SOLUTIONS MANUAL

for
Stochastic Modeling:
Analysis and Simulation

Barry L. Nelson
The Ohio State University

April 2, 2002
Contents

Preface iii

2 Sample Paths 1

3 Basics 5

4 Simulation 25

5 Arrival-Counting Processes 31

6 Discrete-Time Processes 51

7 Continuous-Time Processes 75

8 Queueing Processes 93

9 Topics in Simulation of Stochastic Processes 119

i
ii CONTENTS
Preface

This manual contains solutions to the problems in Stochastic Modeling: Analysis and Simu-
lation that do not require computer simulation. For obvious reasons, simulation results de-
pend on the programming language, the pseudorandom-number generators and the random-
variate-generation routines in use. The manual does include pseudocode for many of the
simulations, however. A companion disk contains SIMAN, SLAM, Fortran and C code for
the simulation “cases” in the text.
I had help preparing this manual. Jaehwan Yang checked many of the solutions. Peg
Steigerwald provided expert LATEX typing of the first draft. And Hasan Jap coded the SLAM,
Fortran and C cases. Please report any errors, typographical or substantial, to me at

voice: (614) 292-0610

fax: (614) 282-7852
e-mail: [email protected]

Barry L. Nelson
September 1994

iii
Chapter 2

Sample Paths

1. The simulation of the self-service system ends at time 129 minutes. Self-service cus-
tomers 7 and 13 experience delays of 1 and 7 minutes, respectively, while full-service
customers 19 and 20 experience delays of 3 and 7 minutes, respectively. All other
customers have delays of 0 minutes.
The simulation of the full-service systems ends at time 124 minutes. Self-service cus-
tomers 7 and 13 experience delays of 1 and 5 minutes, respectively, while full-service
customer 20 experiences a delay of 1 minute. All other customers have delays of 0
minutes.

2. The sample-average gap is 6 minutes. The histogram may resemble an exponential

distribution.

3. If the modification does not change any service times, then the new simulation will be
the same as the simulation of the self-service system with the following exception:
At time 84, self-service customer 13 will switch to the full-service queue, shortening
the customer’s delay to 5 minutes.

Other troublesome assumptions:

Some full-service customers might use self-service if the full-service queue is excessively
long.
We have not accounted for how self-service customers pay for their jobs.

4. See Exercise 3.

1
2 CHAPTER 2. SAMPLE PATHS

5.
Inputs: arrival times service times
Logic: first-come-first-served
system events: arrival finish

part old service times new service times

1 8−5=3 2.7
2 11 − 8 = 3 2.7
3 17 − 11 = 6 5.4
4 20 − 17 = 3 2.7
time system event next next number of
arrival finish parts
0 – 5 – 0
5 arrival 6 5+2.7 = 7.7 1
6 arrival 10 7.7 2
7.7 finish 10 7.7+2.7 = 10.4 1
10 arrival 15 10.4 2
10.4 finish 15 10.4+5.4 = 15.8 1
15 arrival – 15.8 2
15.8 finish – 15.8+2.7 = 18.5 1
18.5 finish – – 0

6.

cust no. arrival finish collate? old copy new copy new
time time finish
0 0.00 0.00 0.0 0.00 0.00 0.00
1 0.09 2.33 0.8 2.24 1.79 1.88
2 0.35 3.09 0.8 0.76 0.61 2.49
3 0.54 4.39 1.1 1.30 1.43 3.92
4 0.92 4.52 0.8 0.13 0.10 4.02
5 2.27 4.62 1.1 0.10 0.11 4.13
6 3.53 4.64 1.1 0.02 0.02 4.16
7 3.66 4.80 0.8 0.16 0.13 4.28
8 4.63 4.86 0.8 0.06 0.05 4.68
9 7.33 8.24 0.8 0.91 0.73 8.06
10 7.61 8.50 1.1 0.26 0.29 8.34
new makespan = 8.34 (or 8.35)
old makespan = 8.50
 3

7. Inputs: Number of hamburgers demanded each day.

Logic: Order only on odd-numbered days. Only order if ≤ 600 patties. Order 1000
minus number in stock.
System event: morning count

day demand
1 (415 + 585) − 704 = 296
2 704 − 214 = 490
3 (214 + 786) − 856 = 144
4 856 − 620 = 236
5 620 − 353 = 267
6 353 − 172 = 181
7 (172 + 828) − 976 = 24
8 976 − 735 = 241
9 735 − 433 = 302
10 433 − 217 = 216
11 (217 + 783) − 860 = 140
12 860 − 598 = 262
13 (598 + 402) − 833 = 167

beginning next size of

day inventory demand order order
1 415 296 2 585
2 704 490 – –
3 214 144 4 786
4 856 236 – –
5 620 267 – –
6 353 181 – –
7 172 24 8 828
8 976 241 – –
9 735 302 – –
10 433 216 – –
11 217 140 12 783
12 860 262 – –
13 598 167 – –
14 431 – – –

No sales were lost. The only change is that no order is placed on day 13.
4 CHAPTER 2. SAMPLE PATHS

8. The available sample path does not permit us to extract the number and timing of
light failures if they remained beyond one year. This is the critical input needed for
the simulation.

9. Inputs: arrival times, processing times

Logic: first-come-first-served
System events: arrival of a job, finishing a job
Here is a portion of a sample path that would result from the sequence of rolls
5, 1, 2, . . .

system number of next next

time event jobs arrival finish
0 – 0 30 ∞
30 arrival 1 60 30+38 = 68
60 arrival 2 90 68
68 finish 1 90 68+20 = 88
88 finish 0 90 ∞
90 arrival 1 120 90+50 = 140

10. Inputs: arrival times, processing times, CPU assignment

Logic: first-come-first-served at each CPU, random distribution to CPUs
System Events: job arrival, job completion at A, job completion at B
Here is a portion of a sample path that would result from the assignment
A, A, B, A, . . .
system number number next next next
time event at A at B arrival A B
0 – 0 0 40 ∞ ∞
40 arrival 1 0 80 40+70 = 110 ∞
80 arrival 2 0 120 110 ∞
110 complete A 1 0 120 110+70 = 180 ∞
120 arrival 1 1 160 180 120+70 = 190
160 arrival 2 1 200 180 190
11. No answer provided.

12. No answer provided.

13. No answer provided.

Chapter 3

Basics

1. (a)
Pr{X = 4} = FX (4) − FX (3)
= 1 − 0.9 = 0.1

(b)
Pr{X = 2} = 1 − Pr{X = 2}
= 1 − 0.3 = 0.7

(c)
Pr{X < 3} = Pr{X ≤ 2}
= FX (2) = 0.4

(d)
Pr{X > 1} = 1 − Pr{X ≤ 1} = 1 − FX (1)
= 1 − 0.1 = 0.9

2. (a) 40 minutes = 40/60 hour = 2/3 hour

Pr{Y > 2/3} = 1 − Pr{Y ≤ 2/3}

= 1 − FY (2/3) = 1 − (1 − e−2(2/3) )
= e−4/3 ≈ 0.264

(b)
Pr{Y > 1} = 1 − Pr{Y ≤ 1}
= 1 − FY (1) = 1 − (1 − e−2(1) )
= e−2 ≈ 0.135

5
6 CHAPTER 3. BASICS

(c) 10 minutes = 1/6 hour

Pr{1/6 < Y ≤ 2/3} = FY (2/3) − FY (1/6)

= (1 − e−4/3 ) − (1 − e−1/3 )
= e−1/3 − e−4/3 ≈ 0.453

(d) 5 minutes = 1/12 hour

Pr{Y ≤ 1/12} = FY (1/12) = 1 − e−2(1/12)

= 1 − e−1/6 ≈ 0.154

3. (a)

d
fX (a) = FX (a)
da


2a/δ 2 , 0 ≤ a ≤ δ
=
0, otherwise

(b) δ
δ
(c) E[X] = 0 a(2a/δ 2 ) da = (2/3)δ

4. (a)
 a
FX (a) = fX (b)db
−∞


0,
 a<0
3 3
= a /β , 0 ≤ a ≤ β


1, β<a

(b) 0
β
(c) E[X] = 0 a(3a2 /β 3 ) da = (3/4) β

5. (a)
 2
E[Y ] = a(3/16 a2 + 1/4)da
0

= 5/4
 7

(b)
 a
FY (a) = fY (b)db
−∞


 0, a<0
= a3 /16 + a/4, 0 ≤ a ≤ 2


1, 2<a
(c) 1 12 , because the density function is larger in a neighborhood of 1 12 than in a
neighborhood of 12 .
6. (a)
Pr{X = 4, X = 1}
Pr{X = 4 | X = 1} =
Pr{X = 1}
Pr{X = 4} 1 − FX (3)
= =
Pr{X = 1} 1 − FX (1)
1 − 0.9
= = 1/9
1 − 0.1
(b)
Pr{X = 4}
Pr{X = 4 | X = 1, X = 2} =
Pr{X = 1, X = 2}
1 − FX (3) 1 − 0.9
= = = 1/6
1 − FX (2) 1 − 0.4

(c)
Pr{X = 2, X ≤ 2}
Pr{X = 2 | X ≤ 2} =
Pr{X ≤ 2}
Pr{X = 2} FX (2) − FX (1)
= =
Pr{X ≤ 2} FX (2)
0.4 − 0.1
= = 3/4
0.4
7. (a)
Pr{Y > 2/3, Y > 1/2}
Pr{Y > 2/3 | Y > 1/2} =
Pr{Y > 1/2}
Pr{Y > 2/3} 1 − FY (2/3) e−2(2/3)
= = = −2(1/2)
Pr{Y > 1/2} 1 − FY (1/2) e
= e−2(1/6) ≈ 0.72
8 CHAPTER 3. BASICS

e−2(1)
(b) Pr{Y > 1 | Y > 2/3} = e−2(2/3)
= e−2(1/3) ≈ 0.51
(c)

Pr{Y < 1/6}

Pr{Y < 1/6 | Y < 2/3} =
Pr{Y < 2/3}
FY (1/6) 1 − e−2(1/6)
= = ≈ 0.38
FY (2/3) 1 − e−2(2/3)

8. (a)

Pr{W = 3, V = 2}
Pr{W = 3 | V = 2} =
Pr{V = 2}
pV W (2, 3) 3/20
= = = 3/12 = 1/4
pV (2) 12/20

(b)

Pr{W ≤ 2, V = 1}
Pr{W ≤ 2 | V = 1} =
Pr{V = 1}
Pr{W = 1, V = 1} + Pr{W = 2, V = 1}
=
pV (1)
2/10 + 1/10 3/10
= = = 3/4
4/10 4/10

(c)

Pr{W = 2} = 1 − Pr{W = 2}
= 1 − pW (2) = 1 − 10/20 = 10/20 = 1/2

9. Joint distribution

X2
0 1
X1 0 0.75 0.05 0.80
1 0.10 0.10 0.20
0.85 0.15
 9

(a)

Pr{X2 = 1, X1 = 0}
Pr{X2 = 1 | X1 = 0} =
Pr{X1 = 0}
0.05
= = 0.0625
0.80
0.10
Pr{X2 = 1 | X1 = 1} = = 0.50
0.20
Clearly X1 and X2 are dependent since Pr{X2 = 1 | X1 = 0} = Pr{X2 = 1 |
X1 = 1}.


100, a = 0
(b) Let g(a) =
−20, a = 1

E[g(X2 ) | X1 = 1] = g(0) Pr{X2 = 0 | X1 = 1}

+ g(1) Pr{X2 = 1 | X1 = 1}
= 100(1 − 0.50) − 20(0.50)
= 40

 = √
10. µ = 2.75 σ = 0.064 se σ
≈ 0.026
6
µ + se
 = 2.776

µ − se
 = 2.724

Assuming a mean of 2.776

Pr{X > 2.90} = 1 − Pr{X ≤ 2.90}

X − 2.776 2.90 − 2.776

= 1 − Pr ≤
0.064 0.064
= 1 − Pr{Z ≤ 1.938}
≈ 1 − 0.974 = 0.026

where Z is a standard-normal random variable.

Assuming a mean of 2.724

2.90 − 2.724
Pr{X > 2.90} = 1 − Pr Z ≤
0.064
= 1 − Pr{Z ≤ 2.75} ≈ 1 − 0.997 = 0.003
10 CHAPTER 3. BASICS

11. (a) µ = 8.08 σ ≈ 3.60

 = √
(b) se σ
≈ 1.61
5

(c)


 0, a < 3.5



 1/5, 3.5 ≤ a < 5.9


 2/5, 5.9 ≤ a < 7.7
F (a) = 

 3/5, 7.7 ≤ a < 11.1





4/5, 11.1 ≤ a < 12.2

1, 12.2 ≤ a

12. (a) µ ≈ 10.98 σ ≈ 4.81

 ≈ 1.08
(b) se
20
(c) Use the definition F (a) = 1
20 i=1 I(di ≤ a)
(d) try normal

13. (a) µ = 0.80 σ ≈ 0.89

 ≈ 0.16
(b) se
(c) 

 0, a<0



 14/30, 0≤a<1

F (a) = 23/30, 1≤a<2



 29/30, 2≤a<3


 1, 3≤a

(d) 

 14/30, a=0



 9/30, a=1


p(a) = 6/30, a=2



 1/30, a=3

 0, otherwise

14. (a) µ ≈ 6.05 σ ≈ 0.23

 ≈ 0.05
(b) se
25
(c) Use the definition F (a) = 1
25 i=1 I(di ≤ a)
(d) try normal

15. Let X be a random variable taking values 0 (to represent a “tail”) and 1 (to represent
a “head”). A plausible model is that X has a Bernoulli distribution with parameter
γ = Pr{X = 1}. Your date will likely show γ > 1/2.
 11

16. Let U be a random variable having the uniform distribution of [0, 1].
Let V = 1 − U . Then

Pr{V ≤ a} = Pr{1 − U ≤ a}
= Pr{U ≥ 1 − a} = a, 0 ≤ a ≤ 1.

Therefore, V and U are both uniformly distributed on [0, 1].

Therefore, Y = − ln(1 − U )/λ = − ln(V )/λ and Y = − ln(U )/λ must have the same
distribution.
a−α
17. (a) U = F (Y ) = β−α
Therefore, Y = (β − α)U + α
algorithm uniform
1. U ← random()
2. Y ← (β − α)U + α
3. return Y
For α = 0, β = 4
U Y
0.1 0.4
0.5 2.0
0.9 3.6
α
(b) U = F (Y ) = 1 − e−(Y /β)
Therefore,
α
1 − U = e−(Y /β)
ln(1 − U ) = −(Y /β)α
Y = β(− ln(1 − U ))1/α

algorithm Weibull
1. U ← random()
2. Y ← β(− ln(1 − U ))1/α
3. return Y
For α = 1/2, β = 1
U Y
0.1 0.011
0.5 0.480
0.9 5.302
12 CHAPTER 3. BASICS

For α = 2, β = 1
U Y
0.1 0.325
0.5 0.833
0.9 1.517
(c) 



1, 0 ≤ U ≤ 0.3



 2, 0.3 < U ≤ 0.4
Y = 3, 0.4 < U ≤ 0.7



 4, 0.7 < U ≤ 0.95


 6, 0.95 < U ≤ 1
Using algorithm discrete inverse cdf with a1 = 1, a2 = 2, a3 = 3, a4 = 4, a5 = 6.
U Y
0.1 1
0.5 3
0.9 4

(d)

0, 0 ≤ U ≤ 1 − γ
Y =
1, 1 − γ < U ≤ 1
algorithm Bernoulli
1. U ← random()
2. if {U ≤ 1 − α} then
Y ←0
else
Y ←1
endif
3. return Y
For γ = 1/4
U Y
0.1 0
0.5 0
0.9 1
(e)

p(a) = F (a) − F (a − 1)
= 1 − (1 − γ)a − (1 − (1 − γ)a−1 )
= (1 − γ)a−1 γ , a = 1, 2, 3 . . .
 13


 1, 0 ≤ U ≤ γ




 2, γ < U ≤ 1 − (1 − γ)2
 .. ..
Y = . .



 a, 1 − (1 − γ)a−1 < U ≤ 1 − (1 − γ)a



 .. ..
. .

algorithm geometric
1. U ← random()
a←1
2. until U ≤ 1 − (1 − γ)a
do
a←a+1
endo
3. Y ← a
4. return Y
For γ = 1/4
U Y
0.1 1
0.5 2
0.9 9
18. (a)


 0, a<α
d 1
f (a) = F (a) =  β−α
, α≤a≤β
da 
0, β<a
α
(b) f (a) = d
da
F (a) = 0 − e−(a/β) (−α(a/β)α−1 (1/β))

α
α β −α aα−1 e−(a/β) , 0 < a
=
0, a≤0

(c) Using the relationship p(a) = F (a) − F (a − 1)



 0.3, a=1



 0.1, a=2

p(a) = 0.3, a=3



 0.25, a=4


 0.05, a=6
14 CHAPTER 3. BASICS

(d)


1 − γ, a = 0
p(a) =
γ, a=1

(e) From the solution of Exercise 17(e)

γ(1 − γ)a−1 , for a = 1, 2, . . .
p(a) =
0, otherwise

19. (a)
 ∞  β
a
E[X] = af (a)da = da
−∞ α β−α
β
a2  β 2 − α2 (β + α)(β − α)

=  = =

2(β − α) α 2(β − α) 2(β − α)
β+α
=
2
(b)
 ∞  ∞
α
E[X] = af (a)da = a α β −α aα−1 e−(a/β) da
−∞ 0
 ∞
α
= a(α/β)(a/β)α−1 e−(a/β) da = (1)
0

Let u = (a/β)α so that du/da = (α/β)(a/β)α−1 and a = βu1/α . Then

 ∞
du −u
(1) = βu1/α e da
0 da
 ∞
= β u1/α e−u du
0
 ∞
= β u(1/α+1)−1 e−u du
0

= βΓ(1/α + 1) = (β/α)Γ(1/α)

(c)

E[X] = a pX (a)
all a

= 1(0.3) + 2(0.1) + 3(0.3) + 4(0.25) + 6(0.05)

= 2.7
 15

(d)

E[X] = a pX (a) = 0(1 − γ) + 1(γ)
all a

= γ

(e)

 ∞

E[X] = a pX (a) = aγ(1 − γ)a−1
all a a=1
∞

= γ a(1 − γ)a−1 (let q = 1 − γ)
a=1
∞
 d a
= γ q
a=1 dq
∞ ∞
d  d 
= γ qa =γ qa
dq a=1 dq a=0
d 0
(since q = 0)
dq

d 1 γ γ 1
= γ = = =
dq 1−q (1 − q) 2 (γ) 2 γ

20. From Exercise 25, Var[X] = E[X 2 ]−(E[X])2 . So if we calculate E[X 2 ], we can combine
it with the answers in Exercise 19.

