Mathematical Optimization 15

Mathematical Optimization
in the Management of Services
Jess Boronico
Dept. of Management/Marketing
School of Business Administration
Monmouth University
West Long Branch, NJ 07764–1898
[email protected]

Introduction
While students versed in mathematics are often exposed to methodologies
involved in solving mathematical models, less often do they see integrated
models that utilize constructs across disciplines, each dependent upon mathe-
matics for the development of optimal solutions to real-world problems. This
paper discusses one such model, where mathematical programming, statistics,
and microeconomics are unified. The model, when expanded, can prove useful
in applied, real-world problems [Boronico 1996].
We illustrate, for a monopolistic service provider facing stochastic demand,
how reliability of service impacts on marginal costs (within which optimal
price is embodied) and capacity decisions. We discuss pricing in a monopoly,
develop the model, and analyze the model to arrive at both optimal capacities
and price. Then we present illustrative examples, followed by supporting
numerical results and concluding comments.

Pricing in a Monopoly
Quality service and consumer satisfaction have become realities for many
service-oriented industries. To implement timely service within these indus-
tries, capacity plans must provide for adequate staffing during both peak-load
and off-peak hours, incorporating probabilistic demand.
The 1980s and 1990s have illustrated how quality management can affect
the profitability of many industries once thought to be immune from competi-
tion, such as the telecommunications industry and postal services [Adie 1989].
Mathematical models, combined with classical economic and statistical the-
ory, yield meaningful results and insights into the nature of optimal operating
policies [Boronico et al. 1995; Boronico et al. 1996]. The results of this paper’s
The UMAP Journal 18 (1) (1997) 15–34. °Copyright
c 1997 by COMAP, Inc. All rights reserved.
Permission to make digital or hard copies of part or all 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. Abstracting with credit is permitted, but copyrights
for components of this work owned by others than COMAP must be honored. To copy otherwise,
to republish, to post on servers, or to redistribute to lists requires prior permission from COMAP.
16 The UMAP Journal 18.1 (1997)

analysis concern two common measures of great importance to all industries:

price and required capacity.
Pricing results, as they apply specifically to monopolistic industries, are im-
portant in discouraging inefficient entry of competitors into the market, which
in turn may lead to misallocation of resources, cream-skimming, and an overall
decrease in welfare [Dobbs and Richards 1991].
While much work has been done in the determination of optimal pricing,
the incorporation of a reliability constraint on service quality has not been in-
vestigated thoroughly. Consider postal services; a typical reliability constraint
would force installed capacity high enough so that the probability of arriving
mail exceeding capacity is less than some prespecified constant. As reliabil-
ity increases, demand increases and the demand curve shifts out in (quantity,
price) space. If reliability decreases, demand decreases, and the demand curve
shifts in.
The explicit determination of price depends on the criterion (e.g., profit
maximization) governing how price is set. As price increases (decreases) there is
a resulting movement along the demand curve, or a shift in quantity demanded,
to be distinguished from the shift in the demand curve brought about through
a change in reliability.
The impact of changes in both reliability and price on the level of service
that consumers choose to consume will depend on the relative strengths of
these two competing effects:
• the increase in demand (the outward shift of the demand curve) caused by
greater reliability, and

• the decrease in quantity demanded (the movement along the new demand
curve) caused by the increase in price attendant with greater reliability.

While the question of optimal reliability level is an important one, of equal im-
portance are the specific issues discussed in this paper, namely, cost-minimizing
capacity plans and welfare-optimal pricing policies that satisfy a prespecified
constraint on quality, with a reliability level determined ex-ante (in advance).

Model Development
Reliability
A monopolist is free to choose between various price/output combinations
(subject to regulation). Monopolists ordinarily attempt to maximize social wel-
fare, that is, the difference between consumers’ aggregate willingness to pay for
a product/service and the total costs involved in providing the required level
of output. Welfare-optimal prices should be equated to reliability-constrained
marginal cost [Crew and Kleindorfer 1992], and we apply this result in the
model developed here to illustrate how these prices are obtained. Similarly,
 Mathematical Optimization 17

optimal capacities should be set to minimize the overall expected cost of meet-
ing the prespecified reliability.
Delivery services and postal systems may have a particular technology for
processing preferential mail (e.g., overnight express mail), such as optical char-
acter recognition systems, barcode sorters, and video encoding technologies.
Arrivals of incoming mail are received at the beginning of given periods, or
tours. For this mail to reach its destination on time, it must be processed within
a critical window during its tour of arrival. Postal services can never guarantee
processing all mail within this window, due to uncertainty in the volume of ar-
rivals. However, the inclusion of a reliability constraint on service allows them
to plan, ahead of time, the required capacity to install to meet certain stated
reliability, such as the overnight standard set for express mail service. When
incoming mail volume exceeds installed capacity, mail must be expedited at
additional cost. This type of the reliability condition may be reworded more
generally as follows:
Capacity decisions must be made so that the expected number of un-
processed arrivals at the end of the planning horizon must be less than or
equal to a prespecified fraction (1 − r) of the expected volume of arrivals
in that period.
The variable r represents the reliability of the system. Although more com-
plex models might involve the actual determination of the optimal reliability
level, reliability is declared ex-ante in the formulation considered here.

