Queueing Theory with Applications . . .

:
Solution Manual
QUEUEING THEORY WITH APPLICA-
TIONS TO PACKET TELECOMMUNICA-
TION: SOLUTION MANUAL

JOHN N. DAIGLE
Prof. of Electrical Engineering
The University of Mississippi
University, MS 38677

Kluwer Academic Publishers

Boston/Dordrecht/London
Contents

List of Figures vii

List of Tables ix
1. TERMINOLOGY AND EXAMPLES 1
2. REVIEW OF RANDOM PROCESSES 9
3. ELEMENTARY CTMC-BASED QUEUEING MODELS 47
3.1 Supplementary Problems 82
4. ADVANCED CTMC-BASED QUEUEING MODELS 101
4.1 Supplementary Problems 133
5. THE BASIC M/G/1 QUEUEING SYSTEM 141
6. THE M/G/1 QUEUEING SYSTEM WITH PRIORITY 199
7. VECTOR MARKOV CHAINS ANALYSIS 217
7.1 Supplementary Problems 232
List of Figures

3.1 Time dependent state probabilities for Exercise 3.22 75

3.2 State diagram for Supplemental Exercise 3.2 85
3.3 State diagram for Supplemental Exercise 3.3 90
3.4 State diagram for Supplemental Exercise 3.4 94
3.5 State diagram for Supplemental Exercise 3.5 96
3.6 State diagram for Supplemental Exercise 3.6 98
4.1 Complementary distributions 111
4.2 Complementary distributions 113
4.3 State Diagram for Problem 4.27 129
4.4 State Diagram for Problem 2a 137
List of Tables

5.1 Occupancy values as a function of the number of units

served for the system of Example 5.8. 170
Chapter 1

TERMINOLOGY AND EXAMPLES

Exercise 1.1 Assume values of x̃ and t̃ are drawn from truncated geo-
metric distributions. In particular, let P {x̃ = n} = 0 and P {t̃ = n} = 0
except for 1 ≤ n ≤ 10, and let P {x̃ = n} = αpx (1 − px )n and
P {t̃ = n} = βpt (1 − pt )n for 1 ≤ n ≤ 10 with px = 0.292578 and
pt = 0.14358.
1. Using your favorite programming language or a spreadsheet, generate a
sequence of 100 random variates each for x̃ and t̃.
2. Plot ũ(t) as a function of t, compute z̃n from (1.3) and w̃n from (1.2) for
1 ≤ n ≤ 20.
3. Compute τ̃n for 1 ≤ n ≤ 20 and verify that w̃n can be obtained from
z̃(t).
4. Compute z̃n from (1.3) and w̃n from (1.2) for 1 ≤ n ≤ 100 and compute
the average waiting times for the 100 customers.

Solution Part 1. We must first determine the constants α and β. Since the
probabilities must sum to unity, we find
10
X 10
X 
P {x̃ = n} = P t̃ = n = 1.
n=1 n=1

Thus, using the formula for the partial sum of a geometric series, we find
h i h i
α (1 − px ) − (1 − px )11 = β (1 − pt ) − (1 − pt )11 = 1,

from which we may solve to find α = 1.45939316 and β = 1.48524892.

2 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

Now, in general,
Pn
P {x̃ ≤ n} = = i} = ni=1 αpx (1 − px )i
i=1 P {x̃
P

= α (1 − px ) − (1 − px )n+1 .

Solving P {x̃ = n} = αpx (1 − px )n for n, we find

ln [(1 − px ) − (1/α)P {x̃ ≤ n}]

n= − 1.
ln (1 − px )

We may therefore select random variates from a uniform distribution and use
these numbers in place of the above formula in place of P {x̃ ≤ n} and then
round up to obtain the variates for x̃. Similarly, by solving P {t̃ = n} =
βpt (1 − pt )n for n we may find the variates for t̃ from
  
ln (1 − pt ) − (1/β)P t̃ ≤ n
n= − 1.
ln (1 − pt )

By drawing samples
 from
a uniform distribution and using those values for
P {x̃ ≤ n} and P t̃ ≤ n in the above formulas, one can obtain the variates
for xn and tn for n = 1, 2, · · · , 100, and then apply the formulas to obtain the
following table:

n xn tn zn wn τn
1 3 6 2 0 6
2 5 1 -1 2 7
3 1 6 -4 1 13
4 1 5 -1 0 18
5 7 2 4 0 20
6 2 3 -1 4 23
7 2 3 -3 3 26
8 4 5 3 0 31
9 2 1 1 3 32
10 1 1 0 4 33
11 2 1 -1 4 34
12 1 3 -3 3 37
13 3 4 -3 0 41
14 4 6 2 0 47
15 7 2 3 2 49
16 2 4 -2 5 53
17 5 4 1 3 57
18 1 4 -6 4 61
19 1 7 -6 0 68
20 4 7 -1 0 75
Terminology and Examples 3

n xn tn zn wn τn
21 1 5 -6 0 80
22 5 7 -1 0 87
23 5 6 2 0 93
24 2 3 -8 2 96
25 4 10 -4 0 106
26 1 8 -1 0 114
27 1 2 -2 0 116
28 4 3 2 0 119
29 3 2 -3 2 121
30 1 6 0 0 127
31 2 1 -2 0 128
32 2 4 -6 0 132
33 1 8 -2 0 140
34 9 3 2 0 143
35 3 7 -6 2 150
36 2 9 -3 0 159
37 5 5 3 0 164
38 3 2 2 3 166
39 4 1 3 5 167
40 2 1 -3 8 168
41 6 5 3 5 173
42 3 3 0 8 176
43 2 3 -2 8 179
44 1 4 -9 6 183
45 5 10 -2 0 193
46 2 7 -8 0 200
47 1 10 -1 0 210
48 1 2 0 0 212
49 2 1 -1 0 213
50 1 3 -4 0 216
51 10 5 8 0 221
52 2 2 -2 8 223
53 3 4 1 6 227
54 1 2 -1 7 229
55 5 2 2 6 231
56 2 3 1 8 234
57 2 1 -1 9 235
58 1 3 -1 8 238
59 1 2 -1 7 240
60 5 2 1 6 242
61 2 4 0 7 246
62 3 2 1 7 248
63 1 2 -1 8 250
64 1 2 -4 7 252
65 5 5 -5 3 257
66 3 10 -2 0 267
67 2 5 0 0 272
68 2 2 -1 0 274
69 1 3 -2 0 277
70 4 3 -2 0 280
4 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

n xn tn zn wn τn
71 1 6 -2 0 286
72 4 3 -1 0 289
73 5 5 -5 0 294
74 1 10 -3 0 304
75 1 4 -9 0 308
76 1 10 0 0 318
77 4 1 -4 0 319
78 4 8 -4 0 327
79 5 8 -1 0 335
80 4 6 3 0 341
81 6 1 5 3 342
82 2 1 -1 8 343
83 3 3 2 7 346
84 1 1 -3 9 347
85 6 4 0 6 351
86 9 6 2 6 357
87 3 7 -3 8 364
88 7 6 3 5 370
89 3 4 2 8 374
90 4 1 2 10 375
91 2 2 -1 12 377
92 3 3 2 11 380
93 5 1 1 13 381
94 3 4 2 14 385
95 3 1 -4 16 386
96 8 7 4 12 393
97 2 4 1 16 397
98 1 1 0 17 398
99 4 1 -4 17 399
100 2 8 0 13 407

Averages 3.05 4.07 -0.98 3.95

Solution Part 2. Figure 2.1 shows the unfinished work as a function of

time. The values of zn and wn are shown in the above Table.
Solution Part 3. The values of τn and wn are shown in the above Ta-
ble.From Figure 2.1, it is readily verified that

wn = u(τn− ).

For example, τ3 = 13, and u(13− ) = 1. Therefore, w3 = 1. That is, if

an arrival occurs at time t, then an amount of time u(t− ) would pass before
service on that customer’s job began. That is, the server would first complete
the work present in the system at time t− , and then would begin the service on
the customer that arrived at time t.
Terminology and Examples 5

Solution Part 4. The average waiting time is shown at the bottom of the
above table. It is found by simply summing the waiting times and dividing by
the number of customers.

10

7.5
u(t)

2.5

0
0 20 40 60 80

time

Figure 2.1 Unfinished work as a function of time.

6 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

Exercise 1.2 Assume values of x̃ and t̃ are drawn from truncated geo-
metric distributions as given in Exercise 1.1.

1. Using the data obtained in Exercise 1.1, determine the lengths of all busy
periods that occur during the interval.
2. Determine the average length of the busy period.
3. Compare the average length of the busy period obtained in the previous
step to the average waiting time computed in Exercise 1.1. Based on the
results of this comparison, speculate about whether or not the average
length of the busy period and the average waiting time are related.

Solution Part 1. A busy period starts whenever a customer waits zero time.
The length of the busy period is the total service time accumulated up to the
point in time at which the next busy period starts, that is, whenever some later
arriving customer has a waiting time of zero. From the table of values given in
the previous problem’s solution, we see that a busy period is in progress at the
time of arrival of C1 00. Therefore we must decide whether or not to count that
busy period. We arbitrarily decide not to count that busy period. The lengths
of the busy periods are given in the following table:
1 9
2 1
3 11
4 10
5 3
6 19
7 1
8 4
9 1
10 5
11 7
12 4
13 1
14 1
15 7
16 1
17 2
18 2
19 1
20 12
21 2
22 26
23 5
24 2
25 1
Terminology and Examples 7

26 1
27 2
28 1
29 44
30 3
31 2
32 2
33 1
34 4
35 1
36 4
37 5
38 1
39 1
40 1
41 4
42 4
43 5

Average 5.209302326

Solution Part 2. The average length of the busy period is determined by

summing the lengths of the busy periods and dividing by the number of busy
periods. The result is shown at the end of the above table.
Solution Part 3. The computed averages are E [w̃] = 3.95 and E [ỹ] =
5.21. We see that based on this very small simulation, the results are not all
that different. It is sometimes possible to establish a definitive relationship
between the two quantities. We will see later on that under certain sets of
assumptions, the busy period and the waiting time have the same value.
Exercise 1.3 For the general case where the packet arrival process has
run lengths, it is seen that the survivor functions decrease with decreasing
run lengths. Determine whether or not there exists an average run length
at which the packet arrival process becomes a sequence of independent
Bernoulli trials. If such a choice is possible, find the value of run length
at which independence occurs. Discuss the result of reducing the run length
below that point. [Hint: A packet arrival occurs whenever the system tran-
sitions into the on state. Thus, if the probability of transitioning into the
on state is independent of the current state, the arrival process becomes a
sequence of independent Bernoulli trials.]
Solution. The arrival process will be Bernoulli if the probability of having a
packet in a slot is independent of the current state. In general, there will be
a packet arrival in a slot if the system is in the on state. Hence we want to
set transition probabilities p01 and p11 equal. Since we want to have a 90%
traffic intensity, we choose p01 = p11 = 0.9/N . Since the run lengths have the
geometric distribution, the average run length will be 1/p10 . Thus, the average
8 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

run length will be

N
N − 0.9.
Similarly the average off time will be
N
0.9.
Chapter 2

REVIEW OF RANDOM PROCESSES

Exercise 2.1 For the experiment described in Example 2.1, specify all
of the event probabilities.
Solution. For the experiment described in Example 2.1, the events are as fol-
lows: ∅, {N }, {S}, {C}, {T }, {S, N }, {C, N }, {T, N }, {C, S}, {T, S},
{T, C}, {C, S, N }, {T, S, N }, {T, C, N }, {T, C, S}, {T, C, S, N }. Since
each of the events is a union of elementary events, which are mutually ex-
clusive, the probability of each event is the sum of the probabilities of its con-
stituent events. For example,
P {T, C, S, N } = P {T }+P {C}+P {S}+P {N } = 0.11+0.66+0.19+0.04 = 1.0.
By proceeding in this manner, we find the results given in the following table:

Event Probability Event Probability

∅ 0.00 {C, S} 0.85
{N } 0.04 {T, S} 0.30
{S} 0.19 {T, C} 0.77
{C} 0.66 {C, S, N } 0.89
{T } 0.11 {T, S, N } 0.34
{S, N } 0.23 {T, C, N } 0.81
{C, N } 0.70 {T, C, S} 0.96
{T, N } 0.15 {T, C, S, N } 1.00

Exercise 2.2 For the experiment described in Example 2.1, there are a
total of four possible outcomes and the number of events is 16. Show that it
is always true that card (Ω) = 2card (S) , where card (A) denotes the car-
dinality of the set A, which, in turn, is the number of elements of A. [Hint:
The events can be put in one-to-one correspondence with the card (S)-bit
binary numbers.]
10 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

Solution. Every event in the event space consists of a union of the elementary
events of the experiment. Thus, every event is characterized by the presence or
absence of elementary events. Suppose an experiment has card (S) elemen-
tary events, then these events can be ordered and put in one-to-one correspon-
dence with the digits of an card (S)-digit binary number, say,

dcard (S)−1 dcard (S)−2 · · · d0 .

If elementary event i is present in some event, the di = 1, else di = 0. Since

there are a total of 2N N -digit binary numbers, there are exactly 2card (S)
events in any experiment. For example, in the experiment of Eample 2.1, we
have four events. Let T , C, S, and N be represented by the binary digits d3 ,
d2 , d1 , and d0 , respectively. Then, 1010 represents the event {T, S}. The total
number of events is 24 = 16.
Exercise 2.3 Repeat the computations of Example 2.2 by constructing
restricted experiments based on SA and SB . [Hint: The probabilities of the
elementary events in the restricted experiments must be normalized by di-
viding the probabilities of the individual elementary events of the restricted
experiment by the sum of the probabilities of the constituent elementary
events of the restricted experiment.]
Solution. The idea is to construct two different experiments, A and B, and
compute P {ω1 } in each of the experiments. The set of all possible outcomes
in experiment A is SA = {s1 , s2 , s4 , s6 , s8 }.
For experiment A, we compute the probabilities of each elementary event to
i
be P {si } = kA 36 , where kA is a normalizing parameter such that
X
P {si } = 1.
si ∈SA

We find KA = 3621 so that P {si } =

i
21 for si ∈ SA and P {si = 0} otherwise.
We then find for experiment A,

1 2 1
P {ω1 } = P {s1 } + P {s2 } + P {s3 } = + +0 = .
21 21 7
i
Similarly, for experiment B, we find P {si } = 15 for si ∈ SB and si = 0
otherwise. Then, for experiment B, we have

0 0 3 1
P {ω1 } = P {s1 } + P {s2 } + P {s3 } = + + = .
15 15 15 5
We recognize these probabilities to be the same as the conditional probabil-
ities previously computed.
Review of Random Processes 11

Exercise 2.4 Suppose ω1 , ω2 ∈ Ω with P {ω1 } = 6 0 and P {ω2 } = 6 0.

Show that if ω1 is statistically independent of ω2 , then necessarily, ω2 is
statistically independent of ω1 . In other words, show that statistical inde-
pendence between events is a mutual property.
Solution. From Bayes’ rule, we have

P {ω1 |ω2 } P {ω2 }

P {ω2 |ω1 } = .
P {ω1 }

If ω1 is statistically independent of ω2 , then P {ω1 |ω2 } = P {ω1 }. Thus,

P {ω1 } P {ω2 }
P {ω2 |ω1 } = = P {ω2 } .
P {ω1 }

Since
P {ω2 |ω1 } = P {ω2 }
is the definition of ω2 being statistically independent of ω1 , the result follows.
Exercise 2.5 Suppose ω1 , ω2 ∈ Ω but P {ω1 } = 0 or P {ω2 } = 0.
Discuss the concept of independence between ω1 and ω2 .
Solution. We assume ω1 and ω2 are two different events. Suppose P {ω1 } = 0
but P {ω2 } =
6 0. Then

P {ω1 ω2 }
P {ω1 |ω2 } = .
P {ω2 }

But, P {ω1 ω2 } ≤ P {ω1 } and P {ω1 ω2 } ≥ 0. Thus, P {ω1 ω2 } = 0, which

implies P {ω1 |ω2 } = 0, and thus, ω1 is independent of ω2 . If P {ω1 } = 0,
then ω1 is equivalent to the null event. Therefore, the probability that any
event has occurred given that the null event has actually occurred is zero; that
is, all subsets of the null event are the null event. Hence no event other than
the null event can be statistically independent of the null event. In words,
if we know the null event has occurred, then we are certain no other event
has occurred. Since the other event is not the null event, its unconditional
probability of occurrence is greater than zero.
Exercise 2.6 Develop the expression for dFc̃ (x) for the random variable
c̃ defined in Example 2.4.
Solution. The random variable c̃ defined in Example 2.4 is discrete, and its
support set is

C = {56.48, 70.07, 75.29, 84.35, 96.75, 99.99, 104.54, 132.22} .

12 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

We then find,
X
dFc̃ (x) = P {c̃ = c} δ(x − c)dx
c∈C
X
= P {c̃ = c} δ(x − c)dx.
c∈{56.48,70.07,75.29,84.35,96.75,99.99,104.54,132.22}
(2.1)

Thus,
8 7 6
dFc̃ (x) = δ(x − 56.48)dx + δ(x70.07)dx + δ(x − 75.29)dx +
36 36 36
5 4 3
δ(x − 84.35)dx + δ(x − 96.75)dx + δ(x − 99.99)dx +
36 36 36
2 1
δ(x − 104.54)dx + δ(x − 132.22)dx
36 36

Exercise 2.7 Find E[c̃] and var (c̃) for the random variable c̃ defined in
Example 2.4.
Solution. By definition,
X
E [c̃] = cP {c̃ = c} ,
c∈C
h i
var (c̃) = E c̃2 − E 2 [c̃],
and h i
E c̃2 = c2 P {c̃ = c} .
X

c∈C

From Example 2.4 and the previous problem, we have the following table:

c P {c̃ = c}
56.48 0.222222222
70.07 0.194444444
75.29 0.166666667
84.35 0.138888889
96.75 0.111111111
99.99 0.083333333
104.54 0.055555556
132.22 0.027777778

Upon performing the calculations, we find E [c̃] = 79.0025, E c̃2

 
=
6562.507681, and var (c̃) = 321.1126743.
Review of Random Processes 13

Exercise 2.8 Let X1 and X2 denote the support sets for x̃1 and x̃2 ,
respectively. Specialize (2.12) where X1 = {5, 6, . . . , 14} and X2 =
{11, 12, . . . , 22}.
Solution. Equation (2.12) is as follows:
n
X
P {x̃1 + x̃2 = n} = P {x̃2 = n − i | x̃1 = i} P {x̃1 = i} .
i=0

First we note that the smallest possible value of x̃1 + x̃2 is 5 + 11 = 16, and
the largest possible value x̃1 + x̃2 can take on is 14 + 22 = 36. We also see that
x̃1 + x̃2 can take on all values between 16 and 36. Thus, we find immediately
that the support set for x̃1 + x̃2 is {16, 17, . . . , 36}. We also see that since x̃1
ranges only over the integers from 5 to 14, P {x̃1 = i} > 0 only if i is over
only the integers from 5 to 14.
Similarly, probx̃2 = n − i > 0 only if n − i ≥ 11 and n − i ≤ 22, so that
probx̃2 = n − i > 0 only if n − 11 ≥ i and n − 22 ≤ i or i ≤ n − 11 and
i ≥ n − 22. Hence, for P {x̃1 = i} > 0, we need i ≥ max {5, n − 22}
and i ≤ min set14, n − 11. We, therefore have the following result. For
n 6= {5, 6, . . . , 36}, P {x̃1 + x̃2 = n} = 0, and for n ∈ {5, 6, . . . , 36},

min {14,n−11}
X
P {x̃1 + x̃2 = n} = P {x̃2 = n − i | x̃1 = i} P {x̃1 = i} .
i=max {5,n−22}

For example,

min {14,18−11}
X
P {x̃1 + x̃2 = 18} = P {x̃2 = n − i | x̃1 = i} P {x̃1 = i}
i=max {5,18−22}
7
X
= P {x̃2 = n − i | x̃1 = i} .
i=5

Exercise 2.9 Derive (2.14) as indicated in the previous paragraph. .

Solution. We wish to find an expression for fx̃1 +x̃2 (x) given the densities
d
fx̃1 (x) and fx̃2 (x). We first note that fx̃1 +x̃2 (x) = dx Fx̃1 +x̃2 (x)(x). But,
Fx̃1 +x̃2 (x)(x) = P {x̃1 + x̃2 ≤ x}. Now, suppose we partition the interval
(0, x] into N equal-length intervals an assume that the random variable x̃1 has
masses at the points i∆X, where ∆X = Nx . Then,
N
X
P {x̃1 + x̃2 ≤ x} = P {x̃1 + x̃2 ≤ x|x̃1 = i∆x} P {x̃1 = i∆X} .
i=1
14 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

But
P {x̃1 = i∆X} ≈ fx̃1 (i∆X)∆X,
so,
N
X
P {x̃1 + x̃2 ≤ x} = P {x̃1 + x̃2 ≤ x|x̃1 = i∆x} fx̃1 ∆X
i=1
XN
= P {x̃2 ≤ x − i∆X|x̃1 = i∆x} fx̃1 (i∆X)∆X.
i=1

In the limit, we can pass i∆X to the continuous variable y and ∆X to dy. The
sum then passes to the integral over (0, x], which yields the result
Z x
P {x̃1 + x̃2 ≤ x} = P {x̃2 ≤ x − y|x̃1 = y} fx̃1 (y)dy,
0
or Z x
Fx̃1 +x̃2 (x) = Fx̃2 |x̃1 (x − y|y)fx̃1 (y)dy.
0
Differentiating both sides with respect to x then yields
Z x
fx̃1 +x̃2 (x) = fx̃2 |x̃1 (x − y|y)fx̃1 (y)dy.
0

Exercise 2.10 Prove Lemma 2.1.

Solution. Recall the definition of memoryless: A random variable x̃ is said to
be memoryless if, and only if, for every α, β ≥ 0,

P {x̃ > α + β|x̃ > β} = P {x̃ > α}.

In order to prove the lemma, we simply show that the exponential distribution
satisfies the definition of memoryless; that is, we compute P {x̃ > α + β|x̃ >
β} and see whether this quantity is equal to P {x̃ > α}. From the definition of
conditional probability, we find
P {x̃ > α + β, x̃ > β}
P {x̃ > α + β|x̃ > β} = .
P {x̃ > β}
But the joint probability of the numerator is just P {x̃ > α + β}, and for the
exponential distribution is

P {x̃ > α + β} = e−λ(α+β) .

Also,
P {x̃ > β} = e−λβ .
Review of Random Processes 15

Thus,
e−λ(α+β)
P {x̃ > α + β|x̃ > β} = = e−λα P {x̃ > α}.
e−λβ
Thus, the definition of memoryless is satisfied and the conclusion follows.
Exercise 2.11 Prove Lemma 2.2. [Hint: Start with rational arguments.
Extend to the real line using a continuity argument. The proof is given
in Feller, [1968] pp. 458 - 460, but it is strongly recommended that the
exercise be attempted without going to the reference.]
Solution. We first prove the proposition that g(n/m) = g(1/m)n for an arbi-
trary integer m. Let T denote the truth set for the proposition. Trivially, n ∈ T ,
so assume (n − 1) ∈ T . Then
   
n n−1
X 1 X 1 1
g(n/m) = g   = g + .
j=1
m j=1
m m

But, by hypothesis of the Lemma 2.2, g(t + s) = g(t)g(s) ∀s, t > 0. Thus,
 
n−1
1 1
X  
g(n/m) = g  g ,
j=1
m m

and by using the inductive hypothesis, we have

n−1  n
1 1 1
  
g(n/m) = g g =g for all integer m, n.
m m m
This completes the proof of the proposition. Now, with m = 1, the result of
the proposition gives g(n) = g(1)n , and with n = m, the same result gives
g(1) = g(1/m)m from which it follows that g(1/m) = g(1)1/m . Substitution
of the latter result into the conclusion of the proposition yields
n−1 
1 1
 
g(n/m) = g g = g(1)n/m for all integer m, n.
m m
If we now take λ = − ln [g(1)], we have

g(n/m) = e−λ(n/m) for all integer m, n.

Since the rationals are dense in the reals, each real number is the limit of a
sequence of rational numbers. Further, g is a right continuous function so that
by analytic continuity,

g(t) = e−λt for all integer t > 0.

16 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

Exercise 2.12 Repeat Example 2.5, assuming all students have a deter-
ministic holding time of one unit. How do the results compare? Would
an exponential assumption on service-time give an adequate explanation of
system performance if the service-time is really deterministic?
Solution. P {Alice before Charlie} must be zero. We see this as follows:
Alice, Bob and Charlie will use the phone exactly one time unit each. Even
if Bob finishes his call at precisely the moment Alice walks up, Charlie has
already begun his. So, Charlie’s time remaining is strictly less than one, while
Alice’s time to completion is exactly one in the best case. Thus, it is impossible
for Alice to finish her call before Charlie. Since assuming exponential hold-
ing times results in P {Alice before Charlie} = 1/4, exponentiality is not an
appropriate assumption in this case.
Exercise 2.13 Prove Theorem 2.2. Theorem2.2. Let x̃ be a non-
negative random variable with distribution Fx̃ (x), and let Fx̃∗ (s) be the
Laplace-Stieltjes transform of x̃. Then
dn ∗ 

2 n
E[x̃ ] = (−1) F (s) .
dsn x̃ s=0

Solution. We first prove the proposition that

dn ∗ ∞
Z
F (s) = (−1)n xn e−sx dFx̃ (x).
dsn x̃ 0

Let T be the truth set for the proposition. By Definition 2.8, Fx̃∗ (s) = E[e−sx ],
so that Z ∞
d ∗
F (s) = (−x)e−sx dFx̃ (x),
ds x̃ 0
and so 1 ∈ T . Now assume that n ∈ T . Then
dn+1 ∗ d dn ∗
 
F (s) = F (s)
dsn+1 x̃ ds dsn x̃
Z ∞
d
 
n n −sx
= (−1) x e dFx̃ (x)
ds 0
Z ∞
n+1
= (−1) xn+1 e−sx dFx̃ (x).
0
(2.2)

This proves the proposition. We may now show

dn ∗ 
 
n R ∞ n −sx 
(−1) Fx̃ (s) = 0 x e dFx̃ (x)
dsn 
s=0

s=0
Review of Random Processes 17
R∞
= 0 xn dFx̃ (s)
= E[x̃n ].
(2.3)

Exercise 2.14 Prove Theorem 2.3. Let x̃ and ỹ be independent, non-

negative, random variables having Laplace-Stieltjes transforms Fx̃∗ (s) and
Fỹ∗ (s), respectively. Then, the Laplace-Stieltjes transform for the random
variable z̃ = x̃ + ỹ is given by the product of Fx̃∗ (s) and Fỹ∗ (s).
Solution. By the definition of Laplace-Stieltjes transform,
Fz̃∗ (s) = E[e−sz̃ ] = E[e−s(x̃+ỹ) ]
= E[e−sx̃ e−sỹ ].
Now, because x̃ and ỹ are independent, h̃1 (x̃) and h̃2 (ỹ) are also indepen-
dent for arbitrary functions h̃1 and h̃2 . In particular, e−sx̃ and e−sỹ are inde-
pendent. Since the expectation of the product of independent random variables
is equal to the product of the expectations,
E[e−sx̃ e−sỹ ] = E[e−sx̃ ]E[e−sỹ ].
That is,
Fz̃∗ (s) = Fx̃∗ (s)Fỹ∗ (s).

Exercise 2.15 Let x̃ be an exponentially distributed random variable

with parameter λ. Find Fx̃∗ (s).
Solution. By definition,
h i Z ∞
Fx̃∗ (s) = E e−sx̃ = e−sx dFx̃ (x).
0

For x̃ exponential, dFx̃ (x) = λe−λx dx, so

Z ∞
Fx̃∗ (s) = e−sx λe−λx dx
0
Z ∞
= λe−(s+λ)x dx
0
λ
Z ∞
= (s + λ)e−(s+λ)x dx
s+λ 0
λ
=
s+λ

Exercise 2.16 Let x̃ be an exponentially distributed random variable

with parameter λ. Derive expressions for E[x̃], E[x̃2 ], and V ar(x̃). [Hint:
Use Laplace transforms.]
18 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

Solution. By Theorem 2.2,

dn ∗ 
E[x̃n ] = (−1)n F (s) ,
dsn x̃ s=0

λ
where by Exercise 2.5, Fx̃∗ (s) = λ+s . Then

d λ
 

E[x̃] = (−1) 
ds λ + s s=0
"  −2 # 
1  1
= (−1) λ  = .
λ+s 
s=0 λ
2
  
E[x̃2 ] = (−1)2 ds
d
2
λ
λ+s 

s=0
" −3 # 
1 2


= 2λ  = .
λ+s 
s=0 λ2

We may now compute Var(x̃) as follows:

Var(x̃) = E[x̃2 ] − E 2 [x̃]

 2
2 1
= 2
−
λ λ
= λ1 .

Note that for the exponential random variable, the mean and variance are equal.
Exercise 2.17 Let x̃ and ỹ be independent exponentially distributed ran-
dom variables with parameters α and β, respectively.

1. Find the distribution of z̃ = min{x̃, ỹ}. [Hint: Note that z̃ = min{x̃, ỹ}
and z̃ > z means x̃ > z and ỹ > z.]
2. Find P {x̃ < ỹ}.
3. Show that the conditional distribution Fz̃|x̃<ỹ (z) = Fz̃ (z).

Solution.
1. First, note that P {z̃ ≤ z} = 1 − P {z̃ > z}. As noted in the hint, z̃ =
min{x̃, ỹ} and z̃ > z means x̃ > z and ỹ > z. Therefore,

P {z̃ > z} = P {x̃ > z, ỹ > z} .

Since x̃ and ỹ are independent,

P {x̃ > z, ỹ > z} = P {x̃ > z} P {ỹ > z}

Review of Random Processes 19

= e−αx e−βx = e−(α+β)x .

Thus,
P {z̃ ≤ z} = 1 − e−(α+β)x ,
so that z̃ is exponentially distributed with parameter λ + α.
2. We compute the required probability by conditioning on the value of x̃ as
follows:
R∞
P {x̃ < ỹ} = 0 RP {x̃ < ỹ|x̃ = x} dFx̃ (x)
= 0∞R P {x < ỹ} dFx̃ (x)
= 0∞ e−βx αe−αx dx
α R∞ −(α+β)x
= α+β 0 (α + β)e dx
α
= α+β .

3. We begin with the definition of Fz̃|x̃<ỹ (z). We have

Fz̃|x̃<ỹ (z) = P {z̃ ≤ z|x̃ < ỹ} .
Thus,
Fz̃|x̃<ỹ (z) = 1 − P {z̃ > z|x̃ < ỹ}
P {z̃>z,x̃<ỹ}
=1− P {x̃<ỹ} .

But, since z̃ = min{x̃, ỹ}

P {z̃ > z, x̃ < ỹ} R ∞= P {x̃ > z, x̃ < ỹ}
= z RP {x̃ < ỹ|x̃ = x} dFx̃ (x)
= z∞R∞P {x < ỹ} dFx̃ (x)
= z e−βx αe−αx dx
α R∞ −(α+β)x
= α+β z (α + β)e dx
α
= α+β e−(α+β)z .
Thus, upon using th eresult of part (ii), we find
P {z̃ > z|x̃ < ỹ} = e−(α+β)z
so that
Fz̃|x̃<ỹ (z) = 1 − e−(α+β)z = Fz̃ (z),
and the proof is complete.
Exercise 2.18 Suppose Albert and Betsy run a race repeatedly. The time
required for Albert to complete the race, ã, is exponentially distributed with
parameter α and the time required for Betsy to complete, b̃, is exponentially
distributed with parameter β. Let ñb denote the number of times Betsy wins
before Albert wins his first race. Find P {ñb = n} for n ≥ 0.
20 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

Solution. First consider P {ñb = 0}, the probability that Albert wins the very
first race. Since nAlbert’so time must be less than Betsy’s
n forohim to win, this
α
probability is P ã < b̃ . By Exercise 2.7 (iii), P ã < b̃ = α+β . Now
consider P {ñb = 1}, the probability that Betsy wins one race and then Albert
wins the second. Both ã and b̃ are exponential random variables; hence they
are memoryless. That is, during the second race they don’t ‘remember’ what
happened during the first race. Thus the times for each race for each runner
are independent of each other. Let ãi , b̃i denote the races times of Albert and
Betsy, respectively, on the ith race. Then
n o
P {ñb = 1} = P ã1 > b̃1 , ã2 < b̃2
n o n o
= P ã1 > b̃1 P ã2 < b̃2
β α
= α+β · α+β .

Repeating this argument for arbitrary n, we see that

n o
P {ñb = n} = P ã1 > b̃1 , ã2 > b̃2 , · · · , ãn > b̃n , ãn+1 < b̃n+1
n o n o n o n o
= P ã1 > b̃1 P ã2 > b̃2 · · · P ãn > b̃n P ãn+1 < b̃n+1
 n
β α
= α+β α+β .
α
That is, ñb is geometric with parameter α+β .
Exercise 2.19 Let {x̃i , i = 1, 2, . . .} be a sequence of exponentially dis-
tributed random variables and let ñ be a geometrically distributed random
variable with parameter p, independent of {x̃i , i = 1, 2, . . .}. Let
ñ
X
ỹ = x̃i .
i=1

Show that ỹ has the exponential distribution with parameter pα.

Solution. First we prove a lemma. Lemma: If x̃i , i = 1, 2, ·P · · are as in the
statement of the exercise and n is a f ixed integer, then ỹ = ni=1 x̃i has the
gamma distribution with parameter (n, α).
Proof: If n = 1, the result holds trivially. So suppose the result holds for
n = N − 1, N ≥ 2. If fPN−1 x̃ (y) is the probability density function of ỹ,
i=1 i
then
(αy)N −2
fPN−1 x̃ (y) = αe−αy .
i=1 i (N − 2)!
PN PN −1
Using the fact that i=1 x̃i = i=1 +x̃N leads to
R∞
fPN x̃i
(y) = 0 fx̃N (y − t)fPN−1 x̃ (t)dt
i=1 i=1 i
Review of Random Processes 21
N−2
αe−α(y−t) αe−αt (αt)
R∞
= 0 (N −2)! dt
N−1
= αe−αy (αy)
(N −1)! .

This proves the Lemma. To complete the exercise, we condition the probability
density function of ỹ on the value of ñ. Recall that here ñ is a random variable
and is independent of x̃i , i = 1, 2, · · ·. Thus,

fỹ (y) = fPñ x̃i

(y)
P∞ i=1
= n=1 x̃i (y)P {ñ = n}
fPn
i=1
N−1
= n=1 αe−αy (αy)
P∞ n−1
(N −1)! p(1 − p)
(αy(1−p))n
= pαe−αy ∞
P
n=0 n!
= pαe−αy eαy(1−p) = pαe−pαy .

Exercise 2.20 This exercise is intended to reinforce the meaning of

Property 2 of exponential random variables. Let x̃ and ỹ denote the two
independent exponential random variables with rates 1 and 2, respectively,
and define z̃ = min{x̃, ỹ}. Using a spreadsheet (or a computer program-
ming language), generate a sequence of 100 variables for each of the ran-
dom variables. Denote the ith variate for x̃ and ỹ by xi and yi , respectively,
and set z̃ = min{x̃, ỹ} for i = 1, 2, . . . , 100.
Let n denote the number of values of i such that xi < yi , let ij denote
the jthe such value and define wj = zi , for j = 1, 2, . . . , n. Compute
P 100
the
sample averages for
P100
the variates; that is
P 100
compute x̄ = (1/100) i=1 xi ,
P 100
ȳ = (1/100) i=1 yi , z̄ = (1/100) i=1 zi , and w̄ = (1/100) j=1 wj .
Compare the results. Is w̄ closer to x̄ or z̄?
Now give an intuitive explanation for the statement, “It is tempting to con-
clude that if one knows the state change was caused by the event having
its interevent time drawn from the distribution Fx̃ (x), then the time to state
change is exponentially distributed with parameter α, but this is false.”
Solution. The following sequences for x̃, ỹ, z̃, and w̃ were generated using a
spreadsheet package:
22 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

n xn yn zn wn

1 1.84397193 0.16714174 0.16714174

2 1.24667042 0.18449179 0.18449179
3 2.2392642 0.07468417 0.07468417
4 0.7575555 0.08376639 0.08376639
5 1.42970352 0.39268839 0.39268839
6 4.01878152 1.34361333 1.34361333
7 3.43923471 0.08348682 0.08348682
8 1.1702069 0.64094628 0.64094628
9 2.46251518 1.1929263 1.1929263
10 2.89695466 0.74196329 0.74196329
11 0.65195416 0.11265824 0.11265824
12 0.38827964 0.77950928 0.38827964 0.38827964
13 0.13928522 0.01412904 0.01412904
14 0.56052792 0.31830474 0.31830474
15 1.98337703 1.1304179 1.1304179
16 0.58175933 0.61897506 0.58175933 0.58175933
17 0.4729621 0.32353766 0.32353766
18 1.18397153 0.1701992 0.1701992
19 0.27687641 0.15029115 0.15029115
20 0.04403085 0.21006062 0.04403085 0.04403085
21 0.1339849 0.25982677 0.1339849 0.1339849
22 0.93645887 0.2989296 0.2989296
23 0.58498602 0.2435585 0.2435585
24 1.20431873 0.05653863 0.05653863
25 1.28216752 2.32710226 1.28216752 1.28216752
26 0.36686726 0.08247806 0.08247806
27 0.73515498 1.20688382 0.73515498 0.73515498
28 0.29300628 0.21725415 0.21725415
29 1.41245136 0.06286597 0.06286597
30 1.43283182 1.14791652 1.14791652
31 0.45688209 0.07873111 0.07873111
32 3.43212736 0.7438511 0.7438511
33 6.08640858 0.53356533 0.53356533
34 2.75348525 0.63894337 0.63894337
35 0.63137432 0.19140255 0.19140255
36 0.17315955 0.64144138 0.17315955 0.17315955
37 0.46050301 0.53641231 0.46050301 0.46050301
38 0.54172938 0.27631063 0.27631063
39 1.81441406 0.55841417 0.55841417
40 0.95500193 0.83148476 0.83148476
Review of Random Processes 23
n xn yn zn wn

41 0.8008909 0.02778907 0.02778907

42 0.02385716 0.67811659 0.02385716 0.02385716
43 0.4176623 0.06398327 0.06398327
44 0.44562071 1.30356611 0.44562071 0.44562071
45 0.93021164 0.58957361 0.58957361
46 0.28880168 0.13168134 0.13168134
47 2.15051176 0.42603754 0.42603754
48 0.38888728 0.33542262 0.33542262
49 0.04647351 0.13146297 0.04647351 0.04647351
50 0.24678196 1.08658913 0.24678196 0.24678196
51 1.29128385 0.12054774 0.12054774
52 1.19616038 0.58372847 0.58372847
53 0.70889554 0.50033913 0.50033913
54 0.03148097 0.19271322 0.03148097 0.03148097
55 2.00060148 0.73581018 0.73581018
56 2.37414692 0.05800158 0.05800158
57 0.5614103 1.02321168 0.5614103 0.5614103
58 0.28880168 0.48428777 0.28880168 0.28880168
59 2.10829625 0.55708733 0.55708733
60 0.00647539 0.28160211 0.00647539 0.00647539
61 1.21198004 0.42532257 0.42532257
62 0.25354251 0.16509961 0.16509961
63 0.25010753 0.31332586 0.25010753 0.25010753
64 0.42281911 1.21254329 0.42281911 0.42281911
65 0.43420203 0.08457892 0.08457892
66 2.10942768 0.48084293 0.48084293
67 0.57224943 0.73587666 0.57224943 0.57224943
68 1.4642691 0.36727522 0.36727522
69 0.92824082 0.34664989 0.34664989
70 0.36374508 0.83064034 0.36374508 0.36374508
71 0.26106201 0.31306894 0.26106201 0.26106201
72 1.86907458 0.1030878 0.1030878
73 0.56825514 0.27826342 0.27826342
74 1.15881452 0.23169248 0.23169248
75 0.01464839 0.00071004 0.00071004
76 0.52810458 0.05496071 0.05496071
77 0.87417563 0.43087249 0.43087249
78 0.03337237 0.20835641 0.03337237 0.03337237
79 1.75585187 0.85560022 0.85560022
80 0.79283399 1.17958999 0.79283399 0.79283399
24 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

n xn yn zn wn

81 0.07335002 2.02886531 0.07335002 0.07335002

82 2.25514443 0.58648222 0.58648222
83 1.81272977 0.14403435 0.14403435
84 0.08817172 0.78172674 0.08817172 0.08817172
85 0.66163598 0.30661797 0.30661797
86 0.74250208 0.40806585 0.40806585
87 0.44464432 0.08091491 0.08091491
88 2.93930994 0.28810271 0.28810271
89 0.06949872 0.08515753 0.06949872 0.06949872
90 0.62340183 0.10592748 0.10592748
91 0.9035322 0.9430666 0.9035322 0.9035322
92 0.13090162 0.06881894 0.06881894
93 0.27739983 0.18747991 0.18747991
94 0.68718355 0.23571032 0.23571032
95 0.37903079 0.38083248 0.37903079 0.37903079
96 0.25523489 0.05314163 0.05314163
97 0.26009176 0.68434451 0.26009176 0.26009176
98 0.41854311 0.90496445 0.41854311 0.41854311
99 0.00704379 0.21641837 0.00704379 0.00704379
100 0.14198992 0.01055528 0.01055528

98.8205643 46.0988392 33.8636937 10.3453931

Mean of xi ′ s Mean of yi ′ s Mean of zi ′ s Mean of wi ′ s

0.98820564 0.46098839 0.33863694 0.33372236

Solution. Observe that the sampled means of x̃, ỹ, and z̃ are close to their
expected values: 0.98820564 as compared to the expected value of 1.0 for x̃;
0.46098839 as compared to the expected value of 0.5 for ỹ; and 0.33863694
for z̃. (Recall that if α and β are the rates for x̃ and ỹ, respectively, then
1
E[z̃] = α+β . In this exercise α = 1 and β = 2; hence E[z̃] = 0.33333333.)
Note that the sampled mean of w̃ is very close to the sampled mean of z̃.
Now, the wi are selected from the samples of z̃. Once the samples of z̃ are
chosen, it does not matter whether the original variates came from sampling the
distribution of x̃ or of ỹ. Hence the same result would hold if w̃ represented
those yi that were less than their corresponding xi values. The distribution of
w̃, and therefore of z̃, does not depend on whether α < β or β < α. Thus z̃ is
not a representation of x̃, and so will not be exponential with rate α.
Review of Random Processes 25

Exercise 2.21 Define counting processes which you think have the fol-
lowing properties:
1. independent but not stationary increments,
2. stationary but not independent increments,
3. neither stationary nor independent increments, and
4. both stationary and independent increments.
What would you think would be the properties of the process which counts
the number of passengers which arrive to an airport by June 30 of a given
year if time zero is defined to be midnight, December 31 of the previous
year?
Solution.
1. Independent, not stationary: Count the number of customers who enter a
store over an hour-long period. Assume the store has infinte capacity. The
number of customers that enter during one 1-hour time interval likely will
be independent of the number that enter during a non-overlapping 1-hour
period. However, because of “peak” hours and slow times, this counting
process is not likely to have stationary increments.

2. Stationary, not independent: Count the number of failures of a light bulb in

any small time interval. Let the failure rate be exponential, so it is memory-
less. Hence the probability it fails in one fixed-length interval is the same as
in any other. This process does not have independent increments, however.
If the light burns out in one increment, then there is a 100% probability it
will not burn out in the next, since it can have only one failure.

3. Neither stationary nor independent: Count the Earth’s population. Certainly

this is not stationary; it is generally agreed that the birth rate increases with
time. Nor is this process independent. The birth of every individual alive
today was dependent upon the births of his or her parents. And their births
were dependent upon the births of their parents, and so on.

4. Both stationary and independent: Count the number of phone calls through
an operator’s switchboard between 3:00 am and 5:00 am in Watseka, Illi-
nois (a small farming community). Then the process can be modeled using
stationary and independent increments.

For the final part of this exercise, assume the airport has an infinite airplane
capacity. Then this process would likely have independent increments. Pas-
sengers usually will not be interested in how many people arrived before them;
26 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

they are concerned in their own, current, flight. Furthermore, this counting
process is not likely to have stationary increments since there will be peak and
slow times for flying. (For example: “red eye” flights are during slow times.)
Exercise 2.22 Show that E[ñ(t + s) − ñ(s)] = λt if {ñ(t), t > 0} is a
Poisson process with rate λ.
Solution.
∞
X
E [ñ(t + s) − ñ(s)] = nP {ñ(t + s) − ñ(s) = n}
n=0
∞
" #
X (λt)n e−λt
= n
n=0
n!
∞
" #
X (λt)n e−λt
= n
n=1
n!
∞
" #
X (λt)n e−λt
= n
n=1
n − 1!
∞
−λt
X (λt)n−1
= (λt) e
n − 1!
n=1
∞
(λt)n
= (λt) e−λt
X

n=0
n!
= (λt) e−λt eλt
= λt
Exercise 2.23 For each of the following functions, determine whether
the function is o(h) or not. Your determination should be in the form of a
formal proof.
1. f (t) = t,
2. f (t) = t2 ,
1
3. f (t) = t 2 ,
4. f (t) = e−at for a, t > 0
5. f (t) = te−at for a, t > 0

Solution.
h
1. limh→0 h = limh→0 1 = 1. Therefore f (t) = t is not o(h).
h2
2. limh→0 h = limh→0 h = 0. Therefore f (t) = t2 is o(h).
1 1 1
h2
3. limh→0 h = limh→0 h− 2 = ∞. Therefore f (t) = t 2 is not o(h).
Review of Random Processes 27
−ah −ah
4. limh→0 e h = limh→0 e h = ∞, since e−ah → 1 as h → 0. Therefore
f (t) = e−at for a, t > 0 is not o(h).
−ah
5. limh→0 he h = limh→0 e−ah = e0 = 1, Therefore f (t) = te−at for
a, t > 0 is not o(h).

Exercise 2.24 Suppose that f (t) and g(t) are both o(h). Determine
whether each of the following functions is o(h).
1. s(t) = f (t) + g(t)
2. d(t) = f (t) − g(t)
3. p(t) = f (t)g(t)
4. q(t) = f (t)/g(t)
Rt
5. i(t) = 0 f (x) dx

Solution.
1. s(t) = f (t) + g(t) is o(h) by the additive property of limits.
2. d(t) = f (t) − g(t) is o(h) by the additive and scalar multiplication proper-
ties of limits. (Here, the scalar is −1.)
We now prove an itermediate result. Lemma: If f (t) is o(h), then f (0) = 0.
proof: Clearly, by the definition of o(h), f (t) is continuous. So suppose
that f (0) is not equal to 0, say f (0) = a, a > 0. Then

f (h)
lim =∞
h→0 h
since a > 0. But f (t) is o(h), so that the required limit is zero. Hence it
must be that a = 0. So if f (t) is o(h) then f (0) = 0.

3. By the Lemma, limh→0 g(h) = limh→0 f (h) = 0. So

f (h)g(h) f (h)
  
lim = lim lim g(h) = 0 · 0 = 0
h→0 h h→0 h h→0

So p(t) = f (t)g(t) is o(h).

4. Since limh→0 f (h) and limh→0 g(h) exist independently,

q(h) limh→0 f (h) 0

lim = =
h→0 h limh→0 g(h)h 0
28 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

The limit is then indeterminant, and whether or not q(h) is o(h) depends
upon the specific form of f (h) and g(h).That is, we now need to know what
f (t), g(t) are to proceed further. For example, if f (t) = t4 , g(t) = t2 , then

q(h) h4
lim = lim 2 = lim h = 0
h→0 h h→0 h h h→0

making q(t) an o(h) function. If, on the other hand, f (t) = t2 , g(t) = t4
then
q(h) h2 1
lim = lim 4 = lim 3 = ∞
h→0 h h→0 h h h→0 h

and here q(t) is not o(h).

5. Since f (t) is continuous, there exists some n in [0, h] such that
Z h
f (x) dx = h · f (n)
0

Therefore
i(h) hf (n)
limh→0 = limh→0 h
h
= limh→0 f (n), n in [0, h]
=0
Rt
by the Lemma above. Hence i(t) = 0 f (x) dx is o(h).

Exercise 2.25 Show that definition 1 of the Poisson process implies def-
inition 2 of the Poisson process.
Solution. Denote Property (j) of Poisson process Definition n by (j)n .

1 (i)2 : Immediate.
2 (ii)2 : Property (ii)1 gives us independent increments; it remains to show
they are also stationary. By (iii)1 , we see that P {ñ(t + s) − ñ(s) = n} is
independent of s. This defines stationary increments.
3 (iii)2 : By (iii)1 ,
1 −λh
P {ñ(h) = 1} = λh 1!
e

= λhe−λh
= λhe−λh
= λhe −λh + λh − λh
 
= λh e−λh − 1 + λh
= g(h) + λh
Review of Random Processes 29

Note that
g(h) λh(e−λh −1)
lim = limh→0 h
h→0 h  
= λlimh→0 e−λh − 1
= λ (1 − 1)
=0

by Exercise 2.15 (v). So g(h) is o(h). Thus

P {ñ(h) = 1} = λh + g(h)
= λh + o(h)

And so (iii)2 holds.

4 (iv)2 : By (iii)1 ,
∞
X (λh)n e−λh
P {ñ(h) ≥ 2} =
n=2
n!

Let n ≥ 2 be arbitrary. Then

(λh)n e−λh
lim = limh→0 λn e−λh hn−1
h→0 n!
= λn limh→0 e−λh hn−1
= λn · 0
=0

since limh→0 eλh = 1 and limh→0 hn−1 = 0, n ≥ 2. And by Exercise

2.16 (i), sums of o(h) functions are o(h). So
∞
X (λh)n eλh
n=2
n!

is o(h). This shows (iv)2 .

Exercise 2.26 Show that Definition 2 of the Poisson process implies

Definition 1 of the Poisson process. [Hint: After satisfying the first two
properties of definition 1, establish that P0 (t) = exp{−λt} where Pn (t) =
P {ñ(t) = n} and then prove the validity of Property 3 of Definition 1 by
induction.]
Solution. Denote propery (j) of Poisson process definition n by (j)n .

