On the use of break quantities in multi{echelon dis-

tribution systems
Rommert Dekker, Hans Frenk, Marcel Kleijn, Nanda Piersma
Erasmus University Rotterdam, The Netherlands

Ton de Kok
Eindhoven University of Technology, The Netherlands

October 1995

Abstract
In multi{echelon distribution systems it is usually assumed that demand is only
satised from the lowest echelon. In this paper we will consider the case where de-
mand can be satised from any level in the system. However, then the problem arises
of how to allocate orders from customers to the dierent locations. A possible way of
dealing with this problem consists of using a so{called break quantity rule. This easy
implementable rule is to deliver every order with a size exceeding the break quantity
from a higher echelon. The use of the break quantity rule now results in a reduction
of the demand variability at the retailer and hence less safety stocks need to be held.
The concept is studied for a two{echelon distribution system, consisting of one ware-
house and one retailer, where the inventory at the retailer is controlled by an order
up to level policy, and where at the warehouse there is enough inventory to satisfy
all orders from the retailer and the customers. For this system an approximation
for the long run average costs as a function of the break quantity is derived, and an
algorithm is presented to determine the cost{optimal break quantity. Computational
results indicate that the break quantity rule can lead to signicant cost reductions.

Keywords: Break quantity rule, inventory, multi{echelon distribution systems.

 Dekker et al./ On the use of break quantities in multi{echelon distribution systems 1

1 Introduction
In multi{echelon distribution systems, management is primarily concerned with meeting
customer demand in a timely way at minimum cost. In practice, customer demand is often
not known in advance, and can exhibit large uctuations over time. Such uctuations
typically occur when the demand process consists of many customers ordering small batches
and few customers ordering large batches. This may lead to high safety stock levels with
large associated costs.
A way to reduce the variability of the demand is by handling large orders in a special way.
We introduce a so{called break quantity to distinguish between small and large orders. In
many cases customers ordering large batches do not have an immediate need for delivery,
since their orders are used as a replenishment for an inventory system, or their order sizes
are larger than necessary because of quantity discounts. Two management procedures, in
which large orders are handled separately, are considered in De Kok [13] and Nass, Dekker
and Van Sonderen{Huisman [15]. The rst procedure is to split the large orders into
smaller batches, which are subsequently delivered. This way the variability of the order
sizes is reduced, and part of future demand will be known beforehand, which will enable a
reduction of the safety stock levels. The second procedure, which we focus on in this paper,
is to supply the large orders, at a later stage, from a higher level in the distribution system.
In this paper it will be shown that this tactical rule leads to a reduction of the safety stocks
at the lowest level of the distribution system, thus decreasing the long run average costs.
The main reasons for this are the fact that the break quantity lters out the peaks in
the demand, and the fact that the total demand for the lowest level decreases. Another
advantage is the reduction of handling and transportation costs, due to the possibility
of delivering large orders directly to the customers, thereby skipping one or more nodes
in the distribution chain. The eect of direct shipping strategies on the transportation
costs has been studied extensively in literature (e.g. Burns et al. [4], Gallego and Simchi{
Levi [9], Sussams [19]). However, the eect on the inventory costs has only been analyzed
for deterministic demand. A last advantage of the break quantity concept is its simplicity:
it can easily be implemented in practice, especially when only statistical information on
future demand is known, and upon order entry immediate feedback on the supply source
can be given. However, besides advantages of handling large orders separately there are
also disadvantages. A negative aspect of the break quantity rule is the (possible) occurrence
of a setup cost for every large order. Another disadvantage is that the response time to the
 Dekker et al./ On the use of break quantities in multi{echelon distribution systems 2

