Optimal Inventory Policies for Assembly Systems under Random Demands

Author(s): Kaj Rosling

Source: Operations Research, Vol. 37, No. 4 (Jul. - Aug., 1989), pp. 565-579
Published by: INFORMS
Stable URL: http://www.jstor.org/stable/171257 .
Accessed: 12/05/2011 10:50

Your use of the JSTOR archive indicates your acceptance of JSTOR's Terms and Conditions of Use, available at .
http://www.jstor.org/page/info/about/policies/terms.jsp. JSTOR's Terms and Conditions of Use provides, in part, that unless
you have obtained prior permission, you may not download an entire issue of a journal or multiple copies of articles, and you
may use content in the JSTOR archive only for your personal, non-commercial use.

Please contact the publisher regarding any further use of this work. Publisher contact information may be obtained at .
http://www.jstor.org/action/showPublisher?publisherCode=informs. .

Each copy of any part of a JSTOR transmission must contain the same copyright notice that appears on the screen or printed
page of such transmission.

JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of
content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms
of scholarship. For more information about JSTOR, please contact support@jstor.org.

INFORMS is collaborating with JSTOR to digitize, preserve and extend access to Operations Research.

http://www.jstor.org
 OPTIMALINVENTORYPOLICIESFOR ASSEMBLYSYSTEMS
UNDER RANDOMDEMANDS
KAJ ROSLING
Linkoping Institute of Technology, Linkoping, Sweden
(Received March 1987; revision received December 1987; accepted April 1988)

This paper considers an inventory model of an assembly system with random demands and proportional costs of
production and stock holding activities. The model is a generalization of the facilities in series model of Clark and Scarf.
Under an assumption on the initial stock levels that in any case should appear in the long run (Long-Run Balance), it is
demonstrated that the assembly system can be remodeled as a series system. Hence, simple reorder policies are optimal
and the computational requirements are drastically reduced. Simple approximate policies are suggested for the generally
short period of time when systems might be out of long-run balance. There is some discussion of extensions.

T his paper examines a periodic review, infinite
horizon model of an assembly inventory system
with proportionalcosts of production. In an assembly
system, a number of components acquired from out-
side vendors are assembled, typically in several stages,
into subassembliesand then, finally, into a single end
product. A generic system is depicted in Figure 1. In
each period, production orders are made for all items.
The ordered amounts are available after a fixed lead-
time, and are then held in stock until assembled into
other items or demanded by customers. The random
customer demands occur only for the end product
and unsatisfied demands are backlogged.
The present model is a direct generalization of the
initial work on multiechelon inventory theory due to
Clark and Scarf (1960). They found an efficient way
to compute an optimal ordering policy for a pure
series system and suggested an approximate way to
handle distributive systems, i.e., systems where a cen-
tral depot serves several retailers.Fukuda (1961) gen-
eralized the approach to allow for disposal of stocks.
The work of Clark and Scarf was computationally
streamlined and generalized to the stationary infinite
horizon case by Federgruen and Zipkin (1984).
Rosling (1989) derived bounds on the optimal reorder
levels. A similar series model under continuous review
has been investigated by Axsater and Lundell (1984).
General multilocation inventory problems with
proportional costs have been studied by Bessler and
Veinott (1966) and Karmarkar (1981, 1987). They
obtained simple form optimal policies under the
restrictive assumption that all lead-times equal zero.
Schmidt and Nahmias (1985) investigated a finite
horizon model of two components assembled into a
single end product. They derived a complicated opti-
mal policy under general assumptions on lead-times
and initial stock levels.
In the present paper, we derive optimal policies for
general assembly systems under a restriction on the
initial stock levels. Our interest is limited to those
stock levels that can prevail in the long run for systems
run by optimal or otherwise reasonable policies.
Under this initial condition, it turns out that the
assembly system can be interpretedas a series system,
and hence, there is a simple optimal policy that can
be calculatedconveniently by Clarkand Scarf's (1960)
approach. This result contrasts most encouraginglyto
the computational intractabilityof a generallyoptimal
policy.
The assumed absence of setup costs in production
limits the applicability of the present model. The
inclusion of such costs in stochastic multiechelon
inventory models is extremely difficult, however.
Clark and Scarf (1962) attempted to include setup
costs in their series model, but they were only able to
derive upper and lower bounds on the minimal cost.
Their approachwas extended to distributivestructures
by Hochstadter (1970) and Rosling (1977). As far as
we know, there has been no substantialcomputational
test of this approximateapproach.Other approximate
approaches,possibly bettertested, have been proposed
e.g., by de Bodt and Graves (1985) for a series system
and by Carlson and Yano (1986) for a two-level
assembly system, but the problem is still far from a
satisfactory theoretical and practical solution. Al-
though the present model includes no setup costs, we

Subject classifications: Inventory/production: long-run inventory position; component ordering and product assembly. Programming, nonlinear theory:
multistage production assembly.

Operations Research 0030-364X/89/3704-0565 $01.25

Vol. 37, No. 4, July-August 1989 565 ? 1989 Operations Research Society of America
566 / ROSLING
CD,
~7
outside
supply ) ~(9-----
lead-time
Figure 1. An assembly system.
hope that our results shed some light on this more
complicated problem.
The remainderof this paper is organized as follows.
In Section 1, the model is presented. In Section 2, the
long-run analysis of the inventory levels is carriedout,
and in Section 3 the series interpretation is demon-
strated. The analysis of these sections relies on the
assumption that all echelon holding costs are positive.
The interestingcase where some echelon holding costs
are nonpositive is handled in Section 4. In Sections 3
and 4, suggestionsare given for applying approximate
policies when the long-run conditions are not satisfied
initially. In Section 5, extensions and generalizations
in various directions and to systems with setup costs
in production are discussed. There are two appendices.
The first presents some alternativebut equivalent cost
assumptions. The second contains proofs that are
omitted from the main text.
1. THE MODEL
There are N items (components, subassemblies at
different levels, the end item). They are denoted i =
1, 2, . .. N. Item 1 is the end item. Each nonend item
has exactly one successor item so that the product
network forms a tree rooted in the end item. The units
are defined so that exactly one unit of each item is
required for the end item. The following parameters
describe the system:
s(i) = the unique immediatesuccessor of item i = I,
2, . N; for notational convenience we define
. .
s(l) = 0;
=
A(i) the set of all successors of item i;
P(i) = the set of immediate predecessors of item i =
0, 1, . .. a N;
B(i) = the set of all predecessorsof item i;
1,= number of periods (lead-time) for assembly (or
delivery) of item i.
Everything occurs at the beginning of periods and
in the following order. First, outstanding orders for
purchase and assembly arrive and new ordering deci-
Z4=2 z 33usoe
4 3 Customer
,,1 > demands
2s=4 21
sions are made. Then, old backlogs are expedited and,
thereafter,customer demands occur for the end prod-
uct. Finally, but still at the beginning of the period,
costs are incurredfor backlog and leftover inventories.
Holding costs are chargedwhen items are in stock but
also when they are in assembly into their successor
item. The following variablesare used:
demands in period t for the end item;
=
demands of different periods are iden-
tically and independently distributed,
f 0;
/(*), 4(.) = density and distribution function for (,;
X= E(Q,),the expected value of (,, X> 0;
Xi, = echelon inventory position of item i in
period t before ordering decisions are
made (= inventory on hand including all
units already assembled into other items,
+ units in assemblyor on order - backlog
of the end item);
Yi,= echelon inventory position of item i after
production decisions are made, Yi, > Xi,;
X',= echelon inventory on hand of item i
in period t before ordering decisions
are made but after assembles arrive =
Yi,_,,- =-,j Jk. (Throughout
J-k this paper,
it is assumed that a sum, ',, equals zero
if a > b.) Note that XI, > Y,;,if i E P(k)
and if 1,= O then X, = YI,.
Although the demand distributions are assumed to
be stationary,this assumption is of little consequence
for the main results of the next sections. The amount
(Y,, - X,,) is ordered for the purchase or assembly of
item i in period t. It arrives after 1iperiods. The cost
of running the system is determined by the following
parameters:
Hi = unit installation holding cost per period of
item i;
hi = unit echelon holding cost per period of item i,
hi,=H -ZkEP(i,H,k;
 Optimal InventoryPolicies for Assembly Systems / 567
p = unit backloggingcost per period of the end item; Problem P
a = period discount factor, 0 < a < 1.
Min E{ a-' al( 1'* hi. Yi,+ ac'(p + HI)
The values of the cost coefficients are restricted
further in Sections 3 and 4 (the Assumption and the
Generalized Assumption, respectively). There is no
explicit proportionalcost of production (or purchase).
I (t - Y,)k/'+(Q)dt)} + constant (2)
Appendix A demonstrates that such costs as well as
such that
several other variations of the cost assumptions
can be handled by modifications of the present x, < Y1i Xl;, for all k E P(i) and all i, t (3)
coefficients.
where
Holding costs are proportional to installation stock
on hand and in assembly into the succeeding item,
i.e., to (X',-,) -(XI,),- ,) for i - 2. Total cost in iS=I-Ik