1 (i)1 : Immediate.
30 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

2 (ii)1 : Immediate by the first half of (ii)2 .

3 (iii)1 : Define
Pn (t) = P {ñ(t) = n}, n≥0
with
Pn (t) = 0, n < 0.
Then Pn (t+h) = P {ñ(t + h) = n}. Conditioning upon the value of ñ(t),
we find
Pn
Pn (t + h) = k=0
Pn
P {ñ(t + h) = n|ñ(t) = k} P {ñ(t) = k}
= k=0 P{ñ(t + h) − ñ(t) = n − k} Pk (t)
= P {ñ(t + h) − ñ(t) = 0} Pn (t)
+P {ñ(t + h) − ñ(t) = 1} Pn−1 (t)
+P {ñ(t + h) − ñ(t) = 2} Pn−2 (t)
+···
+P {ñ(t + h) − ñ(t) = n} P0 (t)

But by (iii)2 , (iv)2 ,

P {ñ(t + h) − ñ(t) = 1} = λh + o(h), and

P {ñ(t + h) − ñ(t) ≥ 2} = o(h).

Thus,

P {ñ(t + h) − ñ(t) = 0} = P {ñ(h) = 0}

= 1 − [P {ñ(h) = 1} + P {ñ(h) ≥ 2}]
= 1 − [(λh + o(h)) + o(h)]
= 1 − λh + o(h).

Observe that
n
X Pn
P {ñ(t + h) − ñ(t) = k} Pn−k (t) = k=2 o(h)Pn−k (t) ≤ o(h) · 1
k=2
= o(h)

since Pn−k (t) is a scalar for all k ≤ n, and any scalar multiplied by a o(h)
function is still o(h). (Think of the limit definition of o(h) to see why this
is true.) Thus

Pn (t + h) = Pn (t) [1 − λh + o(h)] + Pn−1 (t) [λh + o(h)] + o(h).

Rearranging the terms,

Pn (t + h) − Pn (t) = −λhPn (t) + λhPn−1 (t) + o(h).

Review of Random Processes 31

We then divide by h and take the limit as h → 0

Pn (t + h) − Pn (t) −λhPn (t) λhPn−1 (t) o(h)

lim = lim + lim + lim ,
h→0 h h→0 h h→0 h h→0 h

so that
′
Pn (t) = −λPn (t) + λPn−1 (t) + 0
= −λPn (t) + λPn−1 (t), n ≥ 1. (2.26.1)

For n = 0,
′
P0 (t) = −λP0 (t).
i.e.,
′
P0 (t) + λP0 (t) = 0.
Observe that
d h λt i
e P0 (t) = 0,
dt
leads to
′
eλt P0 (t) + λeλt P0 (t) = 0
and
P0 (t) = Ke−λt . (2.26.2)
But by (i)2 , P0 (0) = 1. Using this result in (2.26.2) we find K = 1, so that

P0 (t) = eλt .

Let T denote the truth set for the Proposition that

(λt)n eλt
Pn (t) = .
n!
Then 0 ∈ T . Now suppose n − 1 ∈ T . That is,
" #
(λh)n−1 eλh
Pn−1 (t) = . (2.26.3)
(n − 1)!

By (2.26.1) from above,

′
Pn (t) = −λPn (t) + λPn−1 (t)

Multiplying by e−λt and rearranging terms:

h ′
i
e−λt Pn (t) + λPn (t) = λeλt Pn−1 (t).
32 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

Hence,
d λt
e Pn (t) = λeλt Pn−1 (t)
dt
Since n − 1 ∈ T , we have by (2.26.3),
(λt)n−1 e−λt
h i
= λeλt (n−1)!
λ(λt)n−1
= (n−1)!
λn t(n−1)
= (n−1)!

Integrating both sides with respect to t:

λn tn
e−λt Pn (t) = n(n−1)! + c
n
= (λt)
n! + c.

Thus,
(λt)n
 
Pn (t) = e−λt +c
n!
But by (i)2 , ñ(0) = 0. So for n > 0, Pn (0) = 0. Therefore,
h n i
Pn (0) = e−λ(0) (λ·0)
n! + c
= 1 (c) = 0.

That is,
(λt)n eλt
Pn (t) = .
n!
Thus, n ∈ T , the Proposition is proved, and all properties of Definition 1
are satisfied.

Exercise 2.27 Show that the sequence of interarrival times for a Poisson
process with rate λ forms a set of mutually iid exponential random variables
with parameter λ.
Solution. Recall that if {ñ(t), t ≥ 0} is a Poisson process with rate λ, then for
all s,
P {ñ(t + s) − ñ(s) = 0} = P {ñ(t) = 0} = e−λt
Now, in order for no events to have occurred by time t, the time of the first
event must be greater than t. That is,

P t̃1 > t = P {ñ(t) = 0} = e−λt



This is equivalent to
P t̃1 ≤ t = 1 − e−λt

Review of Random Processes 33

Observe that this is the cumulative distribution function for an exponential ran-
dom variable with parameter λ. Since the c.d.f. is unique, t̃1 is exponentially
distributed with rate λ.
Note that the second interarrival time begins at the end of the first interarrival
time. Furthermore, the process has stationary and independent increments so
that t̃2 has the same distribution as t̃1 and is independent
 of t̃1 .
Repeating these arguments for n ≥ 3, we see that t̃1 , t̃2 , t̃3 , · · · are inde-
pendent exponential variables with parameter λ.
Exercise 2.28 Show that
d λ(λt)n−1 e−λt
P {s̃n ≤ t} = .
dt (n − 1)!

[Hint: Start by noting s̃n ≤ t ⇐⇒ ñ(t) ≥ n].

Solution. First note that s̃n ≤ t if and only if ñ(t) ≥ n. Hence ñ(t) a Poisson
process that
∞
X (λt)k e−λt
P {s̃n ≤ t} = P {ñ(t) ≥ n} = ,
k=n
k!

since {ñ(t), t ≥ 0} is a Poisson process with rate λ. But this implies

d d (λt)k e−λt
P∞
P {s̃n ≤ t} = dt k=n
k!
dt
P∞ d h (λt)k e−λt i
= k=n dt k!
P∞ h kλk tk−1 e−λt λ(λt)k e−λt
i
= k=n k! − k!
P∞ h kλk tk−1 e−λt i P∞ h λ(λt)k e−λt i
= k=n k! − k=n k!
P∞ λ(λt)k−1 e−λt P∞ λ(λt)k e−λt
= k=n (k−1)! − k=n k! .
Observe that we may separate the first summation as
∞
" #
X kλk tk−1 e−λt λ(λt)n−1 e−λt
= (n−1)!
k=n
k!
P∞ λ(λt)k−1 e−λt P∞ λ(λt)k e−λt
+ k=n+1 (k−1)! − k=n k! .
Thus, upon reindexing the first summation,
d λ(λt)n−1 e−λt P∞ λ(λt)k e−λt P∞ λ(λt)k e−λt
P {s̃n ≤ t} = (n−1)! + k=n k! − k=n k!
dt
λ(λt)n−1 e−λt
= (n−1)!

Exercise 2.29 Show that Definition 3 of the Poisson process implies

Definition 1 of the Poisson process.
34 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

Solution.
(i) Clearly, since ñ(0) = max {n : s̃n ≤ 0}, ñ(0) = 0.

(ii) This follows immediately from the independence of t̃i , i ≥ 1 . Fur-
thermore, because these are exponential random variables and thus are
memoryless, ñ has stationary increments.

(iii) Using Exercise 2.20, note that

P {ñ(t) = n} = P {ñ(t) ≥ n} + P {ñ(t) ≥ n + 1}

= P {s̃n ≤ t} − P {s̃n+1 ≤ t} .
d λ(λt)n−1 e−λt λ(λt)n e−λt
P {ñ(t) = n} = (n−1)! − n!
dt
But the right-hand side is:
" #
d (λt)n e−λt
.
dt n!

Therefore
(λt)n e−λt
P {ñ(t) = n} =
n!
and the proof is complete.

Exercise 2.30 Let ñ1 and ñ2 be independent Poisson random variables
with rates α and β, respectively. Define ñ = ñ1 + ñ2 . Show that ñ has the
Poisson distribution with rate α + β. Using this result, prove Property 1 of
the Poisson process.
Solution. From the definition of the Poisson distribution,

αn e−n
P {ñ1 = n} =
n!
and
β n e−n
P {ñ2 = n} =
n!
Now, condition on the value of ñ2 to find:
  
Pn 
P {ñ = n} = k=0 P ñ1 + ñ2 = n, ñ2 = k P {ñ2 = k}


= Pnk=0 P {ñ1 + k = n} P {ñ2 = k}

P

= nk=0 P {ñ
h1= n − k}i hP {ñ 2 =
i k}
n−k −α
(α) e k −β (β) e
P n
= k=0 (n−k)! k!
Review of Random Processes 35

since both ñ1 , ñ2 are Poisson.

e−(α+β) Pn n!(α)n−k (β)k
= n! k=0 (n−k)!k!
e−(α+β) Pn n n−k β k
= n! k=0 (k ) α
−(α+β)
= e n! (α + β)n
n −(α+β)
= [(α+β)·t]n!e
This shows ñ = ñ1 + ñ2 is Poisson with rate α + β.
To prove Property 1, define {ñ1 (t), t ≥ 0} and {ñ2 (t), t ≥ 0} to be Pois-
son processes with rates α and β, respectively, and ñ(t) = ñ1 (t) + ñ2 (t).
Then, from Definition 1 of the Poisson process, ñ1 (t) and ñ2 (t) are Poisson
distributed with rates αt and βt, respectively. Thus, by the result just shown,
ñ(t) is Poisson distributed with parameter (α + β)t; this settles property (iii)
of Definition 1 of the Poisson process.
Since, by property (i) of Definition 1 of the Poisson process , ñ1 (0) =
ñ2 (0) = 0, and ñ(0) = ñ1 (0) + ñ2 (0), we find ñ(0) = 0; this settles property
(i) of Definition 1 of the Poisson process.
It remains to show that {ñ(t), t ≥ 0} has independent increments. Consider
two non-overlapping intervals of time, say (t0 , t1 ) and (t2 , t3 ). Now,
ñ(t1 ) − ñ(t0 ) = [ñ1 (t1 ) − ñ1 (t0 )] + [ñ2 (t1 ) − ñ2 (t0 )],
and
ñ(t3 ) − ñ(t2 ) = [ñ1 (t3 ) − ñ1 (t2 )] + [ñ2 (t3 ) − ñ2 (t2 )].
Since the random variables [ñ1 (t1 ) − ñ1 (t0 )] and [ñ2 (t1 ) − ñ2 (t0 )] are inde-
pendent, and [ñ1 (t3 )− ñ1 (t2 )] and [ñ2 (t3 )− ñ2 (t2 )] are independent, it follows
that [ñ(t1 ) − ñ(t0 )] and [ñ(t3 ) − ñ(t2 )] are independent. Thus, {ñ(t), t ≥ 0}
has independent increments, and the proof of Property 1 of Poisson processes
is complete.
Exercise 2.31 Suppose an urn contains ñ balls, where ñ is a Poisson
random variable with parameter λ. Suppose the balls are either red or green,
the proportion of red balls being p. Show that the distribution of the number
of red balls, ñr , in the urn is Poisson with parameter pλ, the distribution of
green balls, ñg is Poisson with parameter (1 − p)λ, and that ñr and ñg are
independent random variables. Use this result to prove Property 2 of the
Poisson process. [Hint: Condition on the total number of balls in the urn
and use the fact that the number of successes in a sequence of n repeated
Bernoulli trials has the binomial distribution with parameters n and p.]
Solution. Condition on the value of ñ, the total number of balls in the urn:
∞
X
P {ñr = r, ñg = g} = P {ñr = r, ñg = g|ñ = k} P {ñ = k}
k=0
36 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

Note that the only way for ñr = r and ñg = g, given that ñ = k, is if their
sum is r + g; that is, if k = r + g. So for k 6= r + g,

P {ñr = r, ñg = g|ñ = k} = 0.

Hence

P {ñr = r, ñg = g} = P {ñr = r, ñg = g|ñ = r + g} P {ñ = k}

Consider a red ball to be a success and a green ball to be a failure. Then we

may think of all of the balls in the urn as being a series of ñ Bernoulli exper-
iments. Hence P {ñr = r, ñg = g|ñ = r + g} is Binomial having parameters
(r + g), p. Then, since ñ is Poisson with rate λ,

P {ñr = r, ñg = g|ñ = r + g} P {ñ = k}

is ! " #
r+g r e−λt (λt)r+g e−λt (λt)r+g
p (1 − p)(r+g)−r =
r (r + g)! (r + g)!
After some rearranging of terms:
" #" #
e−pλt (pλt)r e−(1−p)λt {(1 − p)λt}g
=
r! g!

Summing over all possible values of ñg , we find

∞
X
P {ñr = r} = P {ñr = r, ñg = g}
g=0
∞
" #
(pλt)r e−pλt −(1−p)λt X {(1 − p)λt}g
 
= e
r! g=0
g!
" #
(pλt)r e−pλt −(1−p)λt (1−p)λt
= e e
r!
(pλt)r e−pλt
=
r!
Therefore ñr is Poisson with rate pλ. Similarly, by summing over all possible
values of ñr , P {ñg = g} is shown to be Poisson with rate (1 − p)λ. Further-
more, since

P {ñr = r, ñg = g} = P {ñr = r} P {ñg = g}

ñr and ñg are independent.

Now show ñr (t) and ñg (t) are Poisson processes. Let {ñ(t), t ≥ 0} be
a Poisson process with rate λ. By Definition 1 of the Poisson process ñ(t)
Review of Random Processes 37

is Poisson distributed with parameter λt. Let each event be recorded with
probability p. Define ñr (t) to be the number of events recorded by time t,
and ñg (t) to be the number of events not recorded by time t. Then by the
result just shown,ñ(t) = ñr (t) + ñg (t), where ñr (t) and ñg (t) are Poisson
distributed with rates pλt and (1 − p)λt, respectively. This proves property
(iii) of Definition 1 of the Poisson process.
By property (iii) of Definition 1 of the Poisson process, ñ(0) = 0. Since
ñr (t) and ñg (t) are non-negative for all t ≥ 0, ñr (0) = ñg (0) = 0. Thus
property (i) of Definition 1 of the Poisson process holds.
It remains to show property (ii) of Definition 1 of the Poisson process. Con-
sider two non-overlapping intervals of time, say (t0 , t1 ) and (t2 , t3 ). Then,
since ñ(t) = ñr (t) + ñg (t),

ñ(t1 ) − ñ(t0 ) = [ñr (t1 ) − ñr (t0 )] + [ñg (t1 ) − ñg (t0 )],

and
ñ(t3 ) − ñ(t2 ) = [ñr (t3 ) − ñr (t2 )] + [ñg (t3 ) − ñg (t2 )].
By the result shown above, ñr (t) and ñg (t) are independent random variables.
That is, [ñr (t1 ) − ñr (t0 )] and [ñg (t1 ) − ñg (t0 )] are independent, and [ñr (t3 ) −
ñr (t2 )] and [ñg (t3 ) − ñg (t2 )] are independent. Furthermore, {ñ(t), t ≥ 0}
is a Poisson process so it has independent increments: ñ(t1 ) − ñ(t0 ) and
ñ(t3 ) − ñ(t2 ) are independent. Since ñr (t) and ñg (t) are independent of
each other across the sums, which are then in turn independent of each other,
ñr (t) and ñg (t) are independent across the intervals. i.e., [ñr (t1 ) − ñr (t0 )] and
[ñr (t3 ) − ñr (t2 )] and independent, and [ñg (t1 ) − ñg (t0 )] and [ñg (t3 ) − ñg (t2 )]
and independent. This proves property (ii) of Definition 1 of the Poisson pro-
cess.
Since all three properties hold, {ñr (t), t ≥ 0} and {ñg (t), t ≥ 0} are Pois-
son processes.
Exercise 2.32 Events occur at a Poisson rate λ. Suppose all odd num-
bered events and no even numbered events are recorded. Let ñ1 (t) be the
number of events recorded by time t and ñ2 (t) be the number of events not
recorded by time t. Do the processes {ñ1 (t), t ≥ 0} and {ñ2 (t), t ≥ 0}
each have independent increments? Do they have stationary increments?
Are they Poisson processes?
Solution. Since only odd numbered events are recorded, the time between
recorded events is the sum of two exponentially distributed random variables
with parameter λ. Hence, the processes {ñ1 (t), t ≥ 0} and {ñ2 (t), t ≥ 0}
are clearly not Poisson. Now, suppose an event is recorded at time t0 . Then,
the probability that an event will be recorded in (t0 , t0 + h) is o(h) because
this would require two events from the original Poisson process in a period of
length h. Therefore, the increments are not independent. On the other hand,
38 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

the event probabilities are independent of t0 because the original process is

Poisson, so the increments are stationary.
Exercise 2.33 Determine the one-step transition probability matrix for
the Markov chain {qk , k = 0, 1, . . .} of Section 1.2.2 for the case where
{ṽk , k = 1, 2, . . .} is assumed to be a sequence of independent, identically
distributed binomial random variables with parameters N and p.
Solution. The dynamical equations are given as follows:

q̃k+1 = (q̃k − 1)+ + ṽk+1 .

By definition, the one-step transition probability is

P {q̃k+1 = j|q̃k = i} for all integers i, j

But,
n o
P {q̃k+1 = j|q̃k = i} = P (q̃k − 1)+ + ṽk + 1 = j|q̃k = i .

In turn,
n o n o
P (q̃k − 1)+ + ṽk + 1 = j|q̃k = i = P (i − 1)+ + ṽk + 1 = j|q̃k = i
n o
= P ṽk + 1 = j − (i − 1)+ |q̃k = i

But, ṽk+1 is independent of q̃k = i. It follows that

n o n o
P ṽk + 1 = j − (i − 1)+ |q̃k = i = P ṽk + 1 = j − (i − 1)+ .

For i = 0 or i = 1,
n o
P ṽk + 1 = j − (i − 1)+ = P {ṽk + 1 = j} .

For i ≥ 2,
n o
P ṽk + 1 = j − (i − 1)+ = P {ṽk + 1 = j + 1 − i} ,

where we note that the previous probability is zero whenever j < i − 1. Define
aj = P {ṽk + 1 = j} for i = 0, 1, . . . , N . Then, we have shown that

a0 a1 a2 a3 a4 a5 a6 a7 ··· aN 0 ···
 
 a0
 a1 a2 a3 a4 a5 a6 a7 ··· aN 0 ···
 .. .. 
P= 0
 a0 a1 a2 a3 a4 a5 a6 ··· aN −1 . ..
 .. .. 
 0
 0 a0 a1 a2 a3 a4 a5 ··· aN −2 . .
.. .. .. .. .. .. .. .. .. .. ..
. . . . . . . . . . .
Review of Random Processes 39

Exercise 2.34 Staring with (2.18), show that the rows of P ∞ must
be identical. [Hint: First calculate P ∞ under the assumption β0 =
[ 1 0 0 . . . ]. Next, calculate P ∞ under the assumption β0 =
[ 0 1 0 . . . ]. Continue along these lines.]
Solution. The general idea is that π is independent of β0 ; that is, the value of
π is the same no matter how we choose the distribution of the starting state.
Suppose we choose β0 = [ 1 0 0 . . . ]. Then, β0 P ∞ is equal to the first
row of P ∞ . Since π = β0 P ∞ , we then have the first row of P ∞ must be π.
if we choose β0 = [ 0 1 0 . . . ], then β0 P ∞ is equal to the second row
of P ∞ . Again, since π = β0 P ∞ , we then have the first row of P ∞ must be
π. In general, if we choose the starting state as i with probability 1, then we
conclude that the i-th row of P ∞ is equal to π. thus, all rows of P ∞ must be
equal to π.
Exercise 2.35 Suppose {x̃k , k = 0, 1, . . .} is a Markov chain such that
 
0.6 0.4
P= .
0.5 0.5

Determine the stationary vector of {x̃k , k = 0, 1, . . .}.

Solution. We have π = πP and πe = 1. Thus,
 
0.6 0.4
[ π1 π2 ] = [ π1 π2 ] ,
0.5 0.5
or
0.4 −0.4
 
[ π1 π2 ] = [0 0].
−0.5 0.5
Also πe = 1 means  
1
[ π1 π2 ] =.
2
Upon substitution of the last equation into the second equation of the matrix
equation, we have
 
0.4 1
[ π1 π2 ] = [0 1].
−0.5 1
1
Adding 2 of the second equation to the first yields
 
0.9 1
[ π1 π2 ] = [ 0.5 1].
0 1
Then dividing the first equation by 0.9 yields
 
1 1
[ π1 π2 ] = [ 59 1].
0 1
40 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

And, finally subtracting the first equation from the second yields
 
1 0
[ π1 π2 ] = [ 59 4
9 ].
0 1
or
[ π1 π2 ] = [ 59 4
9 ].

Exercise 2.36 Develop the one-step state probability transition matrix

for the special case {ṽk , k = 0, 1, . . .} for the special case of N = 5.
Solution. We wish to find P {ṽk+1 = j|ṽk = i} for i, j ∈ {0, 1, 2, 3, 4, 5}.
Now, ṽk is the number of sources in the on state during interval k. Thus, if
ṽk = i, then i sources are in the on state during period k and 5 − i sources are
in the off state during that period. Let ãi denote the number of sources that are
in the on state in period k + 1 if i sources are in the on state in period k and b̃i
denote the number of sources that are in the on state in period k + 1 if i sources
are in the off state in period k. HTen, the number of on sources during period
k + 1 given i on sources in period k is equal to ãi + b̃5−i . Thus,
n o
P {ṽk+1 = j|ṽk = i} = P ãi + b̃5−i = j .

But,
n o i
X n o
P ãi + b̃5−i = j = P ãi + b̃5−i = j|ãi = ℓ P {ãi = ℓ}
ℓ=0
i
X n o
= P b̃5−i = j − ℓ P {ãi = ℓ} ,
ℓ=0

the latter step following by independence

n among theobehavior of the sources.
In the summation, if ℓ > j, then P b̃5−i = j − ℓ = 0, so we may ter-
minate the summationn at min {i,
o j}. In addition, j − ℓ > 5 − i or ℓ <
j + i − 5, then P b̃5−i = j − ℓ = 0. Thus, we may begin the summation at
max {0, j + i − 5}. Thus
min {i,j} n o
X
P {ṽk+1 = j|ṽk = i} = P b̃5−i = j − ℓ P {ãi = ℓ} .
ℓ=max {0,j+i−5}

Since each of the off source turns on with probability p01 .

!
n o 5 − i j−ℓ (5−i)−(j−ℓ)
P b̃5−i = j − ℓ = p p .
j − ℓ 01 00
Review of Random Processes 41

Similarly, since each of the on source remains on with probability p11 .

!
i ℓ i−ℓ
P {ãi = ℓ} = p p .
ℓ 11 10

The transition probabilities can now be obtained by performing the indicate

summations. Some examples are

P {ṽk+1 = 0|ṽk = 0} = p500 ,

P {ṽk+1 = 1|ṽk = 0} = 5p01 p400 ,

and
3
X n o
P {ṽk+1 = 3|ṽk = 4} = P b̃1 = 3 − ℓ P {ã3 = ℓ}
ℓ=2
n o n o
= P b̃1 = 1 P {ã4 = 2} + P b̃1 = 0 P {ã4 = 3}
! !
4 2 2 4 3
= p01 p11 p10 + p00 p p10
2 3 11
= 6p01 p211 p210 + 4p00 p311 p10

Exercise 2.37 For the example discussed in Section 1.2.2 in the case
where arrivals occur according to an on off process, determine whether or
not {q̃k , k = 0, 1, . . .} is a DPMC. Defend your conclusion mathematically;
that is show whether or not {q̃k , k = 0, 1, . . .} satisfies the definition of a
Markov chain.
Solution. We are given

q̃k+1 = (q̃k − 1)+ + ṽk+1 ,

and we want to know whether or not

P {q̃k+1 = j|q̃0 = i0 , q̃1 = i1 , . . . , q̃k = i} = P {q̃k+1 = j|q̃k = i} .

For the case where the arrival process is an independent sequence of random
variables, the answer is “Yes.” The sequence of values of q̃k does reveal the
number of arrivals in each interval, but the probability mass function for the
queue length at epoch k + 1 can be computed solely form the value of infor-
mation is not needed to compute the new value of q̃k . For the case where the
arrival process is on-off, the sequence of queue lengths reveals the sequence
of arrivals. For example, if q̃k = 7 and q̃k−1 = 3, then we know that there
were 5 arrivals during interval k, which means that there were 5 sources in the
on state during interval k. Thus, the distribution of the number of new arrivals
42 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

during interval k + 1, and therefore the distribution of q̃k+1 , is dependent upon

the values of q̃k and q̃k−1 . Thus, if we know only that q̃k = i, we cannot
compute the distribution of q̃k+1 . Hence for the general on-off arrival process,
{q̃k , k = 0, 1, . . .} is not a DPMC.
Exercise 2.38 Suppose {x̃(t), t ≥ 0} is a time-homogeneous CTMC
having infinitesimal generator Q defined as follows:

 −λ,

if j = i,
Qij = λ, if j = i + 1,
0, otherwise.


Show that {x̃1 (t), t ≥ 0} is a Poisson process. [Hint: Simply solve the
infinite matrix differential equation term by term starting with P00 (t) and
completing each column in turn.]
Solution. Solve the infinite matrix differential equation:
d
P (t) = P (t)Q.
dt
Due to independent increments, we need only compute the first row of the
matrix P (t). Begin with P0,0 (t):
d
P0,0 (t) = −λP0,0 (t).
dt
We see immediately that
P0,0 (t) = e−λt .
Now solve for the second column of P (t):
d
P0,1 (t) = λP0,0 (t) − λP0,1 (t)
dt
= λe−λt − λP0,1 (t).

This solves to
P0,1 (t) = λte−λt
Repeating this process for each column of P (t), we find the solution of P (t)
to be
e−λt (λt)n
P0,n (t) =
n!
Observe that this probability is that of a Poisson random variable with param-
eter λt. Hence the time between events is exponential. By Definition 2.16,
{x̃(t), t ≥ 0} is a Poisson process. Furthermore, this is strictly a birth process;
that is, the system cannot lose customers once they enter the system. This
should be apparent from the definition of Q.
Review of Random Processes 43

Exercise 2.39 Let {x̃(t), t ≥ 0} be a CTMC such that

−2.00 1.25 0.75
 

Q = 0.50
 −1.25 0.75  .
1.00 2.00 −3.00

1. Solve for P (∞) directly by solving P (∞)Q = 0 P( ∞)e = 1.

2. Solve for π for the DPMC embedded at points of state transition using
(2.24).
3. Find P for for the DPMC embedded at points of state transition.
4. Show that the value of π found in part 2 of this problem satisfies π = πP
for the P found in part 3 of this problem.

Solution.
1.
−2.00 1.25 0.75
 

[ P (∞)0 P (∞)1 P (∞)2 ]  0.50 −1.25 0.75  = [ 0 0 0 ] ,

1.00 2.00 −3.00
and
1
 

[ P (∞)0 P (∞)1 P (∞)2 ]  1  = 1

1
yield
−2.00 1.25 1
 

[ P (∞)0 P (∞)1 P (∞)2 ] 0.50

 −1.25 1 = [0 0 1].
1.00 2.00 1
Then dividing the first equation by -2 and the second equation by -1.25
yields
1 −1 1
 

[ P (∞)0 P (∞)1 P (∞)2 ]  − 14 1 1 = [0 0 1].

− 21 − 85 1
Then adding the first equation to the second and subtracting the first equa-
tion from the third yields:
1 0 0
 

[ P (∞)0 P (∞)1 P (∞)2 ] − 14

 3
4
5
4
 = [0 0 1].
− 12 21
− 10 3
2
44 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

Dividing the second equation by 3/4 yields

1 0 0
 

[ P (∞)0 P (∞)1 P (∞)2 ] − 14

 1 5
4
 = [0 0 1].
− 12 − 14
5
3
2

Subtracting 5/4 of the second equation from the third yields

1 0 0
 

[ P (∞)0 P (∞)1 P (∞)2 ]  − 14 1 0 = [0 0 1].

− 21 − 14
5 5
Dividing the third equation by 5 yields

1 0 0
 

[ P (∞)0 P (∞)1 P (∞)2 ]  − 14 1 0 = [0 0 0.2 ] .

− 21 − 14
5 1
Then adding 14/5 times the third equation to the second yields

1 0 0
 

[ P (∞)0 P (∞)1 P (∞)2 ]  − 14 1 0 = [0 14

25 0.2 ] .
− 21 0 1
Finally adding 1/4 of the second equation and one half of the third equation
to the first equation yields
6 14 5
[ P (∞)0 P (∞)1 P (∞)2 ] = [ 25 25 25 ].

2. We have from (2.24)

P
i6=j Pi (∞)Qij
πj = P P .
ℓ i6=ℓ Pi (∞)Qiℓ

Plugging in numbers, we find

X 14 1 5 12
Pi (∞)Qi0 = + = ,
i6=0
25 2 25 25

X 6 5 5 70
Pi (∞)Qi1 = + 2= ,
i6=1
25 4 25 100

and
X 6 3 14 3 60
Pi (∞)Qi1 = + = ,
i6=2
25 4 25 4 100
Review of Random Processes 45

Thus,
48 48 70 60
π0 = = . π1 = , and π2 = .
48 + 70 + 60 178 178 178

3. From

Qij
P {x̃k+1 = j|x̃k = i} = ,
−Qii
for i 6= j and P {x̃k+1 = j|x̃k = j} = 0, we have immediately
5 3
0
 
8 8
P =  52 0 3
5
.
1 2
3 3 0

4. We wish to show that the value of π computed in Part 2 satisfies π = πP.

We have

5 3
0
 
8 8
48 70 60
[ 178 178 178 ]  25 0 3
5
 = 70 2
[ 178 5 +
60 1
178 3
48 5
178 8 + 60 2
178 3
48 3
178 8 + 70 3
178 5 ]
1 2
3 3 0
48 70 60
= [ 178 178 178 ].
Chapter 3

ELEMENTARY CONTINUOUS-TIME
MARKOV CHAIN-BASED QUEUEING MODELS

Exercise 3.1 Carefully pursue the analogy between the random walk
and the occupancy of the M/M/1 queueing system. Determine the proba-
bility of an increase in the queue length, and show that this probability is
less than 0.5 if and only if λ < µ.
Solution. Consider the occupancy of the M/M/1 queueing system as the posi-
tion of a random walker on the nonnegative integers, where a wall is erected
to the left of position zero. When there is an increase in the occupancy of the
queue this is like the walker taking a step to the right; a decrease is a step to
the left. However, if the queue is empty (the walker is at position zero), then a
‘decrease’ in the occupancy is analogous to the walker attempting a step to the
left – but he hits the wall and so remains at position zero (queue empty).
Let λ be the rate of arriving customers and µ be the rate of departing cus-
tomers. Denote the probability of an arrival by p+ and a departure by p− .
Then p+ is simply the probability that an arrival occurs before a departure.
λ
From Chapter 2, this probability is λ+µ . It follows that p+ < 21 if and only if
λ < µ.
Exercise 3.2 Prove Theorem 3.1 and its continuos analog
Z ∞
E[x̃] = P {x̃ > x}dx.
0
Solution.
∞
X
E [ñ] = nP {ñ = n}
n=0
∞ n
!
X X
= 1 P {ñ = n}
n=0 k=1
48 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

∞ n
!
X X
= P {ñ = n}
n=0 k=1

Changing the orders of summation

 
∞ n ∞ ∞
!
X X X X
P {ñ = n} =  P {ñ = k}
n=0 k=1 n=0 k=n+1
X∞
= P {ñ > n}
n=0

For the continuous analog, we have

Z ∞
E [x̃] = xfx̃ (x)dx.
0

But, Z x
x= dy,
0
so Z ∞Z x
E [x̃] = dyfx̃ (x)dx.
0 0
Changing order of integration yields
Z ∞Z ∞ Z ∞ Z ∞
E [x̃] = fx̃ (x)dxdy = P {x̃ > y} dy = P {x̃ > x} dx.
0 y 0 0

Exercise 3.3 Prove Theorem 3.2.

Solution. The theorem is as follows: Suppose x̃ and ỹ are any two nonegative
random variables. Then E [min {x̃, ỹ}] ≤ min {E [x̃] , E [ỹ]}.
First we consider the continuous case. Define z̃ = min {x̃, ỹ}. Then, z̃ > z
if, and only if, {x̃ > z, ỹ > z}. But,
Z ∞ Z ∞
E [z̃] = P {z̃ > z} dz = P {x̃ > z, ỹ.z} .
0 0

But, P {x̃ > z, ỹ.z} ≤ P {x̃ > z} and P {x̃ > z, ỹ.z} ≤ P {ỹ > z} because
the event {x̃ > z, ỹ > z} ⊆ {x̃ > z} and {x̃ > z, ỹ > z} ⊆ {ỹ > z}. There-
fore, Z ∞ Z ∞
E [z̃] ≤ P {x̃ > z} and E [z̃] ≤ P {ỹ > z} .
0 0
Equivalently,
E [z̃] ≤ E [x̃] and E [z̃] ≤ E [ỹ] .
Elementary CTMC-Based Queueing Models 49

The previous statement is equivalent to

E [min {x̃, ỹ}] ≤ min {E x̃, E [ỹ]}.

The discrete and mixed cases are proved in the same way.

Exercise 3.4 Suppose customers arrive to a system at the end of every

even-numbered second and each customer requires exactly one second of
service. Compute the stochastic equilibrium occupancy distribution, that
is, the time-averaged distribution of the number of customers found in the
system. Compute the occupancy distribution as seen by arriving customers.
Compare the two distributions. Are they the same?

Solution. Since a customer arrives every two seconds and leaves after one
second, there will either be one customer in the system or no customers in the
system. Each of these events is equally likely since the time intervals that they
occur in is the same (1 second). Hence

E [s̃] = E [s̃|ñ = 0] P {ñ = 0} + E [s̃|ñ = 1] P {ñ = 1}

1 1
= 0· +1·
2 2
1
=
2

On the other hand, because a customer finishes service at the end of every
odd-numbered second, she has left the system by the time the next cusomer
arrives. Hence an arriving customer sees the system occupancy as empty; that
is, as zero. Obviously this is not the same as E [s̃] and we conclude that the
distributions as seen by an arbitrary arriving customer and that of an arbitrary
observer cannot be the same.
50 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

Exercise 3.5 For the ordinary M/M/1 queueing system, determine the
limiting distribution of the system occupancy
1. as seen by departing customers, [Hint: Form the system of equa-
tions πd = πd Pd , and then solve the system as was done to obtain
P {ñ = n}.]
2. as seen by arriving customers, and [Hint: First form the system of equa-
tions πa = πa Pa , and then try the solution πa = πd .]
3. at instants of time at which the occupancy changes. That is, embed a
Markov chain at the instants at which the occupancy changes, defining
the state to be the number of customers in the system immediately fol-
lowing the state change. Define π = [ π0 π1 · · · ] to be the stationary
probability vector and P to be the one-step transition probability matrix
for this embedded Markov chain. Determine π, and then compute the
stochastic equilibrium distribution for the process {ñ(t), t ≥ 0} accord-
ing to the following well known result from the theory of Markov chains
as discussed in Chapter 2:
πi E[s̃i ]
Pi = P
∞ ,
πi E[s̃i ]
i=0

where s̃i denotes the time the systems spends in state i on each visit.
Observe that the results of parts 1, 2, and 3 are identical, and that these are
all equal to the stochastic equilibrium occupancy probabilities determined
previously.
Solution.
1. Let ñd (k) denote the number of customers left by the k − th departing
customer. Then the (i, j) − th element of Pd is given by

P {ñd (k + 1) = j|ñd (n) = i} .

This probability is given by the probability that exactly j − (i − 1)+ cus-

tomers arrive during the k − th service time. For j − (i − 1)+ < 0,

P {ñd (n + 1) = j|ñd (n) = i} = 0,

since you can’t have a negative number of arrivals. Otherwise, letting A

denote the event of an arrival before the k − th service completion,
j−(i−1)+
λ µ

 +
P j − (i − 1) A = .
λ+µ λ+µ
Elementary CTMC-Based Queueing Models 51

This results in
  2 
µ λ µ λ µ
 λ+µ λ+µ λ+µ λ+µ λ+µ ...
 2
 µ µ µ
 
λ λ
 λ+µ λ+µ λ+µ λ+µ λ+µ ...

Pd =  0 µ λ µ
 
 λ+µ λ+µ λ+µ ... 
µ
 0 0 ...
 
λ+µ
.. .. ..
 
..
. . . .
Using this matrix, form the system of equations πd = πd Pd . Then, for
i = 0,
∞
X
πd0 = πdi Pdi0
i=0
∞
X
= πd0 Pd00 + πd1 Pd10 + πdi Pdi0
i=2
µ λ µ
   
= πd0 + πd1 + 0.
λ+µ λ+µ λ+µ
Hence πd1 = µλ πd0 . Substitute this into the equation for πd1 :
P∞
πd1 = i=0 πdi Pdi1
∞
X
= πd1 Pd01 + πd1 Pd11 + πd2 Pd21 πdi Pdi1
i=3
µ λ λ λ µ
     
= πd0 + πd0
λ+
µ λ+ µ µ λ+µ λ+µ
µ
+πd2 +0
λ+µ
 2
λ
which gives the solution πd2 = µ πd0 . Repeating this process we see
 j
λ
that πdj = µ πd0 , j = 0, 1, 2, . . ..
P∞
Now use the normalizing constraint j=0 πdj = 1. Then for λ < µ,
∞  j
!
X λ 1
πd0 = πd0 λ
= 1.
j=0
µ 1− µ
 j  
Thus πd0 = 1 − µλ , and πdj = λ
µ 1− λ
µ .

2. Let ña (k) denote the number of customers as seen by the k − th arriving
customer. Then the (i, j) − th element of Pa is given by
P {ña (k + 1) = j|ñd (n) = i} .
52 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

For j = 0, 1, . . . i+1 this probability is given by the probability that exactly

i + 1 − j customers arrive during the k − th service time:
1−j
µ
 
P {ña (k + 1) = j|ñd (n) = 0} = ,
λ+µ
i+1−j
λ µ

P {ña (k + 1) = j|ñd (n) = i} = , i>0
λ+µ λ+µ
For j > i + 1, this probability is equal to zero since only departures may
occur between arrivals. This results in
µ λ
0 0 0 ...
 
λ+µ λ+µ
  µ 2 λ µ λ


 λ+µ λ+µ λ+µ λ+µ 0 0 ...

Pa = 
 µ 3
 
λ

µ
2
λ µ λ

0 ...

 λ+µ λ+µ λ+µ λ+µ λ+µ λ+µ
 
.. .. .. .. .. . .
. . . . . .

Form the system of equations πa = πa Pa . Then, for j = 0,

∞
X
πa0 = πai Pai0
i=0
∞ i
µ X µ

= πa .
λ + µ i=0 i λ+µ

Now, to show that πa = πd , it suffices to show that πd is a solution to

πa = πa Pa . Substitute πd0 into the equation just found for πa0 :
∞  i  i
µ X µ−λ λ µ
πa0 =
λ + µ i=0 µ µ λ+µ
∞ i
µ−λX λ

=
λ + µ i=0 λ + µ
µ−λ λ+µ
= ·
λ+µ µ
µ−λ
=
µ
which is the same result found for πd0 in part (a).
Now generalize to πaj , j > 0, and substitute πdi for πai :
∞ i
λ X µ

πaj = πaj−1+i
λ + µ i=0 λ + µ
Elementary CTMC-Based Queueing Models 53
∞ 
  j−1 X i
λ µ−λ λ λ

=
λ+µ µ µ λ+µ
  j−1 i=0
λ µ−λ λ λ+µ

=
λ+µ µ µ µ
µ−λ λ j
 
= .
µ µ
Thus πa = πd .

3. Determine P, the one-step transition probability matrix. If the system is in

state 0 at time k, then with probability 1 it will be in state 1 at time (k + 1).
The probability of it being in any other state at time (k + 1) is 0. If the
system is in state i, (i 6= 0) at time k, then the next event will either be a
departure or an arrival. If it’s a departure, the system wil be in state (i − 1)
µ
at time (k + 1). This happens with probability λ+µ . If the next event is an
arrival, then the system will be in state (i + 1) at time (k + 1). This will
λ
occur with probability λ+µ . Thus,
 
0 1 0 0 0 ...
 µ 0 λ
0 0 ...
 λ+µ λ+µ 
µ λ
P= 0 0 0 ...
 
λ+µ λ+µ


µ λ
 0 0 0 ...
 
λ+µ λ+µ
.. .. .. .. ..
 
..
. . . . . .
P∞
Form the system of equations πj = i=0 πj Pij . Then for j = 0:
∞
µ
  X
π0 = 0 + π1 + πj · 0.
λ+µ i=2

That is, π1 = π0 λ+µ

P∞
µ . Substituting this into π1 =
   i=0 πj P1j gives the
λ+µ λ
solution π2 = π0 µ µ . Repeating this process we see that for j ≥ 1,
 j−1
λ+µ λ
πj = π0 .
µ µ
Now use the normalizing constraint to find π0 :
∞   j−1
λ+µ λ
X 
1 = π0 + π0
j=0
µ µ
" !#
λ+µ 1
 
= π0 1 + λ
.
µ 1− µ
54 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

i.e., π0 = (µ − λ)/2µ. And so for j ≥ 1,

  j−1
λ+µ λ µ−λ

πj = .
µ µ 2µ
To compute Pi , first determine E[s̃i ], the expected amount of time the sys-
tem spends in state i on each visit. Note that s̃i is an exponential random
variable, since the time spent in state i is determined by when the next ar-
rival or departure may occur. The rate for s̃0 is λ, since only an arrival
may occur. And the rates for all other s̃i , i > 0, is λ + µ, since either an
arrival or a departure may occur. Thus by the properties of the exponential
distribution,
1
E[s̃0 ] =
λ
1
E[s̃i ] = , i > 0.
λ+µ
Then
∞ ∞    i−1 
λ+µ λ µ−λ 1
X X  
πi E[s̃i ] = π0 E[s̃0 ] +
i=0 i=1
µ µ 2µ λ+µ
" #
µ−λ 1 µ−λ 1
= · + 2 λ
2µ λ 2µ 1− µ
1
= .
2λ
Using this result and the πj′ s obtained above in

πi E[s̃i ]
Pi = P
∞
πi E[s̃i ]
i=0

we get the solutions

µ−λ 1
2µ · λ λ
P0 = 1 =1−
µ
 2λ   i−1   
λ+µ λ µ−λ 1
µ µ 2µ λ+µ
Pi = 1
 j  2λ
λ λ

= 1− .
µ µ
This is precisely the results of parts (a) and (b).
Elementary CTMC-Based Queueing Models 55

Exercise 3.6 Using Little’s result, show that the probability that the
server is busy at an arbitrary point in time is equal to the quantity (λ/µ).
Solution. Let B denote the event that the server is busy. Define the system to
be the server only (i.e., no queue). Then

P {B} = P {ñ = 1}

where ñs is the number of customers in the system (i.e., ñs is the number in
service). Hence

E[ñ] = 0 · P {ñ = 0} + 1 · P {ñ = 1}

= P {ñ = 1}
= P {B} .

Furthermore, the average waiting time in the system is just the average service
time. That is, E[w̃] = µ1 . By Little’s Result, E[ñ] = λE[w̃]. Thus

λ
P {B} = E[ñ] = .
µ

Exercise 3.7 Let w̃ and s̃ denote the length of time an arbitrary cus-
tomer spends in the queue and in the system, respectively, in stochastic
equilibrium. Let Fs̃ (x) ≡ P {s̃ ≤ x} and Fw̃ (x) ≡ P {w̃ ≤ x}. Show that

Fs̃ (x) = 1 − e−µ(1−ρ)x , for x ≥ 0,

and
Fw̃ (x) = 1 − ρe−µ(1−ρ)x , for x ≥ 0,
without resorting to the use of Laplace-Stieltjes transform techniques.
Solution. To compute Fs̃ (x),condition on the value of ñ, the number of cus-
tomers in the system:
∞
X
Fs̃ (x) = P {s̃ ≤ x|ñ = n} P {ñ = n}, (3.6.1)
n=0

where P {ñ = n} = ρn (1 − ρ).

If a customer arrives at time t and finds n customers in the system, then the
probability that she departs the system by time t + x equals the probability that
there will be n + 1 service completions in the interval (t, t + x]: the original
n customers plus herself. This probability is P {s̃n+1 ≤ x}. Recall Exercise
2.20:
d µ(µx)n e−µx
P {s̃n+1 ≤ x} = .
dx n!
56 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

By differentiating and then integrating both sides of (3.6.1),

∞
µ(µα)n e−µα n
xX
Z
Fs̃ (x) = ρ (1 − ρ)dα
0 n=0 n!
∞
(µρα)n
Z x
−µα
X
= µ(1 − ρ)e dα
0 n=0
n!
Z x
= µ(1 − ρ)e−µ(1−ρ)α dα
0
= 1 − e−µ(1−ρ)x , x ≥ 0.

Now find Fw̃ (x):

Fw̃ (x) = P {w̃ ≤ 0} + P {w̃ < x} , x>0

= P {w̃ ≤ 0} + intx0 fw̃ (α)dα, x > 0.

But P {w̃ ≤ 0} = P {w̃ = 0} = P {ñ = 0} = 1 − ρ, so

Z x
Fw̃ (x) = (1 − ρ) + fw̃ (α)dα, x > 0.
0

Condition fw̃ on the value of ñ, the number of customers in the system and use
the fact that the system is in stochastic equilibrium:
∞
µ(µα)n−1 e−µα n
xX
Z
Fw̃ (x) = (1 − ρ) + ρ (1 − ρ)dα
0 n=1 (n − 1)!
∞
(µρα)n
Z x
µρ(1 − ρ)e−µα
X
= (1 − ρ) + dα (reindexing)
0 n=0
n!
Z x
= (1 − ρ) + µρ(1 − ρ)e−µ(1−ρ)α dα
0
−µ(1−ρ)x
= 1 − ρe , x ≥ 0.

Exercise 3.8 M/M/1 Departure Process. Show that the distribution of

an arbitrary interdeparture time for the M/M/1 system in stochastic equi-
librium is exponential with the same parameter as the interarrival-time dis-
tribution. Argue that the interdeparture times are independent so that the
departure process for this system is Poisson with the same rate as the arrival
process (Burke [1956]). [Hint: Use the fact that the Poisson arrival sees the
system in stochastic equilibrium. Then condition on whether or not the ith
departing customer leaves the system empty.]
Solution. Observe that the Poisson arrivals will see the system in stochastic
equilibrium. Using this fact, condition on whether or not the i − th departing
customer leaves the system empty. Now, if the system is not left empty, the
Elementary CTMC-Based Queueing Models 57

time to the next departure is simply the time it takes for the (i+1)−st customer
to complete service. This is Poisson with parameter µ. If the system is left
empty, however, the time to the next departure is the time to the next arrival
plus the time for that arrival to complete service. Denoting the event that the
system is left empty by A,

n o
P d˜ ≤ d|A = P {ã + x̃ ≤ d} .

Recall that in stochastic equilibrium the probability that the system is empty is
(1 − ρ). Then if B denotes the event that the system is not left empty,

n o n o n o
P d˜ ≤ d = P d˜ ≤ d|B P {B} + P d˜ ≤ d|A P {A}
= ρP {x̃ ≤ d} + (1 − ρ)P {ã + x̃ ≤ d} .

Sum over all possible values of ã to get P {ã + x̃ ≤ d}, noting that P {ã + x̃ ≤ d} =
0 if ã > d. Hence

n o   Z d
P d˜ ≤ d = ρ 1 − e−µd + (1 − ρ) P {ã + x̃ ≤ d|ã = a} dFã .
0

Since ã and x̃ are independent of each other, this is

n o   Z d
P d˜ ≤ d = ρ 1−e −µd
+ (1 − ρ) P {a + x̃ ≤ d} λe−λa da
  Z0 d  
−µd
= ρ 1−e + (1 − ρ) 1 − e−µ(d−a) λe−λa da
0
= 1 − e−λd .

This shows the departure process d˜ is Poisson with parameter λ. That is, the
departure process occurs at the same rate as the arrival process, which we know
to be independent. Because of this ‘rate in equals rate out’ characteristic, this
implies that the interdeparture times are also independent. If the interdepar-
ture times were not independent, then its distribution could not be that of the
interarrival times.
58 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

Exercise 3.9 M/M/1 with Instantaneous Feedback. A queueing system

has exogenous Poisson arrivals with rate λ and exponential service with rate
µ. At the instant of service completion, each potentially departing customer
rejoins the service queue, independent of system state, with probability p.
1. Determine the distribution of the total amount of service time the server
renders to an arbitrary customer.
2. Compute the distribution of the number of customers in the system in
stochastic equilibrium. How does your solution compare to that of the
M/M/1 queueing system? What explains this behavior? [Hint: Consider
the remaining service time required for each customer in the queue. Sup-
pose customers that required additional increments of service returned
immediately to service rather than joining the tail of the queue. What
would be the effect on the queue occupancy?]
3. Argue that the departure process from the system is a Poisson process
with rate λ.
4. Compute the average sojourn time for this system and comment on com-
putation of the distribution of the sojourn time.

Solution.

1. Let t̃ be the total amount of service time rendered to an arbitrary customer.

Condition on ñ, the number of times the customer joins the queue.
∞
X 
Ft̃ (x) = P t̃ ≤ x|ñ = n P {ñ = n}. (3.8.1)
n=1

Now, the probability of the customer entering the queue n times is simply
pn−1 (1 − p) since he returns to the queue (n − 1) times with probability
pn−1 and leaves after the n-th visit with probability (1
 − p). Furthermore,

if the customer was in the queue n times, then P t̃ ≤ x|ñ = n is the
probability that the sum of his n service times will be less than x. That is,
(∞ )
 X
P t̃ ≤ x|ñ = n = P x̃i ≤ x
i=1
= P {s̃n ≤ x} , x̃i ≥ 0.

Recall Exercise 2.20:

d µ(µx)n e−µx
P {s̃n+1 ≤ x} = .
dx n!
Elementary CTMC-Based Queueing Models 59

Then, upon differentiating and then integrating both sides of (3.8.1),

∞
µ(µα)n−1 e−µα n−1
xX
Z
Ft̃ (x) = p (1 − p)dα
0 n=1 (n − 1)!
∞
(µpα)n
Z x
µe−µα (1 − p)
X
= dα
0 n=1
n!
Z x
= µe−µα (1 − p)eµpα dα
0
= 1 − e−µ(1−p)x , x ≥ 0.

