.

MANAGEMENT SCIENCE
V(d. 12, No. 3, November, 1965
Printed in U.S.A.

OPTIMAL POLICY FOR A MULTI-PRODUCT, DYNAMIC,

NONSTATIONARY INVENTORY PROBLEM *t
ARTHUR F. VEINOTT, JR.
Stanford University
This paper is concerned with a multi-product dynamic nonstationary
inventory problem in which the system is reviewed at the beginning of each
of a sequence of periods of equal length. The model has the following features.
There is a general demand process with no stationarity or independence as-
sumptions, partial or complete backlogging of unfilled demand, afixednon-
negative delivery lag (which may be positive only under complete backlog-
ging), a nonstationary linear ordering cost, a nonstationary holding and short-
age cost function, discounting of future costs, and nonstationary restric-
tions like budget and storage limitations. The objective is to choose an order-
ing policy that minimizes the expected discounted costs over an infinite time
horizon. Conditions are given that ensure that the base stock ordering policy
is optimal and that the base stock levels in each period are easy to calculate.

1. Introduction
We consider a dynamic nonstationary multi-product inventory model [3]
in which the demands in each of a sequence of periods of equal length are ran-
dom vectors. The demand vectors are not necessarily independent or identically
distributed. We permit partial or complete backlogging of unfilled demand and
deterioration of stock in storage. There is a fixed nonnegative delivery lag for
orders which is permitted to be positive only when unfilled demand is backlogged
and there is no deterioration of stock in storage. Nonstationary restrictions are
pennitted on the stock on hand and on order after ordering in each period. Budget
and storage limitations are examples of such restrictions. There is a linear order-
ing cost and a holding and penalty cost in each period. These costs may vary
over time. Future costs are discounted. An ordering policy that minimizes the
expected discounted cost over the infinite time hori2:on is sought.
Our purpose is to develop sufficient conditions ensuring that the optimal
policy takes a simple form and that the parameters of that policy are easy to
compute. There are two main results. The first of these may be described as
follows for the case where the demand vectors in successive periods are stochasti-
cally independent and where there is no delivery lag. Suppose for each period i
that we are able to determine a vector of inventories iji, called the hase stock
level, that minimizes the expected cost incurred in period i when considered in
isolation from all other periods. We show that if it is possible to order in periods
1, 2, • • • so that the inventory levels after ordering in those periods are respec-

* Received September 1964.

t This research was supported by the Office of Naval Research under Contract Nonr-
225(77) (NR-347-010) and by the National Science Foundation under Grant GP-1625.
Reproduction in whole or in part is permitted for any purpose of the United States Govem-
ment.
206
 A MTTLTI-PBODtJCT, DYNAMIC, NONSTATIONART INVENTORY PROBLEM 207

tively yi, yi, • • • , then that ordering policy is optimal for the infinite stage
model. This result typically solves the problem where the initial amount of in-
ventory on hand is sufficiently small. If there is a sufficiently large initial amount
of inventory on hand, however, it will not be possible to order so that the in-
ventory level after ordering in period 1 is §i. In this situation one special case
of our second main result states that if all items must be stocked in fixed propor-
tions and if certain other conditions are satisfied, then the optimal policy may be
described as follows. Order so as to attain yt in period t, if this is possible, and
order the smallest feasible amount (consistent with proportional stockage) in
period t otherwise. For the usual single product problems, this jx)licy requires
that if in period i the initial inventory on hand is less than yi, one orders up to
yi ; otherwise one does not order in period i. Earlier results of this type for the
single product problem are given in [lb], [lc], [2], [4], [5], [6], and [8].
We do not use the functional equation approach of dynamic programming in
our proofs. We rely instead on a direct analysis of the underlying stochastic
process (the inventory process).
One limitation of our model is that there is nofixedcharge for placing an order.
However, our model provides for complexities that have not yet been success-
fully dealt with where there is a fixed charge for ordering.
The paper is organized as follows. The problem is formulated in Section 2 for
the case where there is no delivery lag and the demand vectors in successive
periods are stochastically independent. The main theorems are given for this
case in Section 3. In Section 4, we reduce the model with a delivery lag to an
equivalent model with no delivery lag. We give conditions under which it is
possible to "factor" a multiproduct problem into several subproblems in Section
5. Finally, in Section 6 our results are extended to the case of a general demand
process in which independence of demands in different periods is not assumed.
In this case the inventory manager can also infiuence the demand process to
some extent.
We will not attempt to give a comprehensive survey of the applications of
our results here. We reserve this for future publications.
2. Model Formulation
Consider a dynamic inventory problem in which there are n products labeled
1, • • • , n. Assume that there are m classes of demands for these products with
m not necessarily equal to n.
The demands occur during an imbounded sequence of periods labeled 1, 2,
• • • . Suppose the demand D^ in class j in period i is a random variable. Let
Di = (Z),i, • • • , Dim) and denote by #,( •) the joint cumulative distribution of
Da , • • • , Dim .' We assume that Di, Di, • • • are independent' but not neces-
sarily identically distributed random vectors with D, assuming values in a set
' Although, for brevity, we always interpret the D.-y in this paper as demands, our formu-
lation pennits some of the D,/ to represent other chance factors, e.g., prices, costs, de-
terioration rates, etc.
• The case where' the D, are dependent is studied in Section 6.
 208 ABTHTTB F. VEINOTT, JB.

