European Journal of Operational Research 278 (2019) 686–698

Contents lists available at ScienceDirect

European Journal of Operational Research

journal homepage: www.elsevier.com/locate/ejor

Decision Support

Revenue maximization in service systems with heterogeneous

customers
Tejas Bodas a,∗, D. Manjunath b
a
LAAS CNRS, 7 Avenue du Colonel Roche, Toulouse 31400, France
b
Department of Electrical Engg, IIT Bombay, Powai, Mumbai 400076, India

a r t i c l e i n f o a b s t r a c t

Article history: In this paper, we consider revenue maximization problems for service systems with heterogeneous cus-
Received 5 March 2018 tomers. To extract meaningful insights, we consider a simplified two queue system where each server
Accepted 25 April 2019
charges an admission price to its customers. The customers differ in their cost for unit delay and this
Available online 3 May 2019
customer heterogeneity is modeled as a random variable with distribution F. On arrival, the customers
Keywords: must choose a server that offers the lowest sum of admission price and the expected delay cost.
Pricing Given the admission prices at the two servers and the explicit knowledge of the distribution F, we
Revenue maximization first characterize the Wardrop equilibrium routing for the heterogeneous customers. We then consider a
Wardrop equilibrium monopoly system where both the servers belong to a single operator. For this system we characterize the
Duopoly revenue maximizing program for the monopoly operator and outline its properties. We then consider the
Queueing system duopoly problem where each server competes with the other server to maximize its revenue rate. For
the special case when the servers have identical service capacity, we obtain the necessary condition for
existence of symmetric Nash equilibrium prices.
For the analysis of the monopoly and the duopoly problem, the heterogeneity distribution F is as-
sumed to be explicitly known. However, for most practical scenarios, the functional form of F may not be
known to the system operator and in such cases, the revenue maximizing prices cannot be determined. In
the last part of the paper, we provide an elementary method to estimate the distribution F. This method
relies on suitably varying the admission prices based on the properties of the Wardrop equilibrium. We
also illustrate this method with some suitable numerical examples.
© 2019 Elsevier B.V. All rights reserved.

1. Introduction
In many service systems, the quality of service received is char-
acterized by the queueing delay that is experienced by the cus-
tomers in the system. Examples of such service systems include
road and transport systems, health-care systems, computer sys-
tems, call centers and communications systems. The customers of
such systems are usually sensitive to the service quality or the de-
lay experienced in these systems. Further they have non-identical
preferences to the delay experienced. It is often beneficial for
the service system to account for these heterogeneous customer
preferences in any optimization concerning the use of system re-
sources. In this paper, we consider the problem of exploiting this
heterogeneous nature of customer preferences for revenue maxi-
mization in parallel server systems.
∗ in the priority queue which in turn depends on the choice of ser-
Corresponding author.
E-mail addresses: tejaspbodas@gmail.com (T. Bodas), dmanju@ee.iitb.ac.in
(D. Manjunath).
Many service systems have emerged that exploit the heteroge-
neous nature of customer preferences and use it to their advan-
tage. As an example, airlines have a separate priority queue which
is available to delay sensitive customers after payment of an ad-
ditional fee. Some airports even offer priority security checks for
passengers willing to pay an additional fee. Passport or visa issu-
ing agencies all over the world offer the facility of quick processing
of their passport or visa with a payment of an additional fee. Sim-
ilarly, amusement parks such as Disneyland or Universal Studios
offer a priority queue for customers willing to pay a higher ad-
mission price. In service systems with such a price priority struc-
ture, the users of the system are faced with a conflicting choice
of whether to pay the additional fee or not. Typically one could
expect that customers with higher sensitivity or cost for unit de-
lay would have a higher willingness to pay the extra fee. However,
this choice for a customer is also influenced by the congestion level
vice by the other customers based on their cost for unit delay thus
complicating the matters.