(a)
 β β
a2 a3 
2 
E[X ] = da = 
α β−α 3(β − α) α
β 3 − α3
=
3(β − α)

Therefore,
 2
β 3 − α3 β+α
Var[X] = −
3(β − α) 2
β 3 − α3 (β + α)2
= −
3(β − α) 4
16 CHAPTER 3. BASICS

4(β 3 − α3 ) 3(β + α)2 (β − α)

= −
12(β − α) 12(β − α)
4(β − α)(β 2 + βα + α2 ) − 3(β 2 + 2βα + α2 )(β − α)
=
12(β − α)
β 2 − 2βα + α2 (β − α)2
= =
12 12
(b)
 ∞
α
E[X 2 ] = a2 αβ −α aα−1 e−(a/β) da
0
 ∞
α
= a2 (α/β)(a/β)α−1 e−(a/β) da
0

Using the same substitution as 19(b) gives

 ∞  ∞
= β 2 u2/α e−u du = β 2 u2/α e−u du
0 0

2β 2
= β 2 Γ(2/α + 1) = Γ(2/α)
α
Therefore,
2β 2
Var[X] = Γ(2/α) − (β/α Γ(1/α))2
α
β2
= (2Γ(2/α) − 1/α (Γ(1/α))2 )
α
(c)

E[X 2 ] = 12 (0.3) + 22 (0.1) + 32 (0.3) + 42 (0.25) + 62 (0.05)

= 9.2

Therefore, Var[X] = 9.2 − (2.7)2 = 1.91

(d) E[X 2 ] = 02 (1 − γ) + 12 (γ) = γ
Therefore, Var[X] = γ − γ 2 = γ(1 − γ)
(e)
∞

E[X 2 ] = a2 γ(1 − γ)a−1
a=1
∞

= γ a2 q a−1 (let q = 1 − γ)
a=1
 17

∞

 d2 a+1
= γ q − aq a−1
a=1 dq 2
∞
d2  1
= γ 2
q a+1 − (from Exercise 19(e))
dq a=1 γ
∞

d2  1 d2 d2
= γ 2 q a+1 − since 2 q 0 = 2 q = 0
dq a=−1 γ dq dq
∞

d2  1 d2 1 1
= γ 2 q − =γ 2
b
−
dq b=0 γ dq 1−q γ
2 1 2 1
= γ − = 2−
(1 − q) 3 γ γ γ

Therefore,
2 1 1 1 1
Var[X] = 2
− − 2 = 2−
γ γ γ γ γ
1−γ
=
γ2

21. No answer provided.

22. You need the solutions from Exercise 17. Using them provides the following inverse
cdf functions
(a) B1 = -LN (1-A1)* 2
(b) B1 = 4* A1
(c) B1 = 2.257* (-LN (1-A1))** (1/2)
All of these distributions have expected value 2, but are otherwise very different.

23. Use exponential distributions for the time data, and parameterize them by using the
fact that 1/λ is the expected value.
interarrival-time gaps
 = 6 minutes (sample average)
1/λ
 = 1/6 customer/minute
Therefore, λ

self-service customers
 = 3 minutes (sample average)
1/λ
 = 1/3 customer/minute
Therefore, λ
18 CHAPTER 3. BASICS

full-service customers
 = 7 minutes (sample average)
1/λ
 = 1/7 customer/minute
Therefore, λ

all customers
 = 4.6 minutes (sample average)
1/λ
 = 1/4.6 customer/minute
Therefore, λ

customer type
Use a Bernoulli distribution with γ = Pr{full-service customer}
γ = 8/20 = 0.4
All of the models except one are very good fits, since in fact the interarrival time gaps,
customer types and conditonal service times were generated from these distributions!
The model for all customers’ service times will not be as good since it is actually a
mixture of two exponential distributions.

24. Suppose X has mass function pX


E[aX + b] = (ac + b)pX (c)
all c
 
= a pX (c) + b pX (c)
all c all c

= aE[X] + b

Suppose X has density function fX

 ∞
E[aX + b] = (ac + b)fX (c)dc
−∞
 ∞  ∞
= a cfX (c)dc + b fX (c)dc
−∞ −∞

= aE[X] + b

25. Suppose X has a density fX

 ∞
Var[X] = (a − E[X])2 fX (a)da
−∞
 19
 ∞  ∞
= a2 fX (a)da − 2E[X] afX (a)da
−∞ −∞
 ∞
+ (E[X])2 fX (a)da
−∞

= E[X 2 ] − 2E[X]E[X] + (E[X])2

= E[X 2 ] − (E[X])2

The proof for discrete-valued X is analogous.

26.
∞

E[X] = a Pr{X = a}
a=0

∞

a 
 
= Pr{X = a}
a=1 i=1
 
∞ 
 ∞ ∞  
 ∞ 
= Pr{X = a} = 
Pr{X = a}
i=1 a=i i=0 a=i+1

∞

= Pr{X > i}
i=0

27. E[X m ] = 0m (1 − γ) + 1m γ = γ
28. Notice that
E[I(Xi ≤ a)] = 0 Pr{Xi > a} + 1 Pr{Xi ≤ a}
= FX (a)

Therefore,
1 m
E[FX (a)] = E[I(Xi ≤ a)]
m i=1
1
= m FX (a) = FX (a)
m
29.
 ∞
E[g(Y )] = g(a)fY (a)da
−∞

= 1 fY (a)da + 0
A

= Pr{Y ∈ A}
20 CHAPTER 3. BASICS

30. (a) FX−1 (q) = − ln(1 − q)/λ for the exponential. To obtain the median we set q = 0.5.
Then η = − ln(1 − 0.5)/λ ≈ 0.69/λ. This compares to E[X] = 1/λ.
Notice that

FX (1/λ) = 1 − e−λ(1/λ)
= 1 − e−1 ≈ 0.63

showing that the expected value is approximately the 0.63 quantile for all values
of λ.
(b) FX−1 (q) = α + (β − α)q for the uniform distribution. For q = 0.5, η = α + (β −
α)0.5 = (β + α)/2 which is identical to the E[X].
31. We want Pr{Y = b} = 1/5 for b = 1, 2, . . . , 5.

Pr{Y = b} = Pr{X = b | X = 6}
Pr{X = b, X = 6}
= (definition)
Pr{X = 6}
Pr{X = b} 1/6
= = = 1/5
Pr{X = 6} 5/6

32. (a)

 
 pY (Z)

Pr{Y = a} = Pr Z = a V ≤
cpZ (Z)
 
pY (Z)
Pr Z = a, V ≤ cpZ (Z) (1)
=   =
pY (Z)
Pr V ≤ cpZ (Z)
(2)


 pY (a)
(2) = Pr V ≤ | Z = a pZ (a)
a∈B cpZ (a)

 pY (a) 1 1
= pZ (a) = pY (a) =
a∈B cpZ (a) c a∈B c


pY (a)
(1) = Pr V ≤ | Z = a pZ (a)
cpZ (a)
pY (a) pY (a)
= pZ (a) =
cpZ (a) c
(1)
Therefore, (2)
= pY (a).
 21

(b) Let T ≡ number of trials until acceptance. On each trial


pY (Z) 1
Pr{accept} = Pr V ≤ =
cpZ (Z) c

Since each trial is independent, T has a geometric distribution with γ = 1/c.

1
Therefore, E[T ] = 1/c = c.
(c) algorithm 17(c)
1. roll a die to obtain Z
2. V ← random()
3. if V ≤ cp(Z)
1/6
then
return Y ← Z
else
goto step 1
endif
0.3
Here c = 1/6 = 9/5 because the largest value of pY is 0.3. Therefore, the expected
number of trials is 9/5 = 1 45 (almost 2).

33. (a)

 
 fY (Z)
Pr{Y ≤ a} = Pr Z ≤ a V ≤
 cfZ (Z)
 
fY (Z)
Pr Z ≤ a, V ≤ cfZ (Z) (1)
=   =
fY (Z)
Pr V ≤ cfZ (Z)
(2)
 ∞

fY (Z)
(2) = Pr V ≤ | Z = a fZ (a)da
−∞ cfZ (Z)
 ∞

fY (a)
= Pr V ≤ fZ (a)da
−∞ cfZ (a)
 ∞
fY (a) 1 ∞ 1
= fZ (a)da = fY (a)da =
−∞ cfZ (a) c −∞ c
 ∞

fY (Z)
(1) = Pr Z ≤ a, V ≤ | Z = b fZ (b)db
−∞ cfZ (Z)
 a
  ∞
fY (b)
= Pr V ≤ fZ (b)db + 0 fZ (b)db
−∞ cfZ (b) a

 a
fY (b) 1 a 1
= fZ (b)db = fY (b)db = FY (a)
−∞ cfZ (b) c −∞ c
22 CHAPTER 3. BASICS

(1)
Therefore, (2)
= FY (a).
(b) Let T ≡ number of trials until acceptance. On each trial


fY (Z) 1
Pr{accept} = Pr V ≤ =
cfZ (Z) c
Since each trial is independent, T has a geometric distribution with γ = 1/c.
1
Therefore, E[T ] = 1/c = c.
(c) Note that fY (a) is maximized at fY (1/2) = 1 12 .
Let fZ (a) = 1, 0 ≤ a ≤ 1
c = 1 12
Therefore, cfZ (a) = 1 12 ≥ fY (a)

1.U ← random()
2.Z←U
3.V ← random()
4.if V ≤ 6Z(1 − Z)/(1 12 ) then
return Y ← Z
else
goto step 1
endif
E[T ] = 1 12
34. (a)
 a
2(b − α)
F (a) = db
α (β − α)2

(a − α)2
= , α≤a≤β
(β − α)2
(X − α)2
U =
(β − α)2
(X − α)2 = U (β − α)2
 √
X = α+ U (β − α)2 = α + (β − α) U

algorithm
1. U ← random()
√
2. X ← α + (β − α) U
3. return X
 23

2 1
(b) f is maximized at a = β giving f (β) = β−α
. Let fZ (a) = β−α
for α ≤ a ≤ β. Set
c = 2 so that cfZ (a) ≥ f (a).
Notice that
2(a−α)
f (a) (β−α)2
= 2
cfZ (a) β−α

a−α
=
β−α

algorithm
1. U ← random()
Z ← α + (β − α)U
2. V ← random()
3. if V ≤ Z−α
β−α
then
return Y ← Z
else
goto step 1
endif
24 CHAPTER 3. BASICS
Chapter 4

Simulation

1. An estimate of the probability of system failure within 5 days should be about 0.082.

2. No answer provided.

3. Let G1 , G2 , . . . represent the interarrival-time gaps between the arrival of jobs. Let
Bn represent the type of the nth job; 0 for a job without collating, 1 for a job with
collating.
Let Xn represent the time to complete the nth job that does not require collating.
Let Zn represent the time to complete the nth job that does require collating.
We model all of these random variables as mutually independent.
Let Sn represent the number of jobs in progress or waiting to start.
Define the following system events with associated clocks:
e0 () (initialization)
S0 ← 0
C1 ← FG−1 (random())
C2 ← ∞

e1 () (arrival of a job)
Sn+1 ← Sn + 1
if {Sn+1 = 1} then
Bn ← FB−1 (random())
if {Bn = 0} then
C2 ← Tn+1 + FX−1 (random())
else

25
26 CHAPTER 4. SIMULATION

C2 ← Tn+1 + FZ−1 (random())

endif
endif
C1 ← Tn+1 + FG−1 (random())

e2 () (finish job)
Sn+1 ← Sn − 1
if {Sn+1 > 0} then
Bn ← FB−1 (random())
if {Bn = 0} then
C2 ← Tn+1 + FX−1 (random())
else
C2 ← Tn+1 + FZ−1 (random())
endif
endif

4. Let Dn represent the number of hamburgers demanded on the nth day. We model
D1 , D2 , . . . as independent.
Let Sn represent the number of patties in stock at the beginning of the nth day.
Define the following system events with associated clocks:
e0 () (initialization)
S0 ← 1000
C1 ← 1

e1 () (morning count)
if n is odd then
if {Sn ≤ 500} then
Sn+1 ← Sn − FD−1 (random()) +(1000 − Sn )
else
Sn+1 ← Sn − FD−1 (random())
endif
else
 27

Sn+1 ← Sn − FD−1 (random())

endif

5. Let Xn represent the time to complete the nth job.

Let Sn represent the number of jobs in process or waiting in the buffer.
Define the following system events and associated clocks:

e0 () (initialization)
C1 ← 30
C2 ← ∞
S0 ← 0

e1 () (arrival of a job)
Sn+1 ← Sn + 1
if {Sn+1 = 1} then
C2 ← Tn+1 + FX−1 (random())
endif
C1 ← Tn+1 + 30

e2 () (finish a job)
Sn+1 ← Sn − 1
if {Sn+1 > 0} then
C2 ← Tn+1 + FX−1 (random())
endif

6. Let Bn be a Bernoulli random variable representing the CPU assignment of the nth
job.
Let S0,n represent the number of jobs at CPU A, and S1,n similar for CPU B.
Define the following system events and associated clocks:
e0 () (initialization)
S0,n ← 0
S1,n ← 0
C1 ← 40
C2 ← ∞
C3 ← ∞
28 CHAPTER 4. SIMULATION

e1 () (arrival of a job)

B ← FB−1 (random())

SB,n+1 ← SB,n + 1

if {SB,n+1 = 1} then

C2+B ← Tn+1 + 70

endif

C1 ← Tn+1 + 40

e2 () (finish at CPU A)

S0,n+1 ← S0,n − 1

if {S0,n+1 > 0} then

C2 ← Tn+1 + 70

endif

e3 () (finish at CPU B)

S1,n+1 ← S1,n − 1

if {S1,n+1 > 0} then

C3 ← Tn+1 + 70

endif

7. No answer provided.

8. (a)
 29

(b)

Ȳ1 = {0(5 − 0) + 1(6 − 5) + 2(9 − 6) + 1(11 − 9)

+ 2(13 − 11) + 3(15 − 13) + 2(16 − 15)
+1(18 − 16) + 0(19 − 18)} ÷ 19
23
= ≈ 1.2 customers waiting for service
19
0(3 − 0) + 1(19 − 3) 16
Ȳ2 = = ≈ 0.84
19 19
utilization of the copier

(c) N19 = 5 arrivals

An estimate of average delay is
 19
0 Y1,t dt 23
= = 4.6 minutes
N19 5
9. Based on the available data, the sample-average interarrival-time gap was about 17.6
seconds, and might be modeled as exponentially distributed.
Let FG be the cdf of the interarrival-time gaps, which we assume to be independent
because students act independently.
Let ∆t be the pick-up interval, in seconds.
Let Sn represent the number of students waiting after the nth state change.
Define the following system events and associated clocks:
e0 () (start of day)
S0 ← 0
C1 ← FG−1 (random())
C2 ← ∆t

e1 () (student arrives)
Sn+1 ← Sn + 1
C1 ← Tn+1 + FG−1 (random())

e2 () (pick up)
30 CHAPTER 4. SIMULATION

Sn+1 ← Sn − min{Sn , 60}

C2 ← C2 + ∆t
record Sn+1
The recorded values of Sn+1 are the number left waiting after each bus pick up. They
can be analyzed to examine the effect of ∆t = 600, 720, 900.
Chapter 5

Arrival-Counting Processes

1. (a)
e−2(2) (2(2))5
Pr{Y2 = 5} = ≈ 0.16
5!
(b)
e−2 2
Pr{Y4 − Y3 = 1} = Pr{Y1 = 1} = ≈ 0.271
1!
(c)
Pr{Y6 − Y3 = 4 | Y3 = 2} = Pr{Y3 = 4}

e−2(3) (2(3))4
= ≈ 0.134
4!
(d)
Pr{Y5 = 4 | Y4 = 2} = Pr{Y5 − Y4 = 2 | Y4 = 2}

e−2 22
= Pr{Y1 = 2} = ≈ 0.271
2!

2. No answer provided.

3. λ = 2/hour. Let t be measured in hours from 6 a.m.

(a)

Pr{Y4 = 9 | Y2 = 6} = Pr{Y4 − Y2 = 3 | Y2 = 6}
= Pr{Y2 = 3}
e−2(2) (2(2))3
= ≈ 0.195
3!

31
32 CHAPTER 5. ARRIVAL-COUNTING PROCESSES

(b) The expected time between arrivals is

1/λ = 1/2 hour

Let G be the gap between two successive arrivals. Then G is exponentially dis-
tributed with λ = 2.

Pr{G > 1} = 1 − Pr{G ≤ 1}

= 1 − (1 − e−2(1) )
= e−2
≈ 0.135

(c) Because of the memoryless property

E[R0 ] = E[G] = 1/2 hour

Pr{R0 ≤ 1/4} = Pr{G ≤ 1/4} = 1 − e−2(1/4)

≈ 0.393
(d)
12 −2(7)
 e (2(7))j
Pr{T13 ≤ 7} = 1 −
j=0 j!

≈ 0.641
(e) Let λ0 be the arrival rate for urgent patients.

λ0 = 0.14λ = 0.28 patients/hour

Pr{Y0,12 > 6} = 1 − Pr{Y0,12 ≤ 6}
6
 e−0.28(12) (0.28(12))j
= 1−
j=0 j!

≈ 0.055

(f) Let λ2 be the overall arrival rate.

λ2 = λ + 4 = 6/hour
 33

Pr{Y2,6 > 30} = 1 − Pr{Y2,6 ≤ 30}

30 −6(6)
 e (6(6))j
= 1− ≈ 0.819
j=0 j!

4. We want the smallest m such that

Pr{Y30 − Y6 > m} ≤ 0.05


m
e−24 24j
Pr{Y30 − Y6 > m} = 1 −
j=0 j!

m = 32 does it.

5. If we assume that the accident rate is still λ = 1/week, then

Pr{Yt+24 − Yt ≤ 20} = Pr{Y24 ≤ 20}

20 −24
 e 24j
= ≈ 0.24
j=0 j!

Therefore, there is nearly a 1 in 4 chance of seeing 20 or fewer accidents even if the

rate is still 1/week. This is not overwhelming evidence in favor of a lower rate.

6. t corresponds to square meters of metal and λ = 1/50 defect/meter2

(a)

Pr{Y200 ≥ 7} = 1 − Pr{Y200 ≤ 6}
1
6
 e− 55 (200) ( 200 )j 50
= 1−
j=0 j!

≈ 0.111

(b) For an out-of-control process λ ≥ 4/50. For λ = 4/50 we want

Pr{Y200 > c} ≥ 0.95

34 CHAPTER 5. ARRIVAL-COUNTING PROCESSES

Pr{Y200 > c} = 1 − Pr{Y200 ≤ c}

4

c
e− 50 (200) ( 4(200) )j
50
= 1−
j=0 j!


c
e−16 (16)j
= 1− ≥ 0.95
j=0 j!

Using trial and error, or the approximation from Exercise 2, c = 9.

Pr{Y200 > 9} ≈ 0.957

Pr{Y200 > 10} ≈ 0.923

We will declare an “in-control” process “out-of-control” if λ = 1/50 and {Y200 >

9}.
1
9
 e− 50 (200) ( 200 )j50
Pr{Y200 > 9} = 1 − ≈ 0.008
j=0 j!

7. Suppose we model the arrival of students to the bus stop as a Poisson arrival process
with expected time between arrivals of
 = 17.6 seconds
1/λ
 = 1/17.6 students/second ≈ 3.4 students/minute.
giving λ
Now we can compute the probability of more than 60 students arriving during each
proposed time interval

Pr{Y15 > 60} = 1 − Pr{Y15 ≤ 60}

60 −3.4(15)
 e (3.4(15))j
= 1−
j=0 j!

