Simultaneous Capacity and Production

Management of Short-Life-Cycle,
Produce-to-Stock Goods Under
Stochastic Demand
Alexandar Angelus • Evan L. Porteus
Strategic Decisions Group, Menlo Park, California
Graduate School of Business, Stanford University, Stanford, California
[email protected] • porteus_ [email protected]

T his paper derives the optimal simultaneous capacity and production plan for a short-
life-cycle, produce-to-stock good under stochastic demand. Capacity can be reduced as
well as added, at exogenously set unit prices. In both cases studied, with and without carry-
over of unsold units, a target interval policy is optimal: There is a (usually different) target
interval for each period such that capacity should be changed as little as possible to bring
the level available into that interval. Our contribution in the case of no carry-over, is a
detailed characterization of the target intervals, assuming demands increase stochastically at
the beginning of the life cycle and decrease thereafter. In the case of carry-over, we establish
the general result and show that capacity and inventory are economic substitutes: The target
intervals decrease in the initial stock level and the optimal unconstrained base stock level
decreases in the capacity level. In both cases, optimal service rates are not necessarily constant
over time. A numerical example illustrates the results.
(Capacity Management; Production Management; Capacity Expansion; Capacity Contraction; Finite
Lifetime; Stochastic Demand; Nonstationary )

1. Introduction example. Thus, we do not label this case the “perish-

This paper addresses simultaneous capacity and pro- able” case.
duction planning for short-life-cycle, produce-to-stock Our nominal example of this case is a new elec-
goods under stochastic demand. Such products arise tronic product, such as a processing chip, with a
life cycle measured in months. A sufficient number
frequently in the electronics industry, but appear in
of small design changes are made to the chip each
many industries. We first analyze the case in which
month that precludes any unsold chips produced in
unsold units (produced) cannot be carried over to
one month being carried over to be sold the follow-
be available for sale in the following period. In this ing month: They must be sold in a secondary market
case, the product is perishable from the perspective for less than full price. Production requires at least
of the firm, in that leftover inventory in one period some dedicated production capacity and there is no
cannot be carried over to the next month. However, such capacity installed at the beginning of the life
the product may not be perishable from the perspec- cycle. Capacity is modeled as an upper limit on the
tive of the the purchaser of the product, who may amount that can be produced in a month. Capac-
use the product for many periods, as in the following ity can be either added or removed each month.

0025-1909/02/4803/0399$5.00 Management Science © 2002 INFORMS

1526-5501 electronic ISSN Vol. 48, No. 3, March 2002 pp. 399–413
 ANGELUS AND PORTEUS
Simultaneous Capacity and Production Management of Short-Life-Cycle Goods

However, proceeds from redeemed capacity are less so unsold chips can be retained and used to meet
than the initial purchase price. Capacity additions demands in the following month. The second differ-
or reductions occur instantaneously and all capacity- ence is that unmet demands in a month are back-
related costs are linear. The demand distributions are logged to be met in the future.
known in advance and the unit costs and revenues are We demonstrate that the optimal capacity plan for
exogenously determined parameters. Our assumption this model is still a target interval policy, but that
of a produce-to-stock environment means that com- the target intervals can depend on the amount of
mitment to the production plan for each month must initial inventory. That is, the optimal capacity lev-
be made prior to observing demand in that month. els for each month cannot be scheduled in advance,
Demands that are unmet become lost sales. Each as they may depend on the extent to which left-
month the firm must decide, simultaneously, how over inventory is available to meet current needs. A
much capacity to have and how much to produce. We major conceptual difference between this case and the
also discuss other applications of this case in §5. first is that this one allows for inventory build-up
It follows from Eberly and Van Mieghem (1997) in prior to peak demand months (because of limited pro-
this case that the optimal capacity plan is a target duction capacity). Thus, the optimal inventory level
interval policy in each month: If initial capacity is in a month, after production but before demand,
below the lower target limit, then bring the capacity can exceed the capacity-unconstrained optimal base
up to that limit. If initial capacity is above the upper stock level in that month. Such an outcome cannot
target limit, then bring capacity down to that limit. arise under our assumptions (starting with zero capa-
Otherwise, make no capacity changes. city) in the first case. We show that capacity and
By assuming that demands stochastically increase inventory are economic substitutes in that the opti-
up to a peak and then decrease, we obtain some mal capacity target levels are decreasing functions of
explicit, straightforward characterizations of the opti- the initial inventory level and the optimal capacity-
mal plans. The optimal capacity plan can have up to unconstrained base stock levels are increasing in the
three phases: An expansion phase at the beginning of capacity level.
the life cycle, a constant phase in the middle, and a
downsizing phase at the end. Under our assumptions,
all but one of the parameters needed to specify the 2. Related Literature
optimal capacity and production plans are given by In the case of deterministic demand, our problem
simple critical fractiles of the demand distributions, as reduces to a linear program that can be solved numer-
in classical newsvendor problems. Furthermore, the ically. Indeed, in this case, much more complexity can
optimal capacity levels can be scheduled in advance, be incorporated into the optimization problem, such
because they do not depend on any demand realiza- as multiple products that compete for production time
tions. (This result depends strongly on our assump- on distinct facilities. See Hax and Candea (1984) for
tion of no carry-over of unsold goods and no backlog- a flavor of the models that are possible in this case.
ging.) In particular, the issue of lead times in capacity See Rao (1976) for treatment of simultaneous capac-
expansions is irrelevant provided identification and ity and production planning, and Rajagopalan and
analysis of the problem can be done sufficiently in Soteriou (1994) and Li and Tirupati (1994) for more
advance (so there is enough time to get the optimal recent developments. It is pretty well accepted that
plan implemented). We also show that it is nonopti- few firms can forecast the future demands for their
mal to maintain a constant service level (probability products with certainty. Thus, it is of interest to study
of not stocking out) over the lifetime of the product. models with uncertain future demands.
Our nominal example of the case in which unsold In the case of stochastic demand in a produce-to-
units can be carried over is similar to the first exam- order environment, the firm observes demand before
ple. The main difference is that there are no mean- production. In this case, demand for capacity is the
ingful design changes to the chip during its life cycle, same as product demand and therefore becomes

400 Management Science/Vol. 48, No. 3, March 2002

 ANGELUS AND PORTEUS
Simultaneous Capacity and Production Management of Short-Life-Cycle Goods

