See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/290441673

Course notes Single item inventory control models

Book · December 2000

CITATIONS READS

0 655

1 author:

Hans Frenk
Sabanci University
144 PUBLICATIONS   1,920 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Regenerative processes and renewal theory View project

Maintenance View project

All content following this page was uploaded by Hans Frenk on 15 January 2016.

The user has requested enhancement of the downloaded file.

Single Item Inventory Control

J.B.G.Frenk

January 4, 2015
Contents

I Preface. 2
1 An Introduction to Inventory Control. 4

2 Deterministic Inventory Control 8

3 The Newsboy Problem. 17

3.1 The newsboy problem with complete back ordering. . . . . . . . . . . . . . 18
3.2 The newsboy problem with lost sales. . . . . . . . . . . . . . . . . . . . . 21
3.3 Computation of the optimal order size and sensitivity analysis. . . . . . . . 22
3.4 Service measures and the newsboy problem. . . . . . . . . . . . . . . . . . 27

4 The Theory of Regenerative Processes. 29

4.1 Regenerative processes and the renewal argument. . . . . . . . . . . . . . . 29

5 Basics of Stochastic Single Item Inventory Control. 49

6 Regenerative Processes and Single Item Inventory Control. 60

6.1 Analysis of the (R,S) model. . . . . . . . . . . . . . . . . . . . . . . . . . 60
6.1.1 Computation of the objectives in a (R,S) model. . . . . . . . . . . . 63
6.2 Analysis of the (R,s,S) model. . . . . . . . . . . . . . . . . . . . . . . . . 68
6.2.1 Computation of the objectives in a (R,s,S) model. . . . . . . . . . . 73
6.3 Analysis of the (R,s,nQ) model. . . . . . . . . . . . . . . . . . . . . . . . 75
6.3.1 Computation of the objectives in a (R,s,nQ) model. . . . . . . . . . 79
6.4 Analysis of the (s,S) model. . . . . . . . . . . . . . . . . . . . . . . . . . . 81
6.4.1 Computation of the objectives in a (s,S) model. . . . . . . . . . . . 87
6.5 Analysis of the (s,nQ) model. . . . . . . . . . . . . . . . . . . . . . . . . . 88
6.5.1 Computation of the objectives in a (s,Q) model. . . . . . . . . . . . 93

1
 Part I

Preface.

2
 In these notes the theory of single item inventory control models with a deterministic
and stochastic demand process is presented. It is shown that the theory of regenerative pro-
cesses is a unifying framework for analyzing all the classical single item inventory control
models. By this observation it is easy to write down the mathematical expressions for cost
functions and restrictions commonly used within the field of inventory control. Both the so-
called cost and service level approach are considered. At this point these notes differ from
the main stream in inventory control since it used to be common within this field to derive
elementary approximations of the cost functions and service level restrictions (cf.[17],[5]).
Although the exact expressions for these functions and restrictions are not given by ele-
mentary functions and so at first sight these expressions are not suitable for quick numer-
ical evaluations (this is the main reasons why simple approximations are used) it is still
possible to achieve this by applying the appropriate tools from numerical mathematics. In
particular, for divisible items, a fast and robust inversion algorithm for generating functions
called the Fast Fourier Transform method is presented in [22],[21] and [20]1 . Finally we
observe that the combination of regenerative processes, Laplace transform inversion and
nonlinear programming techniques and the availability of computer packages like MAT-
LAB or MAPLE serves as a very powerful toolbox to solve applied probability models
occurring within Operational Research. Although these notes only discuss their interaction
in the field of inventory control a similar observation also holds for other fields such as
Queuing theory, Reliability theory and Insurance mathematics.

1 For an excellent book on the theory behind numerical mathematics and in particular the FFT method see[28].

3
Chapter 1

An Introduction to Inventory
Control.

Inventory may be considered as an accumulation of a product that will be used to satisfy

future demand for that product. It requires a policy to manage that inventory. This policy
may involve to tell when to order, how much to order, which products to order. Obviously
there are costs involved related to the previous actions. As an example we mention ordering
costs, transportation costs, holding costs, costs related with shortage and so on. Usually the
choice of the policy will be done with the objective of cost minimization. Since the begin-
ning of this century inventory problems are being treated with mathematical techniques.
The first classical model of inventory control is the deterministic Economic Order Quantity
model (EOQ model) described in 1913 by Harris (cf.[14],[8]). For a discussion of the roots
of this model the reader is referred to Erlenkotter (cf.[3]). Only after World War II the
research started into stochastic inventory models. In these inventory models it is assumed
that the demand process is a stochastic process. Before discussing single item inventory
models with a stochastic or deterministic demand process a general introduction into the
field of inventory control will be given. Inventory control includes several kind of problems
such as Statistics (data analysis, inference, parameter estimation, etc), Computer Science
(to maintain a record of the inventory in an adequate data base) and Operational Research
(modeling and determination of an optimal or a reasonable order policy). Different types
of inventory systems may be considered such as pure inventory systems where only the in-
ventory itself is taken into account. Other systems are production inventory systems, where
production interactions are included and distribution inventory systems which involves the
allocation of the available inventory and so on. In these notes we will study pure inventory
systems. Each particular problem has its own characteristics and below we list the most
important ones (cf. [30]).