≈ 0.094
Pr{Y12 > 60} ≈ 0.002
Pr{Y10 > 60} ≈ 0.000

Ignoring students left behind on one pick up that add to the next pick up, we see that
there is nearly a 1 in 10 chance of filling the bus when pick up is every 15 minutes.
The carryover will only make the problem worse. Pick up every 12 minutes effectively
eliminates the problem. Every 10 minutes is more often than needed.
 35

8. (a) Restricted to periods of the day when the arrival rate is roughly constant, a
Poisson process is appropriate to represent a large number of customers acting
independently.
(b) Not a good approximation, since most arrivals occur during a brief period just
prior to the start, and only a few before or after this period. Therefore, arrivals
do not act independently.
(c) Not a good approximation if patients are scheduled. We do not have independent
increments because patients are anticipated. (May be a good approximation for
a walk-in clinic, however.)
(d) Not a good approximation because the rate of finding bugs will decrease over
time.
(e) Probably a good approximation since fires happen (largely) independently, and
there are a large number of potential customers (buildings).

9. (a)

c = 60 + 72 + 68 = 200 total arrivals

d = 3 + 3 + 3 = 9 total hours observed
 = c/d ≈ 22 customers/hour
λ


λ
 =
se ≈ 1.6 customers/hour
d

(b) Let Yt ≡ number of arrivals by time t, and suppose λ = 22/hour.

56 −22(3)
 e (22(3))j
Pr{Y3 ≤ 56} = ≈ 0.12
j=0 j!

Since this probability is rather small, we might conclude that Fridays are different
(have a lower arrival rate).
At λ = 22 + se = 23.6

Pr{Y3 ≤ 56} ≈ 0.04

and at λ = 22 − se
 = 20.4

Pr{Y3 ≤ 56} ≈ 0.29

So less than or equal to 56 could be quite rare if λ = 23.6. This is further evidence
that Fridays could be different.
36 CHAPTER 5. ARRIVAL-COUNTING PROCESSES

10. Let

λ0 = 8/hour
λ1 = 1/18/hour p1 = 0.005
λ2 = 1/46/hour p2 = 0.08
32
(a) λ = λ0 + λ1 + λ2 = 8 414 surges/hour
By the superposition property, the arrival of all surges is also a Poisson process.
Therefore, E[Y8 ] = λ8 ≈ 64.6 surges.
(b) By the decomposition property, the “small” and “moderate” surges can be de-
composed into Poisson processes of computer-damaging surges.

λ10 = p1 λ1 = 1/3600/hour
λ20 = p2 λ2 = 1/575/hour

These processes can be superposed to give λ3 = λ10 + λ20 ≈ 0.0020. Therefore,

E[Y8 ] = 8λ3 ≈ 0.016 computer-damaging surge.
(c) Using the arrival rate from part (b)

Pr{Y8 = 0} = e−λ3 8 ≈ 0.98

11. No answer provided.

(0)
12. Let {Yt ; t ≥ 0} be a Poisson arrival process representing the arrival of requests that
require printing; λ(0) = 400/hour.
(1)
{Yt ; t ≥ 0} similarly represents the arrival of requests that do not require printing;
λ(1) = 1000/hour.
(0) (1)
(a) Yt = Yt + Yt is Poisson with rate λ = λ(0) + λ(1) = 1400/hour.
∞
 e−1400(1.5) (1400(1.5))m
Pr{Y1.5 > 2000} = ≈ 0.985
m=2001 m!
(A) (B)
(b) Let Yt and Yt represent the arrivals to computers A and B, respectively.

(A)
Yt is Poisson with λ(A) = (1/2)λ = 700

(B)
Yt is Poisson with λ(B) = (1/2)λ = 700
and they are independent, therefore,
 37

(A) (B)
Pr{Y1.5 > 1000, Y1.5 > 1000}
(A) (B)
= Pr{Y1.5 > 1000} Pr{Y1.5 > 1000}

∞
2
 e−700(1.5) (700(1.5))m
= ≈ 0.877
m=1001 m!

13. Let {Yt ; t ≥ 0} model the arrival of autos with λ = 1/minute. Then {Y0,t ; t ≥ 0} which
models the arrival of trucks is Poisson with λ0 = (0.05)λ = 0.05/min, and {Y1,t ; t ≥ 0}
which models all others is Poisson with λ1 = (0.95)λ = 0.95/min.

(a)
∞
 e−(0.05)(60) [(0.05)(60)]m
Pr{Y0,60 ≥ 1} =
m=1 m!

e−3 (3)0
= 1− = 1 − e−3 ≈ 0.95
0!
(b) Since Y0,t is independent of Y1,t what happened to Y0,t is irrelevant. Therefore,
10 + E[Y1,60 ] = 10 + 60(.95) = 67.
(c)

Pr{Y0,60 = 5 | Y60 = 50}

Pr{Y0,60 = 5, Y60 = 50}
=
Pr{Y60 = 50}
1
= Pr{Y0,60 = 5, Y1,60 = 45} e−60 (60)50
50!

50!
= Pr{Y0,60 = 5} Pr{Y1,60 = 45}
e−60 (60)50
e−3 (3)5 e−57 (57)45 50!
=
5! 45! e (60)50
−60

 5  45
50! 3 57
=
5!45! 60 60
≈ 0.07

14. The rate of errors after the nth proofreading is

λ 1
λn = = error/1000 words
2n 2n
38 CHAPTER 5. ARRIVAL-COUNTING PROCESSES

We want Pr{Y200 = 0} ≥ 0.98

Pr{Y200 = 0} = e−λn (200)

200
= e− 2n ≥ 0.98
200
⇒− ≥ ln(0.98)
2n
200
2n ≥ − ≈ 9900
ln(0.98)
⇒ n = 14 times
since 213 = 8192
214 = 16, 384

15. (a) The rate at which sales are made does not depend on the time of year; there is
no seasonal demand. Also, the market for A, B and C is neither increasing nor
decreasing.
(b) Let Yt ≡ total sales after t weeks.
By the superposition property Yt is a Poisson process with rate λ = 10 + 10 =
20/week.

∞
e−20(1) (20(1))n

Pr{Y1 > 30} =
n=31 n!
30
 e−20 20n
= 1− ≈ 0.013
n=0 n!

(c) Let P ≡ person hours/order.

Then

E[P ] = 25(0.2) + 15(0.7) + 40(0.1)

= 19.5 person-hours

Since E[Y4 ] = 20(4) = 80 orders we expect 80(19.5) = 1560 person-hours per

month.
(B)
(d) Let Yt ≡ Louise’s sales of B after t weeks.
(B)
The decomposition property implies that Yt is Poisson with rate λ(B) = (0.7)10 =
7 per week.
 39

(B) (B) (B)

Pr{Y2 − Y1 > 5, Y1 > 5}
(B) (B) (B)
= Pr{Y2 − Y1 > 5} Pr{Y1 > 5}
(B) (B)
= Pr{Y1 > 5} Pr{Y1 > 5}
 5
2
 e−7 7n
= 1− ≈ 0.489
n=0 n!

16. (a) 10 seasons = 30 months

 =8
λ ≈ 0.27 hurricanes/month
30


λ
 =
se ≈ 0.09 hurricanes/month
30

(b) Pr{Y3 = 0} = e−λ(3) ≈ 0.45

17. (a) The rate at which calls are placed does not vary from 7 a.m.–6 p.m.
(D)
(b) Let Yt ≡ number of long-distance calls placed by hour t, where t = 0 corre-
(D)
sponds to 7 a.m. Then Yt is Poisson with rate λ(D) = (1000)(0.13) = 130/hour.

(D) (D) (D)

Pr{Y8 − Y7 > 412} = Pr{Y1 > 412}

∞
 e−130 (130)n
=
n=413 n!
412 −130
 e 130n
= 1−
n=0 n!
≈0

(Notice that t = 8 corresponds to 3 p.m.)

(c) Expected number of local calls is
(L)
E[Y11 ] = λ(L) (11) = (870)(11) = 9570 calls

Expected number of long-distance calls is

(D)
E[Y11 ] = λ(D) (11) = 130(11) = 1430 calls

expected revenue = (0.08) (9570) + (0.13) (1430) = $951.50

40 CHAPTER 5. ARRIVAL-COUNTING PROCESSES

(d)
(D)
Pr{Y1 > 300 | Y1 = 1200}

1200
  
1200
= (0.13)n (0.87)1200−n ≈ 0
n=301 n

(T )
18. Let {Yt ; t ≥ 0} represent arrival of trucks to the restaurant.

Poisson with rate λ(T ) = 10(0.1) = 1/hour

(C)
Let {Yt ; t ≥ 0} represent arrival of cars to the restaurant.

Poisson with rate λ(C) = 20(0.1) = 2/hour

(T ) (C)
(a) Yt = Yt + Yt is Poisson with rate λ = λ(T ) + λ(C) = 3/hour

E[Y1 ] = λ(1) = 3 cars and trucks

(b) Pr{Y1 = 0} = e−3(1) ≈ 0.05

(c) Let C ≡ number of passengers in a car.

E[C] = (0.3)(1) + (0.5)(2) + (0.2)(3) = 1.9

Let P ≡ number of passengers arriving in 1 hour.

 
(T ) (C)
E[P ] = E Y1 + E[C]Y1

= 1 + (1.9)2 = 4.8 passengers

19. Traffic engineer’s model: {Yt ; t ≥ 0} models the number of accidents and is Poisson
with rate λ = 2/week.

(a)
∞
 e−2(2) (2(2))m
Pr{Y2 ≥ 20} =
m=20 m!
19
e−4 4m

= 1− ≈ 1.01 × 10−8
m=0 m!
 41

(b)

Pr{Y52 − Y48 = 20, Y48 = 80 | Y52 = 100}

Pr{Y52 − Y48 = 20, Y48 = 80, Y52 = 100}
=
Pr{Y52 = 100}
Pr{Y52 − Y48 = 20} Pr{Y48 = 80}
=
Pr{Y52 = 100}
Pr{Y4 = 20} Pr{Y48 = 80}
=
Pr{Y52 = 100}
e−2(4) (2(4))20 e−2(48) (2(48))80 100!
= −2(52) (2(52))100
20! 80! e
 20  80   20  80
100! 4 48 100 1 12
= =
20!80! 52 52 20 13 13
≈ 5 × 10−5

(c) Pr{Y2 = 0} = e−2(2) ≈ 0.02

20. (a)
 t
Λ(t) = λ(a)da
0
 t
Λ(t) = 1da = t , for 0 ≤ t < 6
0
 t
Λ(t) = Λ(6) + 2da = 6 + 2a |t6
0

= 6 + 2t − 12 = 2t − 6 , for 6 ≤ t < 13
 t
Λ(t) = Λ(13) + (1/2)da
13

= 20 + (1/2)a |t13 = 20 + (1/2)t − 13/2

= (1/2)(t + 27), for 13 < t ≤ 24

(b) 8 a.m. → hour 2

2 p.m. → hour 8

Λ(8) − Λ(2) = (2(8) − 6) − 2

=8
42 CHAPTER 5. ARRIVAL-COUNTING PROCESSES

∞
 e−8 8m
Pr{Y8 − Y2 > 12} =
m=13 m!

12
 e−8 8m
= 1− ≈ 0.06
m=0 m!

E[Y8 − Y2 ] = 8 patients
(c) 10 a.m. → hour 4
Λ(4) − Λ(2) = 2

Pr{Y4 = 9 | Y2 = 6} = Pr{Y4 − Y2 = 3 | Y2 = 6}
= Pr{Y4 − Y2 = 3}
e−2 23
= ≈ 0.18
3!
(d)
Pr{Y1/4 > 0} = 1 − Pr{Y1/4 = 0}

e−1/4 (1/4)0
= 1−
0!

= 1 − e−1/4 ≈ 0.22

(since Λ(1/4) = 1/4)

(e) 1 p.m. → hour 7

Λ(7) = 2(7) − 6 = 8

∞
e−8 8m

Pr{Y7 ≥ 13} = ≈ 0.06
m=13 m!

(from part (b))

21. (a) Compute the average number of arrivals (and its se) for each hour
hour average (se) estimates
1 144.2 (12.3) λ1
2 229.4 (7.2) λ2
3 382.6 (7.9) λ3
4 96.0 (2.7) λ4
 43



144, 0≤t<1 (after rounding)

229, 1≤t<2
λ(t) =


 383, 2≤t<3

96, 3≤t≤4
(b)
 t
Λ(t) = λ(a)da
0
 t
Λ(t) = 144da = 144t 0≤t<1
0
 t
Λ(t) = Λ(1) + 229da
1

= 144 + 229(t − 1)
= 229t − 85 1≤t<2
 t
Λ(t) = Λ(2) + 383da
2

= 373 + 383(t − 2)
= 383t − 393, 2 ≤ t < 3
 t
Λ(t) = Λ(3) + 96da
3

= 756 + 96(t − 3)
= 96t + 468 3≤t≤4

(c)

E[Y3.4 − Y1.75 ]
= Λ(3.4) − Λ(1.75)
= (96(3.4) + 468) − (229(1.75) − 85)
= 478.65 ≈ 479 cars

(d)

Pr{Y3.4 − Y1.75 > 700}

∞
 e−479 (479)m
=
m=701 m!
700
 e−479 (479)m
= 1− ≈0
m=0 m!
44 CHAPTER 5. ARRIVAL-COUNTING PROCESSES

22. The algorithms given here are direct consequences of the definitions, and not necessarily
the most efficient possible.

(a) Recall that the inverse cdf for the exponential distribution with parameter λ is

X = − ln(1 − U )/λ
algorithm Erlang
a←0
for i ← 1 to n
do
a ← a − ln (1-random())/λ
enddo
return X ← a
(b) algorithm binomial
a←0
for i ← 1 to n
do
U ← random()
if {U ≥ 1 − γ} then
a←a+1
endif
enddo
return X ← a
(c) algorithm Poisson
a←0
b ← − ln(1-random())/λ
while {b < t}
do
b ← b − ln(1-random())/λ
a←a+1
enddo
return X ← a
(d) algorithm nspp
t ← Λ(t)
a←0
b ← − ln(1-random())
 45

while {b < t}
do
b ← b − ln(1-random())
a←a+1
enddo
return X ← a
The first step in algorithm nspp converts the t to the time scale for the rate-1
stationary process.

23.
E[Yt+∆t − Yt ]
= E[Y∆t ] (time stationarity)
∞
 e−λ∆t (λ∆t)k
= k (by definition)
k=0
k!
∞
 e−λ∆t (λ∆t)k
=
k=1 (k − 1)!
∞
 (λ∆t)k−1
= e−λ∆t (λ∆t)
k=1 (k − 1)!
∞

−λ∆t (λ∆t)j
= e (λ∆t)
j=0 j!

= e−λ∆t (λ∆t)eλ∆t
= λ∆t

24.
Pr{Ht = t} = Pr{Yt = 0} = e−λt
Pr{Ht ≤ a} = Pr{Yt − Yt−a ≥ 1} ( at least 1 arrival between t − a and t )
= 1 − Pr{Yt − Yt−a < 1}
= 1 − e−λa
Pr{Ht = t} = e−λt
Pr{Ht ≤ a} = 1 − e−λa , 0 ≤ a ≤ t

Ht

* ** * * * -
0 t−a t
6 6

arrival {Yt − Yt−a ≥ 1} ⇒ {Ht ≤ a}

46 CHAPTER 5. ARRIVAL-COUNTING PROCESSES

Notice that limt→∞ Pr{Ht = t} = limt→∞ e−λt = 0. Therefore, limt→∞ Pr{Ht ≤ a} =

1 − e−λa , a ≥ 0.

25.

E[Lt ] = E[Ht + Rt ]
= E[Ht ] + E[Rt ]
= E[Ht ] + E[G] (memoryless)
> E[G]

since certainly E[Ht ] > 0.

26. Now pG (a) = 1/60 for a = 1, 2, . . . , 60. Therefore, δ = E[G] = 61/2 = 30 12 months.

η = E[R] = E[d(eiG − 1)]

60

= d(eia − 1)pG (a)
a=1


60

1 
=d eia − 1
60 a=1

For the specific case i = 0.06/12 = 0.005

 
70.14
≈d − 1 = d(0.169)
60

= 50, 700 for d = 300, 000

Therefore the long-run rate of return is

η
≈ $1, 662.29 per month
δ

≈ $19, 947.54 per year

This is about $52 per year less than the other model.
 47

27. We suppose that the time to burn-out of lights are independent (almost certainly
true) and identically distributed (true if same brand of bulb and same usage pattern).
Therefore the replacements form a renewal arrival process.
Estimate δ by δ = 248.7 hours (sample average).
Therefore, the long-run replacement rate is

1
≈ 4.021 × 10−4 bulbs/hour
δ

≈ 35 bulbs/year

28. We model it as a renewal-reward process. Therefore, the modeling approximations are:

the times between orders are independent and time-stationary random variables, and
the dollar amount of the orders are also independent and time-stationary random vari-
ables. These approximations are reasonable except on special holidays (e.g., Mother’s
Day) when the rate and size of orders changes.
δ = 15 minutes
η = $27
The long-run revenue/minute is η/δ = $1.8/minute.

29. As given in the hint,

 c  30
δ = a fX (a)da + c fX (a)da
0 c

c3 c2
= +c 1−
1350 900
c3
= c− days
2700

The reward (cost) depends on whether failure occurs before c.

 c  30
η = 1000 fX (a)da + 300 fX (a)da
0 c

7c2
= + 300
9

Therefore, the long-run cost is

48 CHAPTER 5. ARRIVAL-COUNTING PROCESSES


η 7c2 + 2700
= −300 per day
δ c(c2 − 2700)
To minimize the cost we take the derivative w.r.t. c and set it equal to 0, then solve
for c. Of the 3 solutions, the only one in the range [0, 30] is c = 15.9 for which
η
≈ $34.46 per day
δ
Compare this to c = 30 for which

η
= $50 per day
δ
Therefore, it is worthwhile to replace early.
30. Suppose we wait for n patrons before beginning a tour. Then δ = n minutes is the
expected time between tours. The expected cost for a tour of size n is

η = 10 + (0 + (0.50) + (2)(0.50) + (3)(0.50)

+ · · · + (n − 1)(0.50))
(n(n − 1))
= 10 + (0.50)
2
(a) The long-run cost/minute is

η 10 (0.50)(n − 1)
= +
δ n 2
A plot as a function of n reveals that η/δ decreases until n = 6, then increases.
At n = 6, η/δ = $2.92 per minute.
(b) δ = 1(6) = 6 minutes

31. To prove the result for decomposition of m processes, do the decomposition in pairs,
as follows:
(i) Decompose the original process into 2 subprocesses with probabilities γ1 and 1 − γ1 .
Clearly both subprocesses are Poisson, the first with rate λ1 = γ1 λ, and the latter with
rate λ = (1 − γ1 )λ.
(ii) Decompose the λ process into two subprocesses with probabilities γ2 /(1 − γ1 ) and
1 − γ2 /(1 − γ1 ). Clearly both subprocesses are Poisson, the first with rate

 γ2
λ2 = λ = γ2 λ
1 − γ1
 49

and the latter with rate  

 γ2
λ = 1− λ
1 − γ1
(iii) Continue by decomposing the λ process, etc.
To prove the result for superposition of m processes, superpose them two at a time:
(i) Superpose the λ1 and λ2 processes to obtain a Poisson process with rate λ = λ1 +λ2 .
(ii) Superpose the λ and λ3 processes to obtain a Poisson process with rate λ =
λ + λ3 = λ1 + λ2 + λ3 .
(iii) Continue
To prove the result for the superposition of m nonstationary processes, we prove that
for the superposition process
Pr{Yτ +∆τ − Yτ = h | Yτ = k}

e−[Λ(τ +∆τ )−Λ(τ )] [Λ(τ + ∆τ ) − Λ(τ )]h

=
h!
For clarity we prove only the case m = 2.
Let Yi,τ be the NSPP with rate Λi (τ ), for i = 1, 2. Therefore,

Pr{Yτ +∆τ − Yτ = h | Yτ = k}

h
= Pr{Y1,τ +∆τ − Y1,τ = , Y2,τ +∆τ − Y2,τ = h − | Y1,τ + Y2,τ = k}
=0


h
= Pr{Y1,τ +∆τ − Y1,τ = } Pr{Y2,τ +∆τ − Y2,τ = h − }
=0


h
e−∆t1 (∆t1 ) e−∆t2 (∆t2 )h−
= (1)
=0
! (h − )!
where we use the fact that Y1 and Y2 are independent NSPPs, and where ∆ti =
Λi (τ + ∆τ ) − Λi (τ ). Let ∆t = ∆t1 + ∆t2 = Λ(τ + ∆τ ) − Λ(τ ). Then (1) simplifies to
   
e−∆t (∆t)h h
h! ∆t1  ∆t2 h− e−∆t (∆t)h
= · 1
h! =0 !(h − )! ∆t ∆t h!