the operative concept. Product demand can also be phenomenon. There is a substantial literature that
viewed as the aggregate demand for a number of dif- addresses the case of stochastic demand in a single-
ferent products that share the same production capac- product, produce-to-stock environment with full
ity. This approach also encompasses service as well backlogging, but with capacity treated as a constraint
as production of goods. The models in this literature rather than a decision. For example, for the case of
can be either discrete or continuous time. They are stationary stochastic demand, Federgruen and Zipkin
often infinite horizon models, with the prospect of (1986a, b) show that the optimal production policy is
the demand for capacity increasing without bound. a base stock policy in each month, for both finite and
They are also more general than our model in that infinite horizon problems, and both average and dis-
they encompass possible economies of scale in capac- counted cost criteria. Aviv and Federgruen (1997) and
ity expansion—the cost of adding capacity is not Kapuściński and Tayur (1998) extend the analysis to
simply proportional to the amount added, which is the nonstationary, but cyclic, demand case.
the assumption made in this paper. Manne (1961)
shows that there is a deterministic equivalent prob-
lem whose solution also solves the stochastic problem. 3. The Case of No Carry-over
The discount rate to use in the deterministic problem The life cycle of the product, and therefore time hori-
depends on the variability of the demand for capac- zon of the problem, is T periods (months, for con-
ity. Bean et al. (1992) provide additional conditions venience): Demands will occur for the product only
under which this result holds. Luss (1982) gives an in these T consecutive months. Demands in distinct
excellent review of the literature on capacity expan- months are statistically independent, but not necessar-
sion under stochastic demand in a produce-to-order ily identically distributed. The demand Dt in month t
environment. Some relevant subsequent contributions is a random variable with the known distribution
t 
include Bean and Smith (1985), Buzacott and Chaouch The sequencing in each month is as follows: (1) The
(1988), Erlenkotter et al. (1989), and Rajagopalan et al. new capacity (maximum production) level is selected
(1998). and becomes available immediately. (2) The feasible
A particularly relevant paper is Rocklin et al. (1984), production quantity is selected and becomes available
which presents conditions under which the optimal for sale. (3) Demand is observed and met to the extent
capacity plan in a produce-to-order environment is a possible, with each sale yielding a fixed strictly pos-
target interval policy. Eberly and Van Mieghem (1997) itive price p per item. We formalize our assumptions
provide a significant generalization of that paper, cov- on what happens at the end of each month, because
ering multidimensional capacity expansion and con- we consider different assumptions in the next section:
traction. That a target interval policy is optimal in Assumption A1. At the end of each month, unsold
our case of no carry-over follows from this paper. units are disposed of (in a secondary market), yielding a
Van Mieghem (1998) and Harrison and Van Mieghem unit return of pD .
(1999) derive additional insights in the case of mul-
Assumption A2. At the end of each month, unmet
tidimensional capacity and many products. Kouvelis
demands become lost sales and incur a unit shortage
and Milner (2001) characterize the optimal capacity
penalty of cS .
plan in a one-product model with two-dimensional
capacity, with a random amount available to supple- That is, there is no carry-over of unsold units to be
ment one type of capacity. offered for sale in a later month. Furthermore, there
In most practical settings in which goods are man- is no backlogging of unmet demands: Not only is the
ufactured, there is at least some commitment to pro- opportunity for a sale lost, but an additional penalty
duction quantities that is made before demand is is incurred, because of a loss of goodwill, expected
observed, creating the possibility of unsold goods reductions in future demand, etc.
(or unused components). It is appropriate to assume The cost of adding capacity is K per unit, and the
a produce-to-stock environment to encompass this return from selling excess capacity is k per unit. (All

Management Science/Vol. 48, No. 3, March 2002 401

 ANGELUS AND PORTEUS
Simultaneous Capacity and Production Management of Short-Life-Cycle Goods

capacity is sold at that price at the end of the time Assumption A7 guarantees, for example, that
t−1 , the
horizon.) A production cost cP is incurred for each inverse distribution function, is a continuous (single-
unit produced, and a capacity overhead cost of hC for valued) function on 0 1. It also assures that demand
each unit of capacity held for a month. This over- will be strictly positive with probability one.
head cost covers the manager’s salary, rent of physical Our decision variables are the optimal capacity
space for the capacity, and all other costs that are asso- level and optimal production level in each month.
ciated with maintaining production capacity, but are Because there is no carry-over and no backlogging,
independent of production volume. The one-month each month begins with zero inventory. Thus, the
discount factor is denoted by  0 <  ≤ 1. Financial state of the system at the beginning of each month
transactions during a month are expressed in begin- is simply the amount of available capacity. Therefore,
ning of month dollars, except for end-of-time horizon let ft z be the maximum expected present value of
transactions, which are effectively assumed to occur the total net return that can be earned in months t on,
at the beginning of month T + 1. given that month t starts with capacity z (available
For convenience, the service level in a month is from the previous month). The optimality equations
defined as the probability of no stockout in that for 1 ≤ t ≤ T are given by:
month. In many circumstances, the fill rate, a com-

ft z = max max t y − cC z − z
monly used measure of service, is a strictly increasing 
z ≥0 y∈ 0z

function of this service level. 

−hC z + ft+1 z   (2)
Let t y denote the expected contribution (gross
return less production and shortage costs) in month where
t if the quantity y ≥0 is produced: 
kx if x ≤ 0
cC x = (3)
t y = pE miny Dt  + pD Ey − Dt + Kx if x ≥ 0;

− cP y − cS EDt − y+  (1) and

fT +1 z = kz for z ≥ 0
where, as usual, x+ = maxx 0 For ease of future
reference, we formalize the following assumptions: 3.1. Preliminary Results
Assumption A3. 0 ≤ k ≤ K. Theorem 1. If Assumptions A1–A7 hold, then the
following hold for all t = 1 2     T .
Assumption A4. hC > 0.
(a) If, after any additions or downsizing, there is capac-
Assumption A5. p + cS > cP + hC + K − k. ity z available in month t the optimal production level in
that month can be expressed as y = minSt  z, where
Assumption A6. cP > pD .
 
−1 p + cS − cP
Assumption A7.
t is differentiable on R+ = 0 , St =
t  (4)
p + cS − p D
and strictly increasing on its support, which is an interval,
and
t 0 = 0 for all t. which is strictly positive and finite.
(b) ft is concave, smooth (continuously differentiable),
Assumption A3 guarantees that capacity is sold for
and has a slope between k and K.
less than the purchase price. Assumption A4 provides
(c) There exist finite, strictly positive, constants Lt and
an incentive to sell off unneeded capacity as soon
Ut in R+ such that Lt ≤ Ut for each t and, starting month
as possible. Assumption A5 ensures that the optimal
t with capacity z the following capacity level is optimal:
capacity levels are always strictly positive before the 
end of the time horizon. Assumption A6 assures that 
Lt if 0 ≤ z ≤ Lt

the capacity unconstrained base stock level is finite. zt = z if Lt ≤ z ≤ Ut
If this assumption does not hold, production will 

U if z > U 
always equal capacity, and the problem simplifies. t t

402 Management Science/Vol. 48, No. 3, March 2002

 ANGELUS AND PORTEUS
Simultaneous Capacity and Production Management of Short-Life-Cycle Goods