The Model
We address two specific issues:
• the price (P ) to charge for a particular service, and
• the level of capacity to install (Q) in order to satisfy demand while providing
adequate service quality.
Arrivals to the system occur in batches at the beginning of the planning
horizon. (If dynamic arrivals are permitted, the problem becomes increasingly
complex due to the required queuing analysis.) The number (X) of arrivals to
the system is governed by a cumulative distribution function F (k) = Pr(X ≤
k), with X a continuous random variable. Units are processed using regular-
time capacity ($b/unit); any excess demand (X > Q) demand is processed
using overtime ($B/unit) or some other form of penalty technology (B > b).
A constraint is appended to guarantee that an ex-ante declared service level
is adhered to with a certain reliability (probability). At optimal capacity, it may
be that the postal service exceeds the ex-ante declared reliability without the
imposition of the reliability constraint; in this case, the reliability constraint
may be used to determine the actual reliability offered by the postal service,
referred to as the implied reliability of the system.
We use the notation and definitions of Table 1.
18 The UMAP Journal 18.1 (1997)

Table 1.
Notation.

Q installed capacity, set in advance

X demand, a random variable
r reliability of the system, chosen ex-ante
C(Q, X) total costs associated with a particular realization of X
H(Q, X) volume of units requiring penalty technology
E expectation operator with respect to the random variable X
b average cost of processing one unit during regular time
B average cost of processing one unit using penalty technology or overtime

The Cost Function

The firm’s objective is to determine Q so as to

Minimize total expected costs while achieving a prespecified reliability of service.

Thus, the objective function (to be minimized) is

£ ¤
E[C(Q, X)] = E bQ + B(X − Q)+ . (1)

We proceed to explain the components of this cost function:

• The term bQ represents the direct cost of regular-time capacity Q at a unit

cost of b. The capacity Q is set in advance, and this cost is incurred even if
the demand X is smaller than the capacity Q.
• The expression B(X − Q)+ is the cost of processing excess demand, which
equals the difference between demand X and the total processing capacity
Q available during the period. The penalty cost per unit is B. If the capacity
exceeds the demand, there is no penalty processing. We use the “+” function,
defined as (
X, if X ≥ 0;
X+ =
0, otherwise.
The amount of penalty processing is 0 when Q ≥ X and is X − Q when
X > Q.

The Reliability Constraint

The reliability constraint must ensure that enough capacity is installed so
that the expected volume of units requiring penalty technology does not exceed
a prespecified percentage of the expected demand for service in that period.
This constraint may be expressed the following way:
£ ¤
E[H(Q, X)] = E (X − Q)+ − (1 − r)X ≤ 0. (2)
 Mathematical Optimization 19

Distribution of the expected value operator and rearrangement yields

£ ¤
E (X − Q)+ ≤ (1 − r)E(X),
where the left-hand side is the volume of demand unprocessed after the period
of interest and the right-hand side is a fraction of the expected demand, with
E(X) the mean demand for service. The explicit evaluation of (2) requires
knowledge of the demand distribution; later we work through two specific
examples.
The general formulation for the problem may be written as
Minimize {E[C(Q, X) | Q ≥ 0, E[H(Q, X)] ≤ 0} . (3)
More specifically, the objective is to choose Q so as to
£ ¤
minimize E[C(Q, X)] = E bQ + B(X − Q)+ (4)
£ ¤
subject to E[H(Q, X)] = E (X − Q)+ − (1 − r)E[X] ≤ 0. (5)

There Is a Global Optimum

Let us for the moment fix X but keep it as a parameter in our functions.
Figure 1 graphs (C(Q, X) for one particular deterministic value of X, with
f1 (Q, X) = bQ and f2 (Q, X) = B(X − Q)+ . For fixed X, the function C(Q, X)
attains its minimum at Q = X (a point at which the derivative of C with respect
to Q does not exist). The shaded region to the right of the vertical is the feasible
region, per constraint (5).

f(Q,X)

H(Q,X)
C(Q,X)

f (Q,X)
1

f 2 (Q,X)

Q
X

Figure 1. Deterministic solutions to the problem (4)–(5).