32. No answer provided.

33. No answer provided.
34. No answer provided.
50 CHAPTER 5. ARRIVAL-COUNTING PROCESSES
Chapter 6

Discrete-Time Processes

1. 2 session 1 2 3 2 3 3 3 3 3 2 3 2 2 3 2 4

Pr{2} = p1 p12 p23 p32 p23 p33 p33 p33 p33 p32 p23 p32 p22 p23 p32 p24
= p1 p12 (p23 )4 (p32 )4 (p33 )4 p22 p24
= 1 (0.95)(0.63)4(0.36)4 (0.4)4 (0.27)(0.1)
≈ 1.7 × 10−6

∗ session 1 2 2 3 3 2 3 3 3 4

Pr{∗} = p1 p12 p22 p23 p33 p32 p23 p33 p33 p34
= p1 p12 p22 (p23 )2 (p33 )3 p32 p34
= 1 (0.95)(0.27)(0.63)2(0.4)3 (0.36)(0.24)
≈ 5.6 × 10−4

2. (a) 2, 2, 3, 2, 4,...
(b) 3, 3, 2, 3, 2, 3, 4,...
3. (a) 1, 1, 1, 1, 1,...
(b) 2, 2, 1, 1, 1, 1, 2,...
4. (a) p12 = 0.4
(2)
(b) p11 = 0.68
(c)p21 p11 = 0.48
5. recurrent: {2, 3, 4}
transient: {1}

51
52 CHAPTER 6. DISCRETE-TIME PROCESSES

6. (a) irreducible, all states recurrent

(b) irreducible subset {2, 5} are recurrent
{1, 3, 4} are transient
(c) irreducible, all states recurrent

7. (a)
6(a)
n=2
 
0.54 0.08 0.09 0.08 0.21
 
 0.27 0.19 0.24 0.02 0.28 
 
 0.08 0.26 0.32 0.18 0.16 
 
 
 0.05 0.10 0.01 0.81 0.03 
0.31 0.15 0.18 0.02 0.34

n=5
 
0.27496 0.16438 0.19033 0.14404 0.22629
 
 0.34353 0.14109 0.16246 0.12082 0.23210 
 
 0.33252 0.11926 0.11720 0.23690 0.19412 
 
 
 0.12455 0.11640 0.06041 0.60709 0.09155 
0.33137 0.14745 0.17200 0.11146 0.23772

n = 20

 
0.2664964441 0.1393997993 0.1400484926 0.2628192095 0.1912360548
 
 0.2675253416 0.1395887257 0.1406686163 0.2602309763 0.1919863403 
 
 0.2632018511 0.1388746569 0.1382171161 0.2707695205 0.1889368557 
 
 
 0.2490755395 0.1364940963 0.1301154884 0.3054031203 0.1789117568 
0.2678558364 0.1396398961 0.1408494124 0.2594398254 0.1922150308

6(b)
n=2
 
0.01 0.55 0.14 0.01 0.29
 
 0 0.74 0 0 0.26 
 
 0 0.35 0.09 0 0.56 
 

 0 0 0.3 0 0.7 

0 0.65 0 0 0.35
 53

n=5
 
0.00001 0.68645 0.00521 0.00001 0.30832
 
 0 0.71498 0 0 0.28502 
 
 0 0.69230 0.00243 0 0.30527 
 

 0 0.6545 0.0081 0 0.3374 

0 0.71255 0 0 0.28745

n = 20

 
1.0 × 10−20 0.7142857122 7.554699530 × 10−11 1.0 × 10−20 0.2857142881
 
 0 0.7142857146 0 0 0.2857142858 
 
 0 0.7142857134 3.486784401 × 10−11 0 0.2857142871 
 
 
 0 0.7142857108 0.0000000001162261467 0 0.2857142895 
0 0.7142857147 0 0 0.2857142859

6(c)
n=2
 
0.03 0.67 0.04 0.15 0.11
 
 0.02 0.74 0 0.08 0.16 
 
 0.03 0.55 0.09 0.24 0.09 
 

 0 0.8 0 0 0.2 

0.01 0.83 0.01 0.03 0.12

n=5
 
0.01739 0.75399 0.00343 0.07026 0.15493
 
 0.01728 0.75624 0.00242 0.06838 0.15568 
 
 0.01737 0.75119 0.00461 0.07201 0.15482 
 
 
 0.0170 0.7578 0.0024 0.0670 0.1558 
0.01736 0.75702 0.00259 0.06860 0.15443

n = 20

 
0.01727541956 0.7564165848 0.002467917087 0.06836130311 0.1554787760
 
 0.01727541955 0.7564165848 0.002467917078 0.06836130306 0.1554787760 
 
 0.01727541956 0.7564165848 0.002467917098 0.06836130312 0.1554787759 
 
 
 0.01727541955 0.7564165848 0.002467917078 0.06836130306 0.1554787759 
0.01727541955 0.7564165848 0.002467917080 0.06836130306 0.1554787759
54 CHAPTER 6. DISCRETE-TIME PROCESSES

7(b)
6(a)
 
0.2616580311
 
 0.1386010363 
 
π=  0.1373056995 
 
 
 0.2746113990 
0.1878238342

6(b)

0.7142857143
π=
0.2857142857

6(c)
 
0.01727541955
 
 0.7564165844 
 
π=  0.002467917078 
 
 
 0.06836130306 
0.1554787759

8. T = {1, 3, 4}
R1 = {2}, R2 = {5}, R3 = {6}
# $
α = 0.7366666667 0.09666666666 0.1666666667

9. No answer provided.

10. In the answer we use all the data. However, it might also be reasonable to have different
models for free throws and shots from the floor.
(a) Let {Sn ; n = 0, 1, . . .} represent the sequence of shots, where M = {0, 1} and 0
corresponds to a miss, 1 corresponds to a made shot.
If each shot is independent with probability p of being made, then Sn is a Markov
chain with one-step transition matrix

1−p p
P=
1−p p

An estimate of p is
30
p = ≈ 0.67
45
 55

with 
 − p)
p(1 
 =
se ≈ 0.07
45

(b) Now let Sn be a Markov chain with


p00 p01
P=
p10 p11

In the sample, let nij be the number of times the transition (i, j) occurred, and ni =
1
ij = nij /ni . The observed values were
j=0 nij . Then p

nij 0 1 ni
0 6 9 15
1 8 21 29

Therefore,

 = 0.40 0.60
P
0.28 0.72
%
ij (1−pij )
p
An estimate of the se of pij is ni
.  is
Therefore, the matrix of se’s

0.13 0.13
 =
se
0.08 0.08

(c) Let Sn be a Markov chain with state space M = {0, 1, 2, 3} corresponding to the
two most recent shots:
0 = miss, miss
1 = miss, made
2 = made, miss
3 = made, made
The observed values were

nij 0 1 2 3 ni
0 2 4 0 0 6
1 0 0 1 8 9
2 5 4 0 0 9
3 0 0 6 13 19
56 CHAPTER 6. DISCRETE-TIME PROCESSES

 
0.33 0.67 0 0
 0 0 0.11 0.89 
 =  
P 
 0.56 0.44 0 0 
0 0 0.32 0.68
 
0.19 0.19 0 0
 0 0 0.10 0.10 
 
 =
se 
 0.17 0.17 0 0 
0 0 0.11 0.11

(d) Assume a made shot, followed by 4 missed shots, followed by a made shot.
 4 p ≈ 0.008
Under model (a) : (1 − p)
Under model (b) : p10 (p00 )3 p01 ≈ 0.011
Under model (c) : In this case we need to know the result of the previous two shots.
If (made, made): p32 p20 (p00 )2 p01 ≈ 0.013
If (miss, made): p12 p20 (p00 )2 p01 ≈ 0.005
Thus, a run of this length is slightly more likely under the first case of model (c).
This is only one small bit of evidence that favors model (c). Essentially, we look to
see if the additional information we account for in model (b) substantially changes the
transition probabilities from model (a). If it does, then we see if the transition proba-
bilities change when moving from (b) to (c). We stop when the additional information
no longer changes the probabilities. For example,

event model (a) model (b) model (c)

make a p = 0.67 p01 = 0.60 p01 = 0.67
shot p11 = 0.72 p21 = 0.44
p13 = 0.89
p33 = 0.68

Notice that model (b) seems to indicate that there is some difference depending on
whether or not Robinson made the most recent shot, compared to model (a). Model
(c) indicates some additional differences based on the last two shots. So maybe there
is a “hot hand,” but because of the large standard errors we cannot be sure.

11. See Chapter 9, Section 9.1.9.

12. See Chapter 9, Section 9.1.3.

13. No answer provided.

 57

14. Case 6.3

Sn represents the destination of the nth rider
M = {1, 2, . . . , m} one state for each region of the city
pij ≡ probability a rider picked up in region i has destination in region j
pi ≡ probability driver starts the day in region i

Case 6.4
Sn represents the nth key typed
 

 

 
M = 1, 2, . . . , 26 27, 28, . . . , 36 37, 38, . . . , 42 43, . . . 
 & '( )& '( )& '( ) & '( ) 
 
letters numbers punctuation space and others

≡ {A, B, . . . , Z, 0, 1, . . . , 9, ; , ·, “, ‘, ?, , space, . . .}

pij ≡ probability next key typed is j given last one was i

pi ≡ probability first letter typed is i

Case 6.5
Sn represents the brand of the nth toothpaste purchase
M = {1, 2, 3, 4, . . . , m} where each state represents a brand of toothpaste
pij ≡ probability a consumer next purchases brand j given they currently use i
pi ≡ probability a consumer uses brand i at the beginning of the study

Case 6.6
Sn represents the high temperature on the nth day
M = {−20, −19, −18, . . . , 109, 110} is temperature in Fahrenheit
pij ≡ probability high temperature tomorrow is j given today’s high is i
pi ≡ probability temperature is i on the first day of the study

Case 6.7
Sn represents the task type of the nth task
M = {1, 2, . . . , m} where each state represents a task such as “move,” “place,”
“change tool,” etc.
pij ≡ probability next task is j given current task is i
pi ≡ probability first task is type i
58 CHAPTER 6. DISCRETE-TIME PROCESSES

Case 6.8
Sn represents accident history for years n − 1 and n
M = {1, 2, 3, 4} as defined in Case 6.8

15. No answer provided.

16. (a)
π2 = 1/4 = 0.25

(b) No answer provided.

(c) Under the binomial model with probability of γ = π2

1
 15!
Pr{Z15 ≤ 1} = π j (1 − π2 )15−j
j=0 j!(15 − j)! 2

= 0.9904 when π2 = 0.01

= 0.0802 when π2 = 0.25

For the Markov-chain model we have to look at all possible sample paths:
• Pr{no defectives}

Pr{no defectives} = p1 (p11 p11 · · · p11 )

= π1 p14
11 = a

• Pr {1 defective} has several cases:

first item defective

Pr = p2 p21 p11 · · · p11 = π2 p21 p13

11 = b

last item defective

Pr = p1 p11 · · · p11 p12 = π1 p13

11 p12 = c

some other item defective

Pr = p1 p11 p11 · · · p12 p21 p11 p11 · · · p11

= π1 p12
11 p12 p21 = d

and there are 13 possible “other ones”

 59

Pr{≤ 1 defective} = a + b + c + 13d

= p12 2
11 (π1 p11 + π2 p21 p11 + π1 p11 p12
+ 13π1 p12 p21 )
= p12 2
11 (π1 (p11 + p11 p12 + 13p12 p21 ) +
π2 p21 p11 )

For (6.11) this ≈ 0.9169

For 16(a) this ≈ 0.1299
Notice that the Markov-chain model gives a lower probability of accepting a good
process, and a higher probability of accepting a bad process.

17. (a) M = {1, 2, 3, 4} = {A, B, C, D} is the location of the AGV

n represents the number of trips

 
0 1/2 1/2 0
 1/3 0 1/3 1/3 
 
P= 
 1/3 1/3 0 1/3 
1/3 1/3 1/3 0

(5)
(b) p24 ≈ 0.1934
(c) π4 = 3/16 = 0.1875

18. (a)
(2, 000, 000)p12 + (2, 000, 000)p22 + (2, 000, 000)p32 = 2, 107, 800

(b) If “today” is 1994, then 16 years have passed

(16) (16) (16)

(2, 000, 000)(p11 + p21 + p31 ) ≈ 2, 019, 566

19. (a) M = {0, 1, 2, . . . , k} is the number of prisoners that remain in the prison
n is the number of months
Let  
i 
b(i, ) = p (1 − p)i−

which is the probability prisoners are paroled if there are i in prison.
Then P has the following form
60 CHAPTER 6. DISCRETE-TIME PROCESSES

 
1 0 0 0 ··· 0
 
 
 

 b(1, 1) b(1, 0) 0 0 ··· 0 

 
 
 


b(2, 2) b(2, 1) b(2, 0) 0 ··· 0 

 
P=  
 
 b(3, 3) b(3, 2) b(3, 1) b(3, 0) ··· 0 
 
 
 
 ... 
 ··· ··· ··· ··· ··· 
 
 
 
b(k, k) b(k, k − 1) b(k, k − 2) b(k, k − 3) · · · b(k, 0)
(m−1)
(b) pk0
(11)
With k = 6, p = 0.1 and m = 12 we obtain p60 ≈ 0.104.
k (m)
(c) =0 pk
6 (12)
With k = 6, p = 0.1 and m = 12 we obtain =0 p6 ≈ 1.7.
(d) For each prisoner, the probability that they have not been paroled by m months
is

q = (1 − p)m
Therefore, 1 − q is the probability they have been paroled before m months.
The prison will close if all have been paroled. Since the prisoners are independent

Pr{all paroled before m} = (1 − q)k = (1 − (1 − p)m )k

20. (a) M = {1, 2, 3} ≡ {good and declared good, defective but declared good, defective
and declared defective}
n ≡ number of items produced
 
0.995 0.005(0.06) 0.005(0.94)
P= 
 0.495 0.505(0.06) 0.505(0.94) 
0.495 0.505(0.06) 0.505(0.94)
(b) π2 = 0.0006
21. (a)

M = {1, 2, 3, 4}
= {(no, no), (no, accident), (accident, no), (accident, accident)}

n counts number of years

 61

 
0.92 0.08 0 0
 0 0 0.97 0.03 
 
P= 
 0.92 0.08 0 0 
0 0 0.97 0.03

(b) π1 ≈ 0.85
(c)

2000(500π1 + 625(π2 + π3 ) + 800π4 )

≈ 2000(500(0.8499) + 625(0.0739 + 0.0739) + 800(0.0023))
≈ $1, 038, 298

22. The proof is by induction.

(1)
For i ∈ A and j ∈ B, fij = pij . Thus,

(1) (0)
FAB = PAA PAB
Now suppose that the result is true for all n ≤ k, for some k > 1.

(k+1) 
fij = Pr{Sk+1 = j, Sk ∈ A, . . . , S2 ∈ A, S1 = | S0 = i}
∈A

= Pr{Sk+1 = j, Sk ∈ A, . . . , S2 ∈ A | S1 = , S0 = i}
∈A

× Pr{S1 = | S0 = i}

= Pr{Sk = j, Sk−1 ∈ A, . . . , S1 ∈ A | S0 = }pi
∈A
 (k)  (k)
= fj pi = pi fj
∈A ∈A

Therefore, in matrix form and using the induction hypothesis

(k+1) (k)
FAB = PAA FAB
(k−1)
= PAA PAA PAB
(k)
= PAA PAB

Therefore, the result is true for n = k + 1, and is thus true for any n.
62 CHAPTER 6. DISCRETE-TIME PROCESSES

23. Let X ≡ the number of times the process is in state j.

Then ∞

X= I(Sn = j)
n=0

and clearly

0, i = j
I(S0 = j) =
1 i=j

Thus,

*∞ +

µij = E[X] = E I(Sn = j)
n=0
∞

= I(i = j) + E[I(Sn = j)]
n=1
∞  
(n) (n)
= I(i = j) + 0(1 − pij ) + 1pij
n=1
∞
(n)
= I(i = j) + pij
n=1

In matrix form

∞
 (n)
M = I+ PT T
n=1
∞
= I+ PnT T
n=1
= (I − PT T )−1

24. T = {1, 2, 3} and

 
0 0.95 0.01
 
PT T =  0 0.27 0.63 
0 0.36 0.40

Therefore,  
1 2.72 2.87
M = (I − PT T )−1 ≈  0 2.84 2.98 


0 1.70 3.46

The expected session length is

 63

µ11 + µ12 + µ13 ≈ 1 + 2.72 + 2.87 ≈ 6.6

25. (a) We first argue that

1, if S1 = j
Vij =
1 + Vhj , if S1 = h = j
Clearly if the process moves to state j on the first step then {Vij = 1}. If the
process moves to some other state h = j, then the Markov property implies that
the future evolution of the process is as if the process started in state h.
Therefore,


m
E[Vij ] = E[Vij | S1 = h] Pr{S1 = h | S0 = i}
h=1

= 1 pij + E[1 + Vhj ]pih
h=j

= pij + (1 + νhj )pih
h=j


m
= pij + pih (1 + νhj ) − pij (1 + νjj )
h=1


m
= pih (1 + νhj ) − pij νjj
h=1

In matrix form

V = P(1 1 + V) − PD
= P(V − D) + 1 1

(b) Since

V = PV − PD + 1 1

π  V = π  PV − π  PD + π  1 1
= π  V − π  D + 1

Therefore,
64 CHAPTER 6. DISCRETE-TIME PROCESSES

π  D = 1
or term by term

πi νii = 1
or

1
νii =
πi
26. Let M = {0, 1, 2, . . . , 10} represent the number of cherries a player has in the basket.
The one-step transition matrix is
 
3/7 1/7 1/7 1/7 1/7 0 0 0 0 0 0
 
 3/7 0 1/7 1/7 1/7 1/7 0 0 0 0 0 
 
 3/7 0 0 1/7 1/7 1/7 1/7 0 0 0 0 
 
 1/7 2/7 0 0 1/7 1/7 1/7 1/7 0 0 0 
 
 
 1/7 0 2/7 0 0 1/7 1/7 1/7 1/7 0 0 
 
P=  1/7 0 0 2/7 0 0 1/7 1/7 1/7 1/7 0 
 
 
 1/7 0 0 0 2/7 0 0 1/7 1/7 1/7 1/7 
 
 1/7 0 0 0 0 2/7 0 0 1/7 1/7 2/7 
 
 1/7 0 0 0 0 0 2/7 0 0 1/7 3/7 
 
 
 1/7 0 0 0 0 0 0 2/7 0 0 4/7 
0 0 0 0 0 0 0 0 0 0 1

Therefore, T = {0, 1, . . . , 9} and R = {10}.

