We study optimal pricing in a single server queue when the customers valuation of service depends on their waiting time. In particular, we consider a very general model, where the customer valuations are random and are sampled from a distribution that depends on the queue length. The goal of the service provider is to set dynamic state dependent prices in order to maximize its revenue, while also managing congestion. We model the problem as a Markov decision process and present structural results on the optimal policy. We also present an algorithm to find an approximate optimal policy. We further present a myopic policy that is easy to evaluate and present bounds on its performance. We finally illustrate the quality of our approximate solution and the myopic solution using numerical simulations.
1 Introduction
The problem of allocating limited resources amongst competing users is traditionally studied using queueing theory and stochastic networks, with a focus on metrics such as throughput and delay. Users and system operators in real world respond to prices and incentives in addition to the queueing metrics. The focus of this article is to study such interplay of queueing and pricing. A user has certain valuation for the service, and the user enters the system only if the price is smaller than their valuation. However, the valuation of the user can decrease based on quality of experience, such as delay and other queueing metrics. The service provider, therefore, employs dynamic pricing and tunes prices in order to use them as control levers for admission control and to manage congestion in the system. Moreover, the prices affect the service provider’s revenue, which they want to maximize. Examples of such systems include spot pricing for the cloud services [28], surge pricing in ridehailing systems such as Uber [8], auctions for web advertisements [15], and highway tolls with increased peak-hour rates [18].
In this article, we abstract away the resources and model them by a single server queue. Several pricing questions in such a simple system are still unanswered in the literature. Such a system can be used to model the aggregate service capacity of a shared service system such as a data center. Customers arrive into this system according to a Poisson process and see the queue length ahead. A customer’s valuation of the service depends on the queue length. Given the queue length, the customer samples its valuation in an independent and identically distributed (i.i.d.) manner from a distribution. The service provider posts a price at each time that depends on the current queue length. If the customer’s valuation is larger than the price, she pays the price, joins the queue, and leaves only after service completion; otherwise, she does not join the queue. All customers in the queue are served in an FCFS manner, and each customer needs an i.i.d. exponential amount of service. The service provider is not aware of the individual valuations of the customers, but knows the distributions of the queue length dependent valuations. We study the problem from the perspective of the service provider, aiming to maximize long run average revenue.
The following are the main contributions of our work.
–
We formulate the revenue maximization problem as a continuous time Markov decision process (MDP) and present structural results on the optimal pricing policy and the optimal revenue.
–
Since a queue has infinite states, finding the exact solution of this MDP involves solving an infinite system of equations. By truncating the state space appropriately, we propose a finite system of equations to obtain an approximate solution and present its structural properties.
–
We present an algorithm to find the approximate solution, and present guarantees on its convergence. We arrive at this algorithm following intricate arguments that translate Bellman’s equation to a fixed point equation in single variable, namely the optimal revenue rate. Counterintuitively, the optimal prices may be non-monotonic in queue lengths.
–
We consider a special case when the valuations are a deterministic function of the queue lengths and present explicit closed form expression for the optimal dynamic prices. Such deterministic valuations were considered in prior work [5, 7] which focused on only specific functions. Our model is therefore, more general, since we not only consider general deterministic valuations, but also allow randomized valuations.
–
We present a myopic policy, prices of which are much simpler to obtain. We show that even such a simple policy has good performance, and present a lower bound on the ratio of average revenue rates under the myopic policy and the optimal policy. We further present numerical simulations to compare the two policies. The simulations also illustrate rich tradeoffs that occur due to interplay of customers’ preference towards smaller delay and the provider’s goals of maximizing revenue and managing congestion.
1.1 Related Work
Since the queue pricing problem arises naturally in a number of situations, it has been studied extensively. The earliest work on pricing in a queue is the seminal paper by Naor [22], in which the entry of customers to a queue was regulated using tolls. After observing the queue size, customers decide to join (or not join) the queue. Such systems are called observable. Customers join if the difference between their valuation of the job and the cost of waiting exceeds the admission price to the queue. This is equivalent to a threshold type policy—if the queue length is smaller than the threshold, the customers join; otherwise they leave. The optimal threshold varies, depending on whether we are maximizing the social utility or the revenue. In [22], it was demonstrated that the socially optimal threshold was higher than the revenue maximization threshold. A number of subsequent works have looked at the queue pricing problem under different constraints and assumptions, and study the effects of different system parameters. An exhaustive survey of this literature is available in [13, 14].
We are interested specifically in the question of maximization of revenue. In [20], the author studies optimal pricing for an \(M/M/s\) queue with finite waiting room. The author shows that the optimal prices are monotonically increasing in the number of customers waiting in the system. Such structural results for the optimal prices are a common result in a number of works. For example, [23] shows that optimal price is monotonically increasing in the number of customers, for a multi server system with no waiting room. Proceeding further along these lines, [7] is interested in maximizing the expected discounted revenue, while keeping the queueing model of [22]. They obtain a revenue optimizing threshold queue length beyond which entries are not allowed into the queue. This threshold can be computed numerically. All customers who see a waiting queue length smaller than this threshold pay a price equal to the difference between their valuation and waiting cost. In [5], an explicit form is derived for the threshold obtained in the previous work, and they characterize the earning rate asymptotically. However, both aforementioned works provide explicit solutions in the case of fixed service valuation (or simple valuation distributions, such as a valuation which takes two values). They do not provide explicit solutions for valuations with continuous support and general distributions.
In [30], the authors look at optimal pricing in finite server queueing systems, but restrict themselves to the sub-optimal class of static prices. These prices are not dependent on the system state (queue length). They study the variation of optimal price with number of servers. While static prices are sub-optimal, they can be close to optimal in some systems. For example, in [4], the authors prove the existence of a static pricing policy that obtains \(78.9\%\) of the optimal profit in a system with multiple reusable resources. This result holds under the assumption that the revenue rate is a concave function of the arrival rate. In [21], the authors obtain the revenue maximizing policy for a queueing system under the assumption that the generalized hazard rate of the valuation distribution, which is the ratio of price times the density of valuation, to the complementary cumulative distribution of valuation is strictly increasing in price. This assumption, however, does not hold in general. There are a number of similar works that provide existential results and structural results for optimal policies, after modelling the revenue maximization problem as a Markov decision problem (MDP) [10, 27, 29]. There are also many works that study the pricing problem in queues in different asymptotic regimes. In [1], an asymptotically optimal price is obtained for customers with fixed valuations. In [16], the authors show that the revenue loss due to randomness is lower for dynamic pricing than static pricing. An approximately optimal price is obtained by solving a diffusion equation in [6].
In our earlier works [17, 25], we study optimal service pricing in multi-server systems in which the service provider charges a time varying service fee aiming at maximizing its revenue rate. The customers that find free servers and service fees lesser than their valuation join for the service else they leave without waiting. We solve the optimal pricing problems using the framework of MDPs and show that the optimal prices depend on the number of free servers. We propose algorithms to compute the optimal prices. We also establish several properties of the optimal prices and the corresponding revenue rates in the case of Poisson customer arrivals.
In this article, we proceed to model the control problem as an MDP and proceed to solve it. We are interested in the optimal prices themselves, rather than structural properties, as in [10, 27, 29]. We obtain the optimal solution as the solution of a fixed point equation. In all the works discussed, the optimal prices are obtained under restrictive assumptions [4, 21], or assume non-random (or less random) service valuations [5, 7]. We also assume a more general structure for the valuation distribution, which can be viewed as a generalization of valuation models such as in [22] and [19]. Hence, the results in this article are more general.
1.2 Organization of the Article
The rest of this article is organized as follows. In Section 2, we describe the system model, formulate the problem as an MDP, and write the Bellman’s equations corresponding to the optimal policy. In Section 3, we solve the Bellman’s equations after truncating the equations. In Section 4, we obtain a myopic policy that is easy to implement and serves as a benchmark for comparison. In Section 5, we solve the Bellman’s equations for the special case when the valuation function is deterministic. Numerical simulations are presented in Section 6. For ease of reading, we have moved some of the lengthier proofs into a separate section, Section 7.