2. Consider those customers who, on their first pass through the server, still
have increments of service time remaining. Suppose that instead of joining
the end of the queue, they immediately reenter service. In effect, they ’use
up‘ all of their service time on the first pass. This will simply rearrange
the order of service. Since the queue occupancy does not depend on order
of service, reentering these customers immediately will have no effect on
the queue occupancy. Observe that the arrival process is now a Poisson
process: since customers complete service now in one pass and don’t join
the end of the queue as they did before, they arrive to the queue only once.
Hence, arrivals are independent. With this is mind, we see that the system
can now be modeled as an ordinary M/M/1 system with no feedback. From
part (a), the total amount of service time for each customer is exponential
with parameter (1−p)µ. The distribution of the number of customers in the
system in stochastic equilibrium is the same as that of an ordinary M/M/1
λ
system, with this new service rate: Pn = (1 − ρ)ρn , where ρ = (1−p)µ .

3. As argued in part(b), we can model this feedback system as one in which

there is no feedback and whose service rate is (1 − p)µ. Since the departure
process does not depend on the actual rate of the server, it will remain a
Poisson process with the same rate as that of the arrival process.

4. Using the results of part(b), model this system as an ordinary M/M/1 queue-
ing system whose service is exponential with rate (1 − ρ)µ. Then E[s̃] =
1 1 λ
(1−p)µ 1−ρ , where ρ = (1−ρ)µ .
60 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

Exercise 3.10 For the M/M/1 queueing system,

1. find E[h̃], the expected number of customers served in busy period, and
2. find E[e−sỹ ], the Laplace-Stieltjes transform of the distribution
of the length of a busy period. Show that (d/dy)Fỹ (y) =
√ √
1/(y ρ)e−(λ+µ)y I1 (2y λµ). A Laplace transform pair,
√ √
s + 2a − s 1
√ √ ⇐⇒ e−at I1 (at),
s + 2a + s t
taken from Mathematical Tables from the Handbook of Physics and
Chemistry, will be useful in accomplishing this exercise.

Solution.
1. Let D denote the event that the first customer completes service before the
first arrival after the busy period has begun, and let A denote the comple-
ment of D. Then
E[h̃] = E[h̃|D]P {D} + E[h̃|A]P {A} .
If D occurs then clearly only the initial customer will complete service
during that busy period; that is, E[h̃|D] = 1. On the other hand, if D
occurs, then the length of the busy period has been shown to be the length
of the interarrival time plus twice the length of ỹ, a generic busy period.
Now, no service will be completed during the interarrival period, but h̃
service completions will take place during each busy period length ỹ. Thus,
E[h̃] = 1 · P {D} + E[h̃ + h̃]P {A}
µ λ
= + 2E[h̃] .
λ+µ λ+µ
Then, if ρ = λµ ,
µ 1
E[h̃] = = .
µ−λ 1−ρ

2. Condition E[e−sỹ ] on whether or not the first customer completes service

before the first arrival after the busy period has begun. Let D, D be the
same as in part (i). Then
E[e−sỹ ] = E[e−sỹ |D]P {D} + E[e−sỹ |A]P {A} .
Now,
E[e−sỹ |D] = E[e−sz̃1 ],
Elementary CTMC-Based Queueing Models 61

and

E[e−sỹ |A] = E[e−s(z̃1 +ỹ+ỹ) ]

 2
= E[e−sz̃1 ] E[e−sỹ ] .

Thus,
 2
E[e−sỹ ] = P {D} E[e−sz̃1 ] + P {A} E[e−sz̃1 ] E[e−sỹ ] .

This implies that

 2
0 = P {A} E[e−sz̃1 ] E[e−sỹ ] − E[e−sỹ ] + P {D} E[e−sz̃1 ].

Substituting the expressions for E[e−sz̃1 ], P {A}, and P {D}, and using
the quadratic formula,
q
(s + λ + µ) ± (s + λ + µ)2 − 4λµ
E[e−sỹ ] = .
2λ
We now must decide which sign gives the proper function. By the definition
of the Laplace-Stieltjes transform, E[e−sỹ ]|s=0 is simply the cumulative
distribution function of the random variable evaluated at infinity. This value
is known to be 1. Then

E[e−0·ỹ ] = 1 q
(λ + µ) ± (λ + µ)2 − 4λµ
=
p 2λ
(λ + µ) ± (λ − µ)2
=
2λ
p
(λ + µ) ± (µ − λ)2
=
2λ
(λ + µ) ± (µ − λ)
= .
2λ
This of course implies that the sign should be negative. i.e.,
q
(s + λ + µ) − (s + λ + µ)2 − 4λµ
E[e−sỹ ] = .
2λ
It remains to show that
d 1
Fỹ (y) = √ e−(λ+µ)y I1 (2y λµ).
p
dy y ρ

We first prove a lemma. Lemma: Let g(t) = e−at f (t). Then G(s) =
62 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

F (s + a).
proof:
Recall that if F (s) is the Laplace transform of f (t), then
Z ∞
F (s) = f (t)e−st dt.
0

Observe that if s is replaced by (s + a) we get

Z ∞
F (s + a) = f (t)e−(s+a)t dt
Z0 ∞
= f (t)e−at e−st dt
Z0 ∞
= g(t)e−st dt
0
= G(s)

by the definition of a Laplace transform. This proves the Lemma. Using

this Lemma, we wish to invert
q
(s + λ + µ) − (s + λ + µ)2 − 4λµ
α(s) =
2λ
given √ √
s + 2a − s 1
√ √ ⇐⇒ e−at I1 (at).
s + 2a + s t
First note
√ √
s + 2a − s
β(s) = √ √
√s + 2a + s  
√ √ √ 
s + 2a − s s + 2a − s
= √ √  √ √ 
s + 2a + s s + 2a − s
p
(s + a) − s(s + 2a)
=
a
where
1
β(s) ⇐⇒ e−at I1 (at).
t
To apply the Lemma to invert α(s), we first need to manipulate α(s) so it
is in the form of β(s). Define
p
(s + a + b) − (s + b)(s + 2a + b)
γ(s) = β(s + b) =
a
Elementary CTMC-Based Queueing Models 63
p
(s + a + b) − (s + 2a + b)2 − a2
= .
a
√
Let (a + b) = (λ + µ) and a2 = 4λµ. Then b = (λ + µ) − 2 λµ.
p
γ(s) = β(s + (λ + µ) −
q2 λµ)
(s + λ + µ) − (s + λ + µ)2 − 4λµ
= √
2 qλµ
 
1  (s + λ + µ) − (s + λ + µ)2 − 4λµ
= .
ρ 2λ

Note that
√ 1
γ(s) = α(s) ρ =⇒ α(s) = √ γ(s),
ρ
Then by applying the Lemma,
L−1 [γ(s)] = e−bt L−1 [β(s)]
1
= e−bt e−at I1 (at)
t
1 −(a+b)t
= e I1 (at).
t
By linearity, af (t) ⇐⇒ aL[f (t)].
1
L−1 [α(s)] = √ L−1 [γ(s)]
ρ
1 1 −(a+b)t
= √ e I1 (at)
ρt
1 1
√ e−(λ+µ)t I1 (2t λµ).
p
=
ρt
Or, replacing t with y,
d 1 1
Fỹ (y) = √ e−(λ+µ)y I1 (2y λµ).
p
dy ρy
This is the desired result.

Exercise 3.11 For the M/M/1 queueing system, argue that h̃ is a stop-
ping time for the sequence {x̃i , i = 1, 2, . . .} illustrated in Figure 3.5. Find
E[h̃] by using the results given above for E[ỹ] in combination with Wald’s
equation.
Solution. Recall that for h̃ to be a stopping time for a sequence of random
variables, h̃ must be independent of x̃h̃+1 . Now, h̃ describes the number of
64 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

customers served in a busy period and so x̃h̃ is the last customer served during
that particular busy period; all subsequent customers x̃h̃+j , j ≥ 1, are served in
another busy period. Hence h̃ is independent of those customers. This defines
h̃ to be a stopping time for the sequence {x̃i , i = 1, 2, · · ·}. We can thus apply
Wald’s equation  
h̃
X
E[ỹ] = E  x̃i  = E[h̃]E[x̃].
i=1

Hence,
E[ỹ]
E[h̃] =
E[x̃]
1/µ 1
 
=
1−ρ µ
1
= .
1−ρ

Exercise 3.12 For the M/M/1 queueing system, argue that E[s̃], the ex-
pected amount of time a customer spends in the system, and the expected
length of a busy period are equal. [Hint: Consider the expected waiting
time of an arbitrary customer in the M/M/1 queueing system under a non-
preemptive LCFS and then use Little’s result.]
Solution. Let B be the event that an arbitrary customer finds the system empty
upon arrival, and let I be the event that the system is found to be idle. Then
E[s̃] = E[s̃|B]P {B} + E[s̃|I]P {I} .
Now, if the system is idle with probability (1 − ρ), then the customer’s sojourn
time will just be her service time. On the other hand, if the system is busy upon
arrival, then the customer has to wait until the system is empty again to receive
service. That is, w̃ in this case is equivalent to ỹ, and so
E[s̃|B] = E[w̃] + E x̃]
1
= E[ỹ] +
µ
1
µ 1
= +
1 − ρ µ 
1 1
= +1 .
µ 1−ρ
Combining the two conditional probabilities, we see that
ρ 1 1−ρ
 
E[s̃] = +1 +
µ 1−ρ µ
Elementary CTMC-Based Queueing Models 65

1 1
 
=
µ 1−ρ
= E[ỹ].

Exercise 3.13 Let s̃LCFS denote the total amount of time an arbitrary
customer spends in the M/M/1 queueing system under a nonpreemptive
discipline. Determine the Laplace-Stieltjes transform for the distribution of
s̃LCFS .
Solution. Condition the Laplace-Stieltjes transform for the distribution of
s̃LCFS on whether or not the service is busy when an arbitrary customer ar-
rives. If the server is idle, the customer will immediately enter service and so
the Laplace-Stieltjes transform of s̃LCFS is that of x̃, the service time. If the
server is busy, however, the customer’s total time in the system will be waiting
time in the queue plus service time. It has already been shown in Exercise 3.11
that in a LCFS discipline the waiting time in queue has the same distribution
as ỹ, an arbitrary busy period. Let B denote the event that the customer finds
the server is busy, and let B c denote the event that the server is idle.
E[e−ss̃LCFS ] = E[e−ss̃LCFS |B c ]P {B c } + E[e−ss̃LCFS |B]P {B}
= (1 − ρ)E[e−sx̃ ] + ρ{E[e−sỹ ] + E[e−sx̃ ]}
= E[e−sx̃ ] + ρE[e−sỹ ] q
µ (s + λ + µ) − (s + λ + µ)2 − 4λµ
= +ρ· .
s+µ 2λ

Exercise 3.14 Determine the Laplace-Stieltjes transform for the length

of the busy period for the M/M/2 queueing system, the system having Pois-
son arrivals, exponential service, two parallel servers, and an infinite waiting
room capacity. [Hint: Condition on whether or not an arrival occurs prior to
the completion of the first service of a busy period. Then note that there is a
very close relationship between the time required to reduce the occupancy
from two customers to one customer in the M/M/2 and the length of the
busy period in the ordinary M/M/1 system.]
Solution. Let ỹ2 denote the length of a generic busy period in the M/M/1
queueing system and suppose that a customer arrives, beginning the busy pe-
riod. Condition E[e−sỹ2 ] on whether or not the customer finishes service be-
fore the next customer arrives. Let A denote the event of an arrival before
completion of service, and let D denote the event the customer completes ser-
vice before the next arrival:
E[e−sỹ2 ] = E[e−sỹ2 |D]P {D} + E[e−sỹ2 |A]P {A} .
If the original customer departs before this arrival, then the length of the busy
period is the minimum of the service time and the interarrival time. Hence we
66 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

see that

E[e−sỹ2 |D] = E[e−sz̃1 ],

µ+λ
which we know to be s+µ+λ . If the next arrival occurs before the original
customer finishes service, then there will be two customers in the system. Call
this state of the system ‘state 2’. The second server will be activated and the
overall service rate of the system will be 2µ. It will continue to be 2µ until
there is only one customer in the system again. Call this state of the system
‘state 1’. Consider f˜2,1 , the time it takes to return to state 1 from state 2. (This
is called the ‘first passage time from state 2 to state 1.’) Think of the time spent
in state 2 as an ordinary M/M/1 busy period, one in which the service rate is
2µ. Then

q
˜ (s + λ + 2µ) − (s + λ + 2µ)2 − 8λµ
E[e−sf2,1 ] = .
2λ

This follows directly from the definition of E[e−sỹ ], the Laplace-Stieltjes trans-
formof a generic busy period in the M/M/1 system, with 2µ substituted in for
the service rate. Once the system returns to state 1 again, note that this is ex-
actly the state it was in originally. Because of the Markovian properties, the
expected length of the remaining time in the busy period should be the same as
it was originally: E[e−sỹ2 ]. With these observations, we see that

˜
E[e−sỹ2 |A] = E[e−s(z̃1 +f2,1 +ỹ2 ) ],

where z̃1 represents the interarrival time between the original customer and the
first arrival after the busy period begins. Hence,

E[e−sỹ2 ] = P {D} E[e−sỹ2 |D] + P {A} E[e−sỹ2 |A]

˜
= P {D} E[e−sz̃1 ] + P {A} E[e−sz̃1 ]E[e−sf2,1 ]E[e−sỹ2 ) ].

˜
Substituting the expressions for E[e−sz̃1 ], E[e−sf2,1 ], P {A}, and P {D}, and
rearranging terms,

2µ
E[e−sỹ2 ] = q .
(s + λ) + (s + λ + 2µ)2 − 8λµ
Elementary CTMC-Based Queueing Models 67

Exercise 3.15 We have shown that the number of arrivals from a Pois-
son process with parameter λ, that occur during an exponentially distributed
service time with parameter µ, is geometrically distributed with parameter
µ/(µ + λ); that is, the probability of n arrivals during a service time is
given by [λ/(λ + µ)]n [µ/(λ + µ)]. Determine the mean length of the busy
period by conditioning on the number of arrivals that occur during the first
service time of the busy period. For example, let ñ1 denote the number of
arrivals that occur during the first service time, and start your solution with
the statement
∞
X
E[ỹ] = E[ỹ|ñ1 = n]P {ñ1 = n}.
n=0
[Hint: The arrivals segment the service period into a sequence of intervals.]
Solution. Condition E[ỹ] on the number of customers who arrive during the
service time of the original customer. Now, if no customers arrive, then the
busy period ends with the completion of the original customer’s service time.
It has already been shown that ỹ|{x̃1 < t̃1 } = z̃1 . Thus,
1
E[ỹ|ñ = 0] = E[z̃1 ] = .
µ+λ
Now suppose that exactly one new customer arrives during the original cus-
tomer’s service period. Then due to the memoryless property of the exponen-
tial distribution, this service time starts over. So the remaining time in this
busy period is equal to the length of a busy period in which there are initially
two customers present, one of whom’s busy period ends service with no new
arrivals. That is,


ỹ|{x̃1 > t̃1 } = z̃1 + ỹ + ỹ|{x̃1 < t̃1 } ,

where the first term, z̃1 , represents the interarrival time between the original
customer and the first arrival. We’ve already show above that the last term of
this sum is equivalent to z̃1 . Hence,

E[ỹ|ñ = 1] = 2E[z̃1 ] + E[ỹ].

Repeating this process, we see that if there are n arrivals during the service
period of the initial customer, then this is equivalent to the length of a busy
period that has (n+1) initial customers present: n customers we know nothing
about (and so have generic busy periods), and 1 customer whose service period
ends with no new arrivals. i.e.

E[ỹ|ñ = n] = nE[z̃1 ] + nE[ỹ] + E[z̃1 ]

= (n + 1)E[z̃1 ] + nE[ỹ].
68 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

Then, using the geometric properties of P {ñ1 = n},

∞
X
E[ỹ] = E[y|ñ1 = n]P {ñ1 = n}
n=0
∞ n
λ µ
X 
= [(n + 1)E[z̃1 ] + nE[ỹ]]
n=0
λ+µ
λ + µ
∞ n−1
µ λ λ
 X 
= E[z̃1 ] + E[ỹ] n
λ+µ λ+µ λ+µ
n=1
µ λ λ+µ 2
  
= E[z̃1 ] + E[ỹ]
λ+µ λ+µ µ
µ 1 λ λ+µ 2
  
= + E[ỹ]
λ+µ λ+µ λ+µ µ
1 λ
= + E[ỹ],
µ µ
which implies
1
µ
E[ỹ] =
1−ρ
Exercise 3.16 Suppose that the arrival and service time distributions are
memoryless, but that their rates depend upon the number of customers in
the system. Let the arrival rate when there are k customers in the system be
λk , and let the service rate when there are k customers in the system be µk .
Show that the dynamical equations are as follows:

 −(λn + µn )Pn (t) + λn−1 Pn−1 (t)



Pn′ (t) = +µn+1 Pn+1 (t), for n > 0;

λ0 P0 (t) + µ1 P1 (t) for n = 0.


Solution. Let Pn (t) denote the probability of having n customers in the system
at time t. Then
∞
X
Pn (t + h) = P {ñ(t + h) = n|ñ(t) = k}Pk (t) (3.15.1)
k=0

where ñ(t+h) = n|ñ(t) = k signifies the event of having n−k arrivals or k−n
departures in the time interval (t, t + h]. But these arrivals and departures are
Poisson, so we may apply the properties of Definition 2 of the Poisson process
(modified to take into account the possibility of deaths. In this case we may
think of Definition 2 as governing changes in state and not simply the event of
births.) Let Di , Ai denote the events of i departures and i arrivals, respectively,
in the time interval (t, t + h]. And note that for i ≥ 2, P {Di } = P {Ai } =
o(h). Hence, using general λ and µ,
P {D2 } = P {A2 } = o(h)
Elementary CTMC-Based Queueing Models 69

P {D1 } = µh + o(h).
P {A1 } = λh + o(h).
P {D0 } = 1 − µh + o(h).
P {A0 } = 1 − λh + o(h).

Then, substituting these in (3.15.1) and using state-dependent arrival and de-
parture rates,

Pn (t + h) = P {A0 } P {D0 } Pn (t)

= +P {A1 } Pn−1 (t) + P {D1 } Pn+1 (t) + o(h)
= [1 − λn h + o(h)][1 − µn h + o(h)]Pn (t)
= +[λn−1 h + o(h)]Pn−1 (t) + [µn+1 h + o(h)]Pn+1 (t) + o(h)
= [1 − (λn + µn )h]Pn (t) + λn−1 hPn−1 (t)
= +µn+1 hPn+1 (t) + o(h).

Rearranging terms and dividing by h:

Pn (t + h) + Pn (t) o(h)
= −(λn + µn )Pn (t)+λn−1 Pn−1 (t)+µn+1 Pn+1 (t)+ .
h h
Now let h → 0:

Pn′ (t) = −(λn + µn )Pn (t) + λn−1 Pn−1 (t) + µn+1 Pn+1 (t).

Finally, note that for n = 0, Pn−1 (t) = 0 = µ0 , so that

P0′ (t) = −λ0 P0 (t) + µ1 P1 (t).

Exercise 3.17 Prove Theorem 3.5.

Solution. Let Q be a (K +1)-dimensional square matrix having distinct eigen-
values σ0 , σ1 , · · · , σK , and define

W (σ) = (σI − Q) .

Then, for detW (σ) 6= 0, (i.e., for σ 6= σi , i = 1, 2, · · · , K), we find

W (σ)W −1 (σ) = I.

Since W −1 (σ) = adj W (σ)/det W (σ),

adj W (σ)W (σ) = det W (σ)I.

Now, W (σ), adj W (σ), and det W (σ) are all continuous functions of σ.
Therefore,
lim adj W (σ)W (σ) = lim detW (σ)I,
σ→σi σ→σi
70 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

so that, since lim detW (σ) = 0,

σ→σi

adj W (σi )W (σi ) = O(K+1)×(K+1) ,

where On×m denotes an n × m matrix of zeroes. Thus, we have

adj (σi I − Q) (σi I − Q) = O(K+1)×(K+1) .

Now, by definition, X is a left eigenvector of Q corresponding to σi if σi X =

XQ or X(σi I − Q) = 0. Therefore, every row of adj (σi I − Q) is a left
eigenvector of Q corresponding to σi . Hence Mi is proportional to the rows
of adj (σi I − Q).

Exercise 3.18 Prove Theorem 3.6.

Solution. By Exercise 3.16, since the rows of adj (σi I − Q) are proportional
to Mi , the left eigenvector of Q corresponding to σi , they must be proportional
to each other. Similary, we may use the same technique of Exercise 3.16 to
show that the columns of adj (σi I − Q) are proportional to each other. Since
W (σ)W −1 (σ) = I, and W −1 (σ) = adj W (σ)/det W (σ), we find that

W (σ)adj W (σ) = det W (σ)I.

By the continuity of W (σ), adj W (σ), and det W (σ),

lim W (σ)adj W (σ) = lim detW (σ)I.

σ→σi σ→σi

This implies, since lim detW (σ) = 0, that

σ→σi

adj W (σi )W (σi ) = O(K+1)×(K+1) ,

where On×m denotes an n × m matrix of zeroes. Thus, we have

(σi I − Q) adj (σi I − Q) = O(K+1)×(K+1) .

Now, by definition, X is an eigenvector of Q corresponding to σi if σi X =

QX or (σi I − Q)X = 0. Therefore, every column of adj (σi I − Q) is an
eigenvector of Q corresponding to σi . Hence the rows of adj (σi I − Q) are
proportional to each other.
Elementary CTMC-Based Queueing Models 71

Exercise 3.19 Let K = 1. Use Definition 2 of the Poisson process to

write an equation of the form
d
[ P0 (t) P1 (t) ] = [ P0 (t) P1 (t) ] Q.
dt
Show that the eigenvalues of the matrix Q are real and nonnegative. Solve
the equation for P0 (t), P1 (t) and show that they converge to the solution
given in Example 3.2 regardless of the values P0 (0), P1 (0). [Hint: First,
do a similarity transformation on the matrix Q, which converts the matrix
to a symmetric matrix Q̂. Then show that the matrix Q̂ is negative semi-
definite.]
Solution. As seen in the proof of Exercise 3.15, we can use Definition 2 of the
Poisson process to write the dynamical equations as:
P0′ (t) = −λ0 P0 (t) + µ1 P1 (t)
P1′ (t) = −µn P1 (t) + λ0 P0 (t).
Then, using the definition
√
q̂i,i+1 = qi,i+1 qi+1,i for i = 0, 1, · · · , K − 1,

we see that Q̂ is as follows:

√
−λ
 
√ µλ
µλ −µ
It is easily seen that the eigenvalues of Q̂ are σ0 = 0 and σ1 = −(λ + µ).
Then, using the equation
 
P (t) = P (0) M diag eσ0 t , eσ1 t , · · · , eσK t M−1 ,

and arbitrary P0 (0) and P1 (0), we see that

" µ λ
#
−λ
 
1 1 0 µ+λ µ+λ
P (t) = P (0) .
1 µ 0 e−(λ+µ)t −1
µ+λ
1
µ+λ

This implies that

µ e−(λ+µ)t
P0 (t) = [P0 (0) + P1 (0)] + [λP0 (0) − µP1 (0)] ,
λ+µ λ+µ
λ e−(λ+µ)t
P1 (t) = [P0 (0) + P1 (0)] + [−λP0 (0) + µP1 (0)].
λ+µ λ+µ
Note that P0 (0) + P1 (0) = 1, and then let t → ∞:
µ
P0 (t) = ,
λ+µ
72 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

µ
P1 (t) = .
λ+µ
Since P0 (0) and P1 (0) were arbitrary, we see that P0 , P1 converge to the solu-
tion given in the example regardless of the values P0 (0), P1 (0).
Exercise 3.20 For the specific example given here show that the equi-
librium probabilities for the embedded Markov chain are the same as those
for the continuous-time Markov chain.
Solution. As already shown, the probability matrix for the discrete-time em-
bedded Markov chain is
1−ρ
 
ρ
Pe = .
1 0
Thus, using the equation πe = πe Pe ,
1
πe 0 =
1+ρ
ρ
πe 1 =
1+ρ
Recall that in the original system, the continuous-time Markov chain was
 
0 1
PO = .
1 0
Now, however, the system can transition from one state back into itself. The
one-step transition probability matrix is then
1−ρ
 
1
PC = .
1 0
Since this is precisely the probability matrix for the discrete-time embedded
Markov chain, we see that
1
πc 0 = πe 0 =
1+ρ
ρ
πc 1 = πe 1 = .
1+ρ
Furthermore, by the very definition of being uniformized, the mean occupancy
times for each state on each entry are equal. That is, E[s̃0 ] = E[s̃1 ] = 1/ν.
This implies
πci E[s̃i ]
PCi = 1
P
πci E[s̃i ]
i=0
Elementary CTMC-Based Queueing Models 73

πci · ν1
=
πc0 · ν1 + πc1 · 1
ν
πc i
=
πc 0 + πc 1
= πc i
= πe i .

Thus we see that the equilibrium probabilities for the embedded discrete-time
Markov chain are equal to the equilibrium probabilities for the continuous-time
Markov chain when we randomize the system.
Exercise 3.21 Show that the equilibrium probabilities for the embedded
Markov chain underlying the continuous-time Markov chain are equal to
the equilibrium probabilities for the continuous-time Markov chain.
Solution. Let the capacity of the finite M/M/1 queueing system be K, K>1.
In this general case, we must take ν to be at least as large as λ + µ, the rate
leaving states 1, 2, · · · , K − 1. Then
 ν−λ λ
0 0 ... 0 0

ν ν
µ ν−(λ+µ) λ

 ν ν ν 0 ... 0 0 

 µ ν−(λ+µ) λ

Pe = 
 0 ν ν ν ... 0 0 .

 .. .. .. .. 
 . . . . ... 0 0 
µ ν−µ
0 0 0 0 ... ν ν

This produces the solution

1−ρ
πe 0 =
1 − ρK+1
λ
π e i = ρi π e 0 , ρ=
µ
By the same reasoning as in the proof of Exercise 3.19, the one-step transition
probability matrix PC is equal to Pe . Hence for i = 0, 1, · · · , K, πei = πci .
And, since the system has been uniformized, the expected amount of time spent
in each state before transition is unity. Hence,
πci E[s̃i ]
PCi = ∞
P
πci E[s̃i ]
i=0
πc i · 1
= ∞
πc i · 1
P
i=0
= πc i
= πe i .
74 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

Exercise 3.22 For the special case of the finite capacity M/M/1 queue-
ing system with K = 2, λ0 = λ1 = 0.8, and µ1 = µ2 = 1,
determine the time-dependent state probabilities by first solving the dif-
ferential equation (3.32) directly and then using uniformization for t =
0.0, 0.2, 0.4, . . . , 1.8, 2.0 with P0 (0) = 1, plotting the results for P0 (t),
P1 (t), and P2 (t). Compare the quality of the results and the relative diffi-
culty of obtaining the numbers.
Solution. For K = 2, the matrix Q is:
−λ λ 0
 

Q =  µ −λ − µ λ 
 0 µ −µ 
−0.8 0.8 0
=  1.0 −1.8 0.8  .
0 1.0 −1.0
The eigenvalues of Q are found to be 0, −.9055728, and −2.6944272, and their
corresponding eigenvectors are proportional to [ 1 1 1 ]T , [ −7.5777088 1 10.5901699 ]T ,
and [ −0.42229124 1 −0.5901699 ]T , respectively. Thus, we find

P (t) = [ P0 (0) P1 (0) P2 (0) ] MeΛt M−1 ,

where
1 −7.5777088 −0.42229124
 

M = 1 1 1 
 1 10.5901699 −0.5901699
1 0 0


eΛt =  0 e−0.9055728t 0 
−2.6944272t
0 0 e
0.4098361 0.3278688 0.2622951


M−1 =  −0.0582906 0.0061539 0.0521367  .

−0.3515454 0.6659772 −0.3144318
The time dependent state probabilities are then

P0 (t) = 0.4098361 + 0.4417094e−0.9055728t + 0.1484545e−2.6944272t

P1 (t) = 0.3278688 + 0.0466325e−0.9055728t − 0.2812364e−2.6944272t
P2 (t) = 0.2622951 − 0.3950769e−0.9055728t + 0.1327818e−2.6944272t .

To complete the exercise, Equation 3.48 was programmed on a computer for

t = 0.0, 0.2, 0.4, · · · , 2.0. The resulting values of P0 (t), P1 (t), and P2 (t) are
shown in Figure 3.1.
For K = 2, it was easier to solve for the equilibrium probabilities by hand.
For larger values of K, however, this would become increasingly difficult. The
Elementary CTMC-Based Queueing Models 75

1.0

0.8
Time dependent state probabilities

P_0(t)
0.6

0.4

P_1(t)

0.2

P_2(t)

0.0
0.2 0.4 0.6 0.8 1.0 1.2 1.4 1.6 1.8 2.0
Time (sec)

Figure 3.1. Time dependent state probabilities for Exercise 3.22

quality of the computer results differed with the values of t and ǫ, where ǫ de-
termines when the summation in Equation 3.48 may be stopped. As t, depedent
on ǫ, increased to threshold value, the equilibrium probabilities improved. And
as ǫ decreased, the values of t were allowed to get larger and larger, bringing
the equilibrium probabilities closer to those determined by hand.
Exercise 3.23 For the special case of the finite-capacity M/M/1 system,
show that for K = 1, 2, . . . ,
ρPB (K − 1)
PB (K) = ,
1 + ρPB (K − 1)

where PB (0) = 1.
Solution. By Equation (3.60), if the finite capacity is K,
1−ρ
 
PB (K) = ρK .
1 − ρK+1
76 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

It follows that if the finite capacity is K − 1,

1−ρ
 
PB (K − 1) = ρK−1 .
1 − ρK
That is,
(1 − ρ)ρK−1 = PB (K − 1)(1 − ρK ).
So
1−ρ
 
PB (K) = ρρK−1
1 − ρK+1
ρPB (K − 1)(1 − ρK )
= .
1 − ρK+1
Furthermore,
1 − ρK 1 − ρK
=
1 − ρK+1 1 − ρK + ρK − ρK+1
1 − ρK
=
1 − ρK + ρK (1 − ρ)
1
= ρρK−1 (1−ρ)
1+ 1−ρK
1
= .
1 + ρPB (K − 1)
Hence,
ρPB (K − 1)
PB (K) = .
1 + ρPB (K − 1)

Exercise 3.24 For the finite-state general birth-death process, show that
for K = 1, 2, . . . ,
(λK /µK )PB (K − 1)
PB (K) = ,
1 + (λK /µK )PB (K − 1)

where PB (0) = 1.
Solution. If the finite capacity of the M/M/1 queueing system is K, then by
combining Equations (3.56) and (3.58),
K−1
 
Q
λi 
1
PK =  n−1   i=0 .
 
Q K
 Q 
K  λi µi

1+
P  i=0
n
 i=1
Q
n=1

µi
i=1
Elementary CTMC-Based Queueing Models 77

Similarly if the finite capacity is K − 1,

K−2
 
Q
λi 
1 
PK−1 =  n−1   i=0 .
 
Q  K−1 
λi
Q
K−1
P   µi
1+  i=0
n
 i=1
Q
n=1

µi
i=1

Now, note that the numerator of PK is equal to

K−2
Q
λi
λK−1 i=0
( ) ,
µK K−1
Q
µi
i=1

and the denominator of PK is

n−1 K−2
 
Q
λi 
Q
K−1 λi
X  i=0 λK−1 i=0
1+ +( ) .

 n µK K−1
 Q
n=1 µi
 Q
µi
i=1 i=1

Dividing both numerator and denominator by

n−1
 
Q
K−1
X λi 
 i=0
1+ ,

 n
 Q
n=1 µi

i=1

we see that the numerator becomes

K−2
Q
λi
λ i=0
( µK−1
K
) K−1
Q
µi
i=1
n−1
,
Q
K−1 λi
P  i=0

1+ n
Q 

n=1 µi
i=1
78 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

which is (λK−1 /µK )PK−1 . And the denominator becomes

 K−2 
Q
 i=0 λi 
 
 K−1
Q 
µi
λK−1 i=1
1+( )  n−1  ,
µK Q
λ
P  i=0 i 
K−1
1+  n
Q 

n=1 µi
i=1

which is 1 + (λK−1 /µK )PK−1 . Thus,

(λK−1 /µK )PB (K − 1)
PB (K) = ,
1 + (λK−1 /µK )PB (K − 1)
where PB (0) = 1.
Exercise 3.25 Let K be arbitrary. Use the balance equation approach to
write an equation of the form
d
P (t) = P (t)Q
dt
where P (t) = [ P0 (t) P1 (t) · · · PK (t) ]. Show that the eigenvalues of
the matrix Q are real and nonpositive.
d
Solution. Observe that dt P (t) is {the rate of arrivals to the system at time t}
minus {the rate of departures from the system at time t}. Therefore, using the
table in Section 3.4 of the text, we may write
µ1 P1 (t) − λ0 P0 (t)
 
 λ0 P0 (t) + µ2 P2 (t) − (λ1 + µ1 ) P1 (t) 
d 
..

P (t) = 
 
dt  . 

 λK−2 PK−2 (t) + µK PK (t) − (λK−1 + µK−1 ) PK−1 (t) 
λK−1 PK−1 (t) − µK PK (t)
d
We may rewrite this as dt P (t) = P (t)Q, where
−λ0 λ0 0
 
 µ1
 −(λ1 + µ1 ) λ1 

Q=
 .. .

 . 
 µK−1 −(λK−1 + µK−1 ) λK−1 
0 µK −µK
Now, Q is a diagonal, real matrix, whose main diagonal is negative and whose
off-diagonal elements are all of the same sign (positive). Hence, its eigenvalues
are all real and nonpositive.
Elementary CTMC-Based Queueing Models 79

Exercise 3.26 Using the concept of local balance, write and solve the
balance equations for the general birth-death process shown in Figure 3.12.
Solution. In Figure 3.12, the initial boundary is between the nodes represent-
ing state 1 and state 2. We see that the rate going from the left side of the
boundary into the right side (that is, from state 1 into state 2) is λ1 . The rate
going from the right side of the boundary into the left side (from state 2 into
state 1) is µ2 . Since the ‘rate in’ must equal the ‘rate out’, we must be have

λ1 P1 = µ2 P2
λ1
i.e. P2 = P1 .
µ2
Moving the boundary between state 0 and state 1, the local balance is
λ0
P1 = P0 .
µ1
Substituting this into the earlier result for P2 , we see that
λ1 λ0
P2 = P0 .
µ2 µ1
Repeating this process by moving the boundary between every pair of states,
we get the general result Pn−1
λi
Pn = Pi=0n P0 .
i=1 µi

Exercise 3.27 Prove (3.66).

Solution. 
dn 
E[x̃(x̃ − 1) · · · (x̃ − n + 1)] = n Fx̃ (z) (3.66)

dz 
z=1

Since Fx̃ (z) = E[z x̃ ], we may differentiate each side with respect to z:
d d
Fx̃ (z) = E[z x̃ ]
dz dz Z
d
= z x dFx̃ (x)
Zdz x
= xz x−1 dFx̃ (x)
x
= E[x̃z x̃−1 ].

Hence,
d
lim Fx̃ (z) = lim E[x̃z x̃−1 ]
z→1 dz z→1
80 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

= E[x̃].

Now use induction on n, assuming the result holds for n − 1. That is, assume
that
dn−1 h
x̃−n+1
i
F x̃ (z) = E x̃(x̃ − 1) · · · (x̃ − n + 2)z .
dz n−1
Then, differentiating this expression with respect to z,
dn d
Z
Fx̃ (z) = x(x − 1) · · · (x − n + 2)z x−n+1 dFx̃ (x)
dz n dz x
Z
= x(x − 1) · · · (x − n + 2)(x − n + 1)z x−n dFx̃ (x)
xh i
= E x̃(x̃ − 1) · · · (x̃ − n + 1)z x̃−n .

Thus,
dn h
x̃−n
i
lim Fx̃ (z) = lim E x̃(x̃ − 1) · · · (x̃ − n + 1)z
z→1 dz n z→1
= E[x̃(x̃ − 1) · · · (x̃ − n + 1)].

Exercise 3.28 Prove (3.67).

Solution. 
1 dn 
P {x̃ = n} = F (z) . (3.67)

x̃
n! dz n


z=0
By definition of expected value,

Fx̃ (z) = E[z x̃ ]

∞
z n P {x̃ = i}.
X
=
i=0

On the other hand, the Maclaurin series expansion for Fx̃ (z) is as follows:
∞

1 dn 
zn.
X
Fx̃ (z) = F (z)

x̃
n! dz n


n=0 z=0

By the uniqueness of power series, we may match coefficients of z n . The result

follows.
Exercise 3.29 Use (3.66) to find E[ñ] and E[ñ2 ].
Solution. Use equation (3.66) with n = 1:

d 
E[ñ(ñ − 1) · · · (ñ − 1 + 1)] = E[ñ] = Fñ (z)

dz 
z=1
Elementary CTMC-Based Queueing Models 81

d P0 
=
dz (1 − ρz) z=1

ρ
= .
1−ρ

To solve for E[ñ2 ], let n = 2:


2 d2 
E[ñ(ñ − 1) · · · (ñ − 2 + 1)] = E[ñ − ñ] = 2 Fx̃ (z)

dz 
z=1
2ρ 2
= .
(1 − ρ)2

But E[ñ2 − ñ] = E[ñ2 ] − E[ñ], so after a little arithmetic we find that

ρ2 + ρ
E[ñ2 ] = .
(1 − ρ)2
82 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

3.1 Supplementary Problems

3-1 Messages arrive to a statistical multiplexing system according to a Poisson
process having rate λ. Message lengths, denoted by m̃, are specified in
octets, groups of 8 bits, and are drawn from an exponential distribution
having mean 1/µ. Messages are multiplexed onto a single trunk having a
transmission capacity of C bits per second according to a FCFS discipline.

(a) Let x̃ denote the time required for transmission of a message over the
trunk. Show that x̃ has the exponential distribution with parameter
µC/8.
(b) Let E[m̃] = 128 octets and C = 56 kilobits per second (kb/s). Deter-
mine λmax , the maximum message-carrying capacity of the trunk.
(c) Let ñ denote the number of messages in the system in stochastic equi-
librium. Under the conditions of (b), determine P {ñ > n} as a func-
tion of λ. Determine the maximum value of λ such that P {ñ > 50} <
10−2 .
(d) For the value of λ determined in part (c), determine the minimum
value of s such that P {s̃ > s} < 10−2 , where s̃ is the total amount
of time a message spends in the system.
(e) Using the value of λ obtained in part (c), determine the maximum
value of K, the system capacity, such that PB (K) < 10−2 .

Solution:

(a) Since trunk capacity is C, the time to transmit a message of length m

bytes is x = 8m/C. Therefore, x̃ = 8m̃/C, or m̃ = x̃C/8. Thus,
8m̃
 
P {x̃ ≤ x} = P ≤x
 C
xC

= P m̃ ≤
8
−µC
x
= 1−e 8 ,

where the last equation follows from the fact that x̃ is exponential
with parameter µ. Therefore, m̃ is exponential with parameter µC/8.
1
(b) Since E[m̃] = 128 octets, µ = 128 . Therefore, x̃ is exponentially
µC 7 7
distributed with parameter 8 = 128 Kbps, or 128 × 103 bps. Then
E[x̃] = 128
7 × 10
−3 sec. Since λ
max E[x̃] = ρ < 1,

1
λ <
E[x̃]
Elementary CTMC-Based Queueing Models 83

7
= × 103
128
= 54.6875 msg/sec.

(c) We know from (3.10) that P {ñ > n} = ρn+1 . Thus, P {ñ > 50} =
ρ51 . Now, ρ51 < 10−2 implies

51 log 10 ρ < −2.

i.e.,
2
log10 ρ < − .
51
2
Hence, ρ = λE[x̃] < 10− 51 = 0.91366, so that

λ < 49.966 msg/sec.

(d) Now, s̃ is exponential with parameter µ(1 − ρ), so P {s̃ > x} =

e−µ(1−ρ)x . Then for P {s̃ > s} < 10−2 , we must have

e−µ(1−ρ)s < 10−2

or
−µ [1 − ρ] s < −2 (ln 10)
i.e.,
2 (ln 10)
s >
µ (1 − ρ)
2 (ln 10)
= 7 3
128 × 10 (1 − 0.91366)
4.60517
=
4.7217
= 0.9753 sec
= 975.3 ms

(e) Recall the equation for PB (K):

1−ρ
 
PB (K) = ρK
1 − ρK+1

We wish to find K such that PB (K) < 10−2 . First, we set PB (K) =
10−2 and solve for K.
1−ρ
 
ρK = 10−2
1 − ρK+1
84 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL
 
(1 − ρ) ρK = 10−2 1 − ρK+1
= 10−2 − 10−2 ρK+1
K −2 K+1
(1 − ρ) ρ + 10 ρ = 10−2
1 − ρ + 10−2 ρ ρK = 10−2
10−2
ρK =
1 − 0.99ρ

Therefore,
−2 (ln 10) − ln (1 − 0.99ρ)
K =
ln ρ
−[2 (ln 10) + ln (1 − 0.99ρ)]
=
ln ρ

For ρ = 0.91366,

−[4.60517 − 2.348874]
K =
−0.09029676
2.256296
=
0.09029676
= 25.799

Therefore, for a blocking probability less than 10−2 , we need K ≥

26. From part (c), note that for the non-blocking system, the value
of K such that P {ñ > K} < 10−2 is 50 for this particular value of
λ. Thus, it is seen that the buffer size required to achieve a given
blocking probability cannot be obtained directly from the survivor
function. In this case, approximating buffer requirements from the
survivor function would have resulted in K = 50, but in reality, only
26 storage spots are needed.

3-2 A finite population, K, of users attached to a statistical multiplexing sys-

tem operate in a continuous cycle of think, wait, service. During the think
phase, the length of which is denoted by t̃, the user generates a message.
The message then waits in a queue behind any other messages, if any, that
may be awaiting transmission. Upon reaching the head of the queue, the
user receives service and the corresponding message is transmitted over a
communication channel. Message service times, x̃, and think times, t̃, are
drawn from exponential distributions with rates µ and λ, respectively. Let
the state of the system be defined as the total number of users waiting and
in service and be denoted by ñ.
Elementary CTMC-Based Queueing Models 85

(a) The first passage time from state i to state i − 1 is the total amount
of time the systems spends in all states from the time it first enters
state i until it makes its first transition to the state i − 1. Let s̃i denote
the total cumulative time the system spends in state i during the first
passage time from state i to state i − 1. Determine the distribution of
s̃i .
(b) Determine the distribution of the number of visits from state i to state
i + 1 during the first passage time from state i to i − 1.
(c) Show that E[ỹK ], the expected length of a busy period, is given by
the following recursion:
1 
1 + λ(K − 1)E [ỹK−1 ]
E[ỹK ] = with E [ỹ0 ] = 0.
µ
[Hint: Use the distribution found in part (b) in combination with the
result of part (a) as part of the proof.]
(d) Let P0 (K) denote the stochastic equilibrium probability that the com-
munication channel is idle. Determine P0 (K) using ordinary birth-
death process analysis.
(e) Let E[ı̃K ] denote the expected length of the idle period for the com-
munication channel. Verify that P0 (K) is given by the ratio of the
expected length of the idle period to the sum of the expected lengths
of the idle and busy periods; that is,
E[ı̃K ]
P0 (K) =
E[ı̃K ] + E[ỹK ]
which can be determined iteratively by
1
P0 (K) = .
1 + [(Kλ)/µ] {1 + (K − 1)λ E[ỹK−1 ]}
That is, show that P0 (K) computed by the formula just stated is iden-
tical to that obtained in part (d).

Kλ (K-1) λ (K-2) λ 2λ λ
0 1 2 • • • K-1 K

Figure 3.2. State diagram for Supplemental Exercise 3.2

Solution:
86 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

(a) The state diagram is as in Figure 3.2. For a typical state, say i, we
wish to know the total amount of time spent in state i before the sys-
tem transitions from state i to i − 1 for the first time, i = 1, 2, · · · , K.
Now, on each visit to state i, the time spent in state i is the minimum
of two exponential random variables: the time to first arrival, which
has parameter (K − i)λ, and the time to first departure, which has
parameter µ. Thus, the time spent in state i on each visit is exponen-
tially distributed with parameter (K − i)λ + µ. Now, the time spent in
state i on the j −th visit is independent of everything, and in particular
on the number of visits to state i before transitioning to state i−1. Let
s̃ij denote the time spent in state i on the j −th visit and ṽi denote the
number of visits to state i during the first passage from i to i−1. Then
Pṽi
s̃i = j=1 s̃ij . Now, the number of visits to state i is geometrically
distributed (see proof of exponential random variables, page 35 of the
text) with parameter µ/ [(K − i)λ + µ] because the probability that a
given visit is the last visit is the same as the probability of a service
completion before arrival (see Proposition 3 of exponential random
variables, page 35 of the text) and this probability is independent of
the number of visits that have occurred up to this point. Thus, s̃i is the
geometric sum of exponential random variables and is therefore expo-
µ
nentially distributed with parameter [(K − i)λ + µ] (K−i)λ+µ = µ,
which follows directly from Property 5 of exponential random vari-
ables, given on page 35 of the text. In summary, the total amount of
time spent in any state before transitioning to the next lower state is
exponentially distributed with parameter µ.

(b) From arguments similar to those of part (a), the number of visits to
state i + 1 from state i is also geometric. In this case,
k
µ (K − i)λ

P {ṽi = k} =
(K − i)λ + µ (K − i)λ + µ
so that the mean number of visits is
(K − i)λ + µ (K − i)λ
−1= for i = 1, 2, · · · , K − 1.
µ µ
For i = K, the number of visits is exactly 1.

(c) First note that ỹK−i , the busy period for a system having K − i cus-
tomers, has the same distribution as the first passage time from state
i + 1 to i. From part (b), the expected minimum number of visits to
state i + 1 from state i is (K − i) µλ . Thus, the total time spent in states
i + 1 and above during a first passage time from state i to i − 1 is
Elementary CTMC-Based Queueing Models 87

(K − i) µλ E[ỹK−i ]. Also, the total time spent in state i is 1/µ, so that

the first passage time from state i to i − 1 is given by
1 λ
E[ỹK−(i−1) ] = + (K − i) E[ỹK−i ]
µ µ
1
= (1 + (K − i)λE[ỹK−i ]) .
µ
With i = 1, we have
1
E[ỹK ] = (1 + (K − 1)λE[ỹK−1 ]) .
µ
We then have
1
E[ỹ1 ] =
µ
1
E[ỹ2 ] = (1 + 1 · λE[ỹ1 ])
µ
1 λ

= 1+
µ µ
1
E[ỹ3 ] = (1 + 2λE[ỹ2 ])
µ
1 λ λ
 
= 1+2 1+
µ µ µ

so that we may compute E[ỹK ] for any value of K by using the re-
cursion, starting with n = 1,
1
E[ỹn ] = (1 + (K − 1)λE[ỹn−1 ]) ,
µ
with E[ỹ0 ] = 0, and iterating to K.
(d) Using ordinary birth-death analysis, we find from (3.58) that
1
P0 =   
P∞ Qn−1 Qn
1+ n=1 i=0 λi i=1 µi
1
=   .
PK Qn−1 Qn
1+ n=1 i=0 λi i=1 µi

Now, µi = µ and λi = (K − i)λ for this particular case. Thus,

1
P0 =   
PK Qn−1 Qn
1+ n=1 i=0 (K − i)λ i=1 µ
88 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

1
= PK  n Q .
λ n−1
1+ n=1 µ i=0 (K − i)
Qn−1 K!
But i=0 (K − i) = K(K − 1) · · · K − (n − 1) = (K−n)! . So

1
P0 = PK  n .
λ K!
1+ n=1 µ (K−n)!

(e) The expression given for P0 (K) is readily converted to the final form
given as follows:
1
Kλ
P0 (K) = 1
Kλ + E[ỹK ]
1
Kλ
=  
1 1
Kλ + µ 1 + λ(K − 1)E[ỹ( K − 1)]
1
=  
Kλ
1+ µ 1 + λ(K − 1)E[ỹ( K − 1)]

We now prove the proposition

K−1  n #
"
1 X (K − 1)! λ
E[ỹK ] = 1+
µ n=1
(K − 1 − n)! µ

proof: Let T denote the truth set for the proposition. With K = 1, we
find E[ỹ1 ] = µ1 , which is clearly correct. Now, suppose (K − 1) ∈ T .
Then, by hypothesis,
K−2
"  n #
1 X (K − 2)! λ
E[ỹK−1 ] = 1+ (∗)
µ n=1
(K − 2 − n)! µ

and, from part (c),

1
E[ỹK ] = [1 + (K − 1)λE[ỹK−1 ]]
µ
which is always true. Substituting (∗) into this last expression gives
K−2  n #!#
" "
1 1 X (K − 2)! λ
E[ỹK ] = 1 + λ(K − 1) 1+
µ µ n=1
(K − 2 − n)! µ
(K − 1) K−1
"  n #
1 X (K − 1)(K − 2)! λ
= 1+λ +
µ µ n=2
(K − 1 − n)! µ
Elementary CTMC-Based Queueing Models 89

K−1
"  n #
1 X (K − 1)! λ
= 1+
µ n=1
(K − 1 − n)! µ

This proves the proposition.

We now substitute the expression for E[ỹK−1 ] into the given expres-
sion
−1
K

P0 (K) = 1+λ (1 + (K − 1)λE[ỹK−1 ])
µ
" K−1  n !#−1
K X (K − 1)! λ
= 1+λ 1+
µ n=1
(K − 1 − n)! µ
" K  n #−1
X K! λ
= 1+ ,
n=1
(K − n)! µ

which is identical to that obtained in part (d).