 65

The expected number of spins is

µ00 + µ01 + · · · + µ09 ≈ 15.8

which is the sum of the first row of

M = (I − PT T )−1

5.642555330 1.237076756 1.468321920 1.508490981 1.699310675

 4.430331574 2.051427403 1.388800859 1.464830124 1.619923520

 4.189308265 0.9910305820 2.185061681 1.373952904 1.557546460


 3.624339776 1.132212235 1.088206223 2.150572935 1.430445774

 3.326241706 0.7736368034 1.211218915 1.044607959 2.189326947
= 
 2.944765281
 0.7518819338 0.8321534392 1.159204127 1.064213430

 2.502145104 0.5682528341 0.7588874323 0.7378123674 1.120907044

 2.094007860 0.4917825124 0.5692601344 0.6742348636 0.6995152842


 1.721601833 0.3844024235 0.4797861159 0.4846075656 0.6262492773
1.404367293 0.3172345401 0.3724060269 0.4081372440 0.4426201777

1.105741430 1.019364871 0.9134948380 0.6768445448 0.5307779547

1.191625516 1.002037338 0.9081983520 0.6745406751 0.5394859830 

1.126134901 1.081735894 0.8876763410 0.6647276567 0.5371821134 


1.104626156 1.008894626 0.9654729638 0.6442056457 0.5318856274 

1.000579145 0.9847916242 0.8926316951 0.7239042017 0.5145580951 

1.792814119 0.8807446128 0.8711229501 0.6584135873 0.6004421813 


0.6677006009 1.639625786 0.7440222642 0.5960365279 0.5210550255 

0.7822967692 0.5336608097 1.520642295 0.5051593083 0.4773941689 


0.4032312936 0.6566735018 0.4237868366 1.301420130 0.3978731088 
0.3814764240 0.2980980700 0.5649684897 0.2410233088 1.212223756

(n)
The distribution of the number of spins is f0,10 , n = 1, 2, . . . with A = {0, 1, . . . , 9} and
(n) (n−1)
B = {10} which is the (1, 1) element of FAB = PAA PAB .
66 CHAPTER 6. DISCRETE-TIME PROCESSES

A table for n from 1 to 25 is

(n)
n f0,10
1 0
2 0
3 .02915451895
4 .05997501041
5 .06538942107
6 .06399544408
7 .06018750691
8 .05585344577
9 .05162197768
10 .04764238654
11 .04395328102
12 .04054295131
13 .03739618725
14 .03449304106
15 .03181517402
16 .02934515930
17 .02706689761
18 .02496550906
19 .02302726490
20 .02123949969
21 .01959053092
22 .01806958294
23 .01666671664
24 .01537276451
25 .01417927081

There are a variety of ways to revise the spinner.

Advanced Exercise
Let Xi ≡ number of spins by player i
Then the total number of spins is

2 min{X1 , X2 } = 2Z

Pr{Z ≤ n} = 1 − Pr{Z > n}

= 1 − Pr{min{X1 X2 } > n}
= 1 − Pr{X1 > n, X2 > n}
 67

= 1 − Pr{X1 > n} Pr{X2 > n}

= 1 − (Pr{X1 > n})2

since spins are independent and identically distributed. Notice that Pr{X1 > n} =
(n)
1 − p0,10 . Then letting T be the total number of spins

Pr{T ≤ 2n} = Pr{Z ≤ n}

(n)
= 1 − (1 − p0,10 )2

which implies that

Pr{T = 2n} = Pr{T ≤ 2n} − Pr{T ≤ 2(n − 1)}

(n) (n−1)
= (1 − p0,10 )2 − (1 − p0,10 )2

This result can be used to calculate the distribution or expectation. You will find that
E[T ] < E[2X1 ].

27. (a)

M = {1, 2, 3, 4, 5}
≡ {insert, withdraw, deposit, information, done}

n represents the number of transactions

 
0 0.5 0.4 0.1 0
 
 0 0 0.05 0.05 0.9 
 
P=  0 0.05 0 0.05 0.9 
 

 0 0.05 0.05 0 0.9 

0 0 0 0 1

(b) Both properties are rough approximations at best. Customers are likely to do
only 1 of each type of transaction, so all of their previous transactions influence
what they will do next. Also, the more transactions they have made, the more
likely they are to be done, violating stationarity.
(n)
(c) Need f15 for n = 1, 2, . . . , 20 with A = {1, 2, 3, 4} and B = {5}.
68 CHAPTER 6. DISCRETE-TIME PROCESSES

(n)
n f15
1 0
2 0.9
3 0.09
4 0.009
5 0.0009
6 0.00009
7–20 nearly 0

(d) We need 100µ12 because µ12 is the expected number of times in the withdraw
state.

T = {1, 2, 3, 4} R = {5}
M = (I − PT T )−1
 
1.0 0.5291005291 0.4338624339 0.1481481481
 0 1.005291005 0.05291005291 0.05291005291 
 
=  
 0 0.05291005291 1.005291005 0.05291005291 
0 0.05291005291 0.05291005291 1.005291005

Therefore, 100µ12 ≈ 100(0.5291) = $52.91.

28. Clearly Pr{H = 1} = 1 − pii = γ. Suppose the result is correct for all a ≤ n for some
n > 1. Then

Pr{H = n + 1} = Pr{Sn+1 = i, Sn = i, Sn−1 = i, . . . , S1 = i | S0 = i}

= Pr{Sn+1 = i, Sn = i, . . . , S2 = i | S1 = i, S0 = i} Pr{S1 = i | S0 = i}
= Pr{Sn = i, Sn−1 = i, . . . , S1 = i | S0 = i}(1 − γ)
= (1 − γ)n−1 γ(1 − γ) = (1 − γ)n γ

from the induction hypothesis. Therefore, it is correct for any a.

(1) (1)
29. Since fjj = pjj , RBB = PBB .
We can write

(2) 
fjj = Pr{S2 = j, S1 = h | S0 = j}
h∈A

= Pr{S2 = j | S1 = h, S0 = j} Pr{S1 = h | S0 = j}
h∈A
 69

= Pr{S1 = j | S0 = h}pjh
h∈A
 
= phj pjh = pjh phj
h∈A h∈A

In matrix form

(2)
RBB = PBA PAB

For n ≥ 3
(n)
fjj = Pr{Sn = j, Sn−1 ∈ A, . . . , S1 ∈ A | S0 = j}

= Pr{Sn = j, Sn−1 ∈ A, . . . , S1 = h | S0 = j}
h∈A

= Pr{Sn = j, Sn−1 ∈ A, . . . , S2 ∈ A | S1 = h, S0 = j}
h∈A

× Pr{S1 = h | S0 = j}

= Pr{Sn−1 = j, Sn−2 ∈ A, . . . , S1 ∈ A | S0 = h}pjh
h∈A
 (n−1)  (n−1)
= fhj pjh = pjh fhj
h∈A h∈A

In matrix form

(n)
RBB = PBA (Pn−2
AA PAB )

In the customer-behavior model

 
0 0.01 0.04
 
PAA =  0 0.40 0.24 
0 0 1

PBA = (0 0.63 0.10)

 
0.95
PAB =  0.36 


0

Therefore
70 CHAPTER 6. DISCRETE-TIME PROCESSES

(5) (5)
RBB = f22 ≈ 0.015
e−2 2a
30. Let p(a) = Pr{Xn = a} = a!
,a = 0, 1, . . . .
For given (r, s), the state space is

M = {r, r + 1, . . . , s}
Careful thinking shows that




p(i − j), j = r, r + 1, . . . , i

pij =  0, j = i + 1, i + 2, . . . , s − 1

 i

1 − h=r pih , j = s
s−1
with the exception that pss = 1 − h=r psh .
The one-step transition matrix for (r, s) = (4, 10) is

0.1353352832 0 0 0

 0.2706705664 0.1353352832 0 0

 0.2706705664 0.2706705664 0.1353352832 0


P=  0.1804470442 0.2706705664 0.2706705664 0.1353352832

 0.09022352214 0.1804470442 0.2706705664 0.2706705664

 0.03608940886 0.09022352214 0.1804470442 0.2706705664

0.01202980295 0.03608940886 0.09022352214 0.1804470442

0 0 0.8646647168

0 0 0.5939941504 

0 0 0.3233235840 


0 0 0.1428765398 

0.1353352832 0 0.0526530176 

0.2706705664 0.1353352832 0.0165636087 

0.2706705664 0.2706705664 0.1398690889
The associated steady-state distribution is
 
0.1249489668
 
 0.1250496100 
 
 0.1256593229 
 
 
π= 

0.1258703295 

 0.1188385832 
 
 0.09050677047 
 
0.2891264170
 71

Therefore, the long-run expected inventory position is


s 10

hπh = hπh ≈ 7.4
h=r h=4
72 CHAPTER 6. DISCRETE-TIME PROCESSES

Similarly, for the (r, s) = (3, 12) policy, P =


0.1353352832 0 0 0 0

 0.2706705664 0.1353352832 0 0 0

 0.2706705664 0.2706705664 0.1353352832 0 0


 0.1804470442 0.2706705664 0.2706705664 0.1353352832 0

 0.09022352214 0.1804470442 0.2706705664 0.2706705664 0.1353352832

 0.03608940886 0.09022352214 0.1804470442 0.2706705664 0.2706705664


 0.01202980295 0.03608940886 0.09022352214 0.1804470442 0.2706705664

 0.003437086558 0.01202980295 0.03608940886 0.09022352214 0.1804470442


 0.0008592716393 0.003437086558 0.01202980295 0.03608940886 0.09022352214
0.0001909492532 0.0008592716393 0.003437086558 0.01202980295 0.03608940886


0 0 0 0 0.8646647168

0 0 0 0 0.5939941504 

0 0 0 0 0.3233235840 


0 0 0 0 0.1428765398 

0 0 0 0 0.0526530176 

0.1353352832 0 0 0 0.0165636087 


0.2706705664 0.1353352832 0 0 0.0045338057 

0.2706705664 0.2706705664 0.1353352832 0 0.0010967192 


0.1804470442 0.2706705664 0.2706705664 0.1353352832 0.0002374476 
0.09022352214 0.1804470442 0.2706705664 0.2706705664 0.1353817816

 
0.09091028492
 
 0.09091052111 
 
 0.09089895276 
 
 
 0.09087291478 
 
 0.09094611060 
π=  
 0.09138954271 
 
 
 0.09154300364 
 
 0.08642895347 
 
 
 0.06582378587 
0.2102759303

12

hπh ≈ 7.9
h=3

Therefore, policy B has the largest long-run inventory position.

 73

31. We show (6.5) for k = 2. The general proof is analogous.

Pr{Sn+2 = j | Sn = i, Sn−1 = a, . . . , S0 = z}

m
= Pr{Sn+2 = j, Sn+1 = h | Sn = i, Sn−1 = a, . . . , S0 = z}
h=1


m
= Pr{Sn+2 = j | Sn+1 = h, Sn = i, . . . , S0 = z}
h=1

× Pr{Sn+1 = h | Sn = i, . . . , S0 = z}

We now apply the Markov property


m
= Pr{Sn+2 = j | Sn+1 = h} Pr{Sn+1 = h | Sn = i}
h=1


m
= Pr{Sn+2 = j, Sn+1 = h | Sn = i}
h=1

= Pr{Sn+2 = j | Sn = i}

We show (6.6) for k = 2. The general proof is analogous.

Pr{Sn+2 = j | Sn = i}

m
= Pr{Sn+2 = j, Sn+1 = h | Sn = i}
h=1


m
= Pr{Sn+2 = j | Sn+1 = h, Sn = i} Pr{Sn+1 = h | Sn = i}
h=1

We now apply the Markov and stationarity properties


m
= Pr{S2 = j | S1 = h} Pr{S1 = h | S0 = i}
h=1


m
= Pr{S2 = j, S1 = h | S0 = i}
h=1

= Pr{S2 = j | S0 = i}
independent of n.
74 CHAPTER 6. DISCRETE-TIME PROCESSES

32.
(n)
pij = Pr{Sn = j | S0 = i}

m
= Pr{Sn = j, Sk = h | S0 = i}
h=1


m
= Pr{Sn = j | Sk = h, S0 = i} Pr{Sk = h | S0 = i}
h=1


m
(k)
= Pr{Sn = j | Sk = h}pih
h=1


m
(k)
= Pr{Sn−k = j | S0 = h}pih
h=1


m
(n−k) (k)
= phj pih
h=1


m
(k) (n−k)
= pih phj
h=1

33. No answer provided.

Chapter 7

Continuous-Time Processes

1. π1 = 5/7, π2 = 2/7

2. (a) (i)
 
0 1 0 0 0
 
 2/5 0 2/5 1/5 0 
 
P=  1/4 0 0 0 3/4 
 

 3/6 1/6 0 0 2/6 

0 0 0 1 0

(ii) all recurrent

(iii)
 175 
 282 
 20 
 141 
 
 10 
π=  
 141 
 
 25 
 282 
 
11
141

(b) (i)
 
0 1 0 0 0
 
 1/2 0 1/2 0 0 
 
P=  1/4 0 0 0 3/4 
 

 0 0 0 0 1  
0 0 0 1 0

(ii) T = {1, 2, 3} R = {4, 5}

# $
5/7
(iii) π = 2/7

75
76 CHAPTER 7. CONTINUOUS-TIME PROCESSES

(c) (i)
 
1 0 0 0 0
 
 1/2 0 1/2 0 0 
 
P=  1/4 0 0 0 3/4 
 

 0 0 0 0 1  
0 0 0 0 1
(ii) T = {2, 3, 4} R1 = {1}, R2 = {5}
(iii) πR1 = 1 πR2 = 1

3. (a) (i) For i = 1, 2, . . . , 5

dpi1 (t)
= −pi1 (t) + 2pi2 (t) + pi3 (t)
dt
dpi2 (t)
= pi1 (t) − 5pi2 (t) + pi4 (t)
dt
dpi3 (t)
= 2pi2 (t) − 4pi3 (t)
dt
dpi4 (t)
= pi2 (t) − 6pi4 (t) + 5pi5 (t)
dt
dpi5 (t)
= 3pi3 (t) + 2pi4 (t) − 5pi5 (t)
dt

(ii)
∞
 (6t)n
P(t) = e−6t Qn
n=0 n!
where  
5/6 1/6 0 0 0
 
 1/3 1/6 1/3 1/6 0 
 
Q=  1/6 0 1/3 0 1/2 
 

 1/2 1/6 0 0 1/3 

0 0 0 5/6 1/6
(b) (i) For i = 1, 2, . . . , 5

dpi1 (t)
= −pi1 (t) + 2pi2 (t) + pi3 (t)
dt
dpi2 (t)
= pi1 (t) − 4pi2 (t)
dt
 77

dpi3 (t)
= 2pi2 (t) − 4pi3 (t)
dt
dpi4 (t)
= −2pi4 (t) + 5pi5 (t)
dt
dpi5 (t)
= 3pi3 (t) + 2pi4 (t) − 5pi5 (t)
dt

(ii)
∞
 (5t)n
P(t) = e−5t Qn
n=0 n!
where  
4/5 1/5 0 0 0
 
 2/5 1/5 2/5 0 0 
 
Q=  1/5 0 1/5 0 3/5 
 

 0 0 0 3/5 2/5 

0 0 0 1 0
(c) (i) For i = 1, 2, . . . , 5

dpi1 (t)
= 2pi2 (t) + pi3 (t)
dt
dpi2 (t)
= −4pi2 (t)
dt
dpi3 (t)
= 2pi2 (t) − 4pi3 (t)
dt
dpi4 (t)
= −2pi4 (t)
dt
dpi5 (t)
= 3pi3 (t) + 2pi4 (t)
dt

(ii)
∞
 (4t)n
P(t) = e−4t Qn
n=0 n!
where  
1 0 0 0 0
 
 1/2 0 1/2 0 0 
 
Q=  1/4 0 0 0 3/4 
 

 0 0 0 1/2 1/2 

0 0 0 0 1
78 CHAPTER 7. CONTINUOUS-TIME PROCESSES

4. We give a simulation for the proposed system.

Let M = {1, 2, 3, 4} be as defined in Case 7.2.
Define the following random variables:
G ≡ time gap between regular calls
H ≡ time gap between calls to the chair
X ≡ time to answer regular calls
Z ≡ time to answer chair’s calls
Define the following system events.
e0 () (start of the day)
S0 ← 1
C1 ← FG−1 (random())
C2 ← FH−1 (random())
C3 ← ∞
C4 ← ∞

e1 () (regular call)
if {Sn = 1} then
Sn+1 ← 2
C3 ← Tn+1 + FX−1 (random())
endif
C1 ← Tn+1 + FG−1 (random())

e2 () (chair call)
if {Sn = 1} then
Sn+1 ← 3
C4 ← Tn+1 + FZ−1 (random())
else
if {Sn = 2} then
Sn+1 ← 4
Z ← FZ−1 (random())
C4 ← Tn+1 + Z
C3 ← C3 + Z
endif
 79

endif
C2 ← Tn+1 + FH−1 (random())

e3 () (end regular call)

Sn+1 ← 1

e4 () (end chair call)

if {Sn = 3} then
Sn+1 ← 1
else
Sn+1 ← 2
endif
5. (a) M = {0, 1, 2, 3} represents the number of vans in use at time t days.
Approximate the time between requests and the time in use as exponentially
distributed. Therefore, we have a Markov-process model with
 
− 87 8
7
0 0
 1
− 23 8
0 
 2 14 7 
G= 
 0 1 − 15
7
8
7

3
0 0 2
− 32
(b)
 
0.1267709745
 0.2897622274 
 
π= 
 0.3311568313 
0.2523099667
The rate at which requests are denied is 8/7 π3 ≈ 0.29 requests/day.
(c) 0π0 + 1π1 + 2π2 + 3π3 ≈ 1.71 vans
6. (7.1) for a and n integer

Pr{X > n + a, X > n}

Pr{X > n + a | X > n} =
Pr{X > n}
Pr{X > n + a}
=
Pr{X > n}
(1 − γ)n+a
=
(1 − γ)n
= (1 − γ)a = Pr{X > a}
80 CHAPTER 7. CONTINUOUS-TIME PROCESSES

(7.2) is proved similarly

Pr{Z > t + b}
Pr{Z > t + b | Z > t} =
Pr{Z > t}
e−λ(t+b)
=
e−λt
= e−λb
= Pr{Z > b}

7. For this model

 
−1/2 0 0 0 1/2 0 0 0

 1 −3/2 0 0 0 1/2 0 0 
 
 0 1 −3/2 0 0 0 1/2 0 
 
 0 0 1 −3/2 0 0 0 1/2 
 
G=  

 0 0 0 1 −1 0 0 0 
 0 0 0 0 1 −1 0 0 
 
 
 0 0 0 0 0 1 −1 0 
0 0 0 0 0 0 1 −1

giving
 
0.1290322581
 
 0.06451612903 
 
 0.09677419355 
 
 0.1451612903 
 
π= 



 0.2177419355 
 0.1532258065 
 
 
 0.1209677419 
0.07258064516

The long-run lost sales rate is

λπ0 ≈ 0.129 sales/day

The average inventory level is

7

jπj ≈ 3.56 units
j=0
 81

8. See Exercise 7 for an analysis of the (3,7) policy.

For the (2,7) policy
 
−1/2 0 0 0 0 1/2 0 0

 1 −3/2 0 0 0 0 1/2 0 
 
 0 1 −3/2 0 0 0 0 1/2 
 
 0 0 1 −3/2 0 0 0 0 
 
G=  

 0 0 0 1 −1 0 0 0 
 0 0 0 0 1 −1 0 0 
 
 
 0 0 0 0 0 1 −1 0 
0 0 0 0 0 0 1 −1
 
0.2394366197
 
 0.05633802817 
 
 0.08450704225 
 
 0.1267605634 
 
π=  
 0.1901408451 
 
 0.1901408451 
 
 
 0.07042253521 
0.04225352113

λπ0 ≈ 0.239
7
j=0 jπj ≈ 3.03
Therefore, (2,7) carries a slightly lower inventory level but with more than double the
lost sales rate.

