Random Variate Generation

Christos Alexopoulos and Dave Goldsman

Georgia Institute of Technology, Atlanta, GA, USA

11/16/17

1 / 114
Outline

1 Introduction
2 Inverse Transform Method
3 Cutpoint Method
4 Convolution Method
5 Acceptance-Rejection Method
6 Composition Method
7 Special-Case Techniques
8 Multivariate Normal Distribution
9 Generating Stochastic Processes

2 / 114
Introduction

Goal: Use U(0, 1) numbers to generate observations (variates) from

other distributions, and even stochastic processes.

Try to be fast, reproducible.

Discrete distributions, like Bernoulli, Binomial, Poisson, and

empirical
Continuous distributions like exponential, normal (many ways),
and empirical
Multivariate normal
Nonhomogeneous Poisson process
Autoregressive moving average time series
Waiting times
Brownian motion
3 / 114
 Inverse Transform Method

Outline

1 Introduction
2 Inverse Transform Method
3 Cutpoint Method
4 Convolution Method
5 Acceptance-Rejection Method
6 Composition Method
7 Special-Case Techniques
8 Multivariate Normal Distribution
9 Generating Stochastic Processes

4 / 114
 Inverse Transform Method

Inverse Transform Method

Inverse Transform Theorem: Let X be a continuous random variable

with c.d.f. F (x). Then F (X) ∼ U(0, 1).

Proof: Let Y = F (X) and suppose that Y has c.d.f. G(y). Then

G(y) = P(Y ≤ y) = P(F (X) ≤ y)

= P(X ≤ F −1 (y)) = F (F −1 (y)) = y. 2

In the above, we can define the inverse c.d.f. by

F −1 (y) = min[x : F (x) ≥ y] y ∈ [0, 1].

This representation can be applied to continuous or discrete or mixed

distributions (see figure).

5 / 114
Inverse Transform Method

6 / 114
Inverse Transform Method

How do we use this result?

Let U ∼ U(0, 1). Then F (X) = U means that the random variable
F −1 (U ) has the same distribution as X.

So here is the inverse transform method for generating a RV X having

c.d.f. F (x):
1 Sample U from U(0, 1).
2 Return X = F −1 (U ).

We’ll do some continuous examples first, then discrete.

x−a
Example: The U(a, b) distribution, with F (x) = b−a , a ≤ x ≤ b.

Solving (X − a)/(b − a) = U for X, we get X = a + (b − a)U . 2

7 / 114
Inverse Transform Method

Example: The Exp(λ) distribution, with F (x) = 1 − e−λx , x > 0.

Solving F (X) = U for X,

1 1
X = − `n(1 − U ) or X = − `n(U ). 2
λ λ

β
Example: The Weibull distribution, F (x) = 1 − e−(λx) , x > 0.

Solving F (X) = U for X,

1 1
X= [−`n(1 − U )]1/β or X= [−`n(U )]1/β 2.
λ λ

8 / 114
Inverse Transform Method

Example: The triangular (0,1,2) distribution has p.d.f.

(
x if 0 ≤ x < 1
f (x) =
2 − x if 1 ≤ x ≤ 2.

The c.d.f. is
(
x2 /2 if 0 ≤ x < 1
F (x) =
1 − (x − 2)2 /2 if 1 ≤ x ≤ 2.
√
If U < 1/2, we solve X 2 /2 = U to get X = 2U .
If U ≥ 1/2, the only root of 1 − (X − 2)2 /2 = U in [1, 2] is
p
X = 2 − 2(1 − U ).
√
Thus, for example, if U = 0.4, we take X = 0.8. 2

Remark: Do not replace U by 1 − U here!

9 / 114
Inverse Transform Method

Example: The standard normal distribution. Unfortunately, the

inverse c.d.f. Φ−1 (·) does not have an analytical form. This is often a
problem with the inverse transform method.

Easy solution: Do a table lookup. E.g., If U = 0.975, then

Z = Φ−1 (U ) = 1.96. 2

Crude portable approximation (BCNN): The following approximation

gives at least one decimal place of accuracy for
0.00134 ≤ U ≤ 0.98865:

U 0.135 − (1 − U )0.135
Z = Φ−1 (U ) ≈ . 2
0.1975

10 / 114
Inverse Transform Method

Here’s a better portable solution: The following approximation has

absolute error ≤ 0.45 × 10−3 :

c0 + c1 t + c2 t2
 
Z = sign(U − 1/2) t − ,
1 + d1 t + d2 t2 + d3 t3

where sign(x) = 1, 0, −1 if x is positive, zero, or negative,

respectively,
t = {−`n[min(U, 1 − U )]2 }1/2 ,
and

c0 = 2.515517, c1 = 0.802853, c2 = 0.010328,

d1 = 1.432788, d2 = 0.189269, d3 = 0.001308.

11 / 114
Inverse Transform Method

In any case, if Z ∼ Nor(0, 1) and you want X ∼ Nor(µ, σ 2 ), just take

X ← µ + σZ.

Easy Example (Inverse Transform): Suppose you want to generate

X ∼ Nor(3, 16), and you start with U = 0.59. Then

X = µ + σZ = 3 + 4Φ−1 (0.59) = 3 + 4(0.2275) = 3.91. 2

12 / 114
Inverse Transform Method

For discrete distributions, it’s often best to construct a table.

Baby Discrete Example: The Bernoulli(p) distribution.

x P (X = x) F (x) U(0, 1)’s
0 1−p 1−p [0, 1 − p]
1 p 1 (1 − p, 1]

If U ≤ 1 − p, then take X = 0; otherwise, X = 1. For instance, if

p = 0.75 and we generate U = 0.13, we take X = 0. 2

Alternately, we can construct the following “backwards” table (which

isn’t strictly inverse transform, but it’s the one that I usually use).
x P (X = x) U(0, 1)’s
1 p [0, p]
0 1−p (p, 1]

If U ≤ p, take X = 1; otherwise, X = 0. 2
13 / 114
Inverse Transform Method

Example: Suppose we have a slightly less-trivial discrete p.m.f.

x P (X = x) F (x) U(0, 1)’s
−1 0.6 0.6 [0.0,0.6]
2.5 0.3 0.9 (0.6,0.9]
4 0.1 1.0 (0.9,1.0]

Thus, if U = 0.63, we take X = 2.5. 2

Sometimes there’s an easy way to avoid constructing a table.

Example: The discrete uniform distribution on {1, 2, . . . , n},

1
P (X = k) = , 1, 2, . . . , n.
n
Clearly, X = dnU e, where d·e is the ceiling function.

So if n = 10 and U = 0.376, then X = d3.76e = 4. 2

14 / 114
Inverse Transform Method

Example: The geometric distribution with p.m.f. and c.d.f.

f (k) = q k−1 p and F (k) = 1 − q k , k = 1, 2, . . . ,

where q = 1 − p. Thus, after some algebra,

   
k `n(1 − U ) `n(U )
X = min[k : 1 − q ≥ U ] = ∼ .
`n(1 − p) `n(1 − p)

For instance, if p = 0.3 and U = 0.72, we obtain

 
`n(0.28)
X = = 4. 2
`n(0.7)

15 / 114
Inverse Transform Method

Remark: Can also generate Geom(p) by counting Bern(p) trials until

you get a success.