period t given , is thus

and
N

i Hj(X'j,I-XI Xj,-=Yi,_,-41-l

*= an optimal solution to problem P.

+ H, *Max(0, X', - ,)+ p * Max(0, (,X',)
A' In the case where a = 1, the averagecost per period
- h
hi (X'j- 1)+ (p + HI) Max(O, ,-X',)
. is minimized instead of the total discounted cost. The
averagecost is constructedby multiplying (2) with the
since Max(0, a) = a + Max(0, -a) and hi = Hi - discountrate (1 - a)/a whereafter a -> 1. The last
constant in (2) (and its average cost analog) equals
ZEkP(i) H,. This reformulation into echelon holding A'
costs is similar to Federgruen and Zipkin and is of -Ehi. X. (i +1)a"i/(1- a) if a <1
critical importance for the subsequent application of
N
Clark and Scarf's (1960) computational approach.
- hi * X(li+ 1) if a = 1. (4)
The linearityof the original holding costs is apparently
necessary for its validity. It still works, however, if a
shortage cost, nonlinear in the amount backlogged,is
2. THE LONG-RUNINVENTORYPOSITION
added to the costs above (cf. Appendix A3).
Recalling that X,+, - +l = Y,- > , the total We need the following additional notation.
expected (E) cost over the infinite horizon can be M= total lead-time for item i and all its successors;
written as M,, = 0 and M, i; + XkE() i, for i =
1 ,2, ... N.
E aE'-' aE '[h1(Y11- X[11
+ 1])] Thus, M, = 11.
The items are supposed to be indexed so that
I M, a M, for all i and
a "/(p+ HI) X (- Y, 1,OI+ ) d() (1)
k< i for all kEA(i).
where fb+(.) is the convolution of k(.) over (1, + 1) L= equivalent lead-time for item i
periods. In (1), the present value of the holding costs
incurred for item i in period t + 1iis accounted for in Li = Mi - Mi-
period t, and similarly, the present value of the costs Note that L, = 11 and i >, Li a 0 for i = 1,
of backlogging in period t + 1, is accounted for in 2, ... N.
period t. Consequently, uncontrollable costs due to
historical decisions are neglected. X = echelon inventory position of item i in period
For the sequential developments, it is convenient to t ordered Li periods ago or earlier
break out from (1) (and sometimes neglect) the con- I -l

stant X - [1,+ 1]. When this is done, the problem may - that- E iL=.

be stated as finding decisions Yj,,dependent on system

history up to period t, that solve problem P. Note thatX~',--Yj,if L,= O.
568 / ROSLING
To furtherclarify the meaning of Xi,, if Z,j, = which it might be sold to a customer (as part of the
order for item i placed j periods ago (i.e., in end product).
period t - j) then
Lemma 1. If the Assumption holds, thenfor any opti-
l -I
mal policy for all i and t
X', = x', + Zj>.
E
j=1.,
Y*= XI ifXi, Min XA,"i
k>i
- echelon inventory position of item i in period
t ordered M,-, = s periods ago or earlier = Xi, -, Y* Mi'nX,' if Xi, Min X,-. (6)
k;>i k>ij
Y,- s _
A,k= s = 0, 1, .. . , Mi. Note that
The proof is in Appendix B. The intuitive message of
it i
Lemma 1 is illuminating. Since X'-Yi is an upper
it Xk' for k E B(i) [by (3)] bound on what possibly can be made available of the
end product within M, periods of time, then it serves
Xi,1'-8= Yi, for,= M, no purpose to produce item i < k in excess of Xk'i.
Such a policy incurs only additional holding costs to
iAt-8 = Xj, forA = M,-I
the contrary and, therefore, cannot be optimal. (It
t= it for, = M_ actually suffices to take the minimum of Lemma 1
over Ik > i, k 4 B(I)/Msj) < Mi}. This is so since
t= i' for =Mi). X11,-' "i X,,'l/i for all k by definition if I E B(k).)
For a given value of ,u both X-'"- and XA'-",k > i
Lemma 2. If the Assumption holds thenfor any opti-
bound what is available of the end product within ,u
mal policy, for all i and t
time periods. Thus, they are equally close to the end
item (and the customers), so to speak.
Our intended series interpretationrequiresthat the Y* >- Min(O, Min X, i). (7)
inventory position of the assembly system be in a
certain state called long-run balance. An optimal policy thus satisfies (on a first-comefirst-
served basis) all customer demands within at most
Definition (Long-Run Balance) M, + 1 periods of time.
The assembly system (its inventory position) is in
Lemma 2 is proved under more general assump-
long-run balance in period t if and only if for i =
tions in Appendix B. The assertion is similar to the
1, 2...,5N- I
well known property of the single-item problem that
XA,' X$~
j for -0, 1, ..., AMi- 1. (5) Y* > 0. The only difference is that in the present
context there is no reason to produce item i in excess
Thus, in long-run balance the inventory positions of Xk'jj1', k > i, as noted in Lemma 1. The last
equally close to the end item increase with i, i.e., with statement of Lemma 2 is similar to the obvious fact
total lead-time. that no demand can be backlogged for more than
We will demonstrate that optimal policies (among I + 1 periods in the single-item model when the order-
others) eventually lead the system into long-run bal- up-to level is nonnegative.
ance and then keep it there. To do this, we restrictthe
values of the cost coefflcients as follows. Theorem 1 (Realization of Long-Run Balance). Any
policy that satisfies the inequalities of Lemmas 1 and
Assumption 2 will eventuallylead the system into long-runbalance
(i) hi>0 forall i and keep it there. This will happen no later than MN
A' + 1periods aftertheperiod whenaccumulateddemand
(ii) E - atlu) < p + H,. exceeds Max, Xi,.