3-3 Traffic engineering with finite population. Ten students in a certain grad-
uate program share an office that has four telephones. The students are
always busy doing one of two activities: doing queueing homework (work
state) or using the telephone (service state); no other activities are allowed
- ever. Each student operates continuously as follows: the student is ini-
tially in the work state for an exponential, rate β, period of time. The stu-
dent then attempts to use one of the telephones. If all telephones are busy,
then the student is blocked and returns immediately to the work state. If a
telephone is available, the student uses the telephone for a length of time
drawn from an exponential distribution with rate µ and then returns to the
work state.

(a) Define an appropriate state space for this service system.

(b) Draw a state diagram for this system showing all transition rates.
(c) Write the balance equations for the system.
(d) Specify a method of computing the ergodic blocking probability for
the system - that is the proportion of attempts to join the service sys-
tem that will be blocked - in terms of the system parameters and the
ergodic state probabilities.
(e) Specify a formula to compute the average call generation rate.
(f) Let µ = 1/3 calls per minute; that is, call holding times have a mean
of three minutes. Compute the call blocking probability as a function
of β for β ∈ (0, 30).
90 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

(g) Compare the results of part (f) to those of the Erlang loss system hav-
ing 4 servers and total offered traffic equal to that of part (f). That
is, for each value of β, there is a total offered traffic rate for the sys-
tem specified in this problem. Use this total offered traffic to obtain a
value of λ, and then obtain the blocking probability that would result
in the Erlang loss system, and plot this result on the same graph as the
results obtained in (f). Then compare the results.

Solution.
(a) Define the state of the system as the number of students using the
telephones. This can have values 0, 1, 2, 3, 4.

(b) See Figure 3.3.

10ß 9ß 8ß 7ß
0 1 2 3 4

1µ 2µ 3µ 4µ

Figure 3.3. State diagram for Supplemental Exercise 3.3

(c) The balance equations can be either local or global; we choose local.

10βP0 = µP1
9βP1 = 2µP2
8βP1 = 3µP3
7βP3 = 4µP4 .

(d) All arrivals that occur while the system is in state 4 are blocked. The
arrival rate while in state 4 is 6β. Over a long period of time T , the
number of blocked attempts is 6βP4 T while the total number of ar-
rivals is (10βP0 + 9βP1 + 8βP2 + 7βP3 + 6βP4 ) T . Taking ratios,
we find that the proportion of blocked calls is
6P4
.
10P0 + 9P1 + 8P2 + 7P3 + 6P4

(e) The average call generation rate is 10P0 β + 9P1 β + 8P2 β + 7P3 β +
6P4 β.
Elementary CTMC-Based Queueing Models 91

(f) First we solve the balance equations.

10β 10 β
 
P1 = P0 = P
µ 1 µ 0
   2
9β 10 β
P2 = P1 = P0
2µ 2 µ
   3
8β 10 β
P3 = P2 = P0
3µ 3 µ
   4
7β 10 β
P4 = P3 = P0
4µ 4 µ
.
So,
4    i
X 10 β
1 = P0 .
i µ
i=0
i.e.,
1
P0 =  
P4 10  β i
i=0 i µ

Therefore,
10 βµ
P1 =  
P4 10  β i
i=0 µ
 i  
10 β i
i µ
Pi =    for i = 0, 1, 2, 3, 4
P4 10 β j
j=0 j µ

The call generation rate is then

4
X
λ = (10 − i)βPi
i=0   
P4 10 iβ
β i=0 (10 − i) i µ
=    .
P4 10 j
β
j=0 j µ

And the call blocking probability is 6P4 βλ . Hence,

  
10 4 β
6β µ
4
PB =   
P4 10 βi
β i=0 (10 − i) µ
i
92 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL
 
10
6
4
=    .
P4 10 4−i
µ
i=0 (10 − i) i β

With µ = 13 , we have
 
10
(3β)i
P4
β i=0 (10 − i)
i
λ(β) =  
10
(3β)j
P4
j=0 j
and  
10
6 (3β)4
4
PB (β) =   .
10
(3β)j
P4
j=0 (10 − j) j

(g) From (3.64) and (3.65), we have

λ
a= = 3λ
µ
and
 
a4
4!
B(4, a) =  
P4 ai
i=0 i!
(3λ)4
 
4!
= P4  (3λ)i 
i=0 i!
(3λ)4
= h
(3λ)2 3 4
i
4! 1 + 3λ + + (3λ) (3λ)
3! + 4!
2!
(3λ)4
=
24 + (24 + [12 + (4 + 3λ)3λ]3λ)3λ
The idea is to compute λ first and then compute β(4, λ) using this
formula. It can be seen that for the Erlang blocking formula, a smooth
traffic assumption would result in a substantial underestimation of the
actual blocking probability. This will always be true for relatively
small numbers of users. As an example, if β = 1, then the actual
blocking probability is approximately 0.1, but the Erland blocking
formula would yield 0.00025, which is off by a factor of 400.
Elementary CTMC-Based Queueing Models 93

3-4 A company has six employees who use a leased line to access a database.
Each employee has a think time which is exponentially distributed with
parameter λ. Upon completion of the think time, the employee needs
the database and joins a queue along with other employees who may be
waiting for the leased line to access the database. Holding times are ex-
ponentially distributed with parameter µ. When the number of waiting
employees reaches a level 2, use of an auxiliary line is authorized. The
time required for the employee to obtain the authorization is exponentially
distributed with rate τ . If the authorization is completed when there are
less than three employees waiting or if the number of employees waiting
drops below two at any time while the extra line is in use, the extra line is
immediately disconnected.

(a) Argue that the set {0, 1, 2, 3, 3r, 3a, 4r, 4r, 5r, 5a, 6r, 6a}, where the
numbers indicate the number of employees waiting and in service, the
letter r indicates that authorization has been requested, and the letter
a indicates that the auxiliary line is actually available for service, is a
suitable state space for this process.
(b) The situation in state 4r is that there are employees waiting and in
service and an authorization has been requested. With the process in
state 4r at time t0 , list the events that would cause a change in the state
of the process.
(c) Compute the probability that each of the possible events listed in part
(b) would actually cause the change of state, and specify the new state
of the process following the event.
(d) What is the distribution of the amount of time the system spends in
state 4r on each visit? Explain.
(e) Draw the state transition rate diagram.
(f) Write the balance equations for the system.

Solution.

(a) Two waiting employees means that there are three employees waiting
and using the system. At a point in time, if an employee requests a
line, then there is a waiting period before the line is authorized. The
idea is that the system behaves differently when the number of wait-
ing persons is increasing than when the number of waiting persons is
decreasing. When the third person arrives, the system is immediately
in the request mode 3a. But, if no other customers arrive before the
line is authorized, the line is dropped, leaving 3 in the system, but
only one line. If a person arrives while the system is in state 3, the
94 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

line is again requested, putting the system in state 4r. If the line au-
thorization is complete while the system is in 4r, 5r, or 6r, then the
system goes into state 4a, 5a, or 6a, respectively.

(b) While in state 4r, the possible next events are customer arrival, cus-
tomer service, line authorization complete, with rates 2λ, µ, and τ ,
respectively.

(c) Let A, S, and AC denote the events customer arrival, customer ser-
vice, line authorization complete, respectively. Then
2λ
P {A} =
2λ + µ + τ
µ
P {S} =
2λ + µ + τ
τ
P {AC} = .
2λ + µ + τ

(d) The distribution of the amount of time spent in state 4r is exponen-

tially distributed with parameter 2λ + µ + τ because this time is the
minimum of 3 exponential random variables with parameter 2λ, µ,
and τ , respectively.

(e) See Figure 3.4 for the state transition rate diagram.

6λ 5λ µ
0 1 2 3
µ τ 3λ
µ µ 4λ 3λ 2λ λ
3a 4a 5a 6a
2µ µ µ µ
τ τ τ
3λ 2λ λ
3r 4r 5r 6r

2µ 2µ 2µ

Figure 3.4. State diagram for Supplemental Exercise 3.4

Elementary CTMC-Based Queueing Models 95

(f) The balance equations are as follows:

state rate leaves rate enters

0 6λP0 = µP1
1 (5λ + µ)P1 = 6λP0 + µP2
2 (4λ + µ)P2 = 5λP1 + µP3 + µP3a + 2µP3r
3 (3λ + µ)P3 = τ P3a
3a (3λ + τ + µ)P3a = 4λP2 + µP4a
3r (3λ + 2µ)P3r = 2µP4r
4a (2λ + τ + µ)P4a = 3λP3 + 3λP3a + µP5a
4r (2λ + 2µ)P4r = 3λP3r + τ P4a + 2µP5r
5a (λ + τ + µ)P5a = 2λP4a + µP6a
5r (λ + 2µ)P5r = 2λP4r + 2µP6r + τ P5a
6a (µ + τ )P6a = λP5a
6r 2µP6r = λP5r + τ P6a

3-5 Messages arrive to a statistical multiplexer at a Poisson rate λ for trans-

mission over a communication line having a capacity of C in octets per
second. Message lengths, specified in octets, are exponentially distributed
with parameter µ. When the waiting messages reach a level 3, the capacity
of the transmission line is increased to Ce by adding a dial-up line. The
time required to set up the dial-up line to increase the capacity is exponen-
tially distributed with rate τ . If the connection is completed when there
are less than three messages waiting or if the number of messages waiting
drops below two at any time while the additional capacity is in use, the
extra line is immediately disconnected.

(a) Define a suitable state space for this queueing system.

(b) Draw the state transition-rate diagram.
(c) Organize the state vector for this system according to level, where the
level corresponds to the number of messages waiting and in service,
and write the vector balance equations for the system.
(d) Determine the infinitesimal generator for the underlying Markov chain
for this system and comment on its structure relative to matrix geo-
metric solutions.

Solution.
(a) Figure 3.5 details two state diagram for the system. Note that the first
passage time from state 2 to state 1 in the first diagram is the same as
96 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

the length of the busy period in the second diagram. i.e., the M/M/1
with service rate 2µ. Thus, f˜21 = E[ỹM/M/1 ] = 2µ
1
/1 − 2µλ
.

µ µ µ
...
0
1,0 1 2 3
...
λ 2λ 3λ

µ µ
...
1 2
0
1,0 ...
2λ 2λ λ

Figure 3.5. State diagram for Supplemental Exercise 3.5

(b) Let A denote the event of an arrival before the first departure from a
busy period, and let Ac denote its complement. Then,

E[ỹ] = E[ỹ|A]P {A} + E[ỹ|Ac ]P {Ac }

= E[z̃ + ỹ2 ]P {A} + E[z̃]P {Ac }
λ µ
= E[z̃ + f˜21 + ỹ] + E[z̃]
λ+µ λ+µ
.

Thus,
1 !
µ 1 λ 2µ
E[ỹ] = + λ
λ+µ λ+µ λ+µ 1− 2µ

i.e.,
1 !
1 λ 2µ
E[ỹ] = + λ
.
µ µ 1− 2µ

1
 
1 1 λ 2µ
(c) c̃ = ĩ + ỹ implies that E[c̃] = λ + µ + µ λ
1− 2µ
. Therefore,

1
λ
P {system is idle} =  1

1 1 λ 2µ
λ + µ + µ λ
1− 2µ
1
=  λ

λ 2µ
1+ µ 1+ λ
1− 2µ
Elementary CTMC-Based Queueing Models 97

1
=  .
λ 1
1+ µ λ
1− 2µ

(d) From an earlier exercise, we know that the total amount of time the
system spends i state 1 before entering state 0 is exponential with
parameter µ because there are a geometric number of visits, and for
each visit the amount of time spent is exponential with parameter (λ+
µ). Thus,
1
µ
P {ñ = 1} =  1

1 1 λ 2µ
λ + µ + µ λ
1− 2µ
λ
= µ .
λ 1
1+ µ λ
1− 2µ

3-6 Consider the M/M/2 queueing system, the system having Poisson arrivals,
exponential service, 2 parallel servers, and an infinite waiting room capac-
ity.

(a) Determine the expected first passage time from state 2 to state 1.
[Hint: How does this period of time compare to the length of the
busy period for an ordinary M/M/1 queueing system?]
(b) Determine the expected length of the busy period for the ordinary
M/M/2 queueing system by conditioning on whether or not an arrival
occurs before the first service completion of the busy period and by
using the result from part (a).
(c) Define c̃ as the length of time between successive entries into busy
periods, that is, as the length of one busy/idle cycle. Determine the
probability that the system is idle at an arbitrary point in time by tak-
ing the ratio of the expected length of an idle period to the expected
length of a cycle.
(d) Determine the total expected amount of time the system spends in
state 1 during a busy period. Determine the probability that there is
exactly one customer in the system by taking the ratio of the expected
amount of time that there is exactly one customer in the system during
a busy period to the expected length of a cycle.
(e) Check the results of (c) and (d) using classical birth-death analysis.
98 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

(f) Determine the expected sojourn time, E[s̃], for an arbitrary customer
by conditioning on whether an arbitrary customer finds either zero,
one, or two or more customers present. Consider the nonpreemptive
last-come-first-serve discipline together with Little’s result and the
fact that the distribution of the number of customers in the system is
not affected by order of service.
Solution.
(a) If the occupancy is n, then there are (n − 1) customers waiting. So,
if there are 3 customers waiting, occupancy is 4. Less than 3 mes-
sages waiting means occupancy is less than 4. We assume that if the
number waiting drops below 3 while attempting to set up the extra
capacity, then the attempt is aborted. In this case, the required states
are 0, 1, 2, 3, 3e, 4, 4e, · · ·, where e indicates extra capacity.
(b) See Figure 3.6.

λ λ λ λ λ
0 1 2 3 4 5

µC µC µC µC T µC T
λ λ
µCe 3e 4e 5e

µCe µCe

Figure 3.6. State diagram for Supplemental Exercise 3.6

(c) Organizing the state vector for this system according to level, we see
that
level 0 : P0
level 1 : P1
level 2 : P2
level n : Pn = [ Pn0 Pn1 ] n>2

And the balance equations for the system are

λP0 = µCP1
(λ + µC) P1 = µP2 + λP0  
µC
(λ + µC) P2 = λP1 + λP3
µCe
Elementary CTMC-Based Queueing Models 99
   
λ + µC 0 µC 0
P3 = P2 [ λ 0 ] + P4
0 λ + µCe 0 µCe

For all remaining n, n ≥ 4,

     
λ + µC + τ 0 λ 0 0 τ
Pn = Pn−1 + Pn
0 λ + µCe 0 λ  0 0 
µC 0
+Pn+1
0 µCe

(d) The infinitesimal generator is readily written by inspection of the bal-

ance equations by simply taking all terms to the right-hand side. If
we let R = (λ + µC) and Re = (λ + µCe ), the resulting generator is
−λ λ 0 0 0 0 0 ...
 
 µC
 −R λ 0 0 0 0 ...
 0
 µC −R λ 0 0 0 ...
 0
 0 µC −R 0 λ 0 ...
 0 0 µCe 0 −Re 0 λ ...
 
 0 0 0 µC 0 −Re − τ τ ...
 
Q̃ = 
 0 0 0 0 µC 0 −Re ...

 
 0 0 0 0 0 µC 0 ...
 
 0 0 0 0 0 0 µCe ...
 
 0 0 0 0 0 0 0 ...
 
 0 0 0 0 0 0 0 ...
 
.. .. .. .. .. .. ..
. . . . . . .

We may also write Q̃ as

B00

B01 0 0 0 0 0 ...
 B10 B11 B12 B13 0 0 0 ...
 
 0 B21 B22 B23 0 0 0 ...
 
 0 0 B32 B33 A0 0 0 ...
Q̃ =  0
 
 0 0 B43 0 A1 A0 ...

 0 0 0 0 A2 A1 A0 ...
 
 0 0 0 0 0 A2 A1 ...
 
.. .. .. .. .. .. ..
. . . . . . .

We see that this latter form is the same as that given on page 140 of
the text, except that the boundary conditions are very complex. The
text by Neuts treats problems of this form under the heading of queues
with complex boundary conditions.
Chapter 4

ADVANCED CONTINUOUS-TIME
MARKOV CHAIN-BASED QUEUEING MODELS

Exercise 4.1 Using Little’s result, determine the average time spent in
the system for an arbitrary customer when the system is in stochastic equi-
librium.
Solution. Let s̃ denote the time spent in the systemP for an arbitrary customer,
and let γ be the net arrival rate into the system, γ = Mi=1 γi . By Little’s result,

E[ñ]
E[s̃] =
γ
where ñ is the total number of customers in the system in stochastic equilib-
rium. Observe that the total number of customers in the system is the sum of
the customers at each node. i.e., ñ = M
P
i=1 ñi . Therefore,

M
" #
1 X
E[s̃] = E ñi
γ i=1
M
1X
= E[ñi ]
γ i=1
M
1X ρi
= ,
γ i=1 1 − ρi

the latter following from (4.11) and (3.11).

Exercise 4.2 Argue that the matrix R is stochastic, and that, therefore,
the vector λ is proportional to the equilibrium probabilities of the Markov
chain for which R is the one-step transition probability matrix.
Solution. The element rij represents the probability that a unit currently at
node i will proceed next to node j. Since the system is closed, a unit must go
102 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

to some node within the network. Therefore,

M
X
rij = 1∀i.
j=1

Hence R is stochastic. From λ[I − R] = 0, we have

λR = λ.

Therefore, the vector λ is the left eigenvector of R corresponding to its eigen-

value at 1, and all solutions λ are proportional to this eigenvector. But, this
equation has exactly the form πR = π, where R is a one-step transition matrix
and π is the stationary probability vector of the Markov chain for which R is
the one-step transition probability matrix. Thus, λ is proportional to π.
Exercise 4.3 Let x denote any integer. Show that
1 for x = −1

1,
I
x
φ dφ =
j2π C 0, otherwise.
by direct integration.
Solution. Suppose x 6= −1. Then

1 1 φx+1
I
φx dφ = · | .
j2π C j2π x + 1 C

Note that the closed path C may be traversed by beginning at the point rejθ
and ending at rej(θ+2π) . The integral is then

1 φx+1 1 h i
· | = rej(θ+2π) |x+1 − rejθ |x+1
j2π x + 1 C j2π
1 h i
= r x+1 ej(θ+2π)(x+1) − r x+1 ejθ(x+1)
j2π
1 h i
= r x+1 ejθ(x+1) · 1 − r x+1 ejθ(x+1)
j2π
= 0.

Now suppose x = −1.

1 1 dφ
I I
−1
φ dφ =
j2π C j2π C  φ
1 
= ln φ
j2π C
rej(θ+2π)
1 
= ln φ
j2π rejθ
Advanced CTMC-Based Queueing Models 103

1
= [ln r + j(θ + 2π) − ln r − jθ]
j2π
1
= j2π = 1.
j2π

Exercise 4.4 Suppose that the expression for g(N, M, φ) can be written
as
r
Y 1
g(N, M, φ) = , (4.1)
i=1
(1 − σi φ)νi
where ri=1 νi = M . That is, there are exactly r distinct singular values of
P

g(N, M, φ)- these are called σ1 ,σ2 ,. . .,σr - and the multiplicity of σi is νi .
We may rewrite (4.35) as
νj
r X
X cij
g(N, M, φ) = j. (4.2)
i=1 j=1 (1 − σi φ)

Show that
(νi −j)
1 1

cij = −
(νi − j)! σi
d(νi −j) νi

[(1 − σ φ) g(N, M, φ)] .

i
dφ(νi −j)

φ=1/σi

Solution. Let m be a fixed but otherwise arbitrary index, 1 ≤ m ≤ r. Multiply

both sides of g(N, M, φ) by (1 − σm φ)νm :
νm
νm
cmj (1 − σm φ)νm −j + (1 − σm φ)νm F (φ),
X
(1 − σm φ) g(N, M, φ) =
j=1

where F (φ) is the remainder of the terms in the sum not involving σm . Now
fix n, 1 ≤ n ≤ νm . Differentiate both sides of this expression and evaluate at
φ = σ1m . Thus,

d(νm −n)

(1 − σm φ)νm g(N, M, φ) = (−σm )(νm −n) cmn (νm − n)!

dφ (ν m −n)
φ= σ1
i

i.e.,

1 −1 (νm −n) d(νm −n)

 
cmn = [(1 − σm φ)νm g(N, M, φ)]
(νm − n)! σm dφ(νm −n)
1 −1 (νi −n) d(νi −n)
  
νi

= [(1 − σ i φ) g(N, M, φ)]  .
(νi − n)! σi dφ(νi −n) 
φ= σ1
m
104 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

Exercise 4.5 Define bnN to be the coefficient of φN in the expansion of

(1 − σi φ)−n . Show that
N +n−1
 
bnN = σiN . (4.3)
N

Solution. Define h(φ) = (1−σi φ)−n . Then by the Maclaurin series expansion
for h,
∞
X h(k) (0)(−σi φ)k
h(φ) = .
k=0
k!

In particular, the N −th term of this sum is

h(N ) (0)(−σi φ)N

.
N!

But h(N ) (0) = (−n)(−n − 1) . . . (−n − M + 1)(1)N , so that the coefficient

of φN is

(−1)N (N + n − 1)!(−σi )N
bnN =
N !(n− 1)!
N +n−1

= σiN .
N

Exercise 4.6 Verify that the probabilities as specified by (4.39) sum to

unity.

Solution. In this exercise, we must first determine the set

5
( )
X
S4,5 = (n1 , n2 , · · · , n5 )| ni = 4 .
i=1

Then for each member (n1 , n2 , n3 , n4 , n5 ) of this set, we’ll substitute n3 , n4 ,

and n5 into 2n3 +n4 4n5 , producing an intermediate result, which we’ll call I.
Note that I is equal to P {n1 , n2 , n3 , n4 , n5 } multiplied by 1497. It suffices,
then, to show that the sum of these intermediate results is equal to 1497. The
following table gives all possible members of the set S4,5 , with the last column
of the table being the intermediate result I. These members were obtained by
finding all possible combinations of n1 , n2 , · · · , n5 with n5 = 4; then finding
all combinations with n5 = 3, etc.
Advanced CTMC-Based Queueing Models 105

i n1 n2 n3 n4 n5 I

1 0 0 0 0 4 256
2 0 0 0 1 3 128
3 0 0 1 3 3 128
4 0 1 0 0 3 64
5 1 0 0 0 3 64
6 0 0 0 2 2 64
7 0 0 2 0 2 64
8 0 0 1 1 2 64
9 0 2 0 0 2 16
10 2 0 0 0 2 16
11 1 1 0 0 2 16
12 0 1 0 1 2 32
13 1 0 0 1 2 32
14 0 1 1 0 2 32
15 1 0 1 0 2 32
16 0 0 3 0 1 32
17 0 0 0 3 1 32
18 0 0 1 2 1 32
19 0 0 2 1 1 32
20 0 1 2 0 1 16
21 1 0 2 0 1 16
22 1 0 0 2 1 16
23 0 1 0 2 1 16
24 0 1 1 1 1 16
25 1 0 1 1 1 16
26 0 2 1 0 1 8
27 2 0 1 0 1 8
28 0 2 0 1 1 8
29 2 0 0 1 1 8
30 1 1 1 0 1 8
31 1 1 0 1 1 8
32 0 3 0 0 1 4
33 3 0 0 0 1 4
34 1 2 0 0 1 4
35 2 1 0 0 1 4
36 0 0 4 0 0 16
37 0 0 0 4 0 16
38 0 0 3 1 0 16
39 0 0 1 3 0 16
106 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

i n1 n2 n3 n4 n5 I

40 0 0 2 2 0 16
41 1 0 3 0 0 8
42 0 1 3 0 0 8
43 1 0 0 3 0 8
44 0 1 0 3 0 8
45 1 0 2 1 0 8
46 0 1 2 1 0 8
47 1 0 1 2 0 8
48 0 1 1 2 0 8
49 0 2 2 0 0 4
50 2 0 2 0 0 4
51 0 2 0 2 0 4
52 2 0 0 2 0 4
53 1 1 2 0 0 4
54 1 1 0 2 0 4
55 1 1 1 1 0 4
56 0 2 1 1 0 4
57 2 0 1 1 0 4
58 0 3 1 0 0 2
59 3 0 1 0 0 2
60 0 3 0 1 0 2
61 3 0 0 1 0 2
62 1 2 0 1 0 2
63 2 1 0 1 0 2
64 1 2 1 0 0 2
65 2 1 1 0 0 2
66 0 4 0 0 0 1
67 4 0 0 0 0 1
68 1 3 0 0 0 1
69 3 1 0 0 0 1
70 2 2 0 0 0 1
It is easy to verify that the last column does indeed sum to 1497. This shows
that the probabilities given by (3.105) sum to unity.

Exercise 4.7 Carefully develop the argument leading from (4.43) to

(4.44).

Solution. In the development of (4.43), N was never assigned a specific num-

ber. Rather, it is a variable and so it simply represents some integer. Hence
(4.43) is true for all integers. Since N − n also represents some integer, we see
immediately that (4.44) follows from (4.43).
Advanced CTMC-Based Queueing Models 107

Exercise 4.8 Using the recursion of (4.43) together with the initial con-
ditions, verify the expression for g(N, 5) for the special case N = 6 numer-
ically for the example presented in this section.
Solution. Letting ρ = 4 as in the example in the chapter, we find that
ρ
ρ1 = ρ2 = = 1,
4
ρ
ρ3 = ρ4 = = 2,
2
ρ5 = ρ = 4.

Then, using (3.109) along with the initial conditions we may compute g(1, 1):

g(1, 1) = g(1, 0) + ρ1 g(0, 1)

= 0 + 1 · 1 = 1.

We then substitute this value into the expression for g(1, 2):

g(1, 2) = g(1, 1) + ρ2 g(0, 2)

= +1 · 1 = 2.

Repeating this for each value of N and M until g(6, 5) is reached, we may
form a table of normalizing constants:
M
0 1 2 3 4 5

0: 0 1 1 1 1 1
1: 0 1 2 4 6 10
2: 0 1 3 11 23 63
N 3: 0 1 4 26 72 324
4: 0 1 5 57 201 1497
5: 0 1 6 120 522 6510
6: 0 1 7 247 1291 27331

where g(6, 5) is in the lower right-hand corner of the table.

Exercise 4.9 Develop an expression for throughput at node M using Lit-
tle’s result and (4.47).
Solution. Define ñsM and s̃sM to be the number in service and time in service,
respectively, at node M . Each node in the system may have only one customer
in service at a time, so that by (4.47),

E[ñM ] = P {ñ > 0}

g(N − 1, M )
= ρM
g(N, M )
108 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

= L,

and
1
E[s̃M ] =
µ
= W.

The throughput may now be found by applying Little’s result to L and W ,

L
γ =
W
g(N − 1, M )
= λM .
g(N, M )

Exercise 4.10 Show that det A(z) is a polynomial, the order of which
is not greater than 2(K + 1).
Solution. The proof is by induction. First consider the case where A(z) is a
1 × 1 matrix. Clearly det A(z) is a polynomial whose order is at most 2. Now
assume that for an n × n matrix, its determinant is at most 2(n + 1). Let A(z)
be an (n + 1) × (n + 1) matrix. We wish to show det A(z) is at most 2(n + 2).
Now, the determinant of A(z) will be
K
X
det A(z) = a0,j A0,j ,
j=0

where a0,j is the element in the jth column of the first row of A(z) and A0,j
represents its respective cofactor. Note that for all j, j = 0, 1, · · · , n, a0,j is of
at most order 2. Furthermore, each A0,j is the determinant of an n × n matrix.
By assumption, A0,j is of order at most 2(n + 1). Hence the product a0,j A0,j
is of at most order 2(n + 2). Thus the result holds for all K ≥ 0.
Exercise 4.11 Repeat the above numerical example for the parameter
values β = δ = 2, λ = µ = 1. That is, the proportion of time spent in
each phase is the same and the transition rate between the phases is faster
than in the original example. Compare the results to those of the example
by plotting curves of the respective complementary distributions. Compute
the overall traffic intensity and compare.
Solution. Substituting β = δ = 2, λ = µ = 1, into the example in Chapter 3,
we find
−β β −2
   
2
Q= = ,
δ −δ 2 −2
   
µ 0 1 0
M= = ,
0 µ 0 1
Advanced CTMC-Based Queueing Models 109

and    
0 0 0 0
Λ= = .
0 λ 0 1
Note that the matrices M and Λ remain unchanged from the example in the
chapter. Then, upon substitution of these definitions into (3.137), we find
1 − 3z
 
2z
A(z) = ,
2z z 2 − 4z + 1
z 2 − 4z + 1 −2z
 
adj A(z) = ,
−2z 1 − 3z
and

det A(z) = −3z 3 + 9z 2 −√7z + 1 √

6 6
= 3(1 − z)(1 + − z)(1 − − z).
3 3

Thus, we find that


−0.3670068
 
 0.2996598
adj A(z) = ,
z=1−
√
6 −0.3670068 0.4494897
3

−2 −2
 

adj A(z) = ,
z=1 −2 −2
and
det A(z) 

= −2.
(1 − z) z=1
Substituting these numbers into (3.140) and (3.141), we find that
−0.0673470
 
2
[0 1 ] = P0 .
+0.0824829 2
Thus, we find that

P0 = [ 0.2752551 0.2247449 ] .
Upon substitution of this result into (3.138), we find after some algebra that

T T
1 0.2752551 − 0.0505103z
 
G0 (z)
= .
G1 (z) 1 − 0.5505103z 0.2247449
After some additional algebra, this result reduces to

[ G0 (z) G1 (z) ] = [ 0.2752551 0.2247449 ]

110 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

∞
(0.5505103)n z n .
X
+ [ 0.1835034 0.2247449 ]
n=1

Thus, for n ≥ 1, we find

[ Pn0 Pn1 ] = [ 0.1835034 0.2247449 ] (0.5505103)n .

The probability generating function for the occupancy distribution can now be
computed from (3.139) or by simply summing G0 (z) and G1 (z). We find
∞
(0.5505103)n z n .
X
Fñ (z) = 0.5 + 0.4082483
n=1

From this probability generating function, we find P0 = 0.5 and

Pn e = 0.4082483 × (0.5505103)n for n ≥ 1.

To compute the traffic intensity we determine the probability that the system
will be busy. This is simply the complement of the probability that the system
is not busy: 1 − P0 . Since P0 did not change in this exercise from the example
in the chapter, the traffic intensity for both problems will be the same:

1 − 0.5 = 0.5

Figure 4.1 shows the complementary distributions of both the original example
and its modification β = δ = 2,λ = µ = 1.
Exercise 4.12 Suppose we want to use the packet switch and transmis-
sion line of the above numerical example to serve a group of users who
collectively generate packets at a Poisson rate γ, independent of the com-
puter’s activity, in addition to the computer. This is a simple example of
integrated traffic. Assuming that the user packets also require exponential
service with rate µ, show the impact the user traffic has on the occupancy
distribution of the packet switch by plotting curves for the cases of γ = 0
and γ = 0.1.
Solution. Since the group of users generate packets independent of the com-
puter’s activity, the overall arrival rate to the transmission line is simple the
sum of the arrivals from the computer and from the group of users. That is,
 
γ 0
Λ= .
0 λ+γ
The matrices Q and M remain unchanged from the example. Now, in the
first case where γ = 0, it should be clear that the matrix Λ will also remain
Advanced CTMC-Based Queueing Models 111

1
P{occupancy exceeds value shown on abscissa}

.1

modified example
.01
original example

.001

.0001
0 3 6 9 12 15
n

Figure 4.1. Complementary distributions

unchanged. i.e., the problem is the same as that in the example. In the second
case, where γ = 0.1, the matrix Λ will be:
 
0.1 0
Λ= .
0 1.1
We repeat the definitions of Q and M:
−β −1
   
β 1
Q= = ,
δ −δ 1 −1
   
µ 0 1 0
M= = .
0 µ 0 1
We proceed in the same manner as in Exercise 3.39 and in the example in the
chapter. Substituting these definitions into (3.137) yields:
.1z 2 − 2.1z + 1
 
z
A(z) = ,
z 1.1z 2 − 3.1z + 1
112 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

1.1z 2 − 3.1z + 1 −z
 
adj A(z) = ,
−z .1z 2 − 2.1z + 1
and

det A(z) = .11z 4 − 2.62z 3 + 6.71z 2 − 5.2z + 1

= 0.11(1 − z)(1.5091141 − z)(21.0225177 − z)(0.2865501 − z).

Thus, we find that


−0.2865501
 
 0.2020168
adj A(z) = ,
z=0.2865501 −0.2865501 0.4064559

−1 −1
 

adj A(z) = ,
z=1 −1 −1
and
det A(z) 

= −0.8.
(1 − z) z=1
Substituting these numbers into (3.140) and (3.141), we find that
−0.0845332
 
2.5
[ 0 1 ] = P0 .
+0.1199059 2.5
Thus, we find that

P0 = [ 0.2346046 0.1653954 ] .
Upon substitution of this result into (3.138), we find after some algebra that
the matrix [G0 (z)G1 (z)] is

T
1 0.2346046 − 0.0739485z

.
(1 − 0.6626404z)(1 − 0.0475680z) 0.1653954 − 0.0047394
After some additional algebra, this result reduces to

[ G0 (z) G1 (z) ] = [ 0.2346046 0.1653954 ]

∞
(0.6626404)n z n
X
+ [ 0.1325209 0.1704812 ]
n=1
∞
(0.0475680)n z n .
X
+ [ 0.1020837 −0.0050858 ]
n=1

Thus, for n ≥ 1, we find

[ Pn0 Pn1 ] = [ 0.1325209 0.1704812 ] (0.6626404)n

Advanced CTMC-Based Queueing Models 113

+ [ 0.1020837 −0.0050858 ] (0.0475680)n .

The probability generating function for the occupancy distribution can now be
computed from (3.139) or by simply summing G0 (z) and G1 (z). We find
∞
(0.6626404)n z n
X
Fñ (z) = 0.4 + 0.3030021
n=1
∞
(0.0475680)n z n .
X
+ 0.0969979
n=1
From this probability generating function, we find P0 = 0.4 and
Pn e = 0.3030021×(0.6626404)n +0.0969979×(0.0475680)n for n ≥ 1.

1
P{occupancy exceeds value shown on abscissa}

.1

.01 gamma = 0.0

gamma = 0.1
.001

.0001

0
0 5 10 15 20 25
n

Figure 4.2. Complementary distributions

Figure 4.2 shows the occupancy distributions for the cases γ = 0, γ = 0.1.
Exercise 4.13 Let Q denote the infinitesimal generator for a finite,
discrete-valued, continuous-time Markov chain. Show that the rows of
adj Q are equal.
114 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

Solution. We first note that the rows of adj Q are all proportional to each other
and, in addition, these are proportional to the stationary vector for the CTMC.
Let the stationary vector be denoted by φ. Then, with ki , 0 ≤ i ≤ K, a scalar,
k0 φ k0 φ0 k0 φ1 ... k0 φK
   
 k1 φ   k1 φ0 k1 φ1 ... k1 φK 
adj Q =  .. = .. .. ..
   

 .   . . . 
kK φ kK φ0 kK φ1 ... kK φK
In addition, the columns of adj Q are all proportional to each other and these
are all proportional to the right eigenvector of Q corresponding to its zero
eigenvalue. Since Qe = 0, this right eigenvector is proportional to e. There-
fore,
j0 j1 ... jK
 
 j0 j1 ... jK 
adj Q = [ j0 e j1 e ... jK e ] =  .. .. ..  ,
 
 . . . 
j0 j1 ... jK
where ji , 0 ≤ j ≤ K, is a scalar. Comparing the two expressions for adj Q,
we see that ki φl = jl for all i. So that if φl 6= 0, then ki = jl /φl for all i. Since
φl cannot be zero for all i unless all states are transient, which is impossible
for a finite state CTMC, it follows that the ki are identical so that the rows of
adj Q are identical and the proof is complete.
Exercise 4.14 Obtain (4.73) by starting out with (4.64), differentiating
both sides with respect to z, postmultiplying both sides by e, and then taking
limits as z → 1.
Solution. Recall Equation (4.64):

G(z)A(z) = (1 − z)P0 M,

where
A(z) = Λz 2 − (Λ − Q + M)z + M.
Now, differentiating both sides of (3.137) with respect to z, we see that
′
G (z)A(z) + G(z) [2Λz − (Λ − Q + M)] = −zP0 M.

Taking limits as z → 1,
′
G (1)Q + G(1) [Λ + Q − M] = −P0 M.

Then postmultiply both sides by e, and noting that Qe = 0 yields the desired
result:
G(1) [M − Λ] = P0 M.
Advanced CTMC-Based Queueing Models 115

Exercise 4.15 Show that the sum of the columns of A(z) is equal to the
column vector (1 − z)(M − zΛ)e so that det A(z) has a (1-z) factor.
Solution. To find the sum of the column vectors of A(z), postmultiply the
matrix by the column vector e:

A(z)e = Λz 2 e − (Λ − Q + M) e + Me
= −z(1 − z)Λe + (1 − z)Me
= (1 − z) (M − zΛ) e.

Exercise 4.16 Show that the zeros of the determinant of the λ-matrix
AlA(z) are all real and nonnegative. [Hint: First, do a similarity trans-
formation, transforming A(z) into a symmetric matrix, Â(z). Then, form
the inner product < Xz , Â(z)Xz >, where Xz is the null vector of Â(z)
corresponding to z. Finally, examine the zeros of the resulting quadratic
equation.]
Solution. First we show that the null values are all positive by transforming
A(z) into a symmetric matrix, Â(z), and then forming the quadratic form of
Â(z). This quadratic form results in a scalar quadratic equation whose roots
are positive and one of which is a null value of A(z). Recall the definition of
A(z):
A(z) = Λz 2 − (Λ − Q + M)z + M,
where

Λ = diag(λ0 , λ1 , · · · , λK ),
M = diag(µ0 , µ1 , · · · , µK ),

and
−β0 β0 0
 
 δ1
 −(β1 + δ1 ) β1 

Q=
 .. .

 . 
 δK−1 −(βK−1 + δK−1 ) βK−1 
0 δK −δK
Now, since Q is tridiagonal and the off-diagonal terms have the same sign, it
is possible to transform Q to a symmetric matrix using a diagonal similarity
transformation. Define
 Qi
Wij = j=0 lj for i = j = 0, 1, · · · , K
0, else
where
l0 = 1,
116 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

and s
δj
lj = for 1 ≤ j ≤ K.
βj−1
Then
Q̂ = W −1 QW,
where Q̂ is the following tridiagonal symmetric matrix:
√
√−β0 β0 δ1 √0
 
 β0 δ1 −(β1 + δ1 ) β1 δ2 
Q̂ =  .
 
..
 . p

0 βK−1 δK −δK

It is then straightforward to show that Q̂ is negative semi-definite. Also, since

the determinant of a product of matrices is equal to the product of the individual
determinants, and W is non-singular, it follows that

detÂ(z) = detW −1 A(z)W.

Furthermore, since Λ, M, and W are all diagonal, it follows that

h i
detÂ(z) = det Λz 2 − (Λ − Q̂ + M)z + M .

Now, suppose that det A(z) = 0. Then det Â(z) = 0, and there exists some
nontrivial Xσ , such that
Â(z)Xσ = 0.
Therefore,
< Xσ , Â(z)Xσ >= 0,
where < ·, · > denotes the inner product. Thus,

< Xσ , Â(z)Xσ >= lσ 2 − (l − p + m)σ + m = 0, (3.44.1)

where

l = < Xσ , ΛXσ >,

m =< Xσ , MXσ >,
p = < Xσ , Q̂Xσ > .

But, Λ and M are positive semi-definite, so l and m are nonnegative. Also, p

is negative semi-definite, so p ≤ 0. The discriminant of the quadratic equation
for (3.44.1) is given by

(l − p + m)2 − 4lm = (l − m)2 − 2p(l + m) + p2 .

Advanced CTMC-Based Queueing Models 117

We see that the right-hand side of this equation is the sum of nonnegative terms,
and so the value of σ must be real. Hence all null values of A(z) are real.
We now show that these null values are positive. We see from above that either
l = 0, or l > 0. In the first case, we find from (3.44.1) that
m
σ= > 0.
m−p
In the second case we find from the quadratic equation that
1
 q 
σ= (l − p + m) ± (l − p + m)2 − 4lm .
2l
p
Clearly, (l − p + m) > (l − p + m)2 − 4lm > 0. Therefore, both roots of
(3.44.1) are positive so that all of the null values of A(z) are positive.
Exercise 4.17 The traffic intensity for the system is defined as the prob-
ability that the server is busy at an arbitrary point in time.
1. Express the traffic intensity in terms of the system parameters and P0 .
2. Determine the average amount of time a customer spends in service us-
ing the results of part 1 and Little’s result.
3. Check the result obtained in part 2 for the special case M = µI.

Solution.
1. Recall (4.74):
G(1)Λe = [G(1) − P0 ] Me.
The left-hand side expresses the average rate at which units enter the service
the system, and the right-hand side expresses the average rate at which units
leave the system. Hence, analogous to the special case of the M/M/1 system
of λ = (1−P0 )µ where ρ = µλ = (1−P0 ), we may write the traffic intensity
using (4.74) as
ρ = [G(1) − P0 ] e.

2. Let T represent time. Then, if [G(1) − P0 ] e is the traffic intensity, [G(1) − P0 ] T e

is the total time spent serving. In addition, the number of customers com-
pleting service over time is [G(1) − P0 ] MT e. This is due to the fact that
the right-hand side of (4.74) represents the average rate at which customers
leave the system. Therefore, by Little’s result, the average service time is
then
[G(1) − P0 ] T e [G(1) − P0 ] e
= .
[G(1) − P0 ] MT e [G(1) − P0 ] Me
118 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

3. If M = µI, then the average service time is

E[x̃] = [G(1) − P0 ] e 1
= ,
[G(1) − P0 ] µe µ
which is a familiar result.

Exercise 4.18 Show that adj A(zi )e is proportional to the null vector
of A(z) corresponding to zi .
Solution. Since A−1 (zi ) = adj A(zi )/det A(zi ), and A(zi )A−1 (zi ) = I, we
find that
A(zi )adj A(zi ) = det A(zi )I.
By the continuity of A(zi ), adj A(zi ), and det A(zi ), and the fact that zi is a
null value of A(z),

lim A(zi )adj A(zi ) = lim detA(zi )I.

σ→σi σ→σi

This implies, since lim detA(zi ) = 0, that

σ→σi

A(zi )adj A(zi ) = O(K+1)×(K+1) ,

where On×m denotes an n × m matrix of zeroes. Thus, we have

A(zi )adj A(zi )e = 0.

This defines adj A(zi ) to be a null vector of A(z) corresponding to zi .

Exercise 4.19 Show that if the system described by (4.57) and (4.60) is
ergodic, then there are exactly K zeros of det AlA(z) in the interval (0, 1).
[Hint: First show that this is the case if δi = 0 ∀ i. Then show that it is
not possible for det A(z)/(1 − z) to be zero for any choice of δi s unless
P0 = 0, which implies no equilibrium solution exists. The result is that the
number of zeros in (0, 1) does not change when the δi change.]
Solution. First note that det A(z) is equal to the determinant of the matrix
obtained from A(z) by replacing its last column by A(z)e. Now, it is easy to
show that
λ0 z 2 − (λ0 + µ0 )z + µ0
 
 λ1 z 2 − (λ1 + µ1 )z + µ1 
 
A(z)e = 

 = (1 − z)(M − Λz)e.

 ... 
λK z 2 − (λK + µK )z + µK
Therefore, det A(z) = (1 − z)det B(z), where B(z) is the matrix obtained by
replacing the last column of A(z) by the column vector (M − Λz)e.
Advanced CTMC-Based Queueing Models 119

Now, we may compute the determinant of B(1) by expanding about its last
column. In doing this, we find that the vector of cofactors corresponding to the
last column in proportional to µK , the left eigenvector of Q corresponding to
its zero eigenvalue. That is,
φK
G(1) = .
φK e
Thus,
detB(1) = αG(1) [M − Λ] e, (4.57.1)
for some α = φK e 6= 0. But from equation (4.74), G(1) [M − Λ] e = π0 Me.
The left-hand side of (4.57.1) is zero only if π0 = 0. But π0 > 0 if an equi-
librium solution exists, so that if an equilibrium solution exists, then B(1) is
nonsingular. Equivalently, if an equilibrium solution exists, then B(z) has no
null value at z = 1.
We note that the null values of a λ-matrix are continuous functions of the
parameters of the matrix. Thus, we may choose to examine the behavior of the
null values of B(z) as a function of the death rates of the phase process while
keeping the birth rates positive. Toward this end, we define
B̂(z, δ1 , δ2 , · · · , δK ).
First, consider det B̂(z, 0, 0, · · · , 0). Since the lower subdiagonal of this
λ-matrix is zero, the determinant of the matrix is equal to the product of the
diagonal elements. These elements are as follows:
λi z 2 − (λi + βi + µi )z + µi , 0 ≤ i < K,
µK − λK z.
The null values of B̂(z, 0, 0, · · · , 0) may therefore be found by setting each of
the above terms equal to zero. Thus, the null values of B̂(z, 0, 0, · · · , 0) may
be obtained by solving the following equations for z:
λi z 2 − (λi + βi + µi )z + µi = 0, 0 ≤ i < K, (4.57.2)
and
µK − λK = 0.
When solving (4.57.2), two possibilities must be considered: λi = 0, and
λi > 0. In the first case,
µi
z= ,
βi + µi
so that for each i, the corresponding null value is between zero and one. In the
latter case, we find that
lim λi z 2 − (λi + βi + µi )z + µi = µi > 0;
z→0
120 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

lim λi z 2 − (λi + βi + µi )z + µi = −βi < 0;

z→1

and that the expression

lim λi z 2 − (λi + βi + µi )z + µi
z→∞

tends to +∞. Therefore, each quadratic equation of (4.57.2) has one root in
(0, 1) and one root in (1, ∞) if λi > 0. In summary, the equations (4.57.2)
yield exactly K null values of B̂(z, 0, 0, · · · , 0) in (0, 1), and between 0 and K
null values of B̂(z, 0, 0, · · · , 0) in (1, ∞), the latter quantity being equal to the
number of positive values of λi .
Now, due to (4.57.1), there exists no (δ1 , δ2 , · · · , δK ) such that 0 = det
B̂(1, δ1 , δ2 , · · · , δK ). In addition, all of the null values of B̂(z, δ1 , δ2 , · · · , δK )
are real if Q is a generator for an ergodic birth and death process. Therefore,
it is impossible for any of the null values of B̂(z, δ1 , δ2 , · · · , δK ) in the inter-
val (0, 1) to assume that value 0 for some value of (δ1 , δ2 , · · · , δK ). Hence,
the number of null values of B̂(1, δ1 , δ2 , · · · , δK ) in (0, 1) is independent of
the death rates. Therefore, B(z), and consequently A(z), has exactly K null
values in the interval (0, 1).
Exercise 4.20 Beginning with (4.86) through (4.88), develop expres-
sions for the joint and marginal complementary ergodic occupancy distri-
butions.
Solution. From (4.86),
−n
X
Pn = Ai zK+i .
zK+1 ∈Z (1,∞)

Therefore,
∞
−j
X X
P {ñ > n} = Ai zK+i
j=n+1 zK+1 ∈Z (1,∞)
∞ j
1
X X 
= Ai
j=n+1
zK+i
zK+1 ∈Z (1,∞)
∞ j
1
X X 
= Ai
j=n+1
zK+i
zK+1 ∈Z (1,∞)
 n+1
1
X zK+i
= Ai 1
1− zK+i
zK+1 ∈Z (1,∞)
 
(n+1)
X zK+i
= Ai  −1
.
zK+1 ∈Z (1,∞)
1 − zK+i
Advanced CTMC-Based Queueing Models 121

Similarly,  
(n+1)
X zK+i
P {ñ > n} e = Ai e  −1
.
zK+1 ∈Z (1,∞)
1 − zK+i

Exercise 4.21 Develop an expression for adj A(zi ) in terms of the outer
products of two vectors using LU decomposition. [Hint: The term in the
lower right-hand corner, and consequently the last row, of the upper trian-
gular matrix will be zero. What then is true of its adjoint?]
Solution. We let A(z) = L(z)U (z), where L(z) is lower triangular and U (z)
is upper triangular. We follow the usual convention that the elements of the
major diagonal of L(z) are all unity. From matrix theory, we know that for
A(z) nonsingular,
adj A = det A(z) A(z)−1
−1 −1
= det A(z)U (z)L (z)
1 1
= det A(z) adj U (z) adj L(z)
det U (z) det L(z)
= adj U (z) adj L(z).
Since adj A(z) is a continuous function of z for an arbitrary matrix G(z), we
then have
adj A(zi ) = adj U (zi )adj L(zi ).
Now, if zi is a simple null value of U (z), the LU decomposition can be
arranged such that the last diagonal element of U (z) is zero. This means the
entire last row of U (z) is zero so that all elements of adj U (z) are zero except
for its last column. This means that adj A(zi ) will be given by the product of
the last column of adj U (zi ) with the last row of adj L(zi ). Let y(zi ) denote
the last column of adj U (zi ) and x(zi ) denote the last row of adj L(zi ). Since
det L(zi ) = 1, we may solve for the last row of adj L(zi ) by solving
x(zi )L(zi ) = fK ,
where fi is the row vector of all zeros except for a 1 in position i. Since the
last column of adj U (z) is unaffected by the elements of the last row of U (z),
we may determine the last column of adj U (zQ i ) in the following way. First,
we replace the (K, K) element of U (zi ) by 1/ K−1 j=0 Ujj (zi ). Call this matrix
Û (zi ) and the note that this matrix is nonsingular and that its determinant is 1.
Now solve the linear system
Û (zi )y(zi ) = eK .
We then have
1
y(zi ) = adj Û (zi )eK .
det Û (zi )
122 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

But det Û (zi ) = 1 and, in addition, the last column of adj Û (zi ) is the same
as the last column of U (zi ). Consequently y(zi ) is identically the last column
of adj U (zi ).
In summary, we solve Û (zi )y(zi ) = eK , x(zi )L(zi ) = fK and then we
have
adj A(zi ) = y(zi )x(zi ).
We note in passing that solving the linear systems is trivial because of the form
of Û (zi ) and L(zi ). We illustrate the above method with an example. Let

2 3 4
 

A= 4 8
 11  .
6 13 18
We may then decompose A as described above:
 
1 0 0
L =
2 1 1
2 3 4
U = 0 2 3
0 0 0
Now, det L is 1, so that adj L = L−1 . That is, if x̂ is the last row of adj L,
then
x̂L = [ 0 0 1 ]
Thus,
x̂ = [ 1 −2 1 ]
1
We now replace the (3, 3) element of U of 2+2 = 41 , so that

2 3 4
 

Û =  0 2 3
1
0 0 4

Since the determinant of this matrix is 1, adj Û = det Û −1 . We may compute

the last column ŷ of adj Û (and, hence, of adj U ) by solving the system
0
 

Û ŷ =  0 
1

It is easy to show that ŷ = [ 1 −6 4 ]T . The product is then

1 −2 1
 

x̂ŷ =  −6 12 −6 
4 −8 4
Advanced CTMC-Based Queueing Models 123

By computing adj A directly by forming the minors of A, we see that the

product x̂ŷ is, indeed, equal to adj A.
Exercise 4.22 Solve for the equilibrium state probabilities for Example
4.5 using the matrix geometric approach. Specify the results in terms of
the matrix AlR. Determine numerically the range of values of τ for which
(4.98) converges. Also, verify numerically that the results are the same as
those obtained above.
Solution. A program was written to compute the different iterations of Equa-
tion (3.172). R0 was taken to be diag(0.5, 0.5). Using differing values of τ
and the parameters from Example 3.7 into (3.172), we find
2.1 −1.0
    
−1 0.1 0.0
Ri = τ R2i−1 + Ri−1 I − τ −1
+τ −1
.
−1 3.1 0.0 1.1
The matrix R was determined to be
 
0.070500 0.029500
R=
0.460291 0.639709

for values of τ between 1.7215 and 2795. Since the vector G(1) is proportional
to the left eigenvector of Q corresponding to the eigenvalue 0, we find

G(1) = [ 0.5 0.5 ] .