Easy Example: Generate X ∼ Geom(1/6). This is the same thing as

counting the number of dice tosses until a 3 (or any particular
number) comes up, where the Bern(1/6) trials are the i.i.d. dice tosses.
For instance, if you toss 6,2,1,4,3, then you stop on the Bernoulli trial
X = 5, and that’s your answer.

But life isn’t always dice tosses. A general way to generate a Geom(p)
is to count the number of trials until Ui ≤ p. For example, if p = 0.3,
then U1 = 0.71, U2 = 0.96, and U3 = 0.12 implies that X = 3. 2

16 / 114
Inverse Transform Method

Remark: If you have a discrete distribution like Pois(λ) with an

infinite number of values, you could write out table entries until the
c.d.f. is nearly one, generate exactly one U , and then search until you
find X = F −1 (U ), i.e., xi such that U ∈ F (xi−1 ), F (xi ) .

x P (X = x) F (x) U(0, 1)’s

 
x1 f (x1 ) F (x1 ) 0, F (x1 )

x2 f (x2 ) F (x2 ) F (x1 ), F (x2 )

x3 f (x3 ) F (x3 ) F (x2 ), F (x3 )
..
.

17 / 114
Inverse Transform Method

e−2 2x
Example: Suppose X ∼ Pois(2), so that f (x) = x! ,
x = 0, 1, 2, . . ..

x f (x) F (x) U(0, 1)’s

0 0.1353 0.1353 [0, 0.1353]
1 0.2707 0.4060 (0.1353, 0.4060]
2 0.2707 0.6767 (0.4060, 0.6767]
3 0.1804 0.8571 (0.6767, 0.8571]
..
.

For instance, if U = 0.313, then X = 1 2.

18 / 114
 Inverse Transform Method

Continuous Empirical Distributions

If you can’t find a good theoretical distribution to model a certain RV,

you may want to use the empirical c.d.f. of the data, X1 , X2 , . . . , Xn ,
number of Xi ’s ≤ x
F̂n (x) ≡ .
n

Note that F̂n (x) is a step function with jumps of height 1/n (every
time an observation occurs).

Good news: Even though X is continuous, the Glivenko-Cantelli

Lemma says that F̂n (x) → F (x) for all x as n → ∞. So F̂n (x) is a
good approximation to the true c.d.f., F (x).

19 / 114
Inverse Transform Method

The ARENA functions DISC and CONT can be used to generate RV’s
from the empirical c.d.f.’s of discrete and continuous distributions,
respectively.

To do so ourselves, we first define the ordered points,

X(1) ≤ X(2) ≤ · · · ≤ X(n) . For example, if X1 = 4, X2 = 1, and
X3 = 6, then X(1) = 1, X(2) = 4, and X(3) = 6.

20 / 114
Inverse Transform Method

Given that you only have a finite number n of data points, we can turn
the empirical c.d.f. into a continuous RV by using linear interpolation
between the X(i) ’s.



 0 if x < X(1)
i−1 x−X(i)
F (x) = n−1 + (n−1)(X(i+1) −X(i) ) if X(i) ≤ x < X(i+1) , ∀i


 1 if x ≥ X(n)

1 Set F (X) = U ∼ U(0, 1). Let P = (n − 1)U and I = dP e.

2 Solve to get X = X(I) + (P − I + 1) X(I+1) − X(I) .

21 / 114
Inverse Transform Method

Example: Suppose X(1) = 1, X(2) = 4, and X(3) = 6. If U = 0.73,

then P = (n − 1)U = 1.46 and I = dP e = 2. So


X = X(I) + (P − I + 1) X(I+1) − X(I)


= X(2) + (1.46 − 2 + 1) X(3) − X(2)
= 4 + (0.46) 6 − 4 = 4.92. 2

22 / 114
Inverse Transform Method

Check (slightly different way):

(
x−1
0 + 2(4−1) if 1 ≤ x < 4 (i = 1 case)
F (x) = 1 x−4
2 + 2(6−4) if 4 ≤ x < 6 (i = 2 case)

Setting F (X) = U and solving for the two cases, we have

(
1 + 6U if U < 1/2
X =
2 + 4U if U ≥ 1/2

Then U = 0.73 implies X = 2 + 4(0.73) = 4.92. 2

23 / 114
Inverse Transform Method

Example (BCNN): We can use an approximate empirical c.d.f. when

dealing with grouped data. Let’s look at a sample of 100 repair times.

interval freq rel freq approx F̂ (x)

0.25 ≤ x ≤ 0.5 31 0.31 0.31
0.5 < x ≤ 1.0 10 0.10 0.41
1.0 < x ≤ 1.5 25 0.25 0.66
1.5 < x ≤ 2.0 34 0.34 1.00

How to construct a realization X from F̂ (x)? Use the fact that

F̂ (X) ∼ U(0, 1) and do inverse transform, i.e., X = F̂ −1 (U ).

24 / 114
Inverse Transform Method

For instance, if U ∈ (0.66, 1.00] (the last two F̂ (x) entries), then
 F̂ −1 (1.00) − F̂ −1 (0.66) 
X = F̂ −1 (0.66) + (U − 0.66)
1.00 − 0.66
 2.0 − 1.5 
= 1.5 + (U − 0.66) = 1.5 + 1.471(U − 0.66).
1.00 − 0.66
Thus, e.g., if U = 0.83, then X = 1.75. 2

25 / 114
Inverse Transform Method

To do this in general,
1 Generate U .
2 Find the F̂ (x) interval in which U lies, i.e., i such that
ri < U≤ ri+1 . In the above example,
r1 r2 r3 r4 r5
0 0.31 0.41 0.66 1.0
3 Let xi be the left endpoint of the ith X-interval, and let ai be the
reciprocal of the slope of the ith interval.
x1 x2 x3 x4 x5
0.25 0.50 1.0 1.5 2.0

a1 a2 a3 a4 a5
0.81 5.0 2.0 1.47 –
4 X = xi + ai (U − ri ).
26 / 114
 Cutpoint Method

Outline

1 Introduction
2 Inverse Transform Method
3 Cutpoint Method
4 Convolution Method
5 Acceptance-Rejection Method
6 Composition Method
7 Special-Case Techniques
8 Multivariate Normal Distribution
9 Generating Stochastic Processes

27 / 114
 Cutpoint Method

Cutpoint Method
Suppose we want to generate from the discrete distribution

P (X = k) = pk , k = a, a + 1, . . . , b

with large b − a. In this case, inverse transform may have to search a

lot of possible values, and may be inefficient. Let

qk = P (X ≤ k), k = a, a + 1, . . . , b.

For fixed m, the cutpoint method of Fishman and Moore computes

and stores the cutpoints
 
j−1
Ij = min k : qk > , j = 1, . . . , m.
m
These cutpoints help us scan through the list of possible k-values
much more quickly than regular inverse transform.
28 / 114
Cutpoint Method

Here is the algorithm that computes the cutpoints . . .