9. To solve this problem we must evaluate p02 (t) for t = 17520 hours (2 years). Using
the uniformization approach with n∗ = 1000, and carrying 10 digits of accuracy in all
calculations,

p̃02 (t) = 0.56

with truncation error ≤ 0.72 × 10−8 .
The reason that this is dramatically larger than the simulation estimate of 0.02 obtained
in Chapter 4 is the difference in repair-time distribution. In Chapter 4 the distribution
is uniform on [4, 48] hours. Here the distribution is exponential on [0, ∞) hours. The
exponential model is pessimistic because it allows repair times longer than 48 hours.

10. Let M = {0, 1, 2} represent the number of failed machines, so that {Yt ; t ≥ 0} is the
number of failed machines at time t hours. When {Yt = 0}, the failure rate is 2(0.01),
since each machine fails at rate 0.01/hour and there are two machines. When {Yt = 1}
82 CHAPTER 7. CONTINUOUS-TIME PROCESSES

the failure rate is 0.02/hour, as stated in the problem. When {Yt = 1 or 2} the repair
rate is 1/24/hour ≈ 0.04/hour.
Therefore,
 
−0.02 0.02 0
G=
 0.04 −0.06 0.02 

0 0.04 −0.04
and

π0 ≈ 0.57, π1 ≈ 0.29, π2 ≈ 0.14

(a) π2 ≈ 0.14 or 14% of the time.

(b) π1 + π2 ≈ 0.43 or 43% of the time.

11. Let Xij be the total time in j given {Y0 = i}.

Let Zj be the time spent in j on a single visit.
Then 
µij = E[Xij ] = E[Zj ]I(i = j) + E[Xkj ]pik
k∈T ,k=i

because if we start in j we must spend Zj there, and we can condition on the first state
visited after state i; the Markov property implies that once the process leaves i for k,
the fact that it started in i no longer matters. But notice that

E[Xkj ] = µkj

Therefore, 
µij = 1/gjj I(i = j) + µkj pik
k∈T ,k=i

Then noting that pik = gik /gii and gjj = gii if i = j, we have

I(i = j)  gik
µij = + µkj
gii k∈T ,k=i gii

or 
gii µij = I(i = j) + gik µkj
k∈T ,k=i

12. Let Zi be the time initially spent in state i.

Then 
νij = E[Vij ] = E[Zi ] + E[Vkj ]pik
k=i,j
 83

by conditioning on the first state visited after i (if it is not j), and recalling that the
Markov property implies that once the process leaves i for k the fact that it started in
i no longer matters. But notice that

E[Vkj ] = νkj

Therefore, 
νij = 1/gii + νkj pik
k=i,j

Then noting that pik = gik /gii


νij = 1/gii + νkj gik /gii
k=i,j

or 
gii νij = 1 + gik νkj
k=i,j

13. (a) We use the embedded Markov chain and holding times to parameterize the Markov
process. The state space and embedded Markov chain are given in the answer to
Exercise 27, Chapter 6.
We take t in minutes, so that 1/ψ1 = 1/2, 1/ψ2 = 1, 1/ψ3 = 2, 1/ψ4 = 1 and ψ5
is ∞. Therefore,
 
−ψ1 0.5ψ1 0.4ψ1 0.1ψ1 0
 


0 −ψ2 0.05ψ2 0.05ψ2 0.9ψ2 

G=  0 0.05ψ3 −ψ3 0.05ψ3 0.9ψ3 
 

 0 0.05ψ3 0.05ψ3 −ψ4 0.9ψ4 
0 0 0 0 0
 
−2 1 0.8 0.2 0
 


0 −1 0.05 0.05 0.9  
=  0 0.025 −0.5 0.025 0.45 
 

 0 0.05 0.05 −1 0.9 
0 0 0 0 0
(b) We might expect the time to perform each type of transaction to be nearly con-
stant, implying that the exponential distribution is not appropriate. The only
way to be certain, however, is to collect data.
(c) We need 1 − p15 (t) for t = 4 minutes. Using the uniformization approach with
n∗ = 16 and carrying 10 digits of accuracy in all calculations,

1 − p̃15 (4) = 1 − 0.88 = 0.12

with truncation error ≤ 0.004.

84 CHAPTER 7. CONTINUOUS-TIME PROCESSES

(d) From Exercise 11


gii µij = I(i = j) + gik µkj
k∈T ,k=i

In this problem T = {1, 2, 3, 4} and we need µ11 + µ12 + µ13 + µ14 . There will be
16 equations in 16 unknowns. To set up these equations it helps to rewrite the
result as


i−1 4

−I(i = j) = gik µkj − gii µij + gik µkj
k=1 k=i+1

for i = 1, 2, 3, 4 and j = 1, 2, 3, 4. Letting x = (µ11 µ21 µ31 µ41 µ12 µ22 . . . µ44 )
we can write the system of equations as

Ax = b

with

−2 1 0.8 0.2 0 0 0 0
 0 0 0 0 −2 1 0.8 0.2


 0 0 0 0 0 0 0 0

 0 0 0 0 0 0 0 0




0 −1 0.05 0.05 0 0 0 0
 0 0 0 0 0 −1 0.05 0.05


 0 0 0 0 0 0 0 0

 0 0 0 0 0 0 0 0
A= 
 0 0.025 −0.5 0.025 0 0 0 0




0 0 0 0 0 0.025 −0.5 0.025
 0 0 0 0 0 0 0 0


 0 0 0 0 0 0 0 0

 0 0.05 0.05 −1 0 0 0 0

 0 0 0 0 0 0.05 0.05 −1


 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0
 85

0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0  

−2 1 0.8 0.2 0 0 0 0  
0 0 0 0 −2 1 0.8 0.2 

0 0 0 0 0 0 0 0  
0 0 0 0 0 0 0 0  
0 −1 0.05 0.05 0 0 0 0  

0 0 0 0 0 −1 0.05 0.05  
0 0 0 0 0 0 0 0  

0 0 0 0 0 0 0 0  
0 0.025 −0.5 0.025 0 0 0 0  
0 0 0 0 0 0.025 −0.5 0.025 


0 0 0 0 0 0 0 0  
0 0 0 0 0 0 0 0  

0 0.05 0.05 −1 0 0 0 0 
0 0 0 0 0 0.05 0.05 −1

 
−1
 0 
 
 
 0 
 
 0 
 
 
 0 
 
 −1 
 
 
 0 
 
 0 
b=  
 0 
 
 
 0 
 
 −1 
 
 
 0 
 
 0 
 
 0 
 
 
 0 
−1
86 CHAPTER 7. CONTINUOUS-TIME PROCESSES

for which the solution is

 
0.5000000000
 0 
 
 
 0 
 
 0 
 
 
 0.5291005291 
 
 1.005291005 
 
 
 0.05291005291 
 
 0.05291005291 
x=  
 0.8677248677 
 
 
 0.1058201058 
 
 2.010582011 
 
 
 0.1058201058 
 
 0.1481481481 
 
 0.05291005291 
 
 
 0.05291005291 
1.005291005

Thus, µ11 + µ12 + µ13 + µ14 ≈ 2.04 minutes.

14. (a) Let M = {0, 1, 2, 3} represent the state of the system, where 0 corresponds to
an empty system; 1 to the voice-mail being busy but the operator idle; 2 to the
operator being busy and the voice mail idle; and 3 to both the voice mail and
operator being busy. Then {Yt ; t ≥ 0} is the state of the system at time t hours.
 
−20 20 0 0
 30 −50 0 20 
 
G= 
 10 0 −30 20 
0 10 50 −60

(b) Need π2 + π3
π  = (17/41, 8/41, 10/41, 6/41)
Therefore π2 + π3 = 16/41 ≈ 0.39 or 39% of the time.

15. Let λ ≡ failure rate, τ ≡ repair rate, and M = {0, 1, 2} represent the number of failed
computers. Then  
−λ λ 0
 
G =  τ −(λ + τ ) λ 
0 0 0

and T = {0, 1} are the transient states.

We need µ00 + µ01 . Using the result in Exercise 11
 87

λµ00 = 1 +λµ10
λµ01 = λµ11
(λ + τ )µ10 = τ µ00
(λ + τ )µ10 = 1 +τ µ01
or
−1 = −λµ00 +λµ10
0= −λµ01 λµ11
0 = τ µ00 −(λ + τ )µ10
−1 = τ µ01 −(λ + τ )µ11

The solution is
* + * +
λ+τ 1
µ00 µ01 λ2 λ
= τ 1
µ10 µ11 λ2 λ

λ+τ 1 2λ+τ
Therefore, E[T T F ] = λ2
+ λ
= λ2
.

System E[T T F ]
A 111,333.33
B 105,473.68
C 100,200.00
D 86,358.28

System A has the largest E[T T F ].

16. (a) Let Hi represent the holding time on a visit to state i.

Let FN |L=i represent the cdf implied by the ith row of P, the transition matrix of
the embedded Markov chain.
1. n ← 0
T0 ← 0
S0 ← s
2. i ← Sn
3. Tn+1 ← Tn + FH−1i (random())
4. Sn+1 ← Sn + FN−1|L=i (random())
5. n ← n + 1
goto 2
(b) Replace step 3 with
3. Tn+1 ← Tn + E[Hi ]
(c) No clocks need to be maintained. When we use 16(b), we only need to generate
Markov-chain transitions and we eliminate variability due to the holding times.
88 CHAPTER 7. CONTINUOUS-TIME PROCESSES

17. Let

N be the number of copies a customer wants

L be the total number of lost sales

Q be the size of the pending order

We assume customers take all available copies if there are fewer available than they
want. We assume only one pending order at a time.

The system events below implement parts (a)–(c).

e0 () (initial inventory level)

S0 ← s (initially s copies)
C1 ← FG−1 (random()) (set clock for first demand)
C2 ← ∞ (no order arrival pending)

e1 () (demand from customer)

N ← FN−1 (random()) (number demanded)

if {Sn ≥ N } then (if enough in inventory)
Sn+1 ← Sn − N (reduce inventory N copies)
else (not enough in inventory)
L ← L + (N − Sn ) (record lost sales)
Sn+1 ← 0 (reduce inventory to zero)
endif

if {Sn+1 ≤ r and C2 = ∞} then (if reorder point and no order pending)

C2 ← Tn+1 + FR−1 (random()) (set clock for lead time)
Q ← s − Sn+1 (set order quantity)
endif

C1 ← Tn+1 + FG−1 (random()) (set clock for next demand)

 89

e2 () (order arrives from manufacturer)

Sn+1 ← Sn + Q (add order to inventory)

if {Sn+1 ≤ r} then (if still below reorder)
C2 ← Tn+1 + FR−1 (random()) (set clock for lead time)
Q ← s − Sn+1 (set order quantity)
endif

18.
∞

Pr{H ≤ t} = Pr{H ≤ t | X = n} Pr{X = n}
n=1

∞

n 
 
= Pr Ji ≤ t γ(1 − γ)n−1
n=1 i=1
n
since X and Ji are independent. But i=1 Ji has an Erlang distribution with parameter
λ and n phases. Therefore,

 
∞ 
 
n−1
e−λt (λt)j 
Pr{H ≤ t} = 1− γ(1 − γ)n−1
 j! 
n=1 j=0

∞ n−1
  e−λt (λt)j
= 1− γ(1 − γ)n−1
n=1 j=0 j
∞
 
n−1
(λt)j
= 1 − γe−λt (1 − γ)n−1
n=1 j=0 j!
∞
 ∞
−λt (λt)j 
= 1 − γe (1 − γ)n−1
j=0 j! n=j+1
 
∞
 ∞
(λt)j   
j−1
= 1 − γe−λt (1 − γ)n − (1 − γ)n 
j=0 j! n=0 n=0

∞

−λt
 (λt)j 1 1 − (1 − γ)j
= 1 − γe −
j=0 j! γ γ
∞
 (λt(1 − γ))j
= 1 − e−λt
j=0 j!

= 1 − e−λt eλt(1−γ)
= 1 − e−λtγ
90 CHAPTER 7. CONTINUOUS-TIME PROCESSES

an exponential distribution.
This result shows that we can represent the holding time in a state (which is exponen-
tially distributed) as the sum of a geometrically distributed number of exponentially
distributed random variables with a common rate. This is precisely what uniformiza-
tion does.

19.
∞
∗t  (n) (g ∗ t)n
pij (t) = e−g qij
n=0 n!
∞
∗t  (n) (g ∗ t)n
= p̃ij (t) + e−g qij
n=n∗ +1 n!

≥ p̃ij (t)
(n)
because g ∗ > 0, t > 0 and qij ≥ 0.
Notice that
∞
∗t  (n) (g ∗ t)n
e−g qij
n=n∗ +1 n!

−g ∗ t

n∗
(n) (g ∗ t)n
= 1−e qij
n=0 n!

∗t 
n∗
(g ∗ t)n
≤ 1 − e−g
n=0 n!

(n)
because qij ≤ 1.

20.

Pr{Z1 ≤ min Zj , H > t}

j=1

= Pr{Z1 ≤ Z2 , . . . , Z1 ≤ Zk , Z1 > t, . . . , Zk > t}

= Pr{t < Z1 ≤ M }
 ∞
= Pr{t < Z1 ≤ M | M = a}λM e−λM a da
t
 ∞
= (e−λ1 t − e−λ1 a )λM e−λM a da
t
 ∞
= e−λ1 t λM e−λM a da
t
 91
 ∞
λM
− (λ1 + λM )e−(λ1 +λM )a da
λ1 + λM t

λM
= e−λ1 t e−λM t − e−(λ1 +λM )t
λ1 + λM

λM
= 1− e−(λ1 +λM )t
λ1 + λM

λ1
= e−λH t
λH
21. (a)
 1 0.95 0.1 0.4 
− 12 12 12 12
 
 0 73
− 1500 73
( 1500 )( 0.63 73
) ( 1500 )( 0.10 ) 
 0.73 0.73 
G=  
 1 1 1 


0 ( 50 )( .36
.60
) − 50 ( 50 )( 0.24
0.60
) 

1
3
0 0 − 13
Notice that for row 2

15 15 1500
1/g22 = = =
1 − p22 0.73 73
so g22 = 73/1500 to account for transitions from state 2 to itself. Similarly for
row 3. The steady state probabilities are the same:
 
0.08463233821
 0.2873171709 
π=



 0.6068924063 
0.02115808455
(b) No answer provided.
(c) The only change in the generator matrix is that the last row becomes (0, 0, 0, 0).
22. Let G1 , G2 , . . . be the times between visits to state 1. Because the process is semi-
Markov, these are independent, time-stationary random variables.
Let Rn be the time spent in state 1 on the nth visit. Then R1 , R2 , . . . are also inde-
pendent with E[Rn ] = τ1 . Therefore we have a renewal-reward process.
From Exercise 25, Chapter 4, the expected number of states visited between visits to 1
is 1/ξ1 . The fraction of those visits that are to state j has expectation ξj /ξ1 . Therefore,

E[Gn ] = τ1 + (ξj /ξ1 )τj
j=1


m
= 1/ξ1 ξj τj
j=1
92 CHAPTER 7. CONTINUOUS-TIME PROCESSES

Thus, the long-run reward rate (fraction of time spent in state 1) is

E[Rn ] ξ1 τ1
π1 = = m
E[Gn ] j=1 ξj τj

A similar argument applies for all other states.

23. We need to show that

ξi /gii ξi /gii
πi =  m = (1)
j=1 ξj /gjj d
satisfies π  G = 0.
Select the jth equation 
πk gkj − πj gjj = 0
k=j

Substituting (1)
 
 ξk /gkk ξj /gjj
gkj − gjj
k=j d d

1  gkj 1
= ξk − ξj
d k=j gkk d
 
1  ?
= pkj ξk − ξj  = 0
d k=j

For the equality to hold, we must have


pkj ξk = ξj (2)
k=j

But for a Markov process pjj = 0, so (2) is equivalent to


m
pkj ξk = ξj
k=1

which is guaranteed because ξ is the steady-state distribution of the embedded Markov

chain.

24. No answer provided.

Chapter 8

Queueing Processes

1. Let µ be the production rate. Sample average production time = 31.9/10 = 3.19
1
minutes. Therefore, µ = 3.19 ≈ 0.31 parts/minute. With 2 machines the rate is

2µ = 0.62 parts/minute.


0, a<0
FX (a) = −0.31a
1−e , a≥0

2.

93
94 CHAPTER 8. QUEUEING PROCESSES

3. • Self-service system
We can approximate the self-service and full-service copiers as independent M/M/1
queues.

self-service queue
 = 1/10 customer/minute
λ
 = 3 minutes so µ
1/µ  = 1/3 customer/minute

ρ = µλ = 3/10
ρ2
wq = λ(1−ρ)
= 9/7 ≈ 1.3 minutes

full-service queue
 = 1/15 customer/minute
λ
 = 7 minutes so µ
1/µ  = 1/7 customer/minute

ρ = 7/15
49
wq = 8
≈ 6.1 minutes

• Full-service system
We can approximate it as a single M/M/2 queue. Based on the sample average of all
customers
 = 1/6 customer/minute
λ
 = 4.6 minutes so µ
1/µ  = 1/4.6 customer/minute

ρ = 2λµ ≈ 0.38
2 ≈ 0.13 customers
π2 ρ
q = (1−ρ)

wq = q ≈ 0.79 minutes

λ
Notice that the full-service system is actually superior for both types of customers.

• Concerns
For the self-service system we modeled the service times of self-service customers as
exponentially distributed, and the service times of full-service customers as exponen-
tially distributed. For the full-service system (M/M/2) we modeled the service times
of all customers together as exponentially distributed. These are incompatible models,
but still o.k. as approximations.

4. (a) We approximate the process of potential arrivals as Poisson, and the service times
as exponentially distributed.
 95

20(1 − i/16), i = 0, 1, 2, . . . , 15
λi =
0, i = 16, 17, . . .


10, i = 1, 2, . . . , 16
µi =
0, i = 17, 18, . . .
The lost-customer rate in state i is 20(i/16). Therefore, the long-run lost customer
rate is

16
 16
20 
η= 20(i/16)πi = iπi
i=0 16 i=0
20
=
16
≈ 10 customers/hour

(b)
16

q = (j − 1)πj ≈ 8.3 customers
j=2

(c) Customers are served at a rate of

20 − η ≈ 10 customers/hour
Therefore, long-run revenue is

(0.50)(10) − 4 = $1 per hour

5. (a) We approximate IE and MS as independent M/M/1 queues.