Part (a) is standard. If there were unlimited capac- Assumption 9 allows us to narrow down the spec-
ity available, then, under our cost assumptions, the ification of the optimal policy. For example, U1 , the
firm would only need to solve a simple newsvendor upper-capacity target in month 1 is irrelevant.
problem for each month of the life cycle and produce The following intuition may prove helpful in
that amount. These quantities, given in (4), form the understanding the results that follow. The demand
base stock levels for the optimal base stock policy: It is distributions are the only things that change over
optimal to produce the amount St if there is sufficient time. Thus, it will not be optimal to invest in more
capacity, and to produce as much as possible if not. capacity in a month than will be used in that month:
The analysis boils down to determining the capacity It is cheaper to delay purchase of that capacity
levels, knowing these base stock levels. until it is needed. Given the optimal capacity level
Karush (1958) shows that t z = maxy∈ 0 z
t y is in month t and that it was fully utilized then, if
concave on R+ . Parts (b) and (c) then follow from the demand stochastically increases in month t + 1
Eberly and Van Mieghem (1997), with smoothness then the unconstrained optimal base stock level will
coming from Assumption A7. That is, the optimal increase, and it cannot be optimal to redeem capac-
policy is an Invest/Stay-Put/Disinvest (ISD) policy, ity. Thus, capacity will continue to be fully utilized. If
which, in our setting of single-dimensional capacity, demand stochastically decreases in month t + 1 then
it cannot be optimal to add capacity: If it were opti-
reduces to what we call a target interval policy: Call
mal to add now, it would have been optimal to add
Lt the lower (capacity) target and Ut the upper (capac-
it in month t which would have added even more
ity) target. It is optimal to make the smallest capacity
value. Because demands will continue to stochasti-
change that puts the capacity level (after the change)
cally decrease over the remaining time horizon, if the
into the target interval. If you start below Lt , you
capacity level is less than the optimal unconstrained
invest to bring capacity up to it. If you start above
base stock level, then all capacity will be utilized,
Ut , you disinvest to bring capacity down to it. If you
whether or not that level is retained or reduced. If
start within the interval, you stay put. Rocklin et al.
that level is more than the optimal unconstrained base
(1984) showed that a target interval policy is optimal stock level, then it is optimal to redeem at least the
in their produce-to-order model. amount that would not be used: It would not be used
If k = K, the target interval is simply a point. in any later month either, so it is more profitable to
Indeed, in this case, the problem decomposes into a sell it now. Hence, in this version of our model, capac-
sequence of simple newsvendor problems in which ity and initial stock level will be identical, so capacity
capacity can be considered to be purchased afresh will be fully utilized, in every month.
each month and fully redeemed at the end of the Theorem 2. Suppose that Assumptions A1–A8 hold.
month.
We now seek explicit solutions for Lt and Ut . (a) The production base stock levels satisfy

3.2. Explicit Solution for Capacity Targets 0 < S1 < S2 < · · · < StP > StP +1 > · · · > ST > 0
To obtain explicit solutions, we make the following (b) The upper-capacity targets satisfy
additional assumptions:
0 < U1 < U2 < · · · < UtP > UtP +1 > · · · > UT > 0
Assumption A8. Demands strictly increase (stochas-
tically) up through a peak month tP , where 1 ≤ tP ≤ T , (c) For t ≥ tP the upper-capacity targets Ut are given by
and strictly decrease thereafter:
1 ≺
2 ≺ · · · ≺
tP   
p + cS − cP − hC − k1 − 

tP +1  · · · 
T , where “≺” and “” represent strict first- Ut =
t−1  (5)
p + cS − p D
order stochastic dominance.
(d) There exists a month tE such that 1 ≤ tE ≤ tP , and
Assumption A9. There is zero capacity available at the the lower-capacity targets strictly increase up to tE and
beginning of month 1. decrease thereafter.

Management Science/Vol. 48, No. 3, March 2002 403

 ANGELUS AND PORTEUS
Simultaneous Capacity and Production Management of Short-Life-Cycle Goods

(e) If tE > 1, the lower-capacity targets are given, for the expansion phase and the first month of the con-
t = 1 2     tE − 1, by stant phase.
  Theorem 3(b) says that if the life cycle of the prod-
p + cS − cP − hC − K1 − 
Lt =
t−1  (6) uct is two or more months, and k < K then it is
p + cS − p D nonoptimal to specify a constant service level as an
(f) There exists a month tC such that tC ≥ tP , Ut ≥ LtE objective over the lifetime of the product. Indeed, in
for t = tE  tE + 1     tC , and, if tC < T , UtC +1 < LtE . The the richest case possible, when 1 < tE ≤ tC < T  which
lower-capacity target LtE satisfies requires T ≥ 3 there will be at least three service lev-
els implemented over time. The service level will be

tC
p + cS − cP − hC − K − tC −tE +1 k
/ constant during the expansion phase, it will decrease
t
t LtE  = 
t=tE p + cS − p D and then increase during the constant phase, and it
(7) will end up at a high constant level during the down-
sizing phase. The service level is lowest during the
where t = t−tE / and  = 1 +  + · · · + tC −tE  peak month, tP  when capacity is most scarce, and is
(g) Lt < St for all t. highest during the downsizing phase, when capacity
(h) Ut < St for t ≥ tP  is least valuable.
The only capacity level needed that is not a sim-
All proofs are in the appendix. Interpretations fol-
ple critical fractile of the demand distribution for that
low the next result.
month is the level LtE  during the constant phase. This
Theorem 3. Suppose that Assumptions A1–A9 target is characterized by (7), which gives the service
hold. level to be attained by the (weighted) average demand
(a) The (optimal) levels of installed capacity, zt  unit distribution, defined as the explicit mixture given of
production, yt  and service,
t yt  are given by the fol- the distributions for months tE  tE + 1     tC  the dis-
lowing deterministic schedules: tribution function is a convex combination of the dis-
 tribution functions for those months.

Lt if 1 ≤ t ≤ tE − 1
 During the expansion phase, whether to add
yt = zt = z if tE ≤ t ≤ tC another unit of capacity becomes the question of


U if t + 1 ≤ t ≤ T whether to add this unit now or next month, as it will
t C
surely be added then. The optimal level is the solu-
and tion, given by (6), to a simple newsvendor problem.
 Similarly, during the downsizing phase, whether to
 p +cS −cP −hC −K1−

  if 1 ≤ t ≤ tE −1 delete another unit of capacity reduces to whether to

 p +cS −pD
reduce it now or next month, which also amounts to a

t yt  =
t LtE  if tE ≤ t ≤ tC

 newsvendor problem, whose solution is given by (5).

 p +c −c −h −k1−
 S P C
 if tC +1 ≤ t ≤ T 
p +cS −pD
3.3. Numerical Example
A new product, with a short life cycle (of 17 months),
(b) If T > 1 and k < K then the optimal service level
is to be produced and sold. Mean demand starts at
is not equal in each month.
2,000 units per month in month 1 and increases by
It is convenient to call periods t = 1 2     tE the 50% each month until month 10, when a peak of
expansion phase, periods t = tE  tE + 1     tC  the con- about 77,000 units is reached. Thereafter, the means
stant phase, and periods t = tC + 1 tC + 2     T the decrease by 1,000 units per month for two months,
downsizing phase. The capacity level strictly increases with the decrease doubling each month thereafter. The
during the expansion phase, stays constant during mean thus drops to about 13,000 units in month 17.
the constant phase, and strictly decreases during the Demand is estimated to be uniformly distributed over
downsizing phase. Note that tE is the last month of 0 2t
 where t is the mean demand in month t

404 Management Science/Vol. 48, No. 3, March 2002

 ANGELUS AND PORTEUS
Simultaneous Capacity and Production Management of Short-Life-Cycle Goods