https://doi.org/10.1016/j.ejor.2019.04.041
0377-2217/© 2019 Elsevier B.V. All rights reserved.
 T. Bodas and D. Manjunath / European Journal of Operational Research 278 (2019) 686–698
Similarly, the firm or the service system offering a separate pri-
ority queue is faced with the problem of determining the admis-
sion fee for this queue. Very often (as in case of amusement parks
or airline security queues), the objective of the firm is to maxi-
mize the revenue generated. Again, the revenue depends on the
rate of users choosing to make this extra payment which again de-
pends on the users cost for unit delay. A similar question arises
in case of two competing firms (such as two competing amuse-
ment parks) with congestion sensitive customers. A higher admis-
sion price makes a firm more attractive to the congestion sensi-
tive customer but risks loosing revenue to the other firm in case
most of the customers are inherently less sensitive to congestion.
In this paper, we are concerned with the impact of such cus-
tomer heterogeneity not only on the equilibrium choice of queue
for the customers but also on the impact of this heterogenous
delay cost for customers on the revenue maximizing prices for
monopolistic (airport security or passport issuer) and competing
firms.
We answer a variety of questions like the ones discussed above,
by considering a simplified two server system where customers
have heterogeneous cost for unit delay. We assume that each
server has an associated queue for the customers to wait and that
the scheduling discipline at each server is work conserving and
does not discriminate between customers on the basis of their
preference for delay. The servers charge an admission price to ev-
ery customer joining its queue. We assume that the queues are
not observable and only the expected delay as a function of the
arrival rate is available. Further the expected delay at any server is
monotone increasing in the arrival rate of customers to that server.
The customers that use the system are strategic and make an in-
dividually optimal queue-joining decision. The customers differ in
their cost for unit delay which is characterized by a random vari-
able with distribution denoted by F. The cost of service at a server
is the sum of the admission price and the delay cost at that server.
To keep the discussion as simple as possible, we additionally as-
sume that customers do not balk from the system without obtain-
ing service.
As our first revenue maximization problem, we consider a
monopoly and assume that the two servers belong to the same
operator. The objective here is to maximize the total revenue rate,
i.e., the sum of the revenue rate from the two servers. As outlined
earlier, our formulation can be used to analyze monopoly systems
such as government agencies issuing documents (passport, visas,
driving license), amusement parks (in small cities that typically
have no competitors), airport security queues etc. It is not diffi-
cult to see that a revenue maximizing strategy for the monopoly
is to keep both the admission prices as high as possible. This is
because customers cannot balk and are hence required to choose
one of the servers for service. For a meaningful formulation for a
system where customers cannot balk, we assume that the admis-
sion price at one of the servers is a priori fixed to a reasonable
value (for example, in the case of the airport security queue, this
is zero). Further, we assume that each server has enough capac-
ity to handle the entire traffic rate without resulting in a system
breakdown. Our interest is to characterize the revenue maximizing
admission price for different examples of the delay functions at the
queue and for different examples of the delay cost distribution F.
As our second revenue maximization scenario, we assume a
duopoly where each server belongs to separate operators and now
has the objective of maximizing its individual revenue rate. While
the no balking assumption simplifies the analysis to some extent
(without losing on any qualitative properties of the problem), we
observe that the heterogeneity of the customers makes the cal-
culation of the asymmetric price equilibrium difficult. As part of
our main result, we focus on the duopoly problem with identi-
cal servers, and using the properties of the Wardrop equilibrium,
687
identify the necessary condition for existence of symmetric Nash
Equilibrium prices at the two servers.
For both the revenue maximization problems, a crucial assump-
tion is that the optimizer of the system has the complete knowl-
edge of the delay cost distribution for the arriving customers. How-
ever, in most practical scenarios, the distribution F(·) characterizing
the delay cost for a customer may not be completely known to the
system. While the revenue maximizing strategy explicitly depends
on F(·), lack of any knowledge of F(·) makes it difficult to ascer-
tain a revenue optimal admission price. As a simple scenario, con-
sider an amusement park where the arriving customers are of two
types, customers with a high delay cost value and those with a
low delay cost per unit time. Now suppose that the precise values
of the delay costs and the proportion of users of each type are not
know. In such a scenario, lack of essential information on the delay
cost structure would severely affect the ability of the amusement
park to optimize its revenue. In scenarios such as the one outlined
above, and in the absence of any sophisticated delay cost revealing
mechanism, it is imperative for the service system to in some way
estimate the unknown distribution F based on the changes in the
equilibrium arrival rates which is in turn based on the changes in
the admission price. In the last section of this paper, we provide
an elementary method to estimate F(·) by varying the admission
prices and observing the change in the equilibrium traffic routing.
We make use of the properties of the Wardrop equilibrium to iden-
tify a procedure to extract missing information on the distribution
F. This information may prove to be beneficial for service systems
for the underlying revenue maximization.
1.1. Related work
Classical monopoly models have been well studied for the case
of single server queues with homogeneous customers. One of the
first work to analyze such a model is Naor (1969). This model con-
siders a single server queueing system where homogeneous cus-
tomers obtain a reward after service completion. The queue is
observable to arriving customers who choose to either join the
queue or balk. For such a system, the revenue maximizing ad-
mission price was first obtained in Naor (1969). Subsequent works
have analyzed the revenue maximization problem for various mod-
els such as a multiserver queue (Edelson & Hilderbrand, 1972),
GI/M/1 queue (Yechiali, 1971), customers with heterogeneous ser-
vice valuations (Larsen, 1998) and queue length dependent prices
(Chen & Frank, 2001). While the above models assume that the
queue lengths are observable, (Edelson & Hilderbrand, 1975) were
the first to consider the revenue maximization problem for the
case when queues are not observable. See (Bradford, 1996; Chen &
Frank, 2004; Masuda & Whang, 1999; Mendelson, 1985; Mendel-
son & Whang, 1990) for some other single server revenue max-
imization models. The key difference between the above litera-
ture and our model is that we assume customers to have hetero-
geneous preference for the delay experienced. This feature makes
our model more meaningful but at the same time very difficult to
analyze.
The earliest work analyzing the duopoly model with heteroge-
neous customers was by Luski (1976) and Levhari and Luski (1978).
Luski (1976) is interested in knowing whether the revenue maxi-
mizing prices set by the two service systems can be equal. It is
observed that when the parameters of the model are such that
the customers have no incentive to balk, the revenue maximizing
prices set by two identical servers are equal. This is however not
the case when some of the customers prefer to balk. In this case,
the equilibrium revenue maximizing prices are not equal. Levhari
and Luski (1978) provide a numerical analysis for the problem in-
troduced in Luski (1976). Armony and Haviv (2003) analyze this
problem for the case when the customers are from a finite number
688 T. Bodas and D. Manjunath / European Journal of Operational Research 278 (2019) 686–698
of classes and each class has a distinct cost for unit delay. A nu-
merical analysis of the Nash equilibrium admission prices between
the two competing servers is provided. Chen and Wan (2003) con-
sider the revenue maximization in a duopoly with a single cus-
tomer class. The service system is modeled by M/M/1 queues and
the customers are allowed to balk from the system. These assump-
tions on the system model allows them to obtain the sufficient
conditions for the existence of Nash equilibrium. Similar condi-
tions were found in Dube and Jain (2008) who consider an N-
player oligopoly with multiclass customers. The customer classes
differ only in their arrival rates and have the same delay cost per
unit time. A differentiated service model is considered by Dube
and Jain (2010) where each player now operates two types of ser-
vices and each service is used by a dedicated class of customers.
Again, the key result in Dube and Jain (2010) is to obtain suffi-
cient condition for Nash equilibrium prices. Mandjes and Timmer
(2007) consider a duopoly model with two customer classes dif-
fering in their delay cost. The utility of a queue is a decreasing
function of the number of customers using this server. Given the
admission prices, they provide an algorithm that determines the
equilibrium number of customers of each class at the two servers.
While the existence and uniqueness of such a customer equilib-
rium is provided, the existence of Nash equilibrium prices is only
conjectured. Ayesta, Anselmi, and Wierman (2011) consider the
oligopoly pricing game for a single customer class and obtain the
necessary and sufficient conditions on the Nash equilibrium prices
when the queues have identical delay functions. A best-response
algorithm is then provided to numerically obtain Nash equilibrium
prices. Fan, Kumar, and Whinston (2009) consider a price compe-
tition model between service providers offering shrink-wrap soft-
ware (SWS) and software as a service (SaaS). They assume a model
where the customers have heterogeneous sensitivity to the imple-
mentation cost of the software but have identical valuations for
the two software services. Additionally, for the SaaS provider, a
cost for service quality is assumed and a choice of the service
quality rate is part of the strategy (along with the price) for the
SaaS provider. As part of their main result, the Nash equilibrium
prices for the two providers is characterized for a very general set-
ting of the parameters. The primary difference between this model
and that of our paper is that the above model does not involve
any negative congestion effects on customers choosing the service
providers. It is the absence of a congestion function that allows
for explicit expressions for the Nash equilibrium prices. See the
books (Hassin, 2016; Hassin & Haviv, 2003) for a comprehensive
study of revenue maximization problems for queueing systems un-
der Wardrop equilibrium. Also see (Afeche, 2013; Katta & Sethur-
man, 2005; Van Mieghem, 1995) for other related work analyzing
the impact on service systems with heterogeneous customers dif-
fering in their delay costs.
Most of the monopoly and duopoly models described above,
make simplifying assumptions on the customer classes to charac-
terize the underlying Wardrop equilibrium (Wardrop, 1952). Addi-
tional simplification of the analysis is obtained by considering con-
vex or M/M/1 type delay functions at the queues. We do not make
such convexity assumptions in this paper. We utilize the structure
of the Wardrop equilibrium for heterogeneous customers that was
characterized in Bodas, Ganesh, and Manjunath (2014) to analyze
the two problems. This structure on the equilibrium allows us to
provide an equivalent revenue maximization formulation for both
the monopoly and the duopoly that is simpler to analyze.
The rest of the paper is organized as follows. In the next sec-
tion, we shall formalize the notations and provide some prelimi-
naries. We then formulate the revenue maximization problems in
Section 3. In Section 4, we consider the monopoly problem and fi-
nally the duopoly problem in Section 5. Finally in Section 6, we
illustrate a mechanism based on admission pricing to estimate F.
2. Preliminaries
We first introduce the notation to be used throughout. Let cj
denote the admission price at Server j where j = 1, 2. The cus-
tomers arrive according to a homogeneous Poisson process with
rate λ. Let Dj (γ j ) denote the delay function associated with queue j
when the queue arrival rate is γ j , where j = 1, 2. Note that γ1 +
γ2 = λ. We assume that Dj is monotone increasing and continu-
ously differentiable in the interior of its domain with a strictly pos-
itive derivative. Additionally we assume that the cost function at
the two server satisfies the inequalities D3− j (0 ) < D j (λ ) < ∞ for
j = 1, 2.
We associate with each arriving customer a continuous ran-
dom variable β that quantifies a customer’s sensitivity to delay
or congestion. We shall assume that the delay sensitivity β for a
customer is a realization of the random variable β. The customer
arrivals constitute a marked Poisson process of intensity λ × F on
R × R+ . We assume that F is an absolutely continuous cumulative
distribution supported on the interval [a, b] of positive reals. We
additionally assume that F(·) is strictly increasing and hence the
density f(·) satisfies f(x) = 0 for x ∈ [a, b].
A customer with delay cost β entering the system must choose
a queue j so as to minimize c j + β D j (γ j ). Here γ j is determined
through the strategies of all customers. We assume that the quan-
tities γ 1 , γ 2 , Dj (·), F(·) and cj , for j = 1, 2 are part of common
knowledge. We also assume that the customers do not have ac-
cess to current or past queue occupancies, or the history of arrival
times. The strategy of a customer is restricted to choosing a server
according to a fixed probability distribution and such joint strate-
gies are represented by a stochastic kernel, denoted by KW . We in-
terpret KW (β , i) as the probability that a customer with delay sen-
sitivity β chooses queue i at equilibrium. For the two server sys-
tem, the equilibrium kernel KW must satisfy the following Wardrop
equilibrium conditions.
K W (β , i ) ≥ 0 implies ci + β Di (γi ) ≤ c3−i + β D3−i (γ3−i ). (1)
In words, this means that if customers with delay cost β choose
Server i at equilibrium, then the expected cost for this customer
at Server i must be at most the expected cost at Server 3 − i for
i = 1, 2. For a kernel KW , note that the arrival rate of customers to
Server j is given by
 b
γj = λ K W ( β , j )d F ( β ).
β =a
We now provide the following theorem that characterizes the
Wardrop equilibrium kernel for the system. (Proof in Appendix)
Theorem 1. Define δ i as the probability distribution that puts unit
mass on i and suppose that the kernel KW satisfies the Wardrop equi-
librium condition of Eq. (1). Then there exists a threshold β 1 with
β 1 ∈ [a, b] such that
• when c1 > c2 (resp. c1 < c2 ), we have

δ1 (resp. δ2 ) for β ∈ (β1 , b],
K W (β , · ) = (2)
δ2 (resp. δ1 ) for β ∈ [a, β1 ].
Further if β 1 ∈ (a, b) then,
c1 + β1 D1 (γ1 ) = c2 + β1 D2 (γ2 ). (3)
• When c1 = c2 , is not unique and any kernel K with γ1 =
KW
γ + is a valid Wardrop equilibrium kernel KW where γ + :=
{γ1 : D1 (γ1 ) = D2 (γ2 )}. 
Refer Figs. 1 and 2 for a representation of the Wardrop equilib-
rium kernel for the case when c1 > c2 and c1 < c2 respectively. Here
 T. Bodas and D. Manjunath / European Journal of Operational Research 278 (2019) 686–698
Fig. 1. Representation of KW with γ1 ∈ [0, γ + ] for c1 > c2 .
Fig. 2. Representation of KW with γ1 ∈ (γ + , λ] for c1 < c2 .
f(·) denotes the underlying density function of the random variable
β while the shaded region identifies the delay cost parameter of
those customers that choose Server 1.
3. Problem formulation
Having characterized the Wardrop equilibrium kernel KW , we
will now formulate the revenue maximization problems for the
monopoly and the duopoly model.
3.1. Monopoly
Let Rj (cj , γ j ) := cj γ j denote the revenue rate at Server j when
the equilibrium arrival rate of customers under kernel KW is γ j for
j = 1, 2. For the monopoly model, let RT (c1 , γ 1 ) denote the revenue
rate for the monopoly service system. Since γ2 = λ − γ1 , it suffices
to express the revenue rate as a function of only γ 1 . Denoting the
difference c1 − c2 by , we have
RT (c1 , γ1 ) := c1 γ1 + c2 γ2 = c2 λ + γ1 .
Note from Theorem 1, that the argument γ 1 is determined by
the kernel KW which in turn depends on the admission prices c1
and c2 . As noted in the introduction, due to the no balking as-
sumption, we additionally assume that the admission price c2 is
apriori fixed. The dependence of the arrival rates on admission
689
price will be made explicit by writing γ 1 as γ 1 (c1 , c2 ) and the rev-
enue optimization problem for the monopoly can now be stated as
follows:
max RT (c1 , γ1 (c1 , c2 )) = c2 λ + γ1 (c1 , c2 )
c1
subject to 0 ≤ c1 ≤ c 1 (P1 )
where c1 is defined as c1 := inf {c : γ1 (c, c2 ) = 0} to ensure a com-
pact domain. Note that our assumption Dj (λ) < ∞ for j = 1, 2 im-
plies that each server has enough capacity to handle the entire
traffic on its own and this ensures that c1 exists. To be able to
solve program P1 using standard optimization techniques, a closed
form expression for γ 1 (c1 , c2 ) is desired. When c1 > c2 , and β 1 ∈ (a,
b), from Theorem 1 it can be seen that
γ1 (c1 , c2 ) = λ(1 − F (β1 )) (4)
where
β1 = {β : c1 + β D1 (λ(1 − F (β ))) = c2 + β D2 (λF (β ))}.
A similar condition follows when c1 < c2 and it can be seen that
an explicit expression for γ 1 (c1 , c2 ) is difficult. Note that we have
not assumed any functional form for Dj and F(·) and for certain
choice of these functions, a closed form expression for γ 1 (c1 , c2 )
is not even possible. Without an analytic expression for γ 1 (c1 , c2 ),
it is difficult to solve the revenue maximization problem. There-
fore we require an alternative approach to solve program P1. One
possible alternative is to let the equilibrium γ 1 (the value of γ 1 at
equilibrium) be the optimization variable and represent other vari-
ables of the system such as c1 , c2 , β 1 as a function of γ 1 . With
slight abuse of notation, we will use cj (γ j ) to denote the admis-
sion price at Server j when the equilibrium arrival rate to Server j
at equilibrium is γ j where j = 1, 2. Similarly, we use β 1 (γ 1 ) to
represent the threshold β 1 corresponding to an equilibrium ar-
rival rate of γ 1 . Note that c1 (γ 1 ) is also a function of c2 since
γ 1 depends on the difference  and not on their individual val-
ues (from Theorem 1 (Eq. (3)). Therefore within the foregoing as-
sumptions on the delay function, for a given c2 and γ 1 ∈ (0, λ)
one can uniquely determine c1 using Eq. (3). We have suppressed
this dependence on c2 to simplify notation and for the monopoly
model c2 (γ2 ) = c2 as c2 is assumed fixed. Further using the def-
inition g1 (γ1 ) := c1 (γ1 ) − c2 , the equivalent revenue optimization
problem for the monopoly is as follows:
max RT (c1 (γ1 ), γ1 ) = c2 λ + g1 (γ1 )γ1
γ1 (P2 )
subject to 0 ≤ γ1 ≤ γ 1 (c2 ) ≤ λ
where γ 1 (c2 ) determines the domain for the feasible values of γ 1
as a function of c2 and g1 (γ 1 ) represents  = c1 − c2 as a function
of γ 1 . γ 1 (c2 ) and g1 (γ 1 ) are characterized formally in Section 4.
Remark 1. While we assume that the admission price c2 for the
monopoly problem is a-priori fixed, its role in the optimization
problem P2 appears distinctly at two places, namely as a constant
term c2 λ in the objective function and in the term γ 1 (c2 ) that de-
fines the domain of the optimization variable γ 1 . As we will see in
Section 4, this domain can be refined further to eliminate the de-
pendence on c2 and the optimization problem (ignoring the con-
stant c2 λ) can be expressed in terms of the price difference func-
tion g1 (γ 1 ) (See program P5). This reassures the intuition that the
optimization problem depends only on the difference in the ad-
mission prices rather than their exact values.
3.2. Duopoly
Now consider the duopoly market with two competing servers
charging admission prices c1 and c2 to their arriving customers.
The objective of Server j is to choose an admission price cj that
690 T. Bodas and D. Manjunath / European Journal of Operational Research 278 (2019) 686–698
maximizes its revenue rate Rj . For this duopoly, the revenue opti-
mization problem for Server j is as follows.
max R j (c j , γ j ) = c j γ j (c j , c j − )
cj
subject to 0 ≤ cj ≤ cj (P3 )
given c j−
where c j− represents the admission price at the server other than
j, i.e., c1− = c2 and c2− = c1 . For the duopoly market, the aim is to
obtain the Nash equilibrium set of admission prices to be charged
at the two servers. We shall denote the Nash equilibrium prices
by the tuple (c1∗ , c2∗ ). Using the notion of the best response func-
tion (see Osborne, 2003 for details), (c1∗ , c2∗ ) can be characterized
as follows. Let Bi (ci− ) denote the admission price at Server i that
maximizes the server revenue Ri for a given value of ci− for i = 1, 2.
Clearly, Bi (ci− ) is the maximizer in program P3 and it is easy to see
that
 