• Planning horizon. It is the time over which the inventory level is controlled. This
horizon may be finite or infinite, deterministic or stochastic.
• Products. The inventory system may involve one or many products. The items
which are stored may differ from each other in many ways and interactions may take
place among the different items. There are products that have to be stocked under
controlled conditions (humidity, temperature, dust proof, etc); some are perishable
or subject to obsolescence;others can be stocked indefinitely exposed to the elements
without deterioration. In case of interactions, some items may be substitutes of each
other, or may compete for limited space or investment capacity.
• Demand process. The demand process may occur continuous in time or it may only
occur at certain fixed points in time. It may consist of discrete sizes {1, 2, ..} or

4
 continuous sizes (0, ∞). As an example we mention the sales of electronic equip-
ment (discrete size) or the sales of gasoline (continuous size). It may be stationary
or non stationary. Moreover the demand may be deterministic or stochastic. If we
are dealing with a stochastic process it may have independent increments (compound
Poisson process) or not (compound renewal process), to be known or unknown. The
most simple assumption is that the demand process is deterministic and stationary
and this assumption holds for the classical Economic Order Quantity model.
• Lead time. This denotes the time that elapses from the point in time that an order
is issued until it arrives at the facility. Basically we may consider deterministic and
stochastic lead times. If a deterministic lead time L is assumed it may be equal to
zero (instantaneous replenishment) or positive. If a stochastic lead time L is taken
into account then the analysis becomes extremely complicated. In such a case it
might happen that orders placed later will arrive earlier at the facility and so issues
about order crossing have to be considered.
• Review process. The inventory process may be controlled either continuously or
periodically. In continuous review, the level of inventory is known at all moments in
time and the reorder decisions can be made at any time. In periodic review the inven-
tory level is known only at discrete points in time. As an example we mention that
at the end of a working day the inventory level can be determined and so the reorder
decisions can be made only at these moments. Observe these moments correspond
with the beginning of a new period. At the same time we might have a continuous
review system but the reorder decisions can only be taken due to outside restrictions
at discrete points in time.
• Shortage. If demand occurs and no stock is available we have what is called a
shortage. The way that the systems reacts to this situation is important to the structure
of the process. Basically there are two possibilities. Either the demand is lost or it is
back ordered and the customer waits until the delivery. In addition, it is possible to
develop models where mixtures of back orders and lost sales are considered. There
are different ways to treat this situation. One is to fix portions of the shortage, a
portion to be lost and another to be back logged. Another is that customers are
willing to wait but only a fixed amount of time.
• Costs. Since usually the goal is to minimize some cost function its characteristics
are very important. The most relevant costs are listed below.

Ordering costs. These are costs incurred by the inventory system itself during the
ordering process and include administrative costs, transportation costs, inspection costs,
testing costs, etc. These costs can be divided into two parts:those that are independent
of the quantity ordered and those which are dependent. The first ones are usually called
setup costs. It is a fixed cost involved with the placement of an order. The other costs are
a function of the order quantity and the most common assumption is that these costs are
proportional to the order quantity. However, in situations of quantity discounts or price
breaks this function can be concave, convex or even more general. It is important to note
that including a setup cost in a model increases the complexity considerably.
Holding costs. This cost is the direct result of carrying inventory in storage and in-
cludes taxes, insurance, opportunity cost, obsolescence costs, rent, operating costs (light,
heating),etc. It is in general very difficult to represent all these costs with great accuracy in
the model and so one needs a simplification. The common assumption is that holding costs
are proportional to the level of the inventory and vary directly with the storage duration, i.e
holding costs are linear. Clearly in a continuous review system we may adapt the costs of
keeping inventory continuously while in a periodic review system it might happen that the
costs of keeping inventory are charged at period ending inventory levels.