We now turn to consideration of X as a random variable. We want to choose

Q to minimize the expected (i.e., average) cost over the distribution of values
of X.
20 The UMAP Journal 18.1 (1997)

Note that C(Q, X) is convex, since for any realization of X,

• bQ is linear and therefore convex;

• the “+” function is convex, since (X − Q)+ = max(X − Q, 0), and the point-
wise maximum of two convex functions is convex;
• the sum of convex functions is itself convex; and

• expectation is a linear operator, hence preserves convexity.

Similarly, H(Q, X) is also convex. Thus, the problem (4)–(5) has the pleasant
structure of minimizing a convex objective function over a convex region.
It is also worth noting that E[C(Q, X)]
• is unbounded from above (just consider the term bQ),

• is bounded from below (being the sum of nonnegative terms) and is greater
than or equal to zero, and
• approaches infinity as Q → ∞.

Thus, standard arguments [Wismer and Chattergy 1978] establish that over the
nonnegative orthant Q ≥ 0, the function E[C(Q, X)] has a global unconstrained
minimum, call it C ∗ , at some value Q∗ . Note that C ∗ is a function of Q∗ and
depends on the probability distribution of X, though not on X itself.

Determining the Optimal Solution

Equations (4)–(5) are differentiable with respect to Q, so that we may use
the Karush-Kuhn-Tucker theorem to determine the optimal solution [Hillier
and Lieberman 1990].∗ In other words, to find the optimal capacity Q∗ and
the optimal cost C ∗ , we incorporate the constraint via a Lagrange multiplier λ.
Defining the Lagrangian to be

L(Q, X, λ) = E[C(Q, X)] + λE[H(Q, X)],

the first-order conditions are given by:

∂L(Q, X) ∂E[C(Q, X)] ∂E[H(Q, X)]

= +λ = 0, (6)
∂Q ∂Q ∂Q
E[H(Q, X)] ≤ 0, λE[H(Q, X)] = 0, and λ ≥ 0. (7)
∗ There is a problem in applying the theory here, since the partial derivatives of C, H, and
hence L with respect to Q do not exist along the line Q = X in the QX plane, corresponding
to the situations of the kink in the graph of C in Figure 1. In fact, however, the theory requires
only differentiability in a neighborhood of the optimum point to conclude that the point is a local
minimum. But a local minimum of a convex function on a convex set is also a global minimum
(this is why we were careful to point out the convexity).
 Mathematical Optimization 21

Each of the partial differentiation operators commutes with the expectation

operator, i.e., · ¸
∂E[L(Q, X)] ∂L(Q; X)
=E ,
∂Q ∂Q
and the same for the partial derivatives with respect to lambda, since the ex-
pectation is integration over a different variable X and piecewise continuity of
the partial derivative allows such interchange [DeGroot 1970, 270].
Note that if the reliability constraint H(Q, X) is inactive at a solution value
Q (i.e., E[H(Q∗ , X)] 6= 0), then λ will be zero, whereas λ will be positive
∗

if H(Q, X) is active. Also note that H is nonincreasing in Q, which implies

that ∂H/∂Q ≤ 0. Together with the fact that λ ≥ 0 and (6), this implies
that ∂C/∂Q ≥ 0 at optimum. Given the assumptions on X and the fact that
expectation is defined for integrals demonstrating absolute convergence, we
may apply the first-order condition given by (6) to the problem (4)–(5) to yield

∂L(Q, X)
= b − T (B + λ) = 0, (8)
∂Q

where
T = Pr{X ≥ Q}. (9)
Solving (7)–(8) will produce optimum capacity Q∗ .

Optimal Price
In addition to the optimal capacity Q∗ and the optimal cost C ∗ , we are also
interested in the optimal price. For a monopolist, optimal price equals average
reliability-constrained marginal optimum cost—that is, the average value of
∂C ∗ /∂X, or · ∗¸
∗ ∂C
MC = E . (10)
∂X
At this point there is something of a conceptual problem. We would like
to examine how C ∗ changes with changes in X, but in fact X has been inte-
grated out in C and L. What then are we doing? In effect, we are considering
what happens to the optimal cost C ∗ if we add a little more demand at each
point of the probability distribution for X—that is, if we move the probability
distribution curve to the right by dX (average demand rises by dX).
To accomplish this, we turn to the different optimization problem without
taking expected values—that is, the problem to

Minimize C(Q, X) = bQ + B(X − Q)+

subject to H(Q, X) = (X − Q)+ − (1 − r)X ≤ 0.
22 The UMAP Journal 18.1 (1997)