B1 (c2 ) := c1 ≥ 0 : c1 γ1 (c1 , c2 ) ≥ c1 γ1 (c1 , c2 )∀c1 ≥ 0
 
B2 (c1 ) := c2 ≥ 0 : c2 γ2 (c1 , c2 ) ≥ c2 γ2 (c1 , c2 )∀c2 ≥ 0
(c1∗ , c2∗ ) = {(c1 , c2 ) : B1 (c2 ) = c1 , B2 (c1 ) = c2 }.
However as argued earlier, the closed form expression for
γ j (c j , c j− ) is not easy to obtain. This makes it difficult to solve
program P3 and obtain the best responses Bi (ci− ) for i = 1, 2. As
a result, obtaining (c1∗ , c2∗ ) is in general not easy. As in the case of
the monopoly program, to obtain (c1∗ , c2∗ ), we need to first refor-
mulate program P3 by letting γ j denote the optimizing variable.
The corresponding optimization problem is as follows:
max R j (c j (γ j ), γ j ) := c j (γ j )γ j
γj
subject to 0 ≤ γ j ≤ γ j (c j − ) ≤ λ
given c j− .
cj (γ j ) can be interpreted as the admission price at Server j that
leads to the equilibrium arrival rate of γ j when the other server
charges c j− . Note again that cj (γ j ) will be a function of c j− but we
do not make this explicit in the notation. Now let γ1∗ (c2 ) denote
the maximizer in program P4 for a given value of c2 . Then the best
response c1 is in fact given by the function c1 (γ1∗ (c2 )). Therefore,
once the function c1 (γ 1 ) is characterized, the best response now
denoted by Bˆ1 (c2 ) satisfies Bˆ1 (c2 ) = c1 (γ1∗ (c2 )). We now have
 
(c1∗ , c2∗ ) = (c1 , c2 ) : Bˆ1 (c2 ) = c1 , Bˆ2 (c1 ) = c2
where Bˆi (ci− ) = ci (γi∗ (ci− )) and as stated earlier, γi∗ (ci− ) is the
maximizer in program P4 for i = 1, 2. It is therefore clear that
(c1∗ , c2∗ ) can be obtained once we have characterized c1 (γ 1 ). We
shall analyze the program P4 in detail in Section 5 and explicitly
characterize the functions cj (γ j ) for j = 1, 2 to be able to obtain
(c1∗ , c2∗ ).
4. Monopoly
In this section, we will analyze the monopoly program P2. To
be able to solve program P2, we need to characterize g1 (γ 1 ). This
procedure is outlined below. From Eq. (3) of Theorem 1, we know
that when β 1 ∈ (a, b), (and hence γ 1 ∈ (0, λ)) we have
 = β1 (D2 (γ2 ) − D1 (γ1 ) ).
Expressing the right hand side of the above equation as a function
of γ 1 and noting that g1 (γ1 ) = c1 (γ1 ) − c2 , we have
g1 (γ1 ) = β1 (γ1 )(D2 (λ − γ1 ) − D1 (γ1 ) )
where β 1 (γ 1 ) represents the threshold β 1 for a kernel that sat- KW
isfies Theorem 1 and corresponds to an equilibrium arrival rate of
γ 1 . Note that g1 (γ 1 ) characterizes the price difference  as a func-
tion of γ 1 . We characterize g1 (γ 1 ) (in Lemma 3) by first character-
izing β 1 (γ 1 ) using Lemmas 1 and 2. For a fixed c2 , we then obtain
γ 1 (c2 ) in Lemma 5 that determines the domain of the optimization
problem. To prove this lemma we need to characterize the unique-
ness of kernel KW for a fixed price difference  (Lemma 4). Finally
we characterize our main monopoly program P5 which only de-
pends on the price difference function g1 (·) and not on the exact
prices at the individual queues.
Recall that we make minimal assumptions on the distribution
F(·) and on the delay cost function Dj (·). For our numerical exam-
ples and also to illustrate the properties of the functions β 1 (·), g1 (·)
and c1 (·), we consider the following examples for F(·) and Dj (·). The
distribution F(·) considered for our numerical examples are
• Uniform distribution over the range [a, b].
• Exponential distribution with mean τ .
Gamma distribution with shape k and scale θ .
•
The delay functions used for numerical examples are
γ
• D j (γ j ) = μj . This corresponds to linear delay.
j
• D j (γ j ) = μ 1−γ and μj > λ. This corresponds to M/M/1 type
j j
delay cost function.
The distribution and the delay cost functions outlined above are
commonly used to model heterogeneous customers and congestion
costs (Refer Bodas et al., 2014; Levhari & Luski, 1978; Luski, 1976).
These examples are not exhaustive and one can assume more com-
plex examples for F and D(·) as long as F is absolutely continuous
and D(·) is monotone increasing in the arrival rate. We now begin
(P4 ) with the following lemma that identifies necessary and sufficient
condition on γ 1 when either c1 ≥ c2 or c1 < c2 .
Lemma 1. γ1 ∈ [0, γ + ] iff c1 ≥ c2 while γ1 ∈ (γ + , λ] iff c1 < c2 .
Proof. See Appendix for proof and refer Figs. 1 and 2 for an illus-
tration of the lemma. 
Next, we express the threshold β 1 of Theorem 1 as a function
of γ 1 . Let β 1 (γ 1 ) denote the value of the threshold β 1 (character-
izing KW ) for a given γ 1 such that γ1 = γ + . We have the following
lemma.
Lemma 2.
  
F −1 λ−λγ1 for 0 ≤ γ1 < γ + ,
β1 (γ1 ) =   (6)
F −1 γλ1 for γ + < γ1 ≤ λ.
where F −1 represents the quantile function or the inverse function of
the distribution F.
Proof. See Appendix for proof. 
Note that β 1 (γ 1 ) is not defined in Lemma 2 when γ1 = γ + .
This is because the Wardrop kernel KW with γ1 = γ + is not unique
and hence need not be characterized by a threshold. To avoid
complications arising due to possibly multiple Wardrop equilibria,
we shall restrict to a threshold type Wardrop equilibrium when
γ1 = γ + (and hence c1 = c2 ). Hence, for c1 = c2 we assume