2 System Model and Preliminaries
We begin with description of the system model, specifically of the stochastic valuation functions. Subsequently, we frame the optimal pricing problem as a continuous time Markov decision problem (CTMDP). We write the Bellman’s equations and derive several properties.
2.1 System Model
We model the system as a single server queue with infinite buffer size. Customers arrive as a Poisson process of rate \(\lambda\). Each arriving customer has a random service time requirement S. We will assume that the service requirements of different customers are i.i.d. exponential with mean \(\frac{1}{\mu }\). The admission price is a function of the number of customers present in the system. We model the service valuation as a non-negative quantity, stochastically decreasing with the number of waiting customers, motivated by the fact that the mean waiting time of a customer is directly proportional to the number of waiting customers. As such, two customers who see the same queue length on arrival will have i.i.d. valuations. If two customers see queue lengths \(i_1\) and \(i_2\) on arrival, where \(i_1\gt i_2\), their valuations \(V_{i_1}\) and \(V_{i_2}\) will be stochastically ordered, i.e.,
Condition (1) implies that on average, customers seeing a higher queue length on arrival will have lower valuation. Let \(G_i(\cdot)\) denote the value distribution for a customer seeing i customers on arrival, i.e., \(G_i(x)=\mathbb {P}[V_i\leqslant x]\). We denote the complementary valuation distributions by \(\overline{G}_i \triangleq 1-G_i\) for each i. We have succinctly represented our system model in Figure 1.
Fig. 1.
We assume that the service provider selects a deterministic service fee of \(u_i\) for an arriving customer that sees i queued customers at its arrival. If the valuation \(V_i\) exceeds this fee, then the customer pays this service fee \(u_i\) to the provider and joins the system; else he leaves without joining the system. That is, given that an incoming arrival sees i queued customers, its probability of joining the system is \(\overline{G}_i(u_i)\) independent of everything else. We denote the sequence of prices as \(\mathbf {u}= (u_0, u_1, \dots)\). Our objective is to find the stationary price sequence that maximizes the long term average revenue generated by the service provider.
2.1.1 Discussion on the Valuation Model.
The valuation model we present here subsumes as a special case other waiting cost models studied in the literature, as the following examples illustrate.
2.2 MDP Formulation
We formulate the pricing problem as a CTMDP [3, chapter 5]. The state of the CTMDP is the number of customers \(Q(t)\) present in the system at time t. This state evolves over the state space \(\mathcal {Q}\triangleq \left\lbrace 0,1,2,3,\dots \right\rbrace\). The action in each state i is to pick the corresponding price \(u_i\), and we seek to find a stationary state dependent policy \(\mathbf {u}=(u_i,i\in \mathcal {Q})\). The state of the system evolves in a Markovian manner in continuous time. The transition rates from state i to \(i+1\) are
Customers pay the price on arrival. Therefore, in state i, the provider receives a reward of \(u_i\) if the next transition is due to an arrival that accepts the price, and does not receive any reward if the transition is due to a departure. Thus, on average, the reward \(g(i,u_i)\) received in state i under an action \(u_i\) is given by
\begin{equation} g(i,u_i) = {\left\lbrace \begin{array}{ll} u_0 & \text{ if } i = 0,\\ \frac{\lambda u_i\bar{G}_i(u_i)}{\nu _i(u_i)} & \text{ if } i \geqslant 1. \end{array}\right.} \end{equation}
(6)
Following the discussion in [3], we observe that the long term average revenue is equal to
where \(t_N\) is the completion time of the Nth transition of the Markov chain. We formally state our problem in the following.
2.3 Bellman’s Equations
First, we obtain Bellman’s equations corresponding to the average reward CTMDP problem defined in Problem 1, and rewrite them as a series of iterative equations. Finding the optimal solution for Problem 1 is equivalent to solving the following set of Bellman’s equations
From the definition of transition rates \(\nu _i(u)\) defined in (4) and the fact that \(\bar{G}_i(u_i) \leqslant 1\), it follows that the transitions rates are uniformly bounded for all states \(i \in \mathcal {Q}\) and prices \(\mathbf {u}:\mathcal {Q}\rightarrow \mathbb {R}_+\), where
By scaling the transition rates with \(\Lambda\), we can uniformize the controlled CTMC. The state transition probabilities are redefined to allow self-transitions such that the resulting dynamics remains unchanged. Specifically, we define the transition probabilities for the jump chain of uniformized controlled CTMC as
We define the following functions that help us understand the properties of optimal pricing and revenue rate for Problem 2. For all \(i \in \mathcal {Q}\), we define
The terms \(u_i(B)\) and \(m_i(B)\) can be interpreted as follows. Suppose the system is in state i (i.e., there are i customers). For any price u that is set, the instantaneous revenue at the next arrival is \(u\bar{G}_i(u)\). Suppose B is the “cost” that captures reduced valuation of the future users due to increase in queue length because of admitting this user, then the expected instantaneous “profit” is \((u-B)\bar{G}_i(u)\), and \(u_i(B)\) is the price that maximizes this expected instantaneous profit and \(m_i(B)\) is the corresponding profit. We note that the cost B is a priori unknown but is yielded by the analysis in Section 3.
Substituting the values of \(\tilde{p}_{ij}(u)\) from (11), the reduced Bellman’s Equation (14) can be rewritten in terms of the functions \(m_i(\cdot)\)’s defined in (16) and scaled differences \(\Delta (\cdot)\)’s defined in (15), as
We show that the valuation distribution \(G_i\)’s being stochastically ordered implies that there exists a distribution function G which is the limit of \(G_i\)’s at all continuity points. Thus, the valuations \(V_i\)’s converge in distribution to a random variable V with the distribution G.
Let \(C \subset [0, \infty)\) be the set of non-negative continuity points of \(G(\cdot)\), and define
We have defined the limiting value \(m(B)\) and limiting price \(u(B)\) assuming that the maximization in (20) has a solution for each B. It can be shown that \(\lim _{i \rightarrow \infty } m_i(B) \geqslant m(B)\) for all B. It is not clear whether \(m_i(B)\) and \(u_i(B)\) converge to \(m(B)\) and \(u(B),\) respectively. However, it is not hard to imagine that these sequences converge under certain smoothness assumptions. We summarize these assumptions below.
This assumption is used in Lemma 30 to argue that if the original problem is non-degenerate, then the k-truncated problems are also non-degenerate for large k.
We have an infinite system of equations from Bellman’s equation for the problem. The properties of optimal pricing sequences depend on the limiting valuation of the customer that sees an arbitrary large number of existing customers. Thus, we need to understand the limit of valuations \(V_i\)’s and the distributions \(G_i\)’s. This is the focus of the first subsection. In the next subsection, we propose a truncation method to obtain an approximate solution. We show that as the truncation threshold gets larger, we converge to the optimal solution to Problem 2 with an infinite system of equations.
3.1 Existence of the Optimal Solution
Degeneracy.
Before proceeding further, we discuss a special case. Recall that \(\mathbf {u}^\ast\) is the optimal ergodic policy and \(\theta ^\ast \triangleq R(\mathbf {u}^\ast)\) is the maximum revenue rate across all the ergodic policies. Let \(\theta ^\ast \leqslant \lambda m(0) = \lambda \max _u u \bar{G}(u)\). The policy \(\mathbf {u}= (u(0),u(0),\dots)\) achieves a revenue rate \(R(\mathbf {u}) \geqslant \lambda u(0) \bar{G}(u(0))\). The inequality follows from the fact that \((V_i: i \in \mathcal {Q})\) are stochastically decreasing. This result implies that the policy \(\mathbf {u}\) achieves a revenue rate greater than or equal to the maximum revenue rate \(\theta ^\ast\) across all ergodic policies. In this case, we call our pricing problem degenerate [11]. As in [11], we will bypass the degenerate case, i.e., we will focus on the case \(\theta ^\ast \gt \lambda m(0)\).