customer increases, since in general the distance from the lowest level of the distribution
system to the customer is smaller than from a higher level (e.g. Axsater [3]). An increase
in the response time to the customers with large orders will induce penalty costs, that
may represent price rebates to persuade customers to accept the longer delivery time. The
idea of using price rebates to compensate longer delivery times was already presented by
Jaikumar and Kasturi Rangan [10, 12]. They analyzed the question of how to set price
rebates that allocate orders to dierent stages in a multi{echelon distribution system. The
rebates dier from one echelon to the other and they depend on the delivery location, so
customers are forced to choose the location from which they take possession of a product.
The main goal of their papers was to structure the optimal buying arrangement such that
customers minimize their costs simultaneously, while the manufacturer maximizes prot.
They did not analyze the eect on the inventories, and they only considered the case where
all orders were known beforehand.
The analysis in this paper is motivated by a case study for a company in Western{Europe
that delivers self{adhesive materials to customers through a multi{echelon distribution
system (Nass, Dekker and Van Sonderen{Huisman [15]). The multi{echelon distribution
system is composed of Plant Service Centres (PSC's), which are associated to production
plants and thus have both production and warehouse facilities, and of Distribution Centres
(DC's), which are only used as stockpoints. In this system the customers, from all over
Europe, could be served either by the nearest PSC or by the nearest DC. To operate this
system the company needed a tactical rule to allocate orders from customers to dierent
locations. The company used the break quantity rule, which was implemented such that
orders with a size smaller than or equal to the break quantity are delivered from the nearest
DC within 48 hours, and orders with a size exceeding the break quantity are delivered from
the nearest PSC within 10 days. Using this concept it was possible to reduce the average
system costs considerably.
In this paper we will study the eects of the break quantity rule for a two{echelon distribu-
tion system consisting of one warehouse and one retailer. The main goal of this paper is to
show the eect of the break quantity rule on the inventory costs at the retailer. Therefore
we make the simplifying assumption that the warehouse can always satisfy replenishment
orders from the retailer and large orders from customers. Hence, we do not take into ac-
count the eect of the break quantity rule on the inventory at the warehouse. However, it
is assumed that additional costs are incurred for delivering a large order separately, which
may include price rebates and setup costs. In the next section an approximation for the
 Dekker et al./ On the use of break quantities in multi{echelon distribution systems 3

minimum long run average system costs as a function of the break quantity are derived.
Moreover, an easy and tractable condition under which the break quantity rule reduces
the total long run average costs is given. In Section 3 an algorithm that determines the
optimal break quantity is presented, while it is assumed that the management puts a re-
striction on the minimum number of customers served by the retailer. Finally, in Section
4 computational results are reported, to illustrate the in uence of the break quantity rule
on the average costs for a wide range of parameter settings. The results indicate that in
many of the situations analyzed, a signicant reduction in the long run average costs can
be obtained. Moreover, it is shown that using the approximated cost function leads to
similar results to those obtained by exact calculation.

2 The model
Consider a single product, two{echelon distribution system, consisting of one warehouse
and one retailer. Assume that at the retailer the inventory is reviewed at discrete time
points and that every review an order is placed to raise the inventory to the order up
to level S . This policy is in general optimal when the xed ordering costs are negligible
(e.g. Scarf [17]). In most of the literature on multi{echelon inventory theory (e.g. Eppen
and Schrage [6], Federgruen and Zipkin [7], Langenho and Zijm [14]) the order up to level
policy is considered, since it was proved by Clark and Scarf [5] that in this case the inventory
problem can be decomposed. Tagaras and Cohen [20] observed that low ordering costs have
become increasingly common with the advent of distribution information systems, and/or
the assumption that other considerations (e.g. routing schedules) dictate ordering and
delivery of materials every review period. Langenho and Zijm [14] noted that in many
practical cases the xed costs are already accounted for at a higher level, e.g. in a multi{
product environment.
If a replenishment order is placed in period t, it will be delivered in period t + L, i.e. there is
a xed lead time L. Let the lead time be an integral multiple of the review time, so that the
review time can be taken as unity. The sequence of events in any period is: review, order,
demand and delivery of the replenishment order that was placed L periods ago. Demand
that cannot be satised from stock on hand will be backlogged, hence the ordering cost
need not be considered, since all demand must be replenished and after every review an
order is placed. At each review a holding cost h is charged for every unit of stock on hand,
and for every unit backlogged a shortage cost p is charged. Orders arrive according to a
 Dekker et al./ On the use of break quantities in multi{echelon distribution systems 4

Poisson process with arrival rate , and the sizes Y1; Y2; : : : of the consecutive orders are
i.i.d. with distribution function F . For convenience we assume that F has a continuous
positive density on its support. Finally, as already mentioned in the introduction, it is
assumed that the warehouse can always satisfy replenishment orders from the retailer and
large orders from customers. Hence, we do not consider the inventory cost of the warehouse.
We will rst focus on the eects of the break quantity rule on the inventory costs at the
retailer. If the break quantity rule is applied, then all orders with size smaller than or equal
to the break quantity are delivered by the retailer, while the large orders are delivered from
the warehouse. In the next gure the distribution system is illustrated.

  
large orders


WAREHOUSE
replenishment
orders
-

RETAILER
small orders
j
-

CUSTOMER

Figure 1: A two{echelon distribution system with the break quantity rule

If the break quantity equals q, the order size of the ith customer with respect to the retailer
equals Y 1f i  g, where
i Y q 8
< 1 if Y  q
1f i  g = : i
Y q
0 if Y > q i

denotes the indicator function.