The Assumption implies that p > 0. For the average Proof. Consider any item i < N. Suppose that period
cost case (a = 1), it states that all echelon holding q(i) is the first period in which there is production of
costs as well as the penalty cost are positive. Discount- item i. By Lemma 1 Y,, S X'i+,,,, and since
ing in ii is intuitively due to the fact that holding costs
,A+ I=Y 1- < A 1 (- ~t(
are chargedfor item i immediately after assembly, i.e.,
v _= l
MAl)periods ahead of the earliest possible period in < -llit/I
j+ l. izt
I - xAj+ l
(1+ l
 Optimal InventoryPolicies for Assembly Systems / 569
also Yi,+I < X'+I+. Repeating this argument In so far as there exist real-life assembly systems that
shows thatY,, < Xj'+, for all t - q(i). Moreover, for literally satisfy the present model, it would be surpris-
0S , M, and t = q(i) + M, - ing to find them out of long-run balance.
,-,

- Yiq- X (r 3. THE EQUIVALENTSERIES SYSTEM

,-, The series interpretationrelies on the assumption that
vl. _v
< Xfi+iI - > =
_Val-p
- l the assembly system is in long-run balance from the
very first period. The interpretation of the definition
of long-run balance when applied to the initial inven-
Thus, (5) of the Definition holds for i for all t > q +
tory position is investigated at the end of this section
M,, and we are done if q(i) < ti _ s + MN - M, + 1
in Corollary4.
for all i, where s is the period when accumulated
demand exceeds Maxi Xi,. This is demonstrated by Theorem 2 (The Equivalent Series System). If the
an inductive argument. Assume that Y,,,3-X'_,+1 i, Assumption holds and the assembly system is initially
and q(i) < t, for all i > k. Consider item k and suppose in long-run balance, then the optimal policies of the
that y,;,,< X,k=,72l This must hold if q(k)> tk since
tr
assembly system are equivalentto those of a pure series
X- +, < 0. Now system where item i immediately succeeds item i + 1
and where the lead-time of item i is Li. The cost
X-1 =
-; s+1 ,<k coefficient(p + H1) of the series system is the same as
for the assembly system, but the holding cost coeffi-
Ik-I lIk-I cients are modified so that hihi*hi Jai for all i.
<- Yi,- ilk
Proof. The cost function (2) can be written as
for all i > k. Since X/,k <0, Lemma 2 implies produc- I X
tion of k with YkIk f
- Mm
Min El ael-I * ato'
( i * (hiaeli-Li)Yi1
and consequently,
-= (,,

q < tl,. Thus, these inequalities hold for all i > k - 1.