3.2 Truncated Problems
We recall the reduced Bellman’s Equations (18) and (19), which characterize the optimal price sequence for the revenue rate maximization problem. This being an infinite system of equations is not amenable to numerical computation. We address this issue by considering the corresponding k-truncated systems defined as follows.
In the following, we derive properties of the optimal pricing sequences and the corresponding revenue rates for the k-truncated problem. We also prove that as the truncation threshold k increases, the corresponding optimal pricing sequences when applied to the original system yield revenue rates arbitrarily close to the optimal revenue rate \(\theta ^\ast\). In Section 3.4, we develop an algorithm to derive the numerical solution to a k-truncated problem.
For notational convenience, we define functions \(\delta _i:\mathbb {R}_+\rightarrow \mathbb {R}\) for all \(\theta \in \mathbb {R}_+\) and \(i \in \mathcal {Q}\), as
Furthermore, the desired fixed point equation for the k-truncated system reduces to \(\delta _k(\theta) = 0\). We characterize the solutions to these equations for all thresholds k in Proposition 9. However, it requires the following assumption.
This assumption is used in Proposition 9 and in Theorem 15 to establish that the k-truncated problems posses unique optimal revenue rates, \(\theta _k\)s, which are non-increasing in k. Consequently, all the results that require Proposition 9 rely on this assumption. It is also used in Lemma 30. This assumption is also satisfied by many valuation sequences. Following is one such instance.
The following theorem characterizes the optimal revenue rate as well as an optimal price sequence for the non-degenerate k-truncated problem.
3.3 Convergence to an Optimal Solution
We now show that the optimal solutions to the k-truncated problem converge monotonically to an optimal solution to the original pricing problem, as the threshold \(k \rightarrow \infty\). We also obtain a bound on the performance (i.e., the revenue rates) of the optimal policy for a k-truncated problem. First, we show that as the truncation threshold k increases, the corresponding optimal pricing sequences \((\mathbf {u}^k: k \in \mathcal {Q})\) when applied to the original system yield revenue rates \(R(\mathbf {u}^k)\) arbitrarily close to the optimal revenue rate \(\theta ^\ast\).
If \(\theta _\infty \leqslant \lambda m(0)\), then \(\theta ^\ast \leqslant \lambda m(0)\), rendering the original pricing problem degenerate. We dispense with this case and focus on the case \(\theta _\infty \gt \lambda m(0)\).
Recall that the optimal policies for the k-truncated problems converge to \(\mathbf {u}^\ast = (u^\ast _i, i \in \mathcal {Q})\). In general, the asymptotic policy \(\mathbf {u}^\ast\) needs not be optimal for the original problem. The following corollary to Theorem 11 states that if \(\lim _{k \rightarrow \infty } \bar{G}_i(u^k_i) = \bar{G}_i(u^\ast _i)\) for all \(i \in \mathcal {Q},\) then \(\mathbf {u}^\ast\) is an optimal policy for the original problem. We know from Remark 7 that, in general, \(\lim _{k \rightarrow \infty } \bar{G}_i(u^k_i) \geqslant \bar{G}_i(u^\ast _i)\).
3.4 Solution to the k-Truncated Problem
Under Assumption 3, we can iteratively obtain the revenue rates \(\theta _k\) and the optimal policies \(\mathbf {u}^k\) for the k-truncated problems for all \(k \in \mathcal {Q}\), as shown in Proposition 9. It warrants solving (58) for each threshold k (see Section 7.3), which is a considerable computational task. In particular, this equation resembles an intricate fixed point equation which can be solved for \(\theta _k\) via iterative methods. In this section, we develop a fixed point iteration that yields \(\theta _k\) for any given k. Note that, from Theorem 11 and Corollary 12, \(R(\mathbf {u}^k)\) and \(\mathbf {u}^k\) are close to the optimal revenue rate and the optimal policy, respectively, for large k.
Recall that the optimal price sequence for the k-truncated system is characterized by \(k+2\) equations, namely (18) for \(i = 0\), (19) for \(i \in [k]\), and \(\Delta (k) = \Delta (k-1)\). We claim the following property of any solution to these equations.
We again express \(\Delta (i)\)’s successively in terms of \(\theta\) and treat these \(k+2\) equations as a fixed point equation in \(\theta\). In particular, we define functions \(\psi _i:\mathbb {R}_+\rightarrow \mathbb {R}\) as follows. Let \(\psi _k(\theta)\) be a solution of the following fixed point equation in \(\Delta :\)
for \(i = 0,1,\ldots ,k-1\). Then, \(\Delta (i) = \psi _i(\theta)\) for all \(i = 0,1,\ldots ,k\) and the above system of equations reduces to \(\lambda m_0(\psi _0(\theta)) = \theta\). We first prove certain properties of (30) before proposing a fixed point iteration and analyzing it in Proposition 15. We can define \(f_k(\Delta)\) to rewrite (30) as
Since \(m_k(\Delta)\) is non-negative, non-increasing, continuous, and convex in \(\Delta\), \(f_k(\Delta)\) is non-decreasing, continuous, and concave in \(\Delta\). Moreover, by definition, \(f_k(\Delta) \leqslant \theta /\mu\) for all \(\Delta\). Hence, for a given \(\theta\), the necessary and sufficient condition for existence of a fixed point is \(f_k(\Delta) \geqslant \Delta\) for some \(\Delta\), or alternatively, \(\lambda m_k(\Delta) + \mu \Delta \leqslant \theta\) for some \(\Delta\). So, defining
we see that (32) has a solution for all \(\theta \geqslant {\underline{\theta }}\) and does not have a solution for \(\theta \lt {\underline{\theta }}\). Observe that \({\underline{\theta }} \leqslant \lambda m_k(0)\). The following proposition summarizes the properties of the fixed points of (32) for \(\theta \geqslant {\underline{\theta }}\).
Notice that if \(\rho \bar{G}_k(0) \gt 1\), then (32) has two solutions, \(D_1(\theta)\) and \(D_2(\theta)\). However, \(D_1(\theta) = D_2(\theta)\) for \(\theta = {\underline{\theta }}\); we refer to the common value as \(\Delta ({\underline{\theta }})\). We set
For this setting, we establish existence and uniqueness of fixed point of \(\theta = \lambda m_0(\psi _0(\theta))\) in Theorem 15. As a part of the proof of the theorem, we also show that this fixed point equation does not have a solution for other potential choices of \(\psi _k(\theta)\). As in Section 2.2, we let \(\theta _k\) denote the unique solution of \(\lambda m_0(\psi _0(\theta)) = \theta\). Recall that if \(\theta _k \le \lambda m_k(0)\), then the policy \(u = (u_k(0),u_k(0),\dots)\) provides at least as much revenue rate as any ergodic policy. We will distinguish this degenerate case and the case \(\theta _k \gt \lambda m_k(0)\) in what follows. Note that if \(\theta \gt \lambda m_k(0)\), then (32) has unique positive root which is increasing in \(\theta\); it is \(D(\theta)\) if \(\rho \bar{G}_k(0) \lt 1\) and \(D_2(\theta)\) if \(\rho \bar{G}_k(0) \ge 1\). We present a fixed point iteration that solves \(\lambda m_0(\psi _0(\theta)) = \theta\) when \(\psi _k(\theta) = D(\theta)\) or \(\psi _k(\theta) = D_2(\theta)\). In particular, the algorithm yields \(\theta _k\) whenever it exceeds \(\lambda m_k(0)\).
