A Power-of-two Ordering Policy for One Warehouse, Multi-Retailer

Systems with Stochastic Demand

Leon Yang Chu Zuo-Jun Max Shen
Department of Industrial & Systems Engineering, University of Florida, Gainesville, FL 32611
Department of Industrial Engineering & Operations Research, University of California, Berkeley, CA 94720
May 15, 2005
Abstract
We study a two-echelon supply chain with one warehouse and N retailers facing stochastic
demand. An easy-to-implement inventory policy, the so-called power-of-two policy, is proposed
to manage inventory for the system. To maintain a certain service level, safety stocks are kept
at the warehouse and each retailer outlet to buer random demand. Our analysis highlights the
important role of warehouse safety stock level, which, in addition to the length of the warehouse
order interval, signicantly aect the length of the retailers order interval. We develop a
polynomial algorithm to nd a power-of-two ordering policy for arbitrary target service level,
whose cost is guaranteed to be no more than
3

2 1.26 times the optimal cost. Such a policy

can be computed in time O(N
3
).
1 Introduction
We consider a distribution system with one warehouse and multiple retailers. Random demand
occurs at the retailers only. The warehouse orders from an outside supplier with unlimited supply
and replenishes the retailers inventories. To maintain appropriate service levels, safety stock is
1
maintained at the supplier and the retailers. The warehouse and each retailer follow a periodic
review inventory policy with certain service requirements. The planning horizon is innite and the
objective is to minimize the long-run average system-wide cost.
Multi-echelon stochastic inventory models have received much attention in the research com-
munity. In terms of exact results, most papers focus on continuous review policy. For example,
Axsater (2000) provides exact results for compound Poisson demand and nonidentical retailers
when the warehouse and the retailers use the (R, nQ) policy. Cheung and Hausman (2000) analyze
exact results for the central warehouse serving nonidentical retailers. Chen and Zheng (1997) pro-
vide exact results when the central warehouse uses continuous review reorder point policy, while
the reorder point is calculated according to echelon stock inventory. For periodic review models,
Liljenberg (1996) and Cachon (2001) provide exact valuation methods for identical retailers under
batch ordering policy.
It is recognized that the structure of an optimal policy for most multi-echelon stochastic inven-
tory systems is either unknown or extremely complex, thus, it is dicult to implement them in
practice. As a result, much research has focused on heuristic policies that are easier to implement.
For excellent reviews of this literature, we refer the readers to Axsater (1993c) and Federgruen
(1993). Recent developments include Lee and Billington (1993), Tempelmeier (1993), Hausman
and Erkip (1994), Chen and Samroengraja (2000).
However, since these policies are only heuristic, a natural question to ask is how much is the gap
between the heuristic solution and the optimal one. The literature only provides some numerical
evidence (e.g., Federgruen and Zipkin 1984a,b and Chen and Zheng 1993b,1994a,b). But no matter
how extensively the numerical experiments have been conducted, there are no guarantees with
respect to the heuristics performance on other problem instances. Therefore, it is much more
desirable to have a heuristic policy with a good worst-case performance, i.e., a good a priori
2
guarantee for its performance on any problem instance.
For a single-echelon continuous-review inventory system with stationary stochastic demand,
Zheng (1992) shows that the maximum relative error is bounded by 0.125 when using the deter-
ministic EOQ formula as a heuristic solution in the (Q, r) policy. Axsater (1996) gives a slightly
better bound of 0.118. Levi, Roundy, and Shmoys (2004b) analyze stochastic inventory prob-
lems with correlated and non-stationary demands. They provide a 2-approximation algorithm for
the periodic-review stochastic inventory control problem and a 3-approximation algorithm for the
stochastic lot-sizing problem.
Unfortunately, for multi-echelon stochastic inventory systems, we have not seen any work in
this area. In contrast, for a wide range of multi-echelon, deterministic inventory systems, it is well
known that a class of easily implementable policies, the so-called power-of-two policies, perform
surprisingly well (see, e.g., Maxwell and Muckstadt 1985, and Roundy 1985,1986). Under a power-
of-two policy, each item is replenished at constant reorder intervals which are power-of-two multiples
of some xed or variable base planning period. An optimal power-of-two policy can be found very
easily and its cost is guaranteed to be within 2% of optimality. Recent works that analyze worst-case
bound performance of other easy-to-implement policies include Chan, Muriel, Shen, Simchi-Levi,
and Teo (2002), Chen (2000), Levi, Roundy, and Shmoys (2004a), and Shen, Simchi-Levi, and Teo
(2002).
In this paper, we apply the power-of-two policy to OWMR models with stochastic demand. To
ensure a certain service level in this random demand environment, safety stocks are needed at the
warehouse and each retailer site. Inventory replenishment decisions are then based on the average
demand rates, similar to the spirits of Zheng (1992) and Axsater (1996). We develop a polynomial
algorithm to nd a close-to-optimal power-of-two ordering policy for arbitrary target service level.
Our approach highlights the important role of warehouse safety stock level, which, in addition to
3
the length of the warehouse order interval, signicantly aect the length of the retailers order
interval.
Since it is often very dicult in practice to assess shortage costs for an external customer, we
adopt a common service denition that species the probability that a demand or a collection of
demands is met. We assume that the demand at each retailer during the lead time is a continuous
random variable D
i
. Let
i
=
_
var(D
i
) be the standard deviation of demand during the lead time.
Then the amount of safety stocks required can be calculated as z
i
, where the value of z depends
on the service requirement and the type of service measurement. For instance, if we specify a 90%
service level, then z = 1.28 for Type 1 service level (Nahmias, 1997). We adopt Type 1 service level
measurement in this paper, and we also discuss how to extend the results to systems with Type 2
service level.
The rest of this paper is organized as follows: Section 2 describes the power-of-two inventory
model. Section 3 discusses the OWMR model without power-of-two constraints. Based on this
relaxation model, we discuss how to construct a close-to-optimal power-of-two ordering policy in
Section 4. Section 5 describes briey how our approach can be extended to systems with Type 2
service level requirements. We conclude in Section 6.
2 Model
We use term facility to refer to either the warehouse or a retailer. Facility 0 is the warehouse and
retailers are facilities 1 through N. Let the holding cost rate and the xed setup cost at retailer
n, n = 1, . . . , N, be h

n
, K
n
, and at the warehouse be h
0
, K
0
. Let the average demand at retailer n
be
n
and the variance
2
n
per unit time. We dene echelon holding cost rate h
n
h

n
h
0
, which
can also take negative values (salvage values). Order lead time is L
0
at the warehouse and L
n
for
each retailer n, n = 1, . . . , N.
4
A power-of-two (POT) policy is one in which each facility n places an order once every T
n
> 0
units of time beginning at t = 0, and the ratio T
n
/T
0
is an integer power of two. In this model, the
retailers orders are synchronized, i.e., all N retailers order in the same periods. Lee et al. (1997)
and Cachon (1999) study the scheduled ordering policies in which the retailers orders are balanced,
i.e., the same number of retailers order at each period. These researches demonstrate that the
scheduled ordering policies reduce supplier demand variance and may lead to lower total supply
chain cost. However, as pointed out by Cachon, it is not clear that balancing the retailer order
intervals will lower costs. With synchronized ordering the supplier can anticipate a large demand
every T
0
periods, so the supplier can arrange to have inventory arrive just before the retailers order.
Therefore, we rst focus on the synchronized ordering policy in this paper, and leave the scheduled
ordering policy as one important future research direction.
An POT inventory control policy for the system can be characterized by an (N + 1)-tuple,

T = (T
0
, T
1
, , T
N
), and the corresponding average cost can be written as
c(

T) K
0
/T
0
+ h
0
s
0
+

1nN
(c
n
(T
0
, T
n
) + h

n
s
n
)
where c
n
(T
0
, T
n
) includes the setup costs and cycle inventory costs associated with retailer n, and
can be approximated by c
n
(T
0
, T
n
) = K
n
/T
n
+

n
2
(h
n
T
n
+ h
0
(T
0
T
n
)) where x y denotes the
larger of x and y. (Note for the deterministic demand case, setup costs and cycle inventory costs
associated with retailer n is K
n
/T
n
+

n
2
(h
n
T
n
+ h
0
(T
0
T
n
)), and we can continue to use this
formula under stochastic demand when we have a high service level and the probability of out-of-
stock can be ignored in cost computation.) The safety stock level s
n
at retailer n can be written
as
n
_
(L
n
+ T
n
)
2
n
=
n

L
n
+ T
n
, where
n
depends on the specied Type 1 service level. To
calculate the safety stock level at the warehouse, let us rst examine how each retailer contributes
5
to the variance of order at the warehouse.
Case 1. T
n
T
0
: Retailer n orders less frequently than the warehouse does. Each time retailer
n orders, the order introduces variance L
0

2
n
to the warehouse.
Case 2. T
n
< T
0
: Retailer n orders more frequently than the warehouse does, and whenever
the warehouse orders, retailer n also orders. After each warehouses order, retailer n would order
T
0
/T
n
1 more times until warehouses next order. In this case, retailer ns orders introduce
variance (L
0
+ T
0
T
n
)
2
n
to the warehouse.
The maximum safety stock level for warehouse occurs when all the retailers order at the same
time, and the corresponding safety stock level s
0
is
0
_
L
0

n

2
n
+

n
(0 (T
0
T
n
))
2
n
, where
0
depends on the specied service level. This square root formula can be used for any normalized
demand distribution (Eppen, 1981), or it can be used as a good pooling approximation for demand
distributions that are not normalized. For single-parameter demand distributions (e.g. Poisson)
or additive two-parameter demand distributions that have a constant variance-to-mean ratio, it is
possible to construct a appropriate safety stock pooling function numerically by plotting the total
amount of stock required to achieve the given customer service level (as a function of the average
lead time demand) then subtracting the average lead time demand. See Caggiano et al. (2002) for
more discussions.
Our problem can be formulated as follows:
Minimize c(