The opportunity cost of capital is high:  = 097 The produced would always equal the same critical frac-
other economic parameters are as follows: K = 6000 tile (about 96% in this example). The optimal stock-
k = 4000 p = 2000 cP = 100 hC = 100 cS = 400 and out probability would therefore be about 4% in each
pD = 0 Figure 1 shows the results graphically. month. Since demand in each month is a simple scal-
This example illustrates all three possible phases of ing of a common (in this case, uniform) probability
a capacity management plan. If initial capacity is zero, distribution, the fill rate (defined as the expected sat-
this plan begins with a purchase of 3,400 units (per isfied demand divided by the expected total demand)
month) of production capacity in month 1, and then in each month would also be the same in each month.
follows the Lt curve until month 10 when the peak That is, the service level should be the same in
capacity of 102,600 is reached. It is optimal to stay at every month. However, once the effects (and costs) of
this level of production capacity through month 15. expanding and/or reducing capacity are considered,
The downsizing phase begins in month 16. the optimal service levels should vary: During the
Due to Assumption A8, capacity will be fully uti- expansion phase, the (optimal) service level is at the
lized every month. However, starting with positive 84% fractile. At the beginning of the constant phase,
initial capacity can lead to a different conclusion. For in month 10, the service level drops to the 67% frac-
instance, in Figure 1, if initial capacity exceeded U1 = tile, and then rises, reaching the level of 84% in month
21300 then the excess over 21,300 would be sold 15. During the downsizing phase, which begins in
right away, in the first month. Because S1 < U1  not month 16, the service level is at the 87% fractile.
all of the capacity available in month one would be
used. Indeed, there would be unused capacity during
the first five months, but not after that, when the base
4. The Case of Carry-over
We now turn our attention to the case of inventory
stock levels exceed 21,300. Furthermore, if the initial
carry-over. We make the following new assumptions.
capacity level were between S1 and U1  say, at 15,000,
then that level would be left unchanged in month one, Assumption B1. At the end of each month t < T 
even though the capacity would not be fully utilized. unsold units are retained for possible sale in the follow-
The suboptimality of using a uniform service level ing month, incurring a unit holding cost of h At the end
over the product’s life is nicely illustrated with this of month T  unsold units are disposed of, yielding a unit
example. If capacity were unlimited during the entire return of pD <p.
life cycle of the product, the optimal base stock lev- Assumption B2. At the end of each month t < T 
els would be produced. In particular, the amount unmet demands are backlogged, incurring a unit shortage
penalty of cS  and can be met later. At the end of month
Figure 1 Base Stock Levels, Capacity Targets, and Optimal Capacity
T  cumulative unmet demands become lost sales as well as
Plan
incurring the unit shortage penalty of cS .
It is convenient to account for the expected present
value of sales revenues as follows: Give full sales
credit in every month to all demands in that month.
Charge 1 − p for each unit of unmet demand in
that month: That sale will be delayed by at least one
month and 1 − p is the cost, in month t dollars, of
that delay. At the end of the time horizon, which we
think of as month T + 1 charge p for every unit of the
backlog then, as none of that backlog can be turned
into a sale. This means that every unmet demand at
the end of month T will be charged 1 − p in that
month for the delay and p in month T + 1 after dis-
counting into month T dollars. In short, every unmet

Management Science/Vol. 48, No. 3, March 2002 405

 ANGELUS AND PORTEUS
Simultaneous Capacity and Production Management of Short-Life-Cycle Goods

demand in month T will be charged p because it will Assumption B6. h + cP > max0 pD .
never be met. Thus, let
Assumptions B5 and B6 assure that the optimal
It y = p − 1 − p + cS
E maxDt − y 0 base stock levels given below are strictly positive and
finite, respectively.
− hE maxy − Dt  0 − cP y (8)
Theorem 4. If Assumptions B1–B7 hold, then the fol-
where the subscript “I” signals the use of inventory lowing hold for all t = 1 2     T .
(carry-over), and define the terminal value function (a) If, after any additions or reductions, there is capacity
fT +1 as follows: z available in month t the optimal level of inventory after
 production is given by a base stock policy with base stock
px if x ≤ 0 level St z which is a decreasing function of z
fT +1 x z = kz + (9)
pD x if x ≥ 0 (b) The optimal capacity policy is a target interval pol-
icy in each month. The lower- and upper-capacity targets,
It is then straightforward to show that It y + cP x Lt x and Ut x are decreasing functions of x that satisfy
equals the expected assigned sales revenues to month Lt x ≤ Ut x for each t and each initial inventory level x.
t less production and expected holding and shortage (c) ft is (jointly) concave, smooth, submodular, and the
costs in that month, given that month t starts with marginal value of capacity at the beginning of month t is
level x of inventory and y units of stock are available between k and K.
after production is complete (so y − x are scheduled
By (a), the optimal base stock level can depend on
for production). Note that the terminal value function
the optimal capacity level selected, z and the optimal
accounts not only for unmet demands that become
inventory level after production should be as close to
lost sales, but returns from scrapping leftover stock
St z as possible (within the feasible interval, which
and selling off all remaining capacity.
is x x + z
 Therefore, the optimal level of inven-
tory after production is maxx minSt z x + z and
4.1. General Results
the optimal production quantity is max0 minSt z −
In this model, compared to the case of no carry-over,
x z The optimal base stock level and the capacity
we have a state space with an additional state vari-
level are (economic) substitutes, in that as the level
able, the inventory level x at the end of the previous of capacity is increased, less inventory is needed (the
month (unsold units if positive, cumulative backlog optimal base stock level decreases). By (b), the opti-
if negative). Accordingly, let ft x z denote the max- mal capacity level in month t can depend on the ini-
imum expected present value of the total net return tial inventory level in that month. In addition, the
that can be earned in months t on, given that month t optimal capacity level and the initial inventory level
starts with inventory level x and capacity z Therefore, are substitutes, in that as the initial inventory level
the optimality equations for 1 ≤ t ≤ T now become: is increased, less capacity is needed (the optimal tar-

 get levels decrease). By (c), ft is submodular (see
ft x z = max

cP x − cC z − z − hC z + t x z  
z ≥0 Topkis 1978, 1998, for details), so that the initial inven-
tory level x and the initial capacity level z are also
where
substitutes.


t x z = max It y + Eft+1 y − Dt  z  Our model exercises the trade-off between capacity
y∈ x x+z

and inventory that is part of what is called aggregate

For convenience, let Assumptions B3, B4, and B7 planning and is studied extensively in a determinis-
denote Assumptions A3, A4, and A7, respectively, tic demand environment (e.g., see Hax and Candea
which remain in effect in this model. We make the fol- 1994). For example, it can be optimal to install a rel-
lowing changes in assumptions about the parameters. atively low level of capacity and to build up inven-
tory prior to the peak month(s), rather than making
Assumption B5. 1 − p + cS > cP . enough to meet peak month demands immediately

406 Management Science/Vol. 48, No. 3, March 2002

 ANGELUS AND PORTEUS
Simultaneous Capacity and Production Management of Short-Life-Cycle Goods

prior to the peak. In this environment, the optimal is charged on unsold units, and the remaining param-
base stock level during the inventory build-up phase eters are unchanged. This problem was solved numer-
is larger than what it would be if there were no capac- ically. Figure 2 presents the optimal capacity policy
ity constraints. (No capacity constraints corresponds in month 10. This policy indicates the interesting way
to having essentially infinite capacity in our context, in which capacity and inventory are substitutes in
with the smallest base stock levels.) Consequently, the month 10. If capacity is low and there is a large back-
service level will be high during the inventory build- log of, say, 300,000 units, then capacity should be
up phase, to be followed by relatively low service lev- increased to the lower-capacity target, which is about
els in some peak demand months. In particular, it is 190,000. However, if month 10 begins with no backlog
clear in this context that the service level will change (and no inventory), then capacity should be increased
over time. Indeed, it is interesting to observe that the only to about 84,000. Every 100,000 units of inventory
service level in a peak month might depend on the reduces the lower-capacity target by about 35,000, and
observed demands in previous months: If demand is this rate is roughly maintained until the lower target
high in those previous months, the level of inventory hits zero. By contrast, the substitution rate when there
at the beginning of that peak month might be lower is excess capacity changes significantly as the initial
than it would have been if demands had been low inventory level grows: Each decrease in the backlog
in those previous months. In short, the optimal ser- by 100,000 decreases the capacity level by nearly the
vice level in a particular month can depend on previ- same amount. However, as the inventory level grows
ous demands, which cannot happen in the case of no positive, the upper-capacity target flattens. For exam-
carry-over. ple, there is very little reduction in the upper-capacity
target in the range of 100,000 to 300,000 units of initial
4.2. Numerical Example inventory. It appears in this example that high lev-
We modify the numerical example solved in §3 to els of excess capacity will be used for one month and
allow for inventory carry-over. A holding cost of h = 2 redeemed in the next month when there is a large

