14/36

Continuous distributions

15/36

Acceptance-Rejection method
If X is N (0, 1), then the density of |X | is given by
2
2
f (x) = ex /2 ,
2

x > 0.

Now the function

g(x) =

2e/ex

majorizes f . This leads to the following algorithm:

1. Generate an exponential Y with mean 1;
2. Generate U from U (0, 1), independent of Y ;
2 /2

3. If U e(Y 1)

, then accept Y ; else reject Y and return to step 1.

4. Return X = Y or X = Y , both with probability 1/2.

Department of Mathematics and Computer Science

Continuous distributions
Box-Muller method
If U1 and U2 are independent U (0, 1) random variables, then
p
X 1 = 2 ln U1 cos(2U2 )
p
X 2 = 2 ln U1 sin(2U2 )
are independent standard normal random variables.
This method is implemented in the function nextGaussian() in java.util.Random

Department of Mathematics and Computer Science

16/36

Discrete distributions

17/36

Discrete version of Inverse Transform Method

Let X be a discrete random variable with probabilities
P(X = xi ) = pi ,

i = 0, 1, . . . ,

pi = 1.

i=0

To generate a realization of X , we first generate U from U (0, 1) and then set X = xi if

i1
X
j=0

Department of Mathematics and Computer Science

pj U <

i
X
j=0

pj.

Discrete distributions

18/36

So the algorithm is as follows:

Generate U from U (0, 1);
Determine the index I such that
I 1
X

pj U <

j=0

I
X

pj

j=0

and return X = x I .
The second step requires a search; for example, starting with I = 0 we keep adding 1 to I until we have found
the (smallest) I such that
I
X
U<
pj
j=0

Note: The algorithm needs exactly one uniform random variable U to generate X ; this is a nice feature if you
use variance reduction techniques.

Department of Mathematics and Computer Science

Discrete distributions
Array method: when X has a finite support
Suppose pi = ki /100, i = 1, . . . , m,
where ki s are integers with 0 ki 100
Construct array A[i], i = 1, . . . , 100 as follows:
set A[i] = x1 for i = 1, . . . , k1
set A[i] = x2 for i = k1 + 1, . . . , k1 + k2 , etc.
Then, first sample a random index I from 1, . . . , 100:
I = 1 + b100U c and set X = A[I ]

Department of Mathematics and Computer Science

19/36

Discrete distributions
Bernoulli
Two possible outcomes of X (success or failure):
P(X = 1) = 1 P(X = 0) = p.
Algorithm:
Generate U from U (0, 1);
If U p, then X = 1; else X = 0.

Department of Mathematics and Computer Science

20/36

Discrete distributions

21/36

Discrete uniform
The possible outcomes of X are m, m + 1, . . . , n and they are all equally likely, so
P(X = i) =

1
,
nm+1

i = m, m + 1, . . . , n.

Algorithm:
Generate U from U (0, 1);
Set X = m + b(n m + 1)U c.
Note: No search is required, and compute (n m + 1) ahead.

Department of Mathematics and Computer Science

Discrete distributions

22/36

Geometric
A random variable X has a geometric distribution with parameter p if
P(X = i) = p(1 p)i ,

i = 0, 1, 2, . . . ;

X is the number of failures till the first success in a sequence of Bernoulli trials with success probability p.
Algorithm:
Generate independent Bernoulli( p) random variables Y1 , Y2 , . . .; let I be the index of the first successful
one, so Y I = 1;
Set X = I 1.
Alternative algorithm:
Generate U from U (0, 1);
Set X = bln(U )/ ln(1 p)c.

Department of Mathematics and Computer Science

Discrete distributions
Binomial
A random variable X has a binomial distribution with parameters n and p if
 
n i
p (1 p)ni ,
i = 0, 1, . . . , n;
P(X = i) =
i
X is the number of successes in n independent Bernoulli trials, each with success probability p.
Algorithm:
Generate n Bernoulli( p) random variables
Y1 , . . . , Yn ;
Set X = Y1 + Y2 + + Yn .
Alternative algorithms can be derived by using the following results.

Department of Mathematics and Computer Science

23/36

Discrete distributions

24/36

Let Y1 , Y2 , . . . be independent geometric( p) random variables, and I the smallest index such that
I +1
X

(Yi + 1) > n.

i=1

Then the index I has a binomial distribution with parameters n and p.

Let Y1 , Y2 , . . . be independent exponential random variables with mean 1, and I the smallest index such that
I +1
X
i=1

Yi
> ln(1 p).
ni +1

Then the index I has a binomial distribution with parameters n and p.