3.4.1 An Iterative Algorithm.
Algorithm 1, displayed below, is an iterative algorithm that generates two sequences \({\underline{\theta }}_n, n \geqslant 0\) and \(\bar{\theta }_n, n \geqslant 0\) starting with \({\underline{\theta }}_0 = \lambda m(0)\) and \(\bar{\theta }_0 = \lambda m_0(\psi _0(\lambda m(0)))\), respectively.
4 A Myopic Policy
We now proceed to construct a simple myopic policy that is easy to implement. This will serve as a performance benchmark and help us characterize the loss of revenue due to sub-optimality. Towards this, we first have the following result about the stability of the Markov chain \((Q(t),t\geqslant 0)\) introduced in Section 2. All the results in this section are subject to Assumption 1. Let the load factor be defined as
This is a standard result from the theory of Markov chains [12, Section 6.11]. From this result, we deduce the following result regarding the stability of our model.
For any price vector \(\mathbf {u}\) that results in the Markov chain \((Q(t),t\geqslant 0)\) positive recurrent, we can rewrite the revenue rate in the following form.
Following Proposition 19, we can rewrite the optimal pricing problem defined in (8) as
Obtaining an analytical solution to the optimal pricing problem in (39) is difficult in general because the stationary distribution has a complex structure, as seen in (36). Define the myopic pricing vector \(\tilde{\mathbf {u}}=(\tilde{u}_0,\tilde{u}_1,\tilde{u}_2,\dots)\) such that per state price is the maximizer of the expected revenue in each state. That is,
While this pricing vector is sub-optimal, we obtain the following bound that shows that the revenue \(R(\tilde{\mathbf {u}})\) under myopic pricing is greater than a fixed fraction of the optimal revenue \(R(\mathbf {p}^\ast)\), for all values of \(\rho\) for which a stationary distribution exists. Following the bound, we show by numerical examples that in some cases, the fraction of revenue of the optimal that can be obtained by the myopic policy is substantial. We observe that the lower bound on the revenue fraction depends only on the myopic policy, and hence, is easily computable.
The myopic pricing scheme is guaranteed to yield at least a positive fraction of the optimal revenue rate. Note that this lower bound was obtained by upper bounding the optimal revenue rate \(R(\mathbf {u}^\ast)\), by using the inequality \(u_{i}^\ast \overline{G}_{i}(u_{i}^\ast) \leqslant \tilde{u}_0\overline{G}_0(\tilde{u}_0).\) Hence, the bound being close to zero may not indicate that the myopic policy is bad, but that the gap between \(u_{i}^\ast \overline{G}_{i}(u_{i}^\ast)\) and \(\tilde{u}_0\overline{G}_0(\tilde{u}_0)\) is large for states \({i}\) which have a high value of \(\pi _{i}(\mathbf {u}^\ast).\) Characterizing this gap would require an analytical description of the behavior of \(\mathbf {u}^\ast\), which is non-trivial in general, as seen from the preceding sections.
We obtain the value of this fraction for an example value function, below. In this case the myopic pricing performs well, yielding more than \(78\%\) of the optimal revenue for all \(\rho \lt 1\).
Fig. 3.
Fig. 4.
Next, we consider the special case when all the value distributions are identical, i.e., \(G_{i} = G\) for all \({i} \in \mathcal {Q}\). In this case, the customer is not sensitive to waiting time. The optimal price in this case turns out to be the myopic price.
Intuitively, if the valuations are insensitive to the queue length, then the operator could just maximize the average reward for each arrival, resulting in a myopic optimal policy. This result suggests that a myopic policy is good in situations where the valuations do not change much with the queue length. We now proceed to solve these equations for the special case of deterministic service valuations.
5 Deterministic Service Valuation
We now consider the special case where the customers’ valuation of the service is a deterministic function of the number of customers in the queue. We assume that the customer valuation for the service is \(v_i\) when there are i jobs in the queue, and the valuation sequence \((v_i, i \in \mathcal {Q})\) is monotonically decreasing. Equivalently, sequence of valuation distribution functions \((G_i(\cdot), i \in \mathcal {Q})\) are Dirac measures with respective atoms at \((v_i, i \in \mathcal {Q})\). The Bellman’s equations are identical to Equations (18) and (19), where for all \(i \in \mathcal {Q}\)
\begin{align} (v_i - \Delta (i))^+ &= \frac{\theta -\mu \Delta (i-1)}{\lambda }, \text{ for all }i \geqslant 1 . \end{align}
(47b)
Unlike Section 3 where we proposed an algorithm to obtain the optimal policy for the k-truncated version of the problem, we now derive the optimal policy for the deterministic valuation problem without the need for any truncation. Consider the limiting equation \((v-\Delta)^+ = \frac{\theta - \mu \Delta }{\lambda }\) involving the asymptotic valuation \(v=\lim _{i\rightarrow \infty } v_i\). Defining \(g(\Delta) \triangleq \frac{\theta - \lambda (v-\Delta)^+}{\mu }\), it can be rewritten as
Let us observe a few properties of the fixed point Equation (48). Since \((v-\Delta)^+\) is non-negative, non-increasing, continuous, and convex in \(\Delta\), it follows that \(g(\Delta)\) is non-decreasing, continuous, and concave in \(\Delta\). From the definition, it follows that \(g(\Delta) \leqslant \theta /\mu\) for all \(\Delta\). In particular, we have \(g(\Delta)-\Delta \leqslant 0\) for all \(\Delta \geqslant \theta /\mu\). If there exists a \(\Delta\) such that \(g(\Delta)-\Delta \geqslant 0\), then from the continuity of \(g(\Delta)-\Delta\) it follows that there must exists a \(\Delta\) such that \(g(\Delta)-\Delta = 0\). Hence, for a given \(\theta\), the necessary and sufficient condition for existence of a fixed point is \(g(\Delta) \geqslant \Delta\) for some \(\Delta\), or equivalently, \(\lambda (v-\Delta)^+ + \mu \Delta \leqslant \theta\) for some \(\Delta\). So, defining \({\underline{\theta }} \triangleq \min _\Delta \lbrace \mu \Delta +\lambda (v-\Delta)^+\rbrace ,\) (48) has a solution for all \(\theta \geqslant {\underline{\theta }}\) but does not have a solution for \(\theta \lt {\underline{\theta }}\). We have the following result about the solutions of this limiting equation.
5.1 Optimal Revenue Rate
Let \((\theta , \Delta (i), i \in \mathcal {Q})\) be a solution to (47a)–(47b) and (48). Let \(K \triangleq \max \left\lbrace i \in \mathcal {Q}: \Delta (i) \leqslant v_i \right\rbrace\). It is possible that \(\Delta (i) \leqslant v_i\) for all i in which case K is infinity. We will first consider the finite K case. In the following we write the solution as \((\theta ^K, \Delta ^K(i), i \in \mathcal {Q})\) to explicitly express the dependence on K. Since \(\Delta ^K(K+1) \gt v_{K+1}\), from (47b), \(\Delta ^K(K) = \frac{\theta ^K}{\mu }.\) Furthermore, \(\Delta ^K(K-1) = \frac{\theta ^K - \lambda (v_K - \Delta ^K(K))^+}{\mu } \leqslant \frac{\theta ^K}{\mu } = \Delta ^K(K),\) and \((v_{K-1} - \Delta ^K(K-1))^+ \geqslant (v_{K-1} - \Delta ^K(K))^+ \geqslant (v_K - \Delta ^K(K))^+.\) Continuing iteratively for \(i = K-2,\dots ,0\), we can infer that
Notice that \(\Delta ^K(i) \leqslant v_i\) for all \(i \leqslant K\). Furthermore, from (47a), \(\Delta ^K(0) = v_0 - \theta ^K/\lambda\), and from (47b),
Let there also be another \(K^{\prime } \gt K\) such that \(\Delta ^{K^{\prime }}(K^{\prime }) \leqslant v_{K^{\prime }}\), or equivalently, \(v_{K^{\prime }} \geqslant \Delta ^{K^{\prime }-1}(K^{\prime }-1)\). This implies that \(v_{K^{\prime }-1} \geqslant \Delta ^{K^{\prime }-1}(K^{\prime }-1)\). Iteratively, we can show that \(v_{K+1} \geqslant \Delta ^K(K)\), also implying that \(v_{K+1} \geqslant \Delta ^K(K+1)\) (see (52)). This contradicts the assumption that \(K = \max \lbrace i:\Delta ^K(i) \leqslant v_i\rbrace\). Hence we must have
for \((\theta ^K, \Delta ^K(i), i \geqslant 0)\) to be a solution.