This puts us in the situation of Silberberg [1978, 170], with X taking the role of
Silberberg’s α.† The envelope theorem (Silberberg [1978, 171, Eq. 6–90; Samuel-
son [1947, 34–35, 243]) tells us that
∂C ∗ ∂L
= .
∂X ∂X
Hence we have
· ¸ · ¸ · ¸
∗ ∂C ∗ ∂L ∂C ∂H
MC = E =E =E +λ
∂X ∂X ∂X ∂X
= T B + λ[T − (1 − r)] = T (B + λ) − (1 − r)λ. (11)

This approach for marginal cost determination is discussed also in Boronico

[1997], where it is applied to a welfare maximization framework.
In passing, we note that the problem (4)–(5) does not include a Ramsey,
or minimum-profit, constraint. Thus, it is possible that the optimal solution
could involve negative profits, a result that a monopolist could not sustain
indefinitely. In the case of a government-run monopoly, such as postal services,
taxpayers could shoulder this loss. On the other hand, in those cases where
reliability-constrained profits are greater than zero, potential entry into the
industry by new firms might be anticipated, unless significant barriers to entry
exist.

Return to Reality: Postal Service

With regard to postal services, signs of commercialization and increased
competition are evident worldwide [Dobbenberg 1993]. Specifically, second-
class (newspapers and magazines), fourth-class (parcels), and express mail
services have witnessed competitive entry in recent years, although nearly
two-thirds of U.S. Postal Service revenues are generated through first-class
mail, which has been imbued with monopoly status [Sherman 1991].
The argument regarding potential welfare gains brought about by compe-
tition within postal services involves issues similar to those encountered in the
telecommunications industry. For example, a major issue concerns the poten-
tial cross-subsidization between high-revenue monopolistic first-class service
and other services open to competition, which would result in higher prices for
first-class mail and lower prices for those segments of the market open to com-
petition, thus discouraging entry by competitors [Dobbs and Richards 1991].
Postal services have already encountered such accusations from greeting-card
publishers, who have historically objected that resulting rates for first-class
delivery were too high, as well as from parcel-post industries, such as United
Parcel Service, who have argued that parcel-post rates are lower than marginal
† The earlier problem in applying the theory persists: The partial derivatives of C, H, and hence

L with respect to Q do not exist along the line Q = X in the QX plane. Again, this obstacle can
be overcome.
 Mathematical Optimization 23

costs. However, since postal services may cross-subsidize in order to fulfill

their universal service obligation to deliver to all addresses, including rural
addresses that may not be otherwise profitable given the existence of uniform
pricing, the emergence of competition may also lead to cream-skimming, which
may erode existing economies of scale or scope. Various alternative solutions
have been proposed regarding these potential difficulties. Cohen et al. [1993]
discuss some potential alternatives regarding social welfare and competition
as they relate to cross subsidization and cream skimming.
Current investigation into postal services has focused on whether the re-
quired scale and scope economies are present to warrant continued monopoly
protection [Panzer 1993]. By repealing monopoly status and creating a so-
called contestable market scenario, the threat of potential entry would encourage
postal services not only to set price equal to marginal cost but also to measure
marginal cost accurately. Technological changes and relative historical indif-
ference on the part of postal services to allocate costs properly has led to the
suggestion that postal services should re-examine how they estimate marginal
cost [Colvin et al. 1993]. Marginal-cost estimation is critical for postal services,
since postal service prices are typically embodied within marginal cost. Issues
involving subadditive cost structures, sustainable price vectors, and cross sub-
sidization all contribute to the ongoing debate over the continued existence of
postal-service monopoly. Results regarding whether various segments within
the vertical structure of postal services warrant monopoly protection, while
still unclear, will determine whether postal monopolies will be repealed in a
manner similar to that witnessed in telecommunications during the past ten
years.

Illustrative Examples
For illustrative purposes in determining optimal capacity and price from the
model (4)–(5), we consider two particular demand distributions, the continuous
uniform distribution and the triangular distribution, both on [l, u].

Uniform Distribution of Demand

The Unconstrained Solution
Consider first the uniform distribution. Appropriate substitution into (9)
yields Z u
dx u−Q
T = Pr{X ≥ Q} = = . (12)
l u−l u−l
Inserting this result into (8) yields

b(u − l)
Q∗ = u − . (13)
B
24 The UMAP Journal 18.1 (1997)

This is the optimal capacity for the unconstrained problem given by (4), or
the constrained problem under the condition that the constraint (5) is inactive
(referred to as the unconstrained solution). Note that Q → l as b → B, which
suggests that installed capacity should approach the minimum demand level
as regular-time costs approach overtime costs. Also note that (13) yields the
following desired result: l ≤ Q ≤ U as long as b < B. In other words, it would
never be optimal to install capacity outside of the feasible range of demand
values, a conclusion that agrees with intuition.