Algorithm CMSET
j ← 0, k ← a − 1, and A ← 0
While j < m:
While A ≤ j:
k ←k+1
A ← mqk
j ←j+1
Ij ← k

29 / 114
Cutpoint Method

Once the cutpoints are computed, we can use the cutpoint method.

Algorithm CM
Generate U from U(0, 1)
L ← bmU c + 1
X ← IL
While U > qX : X ← X + 1

In short, this algorithm selects an integer L = bmU c + 1 and starts

the search at the value IL . Its correctness results from the fact that

P (IL ≤ X ≤ IL+1 ) = 1 (Im+1 = b).

30 / 114
Cutpoint Method

Let E(Cm ) be the expected number of comparisons until Algorithm

CM terminates. Given L, the maximum number of required
comparisons is IL+1 − IL + 1. Hence,

I2 − I1 + 1 Im+1 − Im + 1
E(Cm ) ≤ + ··· +
m m
b − I1 + m
= .
m

Note that if m ≥ b, then E(Cm ) ≤ (2m − 1)/m ≤ 2.

31 / 114
Cutpoint Method

Example: Consider the distribution

k 1 2 3 4 5 6 7 8
pk 0.01 0.04 0.07 0.15 0.28 0.19 0.21 0.05
qk 0.01 0.05 0.12 0.27 0.55 0.74 0.95 1
For m = 8, we have the following cutpoints:
I1 = min[i : qi > 0] = 1
I2 = min[i : qi > 1/8] = 4
I3 = min[i : qi > 2/8] = 4
I4 = min[i : qi > 3/8] = 5 = I5
I6 = min[i : qi > 5/8] = 6
I7 = min[i : qi > 6/8] = 7 = I8
I9 = 8
For U = 0.219, we have L = b8(0.219)c + 1 = 2, and
X = min[i : I2 ≤ i ≤ I3 , qi ≥ 0.219] = 4. 2
32 / 114
 Convolution Method

Outline

1 Introduction
2 Inverse Transform Method
3 Cutpoint Method
4 Convolution Method
5 Acceptance-Rejection Method
6 Composition Method
7 Special-Case Techniques
8 Multivariate Normal Distribution
9 Generating Stochastic Processes

33 / 114
 Convolution Method

Convolution Method

Convolution refers to adding things up.

Example: Binomial(n, p).

Pn
If X1 , . . . , Xn ∼ i.i.d. Bern(p), then Y = i=1 Xi ∼ Bin(n, p).

We already know how to get Bernoulli RVs via Inverse Transform:

Suppose U1 , . . . , Un are i.i.d. U(0,1). If Ui ≤ p, set Xi = 1;
otherwise, set Xi = 0. Repeat for i = 1, . . . , n. Add up to get Y . 2

For instance, if Y ∼ Bin(3, 0.4) and U1 = 0.63, U2 = 0.17, and

U3 = 0.81, then Y = 0 + 1 + 0 = 1. 2

34 / 114
Convolution Method

Example: Triangular(0,1,2).

It can be shown that if U1 and U2 are i.i.d. U(0, 1), then U1 + U2 is

Tria(0,1,2). (This is easier — but maybe not faster — than our inverse
transform method.) 2

35 / 114
Convolution Method

Pn Erlangn (λ). If X1 , . . . , Xn ∼ i.i.d. Exp(λ), then

Example:
Y = i=1 Xi ∼ Erlangn (λ). By inverse transform,
n n   Y n 
X X −1 −1
Y = Xi = `n(Ui ) = `n Ui .
λ λ
i=1 i=1 i=1

This only takes one natural log evaluation, so it’s pretty efficient. 2

Example: A crude “desert island” Nor(0,1) approximate generator

(which I wouldn’t use).

Suppose that U1 , . . . , Un are i.i.d. U(0,1), and let Y = ni=1 Ui .

P
For large n, the CLT implies that Y ≈ Nor(n/2, n/12).

In particular, let’s choose n = 12, and assume that it’s “large.” Then
12
X
Y −6 = Ui − 6 ≈ Nor(0, 1). 2
i=1 36 / 114
Convolution Method

Other convolution-related tidbits:

Did you know. . . ?

Pn
If X1 , . . . , Xn are i.i.d. Geom(p), then i=1 Xi ∼ NegBin(n, p).
Pn 2
If Z1 , . . . , Zn are i.i.d. Nor(0,1), then i=1 Zi ∼ χ2 (n).

If X1 , . . . , Xn are i.i.d. Cauchy, then X̄ ∼ Cauchy (this is kind of like

getting nowhere fast!).

37 / 114
 Acceptance-Rejection Method

Outline

1 Introduction
2 Inverse Transform Method
3 Cutpoint Method
4 Convolution Method
5 Acceptance-Rejection Method
6 Composition Method
7 Special-Case Techniques
8 Multivariate Normal Distribution
9 Generating Stochastic Processes

38 / 114
 Acceptance-Rejection Method

Acceptance-Rejection Method

Motivation: The majority of c.d.f.’s cannot be inverted efficiently. A-R

samples from a distribution that is “almost” the one we want, and then
adjusts by “accepting” only a certain proportion of those samples.

Baby Example: Generate a U( 23 , 1) RV. (You would usually do this

via inverse transform, but what the heck!) Here’s the A-R algorithm:

1. Generate U ∼ U(0, 1).

2. If U ≥ 32 , ACCEPT X ← U . O’wise, REJECT and go to Step 1.

39 / 114
Acceptance-Rejection Method

Notation: Suppose we want to simulate a continuous RV X with p.d.f.

f (x), but that it’s difficult to generate directly. Also suppose that we
can easily generate a RV having p.d.f. h(x) ≡ t(x)/c, where t(x)
majorizes f (x), i.e.,

t(x) ≥ f (x), x ∈ R,

and Z ∞ Z ∞
c≡ t(x) dx ≥ f (x) dx = 1,
−∞ −∞
where we assume that c < ∞.

40 / 114
Acceptance-Rejection Method

Theorem (von Neumann 1951): Define g(x) ≡ f (x)/t(x) and note

that 0 ≤ g(x) ≤ 1 for all x. Let U ∼ U(0, 1), and let Y be a RV
(independent of U ) with p.d.f. h(y) = t(y)/c. If U ≤ g(Y ), then Y
has (conditional) p.d.f. f (y).

This suggests the following “acceptance-rejection” algorithm . . .

Algorithm A-R
Repeat
Generate U from U(0, 1)
Generate Y from h(y) (independent of U )
f (Y ) f (Y )
until U ≤ g(Y ) = t(Y ) = c h(Y )
Return X ← Y

41 / 114
Acceptance-Rejection Method

42 / 114
Acceptance-Rejection Method

Proof that X has p.d.f. f (x).