The inductive assumption holds for N since tA =
s + 1, X,1,+, < 0 and so, by Lemma 2, YN.+ >- 0. + a,'X . (p + HI) f ( - Y1}iKv+G()
dS)}
Thus the assumption holds for all i. The assertion
follows because accumulated demand must eventually + constant.
exceed any bound as X > 0. This verifies the modified cost coefficients. Since the
Note that all optimal policies realize long-run bal- system is in long-run balance, and will remain so
ance if the Assumption holds. under any optimal policy by Theorem 1, X,, =
XiDwI-i\i+ < X jA'liMj+lxA i Xl;, for all
The expected time periods until accumulated de-
mand exceeds Maxi Xi, is greater than Maxi Xi,/ X, i, t and k E P(i). Because of Lemma 1, (3) thus may
but for large values of Maxi Xi, this is a fairly accurate be replacedby X,, Y < Xj',,. Problem P is restated
estimate. In a well run system, one expects Xi, to equal as a series system of the assertedkind.
the expected demand over the total lead-time, MA, The constant (4) is unaffected by the series re-
plus some safety stock. One should, of course, expect formulation, and therefore, is calculated in terms of
such a system to be in long-run balance but even if it the original cost coefficients.
is not, then it can hardly take more than 3(MN + 1)
time periods, say, until it is. Still, the bound of the Corollary 2. Under the assumptions of Theorem 2,
theorem is presumably not very sharp. there exist numbers Si, i = 1, 2, . .. N, such that the
The modest requirements on a policy to lead the following order-up-topolicy is optimalfor all i and t
system into long-run balance can be relaxed further Y*= Min(S, X'? ) if Xi, Si
as noted in the following corollary. Its proof is con-
tained basically in the proof of Theorem 1 and is
Y*l = Xi, if Xi,I>.s S.
thereforeomitted.
The numbers,S,, are convenientlyobtained by solving
Corollary 1. Theorem 1 still holds if Mink;>i
X2'k1 is the series system of Theorem 2 by Clark and Scarf's
replacedby X , in the inequality of Lemma 1. (1960) procedure.
570 / ROSLING
Proof. Federgruen and Zipkin developed Clark and
Scarf's procedure for a two-level series system to the
stationary infinite horizon case for h,, h2, p > 0 and
0 < a < 1. The optimal policy is of the stated kind.
Their proof generalizes directly to the N-level case
provided that our Assumption holds.
In the present context, Clark and Scarf's procedure
for obtaining Si, i = 1, 2, ... N, may be defined as
Gi(Si) = Min Gi(Y) (8)
y
where
GI(Y) = h Y+ (p + HI) f (Q- Y)0i'`(t) dt (9)
and
Gi(Y)= hiY
+ 'i-l'r
Gi' , (Min[Si-1, Y- p)kL() dt
for i > 1. (1O)
k'(J) and 0 +1(-) are the convolutions of 0(-) over
the Li and Li + 1 periods, respectively.
G,\(S, ) plus the constant (4) is the minimal average
cost per period for the assembly system when a = 1.
The numbers, Si, obtained from (8)-(10) do not
necessarily satisfy Si+, > Si for all i, but since Yi, <
Si+,, there is no loss of optimality in modify-
ing the numbers by letting Si -- Min[Si+,, S] for
i = N- 1, N- 2, ..., 1. Thereafter, SI+ S, > 0
for all i.
The results of Theorem 2 and its corollary are
criticallydependent on the assumption that the initial
inventory levels are in long-run balance. If item i - 1
is seriously overstocked but item i is not, then it is
generally advantageous to have Yi, > Si, and so, the
order-up-to policy will not be optimal. The compli-
cated, generally optimal policy was derived by
Schmidt and Nahmias (1985) for a two-level system
with three items (summarized in their Figure 3). This
policy seems to force the system more rapidly into
balance than the order-up-to policy of Corollary 2,
but the actual difference in costs is not known. There-
fore, it might be that the order-up-to policy is an
excellent approximation when the system is out of
balance.To ascertainfeasibility,the order-up-topolicy
is modified so that
Yi, = Min(Si, Min
XmI-A') if Xi, S Si (11)
and
Yil = Xi, for X, I>, S,. (12)
By this modification, the approximatepolicy does not
violate the inequalities for an optimal one given by
Lemmas 1 and 2, and it will eventually coincide with
an optimal policy.
With arguments similar to those in the proof of
Theorem 1, it is straightforwardto demonstrate that
if Xi, < Si for all i (here Si, Si,+,),then the approximate
policy will lead the system into long-run balance
within MA,+ 1 periods. Moreover, in the next period
accumulated demand exceeds Max1(X, - Si), and we
have X1,? Si for all i. Thus, unless there is very serious
overstocking, it will not take very long until the ap-
proximate policy becomes optimal.
When the equivalent series system is constructed,
there is a useful observation to be made. It is stated in
the following corollary.
Corollary 3. If Li = 0 for some i > 2 then there are
optimal order-up-tolevels such that Si = Si-,. Items i
and i - 1 thereforemay be aggregated when the order-
up-to levels are computed. The lead-time of the aggre-
gated item is then Li-, and the holding cost coefficient
is hi-, + hia-'-i'. Here it is assumed that hi and hi-
are modified according to Theorem 2.
Proof. The order-up-to levels are independent of the
initial stock levels, and hence, we may assume that
Xl,; ? SI, for all k. So, X,, < S, for all k and t by
Corollary 2. Note that Li = 0 implies that X' = Yj,.
Thus, Y1-,,,= Min(S_, YJ,)but then, by the same
arguments used in the proof of Lemma 1, it cannot
be optimal to have Yi,> Si-,. So, S, = Si-, and Yi,=
Yi1 , for all t. Since i > 2, the cost (2) for the series
system, related to Yi,is
atl-I * a'i * hj* Yj,-='-'- hj* Yj_1, since Li=0.
The total holding costs related to Yi1, is then
a-(hi, * Yi 1, + aLi- * hi- Yi-i,)
= a -1I al'j-l . * (hi-, + hi*
Since Yi,= Yi,, the two sets of constraints (3)
Yi_,^ = Y,, and Y,, < X'+,I
can be replacedby Y,1 ' Xf,?,. Thus, the conclusion
follows.
Once, again, constant (4) is unaffected. The proposed
aggregation is not literally possible to implement in
decision making until both Xi, < Si-, and Xi,., ? Si-,.
 Optimal InventoryPolicies for Assembly Systems / 571
Furtherdiscussion on the important notion of long- The requirement( 14) needs a comment for the case
run balance is required. The Definition is, unfortu- where i $ s(i + 1) but Mi MAl(i?,) (thus, Li = 0);
nately, somewhat confusing, especially when applied (14) should be checked for g = Mi- 1 and Mi but
to the initial inventory position. Given Xi, it is appar- by (13) it holds for Mi - 1. For ,t = Mi, (14) reads
ently of no consequence when this quantity was or- Y,, < Xi. ,, and so, whether Xi, < XL+, is to be
dered. X$-/ simply does not appear in problem P for checked. The conditions of the corollary are satisfied
g< M,(j). One can thus expect that long-run balance for the rather interesting case where all initial instal-
be defined with reference only to X,'I- for , lation stocks on hand and on order are zero except
This is indeed the case and the definition also can be possibly for the end item.
simplified as noted in the following corollary. We close this section by applying Theorem 2 and
Corollary 3 to the assembly system of Figure 1, as
Corollary 4. Theorems 1 and 2 still hold if the defi- shown in Figure 2. The conditions for initial long-run
nition of long-run balance is changed so that X',- for balance according to Corollary 4 are X6, 7
,u M.(i) - 1 are replaced by artificial constructs XSX6 Xe t;X4, 'I, X24 X X21I, X34 s31 X3 I
defined recursively for i = 1, 2, ., N by X41 XX and X2 , X31. Note that items 4
X5;
. and 5 may interchange indices. Hence, the require-
Xj,'- X '-,,uI - , 1, ..,M.(j)1. (13)
ments may be relaxed further.
Requirement(5) can then be simplified so that
X I8, < X I8,
itl i+ 1.1
4. A MORE GENERALASSUMPTIONON
only for ,t = MA-1,
.- . ., M,(j+) (14) COSTS
and only fori=N- 1,N-2, ...,2suchthati # At the end of this section (Theorem 4), it is argued
s(i + 1). that the following Generalized Assumption covers all
cases of practicalinterest for a long-run analysis.
Proof. We first show that the new constructs are
feasible in P, following (3) and the definition of Generalized Assumption (GA)
X,"'- so that Xj,'-y : X,'-y+' and X',ju < XAI'-u for (i) hi a "Ju) +
.
E hkaj-l'(k))> 0 for all i
k E P(i). The first inequality follows from the fact (kc)(i)

that it holds for i = 1, and if it holds for i, then it

(i) hj
hi a
oe-Zu) < p +H,.
holds for i + 1 since if i = s(i + 1) then

Basically,the sum of the echelon holding costs of item

- Xil=
X;2,"i+l
_A Aj+, -
_
X"-Ij+,
by (13) and if i $ s(i + 1), then by (14) X"27',
(i + 1 ) >: X ,,- s (i
X"y'1 (i + 1)+' by (13). The conclusion follows since
X/'4'i- XY'- by (13) for ,t < Mx(j+,).The second
inequality follows because i < k and by (13) and (14),
X"'-,u<, XjM7$ for =O, 1,. .M.,M- 1. Thus, the new
definition implies the old one, and so, Theorem 2 still
holds. To demonstrate that Theorem 1 holds, it suf-
fices to show that the old definition implies the new
one. This is so because (14) follows directly if MA,(i)
> M4(j+,), then (14) still follows
MI,(j+,), and if MAi~,)
because by (13), X',- does not increase.
i and all its predecessors are positive. Thus, it is
allowed that hi < 0 for some i. However, hi > 0 for
purchaseditems (B(i) = 0) and p > 0 is.still implied.
XJ1
+ 1 ) Al"- (i +
01) Negative echelon holding costs are interestingfrom
a practical point of view. There exist numerous real-
life production systems where the sum of the instal-
lation holding cost coefficients for the components is
larger than the same coefficient for their assembly.
Consider, for example, meat or rubber before and
after cooking or vulcanization. The components here
have the more distinctivecharacterof perishablegoods
than their assemblies, and so, they may require more
expensive storage facilities. As a consequence, the
echelon holding cost coefficient for the assembliesmay

L7-1 L6=1 L4=2 L3=2 L2=1 L.-=I

outs ide custo
supply 7 6 4(+5 2 D >deman

Figure 2. Equivalent series system of Figure 1.