T) K
0
/T
0
+ h
0
s
0
+

1nN
(c
n
(T
0
, T
n
) + h

n
s
n
)
Subject to T
n
> 0 for n = 0, 1, , N
T
n
/T
0
is an integer power of two
6
where
c
n
(T
0
, T
n
) := K
n
/T
n
+

n
2
(h
n
T
n
+ h
0
(T
0
T
n
)),
s
n
:=
n

L
n
+ T
n
, and
s
0
:=
0
_
L
0

n

2
n
+

n
(0 (T
0
T
n
))
2
n
,
3 Relaxation of the OWMR Model
Now let us rst drop the power-of two ratio constraints that specify T
n
/T
0
must be an integer
power of two, and minimize the corresponding relaxation problem. Let ( denote the retailer set
with T
n
> T
0
, c the retailer set with T
n
= T
0
, and L the retailer set with T
n
< T
0
. We solve this
relaxation problem by discomposing it into smaller subproblems.
3.1 Determining (, c, and L
Note if
n
= 0, for all n = 1, . . . , N, our problem reduces to the original problem in Roundy (1985),
and we can follow his approach to decide whether a retailer belongs to (, c, or L. In the following,
we assume
n
> 0 for n = 1, . . . , N.
Let f
n
(x) = K
n
/x +

n
2
h

n
x + h

L
n
+ x. For T
n
T
0
, f
n
(T
n
) = c
n
(T
0
, T
n
) + h

n
s
n
characterizes the cost at retailer n. By Corollary 1 in Appendix, the rst order condition of
f
n
(x) has unique solution, which minimizes f
n
(x). Denote this minimum solution

n
. Dene
g
n
(x) = K
n
/x +

n
2
h
n
x + h

L
n
+ x For T
n
T
0
, g
n
(T
n
) = c
n
(T
0
, T
n
) + h

n
s
n


2
h
0
T
0
characterizes the cost at retailer n. By Lemma 5 in Appendix, g

n
(x) = 0 has at most two solutions,
and g
n
(x) has at most one local minimum. If g
n
(x) has one local minimum, denote it as
n
; if
g(x) has no local minimum, then g
n
(x) is an decreasing function and we dene
n
= . f

n
(x) is
negative in (0,

n
) and positive in (

n
, ), while g

n
(x) is negative on (0,
n
). g

n
(
n
) = 0 if
n
< .
Lemma 1
n
>

n
for all n.
Proof: Since g

n
(x) = f

n
(x)

n
2
h
0
< f

n
(x), and f

n
(x) < 0 on (0,

n
), we know that g

n
is negative
7
on (0,

n
), and g

n
(

n
) < 0. As either
n
= or g

n
(
n
) = 0, we conclude that
n
>

n
.
To determine whether n belongs to (, c, or L, that is, whether T
n
is greater, equal to, or less
than T
0
, we rst need to know how T
n
inuences the total cost. Note there are four terms in the
cost function that contain T
n
: K
n
/T
n
,

n
2
(h
n
T
n
+h
0
(T
0
T
n
)), h

L
n
+ T
n
, and h
0
s
0
, where
s
0
=
0
_
L
0

n

2
n
+

n
(0 (T
0
T
n
))
2
n
.
We observe that
If T
n
T
0
, the warehouse stock level s
0
is independent of T
n
, and the total cost is a function
of T
n
that varies according to f
n
(T
n
), that is,
c(

T)
T
n
=
f
n
(T
n
)
T
n
.
If T
n
T
0
, s
0
is a decreasing function of the order interval T
n
. We may explicitly express
the relationship by writing s
0
(T
n
) instead of s
0
. The total cost varies according to g
n
(T
n
) +
h
0
s
0
(T
n
), that is,
c(

T)
T
n
=
(g
n
(T
n
)+h
0
s
0
(T
n
))
T
n
.
Proposition 1 n ( if and only if T
0
<

n
.
Proof: Let T

n
be the optimal T
n
for given T
0
.
only if
=
: suppose T
0

n
, f
n
(T
n
) is a strictly increasing function on (T
0
, ). Since the total cost
changes according to f
n
(T
n
), we can decrease the total cost by setting T

n
equal to T
0
if T

n
> T
0
.
Thus, T

n
T
0
, and n L c.
if
=
: suppose T
0
<

n
, we show that T

n
=

n
> T
0
and n (. Recall that when T
n
T
0
,
the total cost varies according to g
n
(T
n
) + h
0
s
0
(T
n
). The derivative of the total cost is (g
n
(T
n
) +
h
0
s
0
(T
n
))

= f

n
(T
n
)

n
2
h
0
h
0

2
0

2
n
2s
0
. Since f
n
(T
n
) is a decreasing function on (0,

n
) and 0 < T
0
<

n
, f

n
(T
n
) 0 for T
n
(0, T
0
), thus (g
n
(T
n
) + h
0
s
0
(T
n
))

< 0 and (g
n
(T
n
) + h
0
s
0
(T
n
)) is a strictly
decreasing function on (0, T
0
). If T

n
< T
0
, we can decrease the total cost by setting T

n
equal to T
0
,
which is an contradiction. Thus, T

n
T
0
, and the total cost changes according to f
n
(T
n
). Since
8

n
minimizes f
n
(T
n
); therefore T

n
=

n
> T
0
, and n (.
Proposition 2 n c if

n
T
0

n
.
Proof: Since

n
T
0
, n L c. T

n
(0, T
0
] minimizes the total cost, which is equivalent to min-
imizing (g
n
(T
n
) + h
0
s
0
(T
n
)) on (0, T
0
]. Since, T
0

n
, g
n
(T
n
) is a decreasing function on (0, T
0
].
Note s
0
(T
n
) is also a decreasing function. Thus, to minimize cost, T

n
equals T
0
, that is, n c.
If the demand is deterministic, our denitions of
n
and

n
are consistent with the corresponding
denitions in Roundy (1985). In this case, n ( if and only if T
0
<

n
; n c if and only if

n
T
0

n
; and n L if and only if
n
< T
0
. However, with stochastic demand, if T
0
>
n
, n
may belong to c or L depending on the value of T
0
, and the level of warehouse safety stock s
0
. We
illustrate this observation graphically in Figure 1 where the boundary function r
n
(T
0
) is used to
determine whether n L or c. We explain how to dene r
n
(T
0
) and eciently enumerate all the
possible partitions in the remainder of this section.
First note that if T
0
>
n
, then T
0
>

n
, T

n
T
0
by Proposition 1, and the total cost
varies according to g
n
(T
n
) + h
0
s
0
(T
n
), where s
0
=
0
_
L
0

n

2
n
+

n
(0 (T
0
T
n
))
2
n
. De-
ne C
n
= (L
0

i

2
i
+

i=n
(0 (T
0
T
i
))
2
i
), then we can also express s
0
as follows: s
0
=

0
_
C
n
+ (0 (T
0
T
n
))
2
n
.
To determine whether n c or L, we need to compare g
n
(T
0
)+h
0
s
0
(T
0
) with g
n
(T
n
)+h
0
s
0
(T
n
)
for T
n
(0, T
0
). Dene g
n
(T
n
) h
0
s
0
(T
n
) +h
0
s
0
(T
0
) +g
n
(T
0
), it suces to compare g
n
(T
0
) and
g
n
(T
n
) on (0, T
0
). Note g
n
(T
n
) depends not only on T
n
, but also on T
0
and C
n
.
In Figure 2, we draw g
n
(T
n
) and g
n
(T
n
) as functions of T
n
on (0, T
0
]. Note that g
n
(T
n
) ap-
proaches positive innity when T
n
goes to zero, while g
n
(T
n
) g
n
(T
0
) for all T
n
(0, T
0
). Thus,
g
n
(T
n
) is either below, or intersecting, or tangent with g
n
(T
n
) on (0, T
0
).
9
0
T
0
s
0

n

n
G
E
L
r
n
(T
0
)
Figure 1: The relationship of n (, c or L with given T
0
and s
0
If g
n
(T
n
) is below g
n
(T
n
) on (0, T
0
), that is, g
n
(T
n
) > g
n
(T
n
) g
n
(T
n
) + h
0
s
0
(T
n
) >
g
n
(T
0
) + h
0
s
0
(T
0
) for T
n
(0, T
0
). This means that T
0
minimizes g
n
(.) + h
0
s
0
(.) on (0, T
n
].
Furthermore, T

n
(0, T
0
] minimizes the total cost, which is equivalent to minimize (g
n
(T
n
) +
h
0
s
0
(T
n
)) on (0, T
0
]. Thus, T

n
= T
0
, and n c.
If g
n
(T
n
) intersects with g
n
(T
n
), that is, g
n
(T
n
) + h
0
s
0
(T
n
) < g
n
(T
0
) + h
0
s
0
(T
0
) for some
T
n
(0, T
0
). T

n
lies in (0, T
0
), and n L.
If g
n
(T
n
) is tangent with g
n
(T
n
) on (0, T
0
) without intersecting with g
n
(T
n
), that is, g
n
(T
n
) +
h
0
s
0
(T
n
) has multiple minimum solutions on (0, T
0
], and n can belong to either c or L.
The above discussion sheds light on how to determine whether n c or L. However, it is still
not easy to implement since g
n
(T
n
) depends not only on T
0
, but also on C
n
, the variance of order
at the warehouse contributed by other retailers. We hope to nd a better way to determine n c
10
T
n
(T
0
, g
n
(T
0
))
g
n
(T
n
)
g
n
(T
n
)