Substituting these result into (3.167) yields

  
0.070500 0.029500
P0 = [ 0.5 0.5 ] I −
0.460291 0.639709
= [ 0.234604 0.165395 ] .

The solution expressed in matrix geometric form as in (3.163) is then

Pn = P0 Rn  n
0.070500 0.029500
= [ 0.234604 0.165395 ] .
0.460291 0.639709

These are the same results as those found in Example 3.7. In particular, we see
that P0 = 0.4.
The following are sample data from the program, where ǫ = 1 × 10−11 , and n
is the number of iterations it took for the program to converge. Note that the
resultant matrix R was the same for each of the sampled τ .
 
0.070500 0.029500
R= ,
0.460291 0.639709
124 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

τ n

2 300

50 3200

200 11800

500 28000

2795 10928
Exercise 4.23 Solve Exercise 3.22 using the matrix geometric approach.
Evaluate the relative difficulty of using the matrix geometric approach to
that of using the probability generating function approach.
Solution. Substituting different values of τ and the parameters from Exercise
3.39 into the computer program from Exercise 3.50, we find
3.0 −2.0
    
−1 0.0 0.0
Ri = τ R2i−1 + Ri−1 I − τ −1
+τ −1
.
−2 4.0 0.0 1.0
The matrix R was determined to be
 
0.0 0.0
R=
0.449490 0.550510
. Since the vector G(1) is proportional to the left eigenvector of Q correspond-
ing to the eigenvalue 0, we find
G(1) = [ 0.5 0.5 ] .
Substituting these result into (3.167) yields
  
0.0 0.0
P0 = [ 0.5 0.5 ] I −
0.449490 0.550510
= [ 0.275255 0.224745 ] .

The solution expressed in matrix geometric form as in (3.163) is then

Pn = P0 Rn  n
0.0 0.0
= [ 0.275255 0.224745 ] .
0.449490 0.550510

These are the same results as those found in Exercise 3.39. In particular, we
see that P0 = 0.5. Note that P0,0 does not play a role in determining Pn since
the first row of R is zero.
Advanced CTMC-Based Queueing Models 125

The following are sample data from the program, where ǫ = 1 × 10−11 , and n
is the number of iterations it took for the program to converge. As in Exercise
3.50, the resultant matrix R was the same for each of the sampled τ .
 
0.0 0.0
R= ,
0.449490 0.550510

τ n

3 300

10 600

50 2400

100 4500

500 20400
For this exercise, K = 1. Since the matrices involved were 2 × 2, using
the probability generating function approach was fairly straight forward and
did not involve a great deal of computation. However, as the dimension of
these matrices increases, the difficulty involved with using this method also
increases. On the other hand, solving the problem using the matrix-geometric
approach involved developing a computer program. Although this is not al-
ways an easy task, this particular program was relatively simple. Furthermore,
the program should be adaptable to different parameters: matrix dimension, τ ,
λi , βi , etc. So although there is initial overhead, it is reusable. Plus, as the di-
mension of the matrices increases even slightly, the matrix geometric method
is much more practical than the probability generating function approach.
Exercise 4.24 Prove the result given by (4.112) for the n-th factorial
moment of ñ.
Solution. First we prove a lemma:

dn
G(z) = n!G(z) [I − zR]−n Rn , n ≥ 1.
dz n

Suppose n = 1. Then using the identity G(z) = P0 [I − zR]−1 ,

d
G(z) = P0 [I − zR]−2 R
dz
= P0 [I − zR]−1 [I − zR]−1 R
= G(z) [I − zR]−1 R.
126 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

Now assume the result holds true for n; we wish to show it holds true for n + 1.

d(n+1) d dn
 
G(z) = G(z)
dz (n+1) dz dz n
d  
= n!G(z) [I − zR]−n Rn ,
dz

by assumption. Thus,

d(n+1)
G(z) = −(n + 1)n! [I − zR]−(n+1) Rn (−R)
dz (n+1)
= (n + 1)!G(z) [I − zR]−(n+1) Rn+1 .

This proves the lemma. Using this result, we evaluate the n−th derivative of

G(z) at z = 1, and recall that Rn = V −1 N n V,

dn
 
G(z) e = n!G(z) [I − zR]−n Rn  e
 
dz n
z=1 z=1
= n!G(1) [I −"R]−n R n
e
n ∞
#
Ri Rn e
Y X
= n!G(1)
j=1 " i=0
n ∞
#
−1
N V V −1 N n Ve
i
Y X
= n!G(1) V
j=1 i=0"
n ∞
#
−1 i
N n Ve
Y X
= n!G(1)V N
j=1 i=0
−1 −n
= n!G(1)V [I − N ] N n Ve.

Now, N is a diagonal matrix, and so [I − N ]−1 is the matrix

1 1
···
 
diag 1−ν0 1−νK .

Furthermore, multiplication of diagonal matrices is again a diagonal matrix,

whose elements are the products of the diagonal elements of the individual
matrices. In particular, powers of diagonal matrices are the diagonal matrix of
powers of the diagonal elements of the original matrix. Thus,

dn
 h n  n i
−1 ν0 νK

G(z) e = n!G(1)ν 1−ν0 , ··· 1−νK Ve.
dz n 
z=1
Advanced CTMC-Based Queueing Models 127

Exercise 4.25 Beginning with (4.113), show that

∞
X
Pi = π.
i=0

[Hint: First, sum the elementsPof the right hand side of (4.113) from i = 1 to
i = ∞. This will yield 0 = [ ∞ i=0 Pi ] [A0 + A1 + A2 ] + θ(P0 , P1 ). Next,
use the fact that, because Q̃ is stochastic, the sum of the elements of each
of the rows of Q̃ must be a zero matrix to show that θ(P0 , P1 ) = 0. Then
complete the proof in the obvious way.]
Solution. We emphasize that we are dealing here with the specific case that the
arrival and service rates are dependent upon the phase of an auxiliary process.
The arrival, service, and phase processes are independent of the level except
that at level zero, there is no service. For a system in which the previous
condition does not hold, the result is in general not true.
The implication of the fact that Q̃ is stochastic is that the sum of the block
matrices at each level is stochastic; that is [A0 + A1 + A2 ], [B1 + A1 + A0 ],
and [B0 + A0 ] are all stochastic, and each of these stochastic matrices is the
infinitesimal generator of the phase process at its level. Thus, in order for the
phase process to be independent of the level, the three stochastic matrices must
be the same. This can be found by comparing the first two to see that B1 = A2
and the first and third to see that B0 = A1 + A2 .
Upon summing the elements of the right hand side of (4.113) from i = 1 to
i = ∞, we find
"∞ #
X
0= Pi [A0 + A1 + A2 ] − P0 [A1 + A2 ] − P1 A2 .
i=0

But, for our special case, A1 + A2 = B0 and A2 = B1 . Thus the previous

summation yields
"∞ #
X
0= Pi [A0 + A1 + A2 ] − P0 B0 − P1 B1 .
i=0

But, from the boundary condition of (4.113), we have P0 B0 − P1 B1 = 0.

Therefore, we have
"∞ #
X
0= Pi [A0 + A1 + A2 ]
i=0

from which we see that

∞
X
Pi = π
i=0
128 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

because both are the unique left eigenvector of [A0 + A1 + A2 ] whose ele-
ments sum to unity.
In a very general case, the arrival and services while in a given phase could
be accompanied by a phase transition. For xample, we would denote by λij
the arrival rate when the process that starts in phase i and results in an increase
in level by 1 and a transition to phase j simultaneously, and define µij in a
parallel way. At the same time, we could have an independently operating
phase process, whose infinitesimal generator is Q.
Let Λ̂ and M̂ denote the more general arrival and service matrices, respec-
tively, and let Λ and M denote the diagonal matrices whose i-th elements
denote the total arrival and service rates while the process is in phase i. It is
easy to show that we then have the following dynamical equations:

0 = P0 (Q − Λ) + P1 M̂ and 0 = Pi−1 Λ̂ + Pi (Q − Λ − M) + Pi+1 M̂.

It is then easy to see that the phase process is independent of the level if and
only if the M = M̂. That is, the service process must not result in a phase
shift. Alternatively, the phase process at level zero could be modified to incor-
porate the phase changes resulting from the service process at all other levels,
but then this would contaminate the definition of the independent phase process
at each level.

Exercise 4.26 With Q̃ defined as in (4.138) and b any vector of proba-

bility masses such that be = 1, show that the matrix S + S 0 b is always an
infinitesimal generator.

Solution. We first note that since Q̃ is stochastic, it follows that Se + S 0 = 0.

Since be = 1, it follows immediately that S 0 be = S 0 . Thus, Se + S 0 be = 0.
But, S 0 b is a matrix whose dimensions are the same as those of S so that
S and S 0 b may be summed term-by-term. Hence we have (S + S 0 b)e =
0. It remains to show that the elements of the diagonal of S + S 0 b are non-
positive and the off diagonal elements are nonnegative. To show this, first note
that the matrix S has these properties. Since Q̃ is stochastic, the magnitude
of the diagonal elements of S must be at least as large as the corresponding
row elements of S 0 , so that the sum of the diagonal elements of S and the
corresponding row elements of S 0 cannot exceed 0. The diagonal elements
of S 0 b cannot exceed the corresponding row elements of S 0 because b is a
vector of probability masses, and hence each element cannot exceed 1. Since
the off-diagonal elements of S and the elements of S 0 b are all nonnegative, the
elements of the result follows.
Advanced CTMC-Based Queueing Models 129

Exercise 4.27 Consider a single-server queueing system having Poisson

arrivals. Suppose upon entering service, each customer initially receives a
type 1 service increment. Each time a customer receives a type 1 service
increment, the customer leaves the system with probability (1 − p) or else
receives a type 2 service increment followed by an additional type 1 service
increment. Suppose type 1 and type 2 service increment times are each
drawn independently from exponential distributions with parameters µ1 and
µ2 , respectively. Define the phase of the system to be 1 if a customer in
service is receiving a type 2 service increment. Otherwise, the system is in
phase 0. Define the state of the system to be the 0 when the system is empty
and by the pair (i, j) where i > 0 is the system occupancy and j = 0, 1
is the phase of the service process. Define Pi = [ Pi0 Pi1 ] for i > 0
and P0 , a scalar. Draw the state diagram, and determine the matrix Q, the
infinitesimal generator for the continuous-time Markov chain defining the
occupancy process for this system.

Solution. The state diagram can be determined by first defining the states in
the diagram which have the from (i, j), where i is the system occupancy and
j is the phase, except for i = 0. It is then a matter of determining transition
rates. Since all phase 0 service completions result in lowering the occupancy
by 1 with probability p, the rate from state (i, 0) to (i − 1, 0) is µ1 (1 − p).
Similarly, the rate from (i, 0) to (i, 1) is µ1 p. This is a direct result of Property
2 of the Poisson Process given on page 44. Since service completions at phase
1 always result in an additional phase 0 service, the transition rate from (i, 1)
to (i, 0) is µ2 . Increases in level occur due to arrivals only, so for every state,
the rate of transition to the next level is λ. The state diagram is detailed in
Figure 4.3.

λ
1,1 2,1 3,1 ...

µ1 p µ2 µ1 p µ2 µ1 p µ2

µ1(1-p) µ1 (1-p) µ1 (1-p)

1,0 2,0 3,0
...
0
1,0 ...
λ λ λ

Figure 4.3. State Diagram for Problem 4.27

130 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

By inspection of the state diagram, we find

−λ

λ 0 0 0 ...
 µ1 (1 − p) −µ1 − λ µ 1p λ 0 ...
−λ −
 

 0 µ 2 µ2 0 λ ...

Q=
 0 µ1 (1 − p) 0 −µ1 − λ µ1 p ...
.

 0 0 0 µ2 −µ2 − λ . . . 


 0 0 0 µ1 (1 − p) 0 ...

.. .. .. .. .. ..
. . . . . .
This compares to the infinitessimal generator for a queueing system with phase
dependent arrivals and service as follows:
Λ = λI,
−µ1 p µ1 p
 
P = ,
 µ2 −µ2 
µ1 (1 − p) 0
M = .
0 0
It is different in the first level, but it is exactly an M/MP/1 system because the
service always starts in phase 0.
Exercise 4.28 Suppose
−µ
   
µ 0 0
S= S = , and Pt (0) = [ 1 0 ] .
0 −µ µ

Find Fx̃ (t) = P {x̃ ≤ t} and fx̃ (t), and identify the form of fx̃ (t). [Hint:
First solve for P0 (t), then for P1 (t), and then for P2 (t) = Pa (t). There is
never a need to do matrix exponentiation.]
Solution. We have
d
P (t) = P (t)Q̃,
dt
or
−µ µ 0
 
d
[ P0 (t) P1 (t) P2 (t) ] = [ P0 (t) P1 (t) P2 (t) ]  0 −µ µ.
dt
0 0 0
d
Thus, dt P0 (t) = µP0 (t), so that P0 (t) = ke−µt . But P0 (0) = 1, so k = 1.
Therefore P0 (t) = e−µt .
d
P1 (t) = µP0 (t) − µP1 (t).
dt
or
d
P1 (t) + µP1 (t) = µP0 (t).
dt
Advanced CTMC-Based Queueing Models 131

Multiplying both sides by eµt ,

d
eµt P1 (t) + µeµt P1 (t) = µeµt P0 (t).
dt
Equivalently,
d h µt i
e P1 (t) = µeµt e−µt = µ.
dt
Therefore, e P1 (t) = µt + k, or P1 (t) = µte−µt + ke−µt . Now, P1 (0) = 0,
µt

so k = 0 and we have P1 (t) = µte−µt .

d
P2 (t) = µP1 (t)
dt
= µ2 te−µt .

Since the probabilities sum to 1, it must be the case that

P2 (t) = 1 − P0 (t) − P1 (t)

= 1 − e−µt µte−µt .

We may check this by evaluating P2 (t) directly.

d d h i
P2 (t) = 1 − e−µt µte−µt
dt dt
= 0 + µe−µt + µ2 te−µt − µe−µt
= µ(µt)e−µt .

Now,

P2 (t) = P {in state 2 at time t}

= P {x̃ ≤ t} .

That is,
Fx̃ (t) = 1 − e−µt − µte−µt ,
and
fx̃ (t) = µ(µt)e−µt .

Exercise 4.29 Starting with (4.139) as given, prove the validity of

(4.140).
Solution. First, find the Laplace-Stieltjes transform of x̃.
Z ∞
Fx̃∗ (s) = e−sx dFx̃ (x)
Z0∞ h i
= e−sx −bSeSx e dx
0
132 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL
Z ∞
= −bS e−[sI−S]x dxe
0 Z ∞
= bS [sI − S]−1 (− [sI − S]) e−[sI−S]x dxe
0
= −bS [sI − S]−1 e.
d n
Now, E[x̃n ] = (−1)n ds ∗
n Fx̃ (s)|s=0 , so

d

E[x̃] = (−1) Fx̃∗ (s)

ds s=0
dn h i 
= (−1) n −bS [sI − S]−1 e 
ds  s=0
h i
= (−1) bS [sI − S]−2 e 
s=0
= −bS −1 e.
We now prove that proposition that
dn ∗
F (s) = −n!(−1)n bS [sI − S]−(n+1) e.
dsn x̃
For n=1,
d ∗
F (s) = bS [sI − S]−2 e
ds x̃
= 1!(−1)(−1)bS [sI − S]−2 e.
So 1 ∈ T , the truth set for the proposition. If (n − 1) ∈ T , then
dn−1 ∗
F (s) = −(n − 1)!(−1)(n−1) bS [sI − S]−n e,
dsn−1 x̃
so that
dn ∗
F (s) = −(n − 1)!(−1)(n−1) bS(−n) [sI − S]−(n+1) e
dsn x̃
= −n!(−1)n bS [sI − S]−(n+1) e.
and the proof of the proposition is complete. Therefore, continuing with the
proof of the exercise,
h i 
n n n −(n+1)
E[x̃ ] = (−1) −n!(−1) bS [sI − S] e 
s=0
= −n!bS(−1)(n+1) e
= n!(−1)n bS −n e.
Advanced CTMC-Based Queueing Models 133

4.1 Supplementary Problems

4-1 Let Q be an (m+1)-square matrix representing the infinitesimal generator
for a continuous-time Markov chain with state space {0, 1, . . . , m}. Let

T T0
 
Q̃ = ,
0 0

where T is an m-square matrix, T 0 is an m × 1 column vector, and the

remaining terms are chosen to conform, be a matrix obtained by replacing
any row of Q by a row of zeros and then exchanging rows so that the
final row is a vector of zeros. Let P (t) = [ Pt (t) Pa (t) ] denote the state
probability vector for the Markov chain for which Q̃ is the infinitesimal
generator, with Pt (t) a row vector of dimension m and Pa (t) a scalar.

(a) Argue that if Q is the infinitesimal generator for an irreducible Markov

chain, then the states 0, 1, . . . , m−1 of the modified chain are all tran-
sient, and state m is an absorbing state.
(b) Prove that if Q is the infinitesimal generator for an irreducible Markov
chain, then the matrix T must be nonsingular. [Hint: Solve for Pt (t),
then prove by contradiction. Make use of the fact that if T is singular,
then T has a zero eigenvalue.]
(c) Show that Pa (t) = 1 − Pt (0) exp {T t}e. [Hint: Use the fact that
Pt (t)e is the probability that the state of the modified Markov chain
is in the set {0, . . . , m − 1} at time t.]
(d) Let x̃ be the time required for the modified Markov chain to reach
state m given an initial probability vector P (0) = [ Pt (0) 0 ], that
is, with Pt (0)e = 1. Argue that P {x̃ ≤ t} = Pa (t), that is,

P {x̃ ≤ t} = 1 − Pt (0) exp {T t}e.

(e) Argue that if Q is the infinitesimal generator for an irreducible Markov

chain, then the matrix T̃ = T +T 0 Pt (0) is the infinitesimal generator
for an irreducible Markov chain with state space {0, 1, . . . , m − 1}.

Solution:

(a) Since the Markov chain is irreducible, all states can be reached from
all other states; therefore, state m is reachable from all other states in
the original chain. Since the dynamics of the system have not been
changed for states {0, 1, · · · , m − 1} in the modified chain, state m
is still reachable from all states. But, once in state m, the modified
system will never leave state m because its departure rate to all other
134 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

states is 0. Therefore, all states other than state m are transient states
in the modified chain and state m is an absorbing state.
(b) Observe that
d
[ Pt (t) Pa (t) ] = Pt (t)T
dt
so that
d
Pt (t) = Pt (0)eT t
dt
Now, we know that
lim Pt (t) = 0
t→0

is independent of Pt (0). This means that all eigenvalues of T must

have strictly negative real parts. But, if all eigenvalues have strictly
negative real parts, the matrix T cannot be singluar.
(c) From the matrix above, d/dtPa (t) = Pt (t)T 0 . Therefore,
d
Pa (t) = Pt0 eT t T 0
dt
Now, Pa (t) is the probability that the system is in an absorbing state
at time t, which is equal to the complement of the probability that the
system is in some other state at time t. Hence,

Pa (t) = 1 − Pt (t)e
= 1 − Pt (0)eT t e

Furthermore, from the differential equation,

Pa (t) = Pt (0)eT t T −1 T 0 + K

But T e + T 0 = 0, so that T −1 T 0 = −e. Thus,

Pa (t) = K − Pt (0)eeT t e,

and
lim Pa (t) = 1 = 1 − 0
t→0

implies that K = 1, and Pa (t) = 1 − Pt (0)eT t e.

(d) The argument is straight-forward. If the chain is in the absorbing
state at time t, then it necessarily has left the transient states by time
t. Thus,

P {x̃ ≤ t} = Pa (t)
= 1 − Pt (0)eT t e
Advanced CTMC-Based Queueing Models 135

(e) The vector T 0 represents the rates at which the system leaves the
transient states to enter the absorbing state. The matrix T 0 Pt (0) is
then a nonnegative m × m matrix. Since [T T 0 ]e = 0, it follows that
T + T0 Pt (0) is a matrix whose diagonal terms are nonpositive. This
is because Tii + Ti0 ≤ 0 so that Tii + Ti0 Pti (0) ≤ 0. This is due to
the fact that Pti (0) ≤ 1. All nondiagonal terms are nonnegative since
all off-diagonal terms of T are nonnegative and T 0 is nonnegative. It
remains only to determine if
T̃e = [ T + T 0 Pt (0) ] e = 0
Now, Pt (0)e = 1, so [T + T 0 Pt (0)]e = T e + T 0 , which we al-
ready know is equal to a zero vector. Thus, T + T 0 Pt (0) is the in-
finitesimal generator for an irreducible Markov chain with state space
{0, 1, . . . , m − 1}.

4-2 Consider an m-server queueing system having Poisson arrivals. Suppose

upon entering service, each customer initially receives a type 1 service
increment. Each time a customer receives a type 1 service increment, the
customer leaves the system with probability (1 − p) or else receives a
type 2 service increment followed by an additional type 1 service incre-
ment. Suppose type 1 and type 2 service increment times are each drawn
independently from exponential distributions with parameters µ1 and µ2 ,
respectively. With the service process defined as in Problem 4-1, suppose
there are m servers. Define the phase of the system to be j if there are j
customers receiving a type 2 service increment, j = 0, 1, · · · , m. Define
the state of the system to be 0 when the system is empty and by the pair
(i, j) where i ≥ 0 is the system occupancy and j = 0, · · · , i is the phase of
the service process. Define Pi = [ Pi0 Pi1 · · · Pi,min{i,m} ] for i > 0
and P0 , a scalar.

(a) Draw the state transition diagram for the special case of m = 3.
(b) Write the matrix balance equations for the special case of m = 3.
(c) Write the matrix balance equations for the case of general values of
m.
(d) Determine the matrix Q, the infinitesimal generator for the continuous-
time Markov chain defining the occupancy process for this system.
(e) Comment on the structure of the matrix Q relative to that for the
phase-dependent arrival and service rate queueing system and to the
M/PH/1 system. What modifications in the solution procedure would
have to be made to solve this problem? [Hint: See Neuts [1981a], pp.
24-26.]
136 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

Solution:

(a) For this exercise, we may think of a customer as staying with the same
server thoughout its service; the server simply does a different task.
See Figure 4.4 for the state diagram of the system.

(b) We first introduce the following notation, where Ik is the k×k identity
matrix:
Λ0 = λ, Λ̂0 = [ λ 0 ] 
λ 0 0
Λ1 = λI2 , Λ̂1 =
0 λ 0
λ 0 0 0


Λ2 = λI3 , Λ̂2 =  0 λ 0 0 
0 0 λ 0
Λ3 = λI
 4
µ1 (1 − ρ) 0 µ1 (1 − ρ)
  
M1 = , M̂1 = ,
0 0  0
2µ1 (1 − ρ) 0 0


M2 =  0 µ1 (1 − ρ) 0  ,
0 0 0
2µ1 (1 − ρ) 0


M̂2 =  0 µ1 (1 − ρ)  ,
0 0
3µ1 (1 − ρ) 0 0 0
 
 0 2µ1 (1 − ρ) 0 0
M3 =  ,
 0 0 µ1 (1 − ρ) 0 
0 0 0 0
3µ1 (1 − ρ) 0 0
 
 0 2µ1 (1 − ρ) 0 
M̂3 =  ,
 0 0 µ1 (1 − ρ) 
0 0  0
0 −2µ1 0

−µ1 ρ
 
0
P1 = , P2 =  −µ2 0 −µ1 ρ  ,
−µ2 0
0 −2µ2 0
0 −3µ1 0 0
 

P3 =  −µ2 0 −2µ1 0  .
0 0 −3µ2 0
With this notation, we may compactly write the balance equations as
follows:
P0 Λ0 = P1 M̂1
P1 [ Λ1 + P1 + M1 ] = P0 Λ̂0 + P2 M̂2
P2 [ Λ2 + P2 + M2 ] = P1 Λ̂0 + P3 M̂3
Advanced CTMC-Based Queueing Models 137

λ (1-p)µ 1

p µ1
1,0 1,1
µ 2

λ (1-p)µ 1 λ (1-p)µ 1

p µ1 pµ1
2,0 2,1 2,2
µ 2
µ 2

λ (1-p)µ 1 λ (1-p)µ 1 λ (1-p)µ 1

p µ1 pµ1 p µ1
3,0 3,1 3,2 3,3
µ 2
µ 2
µ 2

λ (1-p)µ 1 λ (1-p)µ 1 λ (1-p)µ 1 λ (1-p)µ 1

p µ1 pµ1 p µ1
4,0 4,1 4,2 4,3

µ 2
µ 2
µ 2

λ (1-p)µ 1 λ (1-p)µ 1 λ (1-p)µ 1 λ (1-p)µ 1

p µ1 pµ1 p µ1

5,0 5,1 5,2 5,3

µ 2
µ 2
µ 2

λ (1-p)µ 1 λ (1-p)µ 1 λ (1-p)µ 1 λ (1-p)µ 1

Figure 4.4. State Diagram for Problem 2a

P3 [ Λ3 + P3 + M3 ] = P2 Λ̂2 + P4 M3 ,
138 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

and for n ≥ 4,
Pn [ Λ3 + P3 + M3 ] = Pn−1 Λ̂3 + Pn+1 M3

(c) To obtain the balance equations for general m, we simply define Λi ,

i = 1, 2, · · · , m; Λ̂i , i = 0, 1, · · · , m − 1; and Pi , Mi , M̂i , i =
0, 1, · · · , m as in part (b). The equations will then follow the same
pattern as in part (b). i.e.,
P0 Λ0 = P1 M̂1 ,
Pn [ Λn + Pn + Mn ] = Pn−1 Λ̂n−1 + Pn+1 Mn+1 ,
where n = 1, 2, · · · , m. For n > m,
Pn [ Λm + Pm + Mm ] = Pn−1 Λ̂m + Pn+1 Mm

(d) Define
∆ = diag[−Λ0 , − [Λ1 + P1 + M1 ) , − (Λ2 + P2 + M2 ) ,
= −h (Λ3 + P3 + M3 ) , − (Λ3 + i P3 + M3 ) , · · ·]
∆u = superdiag Λ̂0 , Λ̂1 , Λ̂2 , Λ3 , Λ3 , Λ3 , · · ·
h i
∆l = subdiag M̂0 , M̂1 , M̂2 , M̂3 , M3 , M3 , · · ·

Then, by inspection of the balance equations in part (c),

Q̂ = ∆ + ∆u + ∆l

(e) This exercise will be harder to solve than the M/PH/1 system; this is
due to the fact that there is memory from service to service. That is,
a new customer starts service at exactly the phase of the previously
completed customer as long as there are customers waiting. In the
M/PH/1 system, the starting phase is selected at random. Since there
is more than one possible starting phase, this problem will be more
difficult to solve. This matrix Q has the form
 
B00 B01 B02 B03 0 ...
B B11 B12 B13 0 ...
 10


B B21 B22 B23 0 ... 
 20 
 B30 B31 B32 B33 A0 ...
 

 0 0 0 B43 A1 A0 ...
 
,
 0 0 0 0 A2 A1 A0 ...
 
 
 0 0 0 0 0 A2 A1 ...
 
 0 0 0 0 0 0 A2 ...
.. .. .. .. .. .. ..
 
..
. . . . . . . .
Advanced CTMC-Based Queueing Models 139

which is similar to the form of the matrix of the G/M/1 type, but with
more complex boundary conditions.
(f) We know that P4 = P3 R, but we do not know P0 , P1 , P2 , P3 . On the
other hand, we know P Q = 0. Now consider only those elements of
P that are multiplied by the B submatrices. That is, P0 , P1 , P2 , P3 , P4 .
Then,
B00 B01 B02 B03
 
 B10 B11 B12 B13 
 
0 = [ P0 P1 P2 P3 P4 ] 
 B20 B21 B22 B23 

B B31 B32 B33 
30
0 0 0 B43
But P4 = P3 R, so that
B00 B01 B02 B03
 
 B10 B11 B12 B13 
 
0 = [ P0 P1 P2 P3 P3 R ] 
 B20 B21 B22 B23 

B B31 B32 B33 
30
0 0 0 B43
Then combining the columns involving P3 , we have
B00 B01 B02 B03
 
 B10 B11 B12 B13 
0 = [ P0 P1 P2 P3 ] 
 B20

B21 B22 B23 
B30 B31 B32 B33 + RB43
= [ P0 P1 P2 P3 ] B(R)

Thus, we see that the vector [ P0 P1 P2 P3 ] is proportional to

the left eigenvector of B(R) corresponding to its zero eigenvalue. We
may then write the zero eigenvalue of B(R) as x and [ P0 P1 P2 P3 ] =
k [ x0 x1 ], where k is a normalizing constant, and x1 is proportional
to P3 . Thus,
∞
kx1 Ri e = i,
X
kx0 e +
i=1
or
kx0 e + kx1 [I − R]−1 e = 1,
so that
1
k=
x0 e + kx1 [I − R]−1 e
Once k is know, we can compute

[ P0 P1 P2 ] = kx0
140 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

P3 = kx1

and
Pn+3 = P3 Rn , n≥1
Chapter 5

THE BASIC M/G/1 QUEUEING SYSTEM

Exercise 5.1 With x̃, {ñ(t), t ≥ 0}, and ỹ defined as in Theorem 5.2,
show that E[ỹ(ỹ − 1) · · · (ỹ − n + 1)] = λn E[x̃n ].
P∞ yP
Solution. Recall that Fỹ (z) = y=0 z {ỹ = y}. Then
∞
dn
[y(y − 1) · · · (y − n + 1)] z y−n P {ỹ = y} ,
X
Fỹ (z) =
dz n ỹ=0

so that
dn


Fỹ (z) = E [ỹ(ỹ − 1) · · · (ỹ − n + 1)] .
dz n 
z=1

On the other hand,

dn ∗ dn ∞ −λ[1−z]
Z
F (λ[1 − z]) = e dFx̃ (x)
dz n x̃ n
Zdz∞ 0
= (λx)n e−λ[1−z] dFx̃ (x),
0

which implies
dn ∗

= λn E[x̃n ].

n
Fx̃ (λ[1 − z])
dz z=1

Since Fỹ (z) = Fx̃∗ (λ[1 − z]), it follows that

E [ỹ(ỹ − 1) · · · (ỹ − n + 1)] = λn E[x̃n ]

Alternatively,
Fỹ (z) = Fx̃∗ (s(z)) ,
142 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

where s(z) = λ[1 − z] implies

n
dn dn ∗ d

Fỹ (z) = Fx̃ (s(z)) s(z)
dz n ds n dz
dn ∗
= F (−λ)n
dsn nx̃
= (−λ) [(−1)n E[x̃n ]]
= λn E[x̃n ].

Exercise 5.2 Prove Theorem 5.3.

Solution. The proof follows directly from the definition of Fx̃ (z).
∞
z i P (x̃ − 1)+ = i
X 
F(x̃−1)+ (z) =
i=0
∞
= z 0 P (x̃ − 1)+ = 0 + z i P (x̃ − 1)+ = i .
 X 

i=1

But (x̃ − 1)+ = 0 when either x̃ = 1 or x̃ = 0, so that

∞
1X
F(x̃−1)+ (z) = P {x̃ = 0} + P {x̃ = 1} + z (i+1) P {x̃ = i + 1}
z i=1
= P {x̃ = 0} + P {x̃ = 1}
1
= + [Fx̃ (z) − P {x̃ = 0} − zP {x̃ = 1}]
z
1 1
 
= 1− P {x̃ = 0} − Fx̃ (z).
z z

Exercise 5.3 Starting with (5.5), use the properties of Laplace trans-
forms and probability generating functions to establish (5.6).
Solution. Observe that if we take the limit of the right-hand side of (5.5)
as z → 1, both numerator and denominator go to zero. Thus we can apply
L’Hôpital’s rule, and then take the limit as z → 1 of the resultant expression,
using the properties of Laplace transforms and probability generating func-
tions. Let α(z) denote the numerator and β(z) denote the denominator of the
right-hand side of (5.5). Then the derivatives are
d d
 
α(z) = −P {q̃ = 0} Fx̃∗ (λ[1 − z]) − (1 − z) Fx̃∗ (λ[1 − z])
dz dz
d d ∗
β(z) = F (λ[1 − z]) − 1.
dz dz x̃
Now take the limits of both sides as z → 1. It is easy to verify using Theo-
d ∗
rem 2.2 that limz→1 dz Fx̃ (λ[1 − z]) = λE[x̃] = ρ. Thus, after some simple
The Basic M/G/1 Queueing System 143

algebra,
P {q̃ = 0}
− = 1.
ρ−1
i.e.,
P {q̃ = 0} = 1 − ρ.

Exercise 5.4 Establish (5.6) directly by using Little’s result.

Solution. Define the system to be the server only. Then there is either one in
the system (i.e., in service) or zero, so that

E[ñ] = 0 · P {ñ = 0} + 1 · P {ñ = 1}

= P {ñ = 1} .

Note that this is the probability that the server is busy. On the other hand, we
have by Little’s result that

E[ñ] = λE[s̃]
= λE[x̃]
= ρ.

i.e., the probability that the server is busy is ρ, or, equivalently, that the proba-
bility that no one is in the system is (1 − ρ). Now, it has already been pointed
out that the Poisson arrival’s view of the system is the same as that of a ran-
dom observer. Thus, P {ñ = n} = P {q̃ = n}. In particular, we see that
P {q̃ = 0} = 1 − ρ.
Exercise 5.5 Batch Arrivals. Suppose arrivals to the system occur in
batches of size b̃, and the batches occur according to a Poisson process at
rate λ. Develop an expression equivalent to (5.5) for this case. Be sure to
define each of the variables carefully.
Solution. It should be easy to see that the basic relationship

q̃n+1 = (q̃n − 1)+ + ṽn+1

will remain valid. i.e., the number of customers left in the system by the (n +
1)-st departing customer will be the sum of the number of customers left by the
n-th customer plus the number of customers who arrive during the (n + 1)-st
customer’s service. It should be clear, however, that the number of customers
who arrive during the (n + 1)-st customer’s service will vary according to the
batch size, and so we need to re-specify ṽn+1 . Now,

Fq̃n+1 (z) = F(q̃n −1) (z)Fṽn+1 (z)

1 1
  
= 1− P {q̃n = 0} + Fq̃n (z) Fṽn+1 (z).
z z
144 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

Taking the limit of this equation as n → ∞,

1 1
  
Fq̃ (z) = 1− P {q̃ = 0} + Fq̃ (z) Fṽ (z)
 z
 z
1
1 − z P {q̃ = 0} Fṽ (z)
= .
1 − 1z Fq̃ (z)Fṽ (z)

But, conditioning upon the batch size and the length of the service, we find that

Fṽ (z) = E[z ṽ ]

Z ∞
∞X   
E z ṽ ñ = n, x̃ = x P {ñ = n|x̃ = x} dFx̃ (x)

=
0 n=0
Z ∞X∞ h Pn i (λx)n
b
= E z i=1 i e−λx dFx̃ (x)
0 n=0
n!
∞ h i (λx)n
( )
Z ∞
b̃
e−λx dFx̃ (x)
X
= E z
0 n=0
n!
Z ∞
λxFb̃ (z)−λx
= e dFx̃ (x)
0
Fx̃∗
 

= λ 1 − Fb̃ (z) .
The Basic M/G/1 Queueing System 145

Exercise 5.6 Using the properties of the probability generating function,

show that
λρ E[x̃2 ]
E[ñ] = ρ +
1 − ρ 2E[x̃]
ρ Cx̃ 2 + 1 
 
= ρ 1 + .


  (5.1)
1−ρ 2


[Hint: The algebra will be greatly simplified if (5.8) is first rewritten as

Fñ (z) = α(z)/β(z),

where
α(z) = (1 − ρ)Fx̃∗ (λ[1 − z]),
and
1 − Fx̃∗ (λ[1 − z])
β(z) = 1 − .
1−z
Then, in order to find
d
lim Fñ (z),
dz
z→1

first find the limits as z → 1 of α(z), β(z), dα(z)/dz, and dβ(z)/dz, and
then substitute these limits into the formula for the derivative of a ratio.
Alternatively, multiply both sides of (5.8) to clear fractions and then differ-
entiate and take limits.]
α(z)
Solution. Following the hint given in the book, we first rewrite Fñ (z) as β(z) ,
where
α(z) = (1 − ρ)Fx̃∗ (λ[1 − z]),

and
1 − Fx̃∗ (λ[1 − z])
β(z) = 1 − .
1−z
d d
It is straight forward to find the limit as z → 1 of dz Fñ , α(z) and dz α(z):

lim α(z) = 1 − ρ
z→1
d
lim α(z) = (1 − ρ)ρ
z→1 dz
d
lim Fñ (z) = E[ñ].
z→1 dz

The limit as z → 1 of β(z) and its derivative are more complicated. Note that
if we take the limit of β(z) that both the numerator and the denominator of the
146 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

fraction go to zero. We may apply L’Hôpital’s rule in this case.

1 − Fx̃∗ (λ[1 − z])
 
lim β(z) = 1 − lim
z→1 z→1 " 1−z
d ∗
#
λ dz Fx̃ (λ[1 − z])
= 1 − lim
z→1 −1
= 1 − λE[x̃]
= 1 − ρ.

Similarly, we compute
d d 1 − Fx̃∗ (λ[1 − z])
 
β(z) = −
dz dz  1−z 
d 1 − Fx̃∗ (s(z))
= −
dz 1−z
d ∗
−(1 − z)λ ds Fx̃ (s(z)) − [1 − Fx̃∗ (s(z))]
= .
(1 − z)2

Upon taking the limit of this ratio and applying L’Hôpital’s rule twice, we see
that
d 2
(1 − z)λ2 ds ∗
" #
d 2 Fx̃ (s(z))
lim β(z) = lim
z→1 dz z→1 2(z − 1)
d 3 2
−(1 − z)λ3 ds ∗ 2 d ∗
" #
3 Fx̃ (s(z)) − λ ds2 Fx̃ (s(z))
= lim
z→1 2
−λ 2
= E[x̃2 ].
2

To find E[ñ], we substitute these limits into the formula for the derivative of
Fñ (z),
d d
d β(z) dz α(z) − α(z) dz β(z)
Fñ (z) = 2
dz β (z).
Thus,
2
(1 − ρ)2 ρ + (1 − ρ) λ2 E[x̃2 ]
E[ñ] =
(1 − ρ)2
λ E[x̃2 ]
2
= ρ+
1−ρ 2
λ2 (E[x̃2 ] − E 2 [x̃] + E 2 [x̃])
= ρ+
1−ρ 2
The Basic M/G/1 Queueing System 147
!
λ2 E[x̃2 ] − E 2 [x̃] + E 2 [x̃]
= ρ+ E 2 [x̃]
1−ρ 2E 2 [x̃]
ρ2 V ar(x̃) + 1
 
= ρ+
1−ρ 2E 2 [x̃]!
ρ 2 Cx̃ 2 + 1
= ρ+
1−ρ 2
ρ Cx̃ 2 + 1 
 
= ρ
1 + ,

 

1−ρ 2

Exercise 5.7 Let δn = 1 if q̃n = 0 and let δn = 0 if q̃n > 0 so that

q̃n+1 = q̃n − 1 + δn + ṽn+1 . Starting with this equation, find E[δ∞ ] and
E[ñ]. Interpret E[δ∞ ]. [Hint: To find E[ñ], start off by squaring both sides
of the equation for q̃n+1 .]
Solution. First take expectations of the equation for q̃n+1 .

E[q̃] = E[q̃] − 1 + E[δ∞ ] + E[ṽ].

By Little’s result, we know that E[ṽ] = λE[x̃] = ρ. Substitute this value into
the above equation to find that E[δ∞ ] = 1 − ρ.
To find E[ñ], we begin by squaring the equation given in the problem state-
ment.

E[q̃ 2 ] = E[q̃ 2 ] + E[ṽ 2 ] + E[δ∞

2
] + 1 + 2E[q̃ṽ] + 2E[q̃δ] + 2E[ṽδ∞ ]
−2E[q̃] − 2E[ṽ] − 2E[δ∞ ].

But E[q̃] = E[ñ], and E[δ∞ 2 ] = E[δ ] since δ can take on values of 0 or 1
∞ n
only. In addition, q̃n and ṽn+1 are independent, as are δn and ṽn+1 . Clearly
their limits are also independent. However, δn and q̃n are not independent.
Furthermore, their product will always be zero; hence their expected value will
also be zero. Using these observations and solving for E[ñ],
2
E[ñ] (2 − 2E[ṽ]) = E[δ∞ ] + E[ṽ 2 ] + 1 + 2E[ṽ]E[δ∞ ] − 2E[ṽ] − 2E[δ∞ ].

That is,
ρ(1 − 2ρ) + E[ṽ 2 ]
E[ñ] = .
2(1 − ρ)
But by Exercise 5.1, E[ṽ 2 ] − E[ṽ] = λ2 E[x̃2 ], so that E[ṽ 2 ] = ρ2 E[x̃2 ] + ρ.
Substituting this into the expression for E[ñ],
ρ
E[ñ] = .
1−ρ
148 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

Exercise 5.8 Using (5.14) and the properties of the Laplace transform,
show that
ρ E[x̃2 ]
E[s̃] = + E[x̃]
1 − ρ 2E[x̃]
ρ Cx̃ 2 + 1
 
=  + 1  E[x̃].
 

  (5.2)
1−ρ 2
Combine this result with that of Exercise 5.6 to verify the validity of Little’s
result applied to the M/G/1 queueing system. [Hint: Use (5.15) rather than
(5.14) as a starting point, and use the hint for Exercise 5.6]
Solution. Begin with (5.15) and rewrite Fs̃∗ (s) as α(s)/β(s), where
1 − Fx̃∗ (s)
 
β(s) = 1 − ρ .
sE[x̃]
d ∗ d
Simple calculations give us the limits of ds Fs̃ (s), α(s), and ds α(s) as s → 0.

lim α(s) = 1 − ρ
s→0
d
lim α(s) = −(1 − ρ)E[x̃]
s→0 ds
d
lim Fs̃ (s) = −E[s̃].
s→0 ds

Taking the limit of the β(s)as s → 0 results in both a zero numerator and
denominator in the second term. We may apply L’Hôpital’s rule in this case to
find
d ∗
" #
− ds Fx̃ (s)
lim β(s) = 1 − ρ lim
s→0 s→0 E[x̃]
E[x̃]
= 1−ρ
E[x̃]
= 1 − ρ.
′
We now find the limit of β (s), where
d ∗
Fx̃ (s) + [1 − Fx̃∗ (s)] E[x̃]
" #
d ρ sE[x̃] ds
β(s) = .
ds E[x̃] s2

Then, taking the limit of this expression as s → 0 and twice applying L’Hôpital’s
rule,
d ∗ d 2 d ∗
∗
" #
d ρ ds Fx̃ (s) + s ds 2 Fx̃ (s) − ds Fx̃ (s)
lim β(s) = lim
s→0 ds E[x̃] s→0 2s
The Basic M/G/1 Queueing System 149

d2 ∗ d 3
∗
" #
ρ ds2 Fx̃ (s) + s ds 3 Fx̃ (s)
= lim
E[x̃] s→0 2
E[x̃2 ]
= ρ .
2E[x̃]

Substituting these expressions into the formula for the derivative of Fs̃∗ (s), we
find that

−(1 − ρ)2 E[x̃] − (1 − ρ) λ2 E[x̃2 ]

−E[s̃] =
(1 − ρ)2
ρ E[x̃2 ]
= −E[x̃] −
1 − ρ 2E[x̃]
ρ (E[x̃2 ] − E 2 [x̃] + E 2 [x̃])
= −E[x̃] −
1−ρ 2E[x̃] !
ρ E[x̃2 ] − E 2 [x̃] + E 2 [x̃]
= −E[x̃] − E[x̃]
1−ρ 2E 2 [x̃]
!
ρ V ar(x̃) + E 2 [x̃]
= −E[x̃] − E[x̃]
1−ρ 2E 2 [x̃]
!
ρ Cx̃ 2 + 1
= −E[x̃] − E[x̃]
1−ρ 2
2
 
ρ Cx̃ + 1
= −E[x̃]  + 1 .

 

1−ρ 2


i.e.,

ρ Cx̃ 2 + 1
 
E[s̃] = E[x̃]  + 1
.
 
 
1−ρ 2


To complete the exercise, combine this result with that of Exercise 5.6.
 
ρ Cx̃ 2 +1 
E[ñ] ρ 1 + 1−ρ 2
=  2

E[s̃] ρ Cx̃ +1
E[x̃]  1−ρ + 1
2
ρ
=
E[x̃]
= λ.

This verifies Little’s result applied to the M/G/1 queueing system.

150 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

Exercise 5.9 Using (5.17) and the properties of the Laplace transform,
show that
ρ Cx̃ 2 + 1
E[w̃] = E[x̃]. (5.3)
1−ρ 2
Combine this result with the result of Exercise 5.6 to verify the validity
of Little’s result when applied to the waiting line for the M/G/1 queueing
system.
Solution. Recall equation (5.17):
(1 − ρ)
Fw̃∗ (s) = h
1−Fx̃∗ (s)
i.
1−ρ sE[x̃]

As in Exercise 5.8, we rewrite Fw̃∗ (s) as (1 − ρ)β −1 (s), where

1 − Fx̃∗ (s)
 
β(s) = 1 − ρ .
sE[x̃]
To compute E[w̃], we need to find
d ∗ d
Fw̃ (s) = −(1 − ρ)β −2 (s) β(s),
ds ds
and its limit as s → 0. We first calculate β(s) at s = 0. Clearly, however, this
would result in both a zero numerator and denominator. Therefore, applying
L’Hôpital’s rule,
d ∗
" #
− ds Fx̃ (s)
lim β(s) = 1 − ρ lim
s→0 s→0 E[x̃]
= 1 − ρ.

d
We now calculate the limit of ds β(s) by applying L’Hôpital’s rule twice. This
results in
d ∗
Fx̃ (s) − [1 − Fx̃∗ (s)]
" #
d ρ −s ds
lim β(s) = lim −
s→0 ds s→0 E[x̃] s2
d ∗ d 2 d ∗
∗
" #
ρ ds Fx̃ (s) + s ds 2 Fx̃ (s) − ds Fx̃ (s)
= lim
s→0 E[x̃] 2s
d2 ∗ d 3
∗
" #
ρ ds2 Fx̃ (s) + s ds 3 Fx̃ (s)
= lim
s→0 E[x̃] 2
E[x̃2 ]
= ρ .
2E[x̃]
The Basic M/G/1 Queueing System 151

Noting that E[w̃] is −d/dsFw̃∗ (s) at s = 0, we may now specify E[w̃] as

d ∗
E[w̃] = − lim F (s)
s→0 ds w̃
E[x̃2 ]
= (1 − ρ)(1 − ρ)−2 ρ
2E[x̃]
ρ E[x̃2 ]
 
=
1 − ρ" 2E[x̃] #
ρ E 2 [x̃] − E 2 [x̃] + E 2 [x̃]
=
1−ρ 2E[x̃]
ρ Cx̃2 + 1
= E[x̃].
1−ρ 2

Combine this result with that of Exercise 5.6 to complete the exercise. That is,
show that
E[ñ] − λE[x̃]
= E[w̃],
λ
where E[ñ] − λE[x̃] = E[ñq ], the expected number of customers in queue.
" #
E[ñ] − λE[x̃] ρ Cx̃ 2 + 1
= E[x̃] 1 + − E[x̃]
λ 1−ρ 2
" #
ρ Cx̃ 2 + 1
= E[x̃] 1 + −1
1−ρ 2
ρ Cx̃ 2 + 1
= E[x̃]
1−ρ 2
= E[w̃].

Exercise 5.10 Using properties of the Laplace transform, show that

E[x̃]
E[ỹ] = . (5.4)
1−ρ

Solution. Recall equation (5.21)

Fỹ∗ (s) = Fx̃∗ [s + λ − λFỹ∗ (s)].

We see that Fỹ∗ (s) has the form f (s) = F (g(s)), so that

d d d
f (s) = F (g(s)) · g(s).
ds dg ds
152 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

This results in the expression

d ∗ d ∗ d  
F (s) = Fx̃ (s) s + λ − λFỹ∗ (s)
ds ỹ ds ds
d ∗ d
 
= Fx̃ (s) 1 − λ Fỹ∗ (s) .
ds ds

d ∗

Evaluate both sides at s = 0, and recall that ds Fỹ (s)
 = −E[ỹ], and
 s=0
d ∗

F
ds x̃ (s)
 = −E[x̃]. Thus,
s=0

−E[ỹ] = −E[x̃] (1 + λE[ỹ]) .