572 / ROSLING
be negative. Below (Theorem 3) we present a proce-
dure that eliminates all items for which hi ? 0 by
aggregatingthem into other items.
Lemma 3. Suppose that the GeneralizedAssumption
holds. If hi - Ofor some i, then there exists an optimal
policy for which
Y* = Min X', for all t.
AGP(i)
A formal proof is found in Appendix B. When hi < 0
there is no cost of holding stocks, and so, as much as
possible is acquired of item i.
For an important special case, the minimum of
Lemma 3 is eventually attained by n E P(i) where n
is the immediate predecessor of i's with the smallest
lead-time. This is a crucial simplification and the key
to our procedure. Its demonstration requires the fol-
lowing Lemma.
Lemma 4. Lemma 2 holds under the Generalized
Assumption.
The proof of Lemma 4 is found in Appendix B.
Lemma 5. Suppose hi < 0 and h, > Ofor all k E P(i).
If the GeneralizedAssumption holds, then there exists
an optimal policy for which eventually
Iv t (,,
for ,u=0, 1, ...,M,k, k,qeP(i),
x$j,'-'J= Xi,7 for , = 0, 1, ..., M,
m,; = i,n.
n is such that Min
k EP(i)
Thus, specifically, Y* -
XI.
Proof. Since the proof is similar to earlier ones, we
just give an outline. The availability of item k E P(i)
affects the availability of other items only through the
availability of item i. Attention therefore can be re-
strictedto the subsystem consisting of i U P(i). It then
follows by arguments similar to those of the proof of
Lemma 1 that any optimal policy for the assembly
system is such that for all k E P(i) and all t
=X, - if X;,1 a,;,; ak, = Min Xl,-k
Yk*,
/eIN(i)
1>1;
X,;,< Y, ak, ifX,;,< ak,.
Thus, as soon as there is some production of item k,
then Y*, s ak, thereafter, and some periods later it
must be that X"', < XZ'- for all At and I > k,
I E P(i). Since X > 0, accumulated demand will
eventually exceed any bound, and so, by Lemma 4,
there will be production of all items. Thus, the first
set of inequalities will eventually be satisfied. The last
set of equalities then follows from Lemma 3.
Theorem 3 (Elimination of i for which hi - 0).
Consider the following aggregation procedure: Pick i
for which hi < 0 and h, > Ofor all k E P(i); n E P(i)
is such that l,, = Mink(,;,) lk. Item i is aggregated into
item n (i.e., eliminated) and thefollowing data modi-
fications are carried out.
--hi,* a- + hi, 1,, --1,, + 1,, s(n) -- s(i)
h,l
and P(n) {- JP(i) - n} U P(n)
h, hk -*a/, and 1,.l-, - l, for k E P(i), k# n.
Theprocedureends when no such i can befound. Then
D = {i/i has been eliminated} and C = {i/i remains}.
If the GeneralizedAssumption holds then
(i) repeated application of the aggregationprocedure
results in a smaller assembly system that satisfies
the Assumption;
(ii) if Si, i E C, are the reorderlevels that eventually
become optimal for the aggregated system (by
Corollary2) then thefollowing policy for the orig-
inal system eventuallyalso becomes optimal
Yi,= Min(Si,X2,j') ifX,, S, | for i E C
k<q
Yil= Xi, ifXt>S, J k=MinjlEC/l>i}
Yi,= X">'i for i E D, n is the item into which i was
aggregated.
Proof. (i) If item i is aggregatedinto item n, then i of
the Generalized Assumption apparentlystill holds for
the aggregated system (with item i excluded). The
procedurecan be repeateduntil hi > 0 for all (remain-
ing) i, or it is impossible to find i with hi < 0 such that
h, > Ofor all k E P(i). In the first case, we are done
and in the second case, there must exist i (in the
aggregated system) for which hi < 0 and P(i) = 0.
Pick such an i. For the original system the GA implies
that hi > 0 if P(i) = 0, and so, some items must be
aggregated into i. Since P(i) = 0 after aggregation,
these items must have been aggregatedwith all their
predecessors. Thus, there exists k such that after re-
peated application of the cost modification rule, the
modified cost coefficient of i is (expressed in the
original coefficients)
hk + a -( > 0
E hj v'W_)-M(k)) by GA(i).
je13(k)