Figure 2: g
n
(T
n
) and g
n
(T
n
) at C
n
= C

n
or L without calculating C
n
. To do this, let us rst investigate how C
n
aects g
n
(T
n
). Note that
for a xed T
0
, if we decrease the value of C
n
, g
n
(T
n
) will shift downward; if we increase the value of
C
n
, g
n
(T
n
) will shift upward; and if we keep on increasing C
n
till it approaches positive innity,
g
n
(T
n
) will converge to g
n
(T
0
) for each T
n
[0, T
0
], that is, g
n
(T
n
) will become a horizontal line.
Since g
n
(T
n
) approaches positive innity as T
n
approaches zero and g
n
(
n
) < g
n
(T
0
) with
n
< T
0
,
g
n
(T
n
) intersects with g
n
(T
n
) on [0, T
0
]. Therefore, when C
n
approaches positive innity, T
n
has
little impact on the the warehouse safety stock level s
0
; and n L if and only if T
0
>
n
as in
Roundy (1985) under this scenario.
For given T
0
, we are interested in the value of C
n
at which n shifts from L to c. If we decrease
the value of C
n
to zero, there are only two possible scenarios:
I) g
n
(T
n
) intersects with g
n
(T
n
) on (0, T
0
).
11
II) g
n
(T
n
) is below or tangent with g
n
(T
n
) on (0, T
0
).
In this case, there exists unique C

n
such that g
n
(T
n
) is tangent with g
n
(T
n
) when C
n
= C

n
.
According to the dierent positions of the tangent points, we have the following two sub-cases:
IIa) T
0
, the end point, is the unique tangent point when we set C
n
= C

n
.
IIb) There exists internal tangent point

T
n
(0, T
0
) when we set C
n
= C

n
. Let

T
n
be the
rightmost tangent point on (0, T
0
) if there are more than one internal tangent point.
Let us rst focus on Scenario IIb) when the tangent point is

T
n
(0, T
0
), and see how we can
determine whether n c or L according to the values of T
0
and C
n
. Based on the relative values
of C
n
and C

n
, we have the following three cases:
C
n
< C

n
: g
n
(T
n
) is below g
n
(T
n
) on (0, T
0
). Recall that g
n
(T
n
) > g
n
(T
n
) g
n
(T
n
) +
h
0
s
0
(T
n
) > g
n
(T
0
) + h
0
s
0
(T
0
) for T
n
(0, T
0
), this means that T
0
minimizes g
n
(.) + h
0
s
0
(.)
on (0, T
n
]. We also know that T

n
minimizes g
n
(T
n
) + h
0
s
0
(T
n
) on (0, T
0
], thus T

n
= T
0
and
n c. The corresponding warehouse safety stock level is s
0
=
0

C
n
<
0
_
C

n
.
C
n
= C

n
: g
n
(T
n
) is tangent with g
n
(T
n
) on

T
n
(0, T
0
). g
n
(T
n
) + h
0
s
0
(T
n
) has multiple
minimum solutions on (0, T
0
]. Since T

n
minimizes g
n
(T
n
) + h
0
s
0
(T
n
) on (0, T
0
], one possi-
ble value of T

n
is T

n
= T
0
and n c, then s
0
=
0
_
C

n
; another possible value is T

n
=

T
n
(or
other tangent points) and n L, then s
0
=
0
_
C

n
+ (T
0

T
n
)
2
n
_
or s
0

0
_
C

n
+ (T
0

T
n
)
2
n
_
.
C
n
> C

n
: g
n
(T
n
) intersects with g
n
(T
n
) on

T
n
(0, T
0
). Since T

n
minimizes g
n
(T
n
) +
h
0
s
0
(T
n
) on (0, T
0
], T

n
< T
0
, and n L. Furthermore, by Lemma 10 in Appendix, the
optimal T

n
is no more than

T
n
, then s
0
=
0
_
C
n
+ (T
0
T

n
)
2
n
>
0
_
C

n
+ (T
0

T
n
)
2
n
.
Note that the optimal warehouse safety stock level s

0
is either smaller or equal to
0
_
C

n
when
n c, or larger or equal to
0
_
C

n
+ (T
0

T
n
)
2
n
when n L, and s

0
will never pick a value
12
within
_

0
_
C

n
,
0
_
C

n
+ (T
0

T
n
)
2
n
_
. Thus, if we can dene a function r
n
(T
0
) whose value
is always between
0
_
C

n
and
0
_
C

n
+ (T
0

T
n
)
2
n
for each retailer n, then for retailer n, if
T
0
>
n
, n c if and only if s

0
r
n
(T

0
) and n L if and only if s

0
> r
n
(T

0
). This way, we can
use T
0
and s
0
to nd out whether n c or L without calculating C
n
, whose value varies with n.
T
n
T
n

slope: h
0

n
2
/2/(C
n
*
+(T
0
T
n
)
n
2
)
1/2

slope: h
0

n
2
/2/(C
n
*
)
1/2

(T
0
,g
n
(T
0
))
g
n
(T
n
)
g
n
(T
n
)

slope: h
0

0
2

n
2
/2/r
n
(T
0
)
Figure 3: The slopes of the tangent lines at T
n
= T
0
and T
n
=

T
n
Next we show such function r
n
(T
0
) does exist. In Figure 3, we draw a line through (T
0
, g
n
(T
0
))
such that this line is tangent with g
n
(T
n
) and lies below g
n
(T
n
) on (0, T
0
). We dene r
n
(T
0
) such
that h
0

2
0

2
n
2r
n
(T
0
)
is the slope of the tangent line.
Lemma 2
0
_
C

n
r
n
(T
0
) <
0
_
C

n
+ (T
0

T
n
)
2
n
Proof: We x C
n
= C

n
. h
0
s
0
(T
n
) is a convex function of T
n
; the slope (derivative with regard
to T
n
) is (h
0
s
0
(T
n
))

= h
0

2
0

2
n
2s
0
(T
n
)
, which is an increasing function of T
n
.
Because g
n
(T
n
) is convex and below g
n
(T
n
) on (0, T
0
] at C
n
= C

n
, the slope at T
0
, h
0

2
0

2
n
2s
0
(T
0
)
=
13
h
0

2
0

2
n
2
0

n
, is no less than the slope of the tangent line h
0

2
0

2
n
2r
n
(T
0
)
. That is,
0
_
C

n
r
n
(T
0
).
Meanwhile, the slope at

T
n
, h
0

2
0

2
n
2s
0
(

T
n
)
= h
0

2
0

2
n
2
0
_
C

n
+(T
0

T
n
)
2
n
, is less than the slope of the tangent
line. That is, r
n
(T
0
) <
0
_
C

n
+ (T
0

T
n
)
2
n
.
Thus,
0
_
C

n
r
n
(T
0
) <
0
_
C

n
+ (T
0

T
n
)
2
n
. Figure 3 shows the graphic proof and how
to represent r
n
(.) as a function of T
0
.
Thus if T
0
>
n
, we can claim that under Scenario IIb), if s
0
r
n
(T
0
), then n c, and if
s
0
> r
n
(T
0
), then n L.
This statement is also true for Scenario IIa), where g
n
(T
n
) is tangent with g
n
(T
n
) at T
0
when
C
n
= C

n
. Under this scenario, r
n
=
0
_
C

n
. When n c, C
n
C

n
, and we have
s
0
=
0

C
n
, which is indeed no more than r
n
=
0
_
C

n
. On the other hand, when n L, that
is C
n
> C

n
, we have s
0
>
0

C
n
, which is larger than r
n
=
0
_
C

n
. Thus, if s
0
r
n
(T
0
),
then n c, and if s
0
> r
n
(T
0
), then n L.
Furthermore, this statement is also true for Scenario I) where g
n
(T
n
) intersects with g
n
(T
n
)
when we decreases C
n
to zero. Under this scenario, n L, T

n
minimizes g
n
(T
n
) + h
0
s
0
(T
n
) on
(0, T
0
). Since g
n
(T
n
) = h
0
s
0
(T
n
) + h
0
s
0
(T
0
) + g
n
(T
0
), T

n
maximizes g
n
(T
n
) g
n
(T
n
) on (0, T
0
),
(T

n
, g
n
(T

n
)) lies above g
n
(T
n
) as g
n
(T
n
) intersects with g
n
(T
n
). Since the tangent line is either
below or tangent with g
n
(T
n
), (T

n
, g
n
(T

n
)) lies above the tangent line. In the following, we show
that the optimal warehouse safety stock level s

0
is greater than r
n
(T
0
). Suppose s

0
is no more than
r
n
(T
0
), then h
0

2
0

2
n
2s

0
, the slope of g
n
(T
n
) at T

n
, is no less than the slope of the tangent line, h
0

2
0

2
n
2r
n
.
Since s
0
(T
n
) is concave, then g
n
(T
n
) is convex, and the slope of g
n
(T
n
) is no less than the slope of
the tangent line on [T

n
, T
0
]. Therefore, g
n
(T
n
) is above the tangent line on [T

n
, T
0
]. However, both
g
n
(T
n
) and the tangent line pass through (T
0
, g
n
(T
0
)). Contradictory! Thus, the optimal warehouse
safety stock level s