Let A be the “Acceptance” event. The c.d.f. of X is

P (A, Y ≤ x)
P (X ≤ x) = P (Y ≤ x|A) = . (1)
P (A)

Then

P (A|Y = y) = P (U ≤ g(Y )|Y = y)

= P (U ≤ g(y)|Y = y)
= P (U ≤ g(y)) (U and Y are independent)
= g(y) (U is uniform). (2)

43 / 114
Acceptance-Rejection Method

By the law of total probability,

Z ∞
P (A, Y ≤ x) = P (A, Y ≤ x|Y = y)h(y) dy
−∞
Z x
= P (A|Y = y)h(y) dy
−∞
Z x
1
= P (A|Y = y)t(y) dy
c −∞
1 x
Z
= g(y)t(y) dy (by (2))
c −∞
1 x
Z
= f (y) dy. (3)
c −∞

44 / 114
Acceptance-Rejection Method

Letting x → ∞, we have
Z ∞
1 1
P (A) = f (y) dy = . (4)
c −∞ c

Then (1), (3), and (4) imply

x
P (A, Y ≤ x)
Z
P (X ≤ x) = = f (y) dy,
P (A) −∞

so that, by the Fundamental Theorem of Calculus, the p.d.f. of X is

Z x
d P (X ≤ x) d
= f (y) dy = f (x). 2
dx dx −∞

45 / 114
Acceptance-Rejection Method

There are two main issues:

The ability to quickly sample from h(y).
c ≥ 1 ought to be as close to 1 as possible — i.e., t(x) must be
“close” to f (x). That’s because
1
P (U ≤ g(Y )) =
c
and the number of trials until “success” [U ≤ g(Y )] is
Geom(1/c), so that the mean number of trials is c.

46 / 114
Acceptance-Rejection Method

Example (Law 2015): Generate a RV with p.d.f.

f (x) = 60x3 (1 − x)2 , 0 ≤ x ≤ 1. Can’t invert this analytically.

Note that the maximum occurs at x = 0.6, and f (0.6) = 2.0736.

Using the majorizing function

0≤x≤1
t(x) = 2.0736,
R1
(which isn’t actually very efficient), get c = 0 t(x) dx = 2.0736, so
t(x)
h(x) = = 1, 0≤x≤1 (i.e., a U(0,1) p.d.f.)
c
and
f (x)
g(x) = = 60x3 (1 − x)2 /2.0736.
t(x)
E.g., if we generate U = 0.13 and Y = 0.25, then it turns out that
U ≤ g(Y ) = 60Y 3 (1 − Y )2 /2.0736, so we take X ← 0.25. 2
47 / 114
Acceptance-Rejection Method

Example (Ross): Generate a standard half-normal RV, with p.d.f.

2 2
f (x) = √ e−x /2 , x ≥ 0.
2π

Using the majorizing function

r
2e −x
t(x) = e ≥ f (x) for all x ≥ 0,
π
we get r Z ∞ r
2e −x 2e
c = e dx = = 1.3155,
π 0 π

t(x)
h(x) = = e−x (easy Exp(1) p.d.f.),
c
and
f (x) 2
g(x) = = e−(x−1) /2 . 2
t(x)
48 / 114
Acceptance-Rejection Method

We can use the half-normal result to generate a Nor(0, 1) variate.

Generate U from U(0, 1).

Generate X from the half-normal distribution.
Return (
−X if U ≤ 1/2
Z =
X if U > 1/2.

Reminder: We can then generate Nor(µ, σ 2 ) RV by using the obvious

transformation µ + σZ.

49 / 114
Acceptance-Rejection Method

Example: The gamma distribution with p.d.f.:

λβ xβ−1 −(λx)β
f (x) = e , x > 0.
Γ(β)

We’ll split the task of generating gamma RV’s via the A-R algorithm
into two cases depending on the magnitude of the shape parameter:
β < 1 and β ≥ 1. . .

50 / 114
Acceptance-Rejection Method

If β < 1, we’ll use the following A-R algorithm with c ≤ 1.39:

Algorithm GAM1
b ← (e + β)/e (e is the base of `n)
While (True)
Generate U from U(0, 1); W ← bU
If W < 1
Y ← W 1/β ; Generate V from U(0, 1)
If V ≤ e−Y : Return X = Y /λ
Else
Y ← −`n[(b − W )/β]
Generate V from U(0, 1)
If V ≤ Y β−1 : Return X = Y /λ 51 / 114
Acceptance-Rejection Method

If β ≥ 1, the value of
pc for the following A-R algorithm decreases
from 4/e = 1.47 to 4/π = 1.13 as β increases from 1 to ∞.

Algorithm GAM2
a ← (2β − 1)−1/2 ; b ← β − `n(4); c ← β + a−1 ; d ← 1 + `n(4.5)
While (True)
Generate U1 , U2 from U(0, 1)
V ← a`n[U1 /(1 − U1 )]
Y ← βeV ; Z ← U12 U2
W ← b + cV − Y
If W + d − 4.5Z ≥ 0: Return X = Y /λ
Else
If W ≥ `n(Z): Return X = Y /λ
52 / 114
Acceptance-Rejection Method

Example: The Poisson distribution with probability mass function

λn
P (X = n) = e−λ , n = 0, 1, . . .
n!

We’ll use a variation of A-R to generate a realization of X. The

algorithm will go through a set of equivalent statements to arrive at a
rule that gives X = n.

Recall that, by definition, X = n if we observe exactly n arrivals

from a Poisson(λ) process in one time unit.

Define Ai as the ith interarrival time from a Pois(λ) process.

53 / 114
Acceptance-Rejection Method

X = n ⇔ See exactly n Pois(λ) arrivals by t = 1

Xn n+1
X
⇔ Ai ≤ 1 < Ai
i=1 i=1
n   n+1
X  −1 
X −1
⇔ `n(Ui ) ≤ 1 < `n(Ui )
λ λ
i=1 i=1
n n+1
! !
−1 Y −1 Y
⇔ `n Ui ≤ 1 < `n Ui
λ λ
i=1 i=1
n
Y n+1
Y
⇔ Ui ≥ e−λ > Ui . (5)
i=1 i=1

54 / 114
Acceptance-Rejection Method

The following A-R algorithm samples U(0,1)’s

Qn+1until (5) becomes true,
−λ
i.e., until the first time n such that e > i=1 Ui .

Algorithm POIS1
a ← e−λ ; p ← 1; X ← −1
Until p < a
Generate U from U(0, 1)
p ← pU ; X ← X + 1
Return X

55 / 114
Acceptance-Rejection Method

Example (BCNN): Apply Algorithm POIS1 to obtain a Pois(2)

variate.
Qn+1
Sample until e−λ = 0.1353 > i=1 Ui .
Qn+1
n Un+1 i=1 Ui Stop?
0 0.3911 0.3911 No
1 0.9451 0.3696 No
2 0.5033 0.1860 No
3 0.7003 0.1303 Yes

Thus, we take X = 3. 2

Remark: An easy argument says that the expected number of U ’s that

are required to generate one realization of X is E[X + 1] = λ + 1.
56 / 114
Acceptance-Rejection Method

Remark: If λ ≥ 20, we can use the normal approximation

X −λ
√ ≈ Nor(0, 1).
λ

Algorithm POIS2 (for λ ≥ 20)

√
α← λ
Generate Z from Nor(0, 1)
Return X = max(0, bλ + αZ + 0.5c) (Note that this employs a
“continuity correction.”)
√
E.g., if λ = 30 and Z = 1.46, then X = b30.5 + 30(1.46)c = 38.