This is a contradiction, and so, the assertion follows.
(ii) If i and n are picked according to the procedure,
then there is eventually an optimal solution such that
Y* = XI, by Lemma 5. Thus, aggregationof i into n
becomes optimal and we only demonstrate that the
data modifications do not interfere with optimality.
By aggregation
/+/,-I
yi.1+1'
='
Yli
- and we, therefore,propose an approximatepolicy that
. =l
and so, the parts of (2) concerning i and n (neglecting
constants) are
a a+/n-1 /*
hi - Y,,, +
- >
p=1
at-l a'I'+l(hi+ h,, a-1j)Y,1,
This verifies the modified cost coefficient for n and
the modified lead-time partly. The lead-time also ap-
pears,however, in (3) for s(i), but XI, of the aggregated
system is identical to X', in the original one because
1,, -- 1i + 1,,. Since s(n) -- s(i), the constraint (3) is left
intact. The modification of h, follows from the mod-
ification of Ik because a/k * hk - Y,, = a'k'1, * hk -
a'- * YA,,.The modification of l,k affects the feasible
set through (3). In terms of the original variables,
Y,l, X',$-U for,l = 1,;-1,, replaces Y,, - X', in (3).
Since Yi,, = - s, the new inequality
implies the old one, but by the first part of Lemma 5,
this does not interferewith optimality in the long run.
Thus, the inclusion of {P(i) - n} in P(n) is verified
and the data modifications do not interferewith long-
run optimality. Repeat this argument for every item
that is aggregatedinto another one. From the above
k7=l
(D7~ ~
L7=1
7u A
X______(i)
o 2f
Figure3. Applicationof Theorem3 to the assemblysystemof Figure1 (h, < 0 and h4
Optimal InventoryPolicies for Assembly Systems / 573
result i, Theorem 1 and Corollary2, the assertionthen
follows.
In Figure 3, Theorem 4 is applied to the assembly
system of Figure 1. The policy of Theorem 4 (ii) is
optimal from the first period if the inventory position
of the aggregatedsystem is initially in long-runbalance
and if, moreover, for i E D, X-,'y = X$"-ufor y =
0, 1, . . ., M- 1. The latter condition is quite strict
is applicable whether or not these requirements are
fulfilled.
For those items that remain after aggregation, we
suggest the policy (1 )-(12) of the last section where
the minimum is taken over the items of the original
> *h'
'- i* Y,
system. For items that have been aggregatedaway and
for which hi - 0, we suggest the policy of Lemma 3.
We are then left with those items that have been
aggregated away and for which originally hi > 0.
Apparently, some other items must have been aggre-
gated into item i before this item was aggregatedaway
(since then hi < 0). Consider these other items. Since
aggregation always is into a predecessor they must
form a subtree of the assembly network. Let I be the
root of the subtree.Aggregation(in this subtree)began
with I and hence h, < 0. We suggest the policy
Yi,= Min X"'-"'i (15)
where Q = Ik > i /k E B(l)}. This policy is an attempt
to ascertain as large a production as possible of item I
and simultaneously to take into account that pro-
duction of item i in excess of X'-"i, k > i, k E B(l),
is useless for this purpose. We expect the suggested
policy eventually to coincide with an optimal one,
although we have no proof to present.
The question remains whether the Generalized
Assumption covers all cases of practical interest for a
long-run analysis.
=6=1
Z6-6
The aggregated
34+ 1 >system
2=
5i} =4
L6=1 L5=4 L2=1 L1=1
The equivalent
3+4+6
1m
_____>
> series system
acK2+ h3 < 0).
574 / ROSLING
Theorem 4 (Practical Necessity of the Generalized
Assumption).
(i) If hi * a -8s) + ZkEB(i) h,; a (k)Ofor some i,
<
then the minimal cost ofproblem P is unbounded
below.
(ii) If i of the Generalized Assumption holds
with semistrict inequality for all items and
a ('m+ XkEB(i) h,k a AI"(k) = 0 for some i
hi* a-
then there is an optimal policy for which X', is
unboundedabovefor all t > MA - M.(i)+ 1.
(iii) If i of the Generalized Assumption holds but ii
does not, then there exists an optimal policy for
which accumulated production is bounded by
Maxk,X,, -Xi, for all i.
The proof of Theorem 4 is in Appendix B.
Case i of the theorem can, of course, only appear in
practice as a consequence of incorrectly specified
model data. When these errorshave been rectified we
are left with case ii. Here item i is apparently a free
good, and so, as much as necessary is acquired of i
and its predecessors. It might take some time before
this quantity arrivesbut thereafteritem i will never be
in short supply, and so, item i and all its predecessors
can be neglected in a long-run analysis. We are left
with case iii of the theorem. Here, the system is not
worth operatingin the long run, and so, this case lacks
interest (although iii does not formally state that pro-
duction eventually ceases, this is an obvious conjec-
ture). Only systems for which the Generalized
Assumption holds are left. Thus, all cases of possible
interest for a long-run analysis are covered.
5. EXTENSIONS
Most real-life assembly systems are disturbed in one
way or the other by stochastic yields and lead-times.
Long-runbalance is then attained at most temporarily
and the series analogy does not hold. Still it might be
reasonableto model such systems as if yields and lead-
times were deterministic and apply the series interpre-
tation to calculate reasonable order-up-to levels. An
approximate policy then can be devised, as proposed
in Section 3 by (1 )-(12) and in Section 4, Theorem
3 for cases that do not satisfy the long-run conditions
initially.
Frequently, real-lifedemands are modeled with sea-
sonal variations. The series analogy directly carries
over to such nonstationary demand distributions but
then the order-up-tolevels Si are replaced by Si,. The
series interpretation generally cannot be expected to
carry over when holding costs or production costs are
nonstationary. If it is temporarily advantageous to
produce item i, then it might be optimal to raise its
inventory level above X1+f,, and so, the series analogy
breaks down.
It is generallynot possible to apply the seriesanalogy
to systems with setup costs in production. This is
obvious when the setup cost of item i is much larger
than for i + 1. Item i is generally produced in much
largerbatches than i + 1, and so, just after production
of item i, one generally expects that Y* > X'?,,
contradictingthe series analogy. However, there may
be special cases where the series analogy may work as
a good approximation. We have in mind cases where
lot sizes increase with i (i.e., with total lead-time) in
such a way that the expected reorderinterval for item
i + 1 is an integer multiple of the same interval for
item i (e.g., a power of two-policy). If item i is about
to be produced but item i + 1 is not, then X', is
generally much larger than Xi,, and so, even after
production, Y* < XJ?, as required. If, on the other
hand, i + 1 is about to be produced, then item i is
soon to be producedand we expect that X' ,, is greater
but rather close to Xi,. If production of item i always
is postponed until the production of i + 1 arrives in
Li+, periods, then the seriesanalogy will hold. Presum-
ably this postponement generally incurs only a minor
deviation from optimal costs. However, if demand
recently has been high, there can be a profit from
producing a lot of item i immediately, just to make
useful the excess few units of item i + 1, X?+,I-X,I
as soon as possible. Thus, we make an approximation
in addition to the strong assumption about lot sizes
being related to total lead-times. These assumptions
are similar to those used by de Bodt and Graves when
attempting to extend their series model to a two-level
assembly system.
APPENDIXA: VARIATIONSOF THE COST
ASSUMPTIONS
1. Sales Prices and Proportional Costs of
Production
7r= unit sales price of the end product, paid at deliv-
ery;
Cj= unit cost of assembly (or purchase) for item i
incurred after assembly (at delivery).
ci is the additional cost of production of item i so that
the cost of its predecessor components are not
 Optimal InventoryPolicies for Assembly Systems / 575

included. Maximizing net profit, the analogy of (1) is 2. Reduced HoldingCosts DuringAssembly
Q,j= reduction of the echelon holding cost per period
E{ Ea`(r -a"X + I d~
YX1 - tI() of item j during assembly of item i jj, A, hi
if i<jand Al, - 0 if i =j.

f
- (- Y, )b`() d] It seems fairly natural that holding costs are different
when items are on the assembly line and in the storage
A'
room. Not only are reductions permittedby the above
- E Ci a'i(Yi, -Xi,) definition but as j = i is allowed, the possibility of
i=1 holding costs incurred from the moment of ordering
is also accounted for.
The reduced echelon holding costs during assembly
are handled indirectly by a reduction in the produc-
tion costs. The present value (at delivery) of the cost
reduction during assembly of one unit of item i is

E A,j(a' + a2 + *-- + a1i)*

The expected revenues of period t are due to the jE=_B(i)uji)

expected demand of period t plus expected backorders

carried over from period t - 1 minus expected back- As demonstrated in the previous section, a propor-
orders left at-~~~~~
ia of
the end e(i,-li
period . I])A[i+
o t. ] tional cost of production, c,, may be eliminated if
Using the relation E(X,,) = Y,_- X and adjusting (1 - a) - ci is added to the echelon holding cost
1]))
for constant terms associated with A, (16) can be coefficient. Thus
rewrittenas
hi <hi-hi A-,(a-li1).

-Ea1I'* [h, +(l 1-o)c1J

With this modification of hi, the model is equivalent
to problem P. Note that when a = 1 there is, in fact,
no modification (except for the constant of (17)).