Let D be the total demand for the retailer during L +1 periods if the break quantity equals
q

q, and let its distribution be given by G . Then it is well{known (see e.g. Karlin [11]) that
q

the expected one{period holding and shortage costs for the above system are given by
Z Z1
L(S; q) = h 0 (S y)dG (y) + p
S
q (y S )dG (y)q (1)
S

Applying standard newsboy arguments (e.g. Porteus [16]), it follows that the optimal order
up to level S  is given by
S  = G 1 ( p +p h ) q
 Dekker et al./ On the use of break quantities in multi{echelon distribution systems 5

with G 1 : [0; 1] ! [0; 1) the inverse function of G , i.e.

q q

G 1 (x) = inf fy  0 : G (y)  xg

q q

If  and 2 respectively denote the mean and variance of the random variable D , it
q q q

follows that (e.g. Tijms [21])

 = (L + 1)E [Y 1f i  g]
q i Y q (2)
2 = (L + 1)E [Y 21f i  g]
q i Y q (3)
Substituting S  in (1) we obtain that the minimum expected one{period holding and
shortage costs are given by
Z1
L(S  ; q ) = h + (p + h)
q p
G 1 (z)dz
q
p+h
Z1
=  (p + h)
q p
G^ 1 dz q (4)
p+h

with G^ the distribution of the normalized random variable q q q , i.e. G^ ( q q )  G (z).

q
D


q
z


q

Unfortunately the integral in the above expression is in general dicult to compute. More-
over, it depends on the value of q. However, since G^ is the distribution of a normalized q

compound Poisson sum of i.i.d. random variables, we may apply the central limit theorem
(Feller [8]) and approximate G^ by the standard normal distribution  with expectation
q

0 and variance 1. The main advantage of using the standard normal distribution is that
for every q the distribution G^ is replaced by the same distribution, and thus the above
q

integral is no longer depending on the break quantity.

Approximating p1+ph G 1(z)dz by R p1+ph  1(z)dz it follows from (3) and (4) that
R
q

q
L(S ; q)  c1 E [Y 21f i g] i Y q (5)
with
q Z1
c1 :=  1 (z)dz
(L + 1)(p + h) p
p+h

= (L + 1)(p + h) p1 exp( 12 ( 1 ( p +p h ))2)

q
(6)
2
Since E [Y 21f i  g] is an increasing function in q one can easily see that the expected one{
i Y q

period holding and shortage costs for the retailer are reduced by the break quantity rule.
However, the break quantity rule also causes extra costs. Assume that the unit cost of
 Dekker et al./ On the use of break quantities in multi{echelon distribution systems 6

delivering large demand by the warehouse is given by c and the xed costs per large order
is given by K . Since the normal ordering costs are not taken into account in the analysis,
the value c should be interpreted as the additional unit cost for not delivering the large
demand from the retailer. In the value of c are included the reduction in transportation
and handling costs due to the fact that the order can be delivered directly to the customer,
and the penalty costs for a possible increase in waiting times for the customers with large
orders. Although it is in general not easy to determine the value of c, we assume that it is
known. Furthermore, observe that c does not necessarily has to be positive. In the value
of K are included the possible occurrence of setup costs for delivering a large order, and a
xed penalty cost per large order. Clearly, the value of K should satisfy K  0.
The order size of the ith customer with respect to the warehouse equals Y 1f i g, so i Y >q

the costs for serving the ith customer from the warehouse are equal to (cY + K )1f i g.
i Y >q

Hence it follows by the memoryless property of the Poisson process that the expected
costs for delivering the large orders separately for a typical period are given by E [(cY + i

K )1f i g] = cE [Y 1f i g] + K (1 F (q)). These observations lead to the following

Y >q i Y >q

conclusion.