δ1 for β ∈ (β1 , b],
K (β , 1 ) =
W
(7)
δ2 for β ∈ [a, β1 ].
(5) 
λ−γ +
We can now define β1 (γ + ) = F −1 λ and the modified
β 1 (γ 1 ) is now as follows.
 T. Bodas and D. Manjunath / European Journal of Operational Research 278 (2019) 686–698
Fig. 3. Illustrating β 1 (γ 1 ) when the servers are not identical.
Fig. 4. Illustrating β 1 (γ 1 ) for the case of identical servers.
   Proof. See Appendix for proof.
F −1 λ−λγ1 for 0 ≤ γ1 ≤ γ + ,
β1 (γ1 ) =  
F −1 γλ1 for γ + < γ1 ≤ λ.
Refer Fig. 3 for a numerical evaluation of Eq. (8) for the case
when F(·) is an exponential distribution with τ = 20. The delay
γ
functions are D j (γ j ) = μj where μ1 = 3.3 and μ2 = 4. Fig. 4 cor-
j γ 1 ( c2 ) =
responds to the case when the two servers are identical, i.e., μ1 =
μ2 = 4 .
Remark 2. Recall our assumption that F(·) is absolutely continu-
ous and strictly increasing in its domain. Further, the support is
[a, b] and hence F(·) is a bijective function whose inverse exists.
In fact F −1 (· ) is continuous and strictly increasing in its domain.
Since F −1 (· ) is continuous and strictly increasign in its arguments,
β 1 (γ 1 ) is continuous and strictly decreasing when 0 ≤ γ1 < γ +
and continuous and strictly increasing when γ + < γ1 ≤ λ. How-
ever at γ1 = γ + , β 1 (γ 1 ) is in general not continuous (Refer Fig. 3).
For the case when the servers are identical, i.e., D1 (γ ) = D2 (γ ) =
D(γ ), we see from the definition of γ + that γ + = λ2 . For this case,
it is easy to see that β 1 (γ 1 ) is continuous at γ1 = γ + , (but not
differentiable). See Fig. 4.
Having obtained β 1 (γ 1 ), we shall now analyze g1 (γ 1 ) that was
defined in Eq. (5). We have the following lemma.
Lemma 3. For 0 ≤ γ 1 ≤ λ, g1 (γ 1 ) is continuous and monotonic de-
creasing in γ 1 . Further, g1 (γ + ) = 0.
Proof. See Appendix for proof. 
See Fig. 5 for a numerical evaluation of g1 (γ 1 ) when F(·) is ex-
ponential with τ = 20 and when the servers have a linear delay
with μ1 = 3.3 and μ2 = 4.
As stated earlier, the domain of the optimization program P2
is determined by γ 1 (c2 ) which is characterized in Lemma 5. Be-
691
Fig. 5. Illustrating g1 (γ 1 ) when the servers are not identical.
fore stating Lemma 5, we shall first characterize the uniqueness
of β 1 , and hence the kernel KW , when c1 = c2 . While Theorem 1,
guarantees existence of a β 1 characterizing kernel KW , it does not
guarantee the uniqueness of β 1 and hence the uniqueness of the
kernel KW . This result will be used in the proof of Lemma 5.
Lemma 4. For a given , β 1 characterizing the kernel KW in
Theorem 1 is as follows:

b i f  ≥ g1 (0 ) or  ≤ g1 (λ ),
β1 = (9)
β1 (γˆ ) i f g1 ( λ ) <  < g1 ( 0 )
where γˆ satisfies  = g1 (γˆ ). For a fixed , the equilibrium γ 1 and
the corresponding β 1 are unique and this implies the uniqueness of
KW .

(8)
We now characterize γ 1 (c2 ) in the following lemma.
Lemma 5.

λ f or c2 ≥ −g1 (λ ),
(10)
γ : g1 (γ ) = −c2 f or c2 < −g1 (λ )
In words, when c2 < −g1 (λ ) we have γ 1 ( c2 ) =
{γ : g1 (γ ) = −c2 } and for any c1 ≥ 0, the equilibrium γ 1 ∈ (γ 1 (c2 ),
λ]. However when c2 ≥ −g1 (λ ), we have γ 1 (c2 ) = λ in which case
for suitable choices of c1 , γ 1 can be made to achieve every point
in [0, λ].
Proof. See Appendix for proof. 
The above lemma also implies that, if c2 ≥ −g1 (λ ), then for
any c1 ∈ (0, c2 + g1 (λ )) the equilibrium γ 1 satisfies γ1 = λ. On
the other hand, if the parameters of the system are such that
c2 + g1 (λ ) < 0, then for any set of admission prices c1 at Server 1,
we have γ 1 < γ 1 (c2 ).
Remark 3. From Lemma 5, when c2 < −g1 (λ ), we have γ 1 (c2 ) =
γ such that g1 (γ ) = −c2 . From Lemma 3, we know that g1 (γ ) < 0
for γ > γ + . Hence when c2 ≥ 0, we have γ 1 (c2 ) ≥ γ + with strict
equality when c2 = 0.
Now recall the revenue maximization problem P2. Define γ1∗ as
the optimizer for this program with the revenue maximizing ad-
mission price at server 1 given by c1 (γ1∗ ). Since RT (c1 (γ1 ), γ1 ) =
c2 λ + (g1 (γ1 )γ1 , γ1∗ must be such that g1 (γ1∗ ) > 0. Now note
from Lemma 5 and the subsequent remark that when c2 ≥ 0, we
have γ 1 (c2 ) ≥ γ + and from Lemma 3 we have g1 (γ 1 ) > 0 for
γ1 ∈ (0, γ + ). These two facts together imply that γ1∗ ∈ (0, γ + ).
692 T. Bodas and D. Manjunath / European Journal of Operational Research 278 (2019) 686–698
γ
Fig. 6. RT as a function of γ 1 when D j (γ j ) = μjj
The term c2 λ in RT (c1 (γ 1 ), γ 1 ) is a constant and hence we have
the following equivalent program for the revenue maximization
problem.
max g1 (γ1 )γ1
γ1
subject to 0 ≤ γ1 ≤ γ +
where g1 (γ 1 ) is given by Eq. (5).
Note from Lemma 3 that g1 (·) is a continuous function of its
domain. Program P5 involves maximizing a continuous function
over a compact set and hence a maximizer γ1∗ exists. The origi-
nal monopoly program P1 has been significantly simplified to the
equivalent program P5. Since g1 (γ 1 ) is strictly decreasing (and
hence quasi-convex), g1 (γ 1 )γ 1 is in fact a product of two quasi-
convex functions. (However the product of quasi-convex functions
need not be a quasi-convex function). One can now use standard
non-linear optimization techniques to obtain γ1∗ . To further under-
stand Program P5, we perform a numerical evaluation of g1 (γ 1 )γ 1
under a combination of examples for the distribution functions F
and the delay functions Dj (γ j ). We reiterate that our assumptions
allow us to consider even more complicated examples for F and
D(·). However for convenience, we restrict to the examples that
were outlined earlier.
Example 1. In this example we shall assume that the D j (γ j ) = μj
j
for j = 1, 2. We assume that μ1 = 3.3 and μ2 = 4. Further, the ar-
rival rate λ = 3 and we consider the following three examples for
the distribution F(·). (1) F has a uniform distribution with support
on [2,6]. (2) F has an exponential distribution with mean τ = 4
and (3) F has a Gamma distribution with the scale k and shape
θ parameters 2 and 2 respectively. Note that β with these three
distributions have the same mean. We plot RT (c1 (γ1 ), γ1 ) = c2 λ +
g1 (γ1 )γ1 as a function of γ 1 in Fig. 6 where we assume c2 = 1.
When F has the uniform distribution, γ1∗ = 0.62. The optimal rev-
enue rate RT (γ1∗ ) = 4.306 while the admission price c1 (γ1∗ ) max-
imizing RT is 3.106. The corresponding values for the exponential
distribution are γ1∗ = 0.44, RT (γ1∗ ) = 4.712 and c1 (γ1∗ ) = 4.89 while
the values for the gamma distribution are γ1∗ = 0.51, RT (γ1∗ ) =
4.532 and c1 (γ1∗ ) = 4.
Example 2. In this example, we assume that D j (γ j ) = μ 1−γ where
j j
again μ1 = 3.3 and μ2 = 4. Note that λ < μj for j = 1, 2. The
choice of F(·) is as in the previous example. A plot of RT (c1 (γ 1 ), γ 1 )
as a function of γ 1 is provided in Fig. 7. When F is uniformly dis-
tributed, γ1∗ = 0.48. The optimal revenue rate RT (γ1∗ ) = 3.83 while
the admission price c1 (γ1∗ ) maximizing RT is 2.72. The correspond-
ing values for the exponential distribution are γ1∗ = 0.33, RT (γ1∗ ) =
4.21 and c1 (γ1∗ ) = 4.67 while the values for the gamma distribu-
tion are γ1∗ = 0.38, RT (γ1∗ ) = 4.04 and c1 (γ1∗ ) = 3.74.
Fig. 7. RT as a function of γ 1 when D j (γ j ) = μ j 1−γ j
We conclude the analysis of the revenue maximization prob-
lem with the following observations made from the two examples
given above.
– We see that for the given examples of F, RT (c1 (γ 1 ), γ 1 ) is a
unimodal function in γ 1 . For the three distribution functions, it
(P5 )
can be shown that F −1 (· ) is differentiable in its arguments. For
such distribution functions with differentiable F −1 (· ), this implies
that g1 (γ 1 ) and hence RT (c1 (γ 1 ), γ 1 ) is differentiable in γ 1 when
0 < γ1 < γ + . From Rolle’s Theorem (Theorem 10.2.7 Tao, 2006),
this implies that there exists a γ1 ∈ (0, γ + ) such that dγT = 0. A
dR
1
γ1∗ satisfying this equation is the revenue maximizing arrival rate
to server 1. The admission price corresponding to this γ1∗ can now
be obtained using Eq. (5).
– For each of the three distributions, note that we have E β =
4. However, the Revenue rate RT as a function of γ 1 is distinct in
all the three cases. This implies that the revenue rate RT depends
on the higher moments of the distribution F and not just on its
mean value.
– Finally, note that RT depends on admission price through the
addition factor of c2 λ. For different values of c2 , the corresponding
γ1∗ does not change. However it is easy to see from Eq. (5) that
c1 (γ 1 ) increases linearly in c2 .
γ
5. Duopoly
In this section, we shall consider program P4 for revenue max-
imization in the duopoly system. Much of the analysis in this sec-
tion follows from that of the previous section. Let γ j , j = 1, 2 de-
note the optimization variable and represent the admission prices
at the respective servers as a function of the arrival rates. Towards
this, we continue with the use of the notation cj (γ j ) for j = 1, 2.
Note that while in the monopoly case, the admission price c2 was
considered fixed, in the duopoly of this section, it is the strategy
for the second server and hence will not be a constant. The rev-
enue function for Server j is given by
R j ( c j ( γ j ), γ j ) = c j ( γ j )γ j
where cj (γ j ) represents the admission price at Server j resulting in
an equilibrium arrival rate of γ j . As noted in the previous section,
cj (γ j ) is a function of c j− , the admission price at the other server.
For a fixed strategy c2 at Server 2, the revenue function R1 (c1 (γ 1 ),
γ 1 ) can be redefined as
R1 (c1 (γ1 ), γ1 ) = (g1 (γ1 ) + c2 )γ1 . (11)
It can be argued as in the previous section that for a fixed c1
R2 (c2 (γ2 ), γ2 ) = (g2 (γ2 ) + c1 )γ2 (12)
 T. Bodas and D. Manjunath / European Journal of Operational Research 278 (2019) 686–698 693

where analyze the Nash equilibrium under the restriction that the two
servers are identical i.e., the average delay at any queue is the
g2 (γ2 ) = β1 (λ − γ2 )(D1 (λ − γ2 ) − D2 (γ2 ) ) (13)
same for the same arrival rate. Under this setting, our interest is
and from Eq. (8), β1 (λ − γ2 ) is as follows: to characterize the symmetric Nash equilibrium such that c1∗ = c2∗ .
  