Figure 2 Optimal Capacity Policy in Month 10

Management Science/Vol. 48, No. 3, March 2002 407

 ANGELUS AND PORTEUS
Simultaneous Capacity and Production Management of Short-Life-Cycle Goods

backlog at the beginning of the month, and that a 5. Discussion

pretty substantial amount of excess capacity will be We have considered the problem of simultaneous
retained in future months if the initial inventory level capacity and production management under non-
is high. Of course, there is no substitution between stationary, stochastic demands in a short-life-cycle,
inventory and capacity when capacity falls into the produce-to-stock environment. Under our assump-
target interval: For a given current capacity level, tions for the case of no inventory carry-over, includ-
there is an interval of initial inventory levels such that ing no initial capacity, the capacity levels over the
no change in capacity is made regardless of how the time horizon can be specified in advance and will be
initial inventory/backlog level is changed within that fully utilized in every month. These levels increase
interval. until a last expansion month, denoted by tE , remain at
In this example, a moderate amount of inven- that level for one or more months, and then decrease
tory build-up takes place prior to the peak months, during the remaining months. Furthermore, once tE
to increase the chances of less total capacity being is known, the capacity levels are explicitly character-
purchased. Consider the scenario in which demand ized. Indeed, all of these capacity levels except one,
in each month equals the expected demand. Under are solutions to simple newsvendor problems: The
the optimal policy (for the stochastic problem), this capacity levels for the expansion phase correspond
scenario leads to a specific realized capacity plan. to a fixed service level (in that month) and the lev-
Figure 3 compares this capacity plan with the one els for the downsizing phase correspond to a differ-
developed for the case of no inventory carry-over. In ent, higher, but still fixed, service level. The capacity
this case, the maximum capacity level is only 64,000, level for the constant phase corresponds to a third
compared to 102,600. Figure 3 also shows what the service level for a more complicated demand distri-
initial inventory level would be in each month under bution, which is an explicit weighted average of the
this scenario: Inventory builds up to a peak of about demand distributions for the months of the constant
68,000 at the beginning of month 9 and then starts to phase. In short, it is optimal to change the service
be drawn down. Even though it is relatively inexpen- level provided to customers over time: Provide the
sive to hold inventory, it is optimal to build capac- lowest service level during the peak month and the
ity (and, therefore, inventory) slowly early in the highest service level in those months when capacity
life cycle, presumably because capacity is relatively is being redeemed.
expensive to maintain over time. The form of the optimal policy in the case of no
inventory carry-over follows from Eberly and Van
Mieghem’s (1997) general results, which apply to a
Figure 3 Comparison of Optimal Capacity Schedules
produce-to-order environment. Our explicit character-
izations of the pertinent target levels can be used in
produce-to-order environments by judiciously setting
the economic parameters. (Set p to be the unit margin
on the product (the actual unit sales price less the unit
production cost) and set pD = cP = 0. In this case, all
capacity is available to meet demand and production
costs are incurred only for units demanded.)
After a similar judicious assignment of parameter
values, the case of no carry-over can also represent
environments in which an organization sells services,
rather than physical goods. One example is a car
rental problem where demand consists of requests for
vehicle rentals and tends to be seasonal, while capac-
ity is the number of vehicles available (to be rented)

408 Management Science/Vol. 48, No. 3, March 2002

 ANGELUS AND PORTEUS
Simultaneous Capacity and Production Management of Short-Life-Cycle Goods

at the beginning of, say, a week. Our model would unit shortage costs of the model capture the economic
fit best for a car rental company that started buying consequences of producing to order: Negative inven-
new vehicles each new model year, starting with no tory levels correspond to existing backlog levels, and
vehicles from the previous model year, and assum- if the optimal base stock levels are nonpositive, then
ing disposition of all old vehicles by the beginning of it is optimal to always produce to order (and never
the next model year. It would also fit best when there have strictly positive inventory levels). If the optimal
were not significant economies of scale in either pur- capacity targets and base stock levels are strictly pos-
chasing or maintenance, because our model assumes itive but small, the only time goods are produced to
such costs are linear. Our model would then sug- stock is when there has been a slow month or two of
gest that customers should receive the best service (in demand and there is sufficient capacity to more than
terms of probability of obtaining a car from a pre- work off the backlog.
ferred rental company) when that rental company is There are many possible extensions to this work.
disposing of its excess cars, following peak demand For example, in the carry-over case, excess inventory,
periods. as well as excess capacity, could be disposed of at a
We have also obtained basic results for the case discount. For some of these extensions, it would be
of inventory carry-over (and backlogging of unmet relatively straightforward to show that target interval
demand). We showed that a target interval policy policies would continue to be optimal, but the explicit
is optimal in managing capacity and that the opti- characterization of the targets in the case of no carry-
mal target levels are decreasing functions of the ini- over would be lost. One example is allowing possibly
tial inventory level. Once the optimal capacity level, random capacity deterioration. Another is allowing
z , for a month is given, the optimal level of inven- the economic parameters to be nonstationary.
tory after ordering is the closest point in the inter- We assume that there is no lead time in adding
val x x + z
to the base stock level St z , which is or reducing capacity. In particular, any capacity that
a decreasing function of z . Once that level is deter- is purchased becomes available for immediate use.
mined, random demand is observed and the next However, when there is no carry-over, the optimal
period begins. The optimal capacity level and level of capacity targets and production base stock levels are
inventory after ordering in that period depend on the deterministic, in that they don’t depend on any actual
realized demand this period. We show that inventory demand realizations. Consequently, we can relax the
and capacity are always economic substitutes and assumption of no delay in adding or reducing capac-
reveal the way in which the substitution takes place: ity. For the results to continue to hold, it is sufficient to
If there is a demand surge in a period, so that initial assume that the delay is deterministic and the prob-
inventory in the next period is lower than expected, lem known sufficiently far in advance that the opti-
the target levels for capacity are increased, so it may mal capacity plan can be implemented. However, this
be optimal to end up with more capacity than what result does not hold in the case of inventory carry-
would have been expected under the optimal policy. over, where the optimal capacity targets and base
Because of this increase in capacity, the optimal base stock levels can depend on previous demands, includ-
stock level decreases. Yet the level of inventory after ing those of the previous period. This important gen-
production may be larger or smaller than expected. eralization to the case of a positive capacity lead time
Similarly, if demand in the previous period was low, with inventory carry-over merits further research.
it may be optimal to end up with less capacity than Our model does not allow for economies of scale in
expected, higher optimal base stock levels, and either capacity changes, which means that our model may
more or less inventory after production. fit best when capacity consists primarily of labor. The
As with other models that assume inventory carry- first step in allowing for economies of scale in capac-
over and backlogging, our model sheds some light ity changes, such as a fixed (set-up) cost for capac-
on whether production should be done to stock or to ity additions or reductions, in addition to the vari-
order when there is a choice and to the extent that the able costs, is to consider only capacity additions, as