0
> r
n
, and we can still use r
n
to determine whether n c or L.
14
Finally, we can claim that if T
0
>
n
, then n c if s
0
r
n
(T
0
), and n L if s
0
> r
n
(T
0
) under
all possible scenarios.
The curve s
0
= r
n
(T
0
) and the line T
0
=

n
for retailer n cut the place into three areas. A total
of N curves and N lines cut the plane into many small areas. Each small area corresponds to one
possible partition of (, c, and L. In the next, we show that there are at most O(n
2
) areas.
Theorem 1 There exists a constant k, such that for any two retailers i and j, r
i
(T
0
) and r
j
(T
0
)
do not intersect for more than k times.
Proof: See Appendix.
By Theorem 1, a curve r
i
(T
0
) intersects no more than k times with another curve r
j
(T
0
), and
a curve for a particular retailer may have at most O(n) intersections. Thus, there will be at most
O(n
2
) intersections, and we can nd the T
0
value for each intersection since each curve r
n
(T
0
) can
be represented by a polynomial by Lemma 11 in the Appendix. By ranking these O(n
2
) values of

n
s and

n
s, we can divide the T
0
axis into O(n
2
) segments. Starting from the leftmost interval,
which corresponds to all n (, as we increase the value of T
0
and move from left to right, we may
introduce new partitions as T
0
crosses

n
,
n
or an intersection point.
The following example illustrates how we can enumerate all the possible partitions eciently.
To keep the example simple, we assume there are only two retailers and r
1
and r
2
intersects only
once at T
12
as shown in Figure 4.
1
,

1
,
2
,

2
, and T
12
divide the T axis into six intervals. We
starts at T
0
= 0 and continuously increase the value of T
0
till it passes T
12
, and we discuss the
possible partition changes during this process.
(0,

1
): the only possible partition is ((, c, L)= (1, 2, , ).
(

1
,
1
): Retailer 1 is moved from ( to c, and we get a new partition ((, c, L)= (2, 1, ).
15
T
0
s
0

1

1

2

2

T
12

{1,2},, {2},{1},
{2},{1},
r
1
(
0
)
r
2
(
0
)

{2},,{1} ,{2},{1}
,{1,2},
,,{1,2}
,{1},{2}
,{1,2},
Figure 4: Example of Enumerating Possible Partitions
(
1
,

2
): A new partition, (2, , 1), is added for the case s
0
> r
1
(T
0
), in which 1 L and
the rest of retailers remains unchanged. Now we have two possible partitions, (2, , 1)
and (2, 1, ).
(

2
,
2
): Retailer 2 is removed from ( to c for all the partitions in the previous interval. Thus,
we have two possible partitions, (, 2, 1) and (, 1, 2, ).
(
2
, T
12
): A new partition, (, , 1, 2), is added for the case s
0
> r
2
(T
n
) in which 2 L and
the rest of retailers remains unchanged. Now we have three possible partitions, (, , 1, 2),
(, 2, 1), and (, 1, 2, ).
(T
12
, ): A new partition (, 1, 2) is added and an old partition (, 2, 1) is removed.
We have three possible partitions, (, , 1, 2), (, 1, 2), and (, 1, 2, ).
In this example, observe that for each of the intervals, we have at most 3 possible partitions as
16
there are at most two r
i
(T
0
) functions separating the partitions. We list the partitions corresponding
to large s
0
(s
0
> r
i
(T
0
)) rst, then the partitions corresponding to small s
0
(s
0
< r
i
(T
0
)). Note no
sorting is needed since the newly introduced partition when T
0
crosses
n
always have large s
0
.
In general, the partition enumeration can be carried out in O(N
2
log N), and the complexity is
determined by sorting the
n
s,

n
s, and the O(N
2
) intersection positions.
3.2 Solution algorithm with known (, c, and L
Recall the average cost of POT policy can be written as:
c(

T) = K
0
/T
0
+ h
0
s
0
+

1nN
(c
n
(T
0
, T
n
) + h

n
s
n
)
For given (, c, and L, we regroup the terms as
K
0
/T
0
+ h
0
s
0
+

nE
(c
n
(T
0
, T
n
) + h

n
s
n
) +

nG
(c
n
(T
0
, T
n
) + h

n
s
n
) +

nL
(c
n
(T
0
, T
n
) + h

n
s
n
)
where

nE
(c
n
(T
0
, T
n
) + h

n
s
n
) =

nE
(K
n
/T
0
+

n
2
h

n
T
0
+ h

L
n
+ T
0
),

nG
(c
n
(T
0
, T
n
) + h

n
s
n
) =

nG
(K
n
/T
n
+

n
2
h

n
T
n
+ h

L
n
+ T
n
), and

nL
(c
n
(T
0
, T
n
) + h

n
s
n
) =

nL
(K
n
/T
n
+

n
2
h
n
T
n
+ h

L
n
+ T
n
) +

nL

n
2
h
0
T
0
For n (, we can solve the optimal T
n
by minimizing K
n
/T
n
+

n
2
h

n
T
n
+h

L
n
+ T
n
, as
the rst order condition characterizes the optimal solution. The solution is exactly

n
we dened in
previous subsection. However, we can not use the same approach to solve T
n
with n L, because
17
s
0
is also a function of these T
n
s. The rst order condition with respect to T
n
is

K
n
T
2
n
+

n
2
h
n
+ h

n
1
2

L
n
+ T
n
h
0

2
0

2
n
2s
0
= 0 (1)
The rst order condition of T
0
gives us

nL

n
2
h
0
+

nE
(
K
n
T
2
0
+

n
2
h

n
+ h

n
1
2

L
n
+ T
0
)
K
0
T
2
0
+ h
0

2
0

nL

2
n
2s
0
= 0 (2)
Equations (1) and (2) may have multiple solutions. According to (1), we can write T
n
as a
function of the warehouse stock level s
0
by Corollary 2 in the Appendix. Furthermore, by Lemma 8
in the Appendix, each T
n
(s
0
) is a strictly convex function of s
0
. According to (2), we can write T
0
as a function of warehouse stock level s
0
by Corollary 3, and T
0
(s
0
) is a strictly concave function of
s
0
by Lemma 9 in the Appendix. Thus, if we know the optimal s
0
, we can determine the optimal
T
0
and T
n
s.
Note that if we know the values of T
0
and T
n
s, we can calculate s
0
by

L
0

2
n
+ (

nL

2
n
)T
0

nL

2
n
T
n
That is, the optimal solution (s
0
, T
0
, T
n
) must also satisfy the equation that denes s
0
. Dene
(s
0
) = s
2
0
=
2
0
(L
0

n

2
n
+(

nL

2
n
)T
0
(s
0
)

nL

2
n
T
n
(s
0
)). (s
0
) is a strictly concave function
of s
0
. At optimality, we must have (s
0
) = s
2
0
, or
_
(s
0
) = s
0
, according to the denition of s
0
.
Since (
_
(s
0
) s
0
)

s
0
=
1
4
2

(s
0
)(

(s
0
))
2

3/2
< 0 for

(s
0
) < 0,
_
(s
0
) s
0
is also a strictly
concave function, and
_
(s
0
) s
0
= 0 has at most two solutions. Via a binary search or a golden
search, we can nd these potential optimal s
0
values, and determine the corresponding T
n
s and T
0
in O(n). By comparing the two solutions, we can select the solution with lower total cost. Thus,
18
given (, c, and L, we can solve the minimization problem eciently. As there are only O(n
2
)
possible partition of (, c, and L, we can solve the relaxation problem in O(n
3
).
4 A Close-to-optimal Power-of-two Ordering Policy
In this section, we will use the optimal solution of the relaxation problem to build a close-to-optimal
power-of-two ordering policy.
We investigate two scenarios: i) the base interval length T
L
is given and xed, nd a power-of-
two ordering policy satisfying T
n
= 2
p
n
T
L
(p
n
Z); ii) we have the exibility to choose the base
interval length T
L
and the corresponding power-of-two ordering policy.
4.1 A Close-to-optimal Power-of-two Ordering Policy with Given T
L
For arbitrary target service level, we can construct an ordering policy satisfying T
n
= 2
p
n
T
L
(p
n

Z), which is no more than

5+1
2
1.618 times the cost of the relaxation problem by the
following procedure:
i) Let T

n
denote the optimal solution of the relaxation problem. Dene q
n
= log
2
(T

n
/T
L
)|,
that is, q
n
is the integer part of log
2
(T

n
/T
L
), and dene r
n
= log
2
(T

n
/T
L
) q
n
.
ii) For n 0 c, if r
n
< log
2
(), set T
n
= 2
q
n
T
L
, otherwise, set T
n
= 2
q
n
+1
T
L
.
iii) For n (, if r
n
0.5, set T
n
= 2
q
n
T
L
, otherwise, set T
n
= 2
q
n
+1
T
L
.
iv) For n L, we consider two cases: q
n
< q
0
and q
n
= q
0
.
q
n
< q
0
: we set T
n
= 2
q
n
T
L
if r
n
< 0.5 and set T
n
= 2
q
n
+1
T
L
otherwise.
q
n
= q
0
:
if r
0
< log
2
(), we set T
n
= 2
q
n
T
L
= T
0
;
19
if r
0
log
2
(), we set T
n
= 2
q
n
T
L
if r
n
1 log
2
()
and set T
n
= 2
q
n
+1
T
L
if r
n
> 1 log
2
().
Following the above procedure, if n (, then T
n
T
0
; if n c, then T
n
= T
0
; and if n L,
then T
n
T
0
. Now we show that each term in the objective function is positive, and the cost ratio
between the constructed ordering policy and the optimal solution for the relaxation problem is no
more than .
The objective function has the following terms:
K
n
T
n
for all n,

n
2
h
n
T
n
for n L,

n
2
h
0
T
0
for n
L,

n
2
h

n
T
n
for n (c, h

L
n
+ T
n
for n (cL, and h
0

0
_
L
0

n

2
n
+

nL
(T
0
T
n
)
2
n
.
To prove that each term in the objective function is positive, it suces to show h
n
> 0 for
n L. Now suppose this is not ture, and h
n
0, then both g
n
(T
n
) and h
0
s
0
(T
n
) are decreasing
functions on (0, T
0
); thus, n ( c. This is a contradiction. Therefore, h
n
> 0 for n L.
Next we show that the cost ratio between the constructed ordering policy and the optimal
solution for the relaxation problem is no more than . The proof of this result replies on the
following two lemmas.
Lemma 3 max
T
n
T

n
,
T

n
T
n
< for all n.
Lemma 4 max
nL

T
0
T
n
T

0
T

n

2
.
The detailed proofs are provided in the Appendix.
Theorem 4.1 The cost ratio between the constructed ordering policy and the optimal solution for
the relaxation problem is no more than
Proof: Lemma 3 says that
T

n
T
n
< for all n. Thus, the cost ratios between
K
n
T
n
and
K
n
T

n
are less than
. Similarly, the ratios between

n
2
h
n
T
n
and

n
2
h
n
T

n
, between

n
2
h
0
T
0
and

n
2
h
0
T

0
, and between
20

n
2
h

n
T
n
and

n
2
h

n
T

n
are all less than . Furthermore, the ratios between h

L
n
+ T
n
and
h

n
_
L
n
+ T

n
are less than

< for all n ( c L.
So far, we have compared all the terms except the warehouse safety stock cost. According
to Lemma 4, it is easy to show that the ratios between h
0