F −1 γλ2 for λ − γ + ≤ γ2 ≤ λ, 5.1. Characterizing a symmetric Nash equilibrium
β1 (λ − γ2 ) = 
−1 λ−γ2
 (14)
F λ for 0 < γ2 < λ − γ + .
In this section, we shall characterize the necessary conditions
It is easy to see that g2 (γ 2 ) is also continuous and strictly de- for the existence of a symmetric Nash equilibrium, i.e., (c1∗ , c2∗ )
creasing in γ 2 . Further, g2 (γ2 ) = 0 when γ2 = λ − γ + . The revenue where c1∗ = c2∗ := c∗ . A natural scenario where such an equilibrium
maximization problem for the duopoly is re-stated as follows: is possible is when the two servers have identical delay functions.
  In this section, we restrict to this case and assume that D j (· ) =
max R j (γ j ) = g j (γ j ) + c j − γj D(· ) for j = 1, 2. As the service systems are identical in their delay
γj
subject to 0 ≤ γ j ≤ γ j (c j − ) ≤ λ (P6 ) characteristics, it is desirable to identify conditions for existence of
given c j− . a symmetric Nash equilibrium. We begin with the following defi-
nition. Define α 1 , α 2 as follows.
For a given c j− , recall that γ j∗ (c j− ) denotes the maximizer of



program P4 and hence of the above program. Also recall that α1 = −γ + dg1γ(1γ1 )

Bˆ j (c j− ) denotes the best response admission price at Server j in
γ1 = γ + (16)
+ dg2 (γ2 )

response to the admission price c j− at the other facility. Then the α2 = −γ γ2

Nash equilibrium set of admission prices, denoted by (c1∗ , c2∗ ), is γ2 = γ +
characterized as follows: Based on these definitions, we have the following lemma. (See
  Appendix for proof.)
(c1∗ , c2∗ ) = (c1 , c2 ) : Bˆ1 (c2 ) = c1 , Bˆ2 (c1 ) = c2 , (15)
Lemma 7. α1 = α2 .
where Bˆ j (c j− ) = g j (γ j∗ (c j− )) + c j− for j = 1, 2.
We begin the analysis for the duopoly problem by first identi- We now have the following theorem, that characterizes the nec-
fying that γ j∗ (c j− ) lies in the interior of the domain. We have the essary condition for a symmetric Nash equilibrium.
following lemma.
Theorem 2. Let c1∗ = c2∗ be a symmetric Nash equilibrium for the
 
γ j∗ (c j− ) / 0, γ j (c duopoly price competition. Then c1∗ = c2∗ = α1 .
Lemma 6. ∈ j− ) .
Proof. Recall that the Nash equilibrium is characterized as
Proof. See Appendix for proof.   
(c1∗ , c2∗ ) = (c1 , c2 ) : Bˆ1 (c2 ) = c1 , Bˆ2 (c1 ) = c2 .
For a given c j− since γ j∗ (c j− ) lies in the interior of the domain,




dR j d2 R j
where Bˆ j (c j− ) = g j (γ j∗ (c j− )) + c j− for j = 1, 2. This implies that
γ j∗ (c j− )
satisfies
dγ j
= 0 and
dγ j2

≤ 0.
γ =γ ∗ γ j =γ j∗ c∗j = g j (γ j∗ (c∗j− )) + c∗j− and since c1∗ = c2∗ , we have g j (γ j∗ (c∗j− )) = 0
j j 
dR d2 R for j = 1, 2. Now from Lemma 3 and symmetry of the servers, this
Define S j (c j− ) := γ j : dγ j = 0, 2j ≤ 0 . Then γ j∗ (c j− ) is ob- implies that γ1∗ (c2∗ ) = γ + and γ2∗ (c1∗ ) = λ − γ + . Since the servers
dγ j j
tained as a solution to the following. are identical, we have γ + = λ2 and hence γ2∗ (c1∗ ) = γ + . Since
γ j∗ (c j− ) is also a solution to program P8, γ j∗ (c j− ) ∈ S(c j− ). From
γ j∗ (c j− ) = arg max R j (γ j ). (P8 )

dR j

γ j ∈S j ( c j − ) the definition of S(c j− ), this implies that dγ j
= 0. Further,
γ j =γ +
From the above discussion, it should be clear that obtaining this implies from the definition of Rj (γ j ) that
the closed form expression for (c1∗ , c2∗ ) satisfying the simultaneous


equations of (15) is, in general, not easy. Note that our analysis dR j
dg j (γ j )


= γ+
+ g j (γ + ) + c∗j− (17)
until now makes minimal assumptions on the distribution func- d γ j γ j =γ + γj γ j =γ +
tion F or on the delay function Dj (·). For certain choices of these
= 0.
functions, it may be difficult to obtain a closed form expression for
γ j∗ (c j− ). The objective function Rj also need not be a concave func- We have g j (γ + ) = 0 and hence from the definition of α j for j =
tion. In that case, a brute force search among all the local maxima 1, 2 we have c∗j− = α j . From Lemma 7, we have α1 = α2 and hence
points needs to be carried out to choose the right γ j∗ (c j− ). c1∗ = c2∗ = α1 . This completes the proof. 
Now recall from the introduction that the duopoly model with Note that the above theorem only provides a necessary condi-
heterogeneous customers has been studied numerically in Levhari tion for the Nash equilibrium pair and we shall soon see that in
and Luski (1978) when the delay function is of the M/M/1 type and fact this condition is not sufficient. We shall now provide a few
the delay cost distributions are Uniform and Pareto. More specifi- examples illustrating the occurrence of symmetric Nash equilibria.
cally, Levhari and Luski (1978) numerically demonstrate the exis- γ
tence and non-existence of Nash equilibrium prices under differ- Example 3. In this example, we assume that D j (γ j ) = μj for j =
j
ent parameter settings. The specific model considered in Levhari 1, 2. Let μ1 = μ2 = 4 while the arrival rate is λ = 3. We suppose
and Luski (1978) is a special case of our model setting since we al- that the distribution F(·) has a uniform distribution with support
low for more general delay functions and delay cost distributions. of [a, b]. We plot R1 (γ 1 ) and c1 (γ 1 ) as a function of γ 1 in Fig. 8.
In this paper, we do not indulge in performing such a numerical The aim of this example is to check whether (c1∗ , c2∗ ) = (α1 , α1 ) is
exercise since we do expect similar reaction curves (best response a symmetric Nash equilibrium. For the set of parameters of this
plots) for other distributions and delay cost functions. On the con- example we have γ + = 1.5 and since
trary, in the rest of this section, we shall explicitly characterize

dg1 (γ1 )