Management Science/Vol. 48, No. 3, March 2002 409

 ANGELUS AND PORTEUS
Simultaneous Capacity and Production Management of Short-Life-Cycle Goods

in Angelus, Porteus, and Wood (2000), who study which both denote the expected net return starting in period t with
a short-life-cycle, produce-to-order model with pos- initial capacity z and making no immediate change in that capac-
itive capacity expansion lead times and correlated ity. In the ensuing results, we assume t is an arbitrary element of
1 2     T  unless otherwise specified. It is convenient in proving
demands.
Theorems 2 and 3 to explore some of the details of the solution
It would be interesting to generalize the model to given by Theorem 1.
cover multidimensional capacity and resources that It follows from direct differentiation that, for y ≥ 0,
are shared across multiple products, as in Eberly and
Van Mieghem (1997) and Harrison and Van Mieghem t y = p + cS − cP  − p + cS − pD 
t y (A4)
(1999). There are a few minor generalizations that
Two consequences of Assumption A9 are that, for 0 < x < 1,
are encompassed by our model. For instance, if there
are resources that are shared across several products 0 <
1−1 x <
2−1 x < · · · <
t−1
P
x
and that can be added or redeemed for the same >
t−1 x > · · · >
T−1 x > 0 (A5)
P +1
unit price, then the multiproduct problem decom-
poses into separate single-product problems, with the and, by (A4), for y ≥ 0,
marginal production cost including the marginal pur-
1 y ≤ 2 y ≤ · · · ≤ tP y ≥ tP +1 y ≥ · · · ≥ T y (A6)
chase cost of these resources. Resources that yield
strictly less when redeemed than they cost when Lemma 1.
added, must be dedicated to a single product for our (a) The functions t and t are concave and smooth on R+ .
model to apply. It is plausible that general purpose (b) t St  = t St  = 0, and, in general,
resources (used for many different products) will be 

t z if 0 ≤ z ≤ St 
able to be bought and sold at nearly the same unit cost  z =
t
0 otherwise
while specialized, dedicated resources will not be.
In summary, even though some of our assumptions (c) The functions gt and ht are concave and smooth and satisfy gt ≤ ht
are relatively restrictive, we believe that this model on R+ .
takes a useful step in the direction of understanding (d) gt Lt  = ht Ut  = 0, and, hence, Lt and Ut maximize gt and ht
better the interaction between capacity and produc- over R+ , respectively.
(e) gt z = −K1 −  − hC + t z for z ≤ Lt+1 .
tion management.
(f) ft 0 = K and limz→ ft z = k, and, in general,

Acknowledgments 
 if 0 ≤ z ≤ Lt
K
The authors thank Luk Van Wassenhove for his constructive guid- ft z = −hc + t z + ft+1

z if Lt ≤ z ≤ Ut (A7)
ance and particularly the anonymous referees for their extremely 

k if z ≥ Ut 
helpful comments and suggestions and for their patience in waiting
for our revisions.
(g) Lt < St .
Appendix A. Proofs for the Case of Proof. It follows from Assumption A7 that t is strictly concave
No Carry-over and smooth. It is standard that the optimal base stock level St is
Assume that Assumptions A1–A9 hold henceforth in this appendix, a maximizer of t . Thus, this maximizer is unique. It follows from
unless indicated otherwise. We first rewrite the optimality equa- Karush (1957) that t is concave: The optimal stock level is the
tions (2) as closest point to St in the interval 0 z
. Part (b) follows as does the
 
ft z = max Kz + max gt z  kz + max ht z  smoothness of t .
z ≥z 0≤z ≤z Part (c): The functions gt and ht are concave and smooth,
where because, by (a) and theorem 1(b), t and ft+1 both are. The inequal-
ity follows from Assumption A3, which implies that the maximizer
gt z = −Kz − hC z + t z + ft+1 z (A1)
of gt is smaller than that of ht . By the optimality equations, it is
ht z = −kz − hC z + t z + ft+1 z (A2) optimal to move up to the maximizer of gt if you start below it,
to move down to the maximizer of ht if you start above it, and to
and, as earlier, t z = max0≤y≤z t y, with t as defined in (1).
stay-put otherwise. Thus, (d) holds.
t z is the maximum contribution in period t, as a function of the
Part (f): (A7) follows from plugging the form of the optimal
available production capacity z. Note that
policy into the optimality equations. It is convenient to demon-
Kz + gt z = kz + ht z (A3) strate that Lt is strictly positive by induction. Suppose Lt+1 > 0.

410 Management Science/Vol. 48, No. 3, March 2002

 ANGELUS AND PORTEUS
Simultaneous Capacity and Production Management of Short-Life-Cycle Goods


Then, by (A7), ft+1 0 = K. Therefore, using (A1) and (b), gt 0 = weak inequality in the inductive step.) The strict inequality holds

−K − hC + t 0 + K, which, by (A4) and Assumption A7, equals because, by Lemma 2,
tP >
tP−1 and tP−1 is strictly concave at
−K − hC + p + cS − cP + K, which, by Assumptions A3 and A5, is
tP−1 . The next equality follows from Lemma 2(c) and Lemma 1(b).
strictly positive. Hence, Lt , the maximizer of gt , is strictly positive. The final equality follows again from Lemma 2(a). Hence, Part (b)
Part (e) then follows from the definition of gt and (A7). holds.
Part (g): By (b), gt St  = −K − hC + ft+1

St , which, by Theo- We now show that, for z ≥ 0,
rem 1(b) is less than −K1 −  − hC , which, by Assumptions A3
1 z ≤ 2 z ≤ · · · ≤ tP z ≥ tP +1 z ≥ · · · ≥ T z (A9)
and A4, is strictly negative. Hence, Lt , the maximizer of gt , which
is concave, must be strictly less than St .
By Lemma 1(b), t z can be expressed as t z = maxt z 0.
Equation (A9) follows from (A6) because the function maxx 0 is
Lemma 2. For each t, let
  increasing in x and, hence, order preserving.
p + cS − cP − hC − k1 −  We next prove that

t =
t−1  (A8)
p + cS − pD
LtP ≥ LtP +1 ≥ · · · ≥ LT  (A10)
Then (a) t 
t  = hC + k1 − , (b) 0 <
t <
2 < · · · <
tP >

tP+1 > · · · >
T > 0, (c)
t < St for all t, and (d) t is strictly concave We present only the inductive step as the initial step is similar.
at
t (i.e., in an open neighborhood of
t . Suppose inductively that Lt+1 ≥ Lt+2 ≥ · · · ≥ LT for an arbitrary t ≥ tP .
We must show that Lt ≥ Lt+1 .
Proof. Part (a) follows directly from (A4). By Assumptions A3,
A4, and A6, the fractile in (A8) is strictly between 0 and 1, so it gt Lt+1  ≥ −K − hc + t+1