S>i of m-vectors. Note, however, that the random variables Da , • • • , Dim may
be dependent. Usually we shall assume that D,- = Rn*, the set of nonnegative
m-vectors of real numbers.
At the beginning of period i there will be a vector a;,- = (xa , • • • , Xin) of initial
inventori^ of each of the n products. If a;,-, is negative, then there is no stock of
product j on hand, and we take —Xi,- to be an amount of backlo^ed demand for
that product. We assume that the possible valura of Xi are restricted to a specified
set Xi of n-vectors of real numbere. For example, if each of the products is in-
divisible, then Xi could be the set of all n-vectors having integral coordinates.
Or we could let X,- = Rn , the set of all n-vectors of real numbers.
The inventory manager is permitted to place an order for stock at the begin-
ning of each period with immediate delivery. Let y,- = (ya, • • • , j/,n) denote
the vector of inventories of each of the n products on hand after ordering in
period i. Here again a negative value of j/,y signifies a backlog. Thus j/< — a;.- is
the vector of order quantities in period i. yi must be chosen from a specified set
Yi of n-vectors of real numbers and must satisfy yi ^ ?.(a;<) where 3,(xj) =
(qaixi), •• • , qin(xi)) is an n-vector of extended real valued fimctions each de-
fined on X,-.' In inventory theory it has been customary to assume that qi(Xi) =
Xi, but other assumptions are sometimes appropriate. For example, if one re-
quires that at least M{ >0) units of each product be ordered in each period, then
Qiixi) = Xi + Mu where u is an n-coordinate vector of ones. On the other hand
if disposal of up to M(>0) units of each product is permitted, then g,(a;,) =
Xi — Mu.
We assimie that the amount of product j on hand at the end of period i is a
specified fimction s,y(y,-, D,) of the vector j/,- of inventories on hand after ordering
in period i and the vector of demands during period i. Let «,(?/,•, D,) =
(saiyi, Di), • • • , Siniyi, Di)). We assume that the range of s, is Z.^.!. Thus
Siiyi, Di) is an admissible vector of initial inventories with which to b^in period
i -\- 1. In the literature of single product problems it has usually been assumed
that all unsatisfied demand in a period is backlogged or that it is all lost. In our
formulation, if type j demand is for product j (so that n = m), then the back-
logging assumption leads to s,y(2/.-, Di) = y,y — D.y while the lost sales assump-
tion leads to s^iyi, 2),) = max (j/,y — Dij, 0).
At the banning of period i the inventory manager is assumed to have ob-
served the vector

representing the history of the process up to the begiiming of period i. He bases

his ordering decision in period i upon Hi.
An ordering policy is a sequence of vector valued functions F = ( Fi, ?2, • • •)
such that at the beginning of period i after having observed the past history Hi,
the manager orders the vector Yi(Hi) — a;,-. We say that the ordering policy

• If M = (u,) and i>( « (vi) are n-vectors of real numbere, we say u ^ o if u< ^ i;< for
aU t. If in addition u 9^ v, we say u > v.
 A MULTI-PRODUCT, DYNAMIC, NONSTATIONARY INVENTORY PROBLEM 209

is feasible if the rai^e of Yi is F,- and if for each possible history Hi, Yi(Hi) S
qi(xi), i = 1, 2, • • • . Hereafter, an ordering policy will be understood to be
feasible unless otherwise stated.
There are three types of costs incurred in a period: ordering, holding, and
shortage. We assume that the cost of ordering a vector 2/1 — x,- in period i is
Ciiyi — x.) where Cj = (c,i, • • • , c.^) is an n-vector of real numbers. c,(y, — x.)
is, of course, the scalar product 23y-i '^niVi} ~ xa). Note that if qaixi) < x,-,,
then y,j — x,y may be negative. In that event Ci,{yij — x.y) is the rebate re-
ceived from disposal of x,-j — 2/<j units of product j in period i. Assume that the
holding and shortage cost in period i is a specified function gi,(2/<, Z),) of the
vector ^, of inventories on hand after ordering in period i and the vector of de-
mands in period i. Finally we suppose that there is a discount factor «< for period
i, 0 g a,. a< is the value at the beginning of period i of one cost unit (e.g., a
dollar) at the beginning of period t + 1.
We assume that for each i, SD^ is a Borel set and that g,( •, •). s,( •, •))?•(•)
are Borel functions. In addition we require ?,(•) to be a Borel function. Thus
Yi(Hi), gi{yi, Di), Si{yi, Di), and g,(x<) are measurable.
For all y, t, and i, let
Wiiy, t) = c, 2/ + giiy, t) — (XiCi+iSiiy, t ) .
Also let

Liiy) = j giiy, t) d^iit),

and

for all y and i. We assume that these integrals exist and are finite for every y.
Let ft = 1 and fii = 11}-} ay, i > 1. Denote by/^(xi | Y) the expected dis-
counted cost over periods 1,2, • • • , N when Xi is the vector of initial inventories
and the policy Y is followed. Then.
(2.1) /;v(xi I y ) = E{z2f-iPilciiyi — Xi) + giiyi, Di)] — /SJ^+ICW+IXAT+I}.
Note that we are assuming that any stock of type j left over after iV periods can
be discarded with a retum of Cw+i,,-. Similarly any backlogged demand for prod-
uct j remaining after iV^ periods is satisfied at the cost Ciir+i,y.
Using the fact that x<+i = s,(2/<, Di) and regrouping terms in (2.1) we get