57 / 114
Acceptance-Rejection Method

Remark: Of course, another way to generate a Pois(λ) is simply to

table the c.d.f. values like we did in an earlier discrete inverse
transform example. This may be more efficient and accurate than the
above methods — which is not to say that the A-R method isn’t clever
and pretty!

58 / 114
 Composition Method

Outline

1 Introduction
2 Inverse Transform Method
3 Cutpoint Method
4 Convolution Method
5 Acceptance-Rejection Method
6 Composition Method
7 Special-Case Techniques
8 Multivariate Normal Distribution
9 Generating Stochastic Processes

59 / 114
 Composition Method

Composition Method
Idea: Suppose a RV actually comes from two RV’s (sort of on top of
each other). E.g., your plane can leave the airport gate late for two
reasons — air traffic delays and maintenance delays, which compose
the overall delay time.

The goal is to generate a RV with c.d.f.

∞
X
F (x) = pj Fj (x),
j=1
P
where pj > 0 for all j, j pj = 1, and the Fj (x)’s are “easy” c.d.f.’s
to generate from.

Generate a positive integer J such that P (J = j) = pj for all j.

Return X from c.d.f. FJ (x).
60 / 114
Composition Method

Proof that X has c.d.f. F (x): By the law of total probability,

∞
X
P (X ≤ x) = P (X ≤ x|J = j)P (J = j)
j=1
∞
X
= Fj (x)pj = F (x). 2
j=1

61 / 114
Composition Method

Example: Laplace distribution (exponential distribution reflected off

of the y-axis)
( (
1 x 1 x
2e , x<0 2e , x<0
f (x) ≡ 1 −x
and F (x) ≡ 1 −x
2e , x > 0 1 − 2e , x > 0

Meanwhile, let’s decompose X into “negative exponential” and

regular exponential distributions:
( (
ex if x < 0 0 if x < 0
F1 (x) ≡ and F2 (x) ≡ −x
1 if x > 0 1−e if x > 0

62 / 114
Composition Method

Then
1 1
F (x) = F1 (x) + F2 (x),
2 2
so that we generate from F1 (x) half the time, and F2 (x) half.

We’ll use inverse transform to solve F1 (X) = eX = U for X half the

time, and F2 (x) = 1 − e−X = U the other half. Then
(
`n(U ) w.p. 1/2
X ← 2
−`n(U ) w.p. 1/2

63 / 114
 Special-Case Techniques

Outline

1 Introduction
2 Inverse Transform Method
3 Cutpoint Method
4 Convolution Method
5 Acceptance-Rejection Method
6 Composition Method
7 Special-Case Techniques
8 Multivariate Normal Distribution
9 Generating Stochastic Processes

64 / 114
 Special-Case Techniques

Special-Case Techniques

Box–Muller Method: Here’s a nice, easy way to generate standard

normals.

Theorem: If U1 , U2 are i.i.d. U(0,1), then

p
Z1 = −2`n(U1 ) cos(2πU2 )
p
Z2 = −2`n(U1 ) sin(2πU2 )

are i.i.d, Nor(0,1).

Note that the trig calculations must be done in radians.

Proof Someday soon. 2

65 / 114
Special-Case Techniques

Some interesting corollaries follow directly from Box–Muller.

Example: Note that

Z12 + Z22 ∼ χ2 (1) + χ2 (1) ∼ χ2 (2).

But

Z12 + Z22 = −2`n(U1 )(cos2 (2πU2 ) + sin2 (2πU2 ))

= −2`n(U1 )
∼ Exp(1/2).

Thus, we’ve just proven that

χ2 (2) ∼ Exp(1/2). 2

66 / 114
Special-Case Techniques

Example: Note that

Z2 /Z1 ∼ Nor(0, 1)/Nor(0, 1) ∼ Cauchy ∼ t(1).

But
p
−2`n(U1 ) sin(2πU2 )
Z2 /Z1 = p = tan(2πU2 ).
−2`n(U1 ) cos(2πU2 )
Thus, we’ve just proven that

tan(2πU ) ∼ Cauchy (and, similarly, cot(2πU ) ∼ Cauchy).

Similarly,

Z22 /Z12 = tan2 (2πU ) ∼ t2 (1) ∼ F (1, 1).

(Did you know that?)

67 / 114
Special-Case Techniques

Polar Method — a little faster than Box–Muller.

1. Generate U1 , U2 i.i.d. U(0,1).

Let Vi = 2Ui − 1, i = 1, 2, and W = V12 + V22 .

2. If W > 1, reject and go back to Step 1.

p
O’wise, let Y = −2`n(W )/W , and accept Zi ← Vi Y , i = 1, 2.

Then Z1 , Z2 are i.i.d. Nor(0,1).

68 / 114
Special-Case Techniques

Order Statistics

Suppose that X1 , X2 , . . . , Xn are i.i.d. from some distribution with

c.d.f. F (x), and let Y = min{X1 , . . . , Xn } with c.d.f. G(y). (Y is
called the first order stat.) Can we generate Y using just one U(0,1)?
Yes! since the Xi ’s are i.i.d., we have

G(y) = 1 − P (Y > y) = 1 − P (min Xi > y)

i
= 1 − P (all Xi ’s > y) = 1 − [P (X1 > y)]n
= 1 − [1 − F (y)]n .

69 / 114
Special-Case Techniques

Now do Inverse Transform: set G(Y ) = U and solve for Y . After a

little algebra, get (don’t be afraid). . .
 
Y = F −1 1 − (1 − U )1/n .

Example: Suppose X1 , . . . , Xn ∼ Exp(λ). Then

G(y) = 1 − (e−λy )n = 1 − e−nλy .

Thus, Y = mini {Xi } ∼ Exp(nλ). (This is not a surprising result if

you’ve taken a stochastic processes class.) In any case,
Y = − nλ1
`n(U ). 2

We can do the same kind of thing for Z = maxi Xi .

70 / 114
Special-Case Techniques

Other Quickies

If X ∼ Nor(µ, σ 2 ), then eX has the lognormal distribution.

χ 2
Pn(n) distribution:
2 2
If Z1 , Z2 , . . . , Zn are i.i.d. Nor(0,1), then
i=1 Zi ∼ χ (n).

t(n) distribution: If Z ∼ Nor(0, 1) and Y ∼ χ2 (n), and Z and Y are

independent, then
Z
p ∼ t(n).
Y /n

Note that t(1) is the Cauchy distribution.

F (n, m) distribution: If X ∼ χ2 (n) and Y ∼ χ2 (m) and X and Y

are independent, then (X/n)/(Y /m) ∼ F (n, m).