the symmetric Nash equilibrium prices for the system, something α1 = −γ +

which has not obtained until now. In the following subsection we γ1 γ1 = γ +
694 T. Bodas and D. Manjunath / European Journal of Operational Research 278 (2019) 686–698
γ
Fig. 8. R1 and c1 (γ 1 ) when D j (γ j ) = μjj and F(·) is uniform over [2,6]
Fig. 9. R1 and c1 (γ 1 ) when D j (γ j ) = μ j 1−γ j and F(·) is exponential with τ = 4
we have α1 = 3. We now set c2 = α1 = 3. Clearly, for a symmet-
ric Nash equilibrium (c1∗ , c2∗ ) = (α1 , α1 ), γ1∗ (c2 ) = γ + = 1.5 must
hold. It is easy to see from Fig. 8 that R1 (γ 1 ) is indeed maximized
when γ1 = γ + implying that γ1∗ (c2 ) = γ + . Further it can be veri-
fied that (c1 (γ1∗ (c2 )))α1 . Clearly, (c1∗ , c2∗ ) = (α1 , α1 ) is a symmetric
Nash equilibrium for this example.
Example 4. With the help of this example, we will illustrate that
the necessary conditions stated in the previous theorem need not
γ
be sufficient. We shall once again assume that D j (γ j ) = μj where
j
μ1 = μ2 = 4. As for the choice of F(·), we consider an exponen-
tial distribution with τ = 4. A plot of R1 (γ 1 ) and c1 (γ 1 ) as a func-
tion of γ 1 is provided in Fig. 9. For this example we start by set-
ting c2 = α1 . However we observe that the best response γ1∗ (c2 ) =
γ + and hence c1 (γ1∗ (c2 )) = c2 . Both these points γ1∗ (c2 ), c1 (γ1∗ (c2 )
and (γ + , α1 ) are represented in Fig. 9. Clearly, (α1 , α1 ) = (c1∗ , c2∗ )
and therefore the sufficiency conditions differ from the necessary
ones.
6. Estimating distribution F
As outlined in the introduction and also from the preceeding
results, it should now be clear that the knowledge of the distribu-
tion F is pivotal for performing any revenue based optimization of
the system. Therefore lack of any information of F would not be
in the interest of the firm that is trying to maximize its revenue.
Recall the example of the amusement park from the introduction.
Suppose tha the customers arriving to the park are such that their
delay cost parameter is either β h or β l (β h > > β l .) Let λ denote
the total arrival rate and let λh and λl denote the proportion of
customers with the two delay costs. Now suppose that β h , λh , β l ,
λl are not know. In such a scenario, to perform revenue maximiza-
tion, the amusement park should indentify these unknown param-
eters in the system. The purpose of this section is to best iden-
tify the missing information in F as much as possible and by using
only elementary procedures. Using properties of the Wardrop equi-
librium, we offer a procedure that involves using only the admis-
sion prices at the servers. In the rest of this section, we shall first
describe a simple procedure to estimate an underlying continuous
distribution function F. Our proposed method is well suited for a
monopoly system when the single service provider has access to
both the admission prices. We then consider the case when β is
a discrete random variable. In this case, the customers are divided
into finite number of classes differing in their values of β. The aim
is to identify the value of β for the different classes along with the
Poisson arrival rates λi for the classes. Refer Bodas and Manjunath
(2011), Bodas, Ganesh, and Manjunath (2011), Bodas et al. (2014),
Armony and Haviv (2003), Mandjes and Timmer (2007) for some
examples of service systems where such discrete customer classes
are considered.
Throughout this section, we shall make the following assump-
tions. We shall assume that the two servers are modeled as M/M/1
queues with service rates μ1 and μ2 and admission prices c1 and
c2 respectively. With this assumption, we have D j (γ j ) = μ 1−γ . It
j j
goes without saying that our analysis will also hold for any delay
cost Dj (·) that is monotonic and strictly increasing in its arguments.
We assume that once the admission prices c1 and c2 at the servers
are announced and that the Wardrop equilibrium is achieved, each
Server j will accurately determine or measure the equilibrium ar-
rival rate γ j and the mean delay cost Dj (γ j ) for j = 1, 2. Hence the
measured values γ j and Dj (γ j ) and the corresponding quantities at
the Wardrop equilibrium will be assumed to be the same. We also
assume that the total arrival rate of customers to the system de-
noted by λ is known a priori and that c1 > c2 , i.e., the admission
price at the first server is higher than the second. Note that since
the distribution F(·) is unknown, the functions β 1 (·), g1 (·), c1 (·) also
cannot be determined and used for our procedure.
From Lemma 4, we know that the threshold β 1 or equivalently
β 1 (γ 1 ) characterizing the Wardrop equilibrium kernel is unique
and it satisfies the following.
c1 + β1 (γ1 )D1 (γ1 ) = c2 + β1 (γ1 )D2 (λ − γ1 ). (18)
We begin by estimating the distributions F that belongs to a pa-
rameterized family, say for example the exponential distribution.
Let the parameter for the exponential distribution be denoted by
α . When c1 and c2 at the two servers are fixed, the equilibrium
γ 1 and γ 2 at the servers is measured immediately. We choose a
c1 , c2 such that γ j > 0 for j = 1, 2. From this, the mean delay cost
Dj (γ j ) for j = 1, 2 is also calculated. Since all the quantities (ex-
cept β 1 ) in Eq. (3) of Theorem (1) are known, the threshold β 1 can
c −c
be determined as β1 = D (γ 1)−D2 (γ ) . Now increase c1 to c11 where
2 2 1 1
c11 = c1 + δ for δ > 0. This decreases the equilibrium γ 1 to say γ1δ .
Let β1δ denote the threshold when the arrival rate to Server 1 is γ1δ .
Again, using the measurements of the arrival rates and the delay
functions β1δ can be determined from Eq. (3). Since γ1δ < γ1 < γ + ,
from Lemma 2, we know that β1 (γ1δ ) > β1 (γ1 ). This implies that
β1δ > β1 . Since the admission prices and the delay function values
at both the queues can be measured, β 1 and β1δ can both be cal-
γ1 −γ1δ
culated from Eq. (18). Clearly, the ratio λ is the probability of
an arriving customer with β ∈ [β1 , β1δ ] and hence
 βδ
1 γ1 − γ1δ
α e−xα dx = . (19)
β1 λ
 T. Bodas and D. Manjunath / European Journal of Operational Research 278 (2019) 686–698 695

Table 1 Table 2
The table indicates the price The estimates z can be
vector (c1 , c2 ), the measured obtained from Eq. (20)
value of γ 1 and the thresh- from the successive
old β obtained from Eq. (3). changes in the admis-
sion prices and the
c1 c2 γ1 β1 corresponding measure-
5.0 5 1.98 2.84 ments of the arrival
5.2 5 1.69 3.04 rates.
5.4 5 1.44 3.20
c1 c1 + δ z
5.6 5 1.23 3.33
5.8 5 1.05 3.44 5 5.2 0.37
6.0 5 0.89 3.53 5.2 5.4 0.39
6.2 5 0.75 3.60 5.4 5.6 0.41
6.4 5 0.63 3.67 5.6 5.8 0.42
6.6 5 0.52 3.37 5.8 6.0 0.44
6.8 5 0.43 3.78 6.0 6.2 0.44
6.2 6.4 0.45
6.4 6.6 0.46
6.6 6.8 0.47
The only unknown quantity is the exponential parameter α which
can now be obtained from the above equation.

Remark 4. Since the exponential distribution has a single param-

eter, the parameter could be obtained using only Eq. (19). For a
parameterized distribution with k parameters, we need k simulta-
neous equations in terms of the underlying parameters. These can
be obtained by following
 the procedure above for k different ad-
mission price c1k at Server 1.

We will now describe a numerical method to obtain a piece-

wise constant approximation for the density function f that is not
necessarily from a parameterized family of distribution functions.
As an example, consider a random variable β supported on the
range [0,4]. Suppose the distribution function is
x2
P ( β ≤ x ) = F (x ) = .
16
Fig. 10. Comparing the estimate of f(·) with the true density function.
The corresponding density function is denoted by f(x) is x/8 for
x ∈ [0, 4]. For this example assume that there are two M/M/1
servers with service rates μ1 = 5 and μ2 = 5, admission prices ini-
A plot comparing the true density function and the estimate is
tially set to c1 = c2 = 5 and the total arrival rate λ = 5. As earlier,
given in Fig. 10. The plot shows that the estimate of the density
we assume that once the admission prices at the servers are an-
function is reasonably accurate and for better estimation, one nat-
nounced, the Wardrop equilibrium is reached instantaneously and
urally required more of such measurement points.
each servers can accurately determine the aggregate arrival rates
There is however a limitation to this method. Note that when
and the mean delay per customer.
c1 = c2 , the corresponding value of β1 = 2.84. Any increase or de-
Increase c1 by δ > 0 and for the admission price vector (c1 +
crease in either c1 or c2 cannot result in a β 1 such that β 1 < 2.84.
δ, c2 ), measure the equilibrium arrival rates and the mean delay This is because, for the underlying distribution we have from
in the queues and calculate the corresponding threshold β 1 using
Eq. (8) that β1 (γ + ) = 2.84 and for any γ ∈ [0, λ] with γ = γ + , we
Eq. (3). Repeat this for a finite number of times, each time increas-
have β1 (γ ) > β1 (γ + ). As a result, the density function f(x) cannot
ing c1 from its previous value by δ . This experiment is denoted in
be estimated for x ≤ 2.84.
Table 1.
Using the earlier notation, we observe from the table that as c1
increases to, say c1 + δ, γ 1 decreases to γ1δ while the threshold β 1 6.1. Estimating discrete distribution
increases (to β1δ ). As earlier, we have
We shall now consider the case where the distribution F is a
 βδ
1 γ1 − γ1δ discrete distribution with M point masses. Thus, there are M cus-
f (x )dx =
β1 λ tomer classes and we will assume that for each Class i, the asso-
ciated waiting cost β i and the arrival rate λi are unknown. Fur-
where the density function f(x) is to be estimated. Assume for all ther, β1 > β2 . . . > βM . See (Bodas et al., 2014) for the analysis of
x ∈ (β1 , β1δ ) that f (x ) = z, where z is a constant. By assuming this, Wardrop equilibrium of such a model. We continue with the as-
we are approximating the density function f(x) for x ∈ (β1 , β1δ ) by sumption that there are two servers each charging an admission
a horizontal line of magnitude z and thus approximating f(x) by price c1 and c2 . We begin by setting c2 = 0 and c1 to an arbitrar-
a piecewise constant function. As δ → 0, the approximation should ily large value such that γ1 = 0 while γ2 = λ. This is represented
converge to the true density function. We now have in part (a) of Fig. 11. It goes without saying that the necessary as-
γ1 − γ1δ sumption is that μ2 > λ. Now start decreasing c1 in steps of size δ
z= (20) and stop at the first instance when γ 1 increases to an arbitrarily
λ(β1δ − β1 ).
small value
. We use the notation c1 and γ1 to denote the ad-
j j

The value of z for a fixed c1 and c1 + δ can be viewed as an esti- mission price and the arrival rate at Server 1 when c1 is decreased
mate for the density function f(x) and obviously z → f(x) as δ → 0. j times by δ , i.e., when c1 = c1 − jδ.γ1 =
implies that the most
j

These values of z for different values of c1 are given in Table 2. sensitive delay class β 1 must now be using Server 1 along with
696 T. Bodas and D. Manjunath / European Journal of Operational Research 278 (2019) 686–698
Fig. 11. Estimating the discrete distribution F
Server 2. Since the delay function at each queue can be measured,
β 1 can be easily determined from the corresponding Wardrop
condition
c1j + β1 D1 (γ1j ) = β1 D2 (γ2j ).
We will now determine λ1 corresponding to this β 1 . Continue de-
creasing c1 . The proportion of Class 1 customers using Server 1
keeps increasing till all Class 1 customers use only Server 1. When
this happens, the corresponding Wardrop equilibrium condition for
some k > j satisfies
 
c1k < β1 D2 (γ2k ) − D1 (γ1k )
and this is represented by part (b) in Fig. 11. For a Class 2 customer
to start using Server 1, the Wardrop equilibrium condition is
c1m = β2 (D2 (γ2m ) − D1 (γ1m ) )
where m > k. Further since m > k, we have
 
β2 (D2 (γ2m ) − D1 (γ1m ) ) < β1 D2 (γ2k ) − D1 (γ1k ) .
and hence for all l such that k < l < m, we have
 