Solving for E[ỹ], this equation becomes

E[x̃]
E[ỹ] = .
1−ρ

Exercise 5.11 For the ordinary M/G/1 queueing system determine E[ỹ]
without first solving for Fỹ (s). [Hint: Condition on the length of the first
customer’s service and the number of customers that arrive during that pe-
riod of time.]
Solution. To determine E[ỹ], condition on the length of the first customer’s
service. Z ∞
E[ỹ] E[ỹ|x̃ = x]dFx̃ (x). (5.11.1)
0
Next, condition this expected value on the number of customers who arrive
during the first customer’s service. i.e.,
∞
X (λx)v −λx
E[ỹ|x̃ = x] = E[ỹ|x̃ = x, ṽ = v] e . (5.11.2)
v=0
v!

Now, given that the length of the first customer’s service is x and the number
of customers who arrive during this time is v, it is easy to see that the length of
the busy period will be the length of the first customer’s service plus the length
of the sub busy periods generated by those v customers. That is,
v
X
E[ỹ|x̃ = x, ṽ = v] = x + E[ỹi ]
i=0

where ỹ0 = 0 with probability 1. Thus,

v
X
E[ỹ|x̃ = x, ṽ = v] = x + E[ỹ]
i=1
The Basic M/G/1 Queueing System 153

= x + vE[ỹ],

since each of the sub busy periods ỹi are drawn from the same distribution. If
we substitute this expression into (5.11.2),
∞
X (λx)v −λx
E[ỹ|x̃ = x] = (x + vE[ỹ]) e
v=0
v!
= x + E[ṽ]E[ỹ]
= x + ρE[ỹ].

To complete the derivation, substitute this result into (5.11.1) and solve for
E[ỹ]. Thus,
Z ∞
E[ỹ] = (x + ρE[ỹ]) dFx̃ (x)
Z0∞
= xdFx̃ (x) + ρE[ỹ]
0
= E[x̃] + ρE[ỹ].

i.e.,
E[x̃]
E[ỹ] = .
1−ρ

Exercise 5.12 M/G/1 with Exceptional First Service. A queueing sys-

tem has Poisson arrivals with rate λ. The service time for the first customer
in each busy period is drawn from the service-time distribution Fx̃e (x), and
the remaining service times in each busy period are drawn from the gen-
eral distribution Fx̃ (x). Let ỹe denote the length of the busy period for this
system. Show that
E[x̃e ]
E[ỹe ] = ,
1−ρ
where ρ = λE[x̃].
Solution. Let ṽe denote the number of customers who arrive during the first
customer’s service. Then it is easy to see that

E[ỹe ] = E[x̃e ] + E[ṽe ]E[ỹ],

where ỹ denotes the length of the busy period for each of the subsequent ṽe
customers. By Little’s result applied to E[ṽe ], E[ṽe ] = λE[x̃e ]. Then, using
the result from Exercise 5.11,
ρ
E[ỹe ] = E[x̃e ] + E[x̃e ]
1−ρ
154 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

E[x̃e ]
= .
1−ρ

Exercise 5.13 For the M/G/1 queueing system with exceptional first ser-
vice as defined in the previous exercise, show that Fỹ∗e (s) = Fx̃∗e (s + λ −
λFỹ∗ (s)).
Solution. Recall that we may write Fỹ∗ (s) as E[e−sỹe ]. As in the proof of
Exercise 5.11, we shall first condition this expected value on the length of the
first customer’s service, and then condition it again on the number of customers
who arrive during the first customer’s service. That is,
Z ∞
−sỹe
E[e ] = E[e−sỹe |x̃e = x]dFx̃ (x), (5.13.1)
0

where
∞
−sỹe
E[e−sỹe |x̃e = x, ṽ = v]P {ṽ = v} dFx̃ (x).
X
E[e |x̃e = x] = (5.13.2)
v=0

Here, x̃e represents the length of the first customer’s service and ṽ is the num-
ber of arrivals during that time. Now, the length of the busy period will be the
length of the first customer’s service plus the length of the sub busy periods
generated by those customers who arrive during the first customer’s service.
Hence, h P∞ i
E[e−sỹe |x̃e = x, ṽ = v] = E e−s(x+ i=0 ỹi ) ,
where ỹ0 = 0 with probability 1. Furthermore, x and v are constants at this
stage, and the ỹi are independent identically distributed random variables, so
that
v
−sỹe −sx
E[e−sỹi ]
Y
E[e |x̃e = x, ṽ = v] = e
i=1
= e−sx E v [e−sỹ ].
Substituting this into (5.13.2),
∞
v
−sỹe −sx −λx
X E[e−sỹ ]λx
E[e |x̃e = x] = e e
0
v!
−sx −λx (−λxE[e−sỹ ])
= e e e .
To complete the proof, substitute this into (5.13.1), and then combine the ex-
ponents of e.
Z ∞
e−x(s+λ−λE[e ) dF (x)
−sỹe −sỹ ]
E[e ] = x̃
0
The Basic M/G/1 Queueing System 155
 
= Fỹ∗e s + λ − λE[e−sỹ ]
 
= Fỹ∗e s + λ − λFỹ∗ (s) .

Exercise 5.14 Comparison of the formulas for the expected waiting

time for the M/G/1 system and the expected length of a busy period for
the M/G/1 system with the formula for exceptional first service reveals that
they both have the same form; that is, the expected waiting time in an ordi-
nary M/G/1 system is the same as the length of the busy period of an M/G/1
system in which the expected length of the first service is given by

E[x̃2 ]
E[x̃e ] = ρ .
2E[x̃]
Explain why these formulas have this relationship. What random variable
must x̃e represent in this form? [Hint: Consider the operation of the M/G/1
queueing system under a nonpreemptive, LCFS, service discipline and ap-
ply Little’s result, taking into account that an arriving customer may find
the system empty.]
Solution. Consider the waiting time under LCFS. We know from Little’s result
that E[w̃LCF S ] is the same as E[w̃F CF S ] because E[ñF CF S ] = E[ñLCF S ].
But E[w̃LCF S |no customers present] = 0 and E[w̃LCF S |at least 1 customer
present] = E[length of busy period started by customer in service] = E[x̃e ]/(1−
ρ). But the probability of at least one customer present is ρ, so that
ρ
E[w̃] = E[x̃e ].
1−ρ
On the other hand, we know that
ρ E[x̃2 ]
E[w̃] = .
1 − ρ 2E[x̃]
Therefore
E[x̃2 ]
E[x̃e ] =
2E[x̃]
represents the expected remaining service time of the customer in service, if
any, at the arrival time of an arbitrary customer.
Exercise 5.15 Show that the final summation on the right-hand side of
(5.28) is K + 1 if ℓ = (K + 1)m + n and zero otherwise.
ℓ−n
Solution. First, suppose that ℓ = (K + 1)m + n. Then K+1 = m, and the
final summation is
K K
−j2πkm
X X
e = 1=K+1
k=0 k=0
156 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

Now suppose that ℓ is not of the above form; that is, ℓ = (K + 1)m + n + r,
0 < r < (K + 1). Then

2πk(ℓ − n) 2πk [m(K + 1) + r] 2πkr

= = 2πkm + ,
K +1 K +1 K +1

and the final summation becomes

K 2πk[m(K+1)+r]
K
2πkr
e−j e−j K+1 ,
X X
K+1 =
k=0 k=0

since ej2πkm = 1 for all integers k, m. Now, note that 0 < r < K + 1 implies
that this summation may be summed as a geometric series. That is,
 2πr K+1
K
X −j 2πkr
1 − e−j K+1
e K+1 = 2πr
k=0 1 − e−j K+1
1 − e−j2πr
= 2πr = 0.
1 − e−j K+1

Exercise 5.16 Argue the validity of the expression for b1 in (5.31).

Solution. First, multiply both sides of (5.30) by (z − z0 ) to obtain

(1 − ρ)(z − 1) ∗
(z − z0 )Fñ (z) = (z − z0 ) F (λ[1 − z]).
z − Fx̃∗ (λ[1 − z]) x̃

Now, taking the limit as z → z0 , the denominator of the right-hand side goes to
zero by definition of z0 . Hence we may apply L’Hôpital’s rule to the right-hand
side:
h i
d
(z − z0 ) dz (1 − ρ)(z − 1)Fx̃∗ (λ[1 − z]) + (1 − ρ)(z − 1)Fx̃∗ (λ[1 − z])
d ∗
.
1− dz Fx̃ (λ[1 − z])

If we evaluate this expression at z = z0 , the first term in the numerator goes to

zero, and so we have that

b1 = lim (z − z0 )Fñ (z)

z→z0
" #
(1 − ρ)(z − 1)Fx̃∗ (λ[1 − z])
= d ∗
.
1 − dz Fx̃ (λ[1 − z]) z=z0
The Basic M/G/1 Queueing System 157

Exercise 5.17 Show that the denominator of the right-hand side of

(5.30) for the probability generating of the occupancy distribution has only
one zero for z > 1. [Hint: From Theorem 5.2, we know that Fx̃∗ (λ[1 − z])
is the probability generating function for the number of arrivals during
a service time. Therefore, Fx̃∗ (λ[1 − z]) can be expressed as a power
series in which the coefficients are probabilities and therefore nonnega-
tive. The function and all of its derivatives are therefore nonnegative for
nonnegative z, and so on. Now compare the functions f1 (z) = z and
f2 (z) = Fx̃∗ (λ[1 − z]), noting that the expression Fñ (z) can have poles
neither inside nor on the unit circle (Why?).]
Solution. From Theorem 5.2, we know that Fx̃∗ (λ[1 − z] is the probability
generating function for the number of arrivals during a service time. Therefore,
Fx̃∗ (λ[1 − z]) can be expressed as
∞
Fx̃∗ (λ[1 − z]) = P {ña = n} z n ,
X

n=0

where each of the P {ña = n} is, of course, nonnegative, and whose sum is
equal to one. Now Fñ (z) is also a probability generating function and so is
bounded within the unit circle for z > 0. Hence it can have no poles, and its
denominator no zeros, in that region. If Fñ (z) does have a zero at z = 0, then
z = Fx̃∗ (z) = 0 implies that P {ña } = 0. That is, the probability that the
queue is ever empty is zero, and so stability is never achieved. Since this is not
possible we conclude that the denominator has no zero at z = 0. This means
that P {ña = 0} > 0.
Consider the graphs of the functions f1 (z) = z and f2 (z) = Fx̃∗ (z). The
graph of f1 (z) is a line which starts at 0 and whose slope is 1. On the other
hand, the graph of f2 (z) starts above the origin and slowly increases towards
d
1 as z → 1, with f1 (z) = f2 (z) at z = 1. At z = 1, dz f2 (z) = ρ < 1,
d
and dz f1 (z) = 1. But the slope of f2 (z) is an increasing function, so that
d
as z increases, dz f2 (z) also increases. This means that eventually, f2 (z) will
d
intersect f1 (z). This will happen at the point z = z0 . However, since dz f2 (z)
d
is increasing, while dz f1 (z) is constant, f2 (z) and f1 (z) will intersect only
once for z > 1.
Exercise 5.18 Starting with (5.29) and (5.32), establish the validity of
(5.33) through (5.36).
Solution. If i = K + 1 in (5.29), then pK+1+n ≈ pK+1 r0n . In addition, if
i = K and n = 1, then pK+1 ≈ pK r0 . Combining these results,

pK+1+n ≈ pK+1 r0n ≈ (pK r0 )r0n = pK r0n+1 . (5.5.18.1)

158 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

To show the validity of (5.34), let n = 0 in (5.29) to get

∞
X
c0,K = p0 + pm(K+1) ,
m=1

which implies
∞
X
c0,K − p0 = pm(K+1) . (5.18.2)
m=1

Now,
∞
X ∞
X
pm(K+1) = pK+1 + pm(K+1) − pK+1
m=1 m=1
X∞
= pK+1 + p(m+1)(K+1)
m=1
X∞
= pK+1 + pm(K+1)+K+1
m=1
∞
X
≈ pK r0 + r0 pm(K+1)+K
m=1
∞
" #
X
= r0 pK + pm(K+1)+K
m=1
= r0 cK,K ,

where the last equality follows from (5.29) with n = K. Substituting this
result into (5.18.2) and solving for r0 ,
c0,K − p0
r0 ≈ . (5.34)
cK,K

To show the validity of (5.35), note that

∞ ∞
pm(K+1)+n ≈ r0n
X X
pm(K+1) .
m=1 m=1

Furthermore, by the definition of cn,K ,

X∞
c0,K = p ,
m=0 m(K+1)

or
∞
X
c0,K − p0 = pm(K+1) .
m=1
The Basic M/G/1 Queueing System 159

Substituting this into (5.29), we find

∞
X
cn,K = pn + pm(K+1)+n
m=1
∞
pm(K+1) r0n
X
= ≈ pn +
m=1
= pn + (c0,K − p0 ) r0n .
Thus, solving for pn ,
pn ≈ cn,K − (c0,K − p0 )r0n . (5.35)
Finally, by the definition of probability generating function,
∞ 
n
X 
Fñ (0) = z pn  = p0 . (5.36)
n=0 z=0

Exercise 5.19 Approximate the distribution of the service time for the
previous example by an exponential distribution with an appropriate mean.
Plot the survivor function for the corresponding M/M/1 system at 95% uti-
lization. Compare the result to those shown in Figure 5.2.
Solution. Figure 5.19.1 illustrates the survivor function for the M/M/1 system,
with an exponential service distribution and ρ = .95. Observe that this function
closely approximates that of the truncated geometric distribution in Figure 5.2
of the text, with a = 1, b = 5000, but that the survivor function is significantly
different for the rest of the curves. Thus, the exponential distribution is a good
approximation only in the special case where the truncated distribution closely
resembles the exponential but not otherwise.
Exercise 5.20 Starting with (5.39), demonstrate the validity of (5.47).
Solution. Beginning with (5.39),
1 − Fx̃∗ (s)
Fx̃∗r (s) =
sE[x̃]
1 − Fx̃∗ (s)
=
λ[1 − z] µ1
 ∞
1 − Fx̃∗ (s) X

= zi
ρ i=0
P∞
By (5.43), Fx̃∗ (s) = i=0 z
iP ,
x̃,i so that
∞ ∞
" #
1
Fx̃∗r (s) = i
zi
X X
1− z Px̃,i
ρ i=0 i=0
160 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

1
P{occupancy exceeds n}

.1

.01

.001
0 25 50 75 100

occupancy, n

Figure 5.19.1 Survivor function for Exercise 5.19.

∞ i
" #
1X
zi 1 −
X
= Px̃,n .
ρ i=0 n=0

But, by applying (5.43) to Fx̃∗r (s), we see that

∞
Fx̃∗r (s) = z i Px̃r ,i
X

i=0
∞ i
" #
1 X
i
X
= z 1− Px̃,n .
ρ i=0 n=0

Matching coefficients of zi , we obtain

i
" #
1 X
Px̃r ,i = 1− Px̃,n .
ρ n=0
The Basic M/G/1 Queueing System 161

Exercise 5.21 Evaluate Px̃,i for the special case in which x̃ has the ex-
ponential distribution with mean 1/µ. Starting with (5.50), show that the
ergodic occupancy distribution for the M/M/1 system is given by Pi =
(1 − ρ)ρi , where ρ = λ/µ.
Solution. Since the interarrival times of the Poisson process are exponential,
rate λ, we have by Proposition 4, page 35, that
i
λ µ
 
Px̃,i =
λ+µ λ+µ
i
ρ 1

= .
1+ρ 1+ρ

Therefore
i i n
1 X ρ
X 
Px̃,n =
i=0
1 + ρ n=0 1 + ρ
i+1
   
ρ
1 1 − 1+ρ
= ,

ρ
1+ρ 1 − 1+ρ

" i+1 #
1 ρ

= 1− .
1+ρ 1+ρ

We now show that Pi = (1 − ρ)ρi where ρ = λ/µ. Let T denote the truth set
for this proposition. Then 0 ∈ T since from (5.94),
(1 − ρ)Px̃,0
P0 = = 1 − ρ.
Px̃,0

Now assume that i ∈ T . We wish to show this implies (i + 1) ∈ T . By

hypothesis,
Pn = (1 − ρ)ρn , n < i,
so that
Pi−n = (1 − ρ)ρi−n n > 0.
Therefore,
i n i 
! n+1
ρ
(1 − ρ)ρi−n
X X X
1− Px̃,n Pi−n =
n=1 m=0 n=1
1+ρ
i  n+1
i+1
X 1
= (1 − ρ)ρ
n=1
1+ρ
162 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL
 2  i+1 
1 1
1+ρ − 1+ρ
= (1 − ρ)ρi+1 
 
1
1−

1+ρ
" i+1 #
(1 − ρ)ρi 1

= 1− .
1+ρ 1+ρ

Thus, using (5.50) and the hypothesis,

( ),
ρi (1 − ρ)ρi 1 1

Pi = (1 − ρ) i+1
+ 1−
(1 + ρ) 1+ρ (1 + ρ)i 1+ρ
= (1 − ρ)ρi ,

and the proof is complete.

Exercise 5.22 Evaluate Px̃,i for the special case in which P {x̃ = 1} =
1. Use (5.50) to calculate the occupancy distribution. Compare the comple-
mentary occupancy distribution (P {N > i}) for this system to that of the
M/M/1 system with µ = 1.
Solution. If x̃ = 1 with probability 1, then it should be clear that the distribu-
tion describing the number of Poisson arrivals is simply a Poisson distribution
with t = 1 (Definition 1, part iii). i.e.,
e−λ λi
Px̃,i = .
i!
To calculate the complementary occupancy probabilities, a computer program
was written using the recursive formula (5.50) using different values of λ. For
λ = 0.1, the distributions were very similar, with the probabilities of the deter-
ministic and exponenitial summing to 1 after n = 6 and n = 9, respectively.
(Accuracy was to within 10−8 .) However, as n increased, the two clearly di-
verged, with the deterministic system summing to 1 faster than that of the
exponential. The graphs of the complementary probabilities vs. those of the
M/M/1 system for λ = 0.5 and λ = 0.9 is shown in Figure 5.22.1.

Exercise 5.23 Evaluate Px̃,i for the special case in which P {x̃ = 21 } =
P {x̃ = 23 } = 12 . Use (5.50) to calculate the occupancy distribution. Com-
pare the complementary occupancy distribution (P {N > i}) for this system
to that of the M/M/1 system with µ = 1.
Solution. As in Exercise 5.5.22, if x̃ is deterministic, then Px̃,i is a weighted
Poisson distribution, with
i i
  e− 12 λ   e− 23 λ
 
1 3
1 2λ 1 2λ
Px̃,i = + .
2 i! 2 i!
The Basic M/G/1 Queueing System 163

.1

λ=.9
.01
P{occupancy exceeds n}

.001

.0001

0 λ=.5
0

0
Exponential

0 D eterm inistic

0
0 2 4 6 8 10 12 14 16

Figure 5.22.1 Survivor function for Exercise 5.22.

The computer program used in Exercise 22 was modified and the new occu-
pancy probabilites calculated. As in Exercise 22, the graphs of both distribu-
tions were very similar for λ = 0.1, with the probabilities of both summing to
unity within n = 9. As λ increased, however, the differences between the two
systems became apparent, with the deterministic complementary distribution
approaching zero before that of the exponential system. The graphs compar-
ing the resultant complementary probabilites to those of the M/M/1 system for
λ = 0.5 and λ = 0.9 is shown in Figure 5.23. 1.

Exercise 5.24 Beginning with (5.59), suppose Fx̃e (x) = Fx̃ (x). Show
that (5.59) reduces to the standard Pollaczek-Khintchine transform equation
for the queue length distribution in an ordinary M/G/1 system.
Solution. Equation 5.59 is as follows:
 
Fq̃ (z) [z − Fã (z)] = π0 z Fb̃ (z) − Fã (z) . (5.59)
164 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

.1

λ=.9
.01
P{occupancy exceeds n}

.001

.0001

0 λ=.5
0

0
Exponential

0 D eterm inistic

0
0 2 4 6 8 10 12 14 16

Figure 5.23.1 Survivor function for Exercise 5.23.

Since Fx̃e (x) = Fx̃ (x), we have Fb̃ (z) = Fã (z). Upon substitution of this
result in (5.59), we have

Fq̃ (z) [z − Fã (z)] = π0 (z − 1)Fã (z).

Putting Fã (z) = Fx̃∗ (λ[1 − z]) leads to

Fq̃ (z) [z − Fx̃∗ (λ[1 − z])] = π0 (z − 1)Fx̃∗ (λ[1 − z]).

Solving the previous equation, we find

(1 − z)π0 Fx̃∗ (λ[1 − z])

Fq̃ (z) = ,
Fx̃∗ (λ[1 − z]) − z

which is the same as (5.5) with π = 1 − P {q̃ = 0}.

The Basic M/G/1 Queueing System 165

Exercise 5.25 Beginning with (5.59), use the fact that limz→1 Fq̃ (z) =
1 to show that
1 − Fã′ (1)
π0 = . (5.5)
1 − Fã′ (1) + Fb̃′ (1)

Solution. Equation 5.59 is as follows:

 
Fq̃ (z) [z − Fã (z)] = π0 z Fb̃ (z) − Fã (z) . (5.59)
First take the derivative of both sides of (5.59) with respect to z to find
d d d d
   
Fq̃ (z) [z − Fã (z)]+Fq̃ (z) 1 − Fã (z) = π0 z Fb̃ (z) + Fb̃ (z) − Fã (z) .
dz dz dz dz
Now, take the limit as z → 1 and recognize that Fq̃ (1) = 1, Fã (1) = 1, and
Fb̃ (1) = 1. We then have
d d d
   
1− Fã (1) = π0 Fb̃ (1) + 1 − Fã (1)
dz dz dz
from which the result follows directly.
Exercise 5.26 Argue rigorously that in order for the M/G/1 queue to be
stable, we must have a0 > 0.
Solution. Since a0 represents the probability that there are no arrivals during
an ordinary service time, there only two choices, a0 = 0 or a0 > 0. The latter
choice indicates that there are at least one arrival during an ordinary service
time and hence the expected number of arrivals during a service time is at least
unity. Suppose the queueing system ever enters level one. Then the dynamics
of returning to level zero are identical to that of an ordinary M/G/1 queueing
system. Therefore, we may consider the ordinary M/G/1 queueing system. In
that case the expected number of arrivals during a service time is the traffic
intensity. Thus, if a0 = 0, the traffic intensity is at least unity and the system is
unstable. Therefore, we know that for stability of the present queueing system,
we ned a0 > 0.
Exercise 5.27 Verify (5.62).
Solution. The restatement of (5.62) is as follows:
z0 D = π0 N
where z0 = [ π1 π2 ... πm+1 ], N = [ b0 − 1 b1 ... bm ] and
−a0 1 − a1 −a2 ··· −am
 
..
 0 −a0 1 − a1 . −am−1 
 
 
D= 0
 .. .

0 −a0 . −am−2 
 .. ..
 
.. .. .. 
 . . . . . 
0 0 0 ··· −a0
166 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

This is derived directly from (5.61), which is as follows:

[ π0 π1 ... πm ] = π0 [ b0 b1 . . . bm ] +
a0 a1 a2 · · · am
 
 0 a0 a1 · · · am−1 
 
[ π1 π2 . . . πm+1 ]  0
 0 a0 · · · am−2  .
 . . . . .
. . . . .

 . . . . . 
0 0 0 ··· a0
We may write the left side of the previous equation as
 
0 Im
π0 [ 1 0 ... 0 ] + [ π1 π2 ... πm+1 ] ,
0 0

where [ 1 0 . . . 0 ] is an m-vector and Im is an m-square identity matrix.

We thus have for (5.59),
 
0 Im
π0 [ 1 0 . . . 0 ] + z0 = π 0 [ b0 b1 ... bm ] +
0 0
a0 a1 a2 ··· am
 

 0 a0 a1 · · · am−1 

z0 
 0 0 a0 · · · am−2 
.
 .. .. .. .. .. 
 . . . . . 
0 0 0 ··· a0

The result follows by noting that N = [ b0 b1 ... bm ] − [ 1 0 . . . 0]

and
a0 a1 a2 ··· am
 

   0 a0
 a1 ··· am−1 
0 Im ···

D= − 0 0 a0 am−2 
.

0 0  . .. .. .. .. 
 .. . . . . 
0 0 0 ··· a0

Exercise 5.28 For i = 1, 2, . . ., show that

∞
(i)
πi+j z j .
X
Fq̃ (z) =
j=0
P∞ j,
Solution. Since Fq̃ (z) = j=0 πj z we see immediately that Fq̃ (z) − π0 =
P∞ j
j=1 πj z . Therefore,

∞ ∞
1
πj z j−1 = πj+1 z j .
X X
[Fq̃ (z) − π0 ] =
z j=1 j=0
The Basic M/G/1 Queueing System 167

Thus, the result holds for i = 1. Assume the result hold for i − 1, then
∞
(i−1)
πi−1+j z j .
X
Fq̃ (z) =
j=0

Now,
(i) 1 h (i−1) i
Fq̃ (z) = Fq̃ (z) − πi−1 ,
z
we have
   
∞ ∞
(i) 1 X 1 X
Fq̃ (z) =  πi−1+j z j − πi−1  =  πi−1+j z j  .
z j=0 z j=1

Hence
∞
(i)
πi+j z j ,
X
Fq̃ (z) =
j=0

and the proof is complete.

Exercise 5.29 Define
(1) 1 (i+1) 1 h (i) i
Fq̃ (z) = [Fq̃ (z) − π0 ] and Fq̃ (z) = Fq̃ (z) − πi , i ≥ 1.
z z
(1)
Starting with (5.68), substitute a function of Fq̃ (z) for Fq̃ (z), then a func-
(2) (1)
tion of Fq̃ (z) for Fq̃ (z), and continue step by step until a function of
(C) (C−1)
Fq̃ (z) is substituted for Fq̃ (z). Show that at each step, one element
of
C−1
πj z j Fã (z)
X

j=0

is eliminated, resulting in (5.69).

Solution. First we solve
(1) 1
Fq̃ (z) = [Fq̃ (z) − π0 ]
z
for Fq̃ (z) to find
(1)
Fq̃ (z) = zFq̃ (z) + π0
and
(i+1) 1 h (i) i
Fq̃ (z) = Fq̃ (z) − πi , i ≥ 1
z
(i)
for Fq̃ (z) to find
(i) (i+1)
Fq̃ (z) = zFq̃ (z) + πi .
168 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

Now, (5.68) is
h i C−1 h i
Fq̃ (z) z C − Fã (z) = πj z C Fb̃,j (z) − z j Fã (z) .
X
(5.68)
j=0

(1)
To begin, we substitute Fq̃ (z) = zFq̃ (z) + π0 into (5.68) to obtain

h i C−1 C−1
(1) C C
z j Fã (z)+z C π0 +π0 Fã (z).
X X
zFq̃ (z) z − Fã (z) = πj z Fb̃,j (z)−
j=0 j=0

Simplifying, we find
h i C−1 C−1
zFq̃ (z)(1) z C − Fã (z) = πj z C Fb̃,j (z) − z j Fã (z) + z C π0 .
X X

j=0 j=1
(5.68.1)
(2) (1)
Similarly, we substitute zFq̃ (z) + πi for Fq̃ (z) in (5.68.1) to find

h i C−1 C−1
z 2 Fq̃ (z)(2) z C − Fã (z) πj z C Fb̃,j (z) − z j Fã (z)
X X
=
j=0 j=1
C C+1
−z π0 − z π1 + π1 zFã (z),
and simplifying leads to
h i C−1 C−1
z 2 Fq̃ (z)(1) z C − Fã (z) = πj z C Fb̃,j (z)− z j Fã (z)−z C π0 −z C+1 π1 .
X X

j=0 j=2
(5.68.2)
Continuing in this way leads to
h i C−1 C−1
z C Fq̃ (z)(C) z C − Fã (z) πj z C Fb̃,j (z) − z j Fã (z)
X X
=
j=0 j=C−1
C−1
z C+j πj + πC−1 z C−1 Fã (z),
X
+
j=0

and simplifying leads to

h i C−1 C−1
z C Fq̃ (z)(C) z C − Fã (z) = πj z C Fb̃,j (z) − z C + jπj ,
X X

j=0 j=0

or
h i C−1 h i
Fq̃ (z)(C) z C − Fã (z) = πj Fb̃,j (z) − z j .
X
(5.69)
j=0
The Basic M/G/1 Queueing System 169

Exercise 5.30 Suppose Fã (z) and Fb̃ (z) are each polynomials of degree
m as discussed in the first part of this section. Define w(z) = 1. Find ν, D,
N , A, and E using (5.72), (5.73), and (5.78). Compare the results to those
presented in (5.62) and (5.66).
Solution. We are considering the case where C = 1. Recall that Fã (z) =
u(z)/w(z). With w(z) = 1, u(z) = Fã (z), Similarly, v(z) = Fb̃ (z). Thus,
d(z) = z − Fã (z) and n0 (z) = n(z) = Fb̃ (z) − 1 for this special case. Now,
m ≥ 1 because otherwise, there would never be any arrivals. Recall that νd and
νn are the degrees of d(z) and n(z), respectively. Since m ≥ 1, νd = νn = m.
Thus,
ν = max {nud , νn + 1} = max {m, m + 1} = m + 1.
Using (5.73), we then have

d1 = 1−a1 , di = −ai , i 6= 1, 0 ≤ i ≤ m and n0 = b0 −1, ni = bi , 1 ≤ i ≤ m.

We then have from (5.76)

−a0 1 − a1 −a2 ··· −am
 
..
 0 −a0 1 − a1 . −am−1 
 
 
D= 0
 .. ,

0 −a0 . −am−2 
 .. ..
 
.. .. 
 . . . . ··· 
0 0 0 ··· −a0
and N = [ b0 − 1 b1 b2 · · · bm ]. From (5.78), where E and A are de-
fined as follows where E = diag (1, 1, . . . , 1, d0 ), and
0 0 ··· ··· dν−1
 
 .. ..
. .

1 0 dν−2 
 
A = 0
 .. .. ,

 1 . . −dν−3 

. .. .. .. .. 
 .. . . . . 
0 0 ··· 1 d1
we find E = diag (1, 1, . . . , 1, −a0 ) and
0 0 ··· ··· −am
 
.. ..
1 0 . . −am−1 
 
 
A = 0
 .. .. .

1 . . −am−2
 .. ..
 
.. .. .. 
. . . . . 
0 0 ··· 1 (1 − a1 )
These are the same results as presented in (5.62) and (5.63).
170 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

Table 5.1. Occupancy values as a function of the number of units served during a service
period for the system analyzed in Example 5.8.

Occupancy C=1 C=2 C=4 C=8

- 1.00000 1.00000 1.00000 1.00000

0 0.9000 0.9493 0.9861 0.9985
1 0.7633 0.8507 0.9453 0.9920
2 0.6350 0.7347 0.8772 0.9762
3 0.5261 0.6233 0.7914 0.9481
4 0.4356 0.5254 0.6997 0.9071
5 0.3606 0.4419 0.6107 0.8547
6 0.2985 0.3715 0.5292 0.7942
7 0.2472 0.3123 0.4569 0.7292

Exercise 5.31 Table 5.5 gives numerical values for the survivor function
of the occupancy distributions shown in Figure 5.5. From this table, deter-
mine the probability masses for the first few elements of the distributions
and then compute the mean number of units served during a service time
for C = 1, 2, 4, and 8. Analyze the results of your calculations.
Solution. The main idea here is that if there are at least C units present in the
system, then the number of units served will be C. Otherwise, there will be
the same as the number of units present, or equivalently, the servers will serve
all units present at the beginning of the service interval. Hence, the expected
number of units served will be
C−1
X
iP {q̃ = i} + CP {q̃ ≥ C} .
i=1
To obtain the probability masses, we simply subtract successive values of the
survivor functions. For example, for the case C = 2, we have P {q̃ = 0} =
1 − 0.9493 = 0.0507, P {q̃ = 1} = 0.9493 − 0.8507 = 0.0986, P {q̃ ≥ 2} =
P {q̃ > 1} = 0.8507. Hence the expected number of services during a service
interval is 1 × 0.0986 + 2 × 0.8507 = 1.8. From this, we see that the system
utilization is 0.9. If we repeat this for C = 4, we will find that the expected
number of services that occur during a service time is 3.6 so that again the
system utilization is 0.9.
Exercise 5.32 Use a busy period argument to establish the validity of
(5.90). [Hint: Consider the M/G/1 system under the nonpreemptive LCFS
service discipline.]
The Basic M/G/1 Queueing System 171

Solution. Let B be the event that an arbitrary customer finds the system busy
upon arrival, and let I represent its complement. Then

E[w̃] = E[w̃|B]P {B} + E[w̃|I]P {I}

= ρE[w̃|B] + E[w̃|I](1 − ρ).

But E[w̃|I] = 0 since the customer enters service immediately if the system is
idle upon arrival. Therefore,

E[w̃] = ρE[w̃|B].

Now, the total amount of waiting time for the arbitrary customer in a LCFS dis-
cipline will be the length of the busy period started by the customer in service
and generated by the newly arrived customers. Hence,
E[x̃e ]
E[w̃|B] = .
1 − ρ]
Multiplying this expression by ρ, the probability that the system is busy upon
arrival, we find that
ρ
E[w̃] = ρE[w̃|B] = E[x̃e ],
1−ρ
which is the desired result.
Exercise 5.33 Show that the Laplace-Stieltjes transform for the distri-
bution of the residual life for the renewal process having renewal intervals
of length z̃ is given by
1 − Fz̃∗ (s)
Fz̃∗r (s) = . (5.98)
s E[z̃]
Solution. By Definition 2.8,
Z ∞
Fz̃∗r (s) = e−sz dFz̃r (z)
Z0∞
1 − Fz̃ (z)
 
−sz
= e dz
0 E[z̃]
1 ∞
Z ∞ Z
= e−sz dFz̃ (y)dz
E[z̃] Z0 Z x
1 ∞ y 
−sz
= e dz dFz̃ (y)
E[z̃] Z0 0
1 ∞ 1
1 − e−sy dFz̃ (y)

=
E[z̃] 0 s
1
= [1 − Fz̃ (s)] .
sE[z̃]
172 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

Exercise 5.34 For an arbitrary nonnegative random variable, x̃, show

that
E x̃n+1
 
E [x̃nr ] = . (5.102)
(n + 1)E[x̃]
Solution. Recall that
[1 − Fx̃ (z)]
fx̃r (z) = .
E[x̃]
We may then write
Z∞
E[x̃r ] = xn fx̃r (x)dx,
0
so that
Z∞
1
E[x̃r ] = xn [1 − Fx̃ (z)] dx
E[x̃]
0
Z∞ Z∞
 
1
= xn  1 − fx̃ (z)dz  dx.
E[x̃]
0 x

Changing the order of integration,

Z∞ Zz
 
1  xn dx fx̃ (z)dz
E[x̃r ] =
E[x̃]
0 0
Z∞ z
1 xn+1 
=  fx̃ (z)dz

E[x̃] n + 1 0
0
E[x̃n+1 ]
= .
(n + 1)E[x̃]

Exercise 5.35 For the M/G/1 system, suppose that x̃ = x̃1 + x̃2 where
x̃1 and x̃2 are independent, exponentially distributed random variables with
parameters µ1 and µ2 , respectively. Show that Cx̃2 ≤ 1 for all µ1 , µ2 such
that E[x̃] = 1.
Solution. First, by the fact that E[x̃] = 1,

1 = E 2 [x̃] = E 2 [x̃1 + x̃2 ]

The Basic M/G/1 Queueing System 173

= E 2 [x̃1 ] + 2E[x̃1 ]E[x̃2 ] + E 2 [x̃2 ]

1 2 1
= 2 + + 2,
µ1 µ1 µ2 µ2

which implies that 1/µ21 + 1/µ22 ≤ 1 since 2/µ1 µ2 ≥ 0. Therefore,

Var(x̃)
Cx̃2 =
E 2 [x̃]
Var(x̃1 + x̃2 )
=
1
= Var(x̃1 ) + Var(x̃2 )
1 1
= + ≤ 1.
µ21 µ22

Exercise 5.36 Compute the expected waiting time for the M/G/1 system
with unit mean deterministic service times and for the M/G/1 system with
service times drawn from the unit mean Erlang-2 distribution. Plot on the
same graph E[w̃] as a function of ρ for these two distributions and for the
M/M/1 queueing system with µ = 1 on the same graph. Compare the
results.
Solution. To compute the expected waiting time, recall equation (5.101):
" #
ρE[z̃] 1 + Cz̃2
E[w̃] = ,
1−ρ 2

where
Var(z̃)
Cz̃2 = .
E 2 [z̃]
For all three systems, note that E[z̃] = 1. Thus, we need only find the variance
(which will be equal to Cz̃2 in this case) of each to compute its expected waiting
time. For the deterministic system, the variance is zero, so that

ρ
ρ 1+0
 
E[w̃d ] = = 2 .
1−ρ 2 1−ρ

For the Erlang-2 system,

Var(ze ) = Var(z1 ) + Var(z2 )

1 1
= 2
+ 2
4µ 4µ
1
= ,
2
174 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

and the mean waiting time for this system is

1 3
" #
ρ 1+ 2 4ρ
E[w̃e ] = = .
1−ρ 2 1−ρ

For the M/M/1 system, Var(z̃m ) = 1/µ2 = 1, so that

ρ 1+1 ρ
 
E[w̃m ] = = .
1−ρ 2 1−ρ

From Figure 5.1, we see that as the coefficient of variance increases, the perfor-
mance of the system deteriorates. i.e., the coefficient of variance and the mean
waiting time are directly proportional. In addition, we see from the figure that
the mean waiting time is also directly proportional to ρ. This should be clear
since if ρ = µλ increases, then the arrival rate also increases, putting a greater
load on the system.

Exercise 5.37 For the M/G/1 system, suppose that x̃ is drawn from the
distribution Fx̃1 (x) with probability p and from Fx̃2 (x) otherwise, where
x̃1 and x̃2 are independent, exponentially distributed random variables with
parameters µ1 and µ2 , respectively. Let E[x̃] = 1. Show that Cx̃2 ≥ 1 for
all p ∈ [0, 1].
Solution. By the definition of Cx̃2 ,

Var(x̃) E[x̃2 ] − E 2 [x̃]

Cx̃2 = = .
E 2 [x̃] E 2 [x̃]

That is, proving Cx̃2 ≥ 1 is equivalent to showing E[x̃2 ] ≥ 2. Let Di denote

the event that distribution i is selected, i = 1, 2. Then

P {x̃ ≤ x} = P {x̃ ≤ x|D1 } P {D1 } + P {x̃ ≤ x|D1 } P {D1 }

= pFx̃1 (x) + (1 − p)Fx̃2 (x),

so that
dFx̃ (x) = pdFx̃1 (x) + (1 − p)dFx̃2 (x).
Therefore,
E[x̃n ] = pE[x̃n1 ] + (1 − p)E[x̃n2 ].
For the exponential distribution with rate µ, E[x̃2 ] = 2/µ2 , and so
2 2 p 1−p
 
E[x̃2 ] = p 2 + (1 − ρ) 2 = 2 2 + .
µ1 µ2 µ1 µ22
The Basic M/G/1 Queueing System 175

100

D eterm inistic
Erlang-2
80 Exponential
E[waiting time]

60

40

20

0
0.80 0.83 0.86 0.89 0.92 0.95 0.98
rho

Figure 5.36.1 Survivor function for Exercise 5.36.

Therefore, showing E[x̃2 ] ≥ 2 is equivalent to showing

p (1 − p)
2 + ≥ 1.
µ1 µ22

i.e.,
p 1−p
2 + − 1 ≥ 0.
µ1 µ22
Now,
p 1−p
E[x̃] = + = 1,
µ1 µ2
so that
p2 2p(1 − p) (1 − p)2
E 2 [x̃] = + + = 1.
µ21 µ1 µ2 µ22
176 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

Thus,
p 1−p p 1 − p p2 2p(1 − p) (1 − p)2
+ −1 = + − 2− −
µ21 µ22 µ12 µ22 µ1 µ1 µ2 µ22
p(1 − p) p(1 − p) 2p(1 − p)
= + −
µ21  µ22 µ1 µ2
1 2 1
= p(1 − p) 2 − +
µ µ1 µ2 µ22
 1 2
1 1
= p(1 − p) − ≥ 0,
µ1 µ2

and the proof is complete.

Exercise 5.38 With x̃ and p defined as in Exercise 5.37, let p = 12 . Find

µ1 and µ2 such that Cx̃2 = 2. Would it be possible to determine p, µ1 , and
µ2 uniquely for a given value of Cx̃2 ? Explain.
Solution. First, we will compute the general case for any p, and then we will
substitute for the specific value of p = 21 . Now, 1 = E[x̃], and Cx̃2 = E[x̃2 ]−1,
so that we have the following two equations
p (1 − p)
+ = 1 (5.38.1)
µ1 µ2

2p 2(1 − p)
+ = Cx̃2 + 1 (5.38.2)
µ21 µ22
Rearranging (5.38.2),
1 (1 − p)/p Cx̃2 + 1
2 + = . (5.38.3)
µ1 µ22 2p
1
Solving equation (5.38.1) for µ1 , and squaring both sides
" #
1 1 2(1 − p) (1 − p)2
= 1 − + .
µ21 p2 µ2 µ22
Substitute this expression into (5.38.3) to yield
1 − 2p Cx̃2 + 1


1 2
− + = 0.
µ22 µ2 1−p
1
And solving this equation for µ2 , we obtain
s
1 p 1
 
Cx̃2 − 1 .


= 1± (5.38.4)
µ2 1−p 2
The Basic M/G/1 Queueing System 177

Substitute this result into (1) to yield

s
1 1−p

Cx̃2 − 1 .


= 1± (5.38.5)
µ1 2p
This shows that one can find µ1 and µ2 given p and Cx̃2 if Cx̃2 ≥ 1. If we take
p
1/µ1 = 1 + 1/2, then r
1 1
=1− .
µ2 2
From (5.38.4) and (5.38.5), we see that it is necessary to specify p in addition to
Cx̃2 to obtain unique values of µ1 and µ2 . This should be clear from equations
(5.38.1) and (5.38.2), which is a system of two equations in three unknowns.
Exercise 5.39 (Ross[1989]) Consider an ordinary renewal process with
renewal interval z̃. Choose a real number c arbitrarily. Now suppose the
renewal process is observed at a random point in time, t0 . If the age of the
observed interval is less than c, define the system to be in an x-period, else
define the system to be in a y-period. Thus, the expected cycle length is
E[z̃], and the expected length of the it x-period is E[min {c, z̃}]. Show that
Z c
E[min {c, z̃}] = [1 − Fz̃ (z)]dz
0

so that
d 1 − Fz̃ (z)
Fz̃ (z) = ,
dz a E[z̃]
as was shown in the previous subsection.
Solution.
Exercise 5.40 Formalize the informal discussion of the previous para-
graph.
Solution. Let Y (t) denote the total length of time the server has spent serving
up to time t, so that Y (t)/t represents the proportion of time the server has
spent serving up to time t. Now let N (t) denote the total number of busy
periods completed up to time t. So long as there has been at least one busy
period up to this time, we may write
Y (t) Y (t) N (t) Y (t) N (t)
= = .
t t N (t) N (t) t
For a fixed t, Y (t) is a random variable. Thus, the long-term average propor-
tion of time the server is busy is
Y (t) Y (t) N (t)
lim = lim .
t→∞ t t→∞ t N (t)
178 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

It remains to show
Y (t) N (t) Y (t) N (t)
lim = lim lim .
t→∞ t N (t) t→∞ N (t) t→∞ t
That is, that these limits exist separately.
We consider the limit of t/N (t), as t → ∞. Let cn = xn + yn denote the
length of the n-th cycle. Then
PN (t) PN (t)+1
t n=0 cn t n=0 cn
lim = lim ≤ ≤ ,
t→∞ N (t) t→∞ N (t) N (t) N (t)
so that
t
lim = E[c̃] = E[x̃] + E[ỹ].
t→∞ N (t)
Therefore,
N (t) 1
lim = .
t→∞ t E[x̃] + E[ỹ]
Since the limits exist separately,

Y (t) Y (t) N (t)

lim = lim lim
t→∞ t t→∞ N (t) t→∞ t
E[ỹ]
= .
E[x̃] + E[ỹ]

5-1 Consider a communication system in which messages are transmitted over

a communication line having a capacity of C octets/sec. Suppose the mes-
sages have length m̃ (in octets), and the lengths are drawn from a geomet-
ric distribution having a mean of E[m̃] octets, but truncated at a and b
characters on the lower and upper ends of the distribution, respectively.
That is, message lengths are drawn from a distribution characterized as
follows:

P {m̃ = m} = kθ(1 − θ)m−1 for a ≤ m ≤ b,

where m̃ is the number of characters in a message and k is a normalizing

constant.

(a) Given that

P {m̃ = m} = kθ(1 − θ)m−1 for a ≤ m ≤ b,

The Basic M/G/1 Queueing System 179

show that h i−1

k = (1 − θ)a−1 − (1 − θ)b ,

θz 1 − [(1 − θ)z](b−[a−1])
E[z m̃ ] = z (a−1) ,
1 − (1 − θ)z 1 − (1 − θ)(b−[a−1])
and
1 (b − [a − 1])(1 − θ)(b−[a−1])
E[m̃] = a − 1 + − .
θ 1 − (1 − θ)(b−[a−1])

(b) Rearrange the expression for E[m̃] given above by solving for θ −1 to
obtain an equation of the form
1
= f (E[m̃], a, b, θ),
θ
and use this expression to obtain a recursive expression for θ of the
form
1
= f (E[m̃], a, b, θi ).
θi+1
(c) Write a simple program to implement the recursive relationship de-
fined in part (b) to solve for θ in the special case of a = 10, b = 80,
and E[m̃] = 30. Use θ0 = E −1 [m̃] as the starting value for the
recursion.
(d) Argue that Fx̃∗ (s) = Fm̃
∗ (s/C), where C is the transmission capacity

in octets/sec.
(e) Using the computer program given in the Appendix, obtain the com-
plementary occupancy distribution for the transmission system under
its actual message length distribution at a traffic utilization of 95%,
assuming a transmission capacity of 30 characters/sec.
(f) Compare this complementary distribution to one obtained under the
assumption that the message lengths are drawn from an ordinary ge-
ometric distribution. Comment on the suitability of making the geo-
metric assumption.

Solution:

(a) From the laws of total probability,

b
X
P {m̃ = m} = 1.
m=a
180 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

And from (1.1),

b b
θ(1 − θ)m−1
X X
P {m̃ = m} = k
m=a m=a
(1 − θ)a−1 − (1 − θ)b
= kθ
h 1 − (1 − θ) i
= k (1 − θ)a−1 − (1 − θ)b = 1.

Therefore, h i−1
k = (1 − θ)a−1 − (1 − θ)b .
To find E[m̃], observe that
b
E[z m̃ ] = z m P {m̃ = m}
X

m=a
b
z m kθ(1 − θ)m−1
X
=
m=a
b
[z(1 − θ)]m−1
X
= kθz
m=a
[z(1 − θ)]a−1 − [z(1 − θ)]b
= kθz
h 1 − z(1 − θ) i
θz (1 − θ)a−1 − z (b−(a−1) (1 − θ)b
= z a−1
1 − z(1 − θ)
1
· a−1
(1 − θ) − (1 − θ)b
a−1 θz 1 − [(1 − θ)z]b−(a−1)
= z .
1 − z(1 − θ) 1 − (1 − θ)b−(a−1)
To compute E[m̃], first note the following probabilities:
1, x < a,


b


P {m̃ = m} , a ≤ x < b, ,
P
P {m̃ > x} =
 m=x+1


 0, x≥b
 1,h x < a,
i
x b
k (1 − θ) − (1 − θ) , a ≤ x < b, ,
=

0, x≥b

so that
∞
X
E[m̃] = P {m̃ > x}
x=0
The Basic M/G/1 Queueing System 181
" b−1 #
(1 − θ)x − (b − a)(1 − θ)b + a
X
= k
"x=a #
(1 − θ)a − (1 − θ)b
= k − (b − a)(1 − θ)b + a
1 − (1 − θ)
kh i
= (1 − θ)a − (1 − θ)b − θ(b − a)(1 − θ)b + a
θ" #
1 (1 − θ)a − (1 − θ)b − θ(b − a)(1 − θ)b
= +a
θ (1 − θ)a−1 − (1 − θ)b
" #
1 (1 − θ) − (1 − θ)b−(a−1) − θ(b − a)(1 − θ)b−(a−1)
=
θ 1 − (1 − θ)b−(a−1)
+a
1 1 + (b − a)(1 − θ)b−(a−1)
= − +a
θ 1 − (1 − θ)b−(a−1)
1 1 − (1 − θ)b−(a−1) + [b − (a − 1)](1 − θ)b−(a−1)
= −
θ 1 − (1 − θ)b−(a−1)
+a
1 [b − (a − 1)](1 − θ)b−(a−1)
= (a − 1) + − .
θ 1 − (1 − θ)b−(a−1)

We may also compute the mean of m̃ by considering the following.

Define m̃1 to be geometric random variable with parameter θ. Then
E[m̃] = a − 1 + E[m̃1 |m̃1 ≤ c], where the latter expression is equal
to the mean of geometric random variable truncated at the point c =
b − (a − 1). Now,
E[m̃1 ] = E[m̃1 |m̃1 ≤ c]P {m̃ ≤ c}
+E[m̃1 |m̃1 ≤ c]P {m̃ > c} . (5.1.1)

But
∞
θ(1 − θ)i−1
X
P {m̃ > c} =
i=c
θ(1 − θ)c
=
1 − (1 − θ)
= (1 − θ)c ,

and
1
E[m̃1 |m̃ > c] = c + .
θ
Thus,
P {m̃ ≤ c} = 1 − (1 − θ)c ,
182 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

and from (S5.1.1), we then have

 
1 1
θ − c+ θ (1 − θ)c
E[m̃1 |m̃1 ≤ c] =
1 − (1 − θ)c
1 c(1 − θ)c
= − .
θ 1 − (1 − θ)c
Therefore,
1 c(1 − θ)c
E[m̃] = a − 1 + −
θ 1 − (1 − θ)c
1 [b − (a − 1)](1 − θ)b−(a−1)
= a−1+ − .
θ 1 − (1 − θ)b−(a−1)

(b) From part (a),

1 [b − (a − 1)](1 − θ)b−(a−1)
= E[m̃] = (a − 1) + ,
θ 1 − (1 − θ)b−(a−1)
or
1 [b − (a − 1)](1 − θi )b−(a−1)
= E[m̃] − (a − 1) + .
θi+1 1 − (1 − θi )b−(a−1)

(c) The computer program resulted in θ = 0.039531.

n o
m̃ m̃
(d) Since x̃ = C, P {x̃ ≤ x} = P C ≤ x = P {m̃ ≤ Cx}, so that
Z ∞ Z ∞
Fx̃∗ (x) = e −sx
dFx̃ (x) = e−sx dFm̃ (Cx).
0 0

Let y = Cx; then x = y/C, and

Z ∞ y
Fx̃∗ (s) = e−s C dFm̃ (y)
0
∗ s
= Fm̃ ( ).
C
(e) See the graph in Figure 1.
(f) The comparison is roughly that shown in Figure 5.1 between the
(10, 80) curve and the (1, 5000) curve. For n small, the two distri-
butions can be said to approximate one another. However, as n in-
creases, the truncated geometric diverges from that of the ordinary
geometric and so the geometric assumption is not valid in this case.

medskip
The Basic M/G/1 Queueing System 183

1
P{Occupancy exceeds n}

.1
geometric
mean 18.70

.01

mean 12.90
(10,80)

.001
0 25 50 75 100

Occupancy, n

Figure P1.1 Survivor function for Supplementary Problem 1.