Next we focus on the case where \(\Delta (i) \leqslant v_i\) for all i and derive the revenue rate in this case. We first consider a truncated system in which \(\Delta (i) = \Delta (K)\) for all \(i \gt K\). We again write the solution as \((\theta ^K, \Delta ^K(i), i \geqslant 0)\) to explicitly express the dependence on K. Notice that \(\Delta (K)\) is still given by (49). But, using (47b) for \(i = K+1\), we also have \(\Delta ^K(K) = (\theta ^K - \lambda v_K)/(\mu - \lambda)\). Combining these two equations
To distinguish the two expressions of the revenue rates in (50) and (53), we refer to them as \(\theta ^K_1\) and \(\theta ^K_2\), respectively, in what follows.
We now derive conditions under which each of the above two solutions arise. For clarity of exposition, we make the following assumption
Following theorem yields the optimal revenue rates and the optimal prices.
6 Numerical Simulations
In all the simulations for random valuations mentioned below, the optimal prices refer to those for k-truncated systems for \(k = 1000\), and the optimal revenue rates refer to the revenue rates corresponding to these prices.
We first consider a system with fixed service rate \(\mu =5\) and valuation distribution \(G_i(u)=1-\exp (-\log (e+i)u)\). We plot the optimal prices versus queue length for three different values of arrival rate \(\lambda\), in Figure 5. The optimal prices are higher for larger arrival rates but decrease monotonically as the number of customers in the queue increases. At higher arrival rates, the service provider can charge higher prices without the risk of losing a customer, because another customer will be available soon. This leads to higher prices, and consequently, higher revenue.
Fig. 5.
Next, we plot the optimal price vector for a system with arrival rate \(\lambda =5\), service rate \(\mu =1,\) and the state dependent value distribution \(G_i(u)=1-e^{-(2-\frac{1}{i+1})u}\). We vary the number of customers in the queue from 0 through 30, in Figure 6. We see that the prices sharply decrease initially as in the previous example but then increase as the queue length further increases. As the queue length becomes larger, customers valuations become very small, and admitting and serving new customers would lower the service provider’s revenue rate. Thus the service provider increases the prices to discourage new customers so that the queue length can be kept small.
Fig. 6.
In Figure 7(a), we plot the revenue versus arrival rate, for the optimal policy as well as the myopic policy in (40). The valuation function is \(G_i(u)=1-\exp (-\log (e+i)u)\), and the service rate \(\mu =5\). Both revenues increase monotonically with arrival rate. The myopic policy performs close to optimal at low values of arrival rate. As arrival rate increases above the service rate (i.e., \(\lambda \gt 5\) or \(\rho \gt 1\)), the two revenues begin to diverge. This is similar to what we had observed in Example 21. From Proposition 20, we can see that at low values of \(\rho\), the stationary distribution will decay quite fast with n. Hence the stationary distribution can be approximated by its first few elements. For these values, the ratio \(\alpha _n(\tilde{\mathbf {u}})\) in (41) will close to one as well. Hence, the myopic policy will give value close to the optimal. The gap between the two revenues increases as arrival rate increases. Hence, at higher arrival rates, the revenue rate can be improved by using the optimal policy.
Fig. 7.
Finally, we also plot the optimal revenue rate for a deterministic valuation function. We consider the deterministic valuation \(v(i)=\frac{1}{\log (e+i)}\), which is the deterministic counterpart of the random value function with distribution \(G_i(u)=1-\exp (-\log (e+i)u)\). We compare the revenue rates for the random and deterministic rates in Figure 7(b). The service rate \(\mu =5\). While both revenues are increasing in arrival rate, as expected, the revenue extracted is higher in the deterministic case. This is because the service provider knows exactly how much each customer is willing to pay, and extracts it completely. There is no loss due to randomness.
We first show the existence of a distribution function \(G(\cdot)\). Observe that for any given u, \(G_i(u)\) are non-decreasing and bounded, and hence \(\lim _{i \rightarrow \infty }G_i(u)\) exists. Moreover, by Helly’s selection theorem, there exists a subsequence of \(G_i(\cdot),i\geqslant 0\), \(G_{i_n}(\cdot), n \geqslant 0\), and a right-continuous, non-decreasing function \(G(\cdot)\) such that \(G(u) = \lim _{n \rightarrow \infty }G_{i_n}(u)\) at all the continuity points u of G. Clearly, \(\lim _{i \rightarrow \infty }G_i(u) = \lim _{n \rightarrow \infty }G_{i_n}(u)\) at all these points. This function has the following properties.
(1)
\(\lim _{u \rightarrow -\infty }G(u) = 0\): For any \(u \lt 0\), which is a continuity point of \(G(\cdot)\), \(G_i(u) = 0\). So \(G(u) = 0\) for any \(u \lt 0\).
(2)
\(\lim _{u \rightarrow \infty }G(u) = 1\): Let us fix an \(\epsilon \lt 1\). There exists a \(u_\epsilon\), a continuity point of \(G(\cdot)\), such that \(G_0(u) \gt 1-\epsilon\) for all \(u \geqslant u_\epsilon\). Since \(G_i(u_\epsilon) \uparrow G(u_\epsilon)\) and \(G(\cdot)\) is also non-decreasing, \(G(u) \gt 1-\epsilon\) for all \(u \geqslant u_\epsilon\). Choosing \(\epsilon\) to be arbitrarily small leads us to the desired result.
We provide the proof of Proposition 6 in this section. Recall that price \(u_i(B)\) attains the maximum value \(m_i(B)\) in (17) for all \(i \in \mathcal {Q}\). Thus, we first understand the properties of \((m_i(B), i \in \mathcal {Q})\), followed by properties of price sequence \((u_i(B), i \in \mathcal {Q})\).
Let \(u = (u_0,u_1,\dots)\) be an arbitrary ergodic policy. For any \(i \geqslant 1\), from the definition of \(m_i(\cdot)\), we have \((u_i - \Delta (i))\bar{G}_i(u_i) \leqslant m_i(\Delta (i)),\) and using (19), \(\lambda \bar{G}_i(u_i)(u_i - \Delta (i)) \leqslant \theta - \mu \Delta (i-1).\) Multiplying both the sides by \(\pi _i\) and using \(\lambda \pi _i\bar{G}_i(u_i) = \pi _{i+1} \mu\) which follows from ergodicity of u, \(\pi _{i+1}\mu (u_i - \Delta (i)) \leqslant \pi _i \theta - \pi _i \mu \Delta (i-1),\) or
Summing (56) and (57) for \(i \geqslant 1\) (summability is ensured by boundedness of \(\Delta (i), i \geqslant 0\)), we obtain \(\mu \sum _{i = 0}^\infty \pi _{i+1} u_i \leqslant \theta .\) Once again using ergodicity of u, we obtain that \(\theta (u) = \lambda \sum _{i = 0}^\infty \pi _i \bar{G}_i(u_i) u_i \leqslant \theta .\) Since the above holds for any arbitrary ergodic policy u, \(\theta ^\ast \leqslant \theta\) as desired. Moreover, replacing the arbitrary policy u with \(u^\ast\), the corresponding inequalities (56) and (57) hold with equality, implying that \(\theta ^\ast = \theta\).