Since the term CiXi is not affected by the choice of Y, we find it convenient to
redefine /» by letting
(2.2) /^(xi \t) =E{ Z f - i ft- Wiiyi, Di)}.
Observe that in passing from (2.1) to (2.2) we have in effect reduced a model
with ordering co^ c* and holding and shortage cost function giiyi, D,) to an
210 ARTHtJB F. VEINOTT, JR.

equivalent model with a zero ordering cost and a "composite" holding and
shortage cost function Wiiyi, Di).
It is convenient to express the expected value of the bracketed term in (2.2)
in a simpler form. To accomplish this notice that j/,- depends functionally only
upon Di, Di, • • , Z)<_i which in tum are stochastically independent of D , .
Thus yi and D,- are stochastically independent so that

Hence (2.2) reduces to

(2.3) fAxil Y) = T.
We assume that there exist constants 7,- with Giiy) ^ 7,- for all y and i, and
2 < 00. This assures us that
(2.4) fixi I Y) = T^
exists for all Y, with + » allowed as a value of the sum (2.4). We interpret
fixi I Y) as the expected discounted cost incurred in periods 1, 2, • • • when Xi is
the vector of initial inventories and the policy Y is followed. The problem is now
to find a policy Y* such that
fixi I F*) = min7/(a;i | Y).
If such a policy exists, it is termed optimal.
For the special case in which the cost functions, demand distributions, and
constraints do not change over time, we drop the subscript i on the following
quantities: c,-,{/,(-, •), TF,( •, •), 3«( •).«•» *i( •), 3D<, Z,-, F,-, s,( •, •), Lii • )>
Gii •). We call this the stationary case. If any of these quantities does depend
upon i, the model is called nonstationary.
3. The Fundamental Theorems
Consider the following two hypotheses.

(3a) For each i there exists a vector y< = iya), say, that minimizes Giiy)
over the set F,-.

(3b) qi+iisiiyi ,t)) ^ yi+i for aU < £ ©.• and f = 1, 2, • • • .

The vector §,- may be thought of as an optimal inventory level for a one period
problem in which Giiy) is the expected cost function. Observe that (3a) is
satisfied if, for example, F,- is compact and Giiy) is lower semi-continuous on
Yi. We may drop the requirement tliat Yi is compact if we assume that there is
ay* £ Yi such that {y | Giiy) ^ Giiy*), y £ Yi] is compact.
The hypothesis (3b) ensures that if one orders so as to attain y,- in period i,
then no matter what the demand in period i, it is possible to order so as to attain
in period i + 1. The condition (3b) is satisfied in the stationary case if,
 A MTJLTI-PBODUCT, DYNAMIC, NONSTATIONABT INVENTOBY PBOBLEM 211

for example, q(x) g x, there is one demand class for each product (so n = m),
SD = Rrn^, and unsatisfied demands are lost or backlogged. To see this observe
that since yi = y (say) for all i (we assume that y ^ 0 in the lost sales case),
q(siy,t)) g s(y,t) ^ y
under the stated assumptions. If in this example we permit the demand distribu-
tions and cost fimctions to vary over time and if q(x) ^ x — (yi — yi+i) for all
i and x, then (3b) holds since
q(s(yi ,t)) ^ s(yi, t) - (y,- - ^.+1) g y , + i .

If we assume further that yi ^ g,+i for aU i, then q satisfies the indicated hy-
pothesis if q(x) = X. This assumption about q is of special importance because it
rules out the possibility of disposal of stock. In subsequent papers we shall give
conditions under which yi ^ jfi.fi for the single and multi-product problems. As
one example of these results we remark that if there is single product, we may
ensure that y< g yi+i for all i by requiring that G'i(y) ^ Gi.^.l(y) for all y and i.
Tjrpical conditions ensuring that this inequality holds for the case of a single
demand class are that the demand distributions be stochastically increasing over
time [5], that Wi(y, t) be convex in y for every fixed t and i, that
(d'/dtdy)Wi(y, t) ^0 for aU y, t, i*
and that (d/dy)Wi(y, t) he nonincreasing in i for each fixed y and t. The
last assumption is satisfied if, roughly speaking, the marginal shortage costs are
nondecreasing over time, the marginal holding costs are nonincreasing over time,
and the purchase costs are "slowly decreasing", or "rapidly increasing", or
"slowly decreasing" and then "rapidly increasing" over time.

Theorem 3.1. If (3a) and (3b) hold, and if qi(x*) g yi, then
Yi*(Hi*) =yi (i = 1, 2, • •. )

where H* = (x*, ••• , x*, y*, •• • ,y*-i,Di, •• • , I>._i) is t h e history u p t o

period i when following Y*. Also

Proof: Let Y be any ordering policy and let ^ i , 2/2, • • • denote the associated
amounts of stock on hand after ordering when the demands are Z)i, D j , • • • .
Since qi(x*) ^ yi and (3b) holds, we have y* = yi for all i. Thus by (3a),
Gi(yi) ^ Gi(yi) for all i so that
f(xi*\Y) = j:
which proves that F* is optimal and completes the proof.