• IE
λ = 20 letters/day
µ = 25 letters/day
w = wq + 1/µ = 4/25 + 1/25 = 1/5 day
• MS
λ = 15 letters/day
w = wq + 1/µ = 3/50 + 1/25 = 1/10 day
(b) We approximate IE + MS as an M/M/2 queue.
λ = 20 + 15 = 35 letters/day
w = wq + 1/µ = 49/1275 + 1/25 = 4/51 ≈ 0.08 day
By forming a typing pool, performance improves for both departments.
96 CHAPTER 8. QUEUEING PROCESSES

6. Adopting the modeling approximations stated in the exercise we have a Markovian

queue with
λi = 45 customers/hour, i = 0, 1, 2, . . .

µ, i = 1, 2,
µi =
2µ, i = 3, 4, . . .
with µ = 30 customers/hour.

(a) We derive the general result when the second agent goes on duty when there are
k customers in the system.
 # $

 λ ,
j
µ
j = 0, 1, . . . , k − 1
dj =

 λj
µk−1 (2µ)j−k+1
, j = k, k + 1, . . .
 # $j

 µ λ
, j = 0, 1, . . . , k − 1
=  # $
 λ j 1
, j = k, k + 1, . . .
µ 2j−k+1

∞ ∞
 
k−1  (λ/µ)j
dj = (λ/µ)j +
j=0 j=0 j=k
2j−k+1
∞
 j

k−1
(λ/µ)k 
j λ
= (λ/µ) +
j=0 2 j=0 2µ


k−1
(λ/µ)k
= (λ/µ)j +
j=0 2(1 − λ/(2µ))

1
π0 = ∞ = 2/23 when k = 3
j=0 dj

(b)
∞

πj = 1 − π0 − π1 − π2
j=3

= 1 − 2/23 − (2/23)(45/30) − (2/23)(45/30)2

= 27/46 ≈ 0.59
 97

(c)
∞

q = π 2 + (j − 2)πj
j=3
∞
= π0 d2 + (j − 2)π0 dj
j=3
∞
 (λ/µ)j
= π0 d2 + π0 (j − 3)
j=3 2j−k+1
 j
(λ/µ)3 
∞
λ
= π0 d2 + π0 j
2 j=0 2µ

(λ/µ)3 λ 1
= π0 d2 + π0
2 2µ (1 − λ/(2µ))2

2 (λ/µ)3 (λ/(2µ))
= π0 (λ/µ) +
2 (1 − λ/(2µ))2

(45/30)3
2 (45/60)
= 2/23 (45/30) +
2 (1 − 45/60)2
= 45/23 ≈ 1.96 customers

(d) We show that the current trigger is adequate.

# $
q 45 1 1
wq = λ
= 23 45
= 23
1 1 1 53
w = wq + µ
= 23
+ 30
= 690
# $
53 159
= λw = 45 690
= 46
≈ 3.5 < 5.

7. Define Di−1 = Xi−1 = 0.

For i = 1 we have D1 = max{0, −G1 } = 0 which is clearly correct since the first
customer arrives to find the system empty.
Suppose (8.36) holds for all i = 1, 2, . . . , n for some n > 1.

Dn+1 = (G1 + · · · + Gn + Dn + Xn ) − (G1 + · · · + Gn+1 ) (1)

which is the difference between when customer n departed and customer n + 1 arrived,
provided (1) is ≥ 0.
Therefore,
Dn+1 = max{0, Dn + Xn − Gn+1 }
and the result holds for n + 1 also.
98 CHAPTER 8. QUEUEING PROCESSES

8. (a) We approximate each system as an M/M/1 queue.

One-person system
λ = 24 cars/hour
µ = 30 cars/hours
wq = 2/15 ≈ 0.13 hour
w = wq + 1/µ = 1/6 ≈ 0.17 hour
Two-person system
µ = 48 cars/hour
wq = 1/48 ≈ 0.02 hour
w = wq + 1/µ = 1/24 ≈ 0.04 hour
(b) Need π0 for each case
One person π0 = 1/5
Two person π0 = 1/2
(c) The fraction of time that the intercom is blocked is

∞

η= πj = 1 − π0 − π1 − π2
j=3

2

= 1− (1 − ρ)ρj
j=0

One person η ≈ 0.51

Two person η ≈ 0.13
So in the one-person system the intercom is blocked over half the time.
(d) Now we model the system as an M/M/1/6 queue.
• One-person system
π6 ≈ 0.066
so the lost-customer rate is λπ6 ≈ 1.6 customers/hour.
• Two-person system
π6 ≈ 0.008
so λπ6 ≈ 0.2 customer/hour.
9. This is one possible approach to the problem:
Since the user population is relatively large, treat the arrival process as Poisson with
rate
λ = (10000)(0.7) = 7000 calls/day
≈ 292 calls/hour
≈ 4.86 calls/minute
 99

We have assumed a 24-hour day, but perhaps a 12 or 18-hour day is better.

1/µ = 2.5 minutes so µ = 0.4 calls/minute.
If we treat the connect times as exponentially distributed, then we have an M/M/s/s
queue, where s is the number of ports.
A worthwhile table is πs vs. s, since πs is the probability a user cannot connect.

s πs
5 0.63
6 0.56
7 0.49
8 0.43
9 0.37
10 0.31
11 0.25
12 0.20
13 0.16
14 0.12
15 0.09
16 0.06
17 0.04
18 0.03
19 0.02
20 0.01

10. (a) If we model the arrival of calls as Poisson, service times as exponentially distrib-
uted, and no reneging, then we have an M/M/2 queue.
λ = 20 calls/hour
1/µ = 3 minutes so µ = 20 calls/hour
λ
(b) To keep up ρ = 2µ
< 1 or λ < 2µ = 40 calls/hour.
(c) We want the largest λ such that
π2 ρ
wq = ≤ 4/60 hour
(1 − ρ)2
By trial-and-error, λ ≈ 30 calls/hour.
(d) We want the largest λ such that
∞
 7

πj = 1 − πj ≤ 0.15
j=8 j=0

By trial-and-error λ ≈ 31 calls/hour.
100 CHAPTER 8. QUEUEING PROCESSES

(e) Let the reneging rate be β = 1/5 call/minute = 12 calls/hour for customers on
hold.


iµ, i = 1, 2
µi =
2µ + (i − 2)β, i = 3, 4, . . .
with µ = 20.

(f) q = ∞j=3 (j − 2)πj
Therefore, we need the πj for j ≥ 3.
 j
 (λ/µ) , j = 0, 1, 2
j!
dj = j
 ,j λ , j = 3, 4, . . .
2µ2 i=3
(2µ+(i−2)β)

1 1 1
π0 = ∞ ≈ 20 ≈ ≈ 0.36
j=0 dj j=0 dj 2.77
n
since j=0 dj does not change in the second decimal place after n ≥ 20.
Therefore

20

q ≈ (j − 2)πj ≈ 0.137 calls on hold
j=3

11. (a) M = {0, 1, 2, . . . , k + m} is the number of users connected or in the wait queue.

λi = λ, i = 0, 1, . . .


iµ, i = 1, 2, . . . , k
µi =
kµ + (i − k)γ, i = k + 1, k + 2, . . . , k + m
m
(b) q = j=k+1 (j − k)πj
(c) λπk+m (60 minutes/hour)
(d) The quantities in (b) and (c) are certainly relevant. Also wq , the expected time
spent in the hold queue.

12. We approximate the system as an M/M/3/20/20 with τ = 1 program/minute, and µ

= 4 programs/minute


(20 − i)τ, i = 0, 1, . . . , 19
λi =
0, i = 20, 21, . . .


iµ, i = 1, 2
µi =
3µ, i = 3, 4, . . . , 20
 101

 ,j−1

 (20−i)τ
 i=0
, j = 1, 2
µj j!
dj = ,j−1

 (20−i)τ
 i=0
j 3−j , j = 3, 4, . . . , 20
3!µ 3
 j ,j−1
 (τ /µ) i=0 (20 − i), j = 1, 2
j!
=
 (τ /µ)j ,j−1
6 33−j i=0 (20 − i), j = 3, 4, . . . , 20
20
j=0 dj ≈ 453.388
 
0.002205615985
 0.01102807993 
 
 
 0.02619168983 
 
 0.03928753475 
 
 
 0.05565734086 
 
 0.07420978784 
 
 0.09276223484 
 
 
 0.1082226072 
 
 0.1172411578 
 
 
 0.1172411579 
 
π=  0.1074710613 
 
 
 0.08955921783 
 
 0.06716941339 
 
 0.04477960892 
 
 
 0.02612143853 
 
 0.01306071927 
 
 
 0.005441966361 
 
 0.001813988787 
 
 
 0.0004534971966 
 
 0.00007558286608 
0.000006298572176

(a) ∞j=3 πj = 1 − π0 − π1 − π2 ≈ 0.96
(b) Need w.

20

= j πj ≈ 8.22 jobs
j=0

20
 20

λeff = λj π j = (20 − j)πj ≈ 9.78
j=0 j=0

w = /λeff ≈ 0.84 minutes

102 CHAPTER 8. QUEUEING PROCESSES

(c)
20

q = (j − 3)πj ≈ 6.20 jobs waiting
j=4

(d)
3π0 + 2π1 + π2 + 0(π3 + π4 + · · · + π20 ) ≈ 0.055 idle computers

(e)
π0 ≈ 0.02 or 2% of the time

(f)
(d)/3 ≈ 0.018 or 1.8% of the time

13. No answer provided.

14. (a) For the M/M/1/c, let ρ = λ/µ.

(λ/µ)j , j = 0, 1, 2, . . . , c
dj =
0, j = c + 1, c + 2, . . .

Therefore,

 c −1
 1
π0 = di = c
i=0 i=0 ρi

πj = π0 dj = π0 ρj , j = 0, 1, . . . , c

(b) For the M/M/c/c

 j
 (λ/µ) , j = 0, 1, . . . , c
j!
dj =

0, j = c + 1, c + 2, . . .

Therefore,
 c −1
 1
π0 = di = c i
i=0 i=0 (λ/µ) /i!

(λ/µ)j
πj = π0 dj = π0 , j = 0, 1, . . . , c
j!
 103

(c) For the M/M/s/c with s ≤ c


 (λ/µ)j

 , j = 0, 1, . . . , s

 j!

dj = (λ/µ)j
 s! sj−s
, j = s + 1, . . . , c



 0, j = c + 1, c + 2, . . .

Therefore,

(λ/µ)j
πj = π0 dj = π0 , j = 0, 1, . . . , s
j!
and
 j−s
(λ/µ)j (λ/µ)s λ
πj = π0 dj = π0 = π 0
s! sj−s s! sµ
= πs ρj−s , j = s + 1, s + 2, . . . , c

Thus,

Pr{L = j}
Pr{L = j | L ≤ s} =
Pr{L ≤ s}
π0 (λ/µ)j
j!
= s (λ/µ)i
i=0 π0 i!

(λ/µ)j /j!
= s i
i=0 (λ/µ) /i!

Also,

Pr{L = s + j}
Pr{L = s + j | L ≥ s} =
Pr{L ≥ s}
πs ρj
= c i−s
i=s πs ρ

ρj
= c−s
i=0 ρi

15. No answer provided.

104 CHAPTER 8. QUEUEING PROCESSES

16.
 t
Λ(t) = λ(t)dt
0
 3
 t ,
3
0 ≤ t < 10
=
 2000 (t−20)3
3
+ 3
, 10 ≤ t < 20

Therefore, the expected number of arrivals in [9, 10] is

271
Λ(10) − Λ(9) = ≈ 90.3
3
which is also the expected number in the M (t)/D/∞ queue at time t = 10.
For the stationary model M/D/∞ with λ = 100 and µ = 1

λ
= = 100
µ
The stationary model predicts congestion to be over 10% greater than it should be.
This happens because λ(t) increases sharply toward its peak rate, then declines, while
the M/D/∞ model uses a constant rate. Thus, the M (t)/D/∞ is not at its peak rate
long enough to achieve so much congestion.

17. Notice that


s−1
Pr{Wq = 0} = πj
j=0

And

∞

Pr{Wq > a} = Pr{Wq > a | L = j} πj
j=0

∞

= Pr{Wq > a | L = j} πj
j=s

since no waiting occurs if there are fewer than s in the system. But, for j ≥ s,


j−s
e−sµa (sµa)n
Pr{Wq > a | L = j} =
n=0 n!

which is Pr{T > a}, where T has an Erlang distribution with parameter sµ and j −s+1
phases. This follows because each of the j −s+1 customers (including the new arrival)
 105

will be at the front of the queue of waiting customers for an exponentially distributed
time with parameter sµ. Therefore,
 
∞  j−s
  −sµa n
e (sµa)
Pr{Wq > a} = π
 n!  j
j=s n=0
 
∞  j−s
  e−sµa (sµa)n  π0 (λ/µ)j
=   s! sj−s
j=s n=0 n!

But notice that

π0 (λ/µ)j (λ/µ)s
= π0 (λ/(sµ))j−s
s! sj−s s!
= πs ρj−s
Therefore,

 
∞  j−s
  e−sµa (sµa)n  j−s
Pr{Wq > a} = πs  
ρ
j=s n=0 n!
 
∞ 
 j
e−sµa (sµa)n  j
= πs  
ρ
j=0 n=0 n!
∞
∞ 
 e−sµa (sµa)n j
= πs ρ
n=0 j=n n!
∞
 ∞
(sµa)n 
= πs e−sµa ρj
n=0 n! j=n

∞


 (sµa)n ρn
−sµa
= πs e
n=0 n! 1−ρ
∞
πs e−sµa  (sµaρ)n
=
1 − ρ n=0 n!
πs e−sµa sµaρ
= e
1−ρ
πs e−(sµ−sµρ)a
=
1−ρ
πs e−(sµ−λ)a
=
1−ρ
106 CHAPTER 8. QUEUEING PROCESSES

Finally,

Pr{Wq > a}
Pr{Wq > a | Wq > 0} =
Pr{Wq > 0}
(πs e−(sµ−λ)a )/(1 − ρ)
= ∞
j=s πj

But ∞ ∞ ∞
  j−s
 πs
πj = πs ρ = πs ρj =
j=s j=s j=0 1−ρ
so
Pr{Wq > a | Wq > 0} = e−(sµ−λ)a
18. Model the system as an M/M/s queue with λ = 1/3 customer/minute, µ = 1/2
customer per minute, and s the number of ATMs.
They want

Pr{Wq > 5 | Wq > 0} = e−(s/2−1/3)5

= exp(−(5/2)s + 5/3)
≤ 0.15

Since exp(−5/2 + 5/3) ≈ 0.43 and exp(−5 + 5/3) ≈ 0.03, 2 ATMs are adequate.
19. (a) Since εa = 1 for the exponential distribution

 
1 + εs
wq = wq (λ, 1, µ, εs , 1) = wq (λ, µ, 1)
2
 
1 + εs ρ2
=
2 λ(1 − ρ)
σ2
But, εs = (1/µ)2
= µ2 σ 2 . Thus,

 
1 + µ2 σ 2 ρ2
wq =
2 λ(1 − ρ)
ρ2 + ρ2 µ2 σ 2
=
2λ(1 − ρ)
ρ 2 + λ2 σ 2
=
2λ(1 − ρ)
 107

(b) For the M/M/1

ρ2
wq (λ, µ, 1) =
λ(1 − p)
For the M/D/1

ρ2 1
wq (λ, 1, µ, 0, 1) = = wq (λ, µ, 1)
2λ(1 − ρ) 2
20. No answer provided.

21. Let osu.edu be station 1, and eng.ohio-state.edu be station 2 in a network of

queues. Then

j=1 j=2
a0j 2 1
µ(j) 3 2
s(j) 1 1
 
0 0.2
R=
0 0
Therefore,
λ(1) = a01 = 2/second
λ(2) = a02 + 0.2 λ(1) = 1.4/second
We approximate it as a Jackson network.
ρ21 (2/3)2
(a) (1)
q = 1−ρ1
= 1−2/3
= 4/3 messages
(1)
wq(1) =
q
λ(1)
= 2/3 second
ρ22 (1.4/2)2
(b) (2)
q = 1−ρ2
= 1−(1.4/2)
≈ 1.63 messages
(2)
wq(2) =
q
λ(2)
≈ 1.2 seconds
(c) (12 K/message) (1.63 messages) = 19.56K
(d) Let h be the inflation factor, so that

λ(1) = ha01 = 2h
λ(2) = ha02 + 0.2λ(1)
= h + 0.2(2h)
= 1.4h
108 CHAPTER 8. QUEUEING PROCESSES

The expected time to final delivery is

1 1
w(1) + w(2) = wq(1) + (1)
+ wq(2) + (2)
µ µ
ρ21 1 ρ22 1
= + + +
λ (1 − ρ1 ) 3 λ (1 − ρ2 ) 2
(1) (2)

(2h/3)2 1 (1.4h/2)2 1
= + + +
2h(1 − 2h/3) 3 1.4h(1 − 1.4h/2) 2

A plot of this function shows that h = 1.21 (or 21% increase) is the most that
can be allowed before it exceeds 5 seconds.

22. (8.24)

µ(2) (1 − ρ1 )ρ2 (1 − ρ2 ) = λ(1 − ρ1 )(1 − ρ2 )

µ(2) ρ2 = λ
λ = λ

(8.25)

λρ1i−1 (1 − ρ1 )(1 − ρ2 ) + µ(2) ρi1 (1 − ρ1 )ρ2 (1 − ρ2 )

= (λ + µ(1) )ρi1 (1 − ρ1 )(1 − ρ2 )
λ + µ(2) ρ1 ρ2 = (λ + µ(1) )ρ1
λ + λρ1 = λρ1 + λ

(8.26)

µ(1) ρ1 (1 − ρ1 )ρj−1 (2)

2 (1 − ρ2 ) + µ (1 − ρ1 )ρ2 (1 − ρ2 )
j+1

= (λ + µ(2) )(1 − ρ1 )ρj2 (1 − ρ2 )

µ(1) ρ1 + µ(2) ρ22 = (λ + µ(2) )ρ2
λ + λρ2 = λρ2 + λ

Finally,

∞ 
 ∞
ρi1 (1 − ρ1 )ρj2 (1 − ρ2 )
i=0 j=0

∞
 ∞

= (1 − ρ1 )(1 − ρ2 ) ρi1 ρj2
i=0 j=0
 109

∞
 1
= (1 − ρ1 )(1 − ρ2 ) ρi1
i=0 1 − ρ2
1 1
= (1 − ρ1 )(1 − ρ2 )
1 − ρ1 1 − ρ2
= 1

23. Let B be a Bernoulli random variable that takes the value 1 with probability r. The
event {B = 1} indicates that a product fails inspection.
The only change occurs in system event e3 .

e3 () (complete inspection)

S2,n+1 ← S2,n − 1 (one fewer product at inspection)

if {S2,n+1 > 0} then (if another product then start)
C3 ← Tn+1 + FZ−1 (random()) (set clock for completion)
endif
B ← FB−1 (random())
if {B = 1} then (product fails inspection)
S1,n+1 ← S1,n + 1 (one more product at repair)
if {S1,n+1 = 1} then (if only one product then start)
C2 ← Tn+1 + FX−1 (random()) (set clock for completion)
endif
endif

24. The balance equations are

rate in = rate out

(2)
(1 − r)µ π(0,1) = λπ(0,0)

λπ(i−1,0) + rµ(2) π(i−1,1) + (1 − r)µ(2) π(i,1) = (λ + µ(1) )π(i,0) , i > 0

µ(1) π(1,j−1) + (1 − r)µ(2) π(0,j+1) = (λ + µ(2) )π(0,j) , j > 0

λπ(i−1,j) + µ(1) π(i+1,j−1) + rµ(2) π(i−1,j+1) + (1 − r)µ(2) π(i,j+1) = δπ(i,j) , i, j > 0

where δ = λ + µ(1) + µ(2) .