Setting \(\delta _0(\theta) = m_0^{-1}(\theta /\lambda) = 0\), we obtain \(\theta _0 = \lambda m_0(0)\). Also observe that \(\delta _0(\cdot)\) is continuous and decreasing on \([0, \lambda m_0(0)]\).
Let us now fix a \(k \in \mathcal {Q}\). Assume that there exist unique \(\theta _0 \geqslant \theta _1 \cdots \geqslant \theta _k\) satisfying \(\delta _i(\theta _i) = 0\) for all \(i \leqslant k\). Further assume that, for all \(i \leqslant k\), \(\delta _i(\cdot)\) are continuous and decreasing on \([0, \theta _i]\). Observe that
The first term on the right-hand side is decreasing in \(\theta\) from the induction hypothesis. We argue that the second term is also decreasing in \(\theta\). In view of \(\phi _k(\theta)\) being decreasing in \(\theta\), it is enough to show that \(m_{k+1}(B)-m_k(B)\) in increasing in B for \(B \geqslant 0\). For \(0 \leqslant B_1 \lt B_2\),
as desired. This implies that the left-hand side is also decreasing in \(\theta\). Combining this with the facts that \(\phi _k(\theta)\) is decreasing in \(\theta\), \(u_{k+1}(B)\) is increasing in B and \(\bar{G}_k(u)\) is non-increasing in u, we infer that \(\delta _{k+1}\) is also decreasing in \(\theta\).
Furthermore, for \(\theta = \theta _k\), the first term on the right-hand side is zero whereas the second term is negative. So the left-hand side is also negative for \(\theta = \theta _k\), implying that \(\delta _{k+1}(\theta _k) \lt 0\). This also implies that there exists a unique \(\theta _{k+1} \lt \theta _k\) such that \(\delta _{k+1}(\theta _{k+1}) = 0\). This completes the induction step, and so, the proof of the first statement. Monotonicity of \(\delta _i(\cdot)\) along with \(\theta _i \geqslant \theta _k\) for all \(i \leqslant k\) also implies that \(\delta _i(\theta _k) = \phi _i(\theta _k) - \phi _{i-1}(\theta _k) \geqslant 0\) for all \(i \leqslant k\). This proves the second statement.
Observe that \((\theta _k,\phi _0(\theta _k),\phi _1(\theta _k),\dots)\) satisfy (18) and (19), and \((\phi _i(\theta _k), i \in \mathcal {Q})\) are bounded. It remains to show that \(u^k\) as defined above is ergodic if \(\theta _k \gt \lambda m_k(0)\). Then, the desired claim follows from Proposition 6.
For ergodicity of \(u^k\), we now show that \(\lambda \bar{G}_n(u^k_k) \lt \mu\) if \(\theta _k \gt \lambda m_k(0)\). Recall that \(\delta _k(\theta _k) = 0\) implies
Furthermore, \(\theta _k \gt \lambda m_k(0)\) implies that \((\mu - \lambda \bar{G}_k(u^k_k)) \phi _k(\theta _k) \gt \lambda (m_k(0) - u^k_k \bar{G}_k(u^k_k)) \geqslant 0\) where the last inequality follows from the definition of \(m_k(\cdot)\). This inequality implies that \(\lambda \bar{G}_k(u^k_k) \lt \mu\) as desired.
From Lemma 30, if \(\theta _\infty \gt \lambda m(0),\) then \(\rho \bar{G}_k(u^k_k) \leqslant \alpha\) for all \(k \geqslant K\) and so \(\rho \bar{G}_i(u^k_k) \leqslant \alpha\) for all \(i \geqslant k\) and \(k \geqslant K\). Hence, for all \(k \geqslant K\), the policies \(\mathbf {u}^k\) are ergodic for the modified systems as well as the original system. For such k, the stationary distribution of the original system under the policy \(\mathbf {u}^k\) is
Using Lemma 30 it can be shown that, for all \(k \geqslant K\), \(\bar{\sigma }(\mathbf {u}^k) - \sigma (\mathbf {u}^k) \leqslant c \alpha ^k\), where \(c \gt 0\) is a constant. Moreover, it can also be shown that
\begin{equation*} e_i(k) \leqslant {\left\lbrace \begin{array}{ll}0, & i \leqslant k, \\ c\beta ^k_i \alpha ^{i}, & i \gt k \\ \end{array}\right.} \end{equation*}
for bounded constants \(\beta ^k_i, i \geqslant 0\). Suppose \(\beta ^k_i \leqslant A\). Then, we easily obtain the following bound \(\theta _k \leqslant R(\mathbf {u}^k) + cA\alpha ^k(\frac{\alpha }{1- \alpha })\), for all \(k \geqslant K\). Combining it with the earlier observation \(R(u^k) \leqslant \theta ^\ast \leqslant \theta _k\) yields
From (59) and (60), the revenue rate under policy \(\mathbf {u}^k\), \(R(\mathbf {u}^k)\), is a continuous function of \((u^k_i, i \in \mathcal {Q})\) and \((\bar{G}_i(u^k_i), i \in \mathcal {Q})\). From Remark 7, \((u^k_i, k \in \mathcal {Q})\) converges to \(u^\ast _i\) for all \(i \in \mathcal {Q}\). Moreover, from the hypothesis of the corollary, \((\bar{G}_i(u^k_i), k \in \mathcal {Q})\) converges to \(\bar{G}_i(u^\ast _i)\) for all \(i \in \mathcal {Q}\). Consequently, \(\lim _{k \rightarrow \infty } R(\mathbf {u}^k) = R(\mathbf {u}^\ast)\). Also, from Theorem 11, \(R(\mathbf {u}^\ast) = \theta ^\ast\), implying that \(\mathbf {u}^\ast\) is an optimal policy.
Furthermore, from Lemma 26, \(m_{k-1}(\Delta (k-1)) = m_{k-1}(\Delta (k)) \geqslant m_k(\Delta (k))\). Hence, continuing iteratively for \(i = K-2,\dots ,0\), we can infer that \(\Delta (0) \leqslant \Delta (1) \leqslant \cdots \leqslant \Delta (k).\) It remains to show that \(\Delta (0) \geqslant 0\). Notice that \(\Delta (1) \geqslant \Delta (0)\) implies (see Lemma 26)
Hence, using (18) and (19) for \(i=1\), we get \(\frac{\theta - \mu \Delta (0)}{\lambda } \leqslant \frac{\theta }{\lambda }\) or \(\Delta (0) \geqslant 0\) as desired.
7.5.2 Properties of the Lower Bound.
This lower bound \({\underline{\theta }}\) has the following properties.
Notice that existence of a fixed point follows from Lemma 31. So only uniqueness has to be established. However, the proof of Lemma 31(a) also shows that \(\mu \Delta + \lambda m_k(\Delta)\) is strictly increasing for all \(\lambda \leqslant \mu /\bar{G}_k(0)\). So, for any \(\theta \geqslant {\underline{\theta }}\), there cannot be two or more solutions to (32). This also implies monotonicity of the solution, \(D(\theta)\), with \(\theta\).