* If Wi{y, t) = Wiiy - t), this condition is equivalent to the requirement that Wi{y, t)
be convex in y.
 212 ARTHUR F. VEINOTT, JR.

The theorem tells us that if it is possible to order in each period so that the
inventory levels after ordering in periods 1, 2, • • • are yi, yi, • • • , then that
policy is optimal.
In Theorem 3.1 we do not require F, to be a convex set and —Giiy) to be
unimodal on F.." However, for well known reasons it is usually extremely diffi-
cult to locate yi unless this property obtains. The simplest way of assuring that
—Giiy) is unimodal is to require that Wiiy, t) be convex in 2/ on F, for each
fixed < c SD,. For then Giiy) is convex (and hence —G,- is unimodal) on F<.
In order to characterize the optimal solution to our inventory problem when
qiixi) $ ^1, we must impose additional assumptions. First, however, we require
some definitions. Aset A C Rn'is said to be linearly ordered by the relation g if
for any two elements y and y' ot A, we have y ^ y' OT y' ^ y. As an example,
when n = 2, the graph of any monotonicaUy increasing real valued function of
one real variable is a linearly ordered set in .R2. An element y* of A is said to be
a minimal element ot Aiiy ^ ytor every y E A.It & minimal element exists, it is
unique.
We may now formulate our assumptions.

(3c) For each i, Yi is closed and linearly ordered by g .

(3d) For each i, Giiy) ^^d Siiy, t) are nondecreasing in yioxye F, and < f SD,
such that y'^yi, and ?i(x) is non-decreasing in x for x e Xi for which g,(x) g yi.

The hypothesis that Yi is closed is satisfied if, for example, F,- equals Rn , R^^,
or the set of n-vectors with integral coordinates. The hypothesis that F,- is
linearly ordered is always satisfied if n = 1. It is satisfied for n > 1 if, for ex-
ample, all products must be stocked in fixed proportions. When this is the case
Yi takes the form
Yi = {y\y = kiZ, z e Z,]

where Z,- is a set of real numbers and ki = (fcjj) is a fixed n-vector of positive
real numbers.
The hypothesis (3d) is satisfied, for example, if —Giiy) is unimodal on F , ,
if there is one demand class for each product, if unsatisfied demand is lost or
backlogged, and if g<(x) = x.
It is easy to show that every closed linearly ordered subset of Rn that
is bounded below has a minimal element. Thus by (3a) and (3c), for each i and
X eXi, the set F< fl {2/12/ ^ sKx), y ^ y^ has a minimal element Wi(x), say.

Observe that for x e Xi for which g,(x) g j / , , we have w,ix) = ^,.

' By this we mean that for any two vectors y and y' in 7,-, the function
9(2) = -Gi(zy -h (1 - z)y')
is a unimodal function of z on the unit interval [0,1].
 A MULTI-PRODUCT, DYNAMIC, NONSTATIONARY INVENTORY PROBLEM 213

Theorem 3.2. If (3a), (3b), (3c), and (3d) hold, then

Y*(Hi*) = -j ^ (z = 1,2,
[wi(Xi*), Qiixi*) $ y,-
where Hi* = (xi*, • • • , x,*,j/i*, • • • ,yt-i, Di, D^, • • • , Z)._i) is the history up
to period i when following Y*.

The ordering policy given in Theorem 3.2 may be described roughly as follows:
if in period i, it is possible to order so as to attain y,, do so; if this is not possible,
order the minimal feasible amount. We call this policy the base stock ordering
policy and refer to y< as the base stock level for period i. We remark that the
ordering policy in Theorem 3.1 is also a base stock policy if gi(xi) g yi.
As an illustration of this policy suppxjse n = 2 and that
= { (MI , = k,jZ for some scalars 0, j = 1, 2}
where ka and ka are fixed positive constants. Also let X, = Ri, SD, = and
qi(x) = X. The optimal ordering policy is depicted in Figure 1.
If x* lies in region I, i.e., x* lies on the line segment 7, and satisfies x* ^ yi,
then neither product is ordered. Whenever a;,* lies in region II, i.e., x,* lies above
the line segment F,- and satisfies xti ^ ya , then product 2 is not ordered and
product 1 is ordered so as to bring the vector of inventories on hand and on
order into Yi. This is illustrated by the point Fi representing the vector of initial
inventories. After ordering, the vector of inventories on hand will be at the head
of the arrow on the dotted line leading from P i . When x,* lies in r ^ o n III, i.e.,
X,* lies below the line segment Fi and satisfies x*i ^ y,i, then product 2 is ordered
and product 1 is not ordered in the manner illustrated by the point Pi. Finally,
if X.* is in region IV, i.e., x.* < ya , j = 1, 2, then the vector of inventories on
hand after ordering is y, as illustrated by the point Pz.
As we have remarked earlier, once the initial inventory vector enters the
closure of rc^on IV, the initial inventories in all future periods will also fall in
the closure of that region. It is also true under mild assumptions (which essen-

FIGURE 1
214 ABTHtTB F. VEINOTT, JB.

tially rule out the possibility of always having a zero demand for some product)
that if the initial inventory in period 1 is in either regions I, II, or III, then after
finitely many periods the closure of region IV will be reached. In this sense I,
II, and III, are "transient regions" while IV is an "equilibrium region". This fact
is illustrated in Figure 2 for the stationary case.
The generalization of the above example to the case of n products leads to a
situation in which there are 2" regions representing all possible combinations of
ordering and not ordering for each product.