Conclusion: The minimum costs for the model with a break quantity 0  q  1 are
approximated by
q
C (q) = c1 E [Y 21f i g ] + c2E [Y 1f i g ] + c3(1 F (q))
i Y q i Y >q (7)
where c1 is given by (6), c2 := c and c3 := K .

By the above expression for the cost function C (q) it is easy to derive a sucient and
tractable condition forqthe break quantity rule to be protable, i.e. that there exists some
q for which C (q) < c1 E [Y 2]. Since E [Y 1f i  g] = R0 x dF (x) one can easily verify that
i i
k
Y q
q k

the derivative of C (q) is given by

 
0 1
C (q) = 2 c1q v(q) c2q c3 f (q)
2 (8)
where   12
v(q) := E [Y 21f i  g]
=
i Y q

A sucient condition for the break quantity rule to be protable is given in the next
theorem.
 Dekker et al./ On the use of break quantities in multi{echelon distribution systems 7

Theorem 2.1 Assume that c1, c2 and c3 are nite{valued and let 0 < u < 1 be given by
 q 
u := c2 + c22 + 2c1c3E [Yi2] 1=2 =c1E [Y 2]
i
1=2 (9)
q
If F (u) < 1, then there exists a q < 1 such that C (q ) < c1 E [Yi2 ].
Proof: Observe that v(q) is decreasing in q and thus v(q)  v(1) = (E [Y 2]) i
1=2
. Since
u uniquely solves the equality
1 c E [Y 2] 1 2 x2 c x c = 0
=

21 i 2 3

for x > 0, we obtain for q > u that

1 c v(q)q2 c q c  1 c E [Y 2] 1 2 q2 c q c > 0
=

21 2 3
21 i 2 3

and so it follows by f (q) > 0 for all q that C 0(q)  0 for q > u.
Moreover, observe that F (u) < 1 implies that there exists a q^ > u such that
q f (^
q) > 0, and
thus C 0(^q) > 0. Hence, there exists a q < 1 such that C (q) < C (1) = c1 E [Y 2]. i2
In the next section an algorithm which determines the optimal break quantity will be
presented.

3 Determining the optimal break quantity

In this section the main focus will be on solving the problem
inf fC (q) : a  q  bg (P)
where C (q) is dened in (7), and a and b are dened as
a := inf fx > 0 : F (x)   g
b := supfx > 0 : F (x) < 1g
with 0 <  < 1. Observe that a is a lowerbound on the break quantity which is set
by the management through a restriction on the percentage of customers served by the
retailer. The motivation for the introduction of this restriction is threefold. First, due
to xed overhead costs at the retailer and for competitive reasons it is desirable that a
large percentage of customers is served through the retailer. Secondly, it is likely that
 Dekker et al./ On the use of break quantities in multi{echelon distribution systems 8

a customer ordering a small batch is less sensitive to a price reduction than a customer
ordering a large batch, implying that the cost for delivering large orders from the warehouse
becomes relatively higher if the break quantity decreases. Hence, the value of c and K may
only hold for relatively large break quantities. Finally, the normal approximation of the
lead time demand is only justied for relatively large break quantities, because for low
break quantities the average number of orders in a period becomes too small and the
central limit theorem may not apply anymore. Observe also that b is an upperbound on
the break quantity, since by denition the size of an order cannot exceed b.
Any q 2 (a; b) solving (P) must satisfy the necessary rst{order optimality condition
C 0(q) = 0, i.e. q must be a stationary point. By Theorem 2.1 we obtain that C 0(q) > 0
for q > u. Since u is nite{valued it follows that the set of possible stationary points is
bounded and given by (a; min[b; u]). For notational convenience we dene b0 := min[b; u].

' $
Before presenting an algorithm which reduces the set of possible optimal points, we need to
introduce the function  : [a; b0] ! [a; b0]. Let the value of (q) for q 2 [a; b0] be determined
by Algorithm 3.1.