71 / 114
 Multivariate Normal Distribution

Outline

1 Introduction
2 Inverse Transform Method
3 Cutpoint Method
4 Convolution Method
5 Acceptance-Rejection Method
6 Composition Method
7 Special-Case Techniques
8 Multivariate Normal Distribution
9 Generating Stochastic Processes

72 / 114
 Multivariate Normal Distribution

Bivariate Normal Distribution

The random vector (X, Y ) has the bivariate normal distribution with
means µX = E[X] and µY = E[Y ], variances σX 2 = Var(X) and

σY2 = Var(Y ), and correlation ρ = Corr(X, Y ) if it has joint p.d.f.

(  2 )
1 − zX (x) + zY2 (y) − 2ρzX (x)zY (y)
f (x, y) = exp ,
2(1 − ρ2 )
p
2πσX σY 1 − ρ2

where zX (x) ≡ (x − µX )/σX and zY (y) ≡ (y − µY )/σY .

For example, heights and weights of people can be modeled as

bivariate normal.

This distribution is easily generalized to the multivariate case.

73 / 114
Multivariate Normal Distribution

74 / 114
 Multivariate Normal Distribution

Multivariate Normal Distribution

The random vector X = (X1 , . . . , Xk )T has the multivariate normal

distribution with mean vector µ = (µ1 , . . . , µk )T and k × k
covariance matrix Σ = (σij ) if it has p.d.f.

(x − µ)T Σ−1 (x − µ)
 
1
f (x) = exp − , x ∈ Rk .
(2π)k/2 |Σ|1/2 2

It turns out that

E[Xi ] = µi , Var(Xi ) = σii , Cov(Xi , Xj ) = σij .

Notation: X ∼ Nork (µ, Σ).

75 / 114
Multivariate Normal Distribution

In order to generate X, let’s start with a vector Z = (Z1 , . . . , Zk ) of

i.i.d. Nor(0,1) RV’s. That is, suppose Z ∼ Nork (0, I), where I is the
k × k identity matrix, and 0 is simply a vector of 0’s.

Suppose we can find the (lower triangular) Cholesky matrix C such

that Σ = CC 0 .

Similar to the one-dimensional result in which a linear function of a

normal RV is still normal, it can be shown that X = µ + CZ is
multivariate normal with mean µ and covariance matrix

Σ ≡ Cov(X) = Cov(µ + CZ)

= C Cov(Z)C 0 = CIC 0 = CC 0 .

76 / 114
Multivariate Normal Distribution

For k = 2, we can derive after a teensy bit of algebra,

 √ 
σ11 0
C =  σ q 2
σ12
.
√ 12
σ11 σ 22 − σ11

Since X = µ + CZ, we have

√
X1 = µ1 + c11 Z1 = µ1 + σ11 Z1
s
σ12 2
σ12
X2 = µ2 + c21 Z1 + c22 Z2 = µ2 + √ Z1 + σ22 − Z2
σ11 σ11

77 / 114
Multivariate Normal Distribution

Example: Suppose we want to generate a bivariate normal random

vector (X1 , X2 ) with mean µ = (2, 5) and covariance matrix
!
2 −1
Σ = .
−1 4

First of all, let’s find the Cholesky matrix C such that Σ = CC 0 ,

 √  !
2 0 1.4142 0
C =  −1 q 1
 = .
√
2
4 − 2 −0.7071 1.8708

Suppose we sample Z1 = 1.2 and Z2 = −0.3. Then

X1 = µ1 + c11 Z1 = 2 + 1.414(1.2) = 3.697

X2 = µ2 + c21 Z1 + c22 Z2
= 5 − 0.707(1.2) + 1.871(−0.3) = 2.742. 2
78 / 114
Multivariate Normal Distribution

The following algorithm computes C for general k:

Algorithm LTM

For i = 1, . . . , k,

For j = 1, . . . , i − 1,
 Pj−1 
cij ← σij − `=1 ci` cj` /cjj

cji ← 0
 1/2
cii ← σii − i−1 2
P
c
`=1 i`

79 / 114
Multivariate Normal Distribution

Once C has been computed, the multivariate normal RV

X = µ + CZ can easily be generated:

1. Generate Z1 , Z2 , . . . , Zk ∼ i.i.d. Nor(0, 1).

Pi
2. Let Xi ← µi + j=1 cij Zj , i = 1, 2, . . . , k.

3. Return X = (X1 , X2 , . . . , Xk ).

80 / 114
 Generating Stochastic Processes

Outline

1 Introduction
2 Inverse Transform Method
3 Cutpoint Method
4 Convolution Method
5 Acceptance-Rejection Method
6 Composition Method
7 Special-Case Techniques
8 Multivariate Normal Distribution
9 Generating Stochastic Processes

81 / 114
 Generating Stochastic Processes

Generating Stochastic Processes

We’ll now talk about. . .

Markov Chains
Poisson Arrivals
Nonhomogeneous Poisson Arrivals
Time Series
* MA(1)
* AR(1)
* ARMA(p, q)
* EAR(1)
* Autoregressive Pareto
M/M/1 Queue Waiting Times
Brownian Motion

82 / 114
Generating Stochastic Processes

Markov Chains

Consider a time series having a certain number of states (e.g., sun /

rain) that can transition from day to day.

Example: On Monday it’s sunny, on Tues and Weds, it’s rainy, etc.

Informally speaking, if today’s weather only depends on yesterday,

then you have a Markov chain.

Just do a simple example. Let Xi = 0 if it rains on day i; otherwise,

Xi = 1. Denote the day-to-day transition probabilities by

Pjk = P (state k on day i | state j on day i − 1), j, k = 0, 1.

83 / 114
Generating Stochastic Processes

Suppose that the probability state transition matrix is

!
0.7 0.3
P = ,
0.4 0.6

e.g., P01 = P (R today  S yesterday) = 0.3.

If it rains on Monday (rainy days and Mondays always get me down),

let’s simulate the rest of the work week.

To do so, we’ll run daily Bern(Pj0 ) trials to determine if it’s going to

rain today given yesterday’s weather, where

Pj0 ≡ P (R today  yesterday’s weather was Xi−1 = j).

84 / 114
Generating Stochastic Processes

j = Xi−1 Pj0 Ui Ui < Pj0 ? R/S

M – – – – R
Tu 0 P00 = 0.7 0.62 Y R
W 0 P00 = 0.7 0.03 Y R
Th 0 P00 = 0.7 0.77 N S
F 1 P10 = 0.4 0.91 N S

85 / 114
Generating Stochastic Processes

Generating Poisson Arrivals

When the arrival rate is a constant λ, the interarrivals of a Poisson(λ)

process are i.i.d. Exp(λ), and the arrival times are:
1
T0 ← 0 and Ti ← Ti−1 − `n(Ui ), i ≥ 1.
λ

Now suppose that we want to generate a fixed number n of PP(λ)

arrivals in a fixed time interval [a, b]. To do so, we note a theorem
stating that the joint distribution of the n arrivals is the same as the
joint distribution of the order statistics of n i.i.d. U(a, b) RV’s.