Proof of Theorem 3.2: If qi(x*) ^ yi, the theorem follows from Theorem 3.1.
Suppose therefore that qi(xi*) ^ yi. Let T denote the smallest int^er such that
qT(xT*) ^ yr .Ii no such T exists, let T = -f oo.
Now consider any arbitrary ordering poKcy F. When the demands are Di,
Di, • • • , the resulting inventories before ordering are Xi ,X2, • • • and the result-
ing inventories after ordering are 2/i, 2/2, • • • (a:i = Xi*). We seek to show that
(3.1) yi < yi* = Wi(xi*) g to.(x.) ^ yi (i = 1, • • • , T - 1).
Observe that since qi(x) is nondecreasing in x for qi(x) $ yi, the same is
true of Wi(x). Evidently (3.1) holds for i = 1. Supposing now that it holds for
t = < - 1, we seek to establish (3.1) for i = i( < r ) . We have
yt < y* = wt(x*) = wt(st-i(yt-i, i><-i))
g to,(sj_i(2/j_i, Dt-i))
— w,(xt) g yt.
The second inequality follows from (3d) and the inductive assumption.
Since Gi(y) is nondecreasing in y for 2/ ^ yi and (3.1) holds.
Gi(yi*) g for i <

FlGtTBE 2
 A MTJLTI-PRODTTCT, DYNAMIC, NONSTATIONARY INVENTORY PROBLEM 215

But for i ^ T, y*= fi so

Giiyi*) ^ Giiyi) for i ^ T.
Hence
fixi\Y) = Zr-i/3,JS(?.(2/.) ^ EU^iEGiiyi*) = / ( x i | F*),
proving the optimality of F*.

The hypothesis that F,- is linearly ordered is quite restrictive for the multi-
product problem. It is natural to replace this hypothesis by the requirement that
Yi be a closed, boimded, convex set and that G,( •) be a continuous convex
function on F,-. Under these new assumptions the optimal policy in period i for
the case of two products is characterized by yi and two functions2,1 (•) and z,^i-)
as illustrated in Figure 3.
In region I one does not order either product; in region II one orders product 1
up to 2;,i(a;*2); in region III one orders product 2 up to zaixti); and in region IV,
one orders both products up to §<.
In general, z,i( •) and Zai •) will not be monotone functions. However, if for
each i Gii •) has nonnegative cross partial derivatives and if F,- has the property
that (max(«i, Vi), niin(if2, Vi)) and (min(wi, Vi), maxiu^i, Vi)) are in F< when-
ever (MI , Ui) and ivi, %) are in Yi, then Zciixn) and zuixti) are nonincreasing
in xti and x*i respectively. This may be called the case of substitute products.
Analogously, if for each i Gii') ^^s nonpositive cross partial derivatives and
if Yi has the property that inin(M, v) and max(w, v) (these are coordinate-wise
extrema) are in F, whenever u and v are in F , , then Zi^ixti) and znixii) are
nondecreasing in x*i and a;*2 respectively. This may be called the case of comple-
 216 ARTHUR F. VEINOTT, JR.

mentary products. Figure 1 illustrates a special case of this result where 3,i( •)
and ze( •) coincide.'
In subsequent papers we shall establish the above results and their analogues
for the n product case. The proofs are based on the functional equation tech-
nique of djTiamic programming and involve induction on the number of periods.
As is the case for applications of Theorem 3.1, one must nornaally require Yi
to be a convex set and Gi{y) to be convex' on Yi when applying Theorem 3.2.
Otherwise it would in general be too difficult to locate yi in a multi-product prob-
lem. If Yi is a convex set and is also linearly ordered, then F,- is necessarily a
line segment in Rn . Thus Yi may be expressed in the form
Yi = {y\y = k^ -\- Vi for some z e Zi}
with fc,(^O) and Vi being fixed elements of Rn and Zi being a closed convex
subset of Ri. We have, of course, already encountered this structure in the ex-
ample discussed above. When F,- takes the above form, we may define
g.(z) = GiikiZ + Vi) (z E Zi)
which is convex in 2 on Z, . Thus when F,- is a convex linearly ordered set, the
n-dimensional minimization problem offindingy,- is reduced to a one dimensional
problem, namely that of minimizing 9i(z)-
Up to this point we have been concerned with the infinite period model.
It is interesting to note, however, that if we set Gi(y) = 0 for all y and i > N,
then for any policy F,
fixi I F) = Lf
is the expected discounted cost overi\r(^l) periods. Of course Theorems 3.1
and 3.2 still hold in this special case. In fact we must then require only that
(3b) hold for t ^ JV - 1, and that (3a), (3c), and (3d) hold for i g AT.* In
these circumstances y,- has the same value for the N{'^i) period and infinite period
models.
Another point of interest concerning Theorems 3.1 and 3.2 is that in each
case the optimal order quantity in a period depends only upon the initial inven-
tory in the period and not otherwise upon the prior history of the process.
This fact is intuitively obvious, of course, but has usually been assumed in
earlier analyses of special cases of the problems considered here.