The Constrained Solution

The optimal capacity for the constrained problem (4)–(5), under the condi-
tion that the constraint (5) is active (referred to as the constrained solution), is
determined by solving (5) as an equality:
Z u
£ ¤ x−Q u+l
E (X − Q) = +
dx = (1 − r) . (14)
l u−l 2

Note that the mean for X, E[X], has been replaced with the mean for the
uniform distribution, (u + l)/2. We solve (14) for Q:
p
Q = u ± (1 − r)(u + l)(u − l).

Since the integral in (14) must yield a positive value, we conclude that Q < u.
Consequently, the unique solution to (14), and the resulting optimal solution
to (4)–(5) when (5) is active, becomes
p
Q = u − (1 − r)(u + l)(u − l). (15)

Implied Reliability in the Unconstrained Case

Since (4) is convex and possesses a unique unconstrained optimal solution
given by (13), next consider the reliability at which the system operates when
this solution applies, or the “implied” reliability level rI that the system attains
without inclusion of a reliability constraint. Substituting (13) into (5) yields

b2 (u − l)
rI = 1 − .
B 2 (u + l)

Note that increases in B relative to b increase the implied reliability. This fact
may be explained by noting that increased overtime cost would lead to a higher
level of installed capacity, which would in turn increase the system’s reliability.

Marginal Cost
For the unconstrained case, set λ = 0. Substitution of (12) and (13) into (11)
eventually yields the result that the marginal cost for this case, call it M Cu ,
 Mathematical Optimization 25

satisfies M C0 = P ∗ = b. Consequently, the optimal price to charge per unit

of service should be set equal to the regular-time processing cost per unit,
namely, b.
To determine marginal cost when the reliability constraint is active, use the
relationships:
E[X] = (u + l)/2,
dE[C(Q, X)]
= b − BT, and
dQ
dE[H(Q, X)]
= −T,
dQ
and determine λ by substituting (14) and (15) into the first-order condition (6),
setting the result to zero. Solving yields
b(u − l)
λ= p − B,
(1 − r)(u + l)(u − l)
which, when substituted into (11), yields the desired marginal cost for the
constrained case, call it M Cc :
r
(1 − r)(u + l)
M Cc = P ∗ = (B + λ) − (1 − r)
u−l
" r #
(u − l)(1 − r)
= b 1− + B(1 − r).
u+l
It is straightforward to illustrate that M Cc converges to the unconstrained
marginal cost M Cu as r approaches the implied reliability r1 and is convex for
all r > r1 . Furthermore, M Cc first declines as r → r1 and then increases again
as r → 1. The marginal cost curve is illustrated in Figure 2. Note that the curve
is not necessarily symmetric about the point (1 + r1 )/2.

The Triangular Distribution

For comparative purposes, consider the following triangular distribution
on [l, u]:

 4(x − l) u+l

 (u − l)2 , if l ≤ x ≤ 2 ;
f (x) = (16)

 4(u − x) u + l
 , if ≤ x ≤ u.
(u − l)2 2
Using (16) to evaluate (9) yields
 µ ¶2

 u − Q u+l

2 u − l , if Q ≥ ;
2
T = µ ¶2 (17)

 u − Q u + l
1 − 2
 , if Q ≤ .
u−l 2
26 The UMAP Journal 18.1 (1997)

Marginal Cost

MC = b
0

Reliability
r
I 1.0

Figure 2. The marginal cost function.

The Unconstrained Solution

Inserting (17) into (8) yields the following result for the unconstrained so-
lution to problem (4)–(5):
 r

 b b 1

u − (u − l) , if ≤ ;
2B B 2
Q= r (18)

 B − b b 1

l + (u − l) , if ≥ .
2B B 2
The first case applies when Q ≥ (u + l)/2. To show this, note that the
unconstrained optimal solution is determined by solving (8), which for λ = 0
yields b/B = T . If b/B ≤ 12 , then T = Pr{X ≥ Q} ≤ 12 , which further implies
that Q ≥ (u + l)/2. A similar argument establishes that the second case applies
when Q ≤ (u + l)/2.
To determine the unconstrained optimal price, substitute (17) into (11), set-
ting λ = 0. This yields the result that P ∗ = M Cu = b.