5-2 Using the properties of the probability generating function, determine a

formula for E[ñ2 ], the second moment of the occupancy distribution for
the ordinary M/G/1 system, in terms of the first three moments of Fx̃ (x),
the service time distribution. Verify the formula for E[ñ] along the way.
[Hint: The algebra will be greatly simplified if (??) is first rewritten as

Fñ (z) = α(z)/β(z),

where
α(z) = (1 − ρ)Fx̃∗ (λ[1 − z]),

β(z) = 1 − ρFx̃∗r (λ[1 − z]),

and Fx̃r (x) is the distribution for the forward recurrence time of the ser-
vice time. Then, in order to find

d2
lim Fñ (z),
z→1 dz 2
184 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

first find the limits as z → 1 of α(z), β(z), dα(z)/dz, dβ(z)/dz,

d2 α(z)/dz 2 , d2 β(z)/dz 2 , and then substitute these limits into the formula
for the second derivative of the ratio.]
Solution: First rewrite (5.8) as
α(z)
Fñ (z) = ,
β(z)
where
α(z) = (1 − ρ)Fx̃∗ (λ[1 − z]),
and
β(z) = 1 − ρFx̃∗r (λ[1 − z]),
where Fx̃r (x) is the distribution for the forward recurrence time of the
service time. Then,
′ ′
d α (z) α(z)β (z)
Fñ (z) = − ,
dz β(z) β 2 (z)
and
′′ ′ ′
d2 α (z) α (z)β (z)
Fñ (z) = −
dz 2 β(z) β 2 (z) !
′ ′ ′′ ′
α (z)β (z) α(z)β (z) 2α(z)[β (z)]2
− + −
β 2 (z) β 2 (z) β 3 (z)
′′ ′ ′ ′′ ′
α (z) 2α (z)β (z) α(z)β (z) 2α(z)[β (z)]2
= − − + .
β(z) β 2 (z) β 2 (z) β 3 (z)
Now, α(1) = 1 − ρ, and β(1) = 1 − ρ, so that
′ d ∗
α (z) = (1 − ρ)(−λ) F (x)
ds x̃
implies
′ d ∗
α (1) = (1 − ρ)(−λ) F (0) = ρ(1 − ρ),
ds x̃
and
′ d ∗
β (z) = −ρ(−λ) F (x)
ds x̃r
implies
′
β (1) = −ρ(−λ)(−E[x̃r ]) = −λρE[x̃r ].
Therefore
d
E[ñ] = Fñ (1)
ds
The Basic M/G/1 Queueing System 185
′ ′
α (1) α(1)β (1)
= −
β(1) β 2 (1)
(1 − ρ)ρ (1 − ρ)λρE[x̃r ]
= +
1−ρ (1 − ρ)2
λρ E[x̃ ] 2
= ρ+
" 1 − ρ 2E[x̃] #
ρ E[x̃2 ]
= ρ 1+
2(1 − ρ) 2E[x̃]
.

In terms of Cx̃2 ,
" #
ρ 1 + Cx̃2
E[ñ] = ρ 1 + .
1−ρ 2
Taking the second derivatives of α(z) and β(z) evaluated at z = 1,

d2 ∗ 
 
′′ 2
= (1 − ρ)λ2 E[x̃2 ],

α (z) = (1 − ρ)(−λ) F (s)
z=1 ds2 x̃ s=0
and
d2 ∗ 
 
′′ 2
= −ρλ2 E[x̃2r ].

β (z) = (−ρ)(−λ) 2
Fx̃ (s)
z=1 ds s=0
1−Fx̃∗ (s)
Now, Fx̃∗r (x) = sE[x̃] . Thus,

d ∗ 1 1 1 d ∗
 
F (s) = − (1 − Fx̃∗ (s)) 2 − F (s)
ds x̃r E[x̃] " s s ds # x̃
d ∗
1 1 − Fx̃∗ (s) + s ds Fx̃ (s)
= − 2
.
E[x̃] s

Applying L’Hôpital’s rule twice,

3 d2
s d 3 F ∗ (s) + ∗
" #
d 1 ds2 Fx̃ (s)
lim Fx̃∗r (s) = − lim ds x̃ ,
s→0 ds E[x̃] s→0 2

so that
E[x̃2 ]
E[x̃r ] = .
2E[x̃]
Next we obtain
d ∗ d 2
2[1 − Fx̃∗ (s) + s ds Fx̃ (s)] + s2 ds ∗
" #
d2 ∗ 1 2 Fx̃ (s)
F (s) = .
ds2 x̃ E[x̃] s3
186 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

By applying L’Hôpital’s rule four times, we find

d2 ∗ E[x̃3 ]
lim Fx̃ (s) = .
s→0 ds2 r
3E[x̃]

Therefore,
E[x̃3 ]
E[x̃2r ] = .
3E[x̃]
Consequently,
′′ −ρλ2 E[x̃3 ]
limz→1 β (z) = .
3 E[x̃]
We may now specify

d2

= E[ñ(ñ − 1)] = E[ñ2 ] − E[ñ]

2
Fñ (z)
dz z=1
2]
2ρ(1 − ρ)[−λρ] E[x̃
2E[x̃]
2 2
= λ E[x̃ ] −
(1 − ρ)2
" #2
2
(1 − ρ)ρλ E[x̃3 ] 2(1 − ρ)λ2 ρ2 E[x̃2 ]
+ +
(1 − ρ)2 3E[x̃] (1 − ρ)3 2E[x̃]
ρ 3 2
E[x̃ ] ρ 3 E[x̃3 ]
= λ2 E[x̃2 ] + +
(1 − ρ) 2E 2 [x̃] 3(1 − ρ) E 3 [x̃]
" #2
2ρ4 E[x̃2 ]
+
4(1 − ρ)2 E[x̃]
E[x̃2 ] ρ3 E[x̃2 ] ρ3 E[x̃3 ]
= ρ2 2 + +
E [x̃] 1 − ρ E 2 [x̃] 3(1 − ρ) E 3 [x̃]
" #2
ρ4 E[x̃2 ]
+ .
2(1 − ρ)2 E[x̃]

Therefore,

E[x̃2 ] ρ3 E[x̃2 ] ρ3 E[x̃3 ]

E[ñ2 ] = ρ2 + + +
E 2 [x̃] 1 − ρ E 2 [x̃] 3(1 − ρ) E 3 [x̃]
( )2
ρ4 E[x̃2 ] ρ2 E[x̃2 ]
+ ρ +
2(1 − ρ)2 E 2 [x̃] 2(1 − ρ) E 2 [x̃]
( )
ρ2 E[x̃2 ] ρ2 E[x̃2 ] ρ3 E[x̃3 ]
= 3 + + + ρ.
2(1 − ρ) E 2 [x̃] 1 − ρ E 2 [x̃] 3(1 − ρ) E 3 [x̃]

Alternatively, we may determine a formula for E[ñ2 ] using an operator

The Basic M/G/1 Queueing System 187

argument as follows. Rewrite ñ = ñ1 + ñ2 , where

−1
Fñ1 (z) = (1 − ρ) 1 − ρFx̃∗r (λ[1 − z])

,

and
Fñ2 (z) = Fx̃∗ (λ[1 − z]).
Then

E[ñ2 ] = E[(ñ1 + ñ2 )2 ]

= E[ñ21 ] + 2E[ñ1 ]E[ñ2 ] + E[ñ22 ]. (S5.2.1)

Now,
d
E[ñj ] = Fñ (z)|z=1
dz j
and
d2


E[ñj (ñj − 1)] = 2
Fñj (z) ,
ds z=1
or, rearranging terms,

d2 d
E[ñ2j ] = 2
Fñj (z)|z=1 + Fñj (z)|z=1
dz dz
d2
= Fñ (z)|z=1 + E[ñj ].
dz 2 j

Then, for ñ1 and ñ2 as defined above, and for s(z) = λ[1 − z],
d ∗
   
−2
−(1 − ρ) 1 − ρFx̃∗r (λ[1 − z])
 
E[ñ1 ] = ρλ F (s(z)) 
ds x̃r 
z=1
1−ρ
= (ρλ)E[x̃r ],
(1 − ρ)2

and
d
 
E[ñ2 ] = −λ Fx̃∗ (s) 

= ρ.
ds z=1
The second moments of ñ1 and ñ2 are thus
2
d ∗
 
−3
E[ñ21 ] = 2(1 − ρ) 1 − ρFx̃∗r (λ[1 − z])

ρλ F (s)
−2
ds x̃r
+(1 −"ρ) 1 − ρFx̃∗r (λ[1#− z])


d2 ∗


· ρλ 2 Fx̃r (s(z))  + E[ñ1 ]
ds z=1
188 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

2(1 − ρ) (1 − ρ) 2 ρλ
= 3
(ρλ)2 E 2 [x̃r ] + 2
ρλ E[x̃2r ] + E[x̃r ],
(1 − ρ) (1 − ρ) 1−ρ

and
( )
d2
E[ñ22 ] = (−λ)2 2 Fx̃∗ (s(z)) 

+ E[ñ1 ]
ds z=1
= λ2 E[x̃2 ] + ρ.

We may now substitute these results into (S5.2.1) to obtain

2 ρ ρλ
E[ñ2 ] = 2
(ρλ)2 E 2 [x̃r ] + λ2 E[x̃2r ] + E[x̃r ]
(1 − ρ) 1−ρ 1−ρ
ρ2 λ
+2 E[x̃r ] + λ2 E[x̃r ] + ρ
1−ρ
!2
2ρ2 λ2 E[x̃2 ] ρλ E[x̃2 ] ρ2 λ E[x̃2 ]
= + + 2
(1 − ρ)2 2E[x̃] 1 − ρ 2E[x̃] 1 − ρ 2E[x̃]
ρλ 2
+λ2 E[x̃2 ] + ρ + E[x̃2r ]
( 1 − ρ )
ρ2 E[x̃2 ] 2ρ2 E[x̃2 ] ρλ2 E[x̃3 ]
= + 3(1 − ρ) + ρ +
(1 − ρ)2 2E[x̃] 2E 2 [x̃] 1 − ρ 3E[x̃]
( )
ρ2 2
E[x̃ ] 2
ρ E[x̃ ] 2 ρλ2 E[x̃3 ]
= + 3 + ρ + .
2(1 − ρ) E 2 [x̃] 1 − ρ E 2 [x̃] 3(1 − ρ) E[x̃]

This is the same expression as was found using the first method.

5-3 Jobs arrive to a single server system at a Poisson rate λ. Each job con-
sists of a random number of tasks, m̃, drawn from a general distribution
Fm̃ (m), independent of everything. Each task requires a service time
drawn from a common distribution, Fx̃t , independent of everything.

(a) Determine the Laplace-Stieltjes transform of the job service-time dis-

tribution.
(b) Determine the mean forward recurrence time of the service-time dis-
tribution using the result of part (a) and transform properties.
(c) Determine the stochastic equilibrium mean sojourn time for jobs in
this system.
(d) Determine the mean number of tasks remaining for a job in service at
an arbitrary point in time, if any.
The Basic M/G/1 Queueing System 189

Solution:
(a) Since the time between interruptions is exponentially distributed with
parameter β, the process that counts the interruptions is a Poisson
process. Hence

(βx)n e−βx
P {ñ = n|x̃ = x} = .
n!
Therefore,
∞ (βx)n e−βx
Z
P {ñ = n} = dFx̃ (x),
0 n!
and
∞
z n P {ñ = n}
X
Fñ (z) =
n=0
∞
(βx)n e−βx
zn∞
X
= dFx̃ (x)
n=0
n!
∞
(zβx)n −βx
Z ∞X
= e Fx̃ (x)
0 n=0 n!
Z ∞
= ezβx e−βx Fx̃ (x).
0

Therefore
Fñ (z) = Fx̃∗ (β[1 − z]),
which, of course, is just the pgf for the distribution of the number of
arrivals from a Poisson process over a period x̃.
Let ñ represent the total number of interruptions. Then c̃ = x̃ +
(b) P
ñ
i=0 x̃si . Therefore,
 Pñ

−sc̃ −s x̃+ x̃
i=0 si
E[e ] = E e
  ñ
 
∞
( P
Z ∞X −s x̃+ x̃si ñ=n,x̃=x
= E e i=0
 

0 n=0
)
P {ñ = n|x̃ = x} dFx̃ (x)
∞
∞X n (βx)n
Z
e−sx Fx̃∗s (s) e−βx dFx̃ (x)

=
0 n=0
n!
Z ∞
−sx βxFx̃∗s (s) −βx
= e e e dFx̃ (x).
0
190 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

Thus,
Fc̃ (s) = Fx̃∗ s + β − βFx̃∗s (s) .

(c) The formula for the M/G/1 busy period is as follows:

 
Fỹ∗ (s) = Fx̃∗ s + λ − λFỹ∗ (s) .

The relationship between the two formulas is that if the service time
of the special customer has the same distribution as the length of a
busy period, then the completion time has the same distribution as the
length of an ordinary busy period. The relationship is explained by
simply allowing the first customer of the busy period to be interrupted
by any other arriving customer. Then each interruption of the first
customer has the same distribution as the ordinary busy period. When
service of the first customer is complete, the busy period is over. Thus,
the busy period relationship is simply a special case of the completion
time.
(d) Since c̃ = x̃ + ñi=0 x̃si ,
P

E[c̃|x] = x + βxE[x̃s ],
which implies
E[c̃] = E[x̃] + βE[x̃]E[x̃s ].
Therefore
E[c̃] = E[x̃](1 + ρs ),
where ρs = βE[x̃s ]. As a check,
d ∗  d d
   
= Fx̃∗ (s)
 
Fc̃ (s) 1 − β Fx̃s (s)  .
ds s=0 ds s=0 ds s=0
It follows that
E[c̃] = E[x̃] (1 + βE[x̃s ]) .
Similarly,
" #
2
d2 d2 ∗ d ∗ d ∗ d2 ∗
 
Fc̃ (s) = F (s) 1 − β F (s) + F (s) −β F (s) .
ds2 ds2 x̃ ds x̃ ds x̃ ds2 x̃
Taking the limit as s → 0,
E[c̃2 ] = E[x̃2 ] (1 + βE[x̃s ])2 + E[x̃]βE[x̃2s ].
(e) Note that the probability that the server is busy is just the expected
value of the number of customer is service. That is, if B represents
the event the the server is busy,
P {B} = λE[c̃] = λE[x̃](1 + ρs )
The Basic M/G/1 Queueing System 191

= ρ(1 + ρs ).

The stability condition is then ρ(1 + ρs ) < 1. Rewriting ρ as λE[x̃]

the stability condition is then
1 1
λ< .
E[x̃] (1 + ρs )

5-4 Consider a queueing system in which ordinary customers have service

times drawn from a general distribution with mean 1/µ. There is a spe-
cial customer who receives immediate service whenever she enters the
system, her service time being drawn, independently on each entry, from
a general distribution, Fx̃s (x), which has mean 1/α. Upon completion
of service, the special customer departs the system and then returns after
an exponential, rate β, length of time. Let x̃si denote the length of the
ith interruption of an ordinary customer by the special customer, and let ñ
denote the number of interruptions. Also, let c̃ denote the time that elapses
from the instant an ordinary customer enters service until the instant the
ordinary customer departs.

(a) Suppose that service time for the ordinary customer is chosen once.
Following an interruption, the ordinary customer’s service resumes
from the point of interruption. Determine P {ñ = n|x̃ = x}, the con-
ditional probability that the number of interruptions is n, and Fñ (z),
the probability generating function for the number of interruptions
suffered by the ordinary customer.

(b) Determine Fc̃∗ (s), the Laplace-Stieltjes transform for c̃ under this pol-
icy. [Hint: Condition on the the length of the service time of the or-
dinary customer and the number of service interruptions that occur.]

(c) Compare the results of part (b) with the Laplace-Stieltjes transform
for the length of the M/G/1 busy period. Explain the relationship
between these two results.
(d) Determine E[c̃] and E[c̃2 ].
(e) Determine the probability that the server will be busy at an arbitrary
point in time in stochastic equilibrium, and the stability condition for
this system.
192 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

Solution:

(a) Since the time between interruptions is exponentially distributed with

parameter β, the process that counts the interruptions is a Poisson
process. Hence
(βx)n e−βx
P {ñ = n|x̃ = x} = .
n!
Therefore,
∞ (βx)n e−βx
Z
P {ñ = n} = dFx̃ (x),
0 n!
and
∞
z n P {ñ = n}
X
Fñ (z) =
n=0
∞
(βx)n e−βx
zn∞
X
= dFx̃ (x)
n=0
n!
∞
(zβx)n −βx
Z ∞X
= e Fx̃ (x)
0 n=0 n!
Z ∞
= ezβx e−βx Fx̃ (x).
0

Therefore
Fñ (z) = Fx̃∗ (β[1 − z]),
which, of course, is just the pgf for the distribution of the number of
arrivals from a Poisson process over a period x̃.
(b) Let
Pñ
ñ represent the total number of interruptions. Then c̃ = x̃ +
i=0 x̃si . Therefore,
 Pñ

−sc̃ −s x̃+ x̃
i=0 si
E[e ] = E e
  ñ
 
∞
( P
Z ∞X −s x̃+ x̃si ñ=n,x̃=x
= E e i=0
 

0 n=0
)
P {ñ = n|x̃ = x} dFx̃ (x)
∞
∞X n (βx)n
Z
e−sx Fx̃∗s (s) e−βx dFx̃ (x)

=
0 n=0
n!
Z ∞
−sx βxFx̃∗s (s) −βx
= e e e dFx̃ (x).
0
The Basic M/G/1 Queueing System 193

Thus,
Fc̃ (s) = Fx̃∗ s + β − βFx̃∗s (s) .

(c) The formula for the M/G/1 busy period is as follows:

 
Fỹ∗ (s) = Fx̃∗ s + λ − λFỹ∗ (s) .

The relationship between the two formulas is that if the service time
of the special customer has the same distribution as the length of a
busy period, then the completion time has the same distribution as the
length of an ordinary busy period. The relationship is explained by
simply allowing the first customer of the busy period to be interrupted
by any other arriving customer. Then each interruption of the first
customer has the same distribution as the ordinary busy period. When
service of the first customer is complete, the busy period is over. Thus,
the busy period relationship is simply a special case of the completion
time.
(d) Since c̃ = x̃ + ñi=0 x̃si ,
P

E[c̃|x] = x + βxE[x̃s ],
which implies
E[c̃] = E[x̃] + βE[x̃]E[x̃s ].
Therefore
E[c̃] = E[x̃](1 + ρs ),
where ρs = βE[x̃s ]. As a check,
d ∗  d d
   
= Fx̃∗ (s)
 
Fc̃ (s) 1 − β Fx̃s (s)  .
ds s=0 ds s=0 ds s=0
It follows that
E[c̃] = E[x̃] (1 + βE[x̃s ]) .
Similarly,
" #
2
d2 d2 ∗ d ∗ d ∗ d2 ∗
 
Fc̃ (s) = F (s) 1 − β F (s) + F (s) −β F (s) .
ds2 ds2 x̃ ds x̃ ds x̃ ds2 x̃
Taking the limit as s → 0,
E[c̃2 ] = E[x̃2 ] (1 + βE[x̃s ])2 + E[x̃]βE[x̃2s ].
(e) Note that the probability that the server is busy is just the expected
value of the number of customer is service. That is, if B represents
the event the the server is busy,
P {B} = λE[c̃] = λE[x̃](1 + ρs )
194 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

= ρ(1 + ρs ).

The stability condition is then ρ(1 + ρs ) < 1. Rewriting ρ as λE[x̃]

the stability condition is then
1 1
λ< .
E[x̃] (1 + ρs )

5-5 Consider a queueing system that services customers from a finite popula-
tion of K identical customers. Each customer, while not being served or
waiting, thinks for an exponentially distributed length of time with param-
eter λ and then joins a FCFS queue to wait for service. Service times are
drawn independently from a general service time distribution Fx̃ (x).

(a) Given the expected length of the busy period for this system, describe
a procedure through which you could obtain the expected waiting
time. [Hint: Use alternating renewal theory.]
(b) Given the expected length of the busy period with K = 2, describe a
procedure for obtaining the expected length of the busy period for the
case of K = 3.

Solution:
(a) First note if B denotes the event that the system is busy,
E[ỹ]
P {B} = .
E[ỹ] + E[ĩ]
Then, by Little’s result,

P {B} = λef f E[x̃],

where λef f is the average arrival rate of jobs to the server. Since the
λ f
customer are statistically identical, λc = ef K is the job generation
rate per customer. Now, a customer generates one job per cycle; hence
λc is equal to the inverse of the expected cycle length. i.e.,
1
λc = .
E[t̃] + E[w̃] + E[x̃]
Therefore, by Little’s result,
E[x̃] E[ỹ]
P {B} = K = .
E[t̃] + E[w̃] + E[x̃] E[ỹ] + E[c̃]
The Basic M/G/1 Queueing System 195

We may then solve for E[w̃] to find

KE[x̃]
E[w̃] = − E[x̃] − E[t̃]
P {B}
E[x̃]
= [K − P {B}] − E[t̃]
P {B}
K
 
= E[x̃] − 1 − E[t̃]
" P {B} #
KE[ỹ] + KE[ĩ]
= E[x̃] − E[t̃]]
E[ỹ]
" #
E[ĩ]
= E[x̃] K + K − 1 + E[t̃].
E[ỹ]

(b) Let ỹK denote the length of a busy period for a system having K
customers, and let Ai denote the event that i arrivals occur during
x̃1 . Further, define ỹij t be the length of a busy period starting with i
customers in the service system and j total customers in the system.
Then
K−1 h i
E[ỹ1K ] E ỹiK |Ai P {Ai } .
X
= E[x̃] +
i=1
In particular, for K = 3,
2 h i
E[ỹ13 ] = E[x̃] + E ỹi3 |Ai P {Ai }
X

i=1
= E[x̃] + E[ỹ13 ]P1 + E[ỹ23 ]P2 .

Now, ỹ23 = ỹ12 + ỹ13 , because the service can be reordered so that one
customer stands aside while the busy period of the second customer
completes and then the first customer is brought back and its busy
period completes. From the point of view of the second customer,
the whole population is only two, but the first customer sees all three
customers. Thus,
 
E[ỹ13 ] = E[x̃] + E[ỹ13 ]P1 + E[ỹ12 ] + E[ỹ13 ] P2 ,

which implies

E[ỹ13 ](1 − P1 − P2 ) = E[x̃] + E[ỹ12 ]P2 .

That is,
E[x̃] + E[ỹ12 ]P2
E[ỹ13 ] = ,
1 − P1 − P2
196 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

where E[ỹ12 ] is the given quantity, namely the length of the busy pe-
riod if the population size is two.
5-6 For the M/G/∞ queueing system, it is well known that the stochastic equi-
librium distribution for the number of busy servers is Poisson with param-
eter λE[x̃], where λ is the arrival rate and x̃ is the holding time.

(a) Suppose that x̃ = i with probability pi for i = 1, 2, 3 with p1 +

p2 + p3 = 1. Determine the equilibrium distribution of the number of
servers that are busy serving jobs of length i for i = 1, 2, 3, and the
distribution of the number of servers that are busy serving all jobs.
(b) Determine the probability that a job selected at random from all of the
jobs in service at an arbitrary point in time in stochastic equilibrium
will have service time i, i = 1, 2, 3.
(c) Calculate the mean length of an arbitrary job that is in service at an
arbitrary point in time in stochastic equilibrium.
(d) Suppose that job service times are drawn from an arbitrary distribu-
tion Fx̃ (x). Repeat part (c).
(e) What can be concluded about the distribution of remaining service
time of a customer in service at an arbitrary point in time in stochastic
equilibrium for the M/G/∞ system?

Solution:

(a) From the problem statement, the equilibrium distribution of the num-
ber of busy servers is Poisson with parameter λE[x̃]. Now, if x̃ = i
with probability pi , then the arrival process consists of 3 independent
arrival processes with arrival rates λpi , i = 1, 2, 3. Since there are
an infinite number of servers, we may consider each arrival stream as
being served by an independent, infinite set of servers. Thus, if we
denote by ñi the number of busy servers for class i, then

(λpi i)ni e−λpi i

P {ñi = n} = .
ni !
The number of jobs in service at an arbitrary point in time is simply
ñ = ñ1 + ñ2 + ñ3 . Since this is the sum of three independent Poisson
random variables, then by Property 1 of the Poisson process (page 44
of the text) we have that ñ is Poisson distributed with parameter λ.
i.e.,
 P n
3 P3
λ i=1 ipi −λ
i=1
ipi
P {ñ = n} = e n!
n!
The Basic M/G/1 Queueing System 197

(λE[x̃])n −λE[x̃]
= e ,
n!
as in the problem statement.
(b) Since ñ is a Poisson random variable that is obtained by the addition
of 3 independent Poisson random variables, it follows that the propor-
tion of calls of type i is simply
λipi ipi
P3 = .
λ i=1 ipi E[x̃]

That is, it is as though there is a Poisson random variable with param-

eter β, and each item is marked with color i with probability ββi .
(c) From part (b), the probability that a job is type i is ipi /E[x̃]. There-
fore, if x̃0 denotes the length of an observed job, P {x̃0 = x} =
ipi /E[x̃], and
n n
X [ipi ] X [i2 pi ]
E[x̃0 ] = i =
i=1
E[x̃] i=1
E[x̃]
2
E[x̃ ]
= .
E[x̃]

(d) By the same argument as in part (c),

xdFx̃ (x)
P {x ≤ x̃0 ≤ x + dx} = ,
E[x̃]
so that
∞ xdFx̃ (x) E[x̃2 ]
Z  
E[x̃0 ] = x = .
0 E[x̃] E[x̃]
(e) The distribution of the remaining service time is the same as the distri-
bution of the remaining service time of a renewal interval in a renewal
process whose interval distribution is Fx̃ (x).
Chapter 6

THE M/G/1 QUEUEING SYSTEM WITH PRIORITY

Exercise 6.1 Argue the validity of (6.1).

Solution.
Exercise 6.2 Derive an expression for the Laplace-Stieltjes transform
of the sojourn-time distribution for the M/G/1 system under the LCFS-PR
discipline conditional on the customer’s service time requirement. [Hint:
See Exercise 5.13].
Solution. For ease of notation, let s̃ represent in this solution only the sojourn
time of an arbitrary customer for a system having the LCFS-PR discipline.
Then, conditioning the Laplace-Stieltjes transform of this variable on the ser-
vice time requirement of the customer,
Z ∞
Fs̃∗ (s) = E[e−ss̃ |x̃ = x]dFx̃ (x).
0
But this conditional may in turn be conditioned on the number of customers
who arrive during the system time of the arbitrary customer. i.e.,
∞
E[e−ss̃ |x̃ = x] = E[e−ss̃ |x̃ = x, ṽ = v]P {ṽ = v}
X

v=0
∞
e−λx (λx)v
E[e−ss̃ |x̃ = x, ṽ = v]
X
= .
v=0
v!
Now, if v customers arrive during the system time of the arbitrary customer,
then the sojourn time will be the service time requirement plus the busy periods
generated by those v customers. But, each of the v customers have the same
busy period distribution ỹ, with ỹ = 0 with probability 1. Hence,
∞ h Pv i e−λx (λx)v
E[e−ss̃ |x̃ = x] = E e−s(x+ ỹ)
X
i=1

v=0
v!
200 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

∞ v
" #
X
−sx
Y
−sỹ e−λx (λx)v
= e E e
v=0 i=1
v!
∞  v e−λx (λx)v
e−sx Fỹ∗ (s)
X
=
v=0
v!
 v
∞ λxFỹ∗ (s)
= e−sx e−λx
X

v=0
v!
−sx −λx λxFỹ∗ (s)
= e e e ∗
= e−x[s+λ−λFỹ (s)] .

Therefore,  
Fs̃∗ (s) = Fx̃∗ s + λ − λFỹ∗ (s) .

Exercise 6.3 Compare the means and variances of the sojourn times for
the ordinary M/G/1 system and the M/G/1 system under the LCFS-PR dis-
cipline.
Solution. To compute the mean of Fs̃∗LCF S−P R (s), recall that

Fs̃∗LCF S−P R (s) = Fỹ∗ (s).

Hence, by (5.22),
E[x̃]
E[s̃LCF S−P R ] = .
1−ρ
Now, since  
Fỹ∗ (s) = Fx̃∗ s + λ − λFỹ∗ (s) ,
we may rewrite the LST of s̃LCF S−P R as

Fs̃∗LCF S−P R (s) = Fx̃∗ (u(s)) ,

where
u(s) = s + λ − λFỹ∗ (s).
Thus,

d2 ∗ d d ∗ d
 
F (s) = Fx̃ (u(s)) u(s)
ds2 ỹ ds du ds 
2
d2 ∗ d d ∗ d2

= F (u(s)) u(s) + F (u(s)) u(s)
du2 x̃ ds du x̃ ds2
2
d2 ∗ d ∗

= Fx̃ (u(s)) 1 − λ F ỹ (s)
du2 ds
The M/G/1 Queueing System with Priority 201
" #
d d2
+ Fx̃∗ (u(s)) −λ 2 Fỹ∗ (s) .
du ds

We then evaluate this expression at s = 0 to find

d2 ∗ 
  
2 2 2
F (s) = E[x̃ ] (1 + λE[ỹ]) + (−E[x̃]) −λE[ỹ ]
ds2 ỹ s=0
2
ρ
  
2
= E[x̃ ] 1 + − E[x̃] −λE[ỹ 2 ]
1−ρ
= E[ỹ 2 ],

where we have used the above result for E[ỹ]. Then solving for E[ỹ 2 ],

E[x̃2 ]
E[ỹ 2 ] = ,
(1 − ρ)3
so that

Var(s̃LCF S−P R ) = Var(ỹ)

= E[ỹ 2 ] − E 2 [ỹ]
E[x̃2 ] E[x̃] 2
 
= −
(1 − ρ)3 1−ρ
E[x̃2 ] − E 2 [x̃] + ρE 2 [x̃]
=
(1 − ρ)3
2
E [x̃] h 2 i
= C x̃ + ρ .
(1 − ρ)3

It remains to compute Var(s̃). Observe that s̃ = w̃ + x̃, where w̃ and x̃ are

independent since the service time in question occurs after the waiting period
is over. Thus Var(s̃) = Var(w̃) + Var(x̃). Rewrite Fw̃∗ (s) as
1−ρ
Fw̃∗ (s) = .
1 − ρFx̃∗r (s)

Then
d2 ∗ d −2 d ∗
 
(1 − ρ) 1 − ρFx̃∗r (s)

F (s) = ρ Fx̃r (s)
ds2 w̃ ds ds 2
d ∗

 ∗ −3
= 2(1 − ρ) 1 − ρFx̃r (s) ρ Fx̃r (s)
ds
−2 d2 ∗
+(1 − ρ) 1 − ρFx̃∗r (s)

ρ 2 Fx̃r (s).
ds
202 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

Upon evaluating this expression at s = 0 and using the relation E[x̃nr ] =

E[x̃n+1 ]/(n + 1)E[x̃], we find that
1−ρ 2 2 1−ρ
E[w̃2 ] = 2 3
ρ E [x̃r ] + ρE[x̃2r ]
(1 − ρ) (1 − ρ)2
!2 
ρ2 E[x̃2 ] ρ E[x̃3 ]

= 2 +
(1 − ρ)2 2E[x̃] 1 − ρ 3E[x̃]
" #2
ρ E[x̃2 ] ρ E[x̃3 ]
 
= 2 +
1 − ρ 2E[x̃] 1−ρ 3E[x̃]
" !#2
ρE[x̃] Cx̃2 + 1 ρ E[x̃3 ]
 
= 2 +
1−ρ 2 1−ρ 3E[x̃]

Now, by (5.17), !
ρE[x̃] Cx̃2 + 1
E[w̃] = ,
1−ρ 2
so that the variance of w̃ is then
Var(w̃) = E[w̃2 ] − E 2 [w̃]
" !#2 
ρE[x̃] Cx̃2 + 1 ρ E[x̃3 ]

= 2 +
1−ρ 2 1 − ρ 3E[x̃]
" !#2
ρE[x̃] Cx̃2 + 1
−
1−ρ 2
" #2
ρE[x̃] Cx̃2 + 1 ρ E[x̃3 ]
 
= + .
1−ρ 2 1−ρ 3E[x̃]

Therefore,
Var(s̃) = Var(w̃) + Var(x̃)
" #2 
ρE[x̃] Cx̃2 + 1 ρ E[x̃3 ]

= + + E[x̃2 ] − E 2 [x̃]
1−ρ 2 1 − ρ 3E[x̃]
" #2 
ρ (1 + Cx̃2 ρ E[x̃3 ]
   
= E 2 [x̃] + Cx̃2 + .
 1−ρ 2  1−ρ 3E[x̃]

Now,
" #
ρ 1 + Cx̃2 E[x̃]
E[s̃F CF S ] − E[s̃LCF S−P R ] = E[x̃] 1 + −
1−ρ 2 1−ρ
The M/G/1 Queueing System with Priority 203
" #
ρ 1 + Cx̃2 1
= E[x̃] 1 + −
1−ρ 2 1−ρ
" #
ρ 1 + Cx̃2 ρ
= E[x̃] −
1−ρ 2 1−ρ
" #
ρE[x̃] 1 + Cx̃ 2
= −1
1−ρ 2
" #
ρE[x̃] Cx̃2 − 1
= .
1−ρ 2

Thus, if Cx̃2 < 1, then E[s̃LCF S−P R ] < E[s̃F CF S ], and if Cx̃2 > 1, then
E[s̃LCF S−P R ] > E[s̃F CF S ]. If Cx̃2 = 1, then E[s̃LCF S−P R ] = E[s̃F CF S ].
This result is clear for exponential service because of Little’s result, which is
reasoned as follows. Since the service time is memoryless, then upon each
arrival, the service time of the customer in service is redrawn again from the
same distribution as all other customers. Thus, all customers are alike, and it
does not matter which order the service occurs; the distribution of the number
of customers in the system is unaffected by order of service. Therefore, the
average number of customers is unaffected by service order. Thus, by Little’s
result, E[s̃LCF S−P R ] = E[s̃F CF S ].We now compare the variances of the two
distributions.
 
∆s̃ Cx̃2 = Var(s̃F CF S ) − Var(s̃LCF S−P R )
" !#2 
 ρ 1 + Cx̃2  ρ E[x̃3 ]
= E 2 [x̃] + Cx̃2 +
 1−ρ 2  1 − ρ 3E[x̃]
E 2 [x̃] h 2 i
− Cx̃ + ρ .
(1 − ρ)3

That is,
(
  E 2 [x̃] ρ2 (1 − ρ)  2
∆s̃ Cx̃2 = 1 + Cx̃2
(1 − ρ)3 4
)
3 ρ E[x̃3 ]
+(1 − ρ) Cx̃2 − Cx̃2 −ρ + .
1 − ρ 3E[x̃]

At Cx̃2 = 0, E[x̃3 ] = E 3 [x̃], so that for all ρ,

( )
  E 2 [x̃] ρ2 (1 − ρ) ρ E[x̃3 ]
∆s̃ Cx̃2 = − ρ +
(1 − ρ)3 4 1 − ρ 3E[x̃]
2
E [x̃]ρ
= − {8 + 4ρ + ρ(1 − ρ)} < 0.
12(1 − ρ)3
204 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

This means that var(s̃LCF S−P R ) > var(s̃F CF S ) if Cx̃2 = 0 for deterministic
service.
At Cx̃2 = 1, with exponential service,
E[x̃2 ] 6E 3 [x̃]
= = 2E 2 [x̃].
3E[x̃] 3E[x̃]
Hence, at Cx̃2 = 1 and for all ρ,
E 2 [x̃] ρ(1 − ρ)4
   
∆s̃ Cx̃2 = 3
+ (1 − ρ)3 − 1 − ρ + 2ρ(1 − ρ)2
(1 − ρ) ( 4
E 2 [x̃]
= ρ − ρ2 + 1 − 3ρ + 3ρ2 − ρ3 − 1 − ρ
(1 − ρ)3
)
2 3
+2ρ − 4ρ + 2ρ

ρE 2 [x̃] n 2
o
= −1 − 2ρ + ρ
(1 − ρ)3
ρE 2 [x̃] n o
= − 3
1 + 2ρ − ρ2 < 0.
(d1 − ρ)

Again, this means that var(s̃LCF S−P R ) > var(s̃F CF S ) at Cx̃2 = 1. i.e., with
exponential service. The question then arises as to whether Var(s̃LCF S−P R ) <
var(s̃F CF S ) for any value of Cx̃2 . This question can be answered in part by con-
sidering
E 2 [x̃] ρ E[x̃3 ]
∆s̃ (Cx̃2 = f (C 2
x̃ ) + ,
(1 − ρ)3 1 − ρ 3E[x̃]
where
ρ2 (1 − ρ)
f (Cx̃2 ) = (1 + Cx̃2 ) + (1 − ρ)3 Cx̃2 − Cx̃2 + ρ.
4
Since x̃ is a nonnegative random variable, we conjecture that E[x̃3 ] is an in-
creasing function of Cx̃2 . Now,
d ρn o
f (Cx̃2 ) = ρ(1 − ρ)Cx̃2 − [6 − 7ρ + 3ρ2 ]
Cx̃2 2 ( )
ρ2 (1 − ρ) 2 6 − 7ρ + 3ρ2
= Cx̃ − .
2 ρ(1 − ρ)

This shows that for sufficiently large Cx̃2 , f (Cx̃2 ) is an increasing function. That
is, if
6 − 7ρ + 3ρ2
Cx̃2 > ,
ρ(1 − ρ)
The M/G/1 Queueing System with Priority 205

then f (Cx̃2 ) is an increasing function. However, the minimum

√ value of (6 −
7ρ + 3ρ2 )/ρ(1 − ρ) √ can be shown to be at ρ = (3 − 3)/2, and at this
2
point, Cx̃ > 5 + 4 3. Therefore, one would expect that var(s̃LCF S−P R ) >
√
var(s̃F CF S )√for Cx̃2 < 5 + 4 3, and that var(s̃LCF S−P R ) < var(s̃F CF S ) for
Cx̃2 < 5 + 4 3.

Exercise 6.4 Compare the probability generating function for the class
1 occupancy distributions for the HOL system to that of the M/G/1 system
with set-up times discussed in Section 6.2. Do they have exactly the same
form? Explain why or why not intuitively.
Solution. Recall Equation (6.36)

1 − ρ1 − ρ2 ρ2 Fx̃∗2r (λ1 [1 − z])  (1 − ρ1 )Fx̃∗1 (λ1 [1 − z])

 
Fñ1 (z) =  + ,

 
∗
 
1 − ρ1 1 − ρ1 1 − ρ1 Fx̃1r (λ1 [1 − z])


and (6.24)
1 ρs
 
Fñs (z) =  Fx̃∗s (λ[1 − z]) + F ∗ (λ[1 − z])
 Fñ (z),
 
1 + ρs 1 + ρs x̃sr


where (6.36) and (6.24) represent the probability generating functions for the
occupancy distribution of HOL Class 1 customers and M/G/1 customers with
set-up times, respectively. These probability generating functions have the
exactly the same form in that they both are the probability generating function
of the sum of two random variables, one of which is the probability generating
function for the number left in the system in the ordinary M/G/1 system. In
each case, the first part of the expression represents the number left due to
arrivals that occur prior to the time an ordinary customer begins service in a
busy period.
In the case of the priority system, either there is a Class 2 customer in service
or there isn’t. If so, then the probability generating function for the number of
Class 1 customers who arrive is the same as for those who arrive during the
residual service time of the Class 2 customer in service, namely Fx̃∗2r (λ1 [1 −
z]). Otherwise, the probability generating function is 1. The probability of no
ρ2
Class 2 customers in service is 1 − 1−ρ 1
.
In the case of set-up, since this is before ordinary customers are serviced,
either the customer in question in the first customer of the busy period or not.
If so, the customer will leave behind all customers who arrive in the set-up
time; otherwise the customer will leave behind only those who arrive after and
during the set-up time. i.e., during the residual of the set-up time. The result
is that in the case of the priority system, the set-up time is either 0 or x̃2r ,
depending upon whether or not a Class 2 customer is in service or not upon
arrival of the first Class 1 customer of a Class 1 busy period. Recall Equation
206 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

(6.36)
1 − ρ1 − ρ2 ρ2 Fx̃∗2r (λ1 [1 − z])  (1 − ρ1 )Fx̃∗1 (λ1 [1 − z])
 
Fñ1 (z) =  + ,

 
∗
 
1 − ρ1 1 − ρ1 1 − ρ1 Fx̃1r (λ1 [1 − z])


and Equation (6.24)

1 ρs
 
Fñs (z) =  Fx̃∗s (λ[1 − z]) + F ∗ (λ[1 − z])
 Fñ (z),
 
1 + ρs 1 + ρs x̃sr


where (6.36) and (6.24) represent the probability generating functions for the
occupancy distribution of HOL Class 1 customers and M/G/1 customers with
set-up times, respectively. These probability generating functions have the
exactly the same form in that they both are the probability generating function
of the sum of two random variables, one of which is the probability generating
function for the number left in the system in the ordinary M/G/1 system. In
each case, the first part of the expression represents the number left due to
arrivals that occur prior to the time an ordinary customer begins service in a
busy period.
In the case of the priority system, either there is a Class 2 customer in service
or there isn’t. If so, then the probability generating function for the number of
Class 1 customers who arrive is the same as for those who arrive during the
residual service time of the Class 2 customer in service, namely Fx̃∗2r (λ1 [1 −
z]). Otherwise, the probability generating function is 1. The probability of no
ρ2
Class 2 customers in service is 1 − 1−ρ 1
.
In the case of set-up, since this is before ordinary customers are serviced,
either the customer in question in the first customer of the busy period or not.
If so, the customer will leave behind all customers who arrive in the set-up
time; otherwise the customer will leave behind only those who arrive after and
during the set-up time. i.e., during the residual of the set-up time. The result
is that in the case of the priority system, the set-up time is either 0 or x̃2r ,
depending upon whether or not a Class 2 customer is in service or not upon
arrival of the first Class 1 customer of a Class 1 busy period.

Exercise 6.5 Derive the expression for Fñ2 (z) for the case of the HOL-
PR discipline with I = 2.
Solution. The probability generating function for the number left in the system
by a departing class 2 customer is readily derived from (6.45). As noted earlier,
the term
λ1 ∗
 
1 − γ2  (1 − z) + λ2 {1 − Fỹ11 (λ2 [1 − z])} 
  
1 + γ1  Fx̃∗2c (λ2 [1 − z]) − z
 


is the probability generating functionfor the distribution of the number of cus-

tomers who arrive during the waiting time of the departing customer while the
The M/G/1 Queueing System with Priority 207

term
Fx̃∗2 (λ2 [1 − z])
is the probability generating functionfor the distribution of the number of cus-
tomers who arrive during the service time of the departing customer. Now, the
distribution of the number of customers who arrive during the waiting time is
the same whether or not servicing of the class 2 customer is preemptive. On the
other hand, the number of class 2 customers who arrive after the time at which
the class 2 customer enters service and the same customer’s time of departure
is equal to the number of class 2 customers who arrive during a completion
time of the class 2 customer, the distribution of which is Fx̃2 c (x). The prob-
ability generating functionof the number of class 2 arrivals during this time is
then

Fx̃∗2 c (s)|s=λ2 [1−z] = Fx̃∗2 (s + λ1 − λ1 Fỹ∗11 (s))|s=λ2 [1−z]

= Fx̃∗2 (λ2 [1 − z] + λ1 − λ1 Fỹ∗11 (λ2 [1 − z]))

The required probabiltiy generating function is then

(1 − z) + λλ12 {1 − Fỹ∗11 (λ2 [1 − z])} 

 
1 − γ2 
Fñ2 (z) =

 

1 + γ1 F ∗ (λ [1 − z]) − z

 

x̃2c 2
Fx̃∗2 (λ2 [1 − z] + λ1 − λ1 Fỹ∗11 (λ2 [1 − z]))

Exercise 6.6 Derive expressions for Fñ1 (z), Fñ2 (z), and Fñ3 (z) for the
ordinary HOL discipline with I = 3. Extend the analysis to the case of
arbitrary I.
Solution. Consider the number of customers left by a class j departing cus-
tomer by separating the class j customers remaining into two sub-classes:
those who arrive during a sub-busy period started by a customer of their own
class are Type 1; all others are Type 2. As before, the order of service is that
the service takes place as a sub-busy period of a sequence of sub-busy periods,
each being started by a Type 2 customer and generated by Type 1 and higher
priority customers. Then, ñj = ñj1 + ñj2 . Since these customers arrive in
non-overlapping intervals and are due to Poisson processes, ñj1 and ñj2 are
independent.
As we have argued previously, ñj1 is the sum of two independent random
variables. The first is the number of Type 1 customers who arrive during the
waiting time of the tagged customer, which has the same distribution as the
corresponding quantity in an ordinary M/G/1 queueing system having traffic
intensity γj and service time equivalent to the completion time of a class j
208 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

customer. The pgf for the distribution of the number of Type 1 customers in
this category is
(1 − γj )
. (6.6.1)
1 − γj Fx̃∗jcr (λj [1 − z])
In turn, Fx̃∗jc (s) = Fỹ∗j (s). That is, the completion time of a class j customer
H
is simply the length of a sub-busy period started by a class j customer and
generated by customers having priority higher than j. Furthermore,
 
Fỹ∗jH (s) = Fx̃∗j s + λH − λH Fỹ∗HH (s) (6.6.2),

where Fỹ∗HH (s) satisfies

 
Fỹ∗HH (s) = Fx̃∗H s + λH − λH Fỹ∗HH (s) , (6.6.3),

which is the LST of the length of a busy period started by a high priority cus-
tomer and generated by customers whose priority is higher than j. Specifically,
j−1
X
λH = λk , (6.6.4)
k=1

and
j−1
1 X
Fx̃∗H (s) = λk Fx̃∗k (s), (6.6.5)
λH k=1
with λH = 0 and Fx̃∗H (s) undefined if j < 2. In addition,

1 − Fx̃∗jc (s)
Fx̃jcr (s) = (6.6.6)
sE[x̃jc ]
and
E[x̃j ]
E[x̃jc ] = , (6.6.7)
1 − σj−1
where
j
X
σj = Pk .
k=1
λj E[x̃j ]
Since γj = 1−σj−1 ,

∗
λj E[x̃j ] 1 − Fx̃jc (s)
γj Fx̃∗jcr (s) = sE[x̃j ]
1 − σj−1
h 1−σi
j−1

λj 1 − Fx̃∗jc (s)
=
s
The M/G/1 Queueing System with Priority 209
h i
λj 1 − Fỹ∗jH (s)
= (6.6.8)
s

Returning to the second component of ñj1 , this is just the number of class j
customers who arrive during the service time of the class j customer. The pgf
for the distribution of the number of such customers is simply Fx̃∗j (λj [1 − z]).
Thus we find on making appropriate substitutions that
(1 − γj )Fx̃∗j (λj [1 − z])
Fñj1 (z) =
1 − γj Fx̃∗jcr (λj [1 − z])
(1 − γj )Fx̃∗j (λj [1 − z])
= h i
λj 1−Fỹ∗ (λj [1−z])
jH
1− λj [1−z]
(1 − γj )(1 − z)Fx̃∗j (λj [1 − z])
=
Fỹ∗jH (λj [1 − z]) − z
(1 − γj )(z − 1)Fx̃∗j (λj [1 − z])
=  
z − Fx̃∗j λj [1 − z] + λH − λH Fỹ∗HH (λj [1 − z])

by (6.6.2). We now turn to the distribution of the number of Type 2 customers

left in the system. Such customers arrive to the system when one of the fol-
lowing three events is occuring:
I, the event that the system is idle.
L, the event that the system is in a completion time of a lower priority cus-
tomer excluding periods of time covered by event E.
H, the event that a higher priority customer is in service excluding periods
covered by events E and L. We also define
E, the event that the system is in a completion time of a class j customer.
The completion time of class j customers is equal to the length of a sub-busy
period of higher priority customers started by a class j customer. Similarly, the
completion time of a lower priority customer excluding periods of completion
times of class j customers is equal to the length of a sub-busy period of a higher
priority customers started by a lower priority customer. Since there can be at
most 1 completion time in progress at any given time, the event probabilities
are readily calculated by applying Little’s result. Hence
λj E[x̃j ] ρj
P {E} = = = γj ,
1 − ρH 1 − ρH
and
λL E[x̃L ] ρL
P {L} = = .
1 − ρH 1 − ρH
210 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

The proportion of time spent during events E and L serving higher priority
customers is readily computed by simply subtracting ρj and ρL from P {E}
and P {L} respectively. We then find
P {H} = ρH − [P {E} − ρj ] −
 [P{L} − ρL ]
ρj ρL

= ρH − − ρj − − ρH ,
1 − ρH 1 − ρH
(1 − ρ)ρH
P {H} = .
1 − ρH
Next we find the probabilities
 1−ρ
P I|Ē =
1 − γj
 (1 − ρ)ρH
P H|Ē =
(1 − ρH )(1 − γj )
 ρL
P L|Ē =
(1 − ρH )(1 − γj )
Now, the number of Type 2 customers left in the system by an arbitrary de-
parture from the system is equal to the number of Type 2 customers left in the
system by an arbitrary Type 2 arrival. This is because each Type 2 arrival is
associated with exactly one Type 1 sub-busy period. Now, the number of Type
2 customers left behind if a Type 1 customer
 arrives to an empty system is 0.
The number left behind for the event H|Ē is simply the number that arrive
in the residual time of a high priority busy period started by a high priority cus-
tomer, the pfg of which is given by Fỹ∗HHr (λj [1 − z]). Similarly, the number

left behind for the event L|Ē is simply the number that arrive in the residual
life of a high priority busy period started by a low priority customer, the pgf
of which is given by Fỹ∗LHr (λj [1 − z]). Now, from exceptional first service
results,  
Fỹ∗HH (s) = Fx̃∗H s + λH − λH Fỹ∗HH (s) ,
 
Fỹ∗LH (s) = Fx̃∗L s + λH − λH Fỹ∗HH (s) .
Also,
1 − Fỹ∗HH (s)
Fỹ∗HHr (s) =
sE[ỹHH ]
1 − Fỹ∗HH (s)
=
s E[x̃
1−ρH
H]
h i
(1 − ρH ) 1 − Fỹ∗HH (s)
= .
sE[x̃H ]
The M/G/1 Queueing System with Priority 211

Similarly,
h  i
(1 − ρH ) 1 − Fx̃∗L s + λH − λH Fỹ∗HH (s)
Fỹ∗LH (s) = .
sE[x̃L ]
We then find that
h i
∗
1−ρ (1 − ρ)ρH  1 − FỹHH (λj [1 − z]) 
Fñj2 (z) = + +
1 − γj (1 − γj )  λj [1 − z]E[x̃H ] 
h  i 
ρL  1 − Fx̃∗ λj [1 − z] + λH − λH Fỹ∗ (λj [1 − z]) 
L HH
.
(1 − γj )  λj [1 − z]E[x̃L ] 

Thus,
h i
(1 − ρ)λH 1 − Fỹ∗HH (λj [1 − z])
(
Fñj (z) = (1 − ρ) + +
λj [1 − z]
h  i )
λL 1 − Fx̃∗L λL [1 − z] + λH − λH Fỹ∗HH (λj [1 − z])
λj [1 − z]
(z − 1)Fx̃∗j (λj [1 − z])
.
z − Fỹ∗jH (λj [1 − z])

We also note that since

Fñj (z) = Fs̃∗j (λj [1 − z]) ,

 h i
 λH 1 − Fỹ∗HH (s) 
Fs̃∗ (s) = (1 − ρ) 1 + +
 s 
   !
1 − Fx̃∗L s + λH − λH Fỹ∗HH (s) Fx̃∗j (s)
λL   .
s 1 − λj [1 − Fỹ∗jH (s)]/s

That is,
n h io
Fs̃∗j (s) = (1 − ρ) s + λH 1 − Fỹ∗HH (s) +
Fx̃∗j (s)
!
h  i
λL 1 − Fx̃∗L s + λH − λH Fỹ∗HH (s)
s − λj + λj Fỹ∗jH (s)
212 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

Exercise 6.7 Extend the analysis of the previous case to the case of
HOL-PR.
Solution. The solution to this problem is similar to that of Exercise 6.5, and
proceeds as follows. The probability generating function for the number left in
the system by a departing class j customer is readily derived from the results
of Exercise 6.6, which are as follows
h i
(1 − ρ)λH 1 − Fỹ∗HH (λj [1 − z])
(
Fñj (z) = (1 − ρ) + +
λj [1 − z]
h  i )
λL 1 − Fx̃∗L λL [1 − z] + λH − λH Fỹ∗HH (λj [1 − z])
λj [1 − z]
(z − 1)Fx̃∗j (λj [1 − z])
,
z − Fỹ∗jH (λj [1 − z])

or
h i
(1 − ρ)λH 1 − Fỹ∗HH (λj [1 − z])
(
Fñ (z) = (1 − ρ) + +
λj [1 − z]
h  i )
λL 1 − Fx̃∗L λL [1 − z] + λH − λH Fỹ∗HH (λj [1 − z])
λj [1 − z]
( )
(z − 1)
Fx̃∗j (λj [1 − z])
z − Fỹ∗jH (λj [1 − z])

Now, the term Fx̃∗j (λj [1 − z]) is the probability generating functionfor the dis-
tribution of the number of customers who arrive during the service time of
the departing class j customer, while the remainder of the expression repre-
sents the probability generating function for the distribution of the number of
customers who arrive during the waiting time of the class j customer. The dis-
tribution of the number of customers who arrive during the waiting time is the
same whether or not servicing of the class j customer is preemptive. On the
other hand, the number of class 2 customers who arrive after the time at which
the class 2 customer enters service and the same customer’s time of departure
is equal to the number of class j customers who arrive during a completion
time of the class j customer, the distribution of which is Fx̃jc (x). The prob-
ability generating function of the number of class j arrivals during this time
is
 
Fx̃∗jc (s) = Fx̃∗j (s + λH − λH Fỹ∗11 (s))
 
s=λj [1−z] s=λj [1−z]
= Fx̃∗j (λj [1 − z] + λH − λH Fỹ∗HH (λj [1 − z])).
The M/G/1 Queueing System with Priority 213

The required probabiltiy generating function is thus

h i
(1 − ρ)λH 1 − Fỹ∗HH (λj [1 − z])
(
Fñ (z) = (1 − ρ) + +
λj [1 − z]
h  i )
λL 1 − Fx̃∗L λL [1 − z] + λH − λH Fỹ∗HH (λj [1 − z])
λj [1 − z]
(
(z ∗
− 1)Fx̃j (λj [1 − z])
)
+ λH − λH Fỹ∗HH (λj [1 − z]).
z − Fỹ∗jH (λj [1 − z])

Exercise 6.8 Suppose that the service time of the customers in an M/G/1
system are drawn from the distribution Fx̃i (x) with probability pi such that
PI
i=1 pi = 1. Determine E[w̃] for this system.
Solution. Let Si denote the even that distribution i is selected. Then, given x̃
is drawn from Fx̃i (x) with probability pi , we find
I
X
E[x̃] = E[x̃|Si ]P {Si }
i=1
XI
= E[x̃i ]pi .
i=1

Similarly,
I
E[x̃2 ] = E[x̃2i ]pi .
X

i=1
Therefore,
I
E[x̃2i ]pi
P
i=1
E[x̃r ] = I
.
P
2 E[x̃i ]pi
i=1
From (5.96), we then have
ρ
E[w̃] = E[x̃r ]
1−ρ
I
E[x̃2i ]pi
P
ρ i=1
= ·
1−ρ I
2
P
E[x̃i ]pi
i=1
214 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

I
λ
E[x̃2i ]pi
X
=
2(1 − ρ) i=1
I
1
λi E[x̃2i ]
X
=
2(1 − ρ) i=1
I
1 X E[x̃2i ]
= λi E[x̃i ]
1 − ρ i=1 2E[x̃i ]
I
1 X
= ρi E[x̃ri ],
1 − ρ i=1

where λi = λpi and ρi = λi E[x̃i ].