Lt+1  + K
follows that
t is uniquely defined, and, by (A5), that (b) holds.  
≥ −K − hc + t+1 Lt+1  + ft+2 Lt+1 
Since hC + k1 −  > 0, it follows from (4) that (c) holds. Part (d)

follows from Lemma 1(b) and Part (c).
= gt+1 Lt+1  = 0
Proof of Theorem 2.
The first inequality uses Lemma 1(e) and (A9). The next inequality
Part (a) follows from (4) and (14).
follows from Theorem 1(b). The last equalities follow from defini-
Part (b): We first use induction to prove that Ut =
t for t ≥ tP .
tions.
We present only the inductive step, because the initial step is so
Part (d): Define tE as the largest t in 2 3     tP  such that
similar. Assume inductively that Ut+1 =
t+1 for some t ≥ tP . Hence, 
gt−1 Lt  < 0. If such a t does not exist, then set tE = 1. This definition
using Lemma 2(b),
t >
t+1 = Ut+1 , from which it follows from

is unambiguous, because the Lt are computed in reverse order, for
Lemma 1(f) that ft+1 
t  = k, and from Lemma 2(c) that t 
t  =
t = T  T − 1    .
t 
t . Thus, by Lemma 2(a), ht 
t  = 0, so
t is a maximizer of 
By (A10) and the definition of tE , it follows that gt−1 Lt  ≥ 0
ht . Since, by Lemma 2(d), ht is strictly concave at
t , we conclude
for t = tE + 1 tE + 2     tP , which implies that the lower targets
that
t = Ut .
decrease in t for t > tE . By definition of tE , if tE > 1, then LtE −1 < LtE .
We have therefore verified Parts (c) and (h), as well as Part (b)
Thus, we need only show that the lower targets strictly increase in
for t ≥ tP . To complete the proof of Part (b), we show by induction
t for t ≤ tE − 1. Toward this end, we first show that they weakly
that
t ≤ Ut < Ut+1 for t = 1 2     tP −1. We present only the initial
increase. Suppose inductively that Lt+1 ≤ Lt+2 and that t + 2 ≤ tP . We
step, as the inductive step is so similar. To show that
tP−1 ≤ UtP−1 ,
shall show that Lt ≤ Lt+1 .
we have
gt Lt+1  ≤ −K − hc + t+1

Lt+1  + K
htP−1 
tP−1  ≥ −k1 −  − hC + tP−1 
tP−1 
  
= −K − hc + t+1 Lt+1  + ft+2 Lt+1  = gt+1 Lt+1  = 0
= −k1 −  − hc + tP−1 
tP−1  = 0
The inequality comes from Lemma 1(c) and (A9). The first equal-
The inequality follows from (A2) and since ftP ≥ k by Theorem 1(b). ity follows from Lemma 1(f) by the inductive assumption. The
The first equality follows from Lemma 1(b) since
tP−1 < StP−1 by equalities follow as above. Thus, the lower targets weakly increase
Lemma 2(c). The last equality follows from Lemma 2(a). in t for t ≤ tE . Thus, Part (e) follows when solving for Lt via gt Lt  =
To show that UtP−1 < UtP , we have 0, using Parts (e), (g), and (b) of Lemma 1 and (A3).
Finally, by our assumptions, the fractile in (A8) is strictly
htP−1 UtP  = −k1 −  − hc + tP−1 UtP  between zero and one, so, by (A5), the lower targets strictly increase
= −k1 −  − hc + tP−1 
tP  in t for t ≤ tE − 1. Hence, the proof of Part (d) is also complete.
Part (e): Thus, for t = 1 2     tE −1, Lt ≤ Lt+1 , so, by Lemma 1(e),
< −k1 −  − hc + tP−1 
tP−1  gt Lt  = −K1 −  − hC + t Lt . Since, by Lemma 1(g), Lt < St , it
= −k1 −  − hc + tP−1 
tP−1  = 0 follows from Lemma 1(b) that

t Lt  = t Lt  = p + cS − cP  − p + cS − pD 
t Lt 

The first equality follows from (A2) and (A7). The next equality
follows from our verification above that UtP =
tP . (This becomes a The result follows from solving gt Lt  = 0.

Management Science/Vol. 48, No. 3, March 2002 411

 ANGELUS AND PORTEUS
Simultaneous Capacity and Production Management of Short-Life-Cycle Goods

Part (f): By (d), LtE ≥ · · · ≥ LT . By the definition of tC , it follows t x z = max at y z
y∈ xx+z

that Lt ≤ LtE ≤ Ut for t = tE  tE + 1     tC and, if tC < T , LtE > UtE +1 .

By (A7), ftE +1 LtE  = k. Thus, by (A7) again, at y z = It y + Eft+1 y − Dt  z

ftE LtE  = −hC + tE LtE  + k and It is as defined in (8).

Outline of Proof of Theorem 4. By (9) and Assumption B1,
= −hC + tE LtE  + k (A11) fT +1 is concave and submodular, with a marginal value of capac-
ity equal to k. Suppose inductively that ft+1 is concave, submodu-
Furthermore, for arbitrary t = tE  tE + 1     tE − 1, we have, by (A7) 2
lar, and ft+1 is between k and K, where parenthetical superscripts
again, that denote partial differentiation. We shall show that these properties
ft LtE  = hC + t LtE  + ft+1

LtE  are preserved in ft and that, in addition, ft is smooth. We verify
Part (f) follows from using (A4) and (A11) recursively. the other parts of the theorem in the process.
Part (g) is given by Lemma 1(g).
Given capacity z, the objective function when seeking the opti-
Proof of Theorem 3 Part (a): Since, by Assumption A9, initial mal level y of inventory after production in a month, is at y z =
capacity is zero, and, by Theorem 1(c), L1 > 0, capacity will be set at It y+Eft+1 y −Dt  z, which is strictly concave, smooth, and does
L1 in month 1. By Theorem 2(d), the lower-capacity targets increase not depend on the initial inventory level x. Let St z denote the
up through month tE , so those targets will be the capacity levels in unique maximizer of this objective function over y ∈ 0 . The
those months. optimal level is therefore given by a base stock policy with base
We show that the capacity level remains constant at LtE in stock level St z: Get as close to St z as possible within the fea-
months t = tE  tE + 1     tC by showing that LtE remains between sible interval x x + z
. Because ft+1 is submodular, so is at , and,
the capacity targets during those months: Indeed, we have therefore, St z is decreasing in z.
Because cC is a convex function and x y z x ≥ 0 z ≥ 0 x ≤ y ≤
LtE ≤ UtE ≤ UtE + 1 ≤ · · · ≤ UtP ≥ UtP +1 ≥ · · · ≥ UtC ≥ LtE x + z is a convex set, it is straightforward to show, by the convexity
preservation theorem under minimization (e.g., Prop. B-4, Heyman
and and Sobel 1984) that t x z and ft x z are (jointly) concave in
LtE ≥ LtE +1 ≥ · · · ≥ LT  x z. We now show that t is submodular. We first write t as
follows: 
We show by induction that the capacity level equals the upper 

at x z if St z < x
target for t = tC + 1 tC + 2     T . In month tC + 1, by definition t x z = at St z z if x ≤ St z ≤ x + z
of tC , we have ztC +1 = UtC +1 because LtE > UtC +1 . The rest follows 

a x + z z if x + z ≤ S z
because, by Theorem 2(b), UtC +1 > UtC +2 > · · · > UT . t t

Part (b): If tC > tE , then yt is constant in at least two consecu- We can therefore write the partial derivative with respect to x as
tive months in which demand in one month strictly dominates that follows:
in the other. Hence, the service levels in those months is strictly 
different. If tC = tE , then the service level in at least one month is  1
at x + z z if x ≤ St z and z ≤ St z − x

given by (7) as p + cS − cP − hC − K − k
/p + cS − pD . The service 1
t x z = 0 if x ≤ St z and z > St z − x


level in at least one other month is given by either (5) or (6), and, a1 x z if x > St z
t
because k < K, both of these strictly differ from that given in (7).