+[p+(l -at)r+HJ]*o/
3. A Fill-RateType Shortage Cost
b = unit cost of backlogging [$/unit].
*f (H - Y - ' 0")dd}
Y'
While p with dimension [$/unit x time] is naturally
N associated with stockout constraints of the type aver-
+ constant. (17) age outstanding backordersper period, b is naturally
(17) is apparently equivalent to (1). To see what cost associated with a fill rate type constraint, average
modifications are actually carried out, write the back- number of demands per period not satisfied directly
logging coefficient as from stock on hand.
i-l Since stock on hand first is used to satisfy old
backlogs, the discounted expected cost for new back-
i=l ~~~i=l logs in period t + 1 is

Recalling that Hl = X,= hi, the modifications are

a1* b * a(f (Q- Yj,O+'+() dt
E a
+ - a)(ir - - a)c
Q - Y, O'(Q)d#) (19)
and (18)
hi +
(1I-O-)ca. where 0'(.) is the convolution of O(*)over 1, periods.
hj
When adding (19) to the other costs of (1), a modi-
Thus, revenues as well as production costs may be fication of ii of the Generalized Assumption is neces-
neglected and problem P applied. sary for our results to carry over. To see the nature of
576 / ROSLING

the modification rewriteii of the GeneralizedAssump- holding costs of (2) are lower for Y than for Y*.
tion as Consider the backlogging costs of (2). They are the
N
hi(a - 1) p same for Y as for Y* if Y, = Y*, where r = t + M, -
E 1-a < i- (20) M,. This will be demonstrated. Pick j, k, n E A(i)
such that j E P(l), j = s(k) and k = s(n). Let u =
The left-hand side is the present value (at the delivery r-4 and w = u-1,;. Then by (24)
of the end product) of all holding costs for the com-
ponents incurred during production. The right-hand Xj .-pj,. = YJ*,-yj,,
side is the present value of the unit cost of backlogging = Max(O,YJ,,-X/_) - Max(O,Xj, -X, i).
over the entire infinite horizon. To take b into
account, (20) is modified so that Similarly, X$'-' -1 = Y* y_
Max(O,Xkl - X;,7') and so,
A
hi(a-Al,') - 1) p (21)
/

i=l 1 -ea 1 a
- X/

As long as p > 0 or a < 1, ii should thus be modified = iMlax(O,

Xj.-X j,. + Max(O,X,, -X*?, ))
to = Max(O,XJ,-X,," 1) since X>j,
S X,,Al by (3).
N

ihi a-'u) < p + H, + (1 - a)b. (22) Extending this analysis to all items of A(i) and when
i is included (note that Yj,= Min(Y*, YF,))we get
However, if p = 0 and a = 1 then (22) is not correct. X',-X = Max(O,X',-I
The limit is taken directly from (21) so that
N
Pick m E A(b) such that m E P(1) and consider the
hi M(i < b. (23) difference
AT -' Xj, = (X/1?/ NJ,)+ (Xjr. -JAr)
Replacement of ii of the GeneralizedAssumption with
(22) and (23) is sufficient for all our results to carry =(X/,), - Xj.) + Max(O,X, - X',jIxI). (25)
over except for the order-up-to policy of Corollary 1
Suppose that XII - XI,?,.Since YV,3 XAf,I,j XAl-A,
and Theorem 3. Since (19) is not convex in Y,,there
Xh>" and so
is no guarantee that (1) will remain unimodal in Yi,
X/ '-
after the addition of (19). Thus, when X'+,, < Si it Xi/' -,lZrj bl 00

might be that an optimal policy requires Y* < X,' by definition and (3).
because of a local minimum. Neglecting this anomaly
is, presumably, an acceptable approximation. Thus, XX,,-Xj =X, -Xj. so that Xj, = Xj, and
= Y*, by (24). Suppose that XI' > X/?. If
= XX',, then Y,, = Y* again. Suppose that X, >
X',
APPENDIXB: OMITTEDPROOFS
X' Then by (25), necessarily, X', > X,". But
Proof of Lemma 1 then by (25) (note that X" ',- X'3,- by definition
and (3))
If an optimal policy, Y* (with states X), violates
the assertion, then for some i and t (and some out- X I - X + X - /
AT =<0X/ -All

comes of demands) Y* > Xi, and Y* > XZ,'-M'i A 1)71 Jl 1)1/ Jl Jl 171 1)? 1

Y 6 ,? x)l ={AI1 T7,A[,. X < 0

Min,k>i X,-'i. Then define an alternative policy
(with states X) which deviates from Y* by
Fit = Max(X,,, X,,'-xli) and
Y,(k), = Min(Y(/), Ak'q)
for all k EI {i}IU A(i), q =t + Mi- I,(k-,. (24)
Policy Y postpones production of item i in period t,
one period ahead. If required for feasibility (by (3)),
some production is also postponed a period for item
k E A(i) in period t + M, - M,;. Note that Y satisfies
the assertion for i and that it is feasible for P. Since
YF,< Yi*, Y,V,- Y*, and hk > 0 for all k and s, the
by assumption.
Thus, XI, < X/ and since Y*, - XI, by (3), Y,, =
Y,*by (24). Hence, the backloggingcosts are the same
for Yand Y* and so, Yis superiorto Y*, contradicting
its optimality and the conclusion follows.
Proof of Lemma 2
The Assumption implies the GeneralizedAssumption,
and so, Lemma 2 follows directly from Lemma 4.
The proof of Lemma 4 requires the following
proposition.
 Optimal InventoryPolicies for Assembly Systems / 577
Proposition and the new policy Yi,(with states XN,)for i < n by
Let I C {1, 2, ..., N} and A = UiE, A(i). If the Yi, = YF, for all t if i i A or
GeneralizedAssumption holds then
if iEA fort<q+Mn-M,
i hia",u) < p + H,.
ic,I Yi,= Yi, + e foriEA andt=q+M,7-M,

Proof. Let J = {i ? A/s(i) E AI and B = Ui-, {B(i) U Yj, = Max(Y1,, Xj,)

fill. A U B = {1, 2, .. ., NI and GA(ii) may be written foriEA andt>q+M,,-M,. (27)
as
The new policy is feasible because n > i implies that
> hiao1 )+ > hia-','i) < p + H,. M,,> M, by definition. It is assumed that E is small
i.1 iElB -
enough so that if i M A, then Yj X '-i for t =
The second term on the left-hand side is q + M,,- M1.The new policy supplies c more units of
the end item in period q + M,/. Before computing the
, {hi * a-'"(i + >
kEBB(i)
h/; a*1Ix(Ak} < 0 by GA(i) cost implications, note that by (3) i M A implies that
ici
s(i) 4 A since
So, the conclusion follows.