0
_
L
0

n

2
n
+

nL
(T
0
T
n
)
2
n
and
h
0

0
_
L
0

n

2
n
+

nL
(T

0
T

n
)
2
n
are no more than

2
= .
We have shown that each term in the objective function of the power-of-two ordering policy is
no more than 1.618 times of the corresponding term in the optimal solution for the relaxation
problem. Since the relaxation problem is a lower bound of optimal periodic ordering policy, we
claim that the cost of the resulting power-of-two ordering policy is no more than 1.618 times the
optimal cost.
4.2 A Close-to-optimal Power-of-two Ordering Policy with Unrestricted T
L
For arbitrary target service level, if the base period T
L
is not xed in advance, we can construct a
power-of-two ordering policy whose cost is no more than
3

2 1.260 times the cost of the relaxation

problem by the following procedure:
i) Let T

n
denote the optimal solution of the relaxation problem. Dene T
L
= T

0
/
3

2, q
n
=
log
2
(T

n
/T
L
)|, that is, q
n
is the integer part of log
2
(T

n
/T
L
), and dene r
n
= log
2
(T

n
/T
L
)q
n
;
ii) For n 0 c, set T
n
= T
L
.
iii) For n (,
K
n
/T

n


n
2
h

n
T

n
: if r
n
1/3, set T
n
= 2
q
n
T
L
, otherwise, T
n
= 2
q
n
+1
T
L
.
K
n
/T

n
<

n
2
h

n
T

n
: if r
n
2/3, set T
n
= 2
q
n
T
L
, otherwise, T
n
= 2
q
n
+1
T
L
.
21
iii) For n L,
K
n
/T

n


n
2
h
n
T

n
: if r
n
1/3, set T
n
= 2
q
n
T
L
, otherwise, T
n
= 2
q
n
+1
T
L
.
K
n
/T

n
<

n
2
h
n
T

n
: if r
n
2/3, set T
n
= 2
q
n
T
L
, otherwise, T
n
= 2
q
n
+1
T
L
.
Following the above procedure, if n (, then T
n
T
0
; if n c, then T
n
= T
0
; and if n L, then
T
n
T
0
. By regrouping the terms, the objective function has the following six combined terms:
(1)
K
0
T
0
,
K
n
T
n
for n c; (2)

n
2
h
0
T
0
for n L,

n
2
h

n
T
n
for n c; (3) (
K
n
T
n
+

n
2
h
n
T
n
) for n L, (4)
(
K
n
T
n
+

n
2
h

n
T
n
) for n (, (5) h

L
n
+ T
n
, and (6) h
0

0
_
L
0

n

2
n
+

nL
(T
0
T
n
)
2
n
. Now
for each combined term in the objective function, we show the cost ratio between the constructed
ordering policy and the optimal solution for the relaxation problem is no more than
3

2.
Term (1),
K
0
T
0
and
K
n
T
n
for n c:
T
0
= T

0
/
3

2 and T
n
= T

n
/
3

2 for n c. The cost ratios are

3

2.
Term (2),

n
2
h
0
T
0
for n L and

n
2
h

n
T
n
for n c:
T
0
= T

0
/
3

2 and T
n
= T

n
/
3

2 for n c. The cost ratios are

1
3

2
<
3

2.
Term (3), (
K
n
T
n
+

n
2
h
n
T
n
) for n L:
We have two scenarios:
K
n
/T

n


n
2
h
n
T

n
: if r
n
1/3, 1
T

n
T
n

2, the cost ratio is smaller than

3

2
if r
n
> 1/3,
T
n
T

n
= 2
1r
n
,
K
n
/T
n
+

n
2
h
n
T
n
K
n
/T

n
+

n
2
h
n
T

2
r
n
1
+2
1r
n
1+1
<
2
2/3
+2
2/3
2
<
3

2
K
n
/T

n
<

n
2
h
n
T

n
: if r
n
2/3,
T

n
T
n
= 2
r
n
,
K
n
/T
n
+

n
2
h
n
T
n
K
n
/T

n
+

n
2
h
n
T

2
r
n
+2
r
n
1+1
<
2
2/3
+2
2/3
2
<
3

2
if r
n
> 2/3, 1
T
n
T

2, the cost ratio is smaller than

3

2
Term (4), (
K
n
T
n
+

n
2
h

n
T
n
) for n (:
We have two scenarios:
22
K
n
/T

n


n
2
h

n
T

n
: if r
n
1/3, 1
T

n
T
n

2, the cost ratio is smaller than

3

2
if r
n
> 1/3,
T
n
T

n
= 2
1r
n
,
K
n
/T
n
+

n
2
h

n
T
n
K
n
/T

n
+

n
2
h

n
T

2
r
n
1
+2
1r
n
1+1
<
2
2/3
+2
2/3
2
<
3

2
K
n
/T

n
<

n
2
h

n
T

n
: if r
n
2/3,
T

n
T
n
= 2
r
n
,
K
n
/T
n
+

n
2
h

n
T
n
K
n
/T

n
+

n
2
h

n
T

2
r
n
+2
r
n
1+1
<
2
2/3
+2
2/3
2
<
3

2
if r
n
> 2/3, 1
T
n
T

2, the cost ratio is smaller than

3

2
Term (5), h

L
n
+ T
n
:
T
n
T

n
2
2/3
,
L
n
+T
n
L
n
+T

n
2
2/3
, and the cost ratios are no more than

2
2/3
=
3

2.
Term (6), h
0

0
_
L
0

n

2
n
+

nL
(T
0
T
n
)
2
n
:
We rst show
T
0
T
n
T

0
T

n
2
2/3
. It suces to prove the case where T
0
> T
n
. Note T
0
/T
n
is an integer power of two; thus, T
n
T
0
/2. Also note that T

0
=
3

2T
0
, T

n
2
2/3
T
n
,
thus
T
0
T
n
T

0
T

21
2
3

22
2/3
< 2
2/3
. So the ratio of the warehouse safety stock cost is less than

2
2/3
=
3

2.
We have shown that each combined term in the objective function of the power-of-two ordering
policy is no more than
3

2 1.260 times of the corresponding term in the optimal solution for the
relaxation problem. Since the relaxation problem is a lower bound of optimal periodic ordering
policy, the cost of the resulting power-of-two ordering policy guarantees to be no more than 1.260
times the optimal cost.
5 Extension to Type 2 Service Level
We have so far focus on Type 1 service levels. The (T
0
, s
0
) partition and enumeration approach
can be used for Type 2 service level to generate the optimal solution for the relaxation problem.
After that, we can have a close-to-optimal solution by round up all the order intervals.
The model formulation for the Type 1 service level in Section 2 can be adapted for the Type
2 service level counterparts. Under Type 2 service level, the safety stock level s
n
(T
n
) at retailer n
23
can be approximated by the solution of equation
n
= 1((
s
n
(T
n
)

) s
n
(T
n
)(
s
n
(T
n
)

))/(
n
T
n
)
(Chase, Aquilano, and Jacobs 1998, Nahmias 2001), where
n
is the target service level for retailer
n, =
n

L
n
+ T
n
is the lead-time variance, and are the density and cumulative probability
function for the standard Normal distribution, respectively. For warehouses safety stock, s
0
(T
0
)
can be approximated by the solution of equation
0
= 1 ((
s
0
(T
0
)

) s
0
(T
0
)(
s
0
(T
0
)