(b)
Recall that \(f_k(\Delta)\) is increasing and concave and upper bounded. Also, \(f_k(0) = (\theta - \lambda m_k(0))/\mu\). Hence, it is enough to show that \(f_k(\Delta) = \Delta\) has two solutions, \(D_1(\theta)\) and \(D_2(\theta)\), \(D_1(\theta) \leqslant D_2(\theta)\), for all \(\theta \gt \lambda m_k(0)\). Monotonicity of \(D_1(\theta)\) and \(D_2(\theta)\) as stated also immediately follows. So, let us consider any \(\theta \gt \lambda m_k(0)\). Since \(f_k(0) \gt 0\) and \(f_k\Delta)\) is upper bounded, \(f_k(\Delta) = \Delta\) assumes a positive solution. We call it \(D_2(\theta)\). To show existence of another solution, let us define \(\bar{\Delta } \triangleq \theta /(\mu - \lambda \bar{G}_k(0)) \lt 0\). Clearly, \(\theta /\mu + \lambda \bar{G}_k(0)\bar{\Delta }/\mu = \bar{\Delta }\), implying \(f_k(\bar{\Delta }) \leqslant \bar{\Delta }\). The last assertion holds since \(-m_k(\Delta) \leqslant \bar{G}_k(0)\Delta\) for all \(\Delta \lt 0\). So we can conclude that there exists a \(D_1(\theta) \in [\bar{\Delta }, 0)\) that also solves \(g(\Delta) = \Delta\).
and, from Lemma 26, we have \(m_{k-1}(\psi _{k-1}(\lambda m_k(0)) \geqslant m_k(\psi _{k-1}(\lambda m_k(0)) = m_k(0).\) Continuing iteratively for \(i = k-2,\dots ,0\), we can infer that
In particular, \(\lambda m_0(\psi _0(\lambda m_k(0))) \geqslant \lambda m_k(0)\).
We next argue that \(\lambda m_0(\psi _0(\theta))\) is decreasing in \(\theta\). Recall that \(\psi _k(\theta) = D(\theta)\) is increasing in \(\theta\). Suppose that, for some i, \(\psi _i(\theta)\) is increasing in \(\theta\). Then, \(m_i(\psi _i(\theta))\) will be decreasing in \(\theta\) and so \(\psi _{i-1}(\theta)\) will also be increasing in \(\theta\). Arguing inductively, we see that \(\psi _0(\theta)\) is increasing in \(\theta\), and finally, that \(\lambda m_0(\psi _0(\theta))\) is decreasing in \(\theta\).
The above two facts, \(\lambda m_0(\psi _0(\lambda m_k(0))) \geqslant \lambda m_k(0)\) and \(\lambda m_0(\psi _0(\theta))\) is decreasing in \(\theta\), together establish both existence and uniqueness of the fixed point of \(\theta = \lambda m_0(\psi _0(\theta))\). In particular, the unique fixed point \(\theta _k \in [\lambda m_k(0), \lambda m_0(\psi _0(\lambda m_k(0)))]\).
(2)
\(\rho \bar{G}_k(0) \geqslant 1\) and \(\Delta ({\underline{\theta }}) \leqslant 0\): Considering \(\psi _k(\theta) = D_2(\theta)\), \(\psi _k(\lambda m_k(0)) = 0\) and \(\psi _k(\theta)\) is increasing in \(\theta\). Hence, as in Case \((1)\), \(\lambda m_0(\psi _0(\lambda m_k(0))) \geqslant \lambda m_k(0)\) and \(\lambda m_0(\psi _0(\theta))\) is decreasing in \(\theta\), establishing existence and uniqueness of the fixed point. Again, the unique fixed point \(\theta _k \in [\lambda m_k(0), \lambda m_0(\psi _0(\lambda m_k(0)))]\).
(3)
\(\rho \bar{G}_k(0) \geqslant 1\) and \(\Delta ({\underline{\theta }}) \gt 0\): Now we consider the following two subcases separately.
(i)
\(\lambda m_0(\psi _0({\underline{\theta }})) \gt {\underline{\theta }}\): Considering \(\psi _k(\theta) = D_2(\theta)\), \(\psi _k(\theta)\) is increasing in \(\theta\). Hence, as in Case \((1)\), \(\lambda m_0(\psi _0(\theta))\) is decreasing in \(\theta\), establishing existence and uniqueness of the fixed point. The unique fixed point \(\theta _k \in [{\underline{\theta }}, \lambda m_0(\psi _0({\underline{\theta }}))]\).
(ii)
\(\lambda m_0(\psi _0({\underline{\theta }})) \leqslant {\underline{\theta }}\): Considering \(\psi _k(\theta) = D_1(\theta)\), we get \(\psi _k(\lambda m_k(0)) = 0\). Again, as in Case \((1)\), \(\lambda m_0(\psi _0(\lambda m_k(0))) \geqslant \lambda m_k(0)\). So, \(\lambda m_0(\psi _0(\theta)) = \theta\) has a solution in \([{\underline{\theta }}, \lambda m_k(0)]\). In the following, we argue via contradiction that \(\lambda m_0(\psi _0(\theta)) = \theta\) has unique solution. Suppose it has two solutions, \(\theta _k\) and \(\theta ^{\prime }_k\) where \(\theta ^{\prime }_k \ge \theta _k\). Then, \(\psi _{k-1}(\theta ^{\prime }_k) = \psi _k(\theta ^{\prime }_k) = D_1(\theta ^{\prime }_k) \leqslant D_1(\theta _k) = \psi _k(\theta _k) = \psi _{k-1}(\theta _k)\). Moreover,
where the inequalities follow from Assumption 3 and the fact that \(\psi _{k-1}(\theta ^{\prime }_k) \le \psi _{k-1}(\theta _k)\). We argue via induction that \(\psi _i(\theta ^{\prime }_k) - \psi _i(\theta _k) \le \psi _{i+1}(\theta ^{\prime }_k) - \psi _{i+1}(\theta _k) \le 0\) and
where the first inequality follows because \(\psi _{i-1}(\theta ^{\prime }_k) \le \psi _{i-1}(\theta _k)\), \(\psi _i(\theta _k) \ge \psi _{i-1}(\theta _k)\), \(u_{i-1}(B)\) is non-decreasing in B and \(\bar{G}_{i-1}\) is a non-increasing function, the second because \(\psi _i(\theta ^{\prime }_k) \ge \psi _{i-1}(\theta ^{\prime }_k)+\psi _i(\theta _k)-\psi _{i-1}(\theta _k)\) and the third from Assumption 3 and the fact that \(\psi _i(\theta ^{\prime }_k) \le \psi _i(\theta _k)\). This completes the induction step. In particular, we observe that
On the other hand, \(\theta _k = \lambda m_0(\psi _0(\theta _k))\) and \(\theta ^{\prime }_k = \lambda m_0(\psi _0(\theta ^{\prime }_k))\), implying that
and so, \(\theta ^{\prime }_k - \theta _k = \lambda (m_0(\psi _0(\theta ^{\prime }_k)) - m_0(\psi _0(\theta _k))) = 0.\) This establishes that \(\lambda m_0(\psi _0(\theta)) = \theta\) has unique solution.
Finally we show that, in all the above cases, \(\theta = \lambda m_0(\psi _0(\theta))\) has no solution for other potential choices of \(\psi _k(\theta)\).
(1)
\(\rho \bar{G}_k(0) \geqslant 1\) and \(\Delta ({\underline{\theta }}) \leqslant 0\): Taking \(\psi _k(\theta) = D_1(\theta)\), \(\psi _k(\theta) \lt 0\) for all \(\theta \gt {\underline{\theta }}\). But, from Lemma 13, for \((\theta _k, \Delta (i), 0 \leqslant i \leqslant k)\) to be a solution to (18)–(19) and (32), \(\Delta (k) = \psi _K(\theta _k) \geqslant 0\). Hence we can conclude that, for \(\psi _k(\theta) = D_1(\theta)\), \(\lambda m_0(\psi _0(\theta)) = \theta\) does not have any solution.
(2)
\(\rho \bar{G}_k(0) \geqslant 1\) and \(\Delta ({\underline{\theta }}) \gt 0\): We again consider the following two subcases separately.
(a)
\(\lambda m_0(\psi _0({\underline{\theta }})) \gt {\underline{\theta }}\): If we take \(\psi _k(\theta) = D_1(\theta)\), then as shown in (65),
for all \(\theta \ge {\underline{\theta }}\). Equivalently, \(\lambda m_0(\psi _0(\theta)) - \theta \geqslant \lambda m_0(\psi _0({\underline{\theta }})) - {\underline{\theta }} \gt 0\) for all \(\theta \geqslant {\underline{\theta }}\). Hence, \(\lambda m_0(\psi _0(\theta)) = \theta\) cannot have a solution.