Exercise 6.9 Conservation Law (Kleinrock [1976]) Under the condi-

tions of Exercise 6.8, suppose the customers whose service times are drawn
from the distributionPFx̃i (x) are assigned priority i and the service discipline
is HOL. Show that Ii=1 ρi E[w̃i ] = ρE[w̃] where E[w̃] is as determined in
Exercise 5.9. Explain the implications of this result. Does the result imply
that the expected waiting time is independent of the priority assignment?
Why or why not? If not, under what conditions would equality hold?
Solution. This exercise refers to nonpreemptive service or ordinary HOL, with
PI
i=1 E[x̃ri ]ρi
E[w̃j ] =
(1 − σj )(1 − σj−1 )
ρE[x̃r ]
= .
(1 − σj )(1 − σj−1 )

We wish to show
ρE[x̃r ]
E[w̃FCFS ] = ,
1−ρ
so that
I
X
ρE[w̃FCFS ] = ρ ρi E[w̃i ]
i=1
I
X ρi
= ρE[x̃r ] .
i=1
(1 − σi )(1 − σi−1 )

Clearly, it is sufficient to show

I
X ρi ρ
= .
i=1
(1 − σi )(1 − σi−1 ) 1−ρ
The M/G/1 Queueing System with Priority 215

We prove the following proposition by induction, and the above result follows
as a special case. Proposition: For all j,
j
X ρi σj
= ,
i=1
(1 − σi )(1 − σi−1 ) 1 − σj
with σ0 = 0. proof: Let T denote the truth set for the proposition. Clearly
1 ∈ T . Suppose j − 1 ∈ T . That is,
j−1
X ρi σj−1
= .
i=1
(1 − σi )(1 − σi−1 ) 1 − σj−1
Then
j j−1
X ρi X ρi ρj
= +
i=1
(1 − σi )(1 − σi−1 ) i=1
(1 − σ i )(1 − σ i−1 ) (1 − σ j )(1 − σj−1 )
σj−1 ρj
= +
(1 − σj−1 ) " (1 − σj )(1 − σj−1# )
1 σj−1 (1 − σj ) + ρj
=
(1 − σj−1 ) (1 − σj )(1 − σj−1 )
" #
1 σj (1 − σj−1 )
=
(1 − σj−1 ) 1 − σj
σj
= .
1 − σj
Thus j ∈ T for all j, and with j = I, we have
I
X ρi σI ρ
= = .
i=1
(1 − σi )(1 − σi−1 ) 1 − σI 1−ρ
This proves the proposition. Continuing with the exercise, we have
I
X
ρE[w̃FCFS ] = ρj E[w̃j ],
j=1
or
I
X ρj
E[w̃FCFS ] = E[w̃j ].
j=1
ρ
That is, E[w̃] can be expressed as a weighted sum of the expected waiting times
of individual classes, where the weighting factors are equal to the proportion
of all service times devoted to the particular class. Thus,
I
X λj E[x̃j ]
E[w̃FCFS ] = E[w̃j ].
j=1
λE[x̃]
216 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

Now, λE[w̃FCFS ] = E[ñq ], and λj E[w̃j ] = E[ñqj ]. Hence

I
X E[x̃j ]
E[ñqFCFS ] = E[ñqj ].
j=1
E[x̃]

But
I
X
E[ñqHOL ] = E[ñqj ], (6.6.1)
j=1

so that a sufficient condition for E[ñHOL ] = E[ñqFCFS ] is that E[x̃j ] = E[x̃]

for all j. Therefore, by Little’s result, a sufficient condition for E[w̃HOL ] =
E[w̃F CF S ] is that E[x̃j ] = E[x̃] for all j. From (6.6.1), we see that
I
X
λE[w̃qHOL ] = λj E[w̃j ],
j=1

which implies
I I
X λj X ρj
E[w̃qHOL ] = E[w̃j ] 6= E[w̃j ] = E[w̃qFCFS ].
j=1
λ j=1
ρ

That is, if the service times are not identical, the average waiting time may not
be conserved.
Chapter 7

VECTOR MARKOV CHAIN ANALYSIS:

THE M/G/1 AND G/M/1 PARADIGMS

Exercise 7.1 Suppose that P {x̃ = 1} = 1, that is, the service time is
deterministic with mean 1. Determine {ak , k = 0, 1, . . .} as defined by
(7.6).
Solution: if P {x̃ = 1} = 1, then we find that dFx̃ (x) = δ(x − 1)dx, where
δ(x) is the Dirac delta function. Thus, from (5.6),
∞ (λx)k −λx
Z
ak = e dFx̃ (x)
Z0∞
k!
(λx)k −λx
= e δ(x − 1)dx
0 k!
λk −λ
= e for k = 0, 1, 2, · · ·
k!

Exercise 7.2 Suppose that x̃ is a discrete valued random variable having

support set X = {x0 , x1 , . . . , xK }, where K is an integer. Define αk =
P {x̃ = xk } for xk ∈ X . Determine P {ak , k = 0, 1, . . .} as defined by (7.6).
In order to get started, let dFx̃ (x) = x∈X αk δ(x − xk ), where δ(x) is the
Dirac delta function.
Solution: By (7.6),
∞ (λx)k −λx
Z
ak = e dFx̃ (x)
k!
Z0∞
(λx)k −λx X
= e αj · δ(x − xj )dx
0 k! x ∈X j
∞ (λx)k e−λx −λx
X Z
= αj e δ(x − xj )dx
xj ∈X 0 k!
218 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

X (λxj )k e−λxj
= αj
xj ∈X
k!
J
X (λxj )k e−λxj
= αj for k = 0, 1, · · ·
j=0
k!

Exercise 7.3 Suppose that x̃ is an exponential random variable with pa-

rameter µ. Determine {ak , k = 0, 1, . . .} as defined by (7.6)).
Solution: By (7.6),
(λx)k −λx
∞
Z
ak = e dFx̃ (x)
Z0∞
k!
(λx)k −λx −λx
= e µe dx
0 k!
[(λ + µ)x]k −(λ+µ)x
k Z ∞
λ

= µ e dx
λ+µ 0 k!
(λ + µ) [(λ + µ)x]k −(λ+µ)x
k Z ∞
µ λ

= e dx
λ+µ λ+µ 0 k!

But, the expression inside the integral is just the density of the Erlang-(k+1)
distribution with parameter (λ + µ), so that the expression integrates to unity.
Therefore,
k
µ λ

ak = , for k = 0, 1, · · ·
λ+µ λ+µ
This result is obvious since ak is just the probability that in a sequence of
independent trials, there will be k failures prior to the first success.
Exercise 7.4 Suppose that x̃ = x̃1 + x̃2 , where x̃1 and x̃2 are exponen-
tial random variables with parameter µ1 and µ2 , respectively. Determine
{ak , k = 0, 1, . . .} as defined by (7.6)).
Solution: First, we determine dFx̃ (x) = fx̃ (x)dx.
Z x
fx̃ (x) = fx̃2 (x − y)fx̃1 (y)dy
Z0x
= µ2 e−µ2 (x−y) µ1 e−µ1 y dy
0 Z x
= µ1 µ2 e−µ2 x e−(µ1 −µ2 )y dy
0

If µ1 = µ2 , we find
fx̃ (x) = µ1 µ2 e−µ2 x x
Vector Markov Chains Analysis 219

= µ22 xe−µ2 x
= µ2 (µ2 x)e−µ2 x .

If µ1 6= µ2 , we have
µ1 µ2 −µ2 x x
Z
fx̃ (x) = e (µ1 − µ2 ) e−(µ1 −µ2 )y dy
µ1 − µ2 0
µ1 µ2 −µ2 x h i
= e 1 − e−(µ1 −µ2 )x
µ1 − µ2
µ1 µ2  −µ2 x
− e−µ1 x

= e
µ1 − µ2
Then, from Exercise 7.3, we find
" k #
k
µ1 µ2 1 µ2 λ 1 µ1 λ
    
ak = −
µ1 − µ2 µ2 λ + µ2 λ + µ2 µ1 λ + µ1 λ + µ1
k
µ1 µ2 λ
 
=
µ1 − µ2 λ + µ2 λ + µ2
k
µ2 µ1 λ
 
− for k = 0, 1, · · ·
µ1 − µ2 λ + µ1 λ + µ1

Exercise 7.5 Show that the quantity Fq̃ (1) corresponds to the stationary
probability vector for the phase process.
Solution: From (5.37), we find

Fq̃ (z) [Iz − PFã (z)] = π0 [z − 1] PFã (z)

So, with z = 1,
Fq̃ (1) [I − PFã (1)] = 0
But, Fã (1) = I, so the above equation becomes

Fq̃ (1) [I − P ] = 0

i.e.,
Fq̃ (1) = Fq̃ (1)P
Therefore, Fq̃ (1) is the stationary probability vector for the Markov Chain
having transition matrix P; that is, for the phase process.

Exercise 7.6 Derive equation (7.38).

Solution: From equation (5,37), we have that

Fq̃ (z) [Iz − PFã (z)] = π0 [z − 1] PFz̃ (z)

220 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

Differentiating with respect to z, we find

′
h ′
i ′
Fq̃ (z) [Iz − PFã (z)]+Fq̃ (z) I − PFã (z) = π0 PFz̃ (z)+π0 [z − 1] PFz̃ (z)

Taking the limit as z → 1, and recalling that Fã (1) = I,

′
h ′
i
Fq̃ (1) [I − P] + Fq̃ (1) I − PFã (1) = π0 P

Postmultiplying by e,
′
h ′
i
Fq̃ (1) [I − P] e + Fq̃ (1) I − PFã (1) e = π0 Pe

That is,
′
1 − Fq̃ (1)PFã (1)e = π0 e

Exercise 7.7 Suppose there are N identical traffic sources, each of

which has an arrival process that is governed by an M -state Markov chain.
Suppose the state of the combined phase process is defined by an M -vector
in which the ith element is the number of sources currently in phase i. First,
argue that this state description completely characterizes the phase of the
arrival process. Next, show
 that the number of states of the phase process is
N +M −1

given by .
N
Solution: First note that the distribution of the number of arrivals from a given
source depends solely upon the current phase of the source. Now, since all
sources operate independently of each other, the distribution of the total num-
ber of arrivals due to all sources is simply the convolution of the distribution
of the number of arrivals from the individual sources. Since the latter depends
solely upon the phase of the individual sources, the totals number of sources in
each phase is sufficient to determine the distribution of the number of arrivals.
there are ni sources in phase i, then the pgf for the number of arrivals
That is, if Q
ni
is simply M i=1 Fi (z). Also, since the dynamics of the sources are identical,
the future evolution depends only upon the number of sources in each phase
and not upon the which sources are in which phase.
Now, if there are a total of N sources and M phases, then a typical state is
represented
PM
by the M -tuple (n1 , n2 , · · · , nM ) with ni ≥ 0 for i = 1, 2, · · · , M
and i=1 ni = N . We wish to show that the number of such states is
N +M −1
 
N
CM = ,
N
which we shalldo by induction on M . That is, we have the proposition CMN =

N +M −1

for all M , N positive integers. Let T denote the truth set of
N
Vector Markov Chains Analysis 221

the proposition. Then

N +1−1
 
C1N = =1 for all N,
N
so M = 1 ∈ T . Assume
N +M −1
 
N
CM = ,
N
for all n = 0, 1, 2, · · · , N , and m = 1, 2, · · · , M −1. i.e., assume 1, 2, · · · , M −
1 ∈ T . If the last element of the M -tuple, xM ,= i, then the remaining elements
sum to N − i so that M
P −1
j=1 xj = N − i. Thus,

N
N N −i
X
CM = CM −1
i=0
N 
N −i+M −1−1
X 
=
N −i
i=0
N 
N − i + (M − 1) − 1 N + (M − 1) − 1
X   
= +
N −i N
i=1
N −1 
N − i + (M − 1) − 1 N − (M − 1) − 1
X   
= +
N −i N
i=0

Now, we must show

N 
N − i + (M − 1) − 1 N +M −1
X   
=
N −i N
i=0

To do this, we prove
N 
N −i+x
  
X N +x+1
= ,
N −i N
i=0

by induction on N .
If N = 0, we have
   
0+x x+1
=1=
0 0
Now assume
N −1 
N −1−i+x N −1+x+1
X   
=
N −1−i N −1
i=0
222 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

Then

N  N 
N −i+x N −i+x
   
X X N +x
= +
N −i N −i N
i=0 i=0
N −1 
N −1−i+x
  
X N +x
= +
N −1−i N
i=0
N −1+x+1
  
N +x
= +
 N−1   N
N +x N +x
= +
N −1 N
(N + x)! (N + x)!
= +
(N − 1)! (x + 1)!  N !x! 
(N + x)! 1 1
= +
(N − 1)! (x + 1)! x + 1 N
(N + x + 1)!
=
N ! (x + 1)! 
N +x+1
=
N

The result now follows with x = M − 2.

Exercise 7.8 Define the matrices P and Fã (a) for the model defined in
Example 7.1 for the special case of N = 3, where the states of the phase
process have the following interpretations shown in Table 7.2. In the table,
if the phase vector is ijk, then there are i sources in phase 0, j sources in
phase 1, and k sources in phase 2.

Solution: If the phase vector is (ijk) then there are i sources in phase 0, j
sources in phase 1, and k sources in phase 2. for i, j, k = 0, 1, 2, 3. Phase
0 corresponds to the phase vector 300; i.e., all sources are in phase 0. Each
source acts independently of all others, and the only transitions from phase 0
are to phase 0 or to phase 1 for individual sources to 0, 1, 2, or 3. For transition
(0− > 0), all sources must remain in the ‘off’ state. Therefore, this transition
has probability β 3 . For (0− > 1), 2 sources remain off andthere  is one
3
transition to the ‘on’ state. Thus, the transition probability is β 2 (1 −
1
β) = 3β 2 (1 − β). Continuing in this manner results in the following table of
Vector Markov Chains Analysis 223

transition probabilities:

From To Probability
300 −> 300 β3
−> 210 2β(1 − β)
−> 120 3β(1 − β)
−> 030 (1 − β)3

210 −> 201 β2

−> 111 2β(1 − β)
−> 021 (1 − β)2

120 −> 102 β

−> 012 (1 − β)

030 −> 003 1

201 −> 300 (1 − α)β 2

−> 210 2β(1 − β)(1 − α) + β 2 α
−> 120 2β(1 − β)α + (1 − β)2 (1 − α)
−> 030 (1 − β)2 α

111 −> 201 β(1 − α)

−> 111 (1 − β)(1 − α) + βα
−> 021 (1 − β)α

021 −> 012 α

−> 102 (1 − α)

102 −> 300 (1 − α)2 β

−> 210 2β(1 − α)α + (1 − β)(1 − α)2
−> 120 βα2 + 2α(1 − β)(1 − α)
−> 030 (1 − β)α2

012 −> 201 (1 − α)2

−> 111 2α(1 − α)
−> 021 α2

003 −> 300 (1 − α)3

−> 210 3α(1 − α)2
−> 120 3α2 (1 − α)
−> 030 α3
224 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

Exercise 7.9 In the previous example, we specified Fã (z) and P. From
these specifications, we have
2
Ai z i .
X
PFã (z) =
i=0

Therefore,
2
Ai [K(1)]i ,
X
K(1) =
i=0
with
β2 2β(1 − β) (1 − β)2 0 0 0
 

 0 0 0 0 0 0
 0 0 0 0 0 0
A0 = 
 β(1 − α)
,
 βα + (1 − β)(1 − α) α(1 − β) 0 0 0
 0 0 0 0 0 0
(1 − α)2 2α(1 − α) α2 0 0 0

0 0 0 0 0 0
 
0
 0 0 β (1 − β) 0
0 0 0 0 0 0
A1 = 
0
,
 0 0 0 0 0
0 0 0 (1 − α) α 0
0 0 0 0 0 0
and
0 0 0 0 0 0
 
0 0 0 0 0 0
 
0 0 0 0 0 1
A2 = 
0

 0 0 0 0 0
0 0 0 0 0 0
0 0 0 0 0 0
Now, suppose we compute K(1) iteratively; that is, we use the formula
2
Ai [Kj−1 (1)]i
X
Kj (1) = for j ≥ 1, (7.1)
i=0

with K0 (1) = 0. Prove that the final three columns of Kj (1) are zero
columns for all j.
Solution: The proof is by induction. Consider the proposition that the last 3
columns of Kj (1) are zero columns for all j ≥ 0. Let T denote the truth set for
this proposition. Since K0 (1) = 0, 0 ∈ T . Also, K1 (1) = A0 , so that 1 ∈ T .
Vector Markov Chains Analysis 225

Now suppose (j − 1) ∈ T . Then,

Kj (1) = A0 + A1 Kj−1 (1) + A2 [Kj−1 (1)]2 .
Now, if D = AB, and the k-th column of B is a zero column, then the k-th
column of D is also 0 because clk = t alt btk , and btk = 0 for all k. Therefore,
P

with A = B = Kj−1 (1), we have [Kj−1 (1)]2 with its last 3 columns equal to
zero, and with A = A2 , B = [Kj−1 (1)]2 , we have A2 [Kj−1 (1)]2 with its last
3 columns equal to zero. Also, with A = A1 , B = Kj−1 (1), we have the last
3 columns of A1 Kj−1 (1) = 0. Since this is also true of A0 , the sum leading to
Kj (1) also has its last 3 columns equal to zero. Hence, j ∈ T , and the proof is
complete.
Exercise 7.10 Suppose K(1) has the form
K00
 
0
K(1) = , (7.2)
K10 0

where K00 is a square matrix. Bearing in mind that K(1) is stochastic, prove
that K00 is also stochastic and that κ has the form [ κ 0 ], where κ0 is the
stationary probability vector for K00 .
Solution: Recall that K(1) stochastic means that K(1) is a one-step transition
probability matrix for a discrete parameter Markov chain. That is, all entries
are nonnegative and the elements of every row sum to unity. Thus, K(1)e = e.
But
K00 0
 
K(1) = ,
K10 0
and
K00 0
  
e
K(1)e =
 K10 0 e
K00 e
=
K
 10 e
e
=
e
Therefore, the element of the rows of K00 sum to unity and are nonnegative
because they come from K(1). Since K00 is square it is a legitimate one-step
transition probability matrix for a discrete paramter Markov chain.
Now, recall that κ is the stationary vector of K(1); thus
κ · K(1) = κ,
with κe = 1 . Therefore,
K00
 
0
[ κ0 κ1 ] = [ κ0 K00 + κ1 K10 0]
K10 0
226 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

= [ κ0 κ1 ] ,

so that κ1 = 0. Thus, κ0 K00 = κ0 , and we find that κ has the form [ κ0 0 ].

Furthermore, since κe = [ κ0 0 ] · [ e e ]T = 1, we find κ0 e = 1, so that κ0
is the stationary probability vector for K00 .

Exercise 7.11 Prove Theorem 7.3.

Solution: By definition,
∞
Ai z i
X
PFã (z) =
i=0

Then, with z = 1,
∞
X
PFã (1) = P = Ai ,
i=0
so that the latter sum is stochastic.
First, note that [K0 (1)]n represents an n-step transition probability matrix
for a discrete parameter Markov chain so that this matrix is stochastic. We
may also show [K0 (1)]n is stochastic by induction on n. Let T be the truth
set for the propostion. Then 0 ∈ T since [K0 (1)]0 = I, which is stochastic.
Now assume that 0, 1, · · · , n − 1 ∈ T . Hence [K0 (1)]n−1 is nonnegative and
[K0 (1)]n−1 e = e. It follows that

[K0 (1)]n = K0 (1) [K0 (1)]n−1

= K0 (1)e = e.

And clearly, [K0 (1)]n is nonnegative for all n. Therefore, if [K0 (1)]n−1 is
stochastic, then so is Kn (1). This proves the proposition.
Now, Ai is a matrix of probabilities and is therefore nonnegative. Thus,
∞
Ai [Kj−1 ]i ,
X
Kj (1) =
i=0

so that Kj (1) is nonnegative for all j ≥ 0. Now,

∞
Ai [Kj−1 ]i e,
X
Kj (1)e =
i=0

and by the above proposition, [Kj−1 ]i e = e. Thus,

∞
X
Kj (1)e = Ai e
i=0
Vector Markov Chains Analysis 227

= Pe = e,

so that Kj (1) is also stochastic.

Exercise 7.12 Suppose that P is the one-step transition probability ma-

trix for an irreducible discrete-valued, discrete-parameter Markov chain.
Define φ to be the stationary probability vector for the Markov chain. Prove
that eφ [I − P + eφ] = eφ and that, therefore, eφ = eφ [1 − P + eφ]−1 .
Solution: Observe that

eφ [I − P + eφ] = eφ − eφP + eφeφ

But φ is the stationary probability vector for P, so that φP = φ and φe = 1.

Therefore,
eφ [I − P + eφ] = eφ − eφ + eφ = eφ.
Thus, given that [I − P + eφ] is nonsingular, we find

eφ = eφ [I − P + eφ]−1

as required.

Exercise 7.13 Define

(1) 1 (i+1) 1 h (i) i
Fq̃ (z) = [Fq̃ (z) − π0 ] and Fq̃ (z) = Fq̃ (z) − πi , i ≥ 1.
z z
(1)
Starting with (7.49), substitute a function of Fq̃ (z) for Fq̃ (z), then a func-
(2) (1)
tion of Fq̃ (z) for Fq̃ (z), and continue step by step until a function of
(C) (C−1)
Fq̃ (z) is substituted for Fq̃ (z). Show that at each step, one element
of
C−1
πj z j A(z)
X

j=0

is eliminated, resulting in (7.51).

Solution: Starting with
(1) 1
Fq̃ (z) = [Fq̃ (z) − π0 ] ,
z
we find
(1)
Fq̃ (z) = zFq̃ (z) + π0 .
For i > 0, we find
(i) (i+1)
Fq̃ (z) = zFq̃ (z) + πi .
228 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

Given
h i C−1 h i
Fq̃ (z) z C I − A(z) = πj z C Bj (z) − z j A(z) ,
X
(7.49)
j=0

we find
h ih i C−1 h i
(1)
z C I − A(z) = πj z C Bj (z) − z j A(z) ,
X
zFq̃ (z) + π0
j=0

or
h i h i C−1 h i
(1)
zFq̃ (z) z C I − A(z) = −π0 z C I − A(z) + πj z C Bj (z) − z j A(z) .
X

j=0

Upon simplifying, we find

h i h i C−1 C−1
(1)
zFq̃ (z) z C I − A(z) = −π0 z C I − A(z) + πj z C Bj (z)− z j A(z).
X X

j=0 j=0

or
h i C−1 C−1
(1)
zFq̃ (z) z C I − A(z) = −π0 z C I + πj z C Bj (z) − z j A(z).
X X

j=0 j=1

(2) (1)
We now substitute zFq̃ (z) + π1 for Fq̃ (z) in the previous equation to find

h ih i C−1 C−1
(2)
z C I − A(z) = −π0 z C I + πj z C Bj (z)− z j A(z).
X X
z zFq̃ (z) + π1
j=0 j=1

Simplifying, we get
h i C−1 C−1
2 (2) C C C C
z j A(z).
X X
z Fq̃ (z) z I − A(z) = −π0 z I−zπ1 z + πj z Bj (z)−
j=0 j=2

Continuing in this manner, we get

h i C−1 C−1
(C)
z C Fq̃ (z) z C I − A(z) = −z C πj z j + πj z C Bj (z),
X X

j=0 j=0

or
h i C−1 h i
(C)
(z) z C I − A(z) = πj Bj (z) − z j ,
X
Fq̃
j=0
Vector Markov Chains Analysis 229

which is the required result.

Exercise 7.14 Beginning with (7.68), develop an expression for the first
and second moments of the queue length distribution.

Solution: Starting with

C−1
z i πi + z C g [I − F z]−1 H,
X
Fq̃ (z) = (7.68)
i=0

we just use the properties of moment generating functions to compute the

required moments. Recall that for any integer valued, nonnegative random
d d2
Fx̃ (z)kz=1 and E x̃2 = dz
 
variable, x̃ we have E [x̃] = dz 2 Fx̃ (z)kz=1 +
d d
dz Fx̃ (z)kz=1 . In this case, E [q̃] = dz Fq̃ (1)e because Fq̃ (z)e is the generat-
ing function for the occupancy distribution. The appropriate derivatives are

C−1
d
iz i−1 πi + Cz C−1 g [I − F z]−1 H + z C gF [I − F z]−2 H,
X
Fq̃ (z) =
dz i=1

and

C−1
d2
i(i − 1)z i−2 πi + C(C − 1)z C−2 g [I − F z]−1 H
X
Fq̃ (z) =
dz 2 i=2
+2Cz C−1 gF [I − F z]−2 H + z C gF 2 [I − F z]−3 H.

We then have

C−1
d
iπi e + Cg [I − F ]−1 He + gF [I − F ]−2 He,
X
E [q̃] = Fq̃ (1)e =
dz i=1

and

h i C−1
e q̃ 2 = i(i − 1)πi e + C(C − 1)g [I − F ]−1 H
X
E
i=2
+2CgF [I − F ]−2 He + gF 2 [I − F ]−3 He + E [q̃] .
230 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

Exercise 7.15 Suppose C = 2. Define

 
A2i A2i+1
Âi = , i = 1, 2, . . . ,
A2i−1 A2i
∞
Âi z i .
X
Â(z) =
i=0

and φ2 to be the stationary vector of Â(1). Suppose

"∞ #
X
ρ = φ2 iÂi e.
i=0

Show that
1
(a) φ2 = 2 [ φ φ ], where φ is the stationary vector of A(1).
(b) "∞ #
X
φ iAi e = 2ρ.
i=0

Solution:
(a) Given  
A2i A2i+1
Âi = , i = 1, 2, . . . ,
A2i−1 A2i
we have  P∞ i P∞ i
A z i=0 A2i+1 z
Â(z) = P∞i=0 2i i ∞ i .
i=0 2i−1 z
A
P
i=0 A2i z
Upon setting z = 1, we have
 P∞ P∞ 
A i=0 A2i+1 .
Â(1) = P∞i=0 2i P∞
i=0 A2i−1 i=0 A2i
Define
∞
X ∞
X
Ae = A2i and Ao = A2i+1 .
i=0 i=0
We then have  
Ae Ao
Â(1) = .
Ao Ae
Now, suppose φ2 = [ φ2e φ2o ] is the stationary vector of Â(1). Then,
we have φ2 = φ2 Â(1) and φ2 e = 1. From φ2 = φ2 Â(1), we have
 
Ae Ao
[ φ2e φ2o ] = [ φ2e φ2o ]
Ao Ae
Vector Markov Chains Analysis 231

from which we have

φ2e = φ2e Ae + φ2o Ao and φ2o = φ2e Ao + φ2o Ae .
Now suppose we set φ2e = φ2o = 12 φ. We then have
1 1 1 1
φ = φAe + φAo = φ [Ae + Ao ] .
2 2 2 2
But, Ae + Ao = A(1), so we have φA(1) = φ and φe = 1 because φ is
the stationary vector of A(1). Therefore, if we set φ2 = 12 [ φ φ ], then
φ2 satisfies φ2 = φ2 Â(1), φ2 e = 1. That is, φ2 is the unique stationary
vector of Â(1).
(b) Given  P∞ P∞
A z i i
Â(z) = P∞i=0 2i i i=0 A2i+1 z ,
∞ i
i=0 2i−1 z
A
P
i=0 A2i z
we have
∞ P∞
d iA z (i−1) (i−1) 
 P
iA z
Â(z) = P∞i=0 2i (i−1) i=0
P∞ 2i+1
(i−1) ,
dz i=1 iA2i−1 z i=0 iA2i z

So that ∞ P∞
d
 P 
iA i=0 iA2i+1 ,
Â(z)kz=1 = P∞i=0 2i P ∞
dz i=1 iA2i−1 i=0 iA2i
Hence,
1 P∞ 1 P∞ P∞
d i=0 2iA2i i=0 (2i + 1)A2i+1 − 12 i=0 A2i+1
  
Â(z)

= 2 2
1 P∞ P∞ 1 P∞
dz z=1
2 i=1 (2i − 1)A2i−1 + 21 i=1 A2i−1 2 i=0 2iA2i

If we now postmultiply by e, we find

P∞ P∞
d 1
i=0 2iA2i e +
1
(2i + 1)A2i+1 e − 21 ∞
P
i=0 A2i+1 e
  
2 2 i=0P
Â(z) e =

1 P∞
i=1 iA2i−1 e + 21 ∞ 1
i=1 A2i−1 e + 2
∞
P
dz z=1
2 i=0 2iA2i e

Then, upon changing the index of summation, we find

1 P∞ 1 P∞
d iAi e − A2i+1 e
  
Â(z) e = 2 Pi=0 2 i=0

1 ∞ 1 P∞
dz z=1
2 i=0 iAi e + 2 i=1 A2i−1 e
1
Then premultiplying by φ2 = 2 [ φ φ ] yields
∞
d  1 X
φ2 Â(z) e = φ iAi e.

dz z=1 2 i=0

Therefore,
∞
X d
φ iAi e = 2φ2 Â(z)kz=1 e = 2ρ.
i=0
dz
232 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

7.1 Supplementary Problems

7-1 Consider a slotted time, single server queueing system having unit service
rate. Suppose the arrival process to the system is a Poisson process mod-
ulated by a discrete time Markov chain on {0, 1} that has the following
one-step transition probability matrix:
1−β
 
β
P= .
1−α α
While the arrival process is in phase i, arrivals to the system occur accord-
ing to a Poisson process with rate λi , i = 0, 1.

(a) Argue that A(z) = B(z) for this system.

(b) Show that
1−β e−λ0 (1−z)
  
β 0
A(z) = .
1−α α 0 e−λ1 (1−z)

(c) Determine Ai for all i ≥ 0.

(d) Suppose that at the end of a time slot, the system occupancy level is
zero and the phase of the arrival process is i, i = 0, 1. Determine the
probability that the system occupancy level at the end of the following
time slot will be k, k = {0, 1, · · ·} and the phase of the arrival process
will be j, j = 0, 1.
(e) Suppose that at the end of a time slot, the system occupancy level is
k > 0 and the phase of the arrival process is i, i = 0, 1. Determine the
probability that the system occupancy level at the end of the following
time slot will be l, l = {k − 1, k, k + 1} and the phase of the arrival
process will be j, j = 0, 1.
(f) Let α = 0.5, β = 0.75, λ0 = 1.2, and λ1 = 0.3. Determine the equi-
librium probability vector for the phase process and ρ for the system.
(g) Write a computer program to determine K(1) for the parameters given
in part (f), and then determine the equilibrium probability vector for
the Markov chain for which K(1) is the one-step transition probability
matrix.
(h) Compute E[q̃], the expected number of packets in the system at the
end of an arbitrary time slot. Compare the result to the equivalent
mean value E[q̃] for the system in which λ0 = λ1 = ρ as computed
in part (f). What can be said about the effect of burstiness on the
average system occupancy?

Solution:
Vector Markov Chains Analysis 233

(a) The matrices Bj represent the number of arrivals that occur in the first
slot of a busy period while the matrices Aj represent the number of
arrivals that occur in an arbitrary slot. Since arrivals are Poisson and
the slots are of fixed length, there is no difference between the two.
(b) The matrix A(z) represents the generating function for the number
of arrivals that occur during a time slot. In particular, if the phase of
the phase process is 0 at the end of a given time slot, then the phase
will be 0 at the end of the next time slot with probability β, and the
distribution of the number of arrivals will be Poisson with parameter
λ0 . Similarly, if the phase is initially 1, then it will be zero with
probability (1 − α) and the distribution of the number of arrivals will
be Poisson with parameter λ0 . Hence
βe−λ0 (1−z) (1 − β)e−λ1 (1−z)
 
A(z) = −λ (1−z)
 (1 − α)e
0 (1 − β)e−λ1 (1−z) 
  −λ
β 1−β e 0 (1−z) 0
= .
1−α α 0 e−λ1 (1−z)

(c) From the formula for the Poisson distribution,

 " λi0 #
1−β 0

β i!
Ai = λi1
.
1−α α 0 i!

(d)

λk0
P {ñ = k, ℘˜ = 0|ñ = 0, ℘˜ = 0} = β
k!
λk1
P {ñ = k, ℘˜ = 1|ñ = 0, ℘˜ = 0} = (1 − β)
k!
λk0
P {ñ = k, ℘˜ = 0|ñ = 0, ℘˜ = 0} = (1 − α)
k!
λk1
P {ñ = k, ℘˜ = 1|ñ = 0, ℘˜ = 1} = α
k!

(e) The level wil be k − 1 if there are no arrivals, this probability matrix
is given by
1−β e−λ0
  
β 0
A0 = .
1−α α 0 e−λ1
Now, l = k if exactly one arrival given by
1−β λ0 e−λ0
  
β 0
A1 = ,
1−α α 0 λ1 e−λ1
234 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

and l = k + 1 if exactly two arrivals given by

 " λ20 e−λ0 #
1−β 0

β 2!
A2 = λ21 e−λ1
1−α α 0 2!

(f) The equilibrium distribution for the phase process is found by solving
1−β
 
β
x=x ,
1−α α
where xe = 1. Also, we can simply take ratios of the mean time in
each phase to the mean of the cycle time. Let ñi be the number of
slots in phase i, i = 0, 1. Thus, P {ñ0 = k} = (1 − β)β k−1 implies
1 1
E[ñ0 ] = 1−β . Similarly, E[ñ1 ] = 1−α . Therefore,

1/(1 − β)
P {phase 0} =
1/(1 − β) + 1/(1 − α)
1−α
=
2−α−β
0.5 2
= = ,
2 − 0.5 − 0.75 3
and
1/(1 − α)
P {phase 1} =
1/(1 − β) + 1/(1 − α)
1−β
=
2−α−β
0.25 1
= = .
2 − 0.5 − 0.75 3
Thus,
2 1
ρ = λ0 + λ1
3 3
2 1
= (1.2) + (0.3) = 0.9
3 3
(g) Since A(z) and B(z) are identical, G(1) and K(1) are also identical.
We may then solve for K(1) by using (5.39). The result is
 
0.486555 0.513445
K(1) = .
0.355679 0.644321
Note that the result is a legitimate one-step transition probability ma-
trix for a discrete parameter Markov chain. The equilibrium may be
Vector Markov Chains Analysis 235

obtained by normalizing any row of the adjoint of I − K(1). The

relevant adjoint, in turn, is
 
0.355679 0.513445
adj I − K(1) = .
0.355679 0.513445
and, if normalized so that the row sums are both unity, the result is

κ = [ 0.409239 0.590761 ] .

(h) We may compute E[q̃] from (5.47). In order to do this, we must first
define each term in the equation. From part g, Fq̃ (1) = [ 0.6667 0.3333 ].
Also,  −λ0 [1−z] 
e 0
Fã (z) =
0 e−λ1 [1−z]
so that, in general,

λj0 e−λ0 [1−z]

" #
(j) 0
Fã (z) = .
0 λj1 e−λ1 [1−z]

In particular,
   
(1) λ0 0 1.2 0
Fã (1) = = .
0 λ1 0 0.3
and
λ20
   
(2) 0 1.44 0
Fã (1) = 2 = .
0 λ1 0 0.09
From (5.41),
h i
(1)
π0 = 1 − Fq̃ (1)Fã (1)e κ
= [ 0.0409239 0.0590761 ] .

All of the terms are now defined. Substituting these into (5.41) yeilds
E[q̃] = 6.319. For the M/G/1 case, the appropriate formula is (6.10),
which is
ρ Cx̃2 + 1 
 
E[ñ] = ρ
1 +
 
 
1−ρ 2

= 4.95.

From this, it can be seen that the effect of burstiness is to increase the
average system occupancy. As has been observed many times before,
236 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

a Poisson assumption is not valid in many cases where an observed

phenomena shows that the traffic process is not smooth.

7-2 The objective of this problem is to develop a closed-form expression for

the mean queue length, E[q̃], for the slotted M/D/1 system having phase-
dependent arrivals and unit service times. Our point of departure is the
expression for the generating function of the occupancy distribution as
given in (7.37), which is now repeated for continuity:

Fq̃ (z) [Iz − PFã (z)] = π0 [z − 1] PFã (z). (7.69)

(a) Differentiate both sides of (7.69) to obtain

h i
(1) (1)
Fq̃ (z) Iz − PFã (z) + Fq̃ (z) [Iz − PFã (z)]
(1)
= π0 [z − 1]PFã (z) + π0 PFã (z) (7.70)

(b) Take limits of both sides of (7.70) as z → 1 to obtain

h i
(1) (1)
Fq̃ (1) [I − P] = π0 P − Fq̃ (1) I − PFã (1) . (7.71)

(c) Define ρ to be the marginal probability that the system is not empty
at the end of a slot. Postmultiply both sides of (7.71) by e, and show
(1)
that ρ = Fq̃ (1)Fã (1)e.
(1) (1)
(d) Add Fq̃ (1)eFq̃ (1) to both sides of (7.71), solve for Fq̃ (1), and
(1)
then postmultiply by PFã (1)e to obtain
(1) (1)
Fq̃ (1)PFq̃ (1) = Fq̃ (1)eρ
(1)
+{π0 P − Fq̃ (1)[I − Fã (1)]} (7.72)
[I − P + eFq̃ (1)]−1 .

Use the fact that eFq̃ (1) [I − P + eFq̃ (1)] = eFq̃ (1), as shown in
Exercise 7.12 in Section 7.3.
(e) Differentiate both sides of (7.70) with respect to z, postmultiply both
sides by e, take limits on both sides as z → 1, and then rearrange
terms to find
(1) (1) (1) 1 (1) (1)
Fq̃ (1)PFã (1)e = Fq̃ (1)e − Fq̃ (1)Fã (1)e
2
(1)
−π0 PFã (1)e (7.73)
Vector Markov Chains Analysis 237
(1)
(f) Equate right-hand sides of (7.72) and (7.73), and then solve for Fq̃ (1)e
to obtain
(1)
E[q̃] = Fq̃ (1)e
1 n1 (2) (1)
= Fq̃ (1)Fã (1)e + π0 PFã (1)e
1−ρ 2
 h i
(1)
+ π0 P − Fq̃ (1) I − Fã (1)
o
(1)
× [I − P + eFq̃ (1)]−1 PFã (1)e .

Solution:
(a) This part is accomplished by simply using the formula
d d d
   
u(z)v(z) = u(z) v(z) + u(z) v(z)
dz dz dz
with u(z) = Fq̃ (z) and v(z) = [Iz − PFã (z)].
(b) Given
h i
(1) (1)
Fq̃ (z) Iz − PFã (z) + Fq̃ (z) [Iz − PFã (z)]
(1)
= π0 [z − 1]PFã (z) + π0 PFã (z) (7.70)
from Part (a) and using the fact that limz→1 Fã (z) = I, the result
follows.
(c) From Part(b),
h i
(1) (1)
Fq̃ (1) [I − P] = π0 P − Fq̃ (1) I − PFã (1) . (7.71)

Now, Ie = e, and Pe = e because P is the one-step transition

probability matrix for a DPMC. The [I − P] e = 0. Considering the
RHS, π0 Pe = π0 e, which is the probability the occupancy is equal
to zero. Also, Fq̃ (1)Ie = Fq̃ (1)e = 1 because Fq̃ (1) is a vector of
probability masses covering all possible states. Considering the term
(1)
Fq̃ (1)PFã (1)e, Fq̃ (1)P = Fq̃ (1) because Fq̃ (1) is the stationary
vector for P; this fact can easily be seen from (7.69) by setting z = 1.
In summary, we have
h i
(1) (1)
0 = π0 e− 1 − Fq̃ (1)Fã (1)e or 1−π0 e = ρ = Fq̃ (1)Fã (1)e.

(d) Starting with

h i
(1) (1)
Fq̃ (1) [I − P + eFq̃ (1)] = π0 P −Fq̃ (1) I − PFã (1) . (7.71)
238 QUEUEING THEORY WITH APPLICATIONS . . . : SOLUTION MANUAL

(1)
we add Fq̃ (1)eFq̃ (1) to both sides to get

(1)
Fq̃ (1) [I − P + eFq̃ (1)]
h i
(1) (1)
= Fq̃ (1)eFq̃ (1) + π0 P − Fq̃ (1) I − PFã (1) .

We now postmultiply both sides by [I − P + eFq̃ (1)]−1 to obtain

(1) (1)
Fq̃ (1) = Fq̃ (1)eFq̃ (1) [I − P + eFq̃ (1)]−1
n h io
(1)
+ π0 P + Fq̃ (1) I − PFã (1)
[I − P + eFq̃ (1)]−1 .

We then note that since

eFq̃ (1) [I − P + eFq̃ (1)] = eFq̃ (1),

it is also true that

eFq̃ (1) = eFq̃ (1) [I − P + eFq̃ (1)] [I − P + eFq̃ (1)]−1 .

In addition, Fq̃ (1)P = Fq̃ (1) because Fq̃ (1) is the stationary vector
of P. Hence,
(1) (1)
Fq̃ (1) = Fq̃ (1)eFq̃ (1)
n h io
(1)
+ π0 P + Fq̃ (1) I − Fã (1)
[I − P + eFq̃ (1)]−1 .
(1)
Upon multiplying both sides by PFã (1)e, we have
(1) (1) (1) (1)
Fq̃ (1)PFã (1)e = Fq̃ (1)eFq̃ (1)PFã (1)e
n h io
(1)
+ π0 P + Fq̃ (1) I − Fã (1)
(1)
[I − P + eFq̃ (1)]−1 PFã (1)e.
(1)
Again using Fq̃ (1)P = Fq̃ (1) and ρ = Fq̃ (1)Fã (1)e leads to
(1) (1) (1)
Fq̃ (1)PFã (1)e = Fq̃ (1)eρ
n h io
(1)
+ π0 P + Fq̃ (1) I − Fã (1)
(1)
[I − P + eFq̃ (1)]−1 PFã (1)e.
Vector Markov Chains Analysis 239

(e) Upon taking the derivative of both sides of (7.70), we find

h i
(2) (1) (1)
Fq̃ (z) [Iz − PFã (z)] + 2Fq̃ (z) I − PFã (z)
(2) (2) (1)
−Fq̃ (z)PFã (z) = π0 [z − 1]PFã (z) + 2π0 PFã (z)

Then multiplying by e and taking the limit as z → 1 yields

h i
(1) (1) (2) (1)
2Fq̃ (1) I − PFã (1) e − Fq̃ (1)PFã (1)e = +2π0 PFã (1),

(2)
where we have used the fact that Fq̃ (1) [I − PFã (1)] e = 0. Upon
(1) (1)
solving for Fq̃ (1)PFã (1)e, we find

(1) (1) (1) 1 (2) (1)

Fq̃ (1)PFã (1)e = Fq̃ (1)e − Fq̃ (1)Fã (1)e − π0 PFã (1)e.
2
(f) The final result is obtained by simple algebraic operations. These
operations do not involve any other equations. We begin with

(1) 1 (2) (1) (1)

Fq̃ (1)e − Fq̃ (1)Fã (1)e − π0 PFã (1) = Fq̃ (1)eρ
n 2 h io
(1) (1)
+ π0 P + Fq̃ (1) I − Fã (1) [I − P + eFq̃ (1)]−1 PFã (1)e.

This yields

(1) 1 (2) (1)

Fq̃ (1)e(1 − ρ) = Fq̃ (1)Fã (1)e + π0 PFã (1)
n 2h io
(1) (1)
+ π0 P + Fq̃ (1) I − Fã (1) [I − P + eFq̃ (1)]−1 PFã (1)e.

The result follows by dividing both sides by (1 − ρ).