Comparison of Results for the Two Distributions

Optimal Price
The optimal price to be charged per unit of service is equal for both de-
mand distributions (despite their difference in variance) for the unconstrained
problem. Note that as (u − l) increases, the variances for both the uniform and
triangular distributions increase, and the optimal capacities—given by (13) for
the uniform case and by (18) for the triangular case—also both increase. How-
ever, it may be shown that when the unconstrained solution obtains, these
capacities are set so that Pr{X ≥ Q} = T = b/B is constant, independent of
the demand distribution’s variance. Since, from (11), we have M Cu = BT and
 Mathematical Optimization 27

hence that M Cu is directly proportional to T , it follows that marginal cost is

therefore constant and equal for both distributions. Whether this result holds
for other symmetric or nonsymmetric distributions is left as a suggestion for
future research, as are the triangular distribution’s constrained solutions.

Optimal Capacity
We now compare the optimal capacities for the uniform distribution and
the triangular distribution. A direct algebraic comparison of (15), (16), and (13)
yields the following interesting result (see Figure 3):

Consider the unconstrained problem (4). If b/B < 1/2 (respectively,

b/B > 1/2), then the installed capacity required if demand is governed
by a uniform distribution over [l, u] exceeds (respectively, is less than) the
capacity required if demand is governed by a triangular distribution over
the same interval.

Optimal Capacity

u
Triangular Solution

l+u
2
Uniform Solution

b
x=
u B
0.5 1.0
u-l

Figure 3. A comparison between uniform and triangular solutions.

This relationship has intuitive appeal. Note that for cases where the penalty
cost is very large relative to the regular-time cost (i.e., b/B < 1/2), a higher vari-
ance encourages greater levels of installed capacity so as to protect against the
chance of excess demand, which in turn increases the use of penalty technology.
One may also explain this result geometrically by defining Pr{X ≥ Q} = Tt
if the demand distribution is triangular, and Pr{X ≥ Q} = Tu if the demand
distribution is uniform. Consider the case where b/B < 1/2. This implies that
Q ≥ (u + l)/2. A direct comparison between Tt and Tu for this case yields
Tu ≥ Tt , as illustrated in Figure 4.
Noting that T represents the probability that demand will exceed capacity,
necessitating overtime, we see that for a fixed level of capacity there is a greater
28 The UMAP Journal 18.1 (1997)

f(X) u+l
For Q >
Triangular Distribution 2
T = II + III > I + III = T
Uniform Distribution u t

2
u-l

1 I
u-l
II
III

X
u+l
l u
2
Q

Figure 4. The triangular and uniform distributions.

chance that overtime will be needed if demand is governed by a uniform dis-

tribution. Alternatively, for any level of risk (specified by T ), the triangular
distribution requires less capacity than the uniform distribution. Also note
that the uniform distribution has a greater variance than the triangular distri-
bution. The greater variance necessitates more capacity in order to obtain the
same level of risk (for b/B < 1/2), hence the same result is obtained. A sim-
ilar argument establishes the alternative claim for the case where b/B > 1/2,
for which the triangular distribution requires more installed capacity than the
uniform distribution.
The results of this section are further elaborated in the following section,
where numerical results are generated.

Numerical Results
For each case, the following base parameters will apply:
Regular-time cost = b = $8.00,
Overtime cost = B = $12.00,
Lower and upper limits for demand: l = 100, u = 200.
We examine sensitivity of the results of the model—optimal capacity and
optimal price—to the impact of various changes.

Sensitivity to Reliability
We vary the reliability level between .85 and 1.0; results for the uniform
distribution are shown in Figure 5. As expected, capacity increases as relia-
bility increases, but only when the reliability constraint is binding (r > .8518).
 Mathematical Optimization 29

Note that the capacity required to meet this increased reliability increases at
an increasing rate with respect to r. In fact, as r → 1, the capacity required
approaches infinity (evaluate limr→1 H(Q, X) directly).
Results regarding price are also as expected; namely, optimal price is con-
stant for nonbinding instances and first decreases, then increases, when the
reliability constraint is active. The optimal price approaches the regular-time
cost as r → 1, the same result that would obtain under deterministic conditions
when meeting a fractile of the demand with certainty [Crew and Kleindorfer
1986].

Capacity( Units) Price ( $)

Price 8.00

7.50

Capacity
133

Reliability
.85 .90 .95 1.00

Figure 5. Impact of reliability on capacity and price.