Step 0 Set i := 0 and q := q i

Step 1 Let q +1 > 0 solveq12 c1v(q )q2+1 c2q +1 c3 = 0,

i i i i

i.e. q +1 :=
i c2 + c22 + 2c1c3v(qi) =c1 v(qi)
Step 2 If q +1 = q
i i then (q) := qi+1 and stop,

& %
if q +1  a then (q) := a and stop,
i

if q +1  b0 then (q) := b0 and stop,

i

otherwise set i := i + 1 and go back to Step 1

Algorithm 3.1: An algorithm to calculate (q)

Observe that 21 c1v(q)q2 c2q c3 = 0 for a < q < b0 is equivalent to C 0(q) = 0. Moreover, if
C 0(q ) < 0 one can easily verify that the value of q +1 > 0 solving 21 c1v(q )q2+1 c2q +1 c3 =
i i i i i

0 must satisfy q +1 > q . Since v(q +1) < v(q ) if q +1 > q it follows that C 0(q +1) < 0, thus
i i i i i i i

q0; q1; : : : is an increasing sequence and its limit point q1, if it exists, satises C 0(q1) = 0.
A similar observation holds for the case where C 0(q ) > 0. We are now able to proof the
i

next lemma that will help us to determine the set of stationary points.
 Dekker et al./ On the use of break quantities in multi{echelon distribution systems 9

Lemma 3.1 Dene I (x; y) := fz : min(x; y)  z  max(x; y)g. Then for any q 2 [a; b0] it
6 0 if y 2 I (q; (q)) n f(q)g. Moreover,
follows that C 0(y) =
8
<  0 if  (q ) = a
>
>
>
C 0((q)) = > = 0 if a < (q) < b0
>
>
:  0 if  (q ) = b0

Proof: From the observations above this lemma it follows that if Algorithm 3.1 starts
with q 2 [a; b0], no point between q and (q) can be a stationary point, and hence the rst
implication is proved. If a < (q) < b then (q) is a limit point of the sequence q0; q1; : : :
generated by Algorithm 3.1 and thus (q) is a stationary point. If q  a and q +1  a
i i

then either q +1 = q and thus C 0(q +1) = 0, or q +1 < q and C 0(y) > 0 for q +1  y  q .
i i i i i i i

A similar observation holds for the upperbound b0. 2

' $
Algorithm 3.2 will reduce the set of possible optimal break quantities on [a; b0]. It is
assumed that the number of stationary points is nite. Hence, there exists an  > 0 such
that if q1 and q2 are stationary points, it follows that jq1 q2j > .

Step 0 Set X := fq : a  q  b0g, and set X  := ;

Step 1 Calculate I (a; (a)) and I (b0; (b0)), set X := X I (a; (a)) I (b0; (b0))
and set X  := X  + f(a); (b0)g
Step 2 Let q be the midpoint of the largest convex subset of X
and calculate (q)

& %
Step 3 Set X := X I (q; (q)) and X  := X  + f(q)g
Step 4 If the length of the largest convex subset of X is smaller than ,
then go back to Step 2, else stop.

Algorithm 3.2: An algorithm to reduce the set of possible optimal points

Taking the element of X  with the lowest associated average costs, we get the optimal break
quantity. This algorithm is a special case of the C{programming algorithm introduced
by Schniedovich [18]. It can easily be veried that the algorithm will terminate within
O(log 0 ) number of iterations.
b

a
 Dekker et al./ On the use of break quantities in multi{echelon distribution systems 10

4 Computational Results
In this section, Algorithm 3.2 will be implemented to solve problem (P) for a wide variety
of parameters. We will analyze the case where the order sizes of the customers are Gamma
distributed. The motivation for this choice is threefold. First, for this distribution only
positive order sizes are allowed, in contrast with for example the normal distribution.
Secondly, the distribution can easily be tted around the mean and the variance of the
order sizes. In many practical situations managers do not have complete information on
the demand distribution, but they only know the mean and variance. Finally, in order to
compute v(q) one only needs to compute the incomplete Gamma function, for which many
numerical approximations are available. We used an approximation given by Abramowitz
and Stegun [1].
In order to determine how much the approximation of the lead time demand distribution
by the normal distribution in uences the results, we also will calculate the exact lead time
demand distribution to determine the exact optimal break quantity and its corresponding
minimum costs. We used Adelson's recursion scheme (Adelson [2]) to determine the cdf of
the compound Poisson distribution. However, since this recursion scheme is only dened for
discrete distributions, we transformed the continuous Gamma distribution into a discrete
distribution in the following way:
8
P (X
<
= x) := : F (x) F (x 1=N ) for x = 1=N; 2=N; : : :
0 otherwise
where N  1 is an integer number. Clearly, the value of N determines the accuracy of
the transformation. For our computations we used N = 5, and to determine the optimal
break quantity a simple enumeration technique was used.
Generally, management will be interested in obtaining a quick estimate of the optimal break
quantity. For this purpose we also analyzed the quality of approximating the optimal break
quantity by u, which is given in Theorem 2.1. This value can be calculated without having
to perform any numerical procedures. Henceforth we will refer to this approximation as
the u{approximation.
We have analyzed the benets of the break quantity rule for many dierent parameter
values. However, the following parameters were always xed:  = 10, L = 2, h = 1,  :=
E [Y ] = 10 and  = 0:75. The other parameters were varied, i.e. p = f5; 10; 15; 20; 25g, c =
i