1
We must show that t x z is decreasing in z. Because at is sub-
modular, the only case that is not obvious is when x ≤ St z and z
Appendix B. Proof for the Case of crosses the line from below St z − x to above it. Hence, we must
Inventory Carry-over 1
show that at x +z z ≥ 0 when z ≤ St z−x. But that holds because
Assume that Assumptions B1–B7 hold henceforth in this appendix, 1
at is concave, at St z z = 0, and x + z ≤ St z.
unless indicated otherwise. The optimality equations can be writ- Because Lt x and Ut x are respective maximizers of gt x z and
ten as ht x z over z ≥ 0, and these objective functions are submodular,

 we can take Lt x and Ut x to be decreasing functions of x.
ft x z = max cP x − cC z − z − hC z + t x z  We can write ft as follows.
z ≥0


= max Kz + max gt x z  kz + max ht x z  
z ≥z 0≤z ≤z cP x − KLt x − z





 − hC Lt x + t x Lt x if z ≤ Lt x
where 

ft x z = cP x − hC z + t x z if Lt x ≤ z ≤ Ut x


gt x z = cP x − Kz − hC z + t x z 


 c x − kUt x − z

 P

ht x z = cP x − kz − hC z + t x z − hC Ut x + t x Ut x if Ut x ≤ z

412 Management Science/Vol. 48, No. 3, March 2002

 ANGELUS AND PORTEUS
Simultaneous Capacity and Production Management of Short-Life-Cycle Goods

We next have the following. Federgruen A., P. Zipkin. 1986a. An inventory model with limited
 production capacity and uncertain demands I. The average-

K if z ≤ Lt x cost criterion. Math. Oper. Res. 11 193–207.

2
f x z = 2
−h +  x z if Lt x ≤ z ≤ Ut x . 1986b. An inventory model with limited production capacity
t
 C

t
k if Ut x ≤ z and uncertain demands II. The discounted-cost criterion. Math.
Oper. Res. 11 208–215.
Thus, since ft is concave, the marginal value of capacity is between Harrison, J. M., J. A. Van Mieghem 1999. Multi-resource invest-
K and k. We must still show that ft is submodular, which holds if ment strategies: Operational hedging under demand uncer-
2 tainty. Eur. J. Oper. Res. 113 17–29.
we can show that ft x z decreases in x for each fixed z. Consider
the most comprehensive case in which z < Lt 0. As x increases, Hax, A., D. Candea. 1984. Production and Inventory Management.
Lt x decreases, and there may be a point, say x1 , such that z = Prentice-Hall Inc., Englewood Cliffs, NJ.
2 Heyman, D., M. Sobel. 1984. Stochastic Models in Operations Research,
Lt x1 . We must show that ft decreases across this boundary.
2
That is, we must show that K ≥ −hC + t x1  Lt x1 . This follows Volume II. McGraw-Hill, New York.
2
since Lt x1  maximizes gt x1  · and gt x1  Lt x1  = −K + hC  + Kapuściński, R., S. Tayur. 1998. A capacitated production-inventory
2 2
t x1  Lt x1  = 0. That ft x z decreases in x while Lt x ≤ z ≤ model with periodic demand. Oper. Res. 46 899–911.
Ut x follows from t being submodular. Finally, there may be a Karush, W. 1958. A theorem in convex programming. Naval Res.
2
point, say x2 , such that z = Ut x2 . That ft decreases across this Logist. Quart. 5 245–260.
2
boundary follows from gt x2  Ut x2  = 0. Kouvelis, P., J. Milner. 2001. Supply chain capacity expansion and
Finally, smoothness of ft follows from examining partial deriva- outsourcing: The dynamic interplay of demand and supply
tives at the boundaries between regions of differing definitions. uncertainty. Research paper, Olin School of Business, Washing-
ton University, St. Louis, MO.
Li, S., D. Tirupati. 1994. Dynamic capacity expansion problem with
multiple products: Technology selection and timing of capacity
References additions. Oper. Res. 42 958–976.
Angelus, A., E. Porteus, S. Wood. 2000. Optimal sizing and tim- Luss, H. 1982. Operations research and capacity expansion prob-
ing of modular capacity expansions. Research paper, Stanford lems: A survey. Oper. Res. 30 907–947.
Business School, Stanford, CA. Manne, A. S. 1961. Capacity expansion and probabilistic growth.
Aviv, Y., A. Federgruen. 1997. Stochastic inventory models with lim- Econometrica 29 633–649.
ited production capacity and varying parameters. Prob. Engrg. Rajagopalan, S., A. C. Soteriou. 1994. Capacity acquisition and dis-
Inform. Sci. 11 107–135. posal with discrete facility sizes. Management Sci. 40 903–917.
Bean, J. C., R. L. Smith. 1985. Optimal capacity expansion over an Rajagopalan, S., M. R. Singh, T. E. Morton. 1998. Capacity expan-
infinite horizon. Management Sci. 31 1523–1532. sion and replacement in growing markets with uncertain tech-
Bean, J. C., J. L. Higle, R. L. Smith. 1992. Capacity expansion under nological breakthroughs. Management Sci. 44 12–30.
stochastic demands. Oper. Res. 40 210–216. Rao, M.R. 1976. Optimal capacity expansion with inventory. Oper.
Buzacott, J. A., A. B. Chaouch. 1988. Capacity expansion with inter- Res. 24 291–300.
rupted demand growth. Eur. J. Oper. Res. 34 19–26. Rocklin, S. M., A. Kashper, G. C. Varvaloucas. 1984. Capacity
Eberly, J. C., J. Van Mieghem 1997. Multi-factor dynamic investment expansion/contraction of a facility with demand augmentation
under uncertainty. J. Econom. Theory 75 345–387. dynamics. Oper. Res. 32 133–147.
Erlenkotter, D., S. Sethi, N. Okada. 1989. Planning for surprise: Topkis, D. 1978. Minimizing a submodular function on a lattice.
Water resources development under demand and supply Oper. Res. 26 305–321.
uncertainty I. The general model. Management Sci. 35 149–163. Topkis, D. 1998. Supermodularity and Complementarity. Princeton
Evans, R. 1967. Inventory control of a multiproduct system with University Press, Princeton, NJ.
a limited production resource. Naval Res. Logist. Quart. 5 Van Mieghem, J. A. 1998. Investment strategies for flexible
137–154. resources. Management Sci. 44 1071–1078.

Accepted by Luk Van Wassenhove; received January 3, 2001. This paper was with the authors 4 months for 1 revisions.

Management Science/Vol. 48, No. 3, March 2002 413