• Unfortunately the functions z,i(-) and ««(•) seem to be very difficult to calculate in
general. One exception to this statement occurs when it is possible to "factor" an n product
problem into n one product problems. See Section 5 for a discussion of this possibility. A
second exception occurs in the stationary case of two substitute products. For this case
EMward Ignall has shown subsequent to the preparation of this paper that z,i(-) and za{-)
are easy to compute.
' One exception to this rule occurs for the single product case where there is one demand
CIEMS, where Wi{y, t) = Wi{y — t), and where the demand densities must be Polya fre-
quency functions. Then if Wi{z) changes sign once from — to -\- as z traverses the real
line from — « to + », it can be shown [la], (lb], [7] that —Gi(y) is unimodal.
• Actually we require only that Si(y, t) be nondecreasing in y in (3d) for i S JV — 1.
 A MULTI-PBODUCT, DYNAMIC, NONSTATIONABY INVENTORY PBOBLEM 217

4. A Delivery Lag
In this section we show that under certain conditions our results can be ap-
plied to a problem in which there is a nonnegative integral lag X in delivery of
each of the n products. A more realistic model would allow the delivery lag to
vary with the product. But this seems to complicate the model considerably.
We suppose that there are n classes of demand, one for each product. In
addition, we require that unsatisfied demand be backlogged and that deteriora-
tion of stock not occur in storage. Except for these restrictions, our model is as
general as that for the case of inamediate delivery.
In order to use the model of this paper, it will be necessary for us to reinter-
pret the symbols. Before ordering in period i there will in general be some stock
on hand and some outstanding orders scheduled to be delivered in periods i + I,
• • • , i + X — 1. Denote by xa the t o ^ amount of product j on hand plus the
total outstanding orders for product j,^Th£n Xi = (xa , • • • , Xin)'isthe vector
of initial amounts of stock on hand and on order at the beginning of period i.
Xi is restricted to a set X,- as before. The manager then places an order for the
vector yi — a;,- to be delivered at the beginning of period i -f X. j/, is required,
as before, to lie in a set Fj and to satisfy y,- ^ qi(xi). In this case the backlog
assumption leads us to assume that 8<(y,-, Di) = yi — Di, which will be the
amoimt of stock on hand and on order at the beginning of period i -\- I.
Let $ij( •) denote the distribution of the sum of the demand vectors in periods
i, • • • ,j, i.e., of Di + • • • + Dy . Now if ^i is the vector of amounts on hand
and on order at the beginning of period i (after ordering), then the vector of
amounts actually on hand at the beginning of period i + \ will be

yi — Di — ••• - Di+x.-i.

The important fact to notice here is that the stock on hand at the beginning of
period i + X depends upon yi but not otherwise upon the amounts delivered in
periods i-\-l, •••,i-{-\ — 1 nor the initial stock on hand at the beginning
of period i. This property is crucial in what follows and does not hold, for ex-
ample, in the lost sales case with X > 0.
Now we may compute the conditional expected holding and penalty cost in
period i + X given that the vector of stock on hand and on order at the begin-
ning of period i is y. Denoting this function by Li(y), we see that

f f gi+Av - k ,
J»i,,+x-i •'Oi+x

where gj is defined as before. Sin is the range of D,-t- • • • + D , , and $,.._i

is a distribution with all mass concentrated at the origin and 5>,-,i_i = {01.
We assume for convenience that the cost Ci(yi — a;.) of ordering the vector
yi — a;< in period i is incurred at the time of delivery in period i + X. In this
case let«,- be the discount factor for period i + X. That is, ai is the value at the
beginning of period i + X of one cost unit at the beginning of period t + X -|- 1.
218 ARTHUR F. VEINOIT, JR.

Let

iiy, t) = ay -t / 9i-\-\iy - z, t) d$,-,m-i(2) - ocid-n (y -

•'»,-,,+x_i
for aU y, t and i where fnit) = EiDi \ Z).4ji = t). Observe that mit) = t if
X = 0 and M.(<) = EiDi) if X > 0. Also let

Giiy) = [ Wiiy, t) d^i

for all y and i. Notice that the definitions of L,-, TT,-, and G,- given above all
reduce to the definitions given in Section 2 when X = 0.
With the above interpretations an ordering policy may be defined as in Sec-
tion 2. Now let firixi I F) be the expected cost incurred in periods X -f 1, • • • ,
X -f i\r all discounted to the beginning of period X -h 1 when Xi is the initial
vector on hand and on order in period 1 and where ? is the ordering policy fol-
lowed. We assume that any stock of type j left over after \ -\- N periods can be
discarded with a return of Cjv^+i,,-. Similarly any backlogged demand for product
j remaining after \ -\- N periods is satisfied at the cost Cn+ij . Thus
Mxi \Y) = ZUPAciEiyi - Xi) -f ELiiyi)]

= HU
Since the bracketed term is not affected by the choice of F, it is convenient to
redefine /w according to (2.3). We may then define fixi \ Y) for the infinite
period model by (2.4). In each case we use the definition of Gi given in this sec-
tion. Thus we have reduced the model with a delivery lag to an equivalent model
with no delivery lag, permitting us to apply Theorems 3.1 and 3.2 to the model
of this section. We emphasize that in appl3dng these theorems where X > 0,
the ordering decision in a period when using a base stock policy is based on the
sum of the stock on hand and on order in the period. Thus for this case yi is
really a desired level of stock on hand and on order rather than simply stock on
hand. However, we retain the terms base stock policy and base stock level for
brevity.
The reduction given above can be carried out under sli^tly more general
conditions. Suppose that in addition to having one demand class for each prod-
uct, that there are supplementary random variables. Also let Di = (Z)/, D/)
where Dt is the vector of demands in the various classes and D / is the vector of
supplementary random variables. Then if we replace Di by Z)/ and replace
i) by

the above development goes through word for word.