))/(
0
T
0
),
where
0
is the target service level for the warehouse, and =
_
L
0

n

2
n
+

n
(0 (T
0
T
n
))
2
n
is the lead-time variance.
Despite the lack of analytic formula for the s
0
(T
0
) and s
n
(T
n
)s, we can dene f
n
(T
n
) = K
n
/T
n
+

n
2
h

n
T
n
+ h

n
s
n
(T
n
) and g
n
(T
n
) = K
n
/T
n
+

n
2
h
n
T
n
+ h

n
s
n
(T
n
) as before. It is straightforward to
show that both f
n
(T
n
) and g
n
(T
n
) are convex functions of T
n
by calculus. (Rao 2003 shows that
f
n
(T
n
) and g
n
(T
n
) are convex without approximating the safety stock level.) Therefore, we can
dene

n
and
n
that minimize f
n
(T
n
) and g
n
(T
n
), respectively. Note that

n
and
n
can be found
via a search algorithm because f
n
(T
n
) and g
n
(T
n
) are convex.
From the previous development, we see that

n
<
n
. Furthermore, n ( if and only if T
0
<

n
,
and n c if

n
T
0

n
. Finally, when T
0
>
n
, we need determine whether n c or L based on
both T
0
and s
0
, that is, we need to dene a boundary function r
n
(T
0
) so that n c if s
0
r
n
(T
0
),
and n L if s
0
> r
n
(T
0
). As before, we draw a line through (T
0
, g
n
(T
0
)) such that this line is
tangent with g
n
(T
n
) and lies below g
n
(T
n
) on (0, T
0
), while the optimal T
n
minimizes g
n
(T
n
) g
n
(T
n
)
on [0, T
0
], where g
n
(T
n
) h
0
s
0
(T
n
) + h
0
s
0
(T
0
) + g
n
(T
0
). It turns out that g
n
(T
n
) is a concave
function on [0, T
0
]; therefore, once again, the slope of the tangent line species a bound function
r
n
(T
0
), which partitions the (T
0
, s
0
) plane. n L if (T
0
, s
0
) is above the bound function r
n
(T
0
);
and n c if (T
0
, s
0
) is below or on the bound function r
n
(T
0
). After that, we can enumerate all
the possible partitions ((, c, L) and nd the optimal solution for the relaxation problem according
to the rst order condition.
24
6 Conclusions
In this paper we have proposed a power-of-two policy to manage inventory in one-warehouse multi-
retailer models with stochastic demand. For any given target service level, we develop a polynomial
algorithm that runs in time O(N
3
) to nd a power-of-two ordering policy whose cost is guaranteed
to be no more than
3

2 1.26 times the optimal cost. To the best of our knowledge, this is the rst
paper that gives a worst-case bound for a heuristic in multi-echelon stochastic inventory systems.
In the future we plan to investigate the possibility of generalizing the results in this paper to
more complex stochastic inventory systems. Interesting directions include the scheduled ordering
policies, adding restrictions/constraints to the ordering policy, and extension to multi-item inven-
tory systems.
A Appendix
Lemma 5 For f(x) =
a
x
+bx+

i
c
i

d
i
+ x with a > 0, c
i
0,

i
c
i
> 0, and d
i
0, there exists
a critical value x

, such that f(x) is a convex function of x on (0, x

), and a concave function of x

on (x

, ).
Proof: f

(x) =
2a
x
3

i
c
i
4
1
(d
i
+x)
3/2
. f

(x) approaches positive innity as x approaches 0 from above,

and f

(x) approaches zero from below as x approaches positive innity. Thus, there exists x

such
that f

(x

) = 0. We prove that there is a unique x

such that f

(x

) = 0 by contradiction. Assume
there exist x
1
and x
2
= x
1
( > 1) such that f

(x
1
) = f

(x
2
) = 0. It follows that
d
i
+x
2
d
i
+x
1
and
c
i
4
1
(d
i
+x
1
)
3/2

3/2 c
i
4
1
(d
i
+x
2
)
3/2
. Thus, f

(x
1
)
3 2a
x
3
2

3/2

i
c
i
4
1
(d
i
+x
2
)
3/2
= (
3

3/2
)
2a
x
3
2
> 0 for
> 1, which is a contradiction. That is, there exists unique x

such that f

(x

) = 0. Furthermore,
f

(x) > 0 on (0, x

), and f

(x) < 0 on (x

, ), i.e., f(x) is a convex function of x on (0, x

), and
f(x) is a concave function of x on (x

, ).
25
Corollary 1 For f(x) =
a
x
+ bx + c

d + x with a > 0, b > 0, c 0 and d 0, the rst order

condition characterizes the optimal solution of the minimization problem.
Proof: If c = 0, f(x) is convex, and the rst order condition characterizes the optimal solution of
the minimization problem. Otherwise, by Lemma 5, we can have at most two values satisfying the
rst order condition: one between 0 and the critical x

, which corresponds to a local minimum, and

one greater than x

, which corresponds to a local maximum. Also observe that f(x) approaches

positive innity as x approaches positive innity, so there is no local maximum on (x

, ). Thus,
the rst order condition has a unique solution, which corresponds to a local minimum. The unique
solution is also global minimum.
Lemma 6 For f(x) =
a
x
+ bx + c

d + x with a > 0, c > 0 and d 0, let x

denote the unique

solution for f

(x) = 0 as in Lemma 5. Consider g(x) f(x) + e

h x on [0, h] with e > 0 and

h > 0. If the optimal solution of minimizing g(x) is an internal point, then it lies between 0 and
x

.
Proof: It suces to prove the case with h > x

. Since e

h x is a concave function, by Lemma 5,

g(x) is concave on (x

, h), and no internal point of (x

, h) would correspond to an optimal solution

of minimizing g(x). Thus, the internal optimal solution must lie between 0 and x

.
Corollary 2 Let f(x) =
a
x
+ bx + c

d + x with a > 0, c 0, and d 0. Assume the optimal

solution of minimizing g(x) f(x) + e

h x with e > 0 and h > 0 is an internal point, then the

optimal x is a function of the derivative of e

h x.
Proof: The optimal solution satises the rst order condition g

(x) = 0. Thus, f

(x) = (e

h x)

is satised at the optimal solution. The result is obvious when c = 0. If c > 0, by Lemma 6, the
solution lies between 0 and x

. f

(x) is positive on (0, x

), thus f

(x) is monotone, and if we know

26
the derivative of e

h x, the optimal solution x can be uniquely determined. The optimal value x

is a monotone decreasing function of this derivative (e

h x)

, since f

(x) is monotone increasing.

Lemma 7 For f(x) =
a
x
+ bx +

i
c
i

d
i
+ x with a > 0, c
i
0,

c
i
> 0, and d
i
0, let x

denote the unique solution for f

(x) = 0 as in Lemma 5. Consider g(x) f(x) + e

h + x with
e > 0. If the optimal solution of minimizing g is an internal point, then it lies between 0 and x

.
Proof: Due to the concavity, proof is similar to the proof of Lemma 6.
Corollary 3 Let f(x) =
a
x
+bx+

i
c
i

d
i
+ x with a > 0, c
i
0, and d
i
0. Assume the optimal
solution of minimizing g(x) f(x) + e

h + x with e > 0 is an internal point, then the optimal x

is a function of the derivative of e

h + x.
Proof: Proof is similar to the proof of Corollary 2.
Lemma 8 For f(x) =
a
x
+bx+c

d + x with a > 0, c 0, and d 0, let x

be the unique solution

for f

(x) = 0 (as in Lemma 5) if c > 0 and let x

be if c = 0. Let T(s) denote the unique

solution in (0, x

) such that f

(T(s))
C
s
= 0, where C > 0 is a constant, then T(s) is a strictly
convex function of s.
Proof: f

(T)[
T=T(s)
=
C
s
> 0, and f

(T) > 0 as T (0, x

).
f

(T) > 0
2a
T
3
>
1
4
c
(d+T)
3/2

6a
T
4
>
3
4
1
T
c
(d+T)
3/2

3
4
c
(d+T)
5/2

3
8
c
(d+T)
5/2
f

(T) =
6a
T
4
+
3
8
c
(d+T)
5/2
< 0
Thus,
dT
ds
=
1
C
(f

(T))
2
f

(T)
< 0. Furthermore, since f

(T) > 0 and f

(T) < 0,
d
2
T
ds
2
=
1
C
2f

(T)(f

(T))
2
(f

(T))
2
f

(T)
(f

(T))
2
dT
ds
> 0
That is, T(s) is a strictly convex function of s.
27
Lemma 9 For f(x) =
a
x
+ bx +

i
c
i

d
i
+ x with a > 0, b > 0, c
i
0 and d
i
0, let x

be the
unique solution for f

(x) = 0 (as in Lemma 5) if

i
c
i
> 0 and let x

be if

i
c
i
= 0. Let T(s)
be the unique solution in (0, x

) such that f

(T(s)) +
C
s
= 0, where C > 0 is a constant, then T(s)
is a strictly concave function of s.
Proof: f

(T)[
T=T(s)
=
C
s
< 0, and f

(T) > 0 as T (0, x

). By the same arguments in Lemma 8,

f

(T) > 0 f

(T) < 0.
dT
ds
=
1
C
(f

(T))
2
f

(T)
> 0,
d
2
T
ds
2
=
1
C
2f

(T)(f

(T))
2
(f

(T))
2
f

(T)
(f

(T))
2
dT
ds
. Since
dT
ds
> 0
and f

(T) < 0, in order to show

d
2
T
ds
2
< 0, it suces to show that 2(f

(T))
2
f

(T)f

(T) > 0.
It is easy to show that
2(f

(T))
2
= 2(
2a
T
3

1
4

i
c
i
(d
i
+ T)
3/2
)
2
=
8a
2
T
6

2a
T
3

i
c
i
(d
i
+ T)
3/2
+
1
8
(

i
c
i
(d
i
+ T)
3/2
)
2
Furthermore,
0 < f

(T) =
6a
T
4

3
8

i
c
i
(d
i
+ T)
5/2

6a
T
4
and
0 < f

(T) =
a
T
2
b
1
2

i
c
i
(d
i
+ T)
1/2
<
a
T
2

1
2

i
c
i
(d
i
+ T)
1/2
for b > 0. Thus
f

(T)f

(T) = (f

(T))(f

(T)) < (
a
T
2

1
2

i
c
i
(d
i
+ T)
1/2
)
6a
T
4
=
6a
2
T
6

3a
T
4

i
c
i
(d
i
+ T)
1/2
and
2(f

(T))
2
f

(T)f

(T) >
2a
2
T
6

2a
T
3

i
c
i
(d
i
+ T)
3/2
+
1
8
(

i
c
i
(d
i
+ T)
3/2
)
2
+
3a
T
4

i
c
i
(d
i
+ T)
1/2
We know
3a
T
4
>
3
8

i
c
i
(d
i
+T)
5/2
for f

(T) =
2a
T
3

1
4

i
c
i
(d
i
+T)
3/2
> 0, and (

i
c
i
(d
i
+T)
5/2
)(

i
c
i
(d
i
+T)
1/2
)
28
(

i
c
i
(d
i
+T)
3/2
)
2
by Cauchy inequality, thus we have
2(f

(T))
2
f

(T)f

(T) >
2a
2
T
6

2a
T
3

i
c
i
(d
i
+ T)
3/2
+
1
8
(

i
c
i
(d
i
+ T)
3/2
)
2
+
3
8
(

i
c
i
(d
i
+ T)
3/2
)
2
= 2(
a
T
3

1
2

i
c
i
(d
i
+T)
3/2
)
2
0
We conclude that
d
2
T
ds
2
< 0, and T(s) is a strictly concave function of s.
Lemma 10 If C
n
> C