(b)
\(\lambda m_0(\psi _0({\underline{\theta }})) \leqslant {\underline{\theta }}\): If we take \(\psi _k(\theta) = D_2(\theta)\), as argued in Case \((1)\) of Part (a), \(\lambda m_0(\psi _0(\theta))\) is decreasing in \(\theta\), and so \(\lambda m_0(\psi _0(\theta)) = \theta\) cannot hold for any \(\theta \gt {\underline{\theta }}\).
(b) When \(\psi _k(\theta) = D(\theta)\) or \(\psi _k(\theta) = D_2(\theta)\), as shown in Case (1) of Part (a), \(\lambda m_0(\psi _0(\theta))\) is decreasing in \(\theta\). Hence, the claim follows from [9, Theorem 2.1].
Now we obtain \(D(\theta)\) under the different conditions stated.
(a)
\(\lambda \lt \mu\): Observe that (48) has a unique solution \((\theta -\lambda v)/(\mu - \lambda)\) if the latter is less than or equal to v, or equivalently, if \(\theta \leqslant \mu v\). Also, (48) has a unique solution \(\theta /\mu\) if \(\theta \gt \mu v\). Combining both these cases yields the desired expression.
(b)
\(\lambda = \mu\): The claim for \(\theta = \mu v\) follows from inspection. If \(\theta \gt \mu v\), from (48), \(\mu (v - \Delta) \leqslant \mu (v - \Delta)^+\), implying that \(\Delta \geqslant v\). In this case, \(\theta = \mu \Delta\) which yields the claimed solution.
Now we consider the following three cases. \(\theta ^\infty _1 \gt \mu v\):. In this case \(\lim _{k \rightarrow \infty } \Delta ^k(k) \gt v\). So, there exist \(k \lt \infty\) such that \(\Delta ^k(k) \gt v_k\) and \(K = \max \lbrace i: \Delta ^i(i) \leqslant v_i\rbrace\) is finite. In this case (47a)–(47b) have a unique solution \((\theta ^K_1, \Delta ^K(i), i \geqslant 0)\). Furthermore, the mean revenue rate \(\theta ^K_1 \leqslant \mu v_K\).
\(\theta ^\infty _1 \leqslant \mu v\):. In this case \(\lim _{k \rightarrow \infty } \Delta ^k(k) \leqslant v\). So, \(\Delta ^i(i) \leqslant v_i\) for all i. Furthermore, the mean revenue rate is \(\theta ^\infty _2\).
\(\lambda \geqslant \mu\) and \(\theta ^K \leqslant \lambda v\):. In this case, following similar arguments as in Section 2.2, the policy \(u = (v_0,v_1,\dots)\) provides a revenue rate \(\lambda v\) which is at least as much as the revenue rate provided by any ergodic policy.
Footnote
1
The subsequent analysis can also be carried out with \(\lim _{k \rightarrow \infty } \theta ^k_i\) replaced by \(\lim \inf _{k \rightarrow \infty } \theta ^k_i\) or \(\lim \sup _{k \rightarrow \infty } \theta ^k_i\) as appropriate if the limit does not exist.
References
[1]
Barış Ata and Tava Lennon Olsen. 2013. Congestion-based leadtime quotation and pricing for revenue maximization with heterogeneous customers. Queueing Systems 73, 1 (2013), 35–78.
Krishna B. Athreya and Soumendra N. Lahiri. 2006. Measure Theory and Probability Theory (1st ed.). Springer Texts in Statistics, Vol. 19. Springer, New York, NY.
Omar Besbes, Adam N. Elmachtoub, and Yunjie Sun. 2022. Static pricing: Universal guarantees for reusable resources. Operations Research 70, 2 (March2022), 1143–1152.
Christian Borgs, Jennifer T. Chayes, Sherwin Doroudi, Mor Harchol-Balter, and Kuang Xu. 2014. The optimal admission threshold in observable queues with state dependent pricing. Probability in the Engineering and Informational Sciences 28, 1 (2014), 101–119.
M. Keith Chen and Michael Sheldon. 2016. Dynamic pricing in a labor market: Surge pricing and flexible work on the uber platform. In Proceedings of the 2016 ACM Conference on Economics and Computation. ACM, New York, NY, 455–455.
R. Divya, Amar Prakash Azad, and Chandramani Singh. 2017. Fair and optimal mobile assisted offloading. In Proceedings of the 2017 15th International Symposium on Modeling and Optimization in Mobile, Ad Hoc, and Wireless Networks. IEEE, Paris, 1–8.
Eugene A. Feinberg and Fenghsu Yang. 2016. Optimal pricing for a GI/M/k/N queue with several customer types and holding costs. Queueing Systems 82, 1 (Feb.2016), 103–120.
Jennifer M. George and J. Michael Harrison. 2001. Dynamic control of a queue with adjustable service rate. Operations Research 49, 5 (Sept.2001), 720–731.
Refael Hassin and Moshe Haviv. 2003. To Queue or Not to Queue: Equilibrium Behavior in Queueing Systems (1st ed.). International Series in Operations Research & Management Science, Vol. 59. Springer, New York, NY.
Donna L. Hoffman and Thomas P. Novak. 2000. Advertising pricing models for the world wide web. In Internet Publishing and Beyond: The Economics of Digital Information and Intellectual Property. Brian Kahin and Hal R. Varian (Eds.), MIT Press, Cambridge, MA, Chapter 2, 45–61.
Ashok Krishnan K. S., Chandramani Singh, Siva Theja Maguluri, and Parimal Parag. 2022. Optimal pricing in multi server systems. Performance Evaluation 154 (April2022), 102282.
Chengcheng Liu and John J. Hasenbein. 2019. Naor’s model with heterogeneous customers and arrival rate uncertainty. Operations Research Letters 47, 6 (Nov.2019), 594–600.
Idriss Maoui, Hayriye Ayhan, and Robert D. Foley. 2007. Congestion-dependent pricing in a stochastic service system. Advances in Applied Probability 39, 4 (Dec.2007), 898–921.
Ioannis Ch. Paschalidis and John N. Tsitsiklis. 2000. Congestion-dependent pricing of network services. IEEE/ACM Transactions on Networking 8, 2 (April2000), 171–184.
Ashok Krishnan K. S., Chandramani Singh, Siva Theja Maguluri, and Parimal Parag. 2020. Optimal pricing in finite server systems. In Proceedings of the 2020 18th International Symposium on Modeling and Optimization in Mobile, Ad Hoc, and Wireless Networks. IEEE, Volos, 1–8.
Xinchang Wang, Sigrún Andradóttir, and Hayriye Ayhan. 2019. Optimal pricing for tandem queues with finite buffers. Queueing Systems 92, 3 (Aug.2019), 323–396.
Serhan Ziya, Hayriye Ayhan, and Robert D. Foley. 2006. Optimal prices for finite capacity queueing systems. Operations Research Letters 34, 2 (March2006), 214–218.
We study the classic setting of envy-free pricing, in which a single seller chooses prices for its many items, with the goal of maximizing revenue once the items are allocated. Despite the large body of work addressing such settings, most versions of ...
How should a seller price her goods in a market where each buyer prefers a single good among his desired goods, and will buy the cheapest such good, as long as it is within his budget? We provide efficient algorithms that compute near-optimal prices for ...
The price for a product may be set too low, causing the seller to leave money on the table, or too high, driving away potential buyers. Contingent pricing can be useful in mitigating these problems. In contingent pricing arrangements, price is ...
Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. Copyrights for components of this work owned by others than the author(s) must be honored. Abstracting with credit is permitted. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. Request permissions from [email protected].