Sensitivity to Variance of the Demand

We increase the demand distribution’s variance without altering its mean,
by simultaneously varying the distribution’s limits in opposite directions. Re-
sults for the uniform distribution are shown in Figure 6. Less capacity is re-
quired as the variance increases for those instances where the unconstrained
solution is optimal (i.e., for [l, u] = [100, 200] and [90, 210]). This is consistent
with the fact that the implied reliability r1 decreases as (u − l) increases, and
hence less capacity is required.
As expected, the optimal price remains fixed at b = 8.0 (since it was shown
that optimal price should be set equal to the “within-period” processing cost
when the unconstrained solution is optimal). For all other instances, the relia-
bility constraint is binding. For these cases, required capacity increases while
optimal price decreases with increases in variabililty. (In general, the capacity
Q∗ required to meet a fixed reliability level is convex when plotted as a function
of the variance of the demand distribution.)
30 The UMAP Journal 18.1 (1997)

Capacity Price
( Units) ($)

$8.00

166

Price

$6.82
133

Capacity
Demand
Limits
[100, 200] [50, 250] [0, 300]

Figure 6. Impact of demand variability at constant mean on capacity and price.

Sensitivity to Average Demand

We vary the mean demand without changing its variance, by changing
both the lower and upper limits on demand simultaneously while maintaining
a constant range. For illustrative purposes, the range was fixed at [µ−50, µ+50],
while µ = E[X] ranged between [150, 250]. For the uniform distribution,
results are shown in Figure 7. Initially, the unconstrained solution does not
require enough capacity to meet the reliability constraint. Hence, the reliability
constraint is active, and we get the expected result: Both optimal capacity and
optimal price increase with increased average demand. However, as the mean
demand increases, the optimal capacity increases at a slower rate than the mean.
But the unconstrained solution increases at the same rate as the mean demand
and hence will eventually overtake the constrained solution. Consequently, for
large enough average demand, we arrive at the unconstrained solution. That
price is increasing at a decreasing rate is consistent with this fact.

Sensitivity to Penalty Cost

We examine what happens as we vary the “within-period” cost b relative
to the overtime cost B = $12 (see Figure 8). When b = B, the unconstrained
solution would suggest setting capacity equal to the lower limit of demand
(i.e., 100). However, at the specified level of reliability of .9, the unconstrained
solution is infeasible, and the optimal capacity is set equal to 145 units. As b
decreases, two items should be noted:
• Optimal capacity, given by the constrained solution, remains constant. This
is due to the fact that the constrained solution is not a function of the cost
parameters b and B.
 Mathematical Optimization 31

Capacity ( Units) Price ( $)

7.80
250 Price

7.57
Capacity

161

Mean
Demand
150 200 250

Figure 7. Impact of mean demand at constant variance on capacity and price.

• The unconstrained optimal capacity (which is initially lower than that which
is required by the reliability constraint) increases.

As b continues to decrease, the unconstrained solution eventually surpasses the

capacity required by the reliability constraint, at which point the unconstrained
solution becomes optimal. Once the unconstrained solution obtains, further
decreases in b lead to increased capacity at optimal.
The intuitive result regarding pricing also applies: Optimal prices drop as
the cost coefficient b decreases. Note that the change in price with respect to b
is larger when the unconstrained solution applies.

Capacity ( Units) Price ( $)

11.01
183
Price

Capacity
145
2.00

Per Unit
Cost$b
2 7 12

Figure 8. Impact of regular time per unit processing cost on capacity and price.
32 The UMAP Journal 18.1 (1997)

Comparative Results for the Two Distributions

Comparative results for the uniform and triangular distribution are shown
in Figure 9. The reliability level is set low enough so that the unconstrained
solutions are optimal for all instances. As ascertained earlier, the optimal ca-
pacity is greater for the uniform solution when b/B < 1/2 and greater for the
triangular distribution when b/B > 1/2.

Capacity ( Units)

200
Uniform

Triangular

100

Per Unit
Cost$b

0 5 10

Figure 9. Impact of regular-time per unit processing cost on capacity.

Concluding Comments
The integration of principles of mathematics, statistics, and economics al-
lows for development of powerful methodologies involving capacity planning
and pricing. In a global economy where the threat of competition is a salient
element in a monopoly’s quest to remain viable, the determination of opti-
mal operating policies is necessary; anything less allows for the emergence of
inefficient entry and a simultaneous decrease in overall welfare.
The determination of true optimal policies involves the incorporation of
many real-world features. For example, in nature, demand is commonly sto-
chastic, although many models visualize demand as deterministic or static.
Furthermore, as illustrated in the section on numerical results, the variability
of the demand distribution plays an important role in operating policies such
as capacity planning and price determination. Consequently, it is not enough
simply to rely on the mean demand when determining policy.
The general basic model discussed here may be applied to two important
monopolistic industries, namely, postal services and fee-for-service roadways.
A number of factors influence the optimal determination of capacity and
price. In general, the illustrative example suggests the intuitive result that
 Mathematical Optimization 33

Increased capacity is obtained by increasing reliability.