Department of Mathematics and Computer Science

Discrete distributions

25/36

Negative Binomial
A random variable X has a negative-binomial distribution with parameters n and p if


n+i 1 n
p (1 p)i ,
i = 0, 1, 2, . . . ;
P(X = i) =
i
X is the number of failures before the n-th success in a sequence of independent Bernoulli trials with success
probability p.
Algorithm:
Generate n geometric( p) random variables
Y1 , . . . , Yn ;
Set X = Y1 + Y2 + + Yn .

Department of Mathematics and Computer Science

Discrete distributions

26/36

Poisson
A random variable X has a Poisson distribution with parameter if
i
P(X = i) = e ,
i!

i = 0, 1, 2, . . . ;

X is the number of events in a time interval of length 1 if the inter-event times are independent and exponentially distributed with parameter .
Algorithm:
Generate exponential inter-event times Y1 , Y2 , . . . with mean 1; let I be the smallest index such that
I +1
X
i=1

Set X = I .

Department of Mathematics and Computer Science

Yi > ;

Discrete distributions

27/36

Poisson (alternative)
Generate U(0,1) random variables U1 , U2 , . . .;
let I be the smallest index such that

IY
+1
i=1

Set X = I .

Department of Mathematics and Computer Science

Ui < e ;

Fitting distributions

28/36

Input of a simulation
Specifying distributions of random variables (e.g., interarrival times, processing times) and assigning parameter values can be based on:
Historical numerical data
Expert opinion
In practice, there is sometimes real data available, but often the only information of random variables that is
available is their mean and standard deviation.

Department of Mathematics and Computer Science

Fitting distributions
Empirical data can be used to:
construct empirical distribution functions and generate samples from them during the simulation;
fit theoretical distributions and then generate samples from the fitted distributions.

Department of Mathematics and Computer Science

29/36

Fitting distributions

30/36

Moment-fitting
Obtain an approximating distribution by fitting a phase-type distribution on the mean, E(X ), and the coefficient
of variation,
X
cX =
,
E(X )
of a given positive random variable X , by using the following simple approach.

Department of Mathematics and Computer Science

Moment-fitting

31/36

Coefficient of variation less than 1: Mixed Erlang

If 0 < c X < 1, then fit an E k1,k distribution as follows. If
1
1
c2X
,
k
k1
for certain k = 2, 3, . . ., then the approximating distribution is with probability p (resp. 1 p) the sum of k 1
(resp. k) independent exponentials with common mean 1/. By choosing


q
1
kp
2
2 ) k 2 c2 ,
kc

k(1
+
c
,
p=

=
X
X
X
E(X )
1 + c2X
the E k1,k distribution matches E(X ) and c X .

Department of Mathematics and Computer Science

Moment-fitting

32/36

Coefficient of variation greater than 1: Hyperexponential

In case c X 1, fit a H2 ( p1 , p2 ; 1 , 2 ) distribution.
Phase diagram for the Hk ( p1 , . . . , pk ; 1 , . . . , k ) distribution:
1
p1

pk

k
k

Department of Mathematics and Computer Science

Moment-fitting

33/36

But the H2 distribution is not uniquely determined by its first two moments. In applications, the H2 distribution
with balanced means is often used. This means that the normalization
p1
p2
=
1
2
is used. The parameters of the H2 distribution with balanced means and fitting E(X ) and c X ( 1) are given by
s
!
2 1
c
1
X
1+
,
p2 = 1 p1 ,
p1 =
2
2
cX + 1
1 =

Department of Mathematics and Computer Science

2 p1
,
E(X )

2 =

2 p2
.
E(X )

Moment-fitting

34/36

In case c2X 0.5 one can also use a Coxian-2 distribution for a two-moment fit.
Phase diagram for the Coxian-k distribution:
1
1

2
p1

1 p1

k
p2

pk 1

1 p2

1 pk 1

The following parameter set for the Coxian-2 is suggested:

1 = 2/E(X ),

Department of Mathematics and Computer Science

p1 = 0.5/c2X ,

2 = 1 p1 .

Moment-fitting

35/36

Fitting nonnegative discrete distributions

Let X be a random variable on the non-negative integers with mean E(X ) and coefficient of variation c X . Then
it is possible to fit a discrete distribution on E(X ) and c X using the following families of distributions:
Mixtures of Binomial distributions;
Poisson distribution;
Mixtures of Negative-Binomial distributions;
Mixtures of geometric distributions.
This fitting procedure is described in Adan, van Eenige and Resing (see Probability in the Engineering and
Informational Sciences, 9, 1995, pp 623-632).

Department of Mathematics and Computer Science

Fitting distributions
Adequacy of fit
Graphical comparison of fitted and empirical curves;
Statistical tests (goodness-of-fit tests).

Department of Mathematics and Computer Science

36/36