f 1; 2; 0; 1; : : : ; 7g, K = f25; 50; 75; 100g and 2 := V AR[Y ] = f50; 100; 150; 200; 250g.
i

By varying 2 for  xed we basically vary the coecient of variation c , dened as c := ,

Y Y


 Dekker et al./ On the use of break quantities in multi{echelon distribution systems 11

which is often used as an indicator of the variability of demand. Observe that for these
parameter settings the maximum
p 2+coecient
p
of variation of the lead time demand, obtained
2
when 2 = 250, is equal to p ( +1) = 10p35030  0:34, and thus the probability of negative


 L


lead time demand when using the normal approximation is negligible.

In total we evaluated the eect of the break quantity rule for 1000 parameter settings.
In Table 1 the cost reduction obtained by the 3 dierent methods is classied. In the
rst column the classes are presented, in the other columns the frequencies obtained by
exact calculation, Algorithm 3.2, and the u{approximation are reported. We mention that
the points (a) and (b0), calculated in Algorithm 3.2, in all cases coincided, hence with
Algorithm 3.2 the solution was quickly determined.

cost frequencies
reduction (%) discrete exact calculation Algorithm 3.2 {approximation
u
[ 0.00, 1.00) 499 598 601
[ 1.00, 5.00) 165 111 109
[ 5.00, 10.00) 66 53 59
[ 10.00, 25.00) 90 75 73
[ 25.00, 1 ) 180 163 158

Table 1: Cost reduction obtained by break quantity rule

One can see that in most of the cases the cost reduction obtained by the break quantity
rule was less than 1%, but in more than 15% of the cases the reduction exceeded 25%.
This implies that if the break quantity rule is protable, the costs can be reduced consid-
erably. We also observe that the results of the u{approximation are close to the results of
Algorithm 3.2.
In table 2 some of the results are presented. We have selected the following 16 representative
cases: p = f10g, c = f 1; 1; 3; 5g, K = f50; 100g and 2 = f100; 200g. In the rst four
columns the values of the parameters are shown. In columns 5, 6 and 7 the optimal break
quantity q, the corresponding percentage of customers served by the retailer F (q), and
the corresponding relative reduction in average costs C:R:(%) are presented, which are
obtained by exact calculations. Note that C:R:(%) is equal to the ratio of the reduction in
average costs due to the break quantity rule, and the average cost without using the break
quantity rule. In columns 8, 9 and 10 similar results are presented, but then obtained by
Algorithm 3.2. Algorithm 3.2 was terminated when the largest \uncovered" subset was
smaller than , which we set equal to 10 3 . Finally, in the last three columns the value of
u, the percentage of customers served by the retailer and the minimum costs are presented,
 Dekker et al./ On the use of break quantities in multi{echelon distribution systems 12

when the break quantity is equal to u.