The steady-state probabilities are
110 CHAPTER 8. QUEUEING PROCESSES

π(i,j) = (1 − ρ1 )ρi1 (1 − ρ2 )ρj2

where
λ(i) λ/(1 − r)
ρi = (i) =
µ µ(i)

Verification follows exactly as in Exercise 22.

25. To remain stable requires

λ(i) 5/(1 − r) 5
ρi = (i) = = < 1 for i = 1, 2
µ µ(i) µ(i) (1 − r)

For i = 1
5
ρ1 = <1
6(1 − r)
Therefore, r < 1/6.
For i = 2

5
ρ2 = <1
8(1 − r)
Therefore r < 3/8. So we require r < 1/6.

26.
(j)
bjk = rjk εd + (1 − rjk )
= rjk ε(j)
a + (1 − rjk )
m 
 
aij
= rjk bij + (1 − rjk )
i=0 λ(j)

m    
aij a0j
= rjk b ij + r jk b0j + (1 − rjk )
i=1 λ(j) λ(j)


m
rij λ(i)
= rjk bij + djk
i=1 λ(j)

m
= rjk cij bij + djk
i=1

27. Approximate the system as a Jackson network of 2 queues with the following parame-
ters:
 111

i a0i µ(i)
1 20 30
2 0 30

 
0 0.9
R=
0 0
Therefore,
λ(1) = a01 = 20
λ(2) = r12 λ(1) = 18
and each station is an M/M/1 queue.
c
(a) For each station individually, we want the minimum c such that j=0 πj ≥ 0.95.
For the M/M/1, πj = (1 − ρ)ρj , so


c
1 − ρc+1
(1 − ρ)ρj = (1 − ρ) ≤ 0.95
j=0 1−ρ

Therefore,

ρc+1 ≥ 0.05
ln(0.05)
c+1 ≥
ln(ρ)
ln(0.05)
c ≥ −1
ln(ρ)

i ρi = λ(i) /µ(i) c
1 2/3 7
2 3/5 5

(b)

(2) = λ(2) w(2) = λ(2) (wq(2) + 1/µ(2) )


(2) ρ22 1
= λ + (2)
λ (1 − ρ2 ) µ
(2)

= 3/2 jobs

(c) ρ2 = 3/5 = 0.6

112 CHAPTER 8. QUEUEING PROCESSES

(d) We now have a network of 3 queues, with queue 3 being the rework station for
station 1.
 
0 0.9 0.1

R= 0 0 0 
0 0.5 0

λ(1) = 20
λ(2) = 0.9 λ(1) + 0.5λ(3)
λ(3) = 0.1 λ(1)
Therefore,
λ(1) = 20
λ(2) = 18 + 1 = 19
λ(3) = 2
For station 1 nothing changes. For station 2

ρ2 = 19/30 ≈ 0.63
c = 6

(2) (2) ρ22 1
= λ +
λ(2) (1 − ρ2 ) µ(2)
= 19/11 ≈ 1.7 jobs

(e) This change has no impact on the results from part (d), but would change the
performance of the rework stations.

28. Model as a Jackson network of infinite-server queues, with a queue corresponding to

each type of transaction.

transaction i a0j µ(j) per hour

log on 1 1000 or 1500 20
fetch 2 0 20
read 3 0 5
log off 4 0 100

 
0 0.79 0.01 0.2
 0 0.17 0.63 0.2 
 
R= 
 0 0.16 0.4 0.44 
0 0 0 0
 113

 
1000
 0 
A1 = 



 0 
0
or  
1500
 0 
A2 = 



 0 
0
 
1.0 0 0 0
 1.197381672 1.510574018 0.4028197382 0 
 
Λ[(I − R) ]−1 A =  A
 1.273917422 1.586102719 2.089627392 0 
1.0 1.0 1.0 1.0
 
1000.0
 1197.381672 
 
=  when A = A1
 1273.917422 
1000.0
or  
1500.0
 1796.072508 
=


 when A = A2
 1910.876133 
1500.0

(i) = λ(i) /µ(i)

i A1 A2
1 50 75
2 60 90
3 255 382
4 10 15
expected number of 375 562
simultaneous users

29. We approximate the job shop as a Jackson network with a single job type.

queue i name
1 casting
2 planer
3 lathe
4 shaper
5 drill
114 CHAPTER 8. QUEUEING PROCESSES

 
0 0.54 0 0.46 0
 
 0 0 0.46 0 0 
 
R=  0 0 0 0 0 
 

 0 0 0 0 1 
0 1 0 0 0

We obtain R by noticing that when a job departs a machine group that serves both
job types, the probability it is a type 1 job is (historically) 460/(460+540) = 0.46.
 
1/11
 
 0 
 
A=  0 
 
 
 0 
0
because all jobs arrive initially to the casting group, and the sample-mean interarrival-
time gap was 11 minutes.

λ(1) = 1/11
λ(4) = (0.54)λ(1)
λ(5) = λ(4) = (0.54)λ(1)
λ(2) = (0.46)λ(1) + λ(5) = λ(1)
λ(3) = (0.46)λ(2) = (0.46)λ(1)

We approximate the expected service times as follows:

1
= (0.46)(125) + (0.54)(235) ≈ 184.4
µ(1)
1
= (0.46)(35) + (0.54)(30) ≈ 32.3
µ(2)
1
= 20
µ(3)
1
= 250
µ(4)
1
= 50
µ(5)

The utilizations are therefore ρi = λ(i) /(si µ(i) ).

 115

i si ρi
1 19 0.88
2 4 0.75
3 3 0.27
4 16 0.77
5 5 0.50

The casting units are the most heavily utilized.

Modeling each group as an M/M/s queue we obtain

i q(i) wq(i) (minutes)

1 3.7 41.1
2 1.3 14.7
3 0.02 0.5
4 0.8 15.8
5 0.1 2.4

Clearly the longest delay occurs at the casting units. This appears to be the place to
add capacity.
The expected flow times for each job type are approximated as
type 1: 41.1 + 125 + 14.7 + 35 + 0.5 + 20 = 263.3 minutes
type 2: 41.1 + 235 + 15.8 + 250 + 2.4 + 50 = 639 minutes
The expected WIP is

5
 5

(i)
= λ(i) w(i)
i=1 i=1

5

= λ(i) (w(i) q + 1/µ(i) )
i=1

≈ 39 jobs

Comments: The approximation might be improved by adjusting for variability. The

sample squared coefficient of variation of the interarrival times is

(23.1)
εa = ≈ 0.19 < 1
(11.0)2

We might also expect εs(i) < 1. These adjustments would reduce q , wq , flow times and
WIP.
116 CHAPTER 8. QUEUEING PROCESSES

30. We have a network of two queues, one for the packer and one for the forklifts, denoted
i = 1 and i = 2 respectively. Define a customer to be 1 case = 75 cans. Therefore,
1/λ(1) = 150 seconds = 2.5 minutes.
1 1
Clearly, λ(2) = λ(1) , µ(1) = 2 minutes, µ(2)
= 3 + 1 = 4 minutes.
If s2 is the number of forklifts, then just to keep up we must have

λ(2) 4
ρ2 = (2)
= <1
s2 µ s2 (2.5)

Therefore, at least s2 = 2 forklifts are required.

To approximate (2)
q and w
(2)
we need to go further. Notice that

ε(1)
a = 0
(3−1)2
1
ε(1) = 12
=
s
22 12
(1)
εd = ε(1)
a (from (A5))
(1)
ε(2)
a = εd = 0 (since all departures from the packer go to the forklifts)
12
ε(2)
s = = 1/4
(3 + 1)

Therefore, for the forklifts we can approximate wq(2) as

wq(2) ≈ wq (λ(2) , ε(2) (2) (2)

a , µ , εs , s2 )

ε(2)
a + εs
(2)
= wq (λ(2) , µ(2) , s2 )
2
= (1/8)wq (1/2.5, 1/4, s2 )

w(2) = wq(2) + 1/µ(2)

and
wq(2)
(2)
q =λ (2)
wq(2) =
2.4

s2 wq(2) (2)
q w(2)
2 0.89 0.35 4.89
3 0.10 0.04 4.10
 117

It appears that there will be very little queueing even with the minimum of 2 forklifts.

31. We first approximate the drive-up window as an M/M/s/4 queue with λ1 = 1/2
customer/minute and µ1 = 1/1.8 customer/minute. (This approximation is rough
because the true service-time distribution is more nearly normal, and εs < 1.)

s π4
1 0.16
2 0.03

π4 is the probability that an arriving car will find the queue full and thus have to park.
Adding a window reduces this dramatically.
The rate at which customers are turned away from the drive-up window is λ1 π4 =
(1/2)π4 . Therefore, the overall arrival rate into the bank is λ2 = 1+(1/2) π4 customers
per minute. This process will not be a Poisson process, but we approximate it as one.
As a first cut we model the tellers inside the bank as an M/M/s2 queue with µ2 = 1/1.4
customer/minute.

(s1 , s2 ) λ2 wq (minutes)
(1, 2) 1.08 1.9
(2, 2) 1.02 1.5
(1, 3) 1.08 0.2

The bank can now decide which improvement in performance is more valuable.
Comment: The approximation for the inside tellers can be improved by using the
GI/G/s adjustment. Clearly, εs = 1.0/(1.4)2 ≈ 0.5. The exact value of εa can also be
computed, but requires tools not used in this book, so set εa = 1 for a Poisson process.

32. No answer provided.

33. No answer provided.

34. No answer provided.

35. No answer provided.

118 CHAPTER 8. QUEUEING PROCESSES
Chapter 9

Topics in Simulation of Stochastic

Processes
√
1. Solve σ / k ≤ 20 which gives k = (110/20)2 ≈ 31 replications.

2–7. No answer provided.

8. To obtain a rough-cut model we will (a) use a 0.5 probability that an item joins the
queue of each inspector, rather than selecting the shortest queue, and (b) treat all
processing times as exponentially distributed (later we will refine (b)).

• Current System

station j name a0j µ(j)

1 repair 1 0.125 0.167
2 repair 2 0.125 0.167
3 inspect 0 0.333

 
0 0 1

R= 0 0 1 

0.1 0.1 0
 
0.125
 
A =  0.125 
0
   
λ(1) 0.156
Λ =  λ  = [(I − R) ] A =  0.156 
 (2)   −1 

λ(3) 0.313

119
120 CHAPTER 9. TOPICS IN SIMULATION OF STOCHASTIC PROCESSES

The utilization of the technicians is

λ(j)
ρ(j) = ≈ 0.93 for j = 1, 2
µ(j)

Treating each station as an M/M/1 queue the expected flow time is


1 1
wq(1) + (1) + wq(3) + (3) (1.25)
µ µ

≈ (85 + 6 + 47 + 3)(1.25) ≈ 141 minutes

since wq(1) = wq(2) and the expected number of cycles through the system is 1/0.8 = 1.25
cycles.

• Proposed System
The repair technicians now become a single station with λ = λ(1) + λ(2) = 0.313 and
λ
µ = 0.167. The utilization is ρ = 2µ ≈ 0.93, so it is unchanged.
Treating the repair station as an M/M/2 queue, the expected flow time is


1 1
wq + + wq(3) + (3) (1.25)
µ µ
≈ (43 + 6 + 47 + 3)(1.25)
≈ 124 minutes

which is a dramatic reduction.

• Refinement
We can refine the approximation by using the ideas in Section 8.10. Notice that

j ε(j)
s
1 0.0625
2 0.0625
3 0.0370

Since the arrival process is Poisson, ε(j)

a = 1 throughout. Therefore, the preceding
values of wq(j) can be modified by the factor (1 + ε(j)
s )/2.
 121

9. To obtain a rough-cut model, first note that the registration function is not really
relevant to the issue of additional bed versus additional doctor, so we will ignore it.
We also (a) ignore the interaction of beds and doctors, (b) treat treatment time as
exponentially distributed, and (c) develop a composite patient and ignore patient type
and priority.

• Additional doctor.
With an additional doctor there is a one-to-one correspondence between bed and doc-
tor.
λ = 1/20 patient/minute
1/µ = 72(0.15) + 25(0.85) ≈ 32 minutes
Using an M/M/3 model
q = 0.3 patients waiting for a doctor and a bed
wq = 6.3 minutes wait for a doctor and a bed

• Additional bed.
In terms of time to wait to see a doctor, the beds are not a constraint. Using an
M/M/2 model
q = 2.8 patients waiting to see a doctor
wq = 57 minutes to see a doctor
The expected number of patients in the system is

= λw = λ(wq + 1/µ) ≈ 4.45

Thus, with 4 beds the expected number waiting for a bed is 4.45 − 4 = 0.45, which is
larger than the 0.3 patients when adding another doctor.
It appears that adding a doctor will be better in terms of time to see a doctor and
time to reach a bed.

10. To obtain a rough-cut model we will replace all random variables by their expected
values and treat it as deterministic.
Suppose the company is open 8 hours per day. Then the period of interest is 90(8) =
720 hours.
The number of lost sales during 1 day is (8 hours) (1 customer/hour) (6 bags/customer)
= 48 bags.
If the company orders s bags, they will run out in
s bags
= s/6 hours
6 bags/hour
122 CHAPTER 9. TOPICS IN SIMULATION OF STOCHASTIC PROCESSES

so a complete cycle is approximately s/6 + 8 hours.

The number of orders for the summer is
720
s/6 + 8

The expected number of lost bags is


720
(48)
s/6 + 8

And the average stock level is

1/2(s/6)s
s/6 + 8

11. To obtain a rough-cut model we will (a) ignore the interaction of loading bays and
forklifts, and (b) treat all service times as exponentially distributed (later we refine
(b)).

• Adding a fork lift.

We first treat the number of bays as unlimited.
λ = 6.5 trucks/hour
1/µ = 15 minutes = 1/4 hour
Based on an M/M/3 model
wq = 0.05 hours = 3 minutes
q = 0.3 trucks
6.5
ρ= λ
3µ
= 12
≈ 0.54 utilization
Now treating the bays as servers we use an M/M/4 model with
λ = 6.5
1/µ = 15 + 15 = 30 minutes = 1/2 hour
wq = 0.41 hour = 25 minutes
q = 2.7 trucks in the lot

Approximate total time: 0.41 + 0.05 + 1/2 = 0.96 hour ≈ 58 minutes.

• Adding a bay
Again start by treating the bays as unlimited, and the forklifts as an M/M/2 with
λ = 6.5
1/µ = 1/4
 123

wq = 0.49 hours
q = 3.2 trucks
λ
ρ= 2µ
= 0.81 utilization
Next treat the bays as servers in an M/M/5 model with
λ = 6.5
1/µ = 1/2
wq = 0.09 hours = 5.4 minutes
q = 0.6 trucks in the lot

Approximate total time: 0.09 + 0.49 + 1/2 = 1.08 hours ≈ 65 minutes.

Adding a bay looks better, but only by about 7 minutes. The values of wq and q can
be refined by using the ideas in Section 8.10. Notice that

εs
forklift 0.01
(1.52 +3(5)2 )
bay (15+15)2
≈ 0.09

12. For n = 1, p(1) = p P = π  P = π  by Equation (6.21).

Suppose the result is true for all n ≤ k, for some k > 1. Then

p(k+1) = p Pk+1
= (p Pk )P
= π P
= π

Therefore, the result is true for all n by induction.

13. No answer provided.

14. No answer provided.

15.  # $ # $
k k
1  k
j=1 Xj h=1 Zh

σXZ = Xi Z i −
k − 1 i=1 k

In this example (using Table 9.2, columns 2 and 4)

k = 10
124 CHAPTER 9. TOPICS IN SIMULATION OF STOCHASTIC PROCESSES
k
i=1 Xi Zi = 644.99
k
j=1 Xj = 80.3
k
h=1 Zh = 80.3
σXZ ≈ 0.02
16.

k 
k
(Vi − V̄ )2 = (Xi − Zi − (X̄ − Z̄))2
i=1 i=1


k
= ((Xi − X̄) − (Zi − Z̄))2
i=1

k 

= (Xi − X̄)2 − 2(Xi − X̄)(Zi − Z̄)
i=1

+(Zi − Z̄)2


k 
k
2
= (Xi − X̄) + (Zi − Z̄)2
i=1 i=1


k
−2 (Xi − X̄)(Zi − Z̄)
i=1

2
= (k − 1)(σX + σZ2 − 2σXZ )

17. There are mostly disadvantages.

• Unless we understand the bias quite well, we have no idea what overload approxi-
mately compensates the underload. Therefore, the bias may still be quite significant,
and we will still have to study it.
• When we attempt to study the bias, we must either study the underload and overload
individually, which is twice as much work, or study them together, in which case the
combined bias process may have very unusual behavior.
An advantage is that we might be able to bound the bias by looking at convergence
from above and below.
18. If θ is an estimator of θ, then the bias of θ is E[θ − θ] where the expectation is with

respect to the distribution of θ.
Let S0 be the initial state, as in the Markov chain. Then, if we sample the initial state
from the steady-state distribution, we proved

 =
E[θ] E[θ | S0 = x] Pr{S0 = x} = θ (1)
x∈M
 125

When we fix the initial state, the expectation of interest is

E[θ | S0 = 1]

a conditional expectation, only 1 term of (1).

Remember that “expectation” is a mathematical averaging process over the possible
outcomes. When we sample the initial state, all initial states are possible. When we
fix the initial state, only 1 initial state is possible. Therefore, the probability distri-
bution of the estimator changes, even though the results of the simulation (outcome)
may be the same.
 
0.429
 
19. π ≈  0.333 
0.238
Therefore

Pr{S = 3} = π3 ≈ 0.238
E[S] = π1 + 2π2 + 3π3 ≈ 1.809
3

Var[S] = (j − 1.809)2 πj ≈ 0.630
j=1

A 1% relative error for Pr{S = 3} implies that

(n)
| 0.238 − p13 |
≤ 0.01
0.238
(n)
Therefore, | 0.238 − p13 |≤ 0.002.
A 1% relative error for E[S] implies that

| 1.809 − E[Sn | S0 = 1] |
≤ 0.01
1.809
Therefore, | 1.809 − E[Sn | S0 = 1] |≤ 0.018 where
3
 (n)
E[Sn | S0 = 1] = j p1j
j=1

A 1% relative error for Var[S] implies that

| 0.630 − Var[Sn | S0 = 1] |
≤ 0.01
0.063
126 CHAPTER 9. TOPICS IN SIMULATION OF STOCHASTIC PROCESSES

Therefore, | 0.630 − Var[Sn | S0 = 1] |≤ 0.006 where

3
 (n)
Var[Sn | S0 = 1] = (j − E[Sn | S0 = 1])2 p1j
j=1

(n)
n p13 E[Sn | S0 = 1] Var[Sn | S0 = 1]
0 0 1.0 0
1 0.2 1.5 0.65
2 0.25 1.67 0.721
3 0.258 1.735 0.711
4 0.256 1.764 0.691
5 0.251 1.780 0.674
6 0.248 1.789 0.662
7 0.245 1.796 0.652
8 0.243 1.800 0.646
9 0.241 1.803 0.641
10 0.240 1.805 0.638
11 .240 1.806 0.636

They do not all converge at the same time.

20. No answer provided.

21. No answer provided.