5
 Penalty costs. It is the cost charged whenever a shortage occurs and this may be charged
in a complete back ordering system or a lost sales system or a combination of both. The
most common ones are listed by the following.
Average number of stock outs.
A stock out occurs if the level of the inventory process drops from a nonnegative value
to a negative value and so this corresponds to the demand of a customer which cannot
be completely fulfilled. The average number Bstockout of stock outs is then defined by
the expected number of stock outs over the planning horizon divided by the length of the
planning horizon. Associated with this the penalty cost is then given by multiplying the
average number Bstockout of stock outs with a fixed value bstockout > 0 represented the
penalty cost per stock out.
The average number of items short.
The average number Bitemshort of items short is given by the expected total number
of items short over the planning horizon divided by the length of the planning horizon.
The penalty costs are now given by multiplying the value Bitemshort by a given positive
value bitemshort . Observe the value bitemshort represents the cost per item short and does
not depend on the duration of the shortage. This cost concept is often used in inventory
systems where items short must be satisfied by overtime production and the associated
value bitemshort is then determined by the additional cost of overtime production per item
short.
Average number of backorders per unit of time.
The average number Bbackorder/time of backorders per unit of time is given by the total
number of backorders per unit of time within the planning horizon divided by the length
of the planning horizon. The associated costs are now given by multiplying the value
Bbackorder/time by a positive value bbackorder . Observe the value bbackorder represents
the penalty cost per backordered item per unit of time. This cost concept is used if the
inventory items are spare parts and each unit of time that such a part is not available results
in a machine not able to operate.
Obviously in the lost sales case the penalty costs are usually proportional to the ordered
quantity only (why?) and comprises a number of different factors like goodwill costs and
lost profit. This means that in this case we are dealing with penalty costs proportional to
the average number of items short. In case the complete backordering regime holds the
penalty costs are mostly proportional to the average number of backorders per unit of time
and so these penalty costs also incorporate the time it takes before a item is backordered.
Clearly in many cases the above penalty costs may not be real costs since factors as loss of
customers and goodwill are included and these costs are extremely difficult to measure. To
avoid this problem also so-called service measures are introduced in the literature.

• Service measure. This denotes a performance measure of the inventory system and
can be defined in various ways. Observe a replenishment cycle is given by the time
evolving between two consecutive replenishments of the inventory system. The most
common ones are given by the following.

The no-stock-out service measure.

This service measure denotes the probability of no stock out during a replenishment
cycle. Observe a stock out is defined as the event that the so called net stock inventory
level drops within the replenishment cycle from a nonnegative value to a negative value.
This measure only takes the appearance of a stock out into consideration and not the size
or duration of the stock out. In a lot of inventory models the no-stock-out service measure
equals the average number of stockouts divided by the expected length of a replenishment
cycle and so this service measure is related to the average number of stockouts. In the

6
literature the measure is denoted by Pno−stockout . It is also called the α-service measure
(cf.[5]).
The fill rate service measure.
This service measure denotes the fraction of demand directly delivered from stock. In
a lot of inventory models this fill rate service measure equals the average number of items
directly delivered divided by the average demand. Moreover, the average number of items
directly delivered equals the average demand minus the average number of items short
and so the fill rate service measure is related to the average number of items short. This
measure is very popular in practice and is denoted by Pf illrate . It is also called the β-
service measure (cf.[5]). A typical example of this measure is given by the condition that
95% should be delivered directly from stock.
The ready rate measure.
This service measure denotes the fraction of time that the inventory level is positive. It is
denoted by Preadyrate and it is often used in the control of inventory systems of equipment
used for emergency purposes.
The above characteristics represent the basic elements of inventory problems. Observe
when dealing with the above characteristics we might face the following problem. In case
the shortage cost can be determined we might use the cost approach (cf.[12]). In this ap-
proach we try to derive the optimal control rule by minimizing the sum of the expected
holding, shortage and ordering costs. In case the shortage costs incorporate the loss of
goodwill of customers this problem can be seen as a long term or strategic optimization
problem. On the other hand, if the shortage costs can not be determined and this is often
the case in a practical situation, we try to determine the optimal control rule by minimizing
the expected holding and ordering cost subject to some service level restriction and this is
often called the service level approach. This type of optimization problem can be seen as
a short term optimization problem. Finally we observe that besides the above cost con-
siderations many other situations can occur like the price as a decision variable, quantity
discounts, competition among firms, uncertainty in the orders received, etc. Real systems
are in general so complex that they cannot be represented with complete accuracy. There-
fore a model with simplifying assumptions must be made. In these notes we will start with
the simplest possible models and by adding all kind of restrictions we will increase the
complexity of those models.

7
Chapter 2

Deterministic Inventory Control

In this chapter we first apply the cost approach to a class of single-item pure inventory
control models over an infinite horizon and a deterministic stationary demand process with
state space E = N ∪ {0} or E = [0, ∞). A deterministic demand process is given by a right
continuous function D : T → E with

D(