Generate i.i.d. U1 , . . . , Un from U(0, 1)

Sort the Ui ’s: U(1) < U(2) < · · · < U(n)

Set the arrival times to Ti ← a + (b − a)U(i)

86 / 114
Generating Stochastic Processes

Nonhomogeneous Poisson Process — nonstationary arrivals

Same assumptions as regular Poisson process except the arrival rate λ

isn’t a constant, so stationary increments doesn’t apply.

Let
λ(t) = rate (intensity) function at time t,
N (t) = number of arrivals during [0, t].

Then Z b 
N (b) − N (a) ∼ Poisson λ(t) dt .
a

87 / 114
Generating Stochastic Processes

Example: Suppose that the arrival pattern to the Waffle House over a
certain time period is a NHPP with λ(t) = t2 . Find the probability
that there will be exactly 4 arrivals between times t = 1 and 2.

First of all, the number of arrivals in that time interval is

Z 2 
2
N (2) − N (1) ∼ Pois t dt ∼ Pois(7/3).
1

Thus,

e−7/3 (7/3)4
2


P N (2) − N (1) = 4 = = 0.120.
4!

88 / 114
Generating Stochastic Processes

Incorrect NHPP Algorithm [it can “skip” intervals with large λ(t)]

T0 ← 0; i ← 0

Repeat

Generate U from U(0, 1)

1
Ti+1 ← Ti − λ(Ti ) `n(U )

i←i+1

Don’t use this algorithm! — the arrival rate λ(Ti ) doesn’t keep pace
with changes in λ(t) that might occur between the current arrival at
time Ti and the next arrival at time Ti+1 .

89 / 114
Generating Stochastic Processes

Whatever shall we do?

The Thinning Algorithm. . .

Assumes that λ? ≡ maxi λ(t) < ∞;
Generates potential arrivals at the rate λ? ; and
Accepts (keeps) a potential arrival at time t with probability
λ(t)/λ? .

90 / 114
Generating Stochastic Processes

The figure below illustrates two potential arrivals at times t1 and t2 .

We keep potential arrival i with probability λ(ti )/λ? .

Let Ti denote the ith arrival that we actually keep. E.g., if we reject
the first potential arrival but keep the second, then T1 ← t2 .
91 / 114
Generating Stochastic Processes

Thinning Algorithm
T0 ← 0; i ← 0
Repeat
t ← Ti
Repeat
Generate U , V from U(0, 1)
1
t←t− λ? `n(U )

until V ≤ λ(t)/λ?
i←i+1
Ti ← t

92 / 114
Generating Stochastic Processes

Example: Let’s simulate NHPP arrivals in which λ(t) = 1 + sin(t/5),

so that λ? = 2. The figure below plots arrival points.

The t-values of the blue dots represent actual arrival times; and the
y-values are placed to illustrate the values of λ(t) at those points.

93 / 114
Generating Stochastic Processes

Remark: Thinning is an Acceptance-Rejection method. Very nice, but

if λ? is significantly higher than most of the values that λ(t) takes,
then thinning can be inefficient.

Remark: Arena and various other simulation languages use a different

method when generating NHPP arrivals from an Arrival Schedule.

That’s because such a schedule gives λ(t) as a step function that only
changes occasionally, say, every hour. In this case, it’s possible to take
clever advantage the exponential’s memoryless property to avoid the
use of thinning.

94 / 114
Generating Stochastic Processes

First-Order Moving Average Process

An MA(1) is a time series process is defined by

Yi = εi + θεi−1 , for i = 1, 2, . . .,

where θ is a constant and the εi ’s are i.i.d. Nor(0, 1) RV’s that are
independent of Y0 .

The MA(1) is a popular tool for modeling and detecting trends.

95 / 114
Generating Stochastic Processes

It’s easy to show that Y1 , Y2 , . . . are all Nor(0, 1 + θ2 ). But the Yi ’s

aren’t independent!

Define the covariance function, Rk ≡ Cov(Yi , Yi+k ),

k = 0, ±1, ±2, . . ..

For the MA(1), we have R0 = Var(Yi ) = 1 + θ2 ,

R1 = Cov(Yi , Yi+1 ) = θ, and Rk = 0 for k ≥ 2. So the covariances
die off pretty quickly.

How to generate? Start with ε0 ∼ Nor(0, 1). Then generate

ε1 ∼ Nor(0, 1) to get Y1 , ε2 ∼ Nor(0, 1) to get Y2 , etc.

96 / 114
Generating Stochastic Processes

First-Order Autoregressive Process

An AR(1) process is defined by

Yi = φYi−1 + εi , for i = 1, 2, . . .,

where −1 < φ < 1, Y0 ∼ Nor(0, 1), and the εi ’s are i.i.d.

Nor(0, 1 − φ2 ) RV’s that are independent of Y0 .

This is used to model lots of real-world stuff.

97 / 114
Generating Stochastic Processes

As defined, the Yi ’s are all Nor(0,1), but (similar to the MA(1)), they
aren’t independent.

The AR(1) has covariance function

Rk = Cov(Yi , Yi+k ) = φ|k| . for all k = 0, ±1, ±2, . . ..

If φ is close to one, you get highly positively correlated Yi ’s. If φ is

close to zero, the Yi ’s are nearly independent.

pgenerate? Start with Y0 ∼ Nor(0, 1) and

How to
ε1 ∼ 1 − φ2 Nor(0, 1) to get Y1 = φY0 + ε1 .
p
Then generate ε2 ∼ 1 − φ2 Nor(0, 1) to get Y2 = φY1 + ε2 , etc.

98 / 114
Generating Stochastic Processes

99 / 114
Generating Stochastic Processes

ARMA(p, q) Process

An obvious generalization of the MA(1) and AR(1) processes is the

ARMA(p, q), which consists of a pth order AR and a qth order MA,
which we will simply define (without stating properties):
p
X q
X
Yi = φj Yi−j + εi + θj εi−j , i = 1, 2, . . . ,
j=1 j=1

where the φj ’s, θj ’s, and Var(εi ), as well as the initial RVs
Y0 , Y−1 , . . . , Y1−p , are chosen so as to assure that the process doesn’t
explode.

Such processes are used in a variety of modeling and forecasting

applications.

100 / 114
Generating Stochastic Processes

First-Order Exponential Autoregressive Process

An EAR(1) process (Lewis 1980) is defined by

(
φYi−1 , w.p. φ
Yi = ,
φYi−1 + εi , w.p. 1 − φ

for i = 1, 2, . . ., where 0 ≤ φ < 1, Y0 ∼ Exp(1), and the εi ’s are i.i.d.

Exp(1) RV’s that are independent of Y0 .

The EAR(1) has the same covariance structure as the AR(1), except
that 0 ≤ φ < 1, that is, Cov(Yi , Yi+k ) = φ|k| for all
k = 0, ±1, ±2, . . ..

101 / 114
Generating Stochastic Processes

102 / 114
Generating Stochastic Processes

Autoregressive Pareto (ARtoP) Process

Now let’s see how to generate a series of correlated Pareto RV’s. First
of all, a RV X has the Pareto distribution with parameters λ > 0 and
β > 0 if it has c.d.f.