We remark in closing that the reduction of this section does not depend upon
the linearity of the ordering cost function as do virtually all of the other results
 A MULTI-PRODUCT, DYNAMIC, NONSTATIONARY INVENTORY PROBLEM 219

of this paper. However, in the more general case, (2.4) must be replaced by
Y) = Zr-i mciiyi - Xi) + ELiiyi)].
For the single product case with g(x) = x, the result of this section is established
(in another way) in [lc].
5. Factoring a Problem into Subproblems
In real inventory systems, the number of products is usuaDy extremely large.
It is therefore of practical importance to give conditions imder which an n prod-
uct problem can be "factored" into two subproblems, labeled 1 and 2, having
ni and 7i2 (ni + 7i2 = n) products respectively. By factoring we mean that the
original problem can be solved by separately solving the two subproblems.
Having factored the original problem into two subproblems, we can then at-
tempt to factor each subproblem, and so on. Thus, factoring a problem into
several subproblems can be viewed as a sequence of factorizations, each in-
volving factoring a single problem into two subproblems. A multi-product prob-
lem should always be factored into as many subproblems as possible before
attempting to solve the problem.
We now show that the following four conditions are sufficient to permit factor-
ing a problem into two subproblems. We suppose that there are functions
Gi', Si', q/ and sets F / such that
(i) Yi = Yi' X Yi', Yi' C R^, ;
(ii) Giiy\ y') = GUV") + Gj'iy'), y'e Yi';
(iii) Biiy\ y\ t) = is>iy\ t), stiy\ t)), y' e Yi', t a S,i ;
(iv) qiix\ x") = iqi\x'), q,\x')), (x', x^) e Xi, x'e fin, ;
for j = 1, 2 and i = 1,2, • • • . One interesting fact to notice is that we do not
require that the random variables factor into two groups, one associated with
each subgroup of products. Neither do we require that demands for products in
one subgroup be stochastically independent of demands for products in the other
subgroup. This latter fact was first pointed out for a special case of our model
in [2, p. 96].
Let Do = 0 and Di = (Do, Di, • • • , Di_i) for i ^ 1. Observe that there is
no loss in generality in restricting our attention to ordering policies F such
that YiiHi) = YiiDi), i.e., policies for which the decisions in period i are
based only upon Di rather than upon the entire history Hi. The reason for this
is that by using the relations j/y = YjiHj) and Xy+i = Sy(yy, Dy) recursively
we may express the 2/y and xy+i(i < i) in terms of Di (and Xi which we take to
be fixed).
Any ordering rule YiiDi) for period i can be partitioned by

where ?/(D,) has ny ij = 1, 2) coordinates. Let F* = (FiS Yt\ •••) and

F* = (?i\ ft, ••*•). Now by virtue of (i), (iii), and (iv), F is a feasible oider-
 220 ARTHUR F. VEINOTT, JR.

ing policy for the n product problem if and only if ? ' and F^ are respectively
feasible for the two subproblems. This means that the set "y of feasible ordering
policies for the n product problem can be expressed in the form
(5.1) "y = "yi X %
where "y, is the set of feasible ordering policies for the /•• subproblem.
Now from (ii) it follows that
(5.2) /(xi I f) = /i(xx' I Y') + fiixi' I Y')
where fi{xi \ Y') = ^t-i PiEGi'{yi') with yi, yi', • • • being the sequence of
inventory levels (after ordering) that result from using F^ with the ji**" subgroup.
Combining (5.1) and (5.2) we see that
rmnrtyfixi \ Y) = min7ie^,/i(xi' | T) + min?!ty,/2(xi' | F ' ) ,
which proves that the claimed factorization is Indeed possible.
The arguments of this section do not hinge on the linearity of the ordering
cost function and may be applied whenever (5.1) and (5.2) hold.
6. The Case of Dependent Demands over Time
In this section we relax the requirement that the demands in successive
periods be independent random variables. Let t)i = (Do, A , • • • , Di-\),
t = 1, 2, • • , where Do = 0. Let 2Do = {0} and ».- = Do X SDi X • • • X 3D<-i.
Let *.(< 12/., E,) = P(Di S < I y., Hi). We assume that
(6.1) Mt\yi,Hi) =Mt\yi,Di)
for all Hi, yi, t, and i. This means that the distribution of demand in period i
depends upon the past history (yi, Hi) only through (yi, Di). As we shall see
subsequently, this assumption is essential if we are to be able to solve the in-
finite period problem by solving a sequence of one period problems in isolation.
For the present problem we must modify our definitions of Li and Cr, as fol-
lows:

iiy,z) = f
and
G,(y,z) =

for all J/ £ F<, 2 £ 2D<, and all i. All other aspects of the model structure remain
as in Section 2. In this event we see as in Section 2 that
(6.2) f(x^ I F) = Z r - i fiiEGi(yi, Di).
In order to formulate the analog of Theorem 3.1, we substitute the following
hypothraes for (3a) and (3b):