Lemma 4. If the GeneralizedAssumption holds, then

for any optimal policy v-1

=Av l,v < iv =- h, r +

Y* >,-Min(O MiXkm for all i and t.
kv>i
Such a policy satisfies (on a first-come first-served
basis) all customer demands within at most MN + 1
periods of time.
Proof. The first assertion is demonstrated by an
inductive argument. Suppose a feasible policy (inter-
preted as a function of demand history only) has been
fixed for all i > n and, given this restriction, optimal
values, Y,,, have been computed for all i < n - 1 and
t. The inductive hypothesis says there is a solution for
which (26) holds for all i - n - 1 regardless of what
,,1
(26)
where I = s(i), r= q + M,, - Ml and v = q + Mn-
Mi. Consequently, i 4 A implies that 1 4 A for all
/ E B(i) and A is similar to the set appearing in the
Proposition. The difference of e additional units in
stock for the new policy will prevail until some pro-
duction is scheduled by the old one by (27). For some
realization of demands, suppose 6 < Eunits of the end
item are producedby Yin period q + M,, + T(possibly,
T = oo). The total gain in costs of backloggingis then
23
a oea'-I
a*o. (p + HI)
/qt+AI,n-/1

feasible policies were fixed for i > n. This (partial) -

= (p + HI) acI+' (1 -o )/(I - a).
optimization is repeated and this time it also includes
i = n and feasible policies are fixed only for i 2 n + For this production to be feasible there must be pro-
1. By the inductive hypothesis, there is a solution for duction of item i E A in period q + M,, + T- Mi or
which (26) holds for all i < n - 1 and t. By a earlier.Thus, an upper bound on the additional hold-
contradictive argument, (26) will also hold for i = n. ing cost is
So, given some fixed policies for i > n, pick an optimal v+T- I
solution Yj, (with states Xi,), i - n, for which (26) 6. >fi fi a'-' a*'i hi
holds for i < n and where for at least some t, say
t = q, = . a+l.' hi .
a iT)/(I(1 - - a)

Ynql< Min (O, X '-).

k>,,
where v = q + M,, - Mi. Thus, there is a net gain for
Note that Y<q<0 so that there are outstanding back- M,(i)= Mi - 1iand by the Proposition
orders for the end item. Define a new policy Yi, for all
Ehia-'"ZSu)< p + H
i - n and t. To do this note that by the inductive
hypothesis for i < n, Y1,>- XAI,i for t = q + Mn- M
Thus, the new policy gives a net gain for all realizations
because X>jl' < 0 by definition. Then define
Yn,,, of demand and the optimality of the original one is
A n= {n U {i< n/Yl, =X$ ,'-fort= q+ M,, - M1 contradicted. A similar but much simpler argument
578 / ROSLING
shows that the inductive hypothesis holds for n = 1. The increase in penalty costs over the same
Thus, the first assertion is demonstrated. The second period is
one is a direct consequence.
b(p + H,)/(1 - a).
Proof of Theorem 4
By assumption, E',, a(-"l(i) > p + H, and so,
(i) Consider the subsystem {i} U B(i) and suppose the new policy is as good as the old one,
that a finite optimal policy exists. Then consider verifying the proposition for i = 1. Consider
an alternative policy that orders one more unit next all i E P(1). If hi > 0, the conclusion
in period 1 of the component, m, with the directly carries over to item i, but if hi < 0,
greatest total lead-time M11, in period M,,, - then there is an independent motive for build-
Ml + 1 of the component, 1, with the next to ing up stocks of item i. Considerthe subsystem
the greatest total lead-time and so on, so that {i} U B(i) and apply the same argument as for
one more unit is ordered of item i in period the final item. The same conclusion follows
M,- Mi + 1. This additional unit is held in because GA(i) holds for i. These arguments
stock forever. The additional cost of this policy carry over to all i.
is for the first M,,, - Mi periods (capitalized to
period M,,1-M1 + 1)

E h a 'k(1 + a + + aAlk-AXl-) a-Wk-_,d ACKNOWLEDGMENT

IkEB(i)
The advice of an anonymous referee considerably
or equivalently improved this paper. The research was supported
by the Swedish National Board for Technical
h, * a1k(a-U"I,k-`i) - 1)/(1 - a).
k G B (i) Development.
The cost of holding item i forever from period
M.- M1+ 1 on is
REFERENCES
h aia+
hi fi hk g a'k/ (1 -a). AXSATER,S., AND P. LUNDELL.1984. In Process-Safety
k EB(i) J Stocks. In Proceedingsof 23rd Conferenceon Deci-
sion and Control,pp. 839-842. Las Vegas.
So, the total cost increase is BESSLER, S. A., AND A. F. VEINOTT.1966. Optimal Policy
for a Dynamic Multiechelon Inventory Model.
alj'(h,+ ,kEBsi) hk,; a- '1s(k;)- l'.si))) Naval. Res. Logist. Quart.13, 355-389.
1 -a <O CARLSON,R. C., AND C. A. YANO. 1986. Safety Stocks
in MRP-Systems with Emergency Setups for Com-
by assumption. ponents. Mgmt. Sci. 32, 403-412.
CLARK,A. J., AND H. SCARF.1960. Optimal Policies for
Thus, the conclusion follows. a Multiechelon Inventory Problem. Mgmt. Sci. 6,
(ii) As above, except that the final inequality is an 475-490.
equality. CLARK, A. J., AND H. SCARF. 1962. Approximate Solu-
(iii) Consider the final item and suppose that in tions to a Simple Multi-Echelon Inventory Problem.
period t, Y*, > X, and accumulated produc- In Studies in AppliedProbabilityand Management
tion exceeds Max,<X,, - X,. Consequently, Science, pp. 88-110. K. J. Arrow, S. Karlin and
accumulated production of item i exceeds H. Scarf (eds.). Stanford University Press, Stanford,
Max,k XI, - X,, in period t - (Ml - M) = t Calif
or earlier.Consider an alternativepolicy Ysuch DE BODT,M. A., AND S. C. GRAVES. 1985. Continuous-
that Yj, = Y-*- 6 for all t > (t, or the first Review Policies for a Multi-Echelon Inventory
Problem with Stochastic Demand. Mgmt. Sci. 31,
earlier period with production > 0) where 6 >
1286-1299.
Ois arbitrarilysmall. The cost saving of having FEDERGRUEN, A., AND P. ZIPKIN. 1984. Computational
6 units less in stock over the entire horizon is Issues in an Infinite-Horizon, Multi-Echelon Inven-
at least (capitalized to t, + L,) tory Model. Opns.Res. 32, 818-836.
A' FUKUDA, Y. 1961. Optimal Disposal Policies. NavalRes.
8 Ii,
hi ",w( - a). Logist. Quart.8, 221-227.
HOCHSTADTER, D. 1970. An Approximation of the Cost
 Optimal InventoryPolicies for Assembly Systems /
Function for a Multi-Echelon Inventory Model.
Mgmt. Sci. 16, 716-727.
KARMARKAR, U. S. 1981. The Multiperiod,Multiloca-
tion InventoryProblem.Opns.Res. 29, 215-228.
KARMARKAR, U. S. 1987. The Multilocation,Multiper-
iod Inventory Problem: Bounds and Approxima-
tions. Mgmt. Sci. 33, 86-94.
ROSLING, K. 1977.A Generalizationof Clarkand Scarf's
Approachto Multi-EchelonInventoryControl. In
Three Essays on Batch Productionand Optimiza-
579
tion, pp. 20- 111. Linkopings Tryckeri AB.
Linkoping,Sweden.
ROSLING,K. 1989. Bounding Procedurefor Clarkand
Scarf's Serial Multi-Echelon Inventory Model.
Working Paper, Department of Managementand
Economics, Linkoping Institute of Technology.
S-581 83 Linkoping,Sweden.
SCHMIDT,C. P., AND S. NAHMIAS. 1985. OptimalPolicy
for a Two-Stage Assembly System Under Random
Demand. Opns.Res. 33, 1130-1145.