β2 (D2 (γ2m ) − D1 (γ1m ) ) < c1l < β1 D2 (γ2k ) − D1 (γ1k ) .
This means that for any c1l satisfying c1k < c1l < c1m ,γ1l and γ2l re-
main unchanged. Clearly in this case λ1 = γ1l . Fig. 11, part (c) rep-
resents the fact that for any c1 > c1m , Class 2 customers use both
the servers at Wardrop equilibrium. Continue this process till all
the λi , β i as well as the number of customer classes M is deter-
mined. It should be noted that the accuracy of our method in-
creases as δ → 0. A downside of a small δ is that the procedure
may take a very long time to discover the system parameters.
7. Summary
In this paper, we have considered the problem of revenue max-
imization in parallel server systems. We specialize with the case
of two servers and first assume the case when both the servers
belong to the same service provider. The admission price at one
of the servers is required to be fixed and the service system can
change the admission price at the other server to maximize its
revenue. We first characterize the Wardrop equilibrium when cus-
tomers are heterogeneous and strategic. We use this characteri-
zation to simplify the revenue maximization program to make it
more amenable to analysis. The equivalent program is easy to in-
terpret and to analyze and provides more insight into the problem.
While it is intuitive that for a fixed c2 , the revenue maximizing c1
should always be greater than c2 , the program allows us to char-
acterize the revenue maximizing c1∗ as a function of c2 .
In the second part of the paper, we consider the duopoly model
where each server competes with the other one to maximize its
revenue. This is a standard game-theoretic problem and the aim is
to identify the Nash equilibrium set of prices. We see however that
since the customers are heterogeneous, the first order necessary
conditions are not easy to solve. Instead, we characterize this Nash
equilibrium for a simplified case when the two servers are identi-
cal in their delay characteristics. In this case we are interested in
the symmetric Nash equilibrium prices. We provide the necessary
condition for this case and identify the Nash equilibrium prices for
different distributions F and delay cost functions D(·).
In both these revenue maximization problems, an important as-
sumption is that the distribution function F is known. We relax this
assumption in Section 6 and provide a procedure to estimate this
distribution. The proposed method is of course preliminary and as-
sumes that one is allowed to change admission price any number
of time to measure the change in the equilibrium arrival rate. Fur-
ther, we have assumed that there is no cost to making such mea-
surements. A more realistic method incorporating these practical
limitations may make the problem more relevant and this is part
of future work.
Acknowledgments
This work was done when the corresponding author was affil-
iated with the Dept. of Electrical Engineering at IIT Bombay, In-
dia. Both authors would like to acknowledge the support of Bharti
Centre for Communication at IIT Bombay, CEFIPRA and IFCAM. The
corresponding author would also like to acknowledge the support
of the French Agence Nationale de la Recherche (ANR) through the
project ANR- 15-CE25-0 0 04 (ANR JCJC RACON).
Appendix
Theorem 1
Proof. The proof of the first part for the case when c1 > c2 follows
from Corollary 4 in Bodas et al. (2014). The proof for c1 < c2 is
along similar lines and follows by symmetry. We now prove the
second part. Consider the case when c1 = c2 and recall the as-
sumption that D1 (0) < D2 (λ) and D2 (0) < D1 (λ).KW must be such
that D1 (γ1 ) = D2 (γ2 ). To see why this must be true, suppose
that this is not true and let D1 (γ 1 ) = D2 (γ 2 ). Customers from the
queue with a higher delay cost will have an incentive to move to
the queue with a lower delay cost. This implies that a KW with
D1 (γ 1 ) = D2 (γ 2 ) is not at equilibrium. Recall the definition γ + :=
{γ1 : D1 (γ1 ) = D2 (γ2 )}. Since, D1 (0) < D2 (λ) and D2 (0) < D1 (λ), we
have 0 < γ + < λ. Now for any kernel K satisfying γ1 = γ + , since
c1 = c2 , the cost for any customer at the two servers is equal.
Hence there is no incentive for any customer to deviate from its
choice of the server. The Wardrop equilibrium kernel KW though
b
not unique must however satisfy λ β =a K W (β , j )dF (β ) = γ + . 
Lemma 1
Proof. We first prove that γ1 ∈ [0, γ + ] implies c1 ≥ c2 . Recall the
definition of γ + that
γ + = {γ1 : D1 (γ1 ) = D2 (γ2 )}.
Since Dj (γ j ) is monotonic and increasing in γ j for j = 1, 2 and
that γ2 = λ − γ1 we have D1 (γ 1 ) ≤ D2 (γ 2 ) for γ1 ∈ [0, γ + ]. Now let
γ1 = 0. Since no customer uses Server 1 at equilibrium, this im-
plies that c1 + β D1 (0 ) > c2 + β D2 (λ ) for all β . Since D1 (0) < D2 (λ)
(assumption) c1 > c2 must be true.
When γ1 = γ + , we will show that c1 = c2 . Suppose this is not
true, i.e., γ1 = γ + while c1 = c2 . γ1 = γ + implies D1 (γ1 ) = D2 (γ2 ).
As c1 = c2 , customers have an incentive to move from the server
with a higher admission price to the one with a lower price. This
implies that γ1 = γ + is not an equilibrium and this is a contradic-
tion.
Now consider γ1 ∈ (0, γ + ) where D1 (γ 1 ) < D2 (γ 2 ). From
Theorem 1, γ1 ∈ (0, γ + ) implies β 1 ∈ (a, b) and hence c1 +
β1 D1 (γ1 ) = c2 + β1 D2 (γ2 ). Since D1 (γ 1 ) < D2 (γ 2 ) we have c1 ≥ c2 .
 T. Bodas and D. Manjunath / European Journal of Operational Research 278 (2019) 686–698
We now prove that if c1 ≥ c2 , then γ1 ∈ [0, γ + ]. We first show
that when c1 = c2 , we have γ1 = γ + . Suppose that when c1 = c2 ,
γ1 = γ + . From the definition of γ + we have D1 (γ 1 ) = D2 (γ 2 ) and
hence customers have an incentive to move from the server with
higher expected delay to the one with lower expected delay. This
implies that when c1 = c2 , γ1 = γ + is not an equilibrium.
Now let c1 > c2 . From Theorem 1 we have either β1 = a or
β1 = b or β 1 ∈ (a, b). The case β1 = a corresponds to the case
when all customers choose Server 2 at equilibrium and this can-
not happen! This is because while c1 > c2 , we have also assumed
D1 (λ) > D2 (0).KW with β1 = a will be possible only if
c1 − c2 ≤ β (D2 (0 ) − D1 (λ ))
for all β ∈ [a, b]. Now this is not possible as the left hand side
is positive while the right hand side is negative. It is straightfor-
ward to see that when β1 = b, we have γ1 = 0 and hence γ1 ∈
[0, γ + ]. When β 1 ∈ (a, b) we have c1 + β1 D1 (γ1 ) = c2 + β1 D2 (γ2 ).
Again, since c1 > c2 , we have D1 (γ 1 ) ≤ D2 (γ 2 ) and this requires
γ1 ∈ (0, γ + ). The proof for γ1 ∈ (γ + , λ] follows along similar lines
and will not be provided. This completes the proof. 
Lemma 2
Proof. From Lemma 1, γ1 ∈ [0, γ + ) implies that c1 > c2 while γ1 ∈
(γ + , λ] implies c1 < c2 . Now from Theorem 1, when c1 > c2 , we
have
 b
γ1 = λ 1dF (β ) = λ(1 − F (β1 )).
β1
Similarly, when c1 < c2 we have
 β1
γ1 = λ 1dF (β ) = λ(F (β1 )).
0
Now β 1 (γ 1 ) defined as the value of threshold β 1 when the equi-
librium arrival rate to Server 1 is γ 1 can be represented as fol-
lows.
 b
β : β λdF (β ) = γ1 for 0 ≤ γ1 < γ , +
β1 (γ1 ) = β
β : a λdF (β ) = γ1 for γ + < γ1 < λ.
Now as seen earlier, F(·) is absolutely continuous and strictly in-
> 0 such that there exists
creasing in its domain. Further, the support is [a, b] and hence F(·)
− c2 = g1 (γ1 ). Now from the monotonicity
is a bijective function whose inverse exists. In fact F −1 (· ) is con-
tinuous and strictly increasing in its domain. The statement of the
lemma now follows. 
Lemma 3
Proof. Recall our assumption that Dj (γ j ) is continuous and mono-
tone increasing in γ j where j = 1, 2. Since γ2 = λ − γ1 ,(D2 (λ −
γ1 ) − D1 (γ1 )) is monotone decreasing in γ 1 for 0 ≤ γ 1 ≤ λ. Re-
call Eq. (8) that determines β 1 (γ 1 ). For 0 ≤ γ1 < γ + , β 1 (γ 1 ) is
continuous and strictly decreasing. The continuity follows from
(· ). Since F (· ) is strictly increasing in its arguments,
thatof F −1 −1
λ−γ1
F −1 λ = β1 (γ1 ) is decreasing in γ 1 . Clearly, g1 (γ 1 ) is mono-
tone decreasing when γ 1 is such that 0 ≤ γ1 < γ +.
When γ 1 is such that γ + < γ1 ≤ λ, from the definition of γ + ,
we have (D2 (λ − γ1 ) − D1 (γ1 )) < 0. In this range of γ 1 , it can be
seen from Eq. (8) that β 1 (γ 1 ) is continuous and increasing in γ 1 .
This again implies that g1 (γ 1 ) is continuous decreasing when γ 1
satisfies γ + < γ1 ≤ λ.
g1 (γ + ) = 0 follows from the definition of γ + where D1 (γ + ) =
D2 (λ − γ + ). The continuity at γ + is obvious from the fact that
g1 (γ + ) = 0 and limγ1 →γ + g1 (γ1 ) = 0. 
Lemma 4
Proof. Suppose  ≥ g1 (0). From the definition of  and from
Eq. (5), this implies that
> 0 such that setting c1 = c2 +
697
c1 − c2 ≥ b ( D2 ( λ ) − D1 ( 0 ) )
≥ β ( D2 ( λ ) − D1 ( 0 ) )
for all β ∈ [a, b]. From the Wardrop equilibrium condition, this im-
plies that K W (β , · ) = δ2 for β ∈ [a, b]. This implies that γ1 = 0 and
from Eq. (2) we have β1 = b. Similarly when,  ≤ g1 (λ) < 0 we
have
c1 − c2 ≤ b ( D2 ( 0 ) − D1 ( λ ) )
≤ β ( D2 ( 0 ) − D1 ( λ ) )
where β ∈ [a, b]. Again, from the Wardrop equilibrium condition,
this implies that K W (β , · ) = δ1 for β ∈ [a, b]. Hence γ1 = λ and
from Eq. (2), we have β1 = b.
Now suppose g1 (λ) <  < g1 (0) where we know that g1 (0) > 0
and g1 (λ) < 0. From Lemma 3, we know that g1 (γ 1 ) is monoton-
ically decreasing in γ 1 . Therefore there exists a unique γ with
0 < γ < λ such that  = g1 (γ ). This proves the uniqueness of γ 1 .
To see how β1 = β1 (γ ) note that  = g1 (γ ) implies that
c1 − c2 = β1 (γ )(D2 (λ − γ ) − D1 (γ ) ).
Now if γ ≤ γ + we have D2 (λ − γ ) > D1 (γ ). In this case,
c1 − c2 ≤ β ( D2 ( λ − γ ) − D1 ( γ ) )
for β ∈ [a, β 1 (γ )]. This means that K W (β , · ) = δ2 for all β ∈ [a,
β 1 (γ )]. Similarly, we have
c1 − c2 ≥ β ( D2 ( λ − γ ) − D1 ( γ ) ) (22)
and K W (β , · )
= δ1 when β ∈ [β 1 (γ ), b]. Similar arguments hold
when γ > γ + and hence β1 = β1 (γ ) when g1 (0) <  < g1 (λ).
From Theorem 1, KW is characterized by β 1 and for a fixed ,
β 1 is unique. This implies uniqueness of KW . It is important to
mention that KW is unique when  = 0 because of the assump-
tions made to ensure β 1 (γ 1 ) well defined at γ1 = γ + . 
Lemma 5
Proof. Suppose c2 satisfies c2 < −g1 (λ ). Assume that c1 = 0 so
(21)
that we have  > g1 (λ). From Lemma 4 this implies that the equi-
librium γ 1 satisfies g1 (γ1 ) =  = −c2 . Let us label this γ 1 as γˆ .
Now increase c1 from c1 = 0 by a small
γ 1 that satisfies  =
of g1 (·) it is clear that the equilibrium γ 1 is decreasing as  in-
creases. This implies that a higher  caused by increasing c1 will
only lead to a γ 1 satisfying γ1 < γˆ . Clearly, for any choice of c1 ≥ 0,
we have γ1 ∈ / [γˆ , λ] and hence for this case γ 1 (c2 ) = γˆ .
Now suppose that −c2 ≤ g1 (λ ). When c1 = 0, this implies
 ≤ g1 (λ) and from Lemma 4 this implies β1 = b with the corre-
sponding γ 1 satisfying γ1 = λ. As we increase c1 , the equilibrium
γ 1 decreases and hence γ 1 satisfies γ 1 ∈ [0, λ]. The compact rep-
resentation now follows. 
Lemma 6
Proof. To reduce the notations, we represent γ j∗ (c j− ) by γ j∗ in
 