parameters exact, N = 5 Algorithm 3.2 u {approximation

p c K  2 q F (q  ) C.R.(%) q F (q  ) C.R.(%) u F (u) C.R.(%)
10 -1 50 100 21 0.878 26.10 22 0.892 24.27 26 0.927 23.06
10 1 50 100 48 0.992 2.26 53 0.995 1.17 55 0.996 1.15
10 3 50 100 81 1.000 0.14 100 1.000 0.01 100 1.000 0.01
10 5 50 100 110 1.000 0.01 153 1.000 0.00 153 1.000 0.00
10 -1 100 100 37 0.975 7.70 39 0.980 6.17 41 0.984 6.06
10 1 100 100 62 0.998 0.83 69 0.999 0.32 70 0.999 0.32
10 3 100 100 90 1.000 0.07 112 1.000 0.01 112 1.000 0.01
10 5 100 100 116 1.000 0.01 161 1.000 0.00 161 1.000 0.00
10 -1 50 200 19 0.832 50.65 20 0.845 48.62 28 0.905 45.40
10 1 50 200 48 0.972 9.60 54 0.980 6.19 63 0.988 5.76
10 3 50 200 96 0.998 1.32 118 0.999 0.30 120 0.999 0.30
10 5 50 200 135 1.000 0.24 185 1.000 0.01 185 1.000 0.01
10 -1 100 200 36 0.942 24.66 38 0.948 21.70 44 0.965 20.66
10 1 100 200 66 0.990 5.13 74 0.994 2.79 79 0.995 2.72
10 3 100 200 106 0.999 0.92 131 1.000 0.18 132 1.000 0.18
10 5 100 200 142 1.000 0.19 194 1.000 0.01 194 1.000 0.01

Table 2: The eect of the break quantity rule for dierent methods
It can be seen that the success of the break quantity rule very much depends on the
value of c. Hence, it is very important that the management has good information on the
additional variable costs for delivering a large order from the warehouse. Observe that the
main positive component of c is the penalty cost that has to be paid to compensate for
the (possibly) longer delivery times. Therefore, the management should try to keep this
cost as low as possible, e.g. by setting agreements with the customer. One can think of
giving each customer with a large order some xed price reduction if the customer accepts
the longer waiting time due to the break quantity rule. For this reason, it is interesting
to analyze the eect of the values of c and K on the cost reduction obtained by the break
quantity rule. In Figure 1 the cost reduction is plotted for values of p and 2 xed at
p = 10 and 2 = 150, and for values of c and K varying from c = [ 2; 7] and K = [25; 100].
Moreover,  is set equal to zero, so no restriction on the percentage of customers served
by the retailer is taken into account.
This gure is based on results obtained by using Algorithm 3.2. It shows that the sensitivity
of the cost reduction for c is much larger than for K . Hence, it seems protable to \buy
o" the penalty cost for delivering large orders separately by giving a xed price reduction
for each large order. Whereas the value of c has a large in uence on the fact whether or not
 Dekker et al./ On the use of break quantities in multi{echelon distribution systems 13

cost reduction (%)

100

50

-2 0-1 0 90 100
1 2 3 60 70 80
c 4 5 6 50 K
7 30 40

Figure 2: The impact of c and K on the cost reduction

the break quantity rule is protable, from Table 2 it is also observed that the coecient of
variation of the order sizes has a big impact on the size of the cost reduction. For instance,
for 2 = 200 the cost reductions are almost twice as big as for 2 = 100. Finally, we
observed that the shortage cost p does not seem to have a signicant in uence on the cost
reduction that can be obtained by using the break quantity rule.
The cost reduction obtained by the normal approximation of the lead time demand in all
cases underestimated the actual cost reduction. Therefore, Algorithm 3.2 will generally
provide a lowerbound on the cost reduction that can be obtained by using the break
quantity rule.

5 Conclusions
In this paper a multi{echelon distribution system is analyzed, where orders from customers
can be delivered from any level in the system. To allocate orders from customers to the
dierent locations, we introduced a so{called break quantity rule. The break quantity
determines whether an order of a customer is small or large, and the break quantity rule is
implemented in the following way: a small order will be delivered from the lowest echelon,
a large order will be delivered from a higher echelon. It is shown that using this rule, the
inventory costs at the lowest echelon can be reduced signicantly. However, in general the
delivery of a large order from a higher echelon will also cause extra costs, e.g. because the
 Dekker et al./ On the use of break quantities in multi{echelon distribution systems 14

delivery lead time may increase. An approximation for the minimum costs as a function
of the break quantity is derived in Section 2, and an algorithm to determine the optimal
break quantity is presented in Section 3. Moreover, an easy and tractable condition under
which the break quantity rule is protable is given. This condition is that with a positive
probability the order size of a customer exceeds the value u, which can be calculated
without having to perform any numerical procedures.
From the computational results in Section 4 it follows that if the management can persuade
the customer to accept the break quantity rule (for a possible xed price reduction and/or
small price reduction per unit demanded), and the demand is erratic, the total costs can be
decreased signicantly. In almost 15% of the cases we considered, the cost reduction is more
than 25%. It is also shown that by using approximated minimum average cost function,
the optimal break quantity and the corresponding cost reduction are well approximated.
Hence, the results obtained by Algorithm 3.2 give good insight into the eect of the break
quantity rule on the distribution system. Finally, it turned out that approximating the
optimal break quantity by u, given in Theorem 2.1, gives satisfactory results.