However, the impact of reliability on optimal price is not as intuitive. Specif-

ically, price is equal to the per-unit processing cost for cases where the reliability
constraint is inactive, i.e., systems whose reliability, at optimum, already ex-
ceeds a previously specified reliability. When the reliability constraint is active
and reliability is increased, optimal price is convex with respect to reliabil-
ity. Optimal price initially drops from the per-unit processing cost b and then
increases toward b as reliability approaches unity.
We noted other distribution-specific results. When demand is uniformly
distributed, increases in demand variability lead to a decrease in required ca-
pacity when the reliability constraint is inactive, but an increase in required
capacity otherwise. Optimal price decreases as demand variability increases.
In summary, this paper has illustrated how a service provider can incor-
porate a reliability attribute into the managerial process of determining both
optimal capacity and price. The extension of the results presented here by in-
corporating other demand distributions, together with an investigation into the
robustness of the solutions and alternative forms for the reliability constraint,
all represent suggested areas for future research.

References
Adie, D. 1989. Monopoly Mail. Piscataway, NJ: Transactions Publishers.
Boronico, J.S. 1996. Reliability-constrained capacity planning and pricing poli-
cies for toll facility operations. International Journal of Transport Economics
233: 98–122.
. 1997. Postal service pricing subject to reliability constraints on
service quality. Pricing Strategy and Practice: An International Journal, forth-
coming.
, M. Crew, and P.R. Kleindorfer. 1995. Postal service in the nineties.
In The Service and Productivity Challenge, edited by Patrick Harker, 471–492.
Boston, MA: Kluwer Academic Press.
, R.M. Kleyle, and A. De Korvin. 1996. An extension of Ramsey-
Boiteaux pricing to services offering limited deferability subject to reliability
constraints and fuzzy demand functions. Managerial Finance, forthcoming.
Cohen, R.H., W.W. Ferguson, and S.S. Xenakis. 1993. Rural delivery and the
universal service obligation: A quantitative investigation. In Crew and
Kleindorfer [1993], 161–176.
Colvin, J.L., D.B. Bradley, and M.A. Smith. 1993. Measuring product costs for
ratemaking: U.S. Postal Service. In Crew and Kleindorfer [1993], 133–158.
Crew, M.A., and P.R. Kleindorfer. 1986. The Economics of Public Utility Regula-
tion. Cambridge, MA: MIT Press.
34 The UMAP Journal 18.1 (1997)

, eds. 1991. Competition and Innovation in Postal Services. Boston, MA:

Kluwer Academic Press.
. 1992. The Economics of Postal Service. Boston, MA: Kluwer Academic
Press.
, eds. 1993. Regulation and the Nature of Postal and Delivery Services.
Boston, MA: Kluwer Academic Press.
DeGroot, M. 1970. Optimal Statistical Decisions. New York: McGraw Hill.
Dobbenberg, P. 1993. The international mail market in the 1990s. In Crew and
Kleindorfer [1993], 207–216.
Dobbs, I., and P. Richards. 1991. Assessing the welfare effects of entry into
letter delivery. In Crew and Kleindorfer [1991], 61–88.
Hillier, F., and G. Lieberman. 1990. Introduction to Mathematical Programming.
New York: McGraw Hill.
Panzer, J.C. 1993. Competition, efficiency, and the vertical structure of postal
services. In Crew and Kleindorfer [1993], 91–105.
Samuelson, Paul Anthony. 1947. Foundations of Economic Analysis. Cambridge,
MA: Harvard University Press.
Sherman, R. 1991. Competition in postal services. In Crew and Kleindorfer
[1991], 191–214.
Silberberg, Eugene. 1978. The Structure of Economics: A Mathematical Analysis.
New York: McGraw-Hill.
Wismer, D., and R. Chattergy. 1978. Introduction to Nonlinear Optimization. New
York: North Holland.

About the Author

Dr. Jess Boronico is a research professor of operations research within the
Business School at Monmouth University in New Jersey. He holds the Ph.D.
in operations research from the Wharton School of Business of the University
of Pennsylvania. He has published several articles regarding decision-making
and quality maintenance in such journals as the European Journal of Operations
Research, the International Journal of Distribution and Logistics Management, and
the Journal of Applied Business Research, and has also co-authored a book, Com-
puter Simulation in Operations Management (Westport, CT: Greenwood Publish-
ing Group, 1996). Dr. Boronico serves as associate editor for several journals,
including Managerial Finance and the Journal of Cost and Decision Making. He
has also consulted for such industries as the Wakefern Food Industry, the New
Jersey Highway Authority, and the U.S. Postal Service.