n
, the optimal T

n
that minimizes g
n
(T
n
) + h
0

0
_
C
n
+ (T
0
T
n
)
2
n
on
(0, T
0
) is no more than

T
n
.
Proof: Recall that

T
n
is the rightmost tangent point on (0, T
0
). Therefore,

T
n
minimizes g
n
(T
n
)
g
n
(T
n
) on (0, T
0
). By denition, g
n
(T
n
) = h
0
s
0
(T
n
) +h
0
s
0
(T
0
) +g
n
(T
0
),

T
n
minimizes (g
n
(T
n
) +
h
0

0
_
C

n
+ (T
0
T
n
)
2
n
) on (0, T
0
). For

T
n
< T
n
< T
0
, we have
g
n
(

T
n
) + h
0

0
_
C

n
+ (T
0

T
n
)
2
n
< g
n
(T
n
) + h
0

0
_
C

n
+ (T
0
T
n
)
2
n
That is, for

T
n
< T
n
< T
0
,
g
n
(T
n
) g
n
(

T
n
) > h
0

0
_
C

n
+ (T
0

T
n
)
2
n
h
0

0
_
C

n
+ (T
0
T
n
)
2
n
Also, for

T
n
< T
n
< T
0
,
h
0

0
_
C

n
+ (T
0

T
n
)
2
n
h
0

0
_
C

n
+ (T
0
T
n
)
2
n
> h
0

0
_
C
n
+ (T
0

T
n
)
2
n
h
0

0
_
C
n
+ (T
0
T
n
)
2
n
29
because C
n
> C

n
. Thus, for

T
n
< T
n
< T
0
,
g
n
(T
n
) g
n
(

T
n
) > h
0

0
_
C
n
+ (T
0

T
n
)
2
n
h
0

0
_
C
n
+ (T
0
T
n
)
2
n
or
g
n
(

T
n
) + h
0

0
_
C
n
+ (T
0

T
n
)
2
n
< g
n
(T
n
) + h
0

0
_
C
n
+ (T
0
T
n
)
2
n
That is, the optimal solution T

n
is no more than

T
n
.
In the next, we show that the curve (r
i
(T
0
)s separate the plane into O(n
2
) areas. Let us dene
a
n
K
n
, b
n


n
2
h
n
, c
n
h

n
, d
n
L
n
, and e
n
h
0

2
0

2
n
/2, then g
n
(x) = a
n
/x + b
n
x +
c
n

d
n
+ x. Let (

T, g
n
(

T)) be the tangent point. According to the denition of r

n
(T
0
), we have
g

n
(

T) =
e
n
r
n
(T
0
)
(3)
g
n
(T
0
) g
n
(

T) = (T
0


T)g

n
(

T) (4)
We rst show the curve r
n
(T
0
) satises some polynomial function.
Lemma 11 There exists a constant C
1
, and a nonzero polynomial P with degree less than C
1
such
that P(r
n
(T
0
), T
0
, a
n
, b
n
, c
n
, d
n
, e
n
) = 0.
Proof: By (3) and (4),
g
n
(T
0
) T
0
e
n
r
n
(T
0
)
= g
n
(

T)

Tg

n
(

T)
=
a
n

T
+ b
n

T + c
n
_
d
n
+

T +
a
n

T
b
n

T
c
n

T
2
_
d
n
+

T
=
2a
n

T
+
c
n
2
2d
n
+

T
_
d
n
+

T
30
Denote x =
_
d
n
+

T, then

T = x
2
d
n
, and the above formulation can be rewritten as
c
n
2
(x
2
+ d
n
)(x
2
d
n
) (g
n
(T
0
) T
0
e
n
r
n
(T
0
)
)(x
2
d
n
)x + 2a
n
x = 0
That is, x satises a quartic equation. By the quartic formula, x equals one of the four roots x
1
,
x
2
, x
3
, and x
4
, all of which can be expressed by mathematical operations of addition, subtraction,
multiplication, division, square root, and cubic root of T
0
, r
n
(T
0
), a
n
, b
n
, c
n
, and d
n
. Let

T
i
= x
2
i
d
n
for i = 1, , 4. Recall that x =
_
d
n
+

T by denition and g

n
(

T) =
e
n
r
n
(T
0
)
in (3); therefore

4
i=1
(g

n
(

T
i
)
e
n
r
n
(T
0
)
) = 0 because x equals one of the four roots x
1
, x
2
, x
3
, and x
4
. Thus, there
is a nonzero function Q satisfying Q(r
n
(T
0
), T
0
, a
n
, b
n
, c
n
, d
n
, e
n
) = 0 that uses only mathematical
operations of addition, subtraction, multiplication, division, square root, and cubic root.
By multiply all the denominators, we can transform Q to a nonzero function R, which satises
R(r
n
(T
0
), T
0
, a
n
, b
n
, c
n
, d
n
, e
n
) = 0 that uses only mathematical operations of addition, subtraction,
multiplication, square root, and cubic root.
Next we show how to transform R to a polynomial P. The key is to transform square root and
cubic root terms to polynomial terms.
We consider the square root operation rst.
Assume we have square root of p(r
n
(T
0
), T
0
, a
n
, b
n
, c
n
, d
n
, e
n
) in expression R. We can write
equation R = 0 as R
1
= R
2

p, where R
1
, R
2
are functions of r
n
(T
0
), T
0
, a
n
, b
n
, c
n
, d
n
, e
n
. By
squaring the equation on both sides, we no longer take square root of term p. As the square
operation does not introduce new square term into the equation, we can get rid of all the
square root operations.
Now we consider the cubic root operation.
Assume we have cubic root of q(r
n
(T
0
), T
0
, a
n
, b
n
, c
n
, d
n
, e
n
) in expression R. We can write
31
equation R = 0 as R
1
+ R
2
3

q + R
3
3
_
q
2
= 0, where R
1
, R
2
, R
3
are functions of r
n
(T
0
), T
0
,
a
n
,b
n
,c
n
,d
n
,e
n
. If R
3
= 0, then we can move R
1
to the right-hand-side of the equation, and
cubing both sides to remove the cubic root operations. If R
3
,= 0, multiply both sides by R
3
,
and we have R
2
2
R
1
R
3
= R
2
3
3
_
q
2
+ R
3
R
2
3

q + R
2
2
. Multiplying both sides by R
3
3

q R
2
,
we get (R
2
2
R
1
R
3
)R
3
3

q = R
2
3
q R
1
R
2
R
3
. By cubing both sides, we no longer take cubic
root of term q.
Thus, we can generate a nonzero polynomial P such that P(r
n
(T
0
), T
0
, a
n
, b
n
, c
n
, d
n
, e
n
) = 0.
The constructive procedure shows that there exists a constant C
1
such that the degree of P is less
than C
1
.
Proof of Theorem 1.
By the proof of Lemma 11, we know that for any retailer n, we can get a nonzero polynomial
P
n
such that P
n
(r
n
(T
0
), T
0
) = 0 and the degree of P
n
is no more than some constant C
1
.
To prove Theorem 1, it suces to show that P
i
(x, y) = 0 and P
j
(x, y) = 0 will not intercept
with each other for more than some constant times when both P
i
and P
j
are nonzero polynomials
with degree no more than C
1
. Without loss of generality, we can assume P
i
(x, y) and (P
i
(x, y))

y
,
the derivative of P
i
(x, y) with respect to y, have no factor in common, and P
j
(x, y) and (P
j
(x, y))

y
,
the derivative of P
j
(x, y) with respect to y, have no factor in common, because factoring out the
common factors does not change the shape of P
i
(x, y) = 0 or P
j
(x, y) = 0.
By treating P
i
(x, y) and P
j
(x, y) as polynomials of variable y, we can factor the polynomials
P
i
and P
j
by iterated division. Thus, P
i
(x, y) = u(x, y)p(x, y) and P
j
(x, y) = v(x, y)p(x, y), where
p(x, y) is a common factor of P
i
(x, y) and P
j
(x, y) while u(x, y) and v(x, y) have no common
polynomial factor.
P
i
(x, y) = 0 and P
j
(x, y) = 0 will intercept with each other if and only if i) u(x, y) and v(x, y)
32
intercept with each other; ii) u(x, y) or v(x, y) intercept with p(x, y); iii) p(x, y) intercepts itself.
i) u(x, y) and v(x, y) may intercept with each other at (x, y) only if u(x, y) = v(x, y) = 0. Since
u(x, y) and v(x, y) have no common factor involving y, by iterated division, (x, y) must satisfy a
series of polynomial equations, the last of which has degree of y equal to zero, that is x must be a
root of a polynomial equation. Since the degrees of u(x, y), v(x, y) are no more than C
1
, the degree
of the resulting polynomial is bounded by some constant number C
2
. Thus, we have at most C
2
possible values of x, each of which corresponds to at most C
1
possible values of y, and there are at
most C
1
C
2
interceptions between u and v.
ii) Since P
i
(x, y) and (P
i
(x, y))