the proof of the lemma. We shall prove that γ1∗ ∈
/ 0, γ 1 (c2 ) and
 
the proof for γ2∗ ∈ / 0, γ 2 (c1 ) is along similar lines. Suppose γ1∗ ∈
{0, λ}. Then from the requirement that γ1∗ = λ − γ2∗ , we have ei-
ther (1) γ1∗ = 0 and γ2∗ = λ or (2) γ1∗ = λ and γ2∗ = 0. First consider
the case when γ1∗ = 0 and γ2∗ = λ. This implies that R1 (c1 (0 ), 0 ) =
0 and hence the revenue made by Server 1 at equilibrium is zero.
Further since this is an equilibrium, there is no incentive for the
server to change the admission price and increase its revenue. We
shall now show that this is not true. From Eq. (5) we know that
for a given c2 , the admission price at Server 1 must be at least
c2 + g1 (0 ) > 0. Now we know that setting c1 = c2 will result in
γ1 = γ + . Now due to the assumption that (1) D1 (0) < D2 (λ) and
(2) D2 (0) < D1 (λ), there exists an

698 T. Bodas and D. Manjunath / European Journal of Operational Research 278 (2019) 686–698

will result in γ1 ∈ (0, γ + ). The revenue earned is non-zero and Bodas, T., Ganesh, A., & Manjunath, D. (2014). Tolls and welfare optimization for mul-
there is clearly an incentive to deviate from any value greater than ticlass traffic in multiqueue systems arXiv:1409.7195.
Bodas, T., & Manjunath, D. (2011). On load balancing equilibria in multiqueue sys-
c2 + g1 (0 ). This implies that γ1∗ = 0 and γ2∗ = λ is not possible. The tems with multiclass traffic. In Proceedings of the NETGCOOP.
proof for γ1∗ = λ and γ2∗ = 0 is along the same lines.  Bradford, R. M. (1996). Incentive compatible pricing and routing policies for multi-
-server queues. European Journal of Operational Research, 89, 226–236.
Lemma 7 Chen, H., & Frank, M. (2001). State dependent pricing with a queue. IIE Transactions,
33(10), 847–860.
Proof. From the definition of g1 (·) in Eq. (5) we have Chen, H., & Frank, M. (2004). Monopoly pricing when customers queue. IIE Transac-
tions, 36(6), 569–581.
dg1 (γ1 ) Chen, H., & Wan, Y. (2003). Price competition of make-to-order firms. IIE Transac-
= β1 (γ1 )(D2 (λ − γ1 ) − D1 (γ1 ) )
γ1 tions, 35(9), 817–832.
  Dube, P., & Jain, R. (2008). N-player Bertrand-Cournot games in queues: Existence
+ β1 (γ1 ) D2 (λ − γ1 ) − D1 (γ1 ) of equilibrium. In Proceedings of the 46th annual allerton conference on commu-
nication, control, and computing, 2008 (pp. 491–498).
where the partial derivatives on the r.h.s. are w.r.t. γ 1 . Now from Dube, P., & Jain, R. (2010). Diffserv pricing games in multi-class queueing network
the definition of γ + and from Eq. (8) we have models. In Proceedings of the international teletraffic congress (ITC-22).

  Edelson, N., & Hilderbrand, D. (1972). Individual and social optimization in a

dg1 (γ1 ) λ−γ+    multiserver queue with a general cost benefit structure. Econometrica, 40,
=F −1
D2 (λ − γ + ) − D1 (γ + ) . 515–528.
γ1 γ1 = γ + λ Edelson, N., & Hilderbrand, D. (1975). Congestion toll for Poisson queuing processes.
Econometrica, 43, 81–92.
Similarly, from the definition of g2 (·) we have Fan, M., Kumar, S., & Whinston, A. B. (2009). Short-term and long-term competition
between providers of shrink-wrap software and software as a service. European
dg2 (γ2 ) Journal of Operational Research, 196(2), 661–671.
= β1 (λ − γ2 )(D1 (λ − γ1 ) − D2 (γ2 ) )
γ2 Hassin, R. (2016). Rational queueing. CRC Press.
  Hassin, R., & Haviv, M. (2003). To queue or not to queue. Kluwer Academic Publishers.
+ β1 (λ − γ2 ) D1 (λ − γ2 ) − D2 (γ2 ) Katta, A. K., & Sethurman, J. (2005). Pricing strategies and service differentiation in
queues—a profit maximisation perspective. Technical Report. Columbia Univer-
where the partial derivatives on the r.h.s are now w.r.t γ 2 . Note sity. CORC TR-2005-4
∂ D1 (γ1 ) ∂ D1 (γ1 ) Larsen, C. (1998). Investigating sensitivity and the impact of information on pricing
that since γ1 = λ − γ2 , we have ∂γ = − ∂γ . Further note decisions in an M/M/1/∞ models. International Journal of Production Economics,
1 2
that since the servers are identical, i.e., D j (· ) = D(· ) for j = 1, 2 365–377.
Levhari, D., & Luski, I. (1978). Duopoly pricing and waiting lines. European Economic
from the definition of γ + we have γ + = λ2 . From Eq. (14) and the Review, 11, 17–35.
fact that λ − γ + = γ + , we have Luski, I. (1976). On partial equilibrium in a queuing system with two servers. The
  Review of Economic Studies, 43, 519–525.
dg2 (γ2 ) λ−γ+    Mandjes, M., & Timmer, J. (2007). A duopoly model with heterogeneous conges-
= F −1 D2 (λ − γ + ) − D1 (γ + ) . tion-sensitive customers. European Journal of Operational Research, 3, 445–467.
γ2 γ2 = γ + λ Masuda, Y., & Whang, S. (1999). Dynamic pricing for network service: equilibrium
and stability. Management Science, 45, 857–869.
This proves that α1 = α2 .  Mendelson, H. (1985). Pricing services: queueing effects. Communications of the ACM,
28(5), 312–321.
References Mendelson, H., & Whang, S. (1990). Optimal incentive-compatible priority pricing
for the M/M/1 queue. Operations Research, 38(5), 870–883.
Afeche, P. (2013). Incentive-compatible revenue management in queueing sys- Naor, P. (1969). The regulation of queue size by levying tolls. Econometrica, 37,
tems: Optimal strategic delay. Manufacturing and Service Operations Manage- 15–24.
ment, 15(3), 423–443. Osborne, M. J. (2003). An introduction to game theory. Oxford University Press, USA.
Armony, M., & Haviv, M. (2003). Price and delay competition between two service Tao, T. (2006). Analysis (Volume 1). Hindustan Book Agency.
providers. European Journal of Operation Research, 147, 32–50. Van Mieghem, J. A. (1995). Dynamic scheduling with convex delay costs: The gen-
Ayesta, U., Anselmi, J., & Wierman, A. (2011). Competition yields efficiency in load eralized c| mu rule. The Annals of Applied Probability, 809–833.
balancing games. In Proceedings of the IFIP performance (pp. 968–1001). Wardrop, J. G. (1952). Some theoretical aspects of road traffic research communica-
Bodas, T., Ganesh, A., & Manjunath, D. (2011). Load balancing and routing games tion networks. Proceedings of the industrial and civil engineering (1, pp. 325–378).
with admission price. In Proceedings of the IEEE conference on decision and Yechiali, U. (1971). On optimal balking rules and toll charges in the GI/M/1 queue.
control. Operations Research, 19, 349–370.