(6a) There exists a vector yi(z) that minimizes Gi(y, z) over y e Yi for each
fixed 2 £ ©i and i = 1,2, • • • .
(6b) qi.^i(Si(yi(z), t)) g y<+i(z, <) for all 2 £ » . , < f »< > and i = 1, 2, • • • .
 A MUI/n-PBODTJC?r, DYNAMIC, NONSTATIONARY INVENTORY PROBLEM 221

Theorem 6.1. If (6a) and (6b) hold, and if qiix*) ^ yi, then
Yi*iHi*) = UDi) ii = 1,2, •••)
where Hi* = ix*, • • • , a;<*, y * , • • • , yf-i, Di, ••• , Z)._i) is t h e history u p
to period i when follomng F*. Also,
xi* I Y*) = Er-i piEGiiUDi), Di).
The proof is an obvious analog of the proof of Theorem 3.1 and so is omitted.
When (6.1) does not hold, the natural generalization of Theorem 6.1 runs as
follows. If (6a) and (6b) hold with Hi* replacing z, and if qiix*) g yiiH*),
then Y*iH*) = yiiH*). However, our proof breaks down in this case. This is
because we cannot be sure that for any ordering policy F with inventories after
ordering yi,yi, • • • and histories Hi, Hi, • • • that
Giiyi, Hi) ^
The reason for this is that Hi and Hi* need not be the same.
In order to formulate the analog of Theorem 3.2, we must replace (3d) by

(6c) For each i, t e Sii, and z £ SD*, Giiy, z) and s.(j/, t) are nondecreasing in y
for y £ Yi such that y ^ ^,(z), and qiix) is nondecreasing in 2; for x £ Xi such
thatg<(a;) $ ^.(z).

Theorem 6.2. If (6a), (6b), (3c), and (6c) hold, then

where Ht = (aJi*, • • • , x*, y*, • • • , yt-i, Di, • • • , Z)<_i) is the history up to

period i when following F,*.
The proof is the same as that of Theorem 3.2 after making the obvious re-
placements.
The above theorems also hold where Di is replaced in (6.1) by a vector Dj
containing any subset of the coordinates of .D,. Then one simply replaces J5,
everyTvhere by D'i. Observe in particular that if D'i = Do, then yiiD'i) is a
constant for each i.
We remark finally that the results of Sections 4 and 5 also carry over to the
present problem. However, in order to obtain those results, we must strengthen
(6.1), requiring instead that
(6.3) $,(< I Vi, Hi) = #.(« I Di)
for all Hi, yi, t, and i. This means in effect that the decision maker has no con-
trol over the demand process.
References
I. ABEOW, K . , KABLIN, S . AND SCABF, H . (eds.). Studies in the Mathematical Theory of
Irwentory and Prodtudion, Stanford University Press, Stanford, California, 1958.
a. S. Karlin, "One Stage Models with Uncertainty," Chapter 8.
222 AKTHXTB F. VBINOIT, JB.

b. S. Karlin, "Optimal Inventory Policy for the Arrow-Harris-Marschak Dynamic

Model," Chapter 9.
c. S. Karlin and H. Scarf, "Inventory Models of the Arrow-Harris-Marschak Type with
Time Lag," Chapter 10.
2. BELLMAN, R . , GLICKSBEBO, I. AND GBOSB, O., "On the Optimal Inventory Equation,"
Management Science, Vol. 2, No. 1 (October 1965), pp. 83-104.
3. DvoRETZKT, A., KiEPEE, J. AND WoLTOwiTz, J., "The Inventory Problem I. Case of
Known Distributions of Demand," Econometrica, Vol. 20, No. 2 (April 1952), pp. 187-
222.
4. iGiiEHABT, D. AND KARLIN, S., "Optimal Policy for Dynamic Inventory Process with
Nonstationary Stochastic Demands," Chapter 8 in Studies in Applied Prcbability arid
Management Science, eds., K. Arrow, S. Karlin, and H. Scarf, Stanford University
Press, Stanford, California, 1962.
5. KAKLIN, S., "Dynamic Inventory Policy with Varjdng Stochastic Demands," Manage-
ment Science, Vol. 6, No. 3 (April 1960), pp. 231-258.
6. , "Optimal Policy for Dynamic Inventory Process with Stochastic Demands Subject
to Seasonal Variations," J. Soc. Induat. Appl. Math., Vol. 8 (1960), pp. 611-629.
7. ScHOENBEEG, I., "On Polya Frequency Functions I. The Totally Positive Functions
and Their Laplace Transforms," Journal D'Analyse MathemeUique, Vol. 1 (1951), pp.
331-374.
8. VBINOTT J R . , A. F., "Optimal Stockage Policies with Non-Stationary Stochastic De-
mands," Chapter 4 in Multistage Inventory Modeh and Techniques, eds., H. Scarf,
D. Gilford, and M. Shelly, Stanford University Press, Stanford, California, 1963.