References
[1] M. Abramowitz and I. Stegun. Handbook of Mathematical Functions. Dover, New
York, 1965.
[2] R.M. Adelson. Compound Poisson distributions. Operations Research Quarterly,
17:73{75, 1966.
[3] S. Axsater. Continuous review policies for multi{level inventory systems with stochas-
tic demand. In S.C. Graves, A.H.G. Rinnooy Kan, and P. Zipkin, editors, Handbooks
in OR & MS, Vol. 4. Elseviers Science Publishers, North{Holland, 1993.
[4] L.D. Burns, R.W. Hall, D.E. Blumenfeld, and C.F. Daganzo. Distribution strategies
that minimize transportation and inventory costs. Operations Research, 33:469{490,
1985.
[5] A.J. Clark and H. Scarf. Optimal policies for a multi{echelon inventory problem.
Management Science, 6:475{490, 1960.
[6] G. Eppen and L. Schrage. Centralized ordering policies in a multi{warehouse sys-
tem with lead times and random demand. In L.B. Schwarz, editor, Multilevel pro-
 Dekker et al./ On the use of break quantities in multi{echelon distribution systems 15

duction/inventory control systems: Theory and practice. North{Holland, Amsterdam,

1981.
[7] A. Federgruen and P. Zipkin. Computational issues in an innite{horizon, multi{
echelon inventory model. Operations Research, 32:818{836, 1984.
[8] W. Feller. An Introduction to Probability Theory and Its Applications, vol. 2. Wiley,
New York, 1971.
[9] G. Gallego and D. Simchi-Levi. On the eectiveness of direct shipping strategy for
the one{warehouse multi{retailer R{systems. Management Science, 36:240{243, 1990.
[10] R. Jaikumar and V. Kasturi Rangan. Price discounting in multi{echelon distribution
systems. Engineering Costs and Production Economics, 19:341{349, 1990.
[11] S. Karlin. Steady state solutions. In K.J. Arrow, S. Karlin, and H. Scarf, editors,
Studies in the mathematical theory of inventory and production. Stanford University
Press, Stanford, Ca., 1958.
[12] V. Kasturi Rangan and R. Jaikumar. Integrating distribution strategy and tactics: a
model and an application. Management Science, 37:1377{1389, 1991.
[13] A.G. de Kok. Demand management in a multi{stage distribution chain. Technical Re-
port TUE/BDK/LBS/93{35, Eindhoven University of Technology, The Netherlands,
1993.
[14] L.J.G. Langenho and W.H.M. Zijm. An analytical theory of multi{echelon produc-
tion/distribution systems. Statistica Neerlandica, 44:149{174, 1990.
[15] R. Nass, R. Dekker, and W. van Sonderen-Huisman. Distribution optimization by
means of break quantities: A case study. Technical Report 9375/A, Econometric
Institute, Erasmus University Rotterdam, 1993.
[16] E.L. Porteus. Stochastic inventory theory. In D.P. Heyman and M.J. Sobel, editors,
Handbooks in OR & MS, vol. 2. Elsevier Science Publishers B.V., North{Holland,
1990.
[17] H. Scarf. A survey of analytic techniques in inventory theory. In H. Scarf, D. Gil-
ford, and M. Shelly, editors, Multistage inventory models and techniques. Stanford
University Press, Stanford, Ca., 1963.
 Dekker et al./ On the use of break quantities in multi{echelon distribution systems 16

[18] M. Schniedovich. C{programming: an outline. Operations Research Letters, 4:19{21,

1985.
[19] J.E. Sussams. Logistics Modelling. Pitman Publishing, London, 1992.
[20] G. Tagaras and M.A. Cohen. Pooling in two{location inventory systems with non{
negligible replenishment lead times. Management Science, 38:1067{1083, 1992.
[21] H.C. Tijms. Stochastic models: An algorithmic approach. Wiley, New York, 1994.