FX (x) = 1 − (λ/x)β , for x ≥ λ.

The Pareto is a heavy-tailed distribution that has a variety of uses in

statistical modeling, and for which

βλ
E[X] = for β > 1 and
β−1

βλ2
Var(X) = for β > 2.
(β − 1)2 (β − 2)

103 / 114
Generating Stochastic Processes

In order to obtain the ARP process, let’s start off with a regular AR(1)
with normal noise,

Yi = ρYi−1 + εi , for i = 1, 2, . . .,

where −1 < ρ < 1, Y0 ∼ Nor(0, 1), and the εi ’s are i.i.d.

Nor(0, 1 − ρ2 ) and independent of Y0 . Note that Y0 , Y1 , Y2 , . . . are
marginally Nor(0,1) but correlated.

Feed this process into the Nor(0,1) c.d.f. Φ(·) to obtain correlated
Unif(0,1) RV’s, Ui = Φ(Yi ), i = 1, 2, . . ..

Now feed the correlated Ui ’s into the inverse of the Pareto c.d.f. to
obtain correlated Pareto RV’s:
λ
Xi = FX−1 (Ui ) = FX−1 (Φ(Yi )) = , i = 1, 2, . . . .
[1 − Φ(Yi )]1/β
104 / 114
Generating Stochastic Processes

M/M/1 Queue

Consider a single-server queue with customers arriving according to a

Poisson(λ) process, standing in line with a FIFO discipline, and then
getting served in an Exp(µ) amount of time.

Let Ii+1 denote the interarrival time between the ith and (i + 1)st
customers; let Si be the ith customer’s service time; and let WiQ
denote the ith customer’s wait before service.

Lindley gives a very nice way to generate a series of waiting times for
this simple example:
Q
Wi+1 = max{WiQ + Si − Ii+1 , 0}.

And similarly, the total time in system, Wi = WiQ + Si , is

Wi+1 = max{Wi − Ii+1 , 0} + Si+1 .
105 / 114
Generating Stochastic Processes

Brownian Motion

Discovered by Robert Brown; analyzed rigorously by Einstein;

mathematicals established by Wiener (also called Wiener process).

Widely used in everything from financial analysis to queueing theory

to statistics to other OR/IE application areas. Incredibly important.

106 / 114
Generating Stochastic Processes

Definition: The continuous-time stochastic process {W(t), t ≥ 0} is a

standard Brownian motion process if:
1 W(0) = 0.
2 W(t) ∼ Nor(0, t).
3 {W(t), t ≥ 0} has stationary and independent increments.

Increments: Anything like W(b) − W(a).

Stationary increments informally means that the distribution of

W(t + h) − W(t) only depends on h.

Independent increments: If a < b < c < d, then W(d) − W(c) is

indep of W(b) − W(a).

107 / 114
Generating Stochastic Processes

The definition gives rise to some cool properties.

Theorem: Cov(W(s), W(t)) = min(s, t).

Proof: Suppose that s < t. Then

 
Cov(W(s), W(t)) = Cov W(s), W(t) − W(s) + W(s)
 
= Cov W(s), W(t) − W(s) + Var(W(s))
= 0 + s = s,

by independent increments and the fact that W(s) ∼ Nor(0, s). 2

108 / 114
Generating Stochastic Processes

How do you generate BM on a computer?

Suppose that Y1 , Y2 , . . . is any sequence of i.i.d. RV’s with mean zero

and variance 1. (To some extent, the Yi ’s don’t even have to be
independent!) Donsker’s Central Limit Theorem says that
bntc
1 X d
√ Yi −→ W(t) as n → ∞,
n
i=1
d
where −→ denotes convergence in distribution as n gets big.

Remark: The regular CLT is just for the case t = 1, so that the thing
converges to W(1) ∼ Nor(0, 1).

Donsker is a much more general process CLT. Now the thing

converges to an entire BM process {W(t)} for all t — not just a
single, wussy Nor(0,1) RV.
109 / 114
Generating Stochastic Processes

Here’s a way to construct BM:

One choice that works well is to take Yi = ±1, each with probability
1/2. Take n at least 100, t = 1/n, 2/n, . . . , n/n, and calculate
W(1/n), W(2/n), . . . , W(n/n).

Another choice is simply to take Yi ∼ Nor(0, 1), i = 1, 2, . . ..

Exercise: Let’s construct some BM! First, pick some “large” value of
n and start with W(0) = 0. Then
i i − 1 Yi
W = W +√ , i = 1, 2, . . . .
n n n

110 / 114
Generating Stochastic Processes

Here are some miscellaneous properties of Brownian Motion:

BM is continuous everywhere, but has no derivatives! (Pretty
deep result.)
R1
Area under W(t) is normal: 0 W(t) dt ∼ Nor(0, 13 ).

A Brownian bridge, B(t), is conditioned BM such that

W(0) = W(1) = 0.

Cov(B(s), B(t)) = min(s, t) − st.

R1 1
0 B(t) dt ∼ Nor(0, 12 ).

111 / 114
Generating Stochastic Processes

Geometric Brownian Motion

The GBM process

σ2
  
S(t) = S(0) exp µ− t + σW(t) , t ≥ 0,
2
is often used to model stock prices, where µ is related to the “drift” of
the stock price, σ is its volatility, and S(0) is the initial price.

In addition, we can use GBM to estimate option prices. E.g., a

European call option C permits its owner, who pays an up-front fee
for the privilege, to purchase the stock at a pre-agreed strike price k,
at a pre-determined expiry date T . Its “value” is
h
+ i
E[C] = e−rT E S(T ) − k ,

where x+ = max{0, x} and µ ← r, the “risk-free” interest rate.

112 / 114
Generating Stochastic Processes

Exercise: Let’s estimate the value E[C] of a stock option. Pick your
favorite values of r, σ, T , k, and off you go!

Lots of ways to actually do this. I would recommend that you directly

simulate multiple simulation replications of the BM W(T ), and then
take the sample average of the e−rT (S(T ) − k)+ values.

But there are other ways: You can just simulate the distribution of
S(T ) directly (it’s lognormal), or you can actually look up the exact
“Black–Scholes” answer (see below).

There are also many, many generalizations of this problem that we

can talk about some other time.

113 / 114
Generating Stochastic Processes

How to Win a Nobel Prize

The Black–Scholes European call option value is

E[C] = e−rT E[S(T ) − k]+
+
σ2
   
−rT
= e E S(0) exp r − T + σW(T ) − k
2
+
Z ∞
σ2 √
  
−rT
= e S(0) exp r − T + σ T z − k φ(z) dz
−∞ 2
√
= S(0)Φ(b + σ T ) − ke−rT Φ(b) (after lots of algebra),
where φ(·) and Φ(·) are the Nor(0,1) p.d.f. and c.d.f., and
σ2 T
rT −
− `n(k/S(0))
b ≡ √ 2
. 2
σ T
Now get your tickets to Norway or Sweden or wherever they give out
the Nobel Prize. . .
114 / 114