y
have no factor in common, and P
j
(x, y) and (P
j
(x, y))

y
have
no factor in common, and there are at most C
1
C
2
interceptions between u(x, y) and p(x, y) by the
same argument. Similarly, there are at most C
1
C
2
interceptions between v(x, y) and p(x, y).
iii) For the case where p(x, y) = 0 intercepts itself at (x, y), the interception can only happen
if y is a multiple root for the given x. Thus, p(x, y) may intercept with itself only if p(x, y) =
(p(x, y))

y
= 0. Note that p(x, y) and (p(x, y))

y
= 0 have no factor in common, because P
i
(x, y) and
(P
i
(x, y))

y
have no factor in common. Since the degrees of p(x, y), (p(x, y))

y
are no more than C
1
,
and p(x, y), (p(x, y))

y
have no factor in common, this can only happen when x satises a polynomial
with degree bounded by the constant number C
2
. Thus, we have at most C
2
possibility of x, each
of which corresponds to at most C
1
possibility of y, and there are at most C
1
C
2
interceptions of
p(x, y).
Thus, by setting k equal to 4C
1
C
2
, we can guarantee that the dierence of the curves (r
i
(T
0
)
r
j
(T
0
)) does not change sign for more than k times for any two retailers i and j.
Proof of Lemma 3: max
T
n
T

n
,
T

n
T
n
< for all n.
We discuss the following three cases:
33
For n 0 c:
T

n
T
n
= 2
r
n
if r
n
< log
2
(), and
T

n
T
n
= 2
r
n
1
if r
n
log
2
(). Since 0 r
n
< 1,

2

T

n
T
n
< , max
T
n
T

n
,
T

n
T
n
< for n 0 c.
For n (,
T

n
T
n
= 2
r
n
if r
n
< 0.5, and
T

n
T
n
= 2
r
n
1
if r
n
0.5, max
T
n
T

n
,
T

n
T
n

2 < for
n (.
For n L, when q
n
< q
0
, same as the case n (, max
T
n
T

n
,
T

n
T
n
< ; when q
n
= q
0
, if
r
0
< log
2
(), we have
T

n
T
n
= 2
r
n
. Since n L and T

n
< T

0
, r
n
< r
0
< log
2
(), and
1
T

n
T
n
< ; if r
0
log
2
(), we have
T

n
T
n
= 2
r
n
if r
n
1 log
2
() and
T

n
T
n
= 2
r
n
1
if
r
n
> 1 log
2
(); and
1

<
T

n
T
n

. Thus, max
T
n
T

n
,
T

n
T
n

2 < for n L.
Thus, max
T
n
T

n
,
T

n
T
n
< for all n.
Proof of Lemma 4: max
nL

T
0
T
n
T

0
T

n

2
.
Note T
n
T
0
for n L. If T
n
= T
0
, then
T
0
T
n
T

0
T

n
<
2
. Now we focus on the case with T
n
< T
0
.
Since T
0
/T
n
is an integer power of two, T
n
T
0
/2. Note T
0
/T

0
<
2

. According to r
0
, we have the
following two scenarios:
r
0
< log
2
(): in this case, T
0
= T

0
/2
r
n
T

0
, while T

n
< T
n
.
T
0
T
n
T

0
T

n
<
T
0
T
n
T
0
T
n
. Since
T
n
T
0
/2,
T
0
T
n
T
0
T
n

21
2
=
2
. Thus,
T
0
T
n
T

0
T

n

2
.
r
0
log
2
(): in this case, T
0
= 2
1r
n
T

0
, we further divide this case into two subcases:
T
n
= T
0
/2: If T

n
T
n
, then q
n
= q
0
, and r
n
1 log
2
() because T
n
= 2
q
n
T
L
. Thus,
T

n
is no more than 2T
n
/. Since T
n
= T
0
/2 and T

0
/T
0


2
,
T
0
T
n
T

0
T

T
0
T
n

2
T
0
2T
n
/
=
21
2/
=
2
. Thus,
T
0
T
n
T

0
T

n

2
.
T
n
T
0
/4: By T

n
/T
n
< and T

0
/T
0


2
,
T
0
T
n
T

0
T

n
<
T
0
T
n

2
T
0
T
n
. Since T
n
T
0
/4,
T
0
T
n

2
T
0
T
n
=
41
2
<
2
. Thus,
T
0
T
n
T

0
T

n

2
.
34
References
Axsater, S. 1993. Exact and approximate evaluation of batch-ordering polices for two-level inven-
tory systems. Oper. Res. 41 (4) 777-785.
. 1996. Using the deterministic EOQ formula in stochastic inventory control. Management
Sci. 42 (6) 830-834.
. 2000. Exact analysis of continuous review (R,Q) policies in two-echelon inventory systems
with compound poisson demand. Oper. Res. 48 (5) 686696.
. 2003. Supply chain operations: serial and distribution inventory systems. In S. C. Graves
and A. G. de Kok (eds.) Supply Chain Management: Design, Coordination and Operation,
Handbooks of Operations Research and Management Science Vol. 11. Elsevier Science Pub-
lishers. North Holland. Amsterdam. 525559.
Cachon, G. P. 1999. Managing supply chain demand variability with scheduled ordering policies.
Management Sci. 45 (6) 843856.
. 2001. Exact evaluation of batch-ordering inventory policies in two-echelon supply chains
with periodic review. Oper. Res. 49 (1) 7998.
Caggiano, K. E., P. L. Jackson, J. A. Muckstadt, J. A. Rappold. 2002. Simple algorithms for
pooling inventory to satisfy customer service requirements at minimum cost. Working paper.
Cornell University.
Chan, L. A., A. Muriel, Z. M. Shen, D. Simchi-Levi, C.-P. Teo. 2002. Eective zero-inventory-
ordering policies for the single-warehouse multiretailer problem with piecewise linear cost struc-
tures. Management Sci. 48 (11) 14461460.
35
Chase, R. B., N. J. Aquilano, F. R. Jacobs. 1998. Production and Operations Management:
Manufacturing and Services. Eighth Ed. McGraw Hill Irwin, New York.
Chen, F. 2000. Eectiveness of (R,Q) policies in one-warehouse multi-retailer systems with deter-
ministic demand and backlogging, Naval Res. Logist. 47 (5) 422439.
. R. Samroengraja. 2000. A staggered ordering policy for one-warehouse multi-retailer
systems. Oper. Res. 48 (2) 281293.
. Y.-S. Zheng. 1997. One-warehouse multi-retailer systems with centralized stock informa-
tion. Oper. Res. 45 (2) 275-287.
Cheung, K., W. Hausman. 2000. An exact performance evaluation for the supplier in a two-echelon
inventory system. Oper. Res. 48 (4) 646-653.
Devermeyer, B., L. Schwarz. 1981. A model for the analysis of system service level in warehouse-
retailer distribution systems: The identical retailer case. TIMS Studies in the Management
Sciences. Vol. 16. 163-193.
Federgruen, A., P. Zipkin. 1984a. Computational issues in an innite-horizon, multi-echelon in-
ventory model, Oper. Res. 32 (4) 818-836.
., . 1984b. Allocation policies and cost approximations for multi-location inventory
systems. Naval Res. Logist. Quart. 31 97129.
Lee, H., K. Moinzadeh. 1987a. Operating characteristics of a two-echelon inventory system for
repairable and consumable items under batch ordering and shipment policy. Naval Res. Logist.
Quart. 34 365-380.
., . 1987b. Two-parameter approximations for multi-echelon repairable inventory mod-
els with batch ordering policy. IIE Trans. 19 (2) 140-149.
., P. Padmanabhan, S. Whang. 1997. Information distortion in a supply chain: the bullwhip
eect. Management Sci. 43 (4) 546-558.
36
Moinzadeh, K., H. Lee. 1986. Batch size and stocking levels in multi-echelon repairable systems.
Management Sci. 32 (12) 1567-1581.
Muckstadt, J. A., R. O. Roundy. 1993. Analysis of multistage production systems. in S. C. Graves,
A. H. G. Rinnooy Kan, and P. H. Zipkin (Eds.) Logistics of Production and Inventory, Hand-
books in Operations Research and Management Science. Vol. 4. Elsevier Science Publishers.
North Holland. Amsterdam. 59132.
Nahmias, S. 2001. Production and Operations Analysis. Fourth Ed. McGraw Hill Irwin, New York.
Rao, U. S. 2003. Properties of the periodic review (R, T) inventory control policy for stationary,
stochastic demand. Manufacturing & Service Operations Management 5 (1) 3753.
Roundy, R. 1985. 98%-eective integer-ratio lot-sizing for one-warehouse, multi-retailer systems.
Management Sci. 31 (11) 14161430.
Schwarz, L. B., B. L. Deuermeyer, R. D. Badinelli. 1985. Fill-rate optimization in a one-warehouse
N-identical retailer distribution system. Management Sci. 31 (4) 488-498.
Svoronos, A., P. Zipkin. 1988. Estimating the performance of multi-level inventory systems. Oper.
Res. 36 (1) 57-72.
Shen Z. J., D. Simchi-Levi, C. P. Teo. 2002. Approximation algorithms for the single-warehouse
multi-retailer problem with piecewise linear cost structures. url: citeseer.nj.nec.com/439759.html.
Zheng, Y. S. 1992. On properties of stochastic inventory systems. Management Sci. 38 (1) 87103.
37