You are on page 1of 26


Articles in Advance, pp. 120

ISSN 0025-1909 (print) ISSN 1526-5501 (online)

Optimal Multiperiod Pricing with Service Guarantees

Christian Borgs
Microsoft Research, New England Lab, Cambridge, Massachusetts 02142,

Ozan Candogan
Fuqua School of Business, Duke University, Durham, North Carolina 27708,

Jennifer Chayes
Microsoft Research, New England Lab, Cambridge, Massachusetts 02142,

Ilan Lobel
Stern School of Business, New York University, New York, New York 10012,

Hamid Nazerzadeh
Marshall School of Business, University of Southern California, Los Angeles, California 90089,

W e study the multiperiod pricing problem of a service firm with capacity levels that vary over time. Cus-
tomers are heterogeneous in their arrival and departure periods as well as valuations, and are fully
strategic with respect to their purchasing decisions. The firms problem is to set a sequence of prices that max-
imizes its revenue while guaranteeing service to all paying customers. We provide a dynamic programming
based algorithm that computes the optimal sequence of prices for this problem in polynomial time. We show
that due to the presence of strategic customers, available service capacity at a time period may bind the price
offered at another time period. This phenomenon leads the firm to utilize the same price in multiple periods,
in effect limiting the number of different prices that the service firm utilizes in optimal price policies. Also,
when customers become more strategic (patient for service), the firm offers higher prices. This may lead to
the underutilization of capacity, lower revenues, and reduced customer welfare. We observe that the firm can
combat this problem if it has an ability, beyond posted prices, to direct customers to different service periods.
Keywords: dynamic pricing; strategic customers; nonlinear programming; dynamic programming
History: Received June 17, 2011; accepted September 3, 2013, by Dimitris Bertsimas, optimization. Published
online in Articles in Advance.

1. Introduction while guaranteeing that each customer who is will-

Dynamic pricing is one of the key tools available to a ing to pay the price at a given period will obtain
service firm trying to match time-varying supply with service. Our main contributions are to provide algo-
time-varying demand. It is, however, a delicate tool rithms that compute the firms optimal pricing pol-
to use in the presence of customers who strategically icy and to characterize the properties of such optimal
time their purchases. As customers change the timing policy.
of their purchases, not only might the firm lose rev- Service guarantees are an important contract fea-
enue, but it might also cause the firms capacity to be ture that are often used when the customers them-
strained in periods where a low price is offered. selves are businesses that rely on the service they pur-
We consider a general formulation of a multiperiod chase to run their own operations. An example of
pricing problem of a service firm trying to maxi- a setting with the aforementioned properties comes
from the cloud computing market where firms sell
mize its revenue while selling service to strategic cus-
computational services to their customers.1
tomers who arrive and depart over time. We assume
In this market, despite demand for service vary-
the firm is constrained by its time-varying service
ing quite significantly over time, customers typically
capacity level and that it wishes to provide service
demand reliability from their providers, in the sense
guarantees to its customers. More specifically, the
firm announces a sequence of prices in advance; each 1
Cloud computing is a large and quickly growing business. The
customer chooses the period with the lowest price combined revenues of this market were estimated to be more than
between her arrival and departure. The sequence of $22 billion in 2010 and are expected to reach $55 billion by 2014
prices is chosen to maximize the revenue of the firm (see Lohr 2011).

Electronic copy available at:

Borgs et al.: Optimal Multiperiod Pricing with Service Guarantees
2 Management Science, Articles in Advance, pp. 120, 2014 INFORMS

that they should be able to purchase service when- attributed to the difficulty of maintaining service level
ever they need it and rationing is not tolerated. guarantees while customers are strategically timing
For instance, consider the case of Dominos Pizza, their purchases. Our work provides the firm with
which is a client of Microsofts Windows Azure cloud techniques for using dynamic pricing in such a con-
computing platform. As explained by Dominos direc- text and, therefore, giving them a tool to better man-
tor of eCommerce: We have daily peaks for din- age their resources and revenues.
ner rush, with Friday night being the biggest. Super There are other examples of service firms that need
Bowl, however, has a peak 50 percent larger than our to set profit-maximizing prices while guaranteeing
busiest Friday night. Windows Azure allows me to service. For instance, the increasingly popular Uber
focus on customer facing functionality, and not have online taxi service offers a flat pricing scheme on
to worry about whether or not I have enough hosting most days of the year, but utilizes dynamic pricing
capacity to support it (Vitek 2009). For Dominos and in high demand days (see Bilton 2012). As explained
many other companies, the key managerial benefit of by its CEO, Uber is aiming to provide a reliable
purchasing cloud computing services is that it per- ride to anybody who needs one, no matter how crazy
mits them to completely ignore the hosting capacity demand is or what is going on in the city (Kalanick
needs of their online businesses. This is only the case 2011). Uber is able to provide such a service guarantee
because the cloud computing providers go to great by, in times of higher demand, conserving resources
lengths to ensure that their service is always available by charging higher prices.
and, therefore, companies who rely on them need not Another interesting application of our work is in
worry about rationing risk.2 the context of electricity markets. As smart meters
The clients of cloud computing services are also (see Federal Energy Regulatory Commission 2008) are
highly heterogeneous with respect to their willingness- beginning to be widely deployed, allowing customers
to-wait for service. Although some companies utilize to immediately respond to price changes, the tech-
cloud computing to run on-demand services and web- niques we develop in this paper will become increas-
sites, and thus always need immediate service, others ingly useful since electricity markets feature many of
use the cloud to run simulations or solve large-scale the elements of our model: capacity, demand, and
optimization problems such as the ones that arise in prices are time varying, and rationing of service is
financial analysis, weather forecasting, and genome highly undesirable.
sequencing. Such clients will typically display more In the industries discussed above, it is mainly the
strategic behavior in their purchasing of cloud ser-
firms responsibility to set prices that ensure that all
vices and wait for lower prices before purchasing ser-
service requests can be accommodated with a limited
vice (e.g., see the AWS case study DNAnexus (Amazon
service capacity. This is in contrast to settings such as
Web Services 2011)).
traditional retailing, where customers are exposed to
Currently, most cloud computing services are sold
rationing risk. In a traditional retail setting, strategic
via static pricing or via a combination of long-term
customers consider the risk of a stock-out and this
contracts and static pricing (e.g., Amazon Web Ser-
incentivizes them to purchase the good earlier. This
vices 2012, Windows Azure 2012). More specifically,
rationing risk mitigates the effect of strategic customer
the customers can purchase computation capacity
behavior on the firms ability to set its own prices. In
(starting at around 10 cents per hour) in a pay-as-you-
our setting, the firms need to offer service guaran-
go model where the per-hour price is constant over
tees places the entire burden of matching supply and
time; this hourly rate is reduced for customers who
pay in advance, via yearly contracts, to reserve capac- demand over time on the firm. The intuition is that
ity. The largest players in this market have mostly the firm should increase its prices when demand is
shied away from selling their higher quality ser- high (or capacity is low) to shift some of the demand
vices via dynamic pricing;3 this could potentially be to the time periods with enough capacity. This is simi-
lar to the pricing scheme used by the major cellphone
carriers that, in order to decongest their networks
Organizations worry about whether utility computing services
will have adequate availability, and this makes some wary of cloud during business hours, often charge lower prices for
computing. 0 0 0Google Search has a reputation for being highly making calls on nights and weekends (e.g., see AT&T
available, to the point that even a small disruption is picked up 2012). The dynamic pricing tools we propose hold the
by major news sources. Users expect similar availability from new promise of helping firms in such industries improve
services... (Armbrust et al. 2010, p. 54).
their resource utilization by better matching supply
The main exception would be Amazons spot market (Ama- and demand over time.
zon Web Services 2012), which is a secondary market run by Ama-
zon to sell the excess capacity of its main platform. Although the
spot market prices fluctuate over time, the exact manner in which
1.1. Our Framework and Contributions
these prices are determined is not publicly available (see Ben- We consider a monopolist that offers service to cus-
Yehuda et al. 2011). tomers over a finite horizon. The firm faces a (possibly

Electronic copy available at:

Borgs et al.: Optimal Multiperiod Pricing with Service Guarantees
Management Science, Articles in Advance, pp. 120, 2014 INFORMS 3

time-varying) capacity constraint at each time period. not too large, and prices can be combined into a pol-
The firms objective is to implement a posted pric- icy via dynamic programming because strategic cus-
ing scheme to maximize its revenue. At time 0, the tomers never wait past a low price to purchase service
monopolist declaresand commits toa sequence of at a future price that is higher.
prices for its service, one for each time period. Given We extend our results to models where the firm
those preannounced prices, customers decide whether does not know its service capacity levels or the
and when to purchase service. The firm needs to solve number of arriving customers. We do so in two
the constrained optimization problem of determining distinct ways. First, we consider a robust optimiza-
the prices that maximize revenue while still fulfilling tion framework (see Ben-Tal and Nemirovski 2002,
all customer purchase requests. Bertsimas and Thiele 2006), where there is uncertainty
Each customer is assumed to be infinitesimal and about the firms capacity and the size of the customer
demands an (also infinitesimal) single unit of ser- population at any given period, and the firm only
knows that these parameters belong to given sets.
vice. The valuation of a given customer for a unit
In this setting, the firm tries to maximize its worst-
of service is drawn from a known distribution. She
case revenues while ensuring that the capacity con-
is also associated with an arrival and a departure
straints are not violated for any realization of demand
time. The arrival time corresponds to the time she
and capacities. Second, we study the model in a
enters the system and the departure time represents stochastic setting where the seller knows the distribu-
her deadline for obtaining service. All customers are tion of the uncertain parameters and is able to obtain
fully strategic about whether and when they purchase additional capacity at a cost; the goal is to determine a
service from the firm. That is, each customer either sequence of (preannounced) prices that maximizes the
refuses to buy service (if her valuation is below any expected profit. Additionally, this model allows for
of the prices offered while she is present) or buys ser- production costs and different value distributions for
vice at the period when she is offered the lowest price customers with differing patience levels. We establish
among all the periods in which she is present; if two that the insights from our baseline model carry on to
periods have the same low price, the customer prefers these general environments. That is, using a dynamic
the earlier one. programming approach similar to our baseline model,
First, we consider a deterministic baseline model we are able to provide algorithms that compute the
where the monopolist knows the total mass of cus- (near) optimal pricing policy in polynomial time.
tomers that arrive at each given time period as well We also consider a related setting where cus-
as their departure periods. The assumption of deter- tomers do not simply choose the earliest time period
ministic demand is justified when the number of cus- with the lowest price to receive service, but rather
tomers is large and fairly predictable, which is often the firm chooses how customers should break ties
the case in the market for cloud computing. This mod- between time periods with equal prices. In this set-
eling choice allows us to study the impact of strate- ting, customers are guaranteed to receive service at
gic customers and time-varying demand and capacity the lowest price available to them, but the firm can
on the optimal sequence of prices, but it deliberately more efficiently use its capacity by scheduling cus-
removes the element of uncertainty from the model. tomers appropriately. Interestingly, our algorithms
The demand being deterministic also implies the opti- can be modified to account for this additional flexi-
bility and solve the optimal pricing problem of the
mality of using a sequence of preannounced prices.
We show later in the paper that many of the insights
Finally, we conduct numerical studies using our
obtained in this simple environment naturally extend
algorithms and obtain further insights on the effect
to general settings that allow for uncertainty in the
of strategic customers and service guarantees on both
model. the firm and its customers. We show that even in
Interestingly, even the solution of this baseline settings with high volatility in service capacity and
model is far from trivial. We show that because of demand, the number of price levels that optimal
the presence of strategic customers, the set of feasi- pricing policy employs is small. For instance, in a
ble price vectors is neither convex nor closed. This 24-period model, the optimal price sequence includes
means no off-the-shelf software can be used to solve four different price levels on average. This shows that
this problem efficiently. Despite these challenges, we even in complex multiperiod settings, the customers
are able to establish that the firms price optimization strategic behavior severely constrains the firms choice
problem is a tractable one and we provide an efficient of price sequence. We also observe that if patient cus-
polynomial-time algorithm for computing the optimal tomers can wait longer for service, both the revenue
pricing policy. This result relies on two crucial ideas: of the firm and the aggregate customer welfare may
the set of prices that the firm needs to consider is decrease. This occurs because the firm is forced to
Borgs et al.: Optimal Multiperiod Pricing with Service Guarantees
4 Management Science, Articles in Advance, pp. 120, 2014 INFORMS

use higher prices to maintain its service guarantees, their time sensitivity, then the optimal sequence of
and consequently the service capacity is underuti- posted prices might also be increasing.
lized. Thus we conclude that in a phenomenon simi- In the aforementioned works, the service provider
lar to Braesss paradox (Basar and Olsder 1999), when uses the customers fear of rationing to extract
customers have additional freedom in choosing the more revenue from strategic customers (see Liu and
time period they purchase service, the overall perfor- van Ryzin 2008). In contrast, in our model the firm
mance of the system may decrease. ensures the customers does not face such risks and
provides service guarantees. Su and Zhang (2009)
1.2. Related Work consider the issue of rationing in the presence of
In this section, we present a brief overview of the strategic customers and find that sellers have an
literature on pricing mechanisms in the presence of incentive to overinsure consumers against the risk of
customers who strategically time their purchases and stockouts, thus showing that providing service guar-
discuss how the results in the literature relate to ours. antees can be in the firms interest.
There is also an extensive literature on dynamic pric- Another paper related to ours is the one by Ahn
ing with myopic customers (see, e.g., Lazear 1986, et al. (2007), which considers joint pricing and produc-
Wang 1993, Gallego and Ryzin 1994, Feng and Gallego tion decisions. Unlike us, they assume that customers
1995, Bitran and Mondschein 1997, Federgruen and are myopic with respect to prices, but, similarly to
Heching 1999). We do not provide a summary of this our model, they assume customers stay in the system
line of literature here, but we refer the reader to excel- for a number of periods unless they make a purchase.
lent surveys by Talluri and van Ryzin (2004), Bitran Interestingly, their analysis also relies on the notion
and Caldentey (2003), Chan et al. (2004), Shen and Su that a low price effectively separates past and future
(2007), and Aviv et al. (2009). through what they call regeneration points.
The study of monopoly pricing in the presence of An altogether different approach to this problem
strategic customers was pioneered by Coase (1972). is the one taken by the dynamic mechanism design
Coase conjectured that in a setting in which a monop- literature. There, the firm offers a direct mechanism
olist sells a durable good to patient customers, if the that allocates its service as a function of customers
monopolist cannot commit to a sequence of posted reports of their private valuations, entry period, and
prices, then the prices would converge to the pro- departure period. See Bergemann and Said (2011) for
duction cost. Later, Stokey (1979, 1981), Gul et al. a survey. The paper closest to this one within this lit-
(1986), and Besanko and Winston (1990) showed that erature is Pai and Vohra (2013), where strategic cus-
a decreasing sequence of prices is optimal for selling tomers arrive and depart over time, but the allocation
durable goods when customers face the trade-off of problem they study is quite dissimilar to the one we
consuming right away versus the possibility of pur- consider.
chasing at the lower prices in the future. They observe The model we consider here differs from many
that customers with high valuations buy in earlier papers in the literature in at least three key aspects:
periods and pay higher prices compared to the low in our model, the firm guarantees service to all pay-
valuation customers. ing customers; therefore, the customers do not face
In the context of revenue management, Aviv and rationing risk. Second, instead of having a fixed
Pazgal (2008) study a model where a monopolist sells inventory at time 0, in our model, the firm has a time-
multiple items over a finite time horizon to strategic varying service capacity, which is nonstorable. Finally,
customers who arrive over time. The authors consider in the previous work, the customers are either present
two classes of pricing strategies: contingent posted from the beginning of the time horizon or arrive over
pricing, where the firms prices may depend on the time but remain until the end of the horizon (or until
remaining inventory, and preannounced posted pric- they make a purchase). In our model, buyers arrive
ing. They observe that commitment (preannounced and depart the system over time.
discount) can benefit the seller when customers are
strategic. Also, ignoring the strategic customer behav- 2. The Baseline Model
ior can lead to significant loss of revenue. Elmaghraby In this section, we formulate the firms revenue max-
et al. (2008) and Dasu and Tong (2010) extend the imization problem, which will be studied in the sub-
analysis to a setting where the seller can reduce the sequent sections. The firm sets a vector of prices over
prices multiple times over the time horizon. Other a finite horizon t = 11 0 0 0 1 T . The prices, denoted by
papers that consider commitment to a pricing pol- p = 4p1 1 0 0 0 1 pT 5, are announced up front, one price
icy include Arnold and Lippman (2001), Levin et al. for each period t.4 Customers arrive and depart over
(2010), and Cachon and Feldman (2013). These works
mainly consider markdown pricing. Su (2007) shows 4
For instance, the horizon of the problem can be chosen as a day,
that if the customers are heterogeneous regarding with periods corresponding to different hours during the day, to
Borgs et al.: Optimal Multiperiod Pricing with Service Guarantees
Management Science, Articles in Advance, pp. 120, 2014 INFORMS 5

time and are infinitesimal. We denote the population any period t. Thus, the firms decision problem is
of customers that arrive at period i and depart at given by
period j by ai1 j . With slight abuse of notation, we also
represent the mass of the population that arrives at sup pt Dt 4p5
period i and departs at period j by ai1 j . p 0 t=1 (OPT-1)
Each customer wants one unit of service5 from
s.t. Dt 4p5 ct 1 for all t 2 811 0 0 0 1 T 91
which she obtains a (nonnegative) value, and cus-
tomers are strategic with respect to the timing of where p 0 is a shorthand notation for pt 0 for
their purchases. Given the vector of prices p, a cus- all t 2 811 0 0 0 1 T 9. The above problem searches for the
tomer from population ai1 j , with value v for the ser- supremum of the objective function instead of the
vice, purchases the service at a time period with the maximum, since the maximum of OPT-1 does not
lowest price between times i and j, if her value is always exist. We demonstrate the nonexistence of an
larger than the lowest price, i.e., if v min`2 i`j 8p` 9. optimal solution in 3, where we also present our
If there is more than one period with the lowest price technique for handling this issue.
in 8i1 0 0 0 1 j9, the customer chooses the earliest period If there were no capacity constraints, the firm could
(with the minimum price) to obtain the service.6 use a single price p at all periods to maximize its rev-
Given a price vector p, we can assign to each pop- enue,7 and P this would result
P in a revenue equal to
ulation ai1 j a service period, denoted by i1 j 4p5. This p41 F 4p55 ij ai1 j . Since ij ai1 j is a constant, we
period has the lowest price among periods in 8i1 0 0 0 1 j9 call p41 F 4p55 the uncapacitated revenue function. We
and is the earliest one (in 8i1 0 0 0 1 j9) with this price. make the following regularity assumption to simplify
Each member of population ai1 j considers purchas- our analysis.
ing service at time i1 j 4p5 and will purchase service Assumption 1. The uncapacitated revenue function
if her value exceeds the price at that period. We call p41 F 4p55 is unimodal. That is, there exists some mo-
the mass of customers that, given prices p, consider nopoly price pM such that p41 F 4p55 is increasing for all
obtaining service at period t as the potential demand at p < pM and decreasing for all p > pM .
time t and denote it by t 4p5. Formally, the potential Note that this assumption implies that pM maxi-
demand is given by mizes p41 F 4p55, and it is satisfied for a wide range
X of distributions, including the uniform, normal, log-
t 4p5 = ai1 j 18t = i1 j 4p591 (1) normal, and exponential distributions.
i1 j2 1itjT We now show, by means of an example, that the set
of feasible prices of OPT-1 is nonconvex.
where 1 is an indicator function.
Each customer assigns a nonnegative value for Example 1. Let the time horizon be T = 3 and
obtaining service. The fraction of customers with assume that a single unit-mass of customers with uni-
value below v is given by F 4v5. For simplicity of pre- form valuations in 601 17 arrive at period 1 and depart
at period 3. Assume that c2 = 0 and c1 1 c3 = 1. Then the
sentation, we assume that F is a continuous function
price vectors 401 0011 15 and 411 0011 05 are both feasi-
and v 2 601 17 for all customers. We also assume that
ble. However, the average of these two price vectors,
customer valuations are independent of their arrival
40051 0011 0055, is infeasible since all customers with
and departure periods, an assumption that we relax in
valuation above 001 seek service at period 2, violat-
7. Hence, given price vector p, the demand at time t, ing the service capacity c2 = 0. Therefore, the set of
denoted by Dt 4p5, is equal to Dt 4p5 = 41 F 4pt 55t 4p50 feasible prices of OPT-1 is nonconvex.
The firms objective is to maximize its revenue,
P The above example illustrates that OPT-1 is a non-
which is given by Tt=1 pt Dt 4p5. However, the firm is
constrained by a service capacity level of ct for each convex optimization problem, and we cannot hope
t 2 811 0 0 0 1 T 9. The firm provides service guarantees to to solve it using off-the-shelf optimization tools. We
its customers, so it must set prices that ensure that show in 5 that despite being nonconvex, there exists
the demand Dt 4p5 does not violate the capacity ct at a polynomial-time algorithm that solves this opti-
mization problem. The construction of this algorithm
relies on the structural properties of this pricing prob-
capture the problem of the firm selecting prices for its next business lem that are explored in 3 and 4.
day. Such an approach would be reasonable when deciding day-
ahead prices for cloud computing services or electricity markets. 7
Because customer valuations are independent of arrival and
A customer that wants multiple units of service could be consid- departure periods, it can be seen from OPT-1 that if there are no
ered as multiple customers in our model. capacity constraints, setting pt = arg maxp p41 F 4p55 for all t maxi-
We relax this assumption in 8. mizes revenue.
Borgs et al.: Optimal Multiperiod Pricing with Service Guarantees
6 Management Science, Articles in Advance, pp. 120, 2014 INFORMS

3. Optimizing over Prices and earlier periods are ranked lower. Namely, if pt = pt0
Rankings and t < t 0 then Rt < Rt0 . It can be seen from the defini-
In this section, we show that there does not always tion of service period i1 j 4p5 (introduced in 2) that in
exist a feasible solution achieving the supremum in OPT-1 for a given price vector p, each population ai1 j
the firms optimization problem. To address this issue, chooses the time period between i and j, with the low-
we construct a closely related optimization problem est customer-preferred ranking to (potentially) receive
where the firm tries to maximize not only prices but service. Hence potential demand t can be expressed
also rankings of the prices. We show that this opti- as a function of this ranking.
mization problem always admits an optimal solu- More formally, for any period t and ranking of
tion that can be used to obtain feasible solutions prices R we define the R-induced potential demand,
arbitrarily close to the supremum of the original denoted by t 4R5, as
We start with an example that shows that the supre- t 4R5 = ai1 j 1 Rt = min 8Rk 9 0 (2)
mum of OPT-1 may not be achieved by a feasible ij
k2 ikj

price vector. The main idea is that since the customers

always seek the lowest price available, the potential Similarly, the R-induced demand, denoted by Dt 4pt 1 R5,
demand function t is a discontinuous function of p; is defined as
thus, the feasible set of OPT-1 is open.
Dt 4pt 1 R5 = 41 F 4pt 55t 4R50 (3)
Example 2. Consider a two-period model with cus-
tomer valuations drawn uniformly from 601 17, capac- It follows from (1), (2), and the definition of
ity levels c1 = 12 and c2 = , and customer populations customer-preferred ranking that for any price vec-
a11 1 = a11 2 = 1 (and a21 2 = 0). Observe that solutions of tor p and customer-preferred ranking RC 4p5, we have
the form 4p1 1 p2 5 = 4 12 1 12 5 are feasible for any > 0: t 4RC 4p55 = t 4p5 and Dt 4pt 1 RC 4p55 = Dt 4p5. That is,
the members of population a11 1 with value above 12 it is possible to express demand (Dt ) in terms of
obtain service at time 1 and the members of popula- the R-induced demand function (Dt ) and customer-
tion a11 2 with value above 12 are served at time 2.
preferred ranking (RC ).
Hence, 4p1 1 p2 5 = 4 12 1 12 5 yields the revenue of 12 Suppose that in OPT-1 the firm could select not
+ 4 12 5 4 12 + 5 = 12 2 . The revenue is decreasing only the vector of prices p but also any rank-
in and as tends to 0 the revenue approaches 12 . ing R consistent with p (potentially different from the
The uncapacitated problem provides an upper bound
customer-preferred ranking), and customers decided
on the revenue obtained, which is 12 . Therefore, the
when to obtain service according to this ranking;
supremum of OPT-1 is equal to 12 . However, 4p1 1 p2 5 =
i.e., each customer chooses the period with the low-
4 12 1 12 5 is not a feasible solution because under this
est ranking between her arrival and departure time.
price vector, both populations will choose the first
Then, the demand at any period is given by Dt 4pt 1 R5,
period for service, and this violates the capacity con-
and the corresponding revenue maximization prob-
straints. Therefore, the feasible set of price vectors is
lem can be formulated as
open and the supremum of OPT-1 is not achieved by
a feasible price vector. T
max pt Dt 4pt 1 R5
The nonexistence of an optimal solution can be p 01 R2P4T 5
addressed by finding (feasible) solutions that are arbi- s.t. Dt 4pt 1 R5 ct
trarily close to the (infeasible) supremum. In the
remainder of this section, we introduce the notion of for all t 2 811 0 0 0 1 T 91 (OPT-2)
rankings and an alternative optimization formulation
that allow us to obtain such solutions for OPT-1 (or Rt < Rt0 ) pt pt0
the optimal solution itself in instances where the opti-
mum is feasible). for all t1 t 0 2 811 0 0 0 1 T 91
We refer to permutations of 811 0 0 0 1 T 9 as rankings.
We use the notation Rt to denote the rank of time where P4T 5 is the set of all possible rankings of
period t under ranking R. We say that a ranking R is 811 0 0 0 1 T 9. Although this problem is different from
consistent with a price vector p if periods with lower OPT-1, the solutions of these problems are closely
rank have lower prices. More precisely, R is consis- related, as we explain next.
tent with p if for all t and t 0 , Rt < Rt0 implies that Since Dt 4pt 1 RC 4p55 = Dt 4p5, it follows that any fea-
pt pt0 . sible solution p of OPT-1 corresponds to a feasible
We define the customer-preferred ranking, denoted by solution of OPT-2 given by 4p1 RC 4p55, and these solu-
RC 4p5, as a ranking consistent with p, such that when tions lead to the same objective values. Additionally,
there are multiple periods with the same price, the it can be seen from (1) and (2) that for a given price
Borgs et al.: Optimal Multiperiod Pricing with Service Guarantees
Management Science, Articles in Advance, pp. 120, 2014 INFORMS 7

vector p and any ranking R consistent with p, the If setting the price equal to pM violates the capacity
(potential) demand levels in OPT-1 and OPT-2 are constraints, then the firm increases its price to the
equal except for periods where the price is equal to minimum price that respects the capacity constraints.
the price offered at another period. Intuitively, unlike Since customers values are bounded by 1, such a
OPT-1, in OPT-2 the firm can choose how the cus- price exists. Thus, it follows that an optimal price
tomers collectively break ties between time periods in 6pM 1 17 can be found. The following proposition
with equal price, by choosing the ranking R prop- shows that this intuition extends to multiperiod
erly, and this may lead to a difference in demand settings.
levels only at such time periods. These observations
Proposition 1. There exists an optimal solution 4p1 R5
can be used to show that OPT-2 always has an opti-
of OPT-2 such that pt pM for all t.
mal solution, and this solution can be used to con-
struct a solution of OPT-1 that is arbitrarily close to To prove this result, we assume that a solution
the supremum. where pt < pM for some t is given, and we raise prices
Lemma 1. The following claims hold: that are below the monopoly price pM in a way that
1. The problem OPT-2 admits an optimal solution maintains the ranking of the prices. This ensures that
4p? 1 R? 5. as the prices increase to pM , the revenue increases
2. Let 4p? 1 R? 5 be an optimal solution of OPT-2. For any while the demand decreases. Thus, it is possible to
> 0, the price vector p? + R? is a feasible solution of obtain a feasible solution that (weakly) improves rev-
OPT-1 and the revenue it obtains converges to the supre- enues and satisfies pt pM .
mum of OPT-1 as tends to 0. Note that conditioned on prices being above the
3. If p is an optimal solution of OPT-1, then 4p1 RC 4p55 monopoly price pM , by the assumption of unimodal-
is an optimal solution of OPT-2. ity of the uncapacitated revenue function, the incen-
tives of the firm and the customers are aligned: both
The proof of this lemma can be found in the online the firm and the customers prefer lower prices over
appendix (available at higher ones. The firm never raises prices to obtain
ilobel/multiperiod-pricing-online-app.pdf). The idea more revenue, only to satisfy capacity constraints.
behind this lemma is that the projection of the set of We next provide a further characterization of the
feasible solutions of OPT-2 onto the set of prices is prices that are used at an optimal solution of OPT-2.
the closure of the feasible set of OPT-1. Therefore, the This characterization significantly narrows down the
optimal prices generated by OPT-2 can be perturbed set of prices that needs to be considered to find an
in a way that maintains the ranking of prices, lead- optimal solution.
ing to a solution of OPT-1 that is arbitrarily close to
the supremum. In the rest of the paper, we focus on Proposition 2. There exists an optimal solution 4p1 R5
the solution of OPT-2, keeping in mind that an opti- of the optimization problem OPT-2 such that for each
mal solution (or a solution arbitrarily close to optimal, period t one of the following three statements is true: pt =
if the optimal solution does not exist) of OPT-1 with pM , pt = 1, or pt = pt for some t, such that ct = Dt 4pt 1 R5
(almost) the same prices can be constructed using this and pt 2 6pM 1 17.
solution. The proof of this proposition follows by showing
that unless the conditions of the proposition hold,
4. Structure of the Optimal Prices the monopolist can modify the prices in a way that
In this section, we explore the structure of optimal increases its profits while maintaining the feasibility
prices in OPT-2. We first show that at all periods of the capacity constraints.8 The third condition of the
the monopolist has incentive to keep the prices at proposition suggests that the price at time t is either
least as high as the monopoly price pM . Then we use such that the capacity constraint at time t is tight or
this observation to study the optimality conditions in that this price is equal to the price offered at another
OPT-2. Exploiting these conditions, we construct a set time period, and the capacity constraint at this other
of prices that contains all the possible candidate opti- time period is tight. Hence, because of the presence of
mal prices. We show that the cardinality of this set is strategic customers, capacity constraints at one period
polynomial in the time horizon T , a result we later use may bind the prices at another period, but this requires
in 5 to obtain a polynomial-time algorithm to solve the prices to be identical at these two periods.
OPT-2. Proofs of the results presented in this section This proposition implies that for an optimal solu-
can be found in the online appendix. tion 4p1 R5 of OPT-2, each entry of the price vector p
To gain some intuition, we first consider the opti-
mal solution in a single period setting. By Assump- 8
We note that if we strengthen Assumption 1 to impose concavity
tion 1, choosing any price p < pM is suboptimal, and on the uncapacitated revenue function, then the proposition can
the firm has incentive to increase its price to pM . be proved using the KarushKuhnTucker conditions.
Borgs et al.: Optimal Multiperiod Pricing with Service Guarantees
8 Management Science, Articles in Advance, pp. 120, 2014 INFORMS

either belongs to 8pM 1 19 or is above pM and satisfies This example suggests that attraction ranges can
the equation be used to determine R-induced potential demand
t 4R5. Assume that 4p1 R5 is a consistent price-ranking
ct = Dt 4p1 R5 = t 4R541 F 4p55 (4)
pair and consider the attraction range of some
for some time period t. However, to characterize the time period k 2 811 0 0 0 1 T 9, denoted by 8 t4k1 R51 0 0 0 1
set of all prices that may appear at an optimal solu- t4k1 R59. As discussed earlier, customers who arrive
tion, we still need to consider all possible rankings. at the system between t4k1 R5 and k (inclusive) and
Although there are T ! possible rankings R, there are who can wait until time k but not beyond time t4k1 R5
a significantly smaller number of prices that satisfy are the only ones who can request service at time k.
equations of the form (4). To formalize this idea, we P Pt4k1R5
Thus, we obtain that k 4R5 = ki= t4k1R5 j=k aij . From
introduce the notion of attraction range, which is a this equation it follows that k 4R5 can immediately
representation of all the populations that choose the be obtained by specifying the attraction range of
same period for service. time period k. By considering all the possible attrac-
Definition 1 (Attraction Range). For a given tion ranges 8 t1 0 0 0 1 t9 corresponding to time period k
consistent price-ranking pair 4p1 R5 the attraction range we conclude that for any ranking R, we have
of a time period k is defined as the largest inter- k 4R5 2 8 ki= t tj=k aij t k t9. Using this observa-
val 8 t1 0 0 0 1 t9 811 0 0 0 1 T 9 containing k such that Rk = tion, it follows that any p satisfying (4) for some R
min`28 t10001t9 R` . and k 4R5 belongs to the set
Assume that the attraction range of time period k
for a consistent price-ranking pair 4p1 R5 is 8 t1 0 0 0 1 t9. 4 1 ck
Lk = max pM 1 F 1 Pk Pt
Since customers choose the time period with the low- i= t j=k aij
est ranking available to them when purchasing ser-
k X
X t
vice, customers who arrive at the system between
ck aij 1 and t k t 0 (5)
periods t and k, and who can wait until time period k i= t j=k
but not beyond time period t, are exactly the ones
who will seek service at period k. Thus, the attraction P P
Here the condition ck ki= t tj=k aij is present since
range concept can be used to identify customers who
are attracted to a particular time period for receiv- F 1 is defined over the domain 601 17. The maximum
ing service (see Example 3). with pM is taken to make sure that all the prices in Lk
are at least equal to pM , which follows from Propo-
Example 3 (Attraction Range). Consider a prob- sition 2. By construction each element of Lk corre-
lem instance with six time periods. Assume that a sponds to an attraction range 8 t1 0 0 0 1 t9. Since there
consistent price-ranking pair 4p1 R5 for this problem are O4T 2 5 attraction ranges (there are O4T 5 values t
is given and the prices at different time periods are and t can take), the cardinality of Lk is O4T 2 5. Thus,
as in Figure 1. Since prices at all time periods are dif- we reach the following characterization of optimal
ferent, there is a unique ranking R consistent with
prices, which is stated without proof because it imme-
these prices. The attraction range of time period 4 in
diately follows from Propositions 1 and 2 and the def-
this example is 821 0 0 0 1 59. Thus, customers who arrive
inition of Lk given in (5).
between time periods 2 and 4 (inclusive) and who
cannot wait beyond time period 5 are the ones who 4 S
Proposition 3. Let L be defined as L = 4 Tk=1 Lk 5 [
seek service at period 4. 8pM 9 [ 819. There exists an optimal solution 4p1 R5 of
OPT-2 such that pt 2 L for all t 2 811 0 0 0 1 T 9. Moreover,
Figure 1 Attraction Range of Period 4 Is 821 0 0 0 1 59 in This Six-Period the cardinality of L is O4T 3 5.
Problem Instance
2 4 5 This proposition implies that without actually solv-
Prices ing OPT-2, it is possible to characterize a superset
of the prices that will be used at an optimal solu-
tion. Moreover, this set has polynomially-many ele-
ments, and it is sufficient for the monopolist to con-
sider these prices when making its pricing decisions.
However, finding the vector of optimal prices could
still be a computationally intractable problem even if
L has small cardinality. In the next section, we show
that this is not case, and we develop a polynomial-
time algorithm that determines the optimal sequence
1 2 3 4 5 6
Periods of prices.
Borgs et al.: Optimal Multiperiod Pricing with Service Guarantees
Management Science, Articles in Advance, pp. 120, 2014 INFORMS 9

5. A Polynomial-Time Algorithm i < k1 k2 < j and (ii) restricting prices to be weakly

In this section, we use the characterization of larger than p. That is,
the optimal prices obtained in 4 to design a
polynomial-timealgorithm for computing the optimal X ij
4i1j1 p5 = max pt Dt 4pt 1R5
sequence of prices. p2LT 1 R2P4T 5
As shown in Proposition 3, an optimal solution of ij
OPT-2 can be obtained by restricting attention to set of s.t. Dt 4pt 1R5 ct
prices L given in Proposition 3. Thus, an optimal solu- for all t 2 8110001T 91 (6)
tion to OPT-2 can be obtained by restricting attention
Rt < Rt0 ) pt pt0
to prices in L and solving the following optimization
problem: for all t1 t 0 2 8110001T 91

pt p for all t 2 8110001T 9
max pt Dt 4pt 1 R5 ij
p2LT 1 R2P4T 5
t=1 where Dt is defined similarly to (3) and denotes the
s.t. Dt 4pt 1 R5 ct demand at time t, assuming ak1 1 k2 = 0 unless i < k1
(OPT-3) k2 < j. Observe that the optimal objective value of
for all t 2 811 0 0 0 1 T 91 OPT-3 is equal to 401 T + 11 05.
For i + 1 > j 1, we assume 4i1 j1 p5 is equal to 0.
Rt < R t 0 ) p t p t 0
On the other hand, for any i1 j such that i + 1 j 1,
for all t1 t 0 2 811 0 0 0 1 T 90 we have

We next show that it is possible to find an optimal 4i1 j1 p5 = max max 84i1 k1 p5 + k 4p5
k28i+110001j19 p2L2 p p
solution of OPT-3 by recursively solving problems
that are essentially smaller instances of itself.
+ 4k1 j1 p59 1 (7)
Consider an optimal solution of OPT-3, denoted by
4p? 1 R? 5. Suppose time period k has the lowest rank- ij
ing; i.e., R?k = 1. In this case, the attraction range of where k 4p5 is given by
k is 811 0 0 0 1 T 9, and k 4R? 5 = ki=1 Tj=k aij . Hence, all 8
k j1
> X X
customers who are present in the system at time k >
> a lm 41 F 4p55p
will seek service at time k. This implies that only >
> l=i+1 m=k
populations ak1 1 k2 , 1 k1 k2 < k can receive service <
ij k j1
at time periods 811 0 0 0 1 k 19 (similarly, only popula- k 4p5 = X X (8)
> if alm 41 F 4p55 ck 1
tions ak1 1 k2 , k < k1 k2 T can receive service at time >
> l=i+1 m=k
periods 8k + 11 0 0 0 1 T 9). Therefore, if the monopolist >
knows pk? and that R?k = 1, it can solve for optimal otherwise.
prices at other time periods by solving two separate
To see why the recursion in (7) holds, consider a solu-
subproblems for time periods 811 0 0 0 1 k 19 and 8k +
tion of (6) and assume that in this solution, k is the
11 0 0 0 1 T 9: maximize the revenue obtained from time time period in 8i + 11 0 0 0 1 j 19 with the lowest ranking
periods 811 0 0 0 1 k 19 assuming only populations ak1 1 k2 and pk p is the corresponding price. Then all popula-
are present (with 1 k1 k2 < k) and similarly for tions that are present in the system at k receive service
time periods 8k + 11 0 0 0 1 T 9. Note that in the solution at this time period. The total mass of these populations
of the subproblems, we need to impose the condition P Pj1
is kl=i+1 m=k alm , since ak1 1 k2 = 0 unless i < k1 k2 < j,
that prices are weakly larger than pk? because other- as can be seen from the definition of 4i1 j1 p5. Since at
wise pl? < pk? for some l, and we obtain a contradiction the optimal solution of (6), the capacity constraints are
to R?k = 1. satisfied, the revenue obtained from time period k is
The above observation suggests that given the time ij
given by k 4pk 5. Since k has the lowest ranking among
period k with the lowest ranking, the pricing problem 8i + 11 0 0 0 1 j 19, only populations ak1 1 k2 such that i <
can be decomposed into two smaller pricing prob- k1 k2 < k can receive service before time k, and the
lems, where the prices that can be offered are lower prices offered at those time periods should be weakly
bounded by the price offered at k. We next exploit larger than pk . It follows from the definition of that
this observation and obtain a dynamic programming the maximum revenue that can be obtained from these
algorithm for the solution of OPT-3. populations (with prices weakly larger than pk ) is
Let 4i1 j1 p5 denote the maximum revenue obtained given by 4i1 k1 pk 5. Similarly, it follows that the maxi-
from an instance of OPT-3 assuming (i) ak1 1 k2 = 0 unless mum revenue that can be obtained from time periods
Borgs et al.: Optimal Multiperiod Pricing with Service Guarantees
10 Management Science, Articles in Advance, pp. 120, 2014 INFORMS

after k equals to 4k1 j1 pk 5. Thus, we conclude that the set of all possible capacity levels by C = t C t
4i1 j1 p5 = 4i1 k1 pk 5 + k 4pk 5 + 4k1 j1 pk 5. The recur- and
Q the set of all possible population matrices by A =
sion in (7) follows since it searches for time period k i1 j Ai1 j . To make dependence of the demand (defined
with the lowest ranking and the corresponding price in Equation (3)) on population size explicit, in this
pk that maximizes the objective of (6). Note that since section we denote the demand at period t, when pop-
k 4p5 = when a capacity constraint is violated, the ulation matrix is given by A 2 A, and the firm uses
solution obtained by solving this recursion also satis- price pt and ranking R 2 P4T 5 by Dt 4pt 1 R1 A5.
fies the capacity constraints. The problem of selecting prices and ranking that
The following theorem shows that a solution of are feasible for all capacity levels in C and popula-
OPT-2, or equivalently a solution of the alternative tion matrices in A, and that maximize the worst-case
formulation in OPT-3, can be obtained by solving for revenue, can be formulated as follows:
prices using the dynamic programming recursion in max M
(7) and constructing a ranking vector consistent with p 01 R2P4T 51 M

these prices. T
s.t. M pt Dt 4pt 1 R1 A5
Theorem 1. The optimal solution of OPT-2 can be t=1
computed in time O4T 6 5.
for all A 2 A1
This theorem, which is proved in the appendix, (OPT-4)
suggests that firms that provide service guarantees Dt 4pt 1 R1 A5 ct
can effectively implement optimal pricing policies by for all t1 ct 2 C t and A 2 A1
solving OPT-2. In the subsequent sections, we show
that this result extends to other settings where there Rt < Rt0 ) pt pt0
is uncertainty about the problem parameters and the for all t1 t 0 2 811 0 0 0 1 T 91
firm has more general objectives.
Our next result shows that this robust optimization
problem is tractable.
6. A Robust Optimization
Proposition 4. The optimal solution of OPT-4 can be
Formulation computed in time O4T 6 5.
The baseline model we considered in previous sec-
tions assumes that both demand and service capacity The optimization problem in OPT-4 is not a special
available over the planning horizon can be fully antic- case of OPT-2 because it uses the population matrix
ipated. This assumption is motivated by our online AL when computing revenue and a different popu-
services application, where the planning horizon is lation matrix AU when computing feasibility. How-
typically counted in hours or, at most, days. Over the ever, with minor modifications, the structural insights
next two sections, we show how to solve the firms and the polynomial-time algorithm from 5 still apply,
problem when this assumption is not valid, by either leading to the result in Proposition 4.
taking a robust or a stochastic view of the demand The robust optimization formulation finds a conser-
or capacity uncertainty. In particular, in this section vative solution that is feasible for all possible values
we introduce a robust optimization formulation of the of uncertain parameters. We next quantify the poten-
firms pricing problem. We show that when there is tial revenue loss due to the uncertainty, when the
uncertainty about either the service capacity levels or solution obtained from this formulation is used for
the size of the customer population, a variant of the pricing. Assume that a solution of OPT-4 is obtained
algorithm of 5 can be used to obtain a solution that using uncertainty sets C and A. Let V ROB 4C1 A1 c1 A5
maximizes revenue while maintaining feasibility for denote the revenue the firm achieves, using this solu-
all possible values of uncertain parameters. Further- tion, when realized parameter values are c 2 C and
more, we can bound the firms worst-case revenue A 2 A. We denote the revenues that could be obtained
loss as a function of the uncertainty in the problem if we knew with certainty c and A by V 4c1 A5. The
parameters. The proofs of this section can be found in following proposition bounds the decrease in the rev-
the online appendix. enues when there is uncertainty, and the solution of
Suppose the firm does not know its service capac- OPT-4 is used to obtain a robust pricing rule.
ity level ct at a given period but only knows that Proposition 5. Suppose ctU 41 + 5ctL for all t 2
it belongs to an interval C t = 6ctL 1 ctU 7. Similarly, the 811 0 0 0 1 T 9 and aUi1 j 41 + 5aLi1 j for all i1 j 2 811 0 0 0 1 T 9.
firm does not know the mass of customers in popu- Then,
lation ai1 j , but instead it knows only that ai1 j 2 Ai1 j = X
6aLi1 j 1 aUi1 j 7. We refer to a collection of population sizes sup 4V 4c1 A5 V ROB 4C1 A1 c1 A55 3 aUi1 j 0
c2C1 A2A
A = 8ai1 j 91i1 jT as a population matrix and represent i1 j
Borgs et al.: Optimal Multiperiod Pricing with Service Guarantees
Management Science, Articles in Advance, pp. 120, 2014 INFORMS 11

This proposition implies that when the uncertainty We make the following assumption throughout this
in the problem parameters is small (i.e., when is section:
small), the revenue loss is also small, provided that a
Assumption 2. For any ranking R and period t, the
solution of OPT-4 is used for pricing. The proposition
functions gt 41 R5 and ht 41 R5 satisfy the following:
also suggests that if a nominal version of the prob-
1. Each customer prefers the time period with the lowest
lem with parameters 4cNOM 1 ANOM 5 is known, and the
rank (among those during which she is present) to (poten-
realized parameters are between 1 and 1 + times
tially) receive service. Hence, the dependence of functions
the nominal ones (i.e., 1 + = 1 + /41 5), then the
gt 2 P4T 5 ! and ht 2 P4T 5 ! on R is through
maximum revenue loss P due to uncertainty is equal to the attraction range of time period t. That is, there exist
44641 + 55/41 55 i1 j aNOM
i1 j .
functions g, h such that

7. A General Framework and an gt 4pt 1 R5 = gt 4pt 1 bt 4R51 et 4R55 and

Approximation Scheme ht 4pt 1 R5 = ht 4pt 1 bt 4R51 et 4R55
In this section, we generalize our baseline model by
incorporating random customer arrivals and capacity where 8bt 4R51 0 0 0 1 et 4R59 is the attraction range of time
levels, production costs, and customer valuations that period t, when ranking R is chosen.
are dependent on arrival and departure periods to the 2. ht 4pt 1 R5 is decreasing in pt , and gt 4pt 1 R5 is Lips-
model. Namely, the distribution of the size and valua- chitz continuous in pt with parameter lt .
tions of each population is known in advance and the
firm determines a sequence of (preannounced) prices The first part of the assumption implies that the
to maximize its expected profit. In order for the firm demand for service at period t and, therefore, the
to be able to satisfy its service guarantees, we consider profit earned at period t, depends only on the price pt
soft capacity constraints in the sense that the firm can and the attraction range generated by the ranking of
exceed the allowable capacity by paying a penalty or prices R. That is, if we modify price pt+1 while leav-
purchasing more capacity. ing the ranking of prices R intact, thus leaving the
At this level of generality, the characterization of attraction range intact, this change in the price vec-
the set of prices in an optimal solution, presented tor p will have no effect on the demand at period t.
in 4, does not hold. However, we can extend our This excludes customer discounting, for example, but
algorithm to obtain a fully polynomial-time approx- is a generalization of the customer behavior model
imation scheme (FPTAS) for the general model. An assumed in earlier sections. The second part of the
FPTAS is an approximation algorithm that for any assumption just implies that demand in a given
> 0 obtains a solution within a factor of 1 of the period decreases continuously in that periods price
optimal solution and is polynomial in the size of the if we maintain a constant ranking of prices. Observe
problem and in 1/. Therefore, using an FPTAS, one that OPT-2 is a special case of OPT-5, when gt 4pt 1 R5 =
can obtain a solution arbitrarily close to the optimal. pt Dt 4pt 1 R5 and ht 4pt 1 R5 = Dt 4pt 1 R5 ct , assuming that
Intuitively, this problem is still tractable since, simi- demand Dt 4pt 1 R5 is Lipschitz continuous in pt , for a
lar to our baseline model, if a time instant has the low- fixed ranking R.
est price in the horizon, then all customers who are For a constant 2 401 15, consider the set of prices
present in the system at this time instant prefer receiv- P = 8k k 2 + 1 k 19 and assume that we seek a
ing service there. Consequently, our main divide-and- solution to OPT-5 by restricting attention to the prices
conquer approach is applicable, and we can reformu- that belong to this set, i.e.,
late the problem as a dynamic programming problem T
following the approach in 5. In this section, we for- max gt 4pt 1 R5
mally explain this idea. The proofs of our results are p2PT 1 R2P4T 5
presented in the online appendix.
s.t. ht 4pt 1 R5 0
We start by introducing an abstract problem that is (OPT-6)
the focus of this section: for all t 2 811 0 0 0 1 T 91
X Rt < R t 0 ) p t p t 0
max gt 4pt 1 R5
p2601 17T 1 R2P4T 5
t=1 for all t1 t 0 2 811 0 0 0 1 T 90
s.t. ht 4pt 1 R5 0 Note that any feasible solution of OPT-6 is feasible
for all t 2 811 0 0 0 1 T 91 in OPT-5. We next show that for small the opti-
mal objective values of these problems are also close.
Rt < Rt0 ) pt pt0 Hence, an optimal solution of OPT-6 can be used to
for all t1 t 0 2 811 0 0 0 1 T 90 provide a near-optimal solution of OPT-5.
Borgs et al.: Optimal Multiperiod Pricing with Service Guarantees
12 Management Science, Articles in Advance, pp. 120, 2014 INFORMS

Lemma 2. Let the optimal solutions of OPT-5 and Lemmas 2 and 3 imply that an approximate solu-
OPT-6 have
P objective values v and v , respectively. Then, tion to OPT-5 can be found in polynomial time pro-
v v Tt=1 lt . vided that gt 4p1 R5 and ht 4p1 R5 can be evaluated in
polynomial time.
We next show that a modified version of the algo-
rithm in 5 can be used to solve OPT-6. Our approach Theorem 2. Assume that for any given t1 p1 R, compu-
is again based on obtaining a solution by recursively tation of gt 4p1 R5 and ht 4p1 R5 takes O4s4T 55 time. An -
solving smaller instances of the problem. For this pur- optimal solution of OPT-5 can be found in O44T 3 s4T 55/2 5
pose, we first define 4i1 j1 p5 to be the maximum time.
utility that can be obtained assuming only popula- The proof immediately follows from Lemmas 2 and
tions ak1 1 k2 , i < k1 k2 < j, are present and the prices 3 and is omitted. In many of the relevant cases (such as
that can be used at periods 8t i < t < j9 are (weakly) revenue maximization subject to capacity constraints
larger than p. It can be seen that the optimal value of as introduced in 2 and 3), for given prices and rank-
OPT-6 is equal to 401 T + 11 05. ings, evaluating constraints and the objective function
We set 4i1 j1 p5 = 0, for i +1 > j 1. Using the same (ht and gt ) can be completed in O415 time. In such set-
argument given in 5 to justify the recursion in (7), it tings, Theorem 2 implies that an approximate solution
follows that for i + 1 j 1, the following dynamic can be obtained in O4T 3 /2 5 time.
programming recursion holds: We conclude this section by showing that this
general framework allows us to find approximately
4i1j1 p5 = max max 84i1k1p5+ k 4p5 optimal prices in polynomial time for problem
k28i+110001j19 p2P 2p p
instances that involve random arrivals and capacity
levels, production costs, and a richer class of cus-
+ 4k1j1p59 1 (10) tomer valuations.
Correlated Valuations. Here, we relax the assumption
where k 4p5 denotes the utility obtained at time k made before that the customers valuations are inde-
with price p, from all populations ak1 1 k2 that can pendent of their arrival and departure periods. Let
receive service at this period and that satisfy i < k1 Fi1 j 4v5 represent the fraction of the ai1 j population that
k2 < j. That is, k 4p5 = gk 4p1 i1 j5, if hk 4p1 i1 j5 0 and values service at most v. We assume that all valua-
k 4p5 = otherwise. tions are in 601 17 and that Fi1 j is differentiable, but we
The intuition behind (10) is similar to the intuition no longer suppose that Assumption 1 holds; i.e., the
of (7): to find 4i1 j1 p5, we search for the time period corresponding uncapacitated revenue function need
with the lowest rank (maximization over k in (10)), not be single peaked. Customer demand at period t
and we search for the best possible price for this time as a function of price pt and the ranking of prices R
period (maximization over p). Since all populations is now given by
that are present at the time period with the lowest X
Dt 4pt 1R5 = ai1 j 41Fi1 j 4pt 5518Rt Rk for all i k j90
ranking (say k) receive service at this time period, the ij
payoff obtained from this time period can be given
by k 4p5. We then solve for prices of subproblems for By choosing gt 4pt 1 R5 = pt Dt 4pt 1 R5 and ht 4pt 1 R5 =
time periods 8i + 11 0 0 0 1 k 19 and 8k + 11 0 0 0 1 j 19. Dt 4pt 1 R5 ct , the corresponding revenue maximiza-
Since the time period with the lowest ranking also tion problem is an instance of OPT-5. Note that for
has the lowest price, we impose the prices for these a fixed R, denoting fi1 j 4p5 = dFi1 j 4p5/dp and assuming
fi1 j is bounded by li1 j 1 we conclude
subproblems to be weakly larger than p. Thus, the
payoffs of the subproblems are given by 4i1 k1 p5 and Dt 4pt 1 R5 X
4k1 j1 p5. Hence, we obtain the recursion in (10) for = ai1 j fi1 j 4pt 518Rt Rk for all i k j9
computing optimal prices in OPT-6. ij
In the following lemma, we use this dynamic pro- ai1 j li1 j 0
gram to construct optimal prices and ranking for the ij
solution of OPT-6 and characterize the computational
Thus, it follows that when fi1 j is bounded for all i,
complexity of the solution. Note that since we are
j, Dt 41 R5 is Lipschitz continuous. This implies that
dealing with general functions gt and ht , our result
gt 41 R5 is also Lipschitz continuous for all t and R.
depends on the computational complexity of evaluat-
Moreover, ht is decreasing in pt (since demand is
ing these functions.
decreasing in pt ). Furthermore, for any t1 p1 R eval-
Lemma 3. Assume that for any given t1 p1 R, compu- uating Dt 4pt 1 R5 and in turn gt 4p1 R5 and ht 4p1 R5
tation of gt 4p1 R5 and ht 4p1 R5 takes O4s4T 55 time. An takes O4T 2 5 time. Thus, Theorem 2 applies, and we
optimal solution of OPT-6 can be found in O44T 3 s4T 55/2 5 conclude that the approximate revenue maximization
time. problem can be solved in O4T 5 /2 5.
Borgs et al.: Optimal Multiperiod Pricing with Service Guarantees
Management Science, Articles in Advance, pp. 120, 2014 INFORMS 13

Production Costs and Soft Capacity Constraints. We that the expectation in E6pt Dt 4pt 1 R5 t 4Dt 4pt 1 R557
also incorporate production costs into the model. We can be evaluated in polynomial time, OPT-7 can be
assume that it costs the firm t 4d5 to provide service solved using the dynamic programming recursion in
to mass d of customers at period t. We make the fol- (10) in polynomial time.
lowing regularity assumption on the production costs:
Assumption 3. Assume that for any period t, the 8. Multiperiod Pricing with
production cost t is a nonnegative, nondecreasing, Customer Scheduling
-Lipschitz continuous function. In the earlier sections, we studied the pricing problem
Along with capturing the cost of producing the ser- of a firm in a setting where the customers choose the
vice to be delivered, the function t can also capture earliest time instant with the lowest price to receive
a soft capacity constraint: if ct represents a capacity service. Consider an example where all customers
level above which any unit produced costs , then we arrive at the initial period and can wait until the end
can capture this by setting t 4d5 = max801 4d ct 59. of the horizon to receive service, but service capac-
Even with soft capacity constraints, we still assume ity is spread over many periods. For any pricing rule,
that the firm provides service guarantees. Whenever in this example all customers receive service at the
the firm is incapable of providing the purchased ser- same time instant (with the lowest price). However,
vice itself, it contracts service delivery out to a third this results in inefficient use of capacity and reduced
party with a unit cost of . revenues.
By letting gt 4pt 1 R5 = pt Dt 4pt 1 R5 t 4Dt 4pt 1 R55 and Motivated by this example, in this section, we con-
ht 4pt 1 R5 = Dt 4pt 1 R5 ct , we obtain an instance of sider the pricing problem in a setting where the firm
OPT-5. From Assumption 3, it follows that gt 4pt 1 R5 is can choose how customers should break ties between
Lipschitz continuous in pt . Since demand is decreas- time instants with equal prices. That is, the customers
ing with price, we observe that ht 4pt 1 R5 decreases still receive service at a time instant with the low-
with price. Thus, Theorem 2 applies, and since est price, but if there are multiple such time instants,
Dt 4p1 R5 (and hence gt 4pt 1 R5, ht 4pt 1 R5) can be eval- the firm schedules how customers should be served.
uated in polynomial time for any given p and R, it Observe that in the example described above, such a
follows that an approximate solution of the problems scheduling of customers would avoid the inefficiency
with production costs and soft capacity constraints created by serving all customers at the same period
can be obtained in polynomial time. (and wasting the capacity in the remaining periods).
Stochastic Arrival and Capacity Processes. Assume that Note that it may not always be in the power of
population sizes 8ai1 j 9i1 j and capacities 8ct 9t are ran- the firm to schedule its customers as described above
dom variables with known distributions. Let E6ai1 j 7 = since at the least this requires knowledge of the arrival
ai1 j and E6ct 7 = ct . In this setting, if a monopolist and departure times (or deadlines) of each individual
wants to guarantee that the total service request does customer, which is not always available. However, in
not exceed the capacity for any realization of the this section we establish that if the firm has the nec-
parameters, it can use the robust optimization frame- essary means to schedule its customers, then it can
work in 6. On the other hand, if the firm has the decide on the optimal prices and schedule by follow-
capability to contract service delivery out whenever ing a dynamic programming approach similar to the
the capacity is exceeded (hence it has soft capacity one discussed in 5.
constraints), it can solve the following expected rev- In this setting, the firms optimization problem can
enue maximization problem: be formulated as follows:
max E6pt Dt 4pt 1 R5 t 4Dt 4pt 1 R557 max pt t
xi1 j
p2P 1R2P4T 5 p2PT 1 x
t=1 t=1 i1 j2 itj
(OPT-7) X
s.t. Rt < Rt0 ) pt pt0 s.t. xi1t j ct
i1 j2 itj
for all t1 t 0 2 811 0 0 0 1 T 90
for all t 2 811 0 0 0 1 T 91
By choosing (OPT-8)
xi1t j pt = xi1t j min 8pk 9
k2 ikj
ht 4pt 1 R5 = 0 and
for all t 2 811 0 0 0 1 T 91
gt 4pt 1 R5 = E6pt Dt 4pt 1 R5 t 4Dt 4pt 1 R5571
X 1
we obtain an instance of OPT-5. Note that if t is Lip- xi1t j = aij
41 F 4pt 55
schitz continuous, then so is gt 4pt 1 R5 = E6pt Dt 4pt 1 R5 t2 itj1 F 4p 5<1

t 4Dt 4pt 1 R557. Thus, under Assumption 3, provided for all i1 j 2 811 0 0 0 1 T 90
Borgs et al.: Optimal Multiperiod Pricing with Service Guarantees
14 Management Science, Articles in Advance, pp. 120, 2014 INFORMS

Here, xi1t j corresponds to the mass of customers that the set of price vectors in S 4i1 j5 such that all prices
belong to population ai1 j and receive service at time t, are (weakly) greater than p and k is the latest period
and P is the set of prices that can be used by the firm. that has price p (i.e., all periods between time k and
For simplicity, in this section we assume that P is a j have a price (strictly) larger than p). If no such
finite set. The first constraint suggests that the capac- feasible set exists, then let S 4i1 j1 p1 k5 = . Addi-
ity constraint is satisfied at all time instants, and the tionally, we define L4i1 j1 p1 k5 = 8t 9 S 2 S 4i1 j1 p1 k5
second one ensures that if a fraction of population ai1 j and pt 4S5 = p9, where pt 4S5 denotes the price at
is scheduled to receive service at time t (i.e., xi1t j > 0), time t in price vector S. This definition suggests that
then the price at time t should be equal to the mini-
L4i1 j1 p1 k5 is the set of all periods that take price
mum price offered from time i up to (and including)
p for some price vector in S 4i1 j1 p1 k5. Using these
time j. The final constraint guarantees that all cus-
definitions, we first provide a characterization of the
tomers that belong to some population ai1 j and have
optimal pricing policy when we restrict attention to
valuation larger than the lowest price between i and
price vectors in S 4i1 j1 p1 k5.
j will receive service.
It is possible to interpret our paper as a two- Lemma 4. Suppose S 4i1 j1 p1 k5 is nonempty and p
stage game where the firm moves first and selects a is (weakly) larger than the monopoly price pM . Then
sequence of prices and the consumers respond in the there exists a price vector S ? 2 S 4i1 j1 p1 k5 such that
second stage. This two-stage game might have mul- pt 4S ? 5 = p for all t 2 L4i1 j1 p1 k5 and S ? maximizes the
tiple equilibria. One equilibrium is the one we call revenue (objective of OPT-8) among all the price vectors
the baseline modelconsumers break ties in favor of S 4i1 j1 p1 k5.
buying service early. Another potential equilibrium is
the one where consumers break ties in favor of buying Now, suppose S 4i1 j1 p1 k5 is nonempty, and S ? 2
at the most favorable period for the firm. The second S 4i1 j1 p1 k5. Observe that if we remove periods in
equilibrium is the one studied in this section. L4i1 j1 p1 k5 from the set of periods between i and j,
Observe that our baseline model can be viewed as we will end up with some intervals (consecutive time
a version of OPT-8 where the firm is restricted to periods belong to the same interval and we may have
breaking ties by assigning the customers to the ear- intervals with only one period). We denote the `th
liest time instant with the lowest price. Hence, if the such interval by 6Ii1` j1 p1 k1 0 1 Ii1` j1 p1 k1 1 7 and note that there
firm has the ability to break the ties favorably, it may are fewer than 4j i5 intervals. We also observe that in
obtain a higher revenue. We next show that in this S ? , only populations that arrive after time 4Ii1 j1 p1 k1 0 15
case the optimal pricing policy (i.e., the solution of and leave before period 4Ii1` j1 p1 k1 1 +15 receive service at
OPT-8) can be obtained by using a dynamic program- the periods in 6Ii1` j1 p1 k1 0 1 Ii1` j1 p1 k1 1 7 because the price at
ming approach that generalizes the one used for the periods Ii1` j1 p1 k1 0 1 and Ii1` j1 p1 k1 1 + 1 is equal to p. Now
solution of our baseline model. In particular, given we can state the recursion for computing 4i1 j1 p5:
a set of prices that can be used by the seller P, this

algorithm finds the optimal solution of OPT-8 in time ij
polynomial in T and size of P. 4i1 j1 p5 = max max k 4p5
k2 i<k<j p2 p p
To compute the optimal revenue in the OPT-8, sim-
ilar to 5, define 4i1 j1 p5 as the maximum revenue + 4Ii1` j1 p1 k1 0 11 Ii1` j1 p1 k1 1 + 11 p5 1 (11)
that can be obtained from all populations arriving `
between periods i and j, using prices weakly greater
where, similar to 5, is given by
than p, with the boundary condition that all cus-
tomers who can wait until or after time j receive ser- 8
> X
vice at a later time instant. In our original model, we >
> a lm 41 F 4p55p
recursively calculated 4i1 j1 p5 by finding a period k < l1 m2 9 t2L4i1 j1 p1 k51 ltm
where it gets the lowest price and highest ranking k 4p5 = (12)
(see 5). In the new model, since a population can be >
> if L4i1 j1 p1 k5 6= 1
split into different periods, instead of finding a sin- if L4i1 j1 p1 k5 = 0
gle period that takes the lowest price, we find a set ij
of such periods. Before presenting the algorithm, we Namely, k 4p5 denotes revenue obtained from the
need a couple of definitions. periods in L4i1 j1 p1 k5, by setting their prices equal to
Let S 4i1 j5 be the set of all feasible price vectors for p p, assuming only the populations between peri-
subproblem 4i1 j5. Namely, S 4i1 j5 is the set of all fea- ods i and j are present in the system. Following the
sible solutions of problem OPT-8 where the mass of same argument in 5, and using Lemma 4, it is easy to
all populations except those arriving between time i show that the recursion calculates the optimal value
and j is assumed to be zero. Define S 4i1 j1 p1 k5 to be for 4i1 j1 p5. Note that (12) suggests that if L4i1 j1 p1 k5
Borgs et al.: Optimal Multiperiod Pricing with Service Guarantees
Management Science, Articles in Advance, pp. 120, 2014 INFORMS 15

can be found in polynomial time, then can be com- whereas ai1 i+s is generated from a uniform distribu-
puted efficiently. Our next result, which is proved in tion between 0 and m2 , where m1 and m2 are simula-
the online appendix, shows that this is the case. tion parameters.
Lemma 5. Set L4i1 j1 p1 k5 can be found in polynomial 9.1. Effects of Capacity Constraints
time. We first consider different capacity regimes and try
Observe that we have O4T 2 P5 subproblems to understand their impact on pricing rules. We con-
4i1 j1 p5. Therefore, by the lemma given above and sider capacity vectors that satisfy one of the following
the recursion in (11), we immediately obtain the fol- cases: (case 1) ct = 1 for all t 2 811 0 0 0 1 T 9, i.e., con-
lowing theorem. stant capacity; (case 2) ct = 1025 0054t 15/4T 15 for
all t 2 811 0 0 0 1 T 9, i.e., capacity decreasing from 1025 to
Theorem 3. The optimal solution of OPT-8 can be 0075 over the horizon; (case 3) ct = 0075 + 0054t 15/
found in polynomial time. 4T 15 for all t 2 811 0 0 0 1 T 9, i.e., capacity increasing
from 0075 to 1025 over the horizon; (case 4) ct = 1025
9. Numerical Insights 4t 15/4T 15 for all t T /2 and ct = 0075 + 4t
In this section, we consider generic instances of the 1 T /25/4T 15 for all t > T /2, i.e., capacity first
firms pricing problem and obtain qualitative insights decreasing from 1025 to 0075 and then increasing back
about the optimal pricing scheme introduced in 2. to the original level, with a midday minimum. The
We first investigate the effect of available capacity first three cases capture the constant, decreasing, and
on the optimal prices used by the firm and estab- increasing capacity settings, respectively. The last one
lish that the prices closely track the capacities; e.g., captures a phenomenon that is typical in cloud com-
decreasing capacities induce increasing prices and puting markets: during peak business hours, part of
vice versa. Then, we focus on how patience level the service capacity is usually unavailable because of
of players affect the outcome and show that as cus- high demand on the servers due to other contracts
tomers become more patient, the firm offers higher and obligations.
prices that leads to underutilization of capacity and We next plot the average price vector over 100 ran-
lowered revenues and customer welfare. In addition, domly generated problem instances (Figure 2). We
we observe that when customers are patient, the firm consider three settings in which (i) the entire popu-
ends up using only a few different prices, thereby lation is impatient (6m1 1 m2 7 = 661 07), (ii) half of the
considerably decreasing the complexity of the pric- population is patient (6m1 1 m2 7 = 631 37), (iii) the entire
ing policy. Finally, we compare the pricing schemes population is patient (6m1 1 m2 7 = 601 67).
we introduce in this paper with static pricing that Figure 2 shows that in all four cases (and for all
is commonly employed in practice and establish that population structures), prices track service capacities.
it is possible to significantly improve the revenues Prices are lower when service capacities are higher
using our algorithms. We conclude by testing the run and vice versa (note that the ranges of the verti-
time of our algorithms and showing that for realis- cal axes in Figure 2 are not the same). Case 2 is to
tic scenarios optimal prices can be computed in a few some extent analogous to a typical revenue manage-
minutes. ment setting. As capacity dwindles toward the end
Unless noted otherwise, we will always consider of the horizon, prices rise accordingly. An interest-
problem instances with 36 time periods, where we ing phenomenon occurs as we move from the graph
focus on the middle 24 periods to avoid potential with impatient customers, on the left, toward the one
boundary effects.9 We let customer valuations be uni- with patient customers, on the right: Prices become
formly distributed between 0 and 1. We assume that both smoother and higher as customers become more
there are two types of populations arriving at each patient. We explore this observation more in the next
time period: (i) impatient (or myopic) customers (who subsection.
are only interested in purchasing service at the period
they arrived), and (ii) strategic (or s-patient) cus- 9.2. Effects of Strategic Behavior
tomers, who are willing to wait up to s periods to We now investigate how strategic behavior of cus-
purchase service. This is captured by setting all ai1 j tomers affects revenues, capacity usage, and customer
equal to 0 unless j = i (myopic customers) or j = i + s welfare as the parameters s (willingness to wait for
(strategic ones). For each i, ai1 i is generated at ran- strategic customers) and m2 /4m1 +m2 5 (fraction of cus-
dom from a uniform distribution between 0 and m1 , tomers who are strategic) change. Our results indicate
that as customers become more patient (and strate-
We note that no significant changes are observed in our results in
gically time their purchases), the monopolist uses
any of the subsections of this section, when the entire time horizon fewer different prices that are on average higher. This
is used for the analysis. leads to inefficient use of the available capacity and
Borgs et al.: Optimal Multiperiod Pricing with Service Guarantees
16 Management Science, Articles in Advance, pp. 120, 2014 INFORMS

Figure 2 Average Price over 24 Time Periods of Interest

(a) Population is impatient (b) Half of population is patient (c) Entire population is patient
0.72 0.80 0.84
0.79 0.83
0.78 0.82
Average price

Average price

Average price
0.68 0.77 0.81
0.76 0.80
0.75 0.79
0.64 0.74 0.78
Case 1
0.73 0.77
0.62 Case 2
Case 3 0.72 0.76
Case 4
0.60 0.71 0.75
0 5 10 15 20 25 0 5 10 15 20 25 0 5 10 15 20 25
Time Time Time

reduces both the revenue of the firm and the welfare that the optimal price in a given time period must
of customers. either belong to 8pM 1 19 or be equal to the price of a
We assume that willingness to wait for strategic time period where the capacity is tight. This implies
customers s belongs to set 801 11 0 0 0 1 89. The capac- that the total number of price levels used might be
ities are generated independently and uniformly smaller than the total number of periods. Our sim-
between 005 and 105 for each time period. Addi- ulation in Figure 3(a) shows that this is indeed the
tionally, the sizes of impatient and patient popula- case. In particular, it shows that the average number
tions are characterized by the following cases: (case of price levels over the 24-period horizon drops both
1) 6m1 1 m2 7 = 651 17; (case 2) 6m1 1 m2 7 = 631 37; (case 3) when a higher fraction of the population is willing
6m1 1 m2 7 = 611 57; (case 4) 6m1 1 m2 7 = 601 67. That is, case to wait for service and when the customers who are
1 captures the scenario, where most of the population willing to wait become more patient. For example, in
is impatient, whereas, case 4 captures the one where case 2, while roughly 14 prices are needed when cus-
all buyers are patient. Since parameters are generated tomers are willing to wait only up to one period, this
randomly, we present our results by averaging them number drops to 8 if they are willing to wait for two
over 100 problem instances. periods and 5 if they are willing to wait for three peri-
We first consider the average price (over the hori- ods. Moreover, this drop in the number of prices is
zon) offered by the monopolist for different cases and more significant if a larger proportion of the popula-
s parameters. We also plot average number of dif- tion is patient. We note that when an optimal solution
ferent prices used by the monopolist for the optimal for the original pricing problem OPT-1 does not exist,
solution of the pricing problems. Proposition 2 stated Lemma 1 suggests using perturbed prices 8pt + Rt 9t

Figure 3 Average Number of Different Price Levels (a) and Average Prices (b)

(a) (b)
18 0.90

Average number of price levels

Average price

Case 1 = [5, 1]
Case 2 = [3, 3]
Case 3 = [1, 5] 0.75
Case 4 = [0, 6]


2 0.65
0 1 2 3 4 5 6 7 8 0 1 2 3 4 5 6 7 8
Patience level Patience level
Borgs et al.: Optimal Multiperiod Pricing with Service Guarantees
Management Science, Articles in Advance, pp. 120, 2014 INFORMS 17

Figure 4 Wasted Capacity, Revenue, and Customer Welfare over 24 Time Periods

(a) % Wasted capacity (b) Revenue of the firm (c) Customer welfare
70 15 4.0
Case 1 = [5, 1]
Percentage of wasted capacity

60 Case 2 = [3, 3] 3.5

Total consumer welfare

Case 3 = [1, 5]
Case 4 = [0, 6] 13
50 3.0

Total revenue
40 2.5
30 2.0

20 1.5
10 1.0

0 7 0.5
0 1 2 3 4 5 6 7 8 0 1 2 3 4 5 6 7 8 0 1 2 3 4 5 6 7 8
Patience level Patience level Patience level

for an arbitrarily small to obtain solutions arbitrarily customer has for the service and her payment). In Fig-
close to the optimal. In such cases, our results indicate ures 4(b) and 4(c), respectively, we plot the average
that the firm uses essentially few prices. Our first revenue and customer welfare for different cases and
conclusion in this set of simulations is that patient s parameters. We establish that as customers become
customers lead to fewer price levels. more patient, both revenue loss and customer welfare
As customers become more patient, the firm reduction become quite significant. For instance, it can
becomes more constrained in the prices it can offer. be seen that for the case m1 = 1, m2 = 5, the revenue
Consequently, to maintain feasibility with fewer and welfare for s = 8 are, respectively, 35% and 75%
prices and sustain its service guarantees, it may need lower compared to a scenario with s = 0. This phe-
to increase the prices at some periods. Recall that the nomenon analogous to Braesss paradox that arises in
prices were already at or above monopoly price to transportation problems (where opening a new road
begin with, so both the firm and the customers lose may lead to higher overall congestion in the trans-
as the prices go up. Even a small increase in cus- portation network). However, the mechanism at work
tomer patience causes a fairly large increase in aver- here is different from the one at Braesss paradox since
age prices (see Figure 3(b)) and, as expected, the effect in our setting the lower welfare is a consequence of
is more pronounced when a larger fraction of the the firms price adjustment (raising prices to maintain
population is strategic. Thus, we also conclude that feasibility of solution).
patient customers lead to higher prices.
9.3. Static Pricing and Multiperiod Pricing
We next focus on the effect of these higher prices
In this subsection, we compare the performance of
on capacity usage, revenue, and customer welfare. At our multiperiod pricing algorithms with the static
first glance, the presence of customers that are more pricing that is commonly used in practice. The three
patient would seem to lead to better use of resources. scenarios considered here are the following: (case 1)
After all, high demand and low supply in one period monopolist can use multiple prices and selects the ser-
followed by low demand and high supply in the next vice time of the customers, as in 8; (case 2) monop-
period could be properly matched if customers are olist can use multiple prices, but customers receive
willing to wait. Indeed this phenomenon does show service in their preferred (earliest) period, as in our
up in our numerical analysis to a small extent, when base model; and (case 3) monopolist is restricted to
customers switch from being completely impatient to using a single price and customers receive service in
willing to wait for one period (see cases 1 and 2 in their preferred period, which is the case closest to
Figure 4(a)). However, we mainly observe the oppo- current practice in the cloud computing industry. We
site effect. As customers become more patient (for assume that half the population is impatient and half
s 1), the firm is forced to use fewer and higher is patient with willingness-to-wait s 2 801 0 0 0 1 89.
prices. These prices lead to inefficient use of the firms Figure 5 shows our result. The cases where the firm
resources. The inefficiency is higher when a larger is most flexible (case 1) and least flexible (case 3) in
fraction of customers is impatient (Figure 4(a)). This terms of its pricing (and scheduling) policies, respec-
phenomenon lowers the firms revenue and simulta- tively, lead to the best and worst cases in terms of
neously reduces customer welfare (i.e., total surplus capacity management and firm profits. The efficiency
of the customers who purchase the service, where sur- gains from better pricing strategies are large, and thus
plus is defined as the difference between the value a the customers are better off when the firm has the
Borgs et al.: Optimal Multiperiod Pricing with Service Guarantees
18 Management Science, Articles in Advance, pp. 120, 2014 INFORMS

Figure 5 Wasted Capacity, Revenue, and Customer Welfare over 24 Time Periods of Interest









most tools in its arsenal. Interestingly, our base algo- (with 3006 GHz Intel Core 2 Duo processor and 4 GB
rithm that uses only pricing to direct customers to 1,067 MHz DDR3 memory). Therefore, the algorithm
periods performs well for low values of patience s, is sufficiently efficient to be implemented in practice.
but its performance degrades as s increases. If the firm
has some mechanism aside from pricing for schedul- 10. Conclusions
ing customers to periods and customers are fairly We study a service firms multiperiod pricing prob-
patient, then the pricing scheme with scheduling lem in the presence of time-varying capacities and
should be used because the performance of this alter- heterogeneous customers that are strategic with
native algorithm improves with customer patience. respect to their purchasing decisions. A distinct fea-
ture of our model is the service guarantees pro-
9.4. Running Time vided by the firm that ensure that any customer
We conclude this section by discussing the running willing to pay the announced service price will be
time of our base algorithm. In this section, we con- able to receive service. Such guarantees are quite
sider problem instances where the horizon length is appealing to customers because they allow them to
given by T 2 8241 481 969. We still assume that the ignore rationing risk, tremendously simplifying the
capacities are drawn from 60051 1057 uniformly at ran- consumers decision-making process. However, pro-
dom. We first focus on problems where there are two viding such guarantees requires the firm to use prices
populations impatient and s-patient, and s-patient that ensure the firm has sufficient service capacity
players can wait for s 2 801 11 0 0 0 1 89 time instants in every period. We propose an efficient algorithm
to receive service. Moreover, the size of impatient to compute revenue-maximizing prices while main-
and patient populations are drawn from 601 m1 7 and taining service guarantees. We show, via numerical
601 m2 7, respectively, uniformly at random. We then simulation, that in a typical instance the optimal pric-
consider a heterogeneous setting, where all population ing policy involves only a few prices and it enables
groups (characterized by an arrival and departure the firm to obtain significantly more revenues than
period) are present and their sizes are drawn at ran- the static pricing schemes that are common in prac-
dom. For each of these randomly generated prob- tice. We show that such algorithms and insights
lem instances, we run our simulations 100 times and generalize to complex versions of the problem with
report the average running time in seconds in Table 1. random arrivals, departures and capacity levels, pro-
Running times increase as either s or T grows, duction costs, and customer valuations that depend
but our results do indicate that in instances of rea- on arrival and departure periods. We also construct
sonable size (up to T = 96), the optimal prices can an algorithm that the firm can use if it is capable
be found in a few minutes using a standard laptop of scheduling customer service times in addition to

Table 1 Average Running Times for Random Problem Instances (in Seconds)

s=0 s=1 s=2 s=3 s=4 s=5 s=6 s=7 s=8 Heterogeneous

T = 24 0.99 3.55 5.88 8.48 10.24 13.23 14.23 18.77 20.10 32.74
T = 48 14.9 51.4 85.0 105.6 141.8 167.3 168.4 165.0 190.7 233.6
T = 96 193 655 868 1,102 1,232 1,304 1,501 1,489 1,550 1,802
Borgs et al.: Optimal Multiperiod Pricing with Service Guarantees
Management Science, Articles in Advance, pp. 120, 2014 INFORMS 19

using dynamic pricing and demonstrate numerically p = 0, and hence the optimal value is equal to 401 T + 11 05.
that such an algorithm could yield even higher rev- Given 4i1 j1 p5, for 0 i j T + 1, one can construct the
enues and better resource utilization for the firm. optimal sequence of prices in this problem using the recur-
sion in (7): We say that k is the solution for 4i1 j1 p5 if the
Acknowledgments right-hand side of Equation (7) takes its maximum at k and
The authors thank Ishai Menache (Microsoft Research), k is the earliest time period that achieves the maximum. Let
Georgia Perakis (MIT), Rakesh Vohra (University of 4k 1 pk 5 be the optimal solution of 401 T + 11 05 in (7). Then
Pennsylvania), and the anonymous associate editor and ref- the price of time period k in the optimal solution of (6) is
erees for their fruitful comments and suggestions. pk , and prices for time periods earlier and later than k can
be obtained by solving for the prices in the subproblems
Appendix 401 k 1 pk 5 and 4k 1 T + 11 pk 5.
We assume that at each step the leftmost subproblem is
Proof of Proposition 2. Note that if for a ranking R, solved first. We say that the time period k , which solves the
we have t 4R5 = 0, then, without loss of generality, we can ith subproblem, has priority i (hence the time period that
let pt = 1. Now, by Proposition 1, we can assume that 4p1 R5 solves 401 T + 11 05 has priority 1). Using these priorities
is an optimal solution of OPT-2 such that pM pt 1 and together with prices, we next construct the ranking vector
pt = 1 if t 4R541 F 4pt 55 = 0 for all t 2 811 0 0 0 1 T 9. We will (consistent with the already obtained prices) that appears
prove that 4p1 R5 is such that pt necessarily satisfies one of in the solution of OPT-3 (or equivalently to (7) with i = 0,
the conditions 13 of the proposition for all t 2 811 0 0 0 1 T 9. By j = T + 1, p = 0). Consider time periods k1 and k2 . If pk1 6=
contradiction, assume that for 4p1 R5 none of the conditions pk2 , it is clear how to rank them: the lower price will have a
13 hold at time t0 . Since conditions 12 do not hold 1 > smaller rank. Now suppose pk1 = pk2 ; then the time period
pt0 > pM . Note that pt0 6= 1 implies that t0 4R541 F 4pt0 55 6= 0, with lower priority receives lower ranking. Note that under
and hence t0 4R5 > 0. Since condition 3 does not hold, for t 2 this ranking, the ranking vector is consistent with prices.
S = 8t 2 811 0 0 0 1 T 9 pt = pt0 9 we have that ct is not tight; i.e., Moreover, when there are multiple time periods with the
same price, the time period that has lower ranking is the one
t 4R541 F 4pt 55 < ct for all t 2 S0 (13)
that is used by the algorithm to solve an earlier subproblem.
Let be a constant such that This implies that the ranking is consistent with the time
8 period each population receives service in the solution of
<pt pM
if pt0 pk for all k 2 811 0 0 0 1 T 91 the recursion (7).
= p max p otherwise. We next characterize the computational complexity of
: t0 t1
8t1 pt <pt 9
1 0 providing a solution to OPT-2. Note that by Proposition 3,
Consider the price vector p, for which pk = pk for k y S, and there exists an optimal solution for OPT-2 with prices that
pk = pk otherwise, for some 0 < < . It follows from belong to set L. It can be seen from (5) that to compute
the definition of that if pi pj for some i1 j 2 811 0 0 0 1 T 9, the prices in this set we need quantities of the form zijk =
Pk Pj 3
then pi pj . Hence, the price vector p is also consistent with t1 =i t2 =k at1 1t2 for all i j k. Note that there are O4T 5
ranking R. Moreover, since 41 F 4p55 is a continuous func- values zijk can take, and each value takes at most O4T 2 5
tion, by (13) we conclude that can be chosen small enough to compute. Thus, all values of zijk and the set L can be
to guarantee that for time periods t 2 S, t 4R541 F 4pt 55 < ct . computed in O4T 5 5 time (the computation time can be fur-
Since 4p1 R5 is feasible and pt = pt for t y S, it also follows ther reduced by exploiting the relation between different
that for t y S, we have t 4R541F 4pt 55 = t 4R541F 4pt 55 ct . values of zijk ; this is omitted because it does not affect our
Consequently, 4p1 R5 is feasible in OPT-2. The definition of final complexity result). Thus, in O4T 5 5 time we can reduce
also suggests that pM < pt < pt = pt0 for t 2 S. It follows by OPT-2 to OPT-3. We characterize the computational com-
the definition of pM and the unimodality of the uncapaci- plexity of the latter problem.
tated revenue function that pt 41 F 4pt 55 < pt 41 F 4pt 55 for Observe that the algorithm relies on characterizing
t 2 S. Thus, since t0 4R5 6= 0 we conclude that the revenue 4i1 j1 p5 for all time periods i j and p 2 L. Since cardinal-
obtained from time periods t 2 S increases under p; i.e., ity of L is O4T 3 5, there are O4T 5 5 values of 4i1 j1 p5 that
X X need to be characterized. These can be computed using the
pt 41 F 4pt 55t 4R5 > pt 41 F 4pt 55t 4R50 (14) condition 4i1 j1 p5 = 0 if i + 1 j 1 and the recursion in
t2S t2S
(7). At each step of the recursion, there are O4T 5 different
P values k can take. On the other hand, for a given value of
Since pt = pt for t y S, it also follows that tyS pt 41
P k, the corresponding optimal pk can be computed in O415:
F 4pt 55t 4R5 = tyS pt 41 F 4pt 55t 4R5. Hence, we conclude
that the overall revenue improves when 4p1 R5 is used. Since for all p 2 L we have p pM , it follows that k 4p5 is
Therefore, we reach a contradiction and 4p1 R5 has to satisfy decreasing in p, provided that p 2 L. Moreover, 4i1 j1 p5 is
one of the conditions 13 of the proposition. also decreasing in p for all i1 j since larger p corresponds
to tighter constraints in (6). Thus, the pk that solves (7)
Proof of Theorem 1. We first describe how the opti- is the smallest p pM that makes the capacity constraint
mal prices and ranking in OPT-2 are obtained, and then we feasible. Therefore, it follows that pk = max8pM 1 F 1 41 ck /
Pk Pj1
consider the computational complexity of the solution. As l=i+1 m=k alm 59, where the latter is the price that makes
explained in the text, given the set L, the optimal solution of the capacity at time k tight. Since by construction both these
OPT-2 can be obtained by solving OPT-3. The solution of the prices belong to L, and elements of L were computed ear-
latter problem is identical to that of (6), with i = 0, j = T + 1, lier, it follows that given k, pk can be constructed in O415.
Borgs et al.: Optimal Multiperiod Pricing with Service Guarantees
20 Management Science, Articles in Advance, pp. 120, 2014 INFORMS

Thus, we conclude that each step of the recursion in (7) Chan LMA, Shen ZJM, Simchi-Levi D, Swann JL (2004) Coor-
can be computed in O4T 5. Thus, the overall complexity of dination of pricing and inventory decisions: A survey and
computing all 4i1 j1 p5 is O4T 6 5. Finally, given all values classification. Simchi-Levi D, Wu SD, Shen Z-J, eds. Handbook
of Quantitative Supply Chain Analysis, International Series in
of 4i1 j1 p5, the construction of the prices that solve (6)
Operations Research & Management Science, Vol. 74 (Springer,
takes O4T 2 5 following the procedure described in the begin- New York), 335392.
ning of the proof: to solve for each pk , an instance of the Coase RH (1972) Durability and monopoly. J. Law Econom. 15(1):
recursion (7) needs to be solved. This takes O4T 5 time, and 143149.
there are O4T 5 prices to be solved for. Similarly constructing Dasu S, Tong C (2010) Dynamic pricing when consumers are strate-
priorities and rankings consistent with these prices takes gic: Analysis of posted and contingent pricing schemes. Eur. J.
Oper. Res. 204(3):662671.
another O4T 5. Thus, the overall complexity of the algorithm
Elmaghraby W, Glc A, Keskinocak P (2008) Designing optimal
is O4T 5 + T 6 + T 2 + T 5 = O4T 6 5. preannounced markdowns in the presence of rational cus-
tomers with multiunit demands. Manufacturing Service Oper.
Management 10(1):126148.
Federgruen A, Heching A (1999) Combined pricing and inventory
References control under uncertainty. Oper. Res. 47(3):454475.
Ahn H, Gms M, Kaminsky P (2007) Pricing and manufactur- Feng Y, Gallego G (1995) Optimal starting times for end-of-season
ing decisions when demand is a function of prices in multiple sales and optimal stopping times for promotional fares. Man-
periods. Oper. Res. 55(6):10391057. agement Sci. 41(8):13711391.
Amazon Web Services (2011) AWS Case Study: DNAnexus. Federal Energy Regulatory Commission (2008) Assessment of
Accessed June 1, 2013, demand response and advanced metering. Staff report, Federal
-studies/dnanexus/. Energy Regulatory Commission, Washington, DC. http://
Amazon Web Services (2012) Amazon EC2 Pricing. Accessed
June 1, 2013, Gallego G, Ryzin GV (1994) Optimal dynamic pricing of inventories
Armbrust M, Fox A, Griffith R, Joseph AD, Katz R, Konwinski A, with stochastic demand over finite horizons. Management Sci.
Lee G, et al. (2010) A view of cloud computing. Comm. ACM 40(8):9991020.
53(4):5058. Gul F, Sonnenschein H, Wilson R (1986) Foundations of dynamic
Arnold MA, Lippman SA (2001) The analytics of search with posted monopoly and the Coase conjecture. J. Econom. Theory 39(1):
prices. Econom. Theory 17(2):447466. 155190.
AT&T (2012) Wireless Home Phone Plans. Accessed June 1, 2013, Kalanick T (2011) NYE surge pricing explained. (blog) (December 31),
.html. -pricing-explained/.
Aviv Y, Pazgal A (2008) Optimal pricing of seasonal products in the Kimball WS (1952) Calculus of Variations by Parallel Displacement
presence of forward-looking consumers. Manufacturing Service (Butterworth, London).
Oper. Management 10(3):339359. Lazear EP (1986) Retail pricing and clearance sales. Amer. Econom.
Aviv Y, Levin Y, Nediak M (2009) Counteracting strategic consumer Rev. 76(1):1432.
behavior in dynamic pricing systems. Netessine S, Tang CS, Levin Y, McGill J, Nediak M (2010) Optimal dynamic pricing of
eds. Consumer-Driven Demand and Operations Management perishable items by a monopolist facing strategic consumers.
Models (Springer, New York), 323352. Production Oper. Management 19(1):4060.
Basar T, Olsder GJ (1999) Dynamic Noncooperative Game Theory, Liu Q, van Ryzin GJ (2008) Strategic capacity rationing to induce
2nd ed. (Society for Industrial Mathematics, Philadelphia). early purchases. Management Sci. 54(6):11151131.
Ben-Tal A, Nemirovski A (2002) Robust optimization methodology Lohr S (2011) The business market plays cloud computing catch-up.
and applications. Math. Programming 92(3):453480. New York Times (April 15),
Ben-Yehuda OA, Ben-Yehuda M, Schuster A, Tsafrir D (2011) 15/business/15cloud.html.
Deconstructing Amazon EC2 spot instance pricing. Proc. 2011 Pai MM, Vohra R (2013) Optimal dynamic auctions and simple
Third Internat. Conf. Cloud Comput. Tech. Sci. (IEEE Computer index rules. Math. Oper. Res. 38(4):682697.
Society, Washington, DC), 304311. Shen ZJM, Su X (2007) Customer behavior modeling in revenue
Bergemann D, Said M (2011) Dynamic auctions: A survey. Wiley management and auctions: A review and new research oppor-
Encyclopedia of Operations Research and Management Science (John tunities. Production Oper. Management 16(6):713728.
Wiley & Sons, Hoboken, NJ). Stokey NL (1979) Intertemporal price discrimination. Quart. J.
Bertsekas D (1999) Nonlinear Programming (Athena Scientific, Econom. 93(3):355371.
Nashua, NH). Stokey NL (1981) Rational expectations and durable goods pricing.
Bertsimas D, Thiele A (2006) A robust optimization approach to Bell J. Econom. 12(1):112128.
inventory theory. Oper. Res. 54(1):150168. Su X (2007) Intertemporal pricing with strategic customer behavior.
Besanko D, Winston WL (1990) Optimal price skimming by a Management Sci. 53(5):726741.
monopolist facing rational consumers. Management Sci. 36(5): Su X, Zhang F (2009) On the value of commitment and availability
555567. guarantees when selling to strategic consumers. Management
Bilton N (2012) Disruptions: Taxi supply and demand, priced by Sci. 55(5):713726.
the mile. New York Times (January 8), http://bits.blogs.nytimes Talluri K, van Ryzin GV (2004) Revenue management under a gen-
.com/2012/01/08/disruptions-taxi-supply-and-demand-priced eral discrete choice model of consumer behavior. Management
-by-the-mile/. Sci. 50(1):1533.
Bitran G, Caldentey R (2003) An overview of pricing models for Vitek J (2009) Windows Azure lessons learned: Dominos Pizza.
revenue management. Manufacturing Service Oper. Management MSDN Blog (November 17),
5(3):203229. archive/2009/11/17/domino-s-demonstrates-tomcat-site
Bitran GR, Mondschein SV (1997) Periodic pricing of seasonal prod- -on-windows-azure.aspx.
ucts in retailing. Management Sci. 43(1):6479. Wang R (1993) Auctions versus posted-price selling. Amer. Econom.
Cachon GP, Feldman P (2013) Price commitments with strategic Rev. 83(4):838851.
consumers: Why it can be optimal to discount more frequently Windows Azure (2012) Windows Azure Pricing. Accessed
than optimal. Working paper, University of Pennsylvania, June 1, 2013.
Philadelphia. calculator/.
Borgs, Candogan, Chayes, Lobel and Nazerzadeh: Optimal Multiperiod Pricing with Service Guarantees
Article submitted to Management Science; manuscript no. (Please, provide the mansucript number!) 1

A. Online Appendix of Optimal Multiperiod Pricing with Service Guarantees

A.1. Appendix to Section 3
Proof of Lemma 1 First note that we can add the constraint 0 p 1 to the problem without loss of optimality,
since customer valuations are bounded by 1. Consequently, it follows that for a given fixed ranking R, the set of
consistent and feasible prices defines a closed and bounded set. Since the objective function is continuous in prices
(for a fixed ranking), we conclude that optimal prices exist for any given ranking R. By maximizing over the finitely
many possible rankings, we conclude that an optimal solution of OPT-2 exists.
For the second claim, observe that if p is a feasible solution of OPT-1, then (p, RC (p)) is a feasible solution of OPT-
2 with the same objective value. Thus, the maximum of OPT-2 is an upper bound on the supremum of OPT-1. Given
an optimal solution (p? , R? ) of OPT-2, and any > 0, p? + R? is a feasible vector of prices that is consistent with the
ranking R? , and hence (p? + R? , R? ) is a feasible solution of OPT-2. This is because, if Rt? < Rt?0 , then p?t p?t0 , and
consequently p?t + Rt? < p?t0 + Rt?0 . Moreover, this inequality also implies that in p? + R? no price is repeated, and
hence the only consistent ranking with this price vector is R? . This implies that R? is the customer-preferred ranking
corresponding to p? + R? , and thus this price vector is feasible in OPT-1 with the same objective value. Since the
objective of OPT-2 is continuous in prices for a fixed ranking R? , the value of (p? + R? , R? ) approaches to that
of (p? , R? ), as goes to 0. Thus for > 0, p? + R? is a feasible solution of OPT-1, value of which converges to
maximum of OPT-2 as goes to 0. Since maximum of OPT-2 is an upper bound on the supremum of OPT-1, it follows
that these values are equal, and p? + R? converges to the supremum of OPT-1, as claimed.
If p is an optimal solution of OPT-1, then its value equals to the supremum value. However, as explained earlier this
value equals to the maximum of OPT-2, and (p, RC (p)) is a feasible solution of this problem with the same value.
Thus, the claim follows.
A.2. Appendix to Section 4
Proof of Proposition 1 Since the valuations are bounded by 1, it is not beneficial to set a price above 1. Now,
suppose (p, R) is a feasible and consistent price ranking. Let p0 be the price vector such that p0t = max{pM , pt }.
We claim that (p0 , R) is both consistent and feasible. For consistency, note that if Rt < Rt0 then pt pt0 . Hence,
max{pM , pt } max{pM , pt0 }. Therefore, (p0 , R) is consistent. Moreover, because we have (weakly) increased the
prices, it is a feasible solution. Finally, observe that the revenue obtained from (p0 , R) is at least equal to the revenue of
(p, R). The reason is t (R) does not change, but the uncapacitated revenue function, p(1 F (p)), increases. Namely,
pt (1 F (pt ))t (R) p0t (1 F (p0t ))t (R).
t t

since by definition pM maximizes p(1 F (p)). Therefore, starting from a feasible solution, we can construct another
one with weakly better objective value, where all prices are weakly above pM , thus the claim follows.
A.3. Appendix to Section 6
L EMMA 6. Let AL be the population matrix with elements aL
i,j and A be the population matrix with elements ai,j .

Then, OPT-4 is equivalent to:


max pt Dt (pt , R, AL )
p 0,R2P (T )
s.t. Dt (pt , R, AU ) cL
t for all t 2 {1, ..., T }

R t < R t 0 ) pt pt 0 for all t, t0 2 {1, ..., T },

Borgs, Candogan, Chayes, Lobel and Nazerzadeh: Optimal Multiperiod Pricing with Service Guarantees
2 Article submitted to Management Science; manuscript no. (Please, provide the mansucript number!)

Proof of Lemma 6: The function Dt (pt , R, A) is weakly increasing in all the elements of the matrix A. There-
fore, for all A 2 A, the tightest constraint among all of constraints of the form M t=1 pt Dt (pt , R, A) is the one
given by AL . At optimal solutions of OPT-4, M should be replaced by the maximum value it can attain, which is
p D (pt , R, AL ). Similarly, the tightest constraint among of the constraints of the form Dt (pt , R, A) ct is the
t=1 t t

one given by AU and cL

t . Thus, the claim follows replacing constraints of this form by Dt (pt , R, A ) ct .

Proof of Proposition 4: The constraint set of OPT-R is identical to that of an instance of OPT-2 with parameters
(c , A ). Additionally, the objective functions of both problems are nonincreasing for all p
pM . Since, Proposition
3 relied on the monotonicity of revenue in prices, and the properties of constraint sets, it follows that for OPT-R, a
set L with O(T 3 ) prices that contains all candidate optimal prices can be constructed (using parameters (cL , AU )).
Thus, we can still use the recursion in (7) to find the optimal sequence of prices. However, ij
k (p) needs to be modified
slightly since in OPT-R the feasibility constraints involve A , whereas, the revenue function involves AL . Therefore,

the recursion in (7) solves OPT-R (again in O(T 6 )), using the following modified definition of kij (p):
8 !
> Xk j 1
X P Pj 1 U
< a L
(1 F (p))p if
a (1 F (p)) cL
ij lm l=i+1 m=k lm k
k (p) =
l=i+1 m=k

1 otherwise.
Proof of Proposition 5 Let c , A denote the capacity and arrivals in a given problem instance. We will show that

V (c , A ) V ROB (C, A, c , A ) (2H + 1)P (AU ). Since, c , and A are arbitrary, the claim then follows from
taking supremum over all c and A .
Let VR (c, A, A ) denote the revenue obtained by (i) offering a price vector consistent with ranking R, (ii) ensuring
that prices are feasible for arrival matrix A and capacity vector c, (iii) having arrival realization A , i.e.,

VR (c, A, A ) = max pt Dt (pt , R, A )

p 0

s.t. Dt (pt , R, A) ct for all t 2 {1, ..., T }

pt pt0 if Rt < Rt0 for all t, t0 2 {1, ..., T }.

Note that imposing the constraint pt pt0 if Rt0 = Rt + 1 (for all t, t0 ) is equivalent to imposing the constraint pt
pt0 if Rt < Rt0 in the above optimization problem, due to the transitivity of the inequalities. Thus, we conclude

VR (c, A, A ) = max pt Dt (pt , R, A )

p 0
s.t. Dt (pt , R, A) ct for all t 2 {1, ..., T }

pt pt0 if Rt0 = Rt + 1 for all t, t0 2 {1, ..., T }.

Let t 0 denote the Lagrange multiplier corresponding to the capacity constraint associated with time t, and
t,t0 0 be the Lagrange multiplier associated with the ranking constraint pt pt0 , assuming Rt0 = Rt + 1. The KKT
conditions (see, for example, Bertsekas (1999)) imply that for all t, the optimal prices satisfy:
@Dt (pt , R, A ) @Dt (pt , R, A)
Dt (pt , R, A ) + pt t + t00 ,t t,t0 = 0, (17)
@pt @pt
where t, t0 , and t00 are such that Rt0 = Rt + 1 and Rt = Rt00 + 1. By the complementary slackness conditions, if two
prices pt and pt0 are different, then t,t0 = 0. Thus, summing the KKT conditions for all periods that have the same
price p (and noting that ranking of such periods are necessarily consecutive), t terms cancel, and we obtain:
X @Dt (p, R, A ) @Dt (p, R, A)
Dt (p, R, A ) + p t = 0.
t : p =p
@p @p
Borgs, Candogan, Chayes, Lobel and Nazerzadeh: Optimal Multiperiod Pricing with Service Guarantees
Article submitted to Management Science; manuscript no. (Please, provide the mansucript number!) 3

By definition Dt (p, R, A) = t (R, A)(1 F (p)), where t (R, A) is the R-induced potential demand for population
matrix A. Hence, using the notation F 0 (p) = f (p), we obtain
[t (R, A )(1 F (p)) p t (R, A )f (p) + (R, A)f (p)] = 0.
t t
t : p t =p

Rearranging terms, this equation leads to

1 F (p)
t (R, A) t = t (R, A ) p t (R, A ),
t : p t =p t : p t =p
f (p) t : p =p t

1 F ( p)
where the inequality follows from the fact that optimal prices are bounded by 1, and f (p)
0. Thus, summing the
above equality over all periods t (or all different price levels p that appear in an optimal solution) we obtain

t (R, A) t t (R, A ) = P (A ) P (AU ), (18)

t=1 t=1
where P (A) = i,j
ai,j . By the complementary slackness conditions, ct = Dt (pt , R, A) = t (R, A)(1 F (pt ))
t (R, A) whenever the Lagrange multiplier t 6= 0. Hence, the above inequality also implies

ct t P (AU ). (19)

We next consider how VR (c, A, A ) changes as c increases and A decreases. The Envelope Theorem (see Kimball
@VR (c,A,A ) @VR (c,A,A )
(1952)) suggests that the derivatives @ct
and @ai,j
are equal to
@VR (c, A, A ) @VR (c, A, A )
= t and = t0 (i,j,R) (1 F (pt0 (i,j,R) )), (20)
@ct @ai,j
where t0 (i, j, R) represents the period t0 that has minimum ranking in R within {i, ..., j}, i.e., the time period popula-
tion ai,j receives service.
cU aU
Observe that by definition VR is increasing in c and decreasing in A. Since t
and i,j
1 + , it follows that
t i,j

0 VR (c , A , A ) VR (cL , AU , A ) VR (cU , AL , A ) VR (cL , AU , A )

VR ((1 + )cL , AL , A ) VR (cL , AL (1 + ), A )
Using the Fundamental Theorem of Calculus (and the notation gR (x) = VR ((1 + x)cL , AL , A )) it follows that
dgR (x) @VR
0 VR ((1 + )cL , A , A ) VR (cL , A , A ) = dx = ((1 + x)cL , AL , A )cL
t dx
x=0 dx x=0 t=1 @ct
((1 + x)cL , AL , A )(1 + x)cL
t dx.
x=0 t=1 @ct

Observing from (20) that @VR

((1 + x)cL , AL , A ) equals to the Lagrange multiplier t for the problem instance with
capacity vector (1 + x)cL , and using (19) and (22), we obtain
VR ((1 + )cL , AL , A ) VR (cL , AL , A ) P (AU )dx = P (AU ). (23)

Following a similar approach, we also obtain

0 VR (cL , AL , A ) VR (cL , (1 + )AL , A ) = (cL , (1 + x)AL , A )ai,j dx
x=0 i,j @ai,j
Z X @VR (24)
(cL , (1 + x)AL , A )(1 + x)ai,j dx
x=0 i,j @ai,j
Borgs, Candogan, Chayes, Lobel and Nazerzadeh: Optimal Multiperiod Pricing with Service Guarantees
4 Article submitted to Management Science; manuscript no. (Please, provide the mansucript number!)

Using (20), it follows that @VR

(cL , (1 + x)AL , A ) = t0 (i,j,R) (1 F (pt0 (i,j,R) )) t0 (i,j,R) , where t denotes the
Lagrange multiplier in a problem instance with parameters cL , (1 + x)AL , A . Thus, using (24) and noting from the
definition of t0 (i, j, R) that t (R, A) = i,j :t0 (i,j,R)=t ai,j , we obtain,
VR (cL , A , A ) L
VR (cL , (1 + )A , A )
(R, AL (1 + x))dx.
t t (25)
x=0 t=1

Thus it follows from (18) that

VR (cL , AL , A ) VR (cL , (1 + )AL , A ) P (AU )dx = P (AU ). (26)

Adding (23) and (26), and using it in the right hand side of (21) it follows that

VR (c , A , A ) VR (cL , AU , A ) 2P (AU ). (27)

Note that by linearity of the objective of (16) in its third argument, and the fact that aL
i,j ai,j ai,j (1 + )ai,j ,

it follows that VR (cL , AU , A ) VR (cL , AU , AL )(1 + ). On the other hand, since maximum price customers can pay
for service is 1, it follows from the definition of P (A) that VR (cL , AU , AL ) P (AL ) P (AU ). Thus, we conclude
VR (cL , AU , A ) VR (cL , AU , AL ) P (AU ). Combining this with (27) we obtain

VR (c , A , A ) VR (cL , AU , AL ) + 3P (AU ). (28)

Maximizing both sides of this inequality over R and noting that maxR VR (c , A , A ) = V (c , A ), we conclude
V (c , A ) maxR VR (cL , AU , AL ) + 3P (AU ). Note that by definition maxR VR (cL , AU , AL ) equals the solution
of OPT-R and V ROB (C, A, c , A ) is larger than this solution (OPT-R gives the worst case profits for optimal prices
that are feasible for all capacities in C, and arrivals in A, whereas V ROB (C, A, c , A ) is the realized profit). Thus, we
conclude V (c , A ) V ROB (C, A, c , A ) + 3P (AU ), and the claim follows.

A.4. Appendix to Section 7

Proof of Lemma 2 Let p? and R? denote an optimal solution of OPT-5. Observe that for all t, the set P \ [p?t , p?t +
) contains a single element. Denote this element by pt .
We first show that p is consistent with ranking R? . Note that if Rt? < Rt?0 then p?t0 p?t . Moreover, since we have
p?t0 + p?t + , and pk is characterized by intersection of [p?k , p?k + ) with P for all k, it follows that pt0 pt , and
hence the consistency claim.
By Assumption 2, ht (p, R? ) is decreasing in p, for any R. Therefore, ({pt }, R? ) is a feasible solution of OPT-5. By
Assumption 2 again, and the fact that pt 2 [p?t , p?t + ) for all t, it follows that
v= gt (p?t , R? ) gt (pt , R? ) + lt . (29)
t t

On the other hand, by construction pt 2 P for all t, thus ({pt }, R? ) is a feasible solution of OPT-6. Hence v
g (pt , R? ), and together with (29), this implies that v v t lt .
t t

Proof of Lemma 3 Construction of optimal prices and ranking, using the dynamic programming recursion in (10)
is identical to the construction given in Theorem 3, and is omitted. In the rest of the proof we characterize the compu-
tational complexity of this construction.
Borgs, Candogan, Chayes, Lobel and Nazerzadeh: Optimal Multiperiod Pricing with Service Guarantees
Article submitted to Management Science; manuscript no. (Please, provide the mansucript number!) 5

In order to solve the recursion in (10) we compute all values of !(i, j, p) by solving O(T 2 |P |) subproblems. At
each step of the recursion we solve for the optimal k and p. Finding these requires at most O(T |P |) trials. Given
a value of p and k, we need to evaluate kij (p). This requires checking if constraints are satisfied in the subproblem
(hence computing hk (p, R)), and evaluating the corresponding objective value (gk (p, R)) in the relevant subproblem.
Thus, computation of kij (p) can be completed in O(s(T )) time, and the overall complexity is O T s2(T ) .

A.5. Appendix to Section 8

Proof of Lemma 4: Let S1 be a revenue maximizing vector in S(i, j, p, k). Suppose there exits a price vector
S2 2 S(i, j, p, k) and period t such that pt (S2 ) = p < pt (S1 ). We show that no such price vector exists, hence proving
the lemma.
Define S 0 to be the price vector such that
p pt (S1 ) = p or pt (S2 ) = p
pt (S 0 ) =
pt (S1 ) Otherwise

To prove the claim, first consider the assignment of the populations to the periods when the price vector is S1 . Let A1
be the set of periods that have price p under S1 and A2 be the set of such periods under S2 . Note that in S 0 , we update
S1 by decreasing the prices of periods in A2 \ A1 to p. Observe that the price change only matters for populations that
are present in the system in a period in A2 \ A1 , but not A2 \ A1 , since it is possible to schedule all the remaining
populations exactly as we did under S1 . Since S2 2 S(i, j, p, k) is feasible, it follows that for S 0 , there exists a feasible
schedule that assigns all populations that are present in a time instant in A2 \ A1 , but not A2 \ A1 , to time instants in
A2 \ A1 , even when the price offered at these periods equals to p. Moreover, because p pM , reducing the prices can
only increase the revenue, implying that S 0 leads to higher revenues than S1 , and contradicting with the assumption
that S1 is a revenue maximizer. Hence, the lemma follows.
Proof of Lemma 5: We say that interval [l, k] satisfies the minimum requirements for having price p if this
interval can serve all customers who can receive service only in this interval at price p. Namely, u=l cu (1
Pk Pk
F (p)) u=l v=u auv . This is a necessary condition for time instants in this interval to have price equal to p.
We claim that the algorithm described in Figure 6 finds set L in polynomial time. To prove the correctness of
the algorithm, first observe that if interval [i + 1, k] satisfies the minimum requirements, then we can serve all the
populations who arrive at the system after time i, and can wait up to (and including time) k, to this interval at price
p. Hence, L = {i + 1, i + 2, , k}. We also show at step 3 of the algorithm that if interval [`, k] does not satisfy the
minimum requirements, then ` does not belong to L(i, j, p, k). To show this consider the largest ` such that [`, k] does
not satisfy the minimum requirements, and assume that ` belongs to L. Note that because we chose the largest such `,
all the populations that arrive at time ` or after that need to be scheduled to periods in [`, k] for service. However, this
period does not satisfy the minimum requirements, and we obtain a contradiction.
Finally, we note that the algorithm is polynomial; the number of recursions is bounded by k i, and we can verify
the minimum requirements in polynomial time.
Borgs, Candogan, Chayes, Lobel and Nazerzadeh: Optimal Multiperiod Pricing with Service Guarantees
6 Article submitted to Management Science; manuscript no. (Please, provide the mansucript number!)


1. If interval [k, k] does not satisfy the minimum requirements:

It is infeasible to assign time instant k price p.
Terminate and Return S = ;.
2. If interval [i + 1, k] satisfies the minimum requirements.
Terminate and Return S = {i + 1, , k }.
3. Let ` be the largest element in [i + 1, k] such that [`, k] does not satisfy the minimum
(Time instant ` cannot receive price p)
(a) Define a new problem instance where period ` is removed, and search for the maximal
set which receives price p in this new problem instance.
(b) Label periods in the new problem as i0 , (i + 1)0 , . . . (k 1)0 . Construct a population
matrix A0 = {a0i,j } and capacity vector c0 for the new problem.
> ai0 ,j 0 j0 < ` 1
<ai0 ,j 0 + ai0 ,` i0 < `, j 0 = ` 1
a0i0 ,j 0 = ai0 ,j 0 +1 i0 < `, j 0 > ` 1
> a`,j 0 +1 + a`+1,j 0 +1 i0 = `
ai0 +1,j 0 +1 ` < i0
c0j j0 < `
c0j 0 =
cj 0 +1 ` j0 < k

(c) Let S 0 = F indL(i0 , k 0 , p, A0 , c0 )

i. If S 0 = ;
It is infeasible to assign price p to any subset of [i, k].
Terminate and Return S = ;.
ii. If S 0 6= ;
Construct the solution from S 0 .

S = {s|s 2 S 0 and s < `, or s l and s 1 2 S 0 }.

Terminate and Return S.
Figure 6 Algorithm FindL(i,k,p,A,c): Finds the maximal subset of (i, k] (excluding time i but including time k)
which can receive price equal to p , where the population matrix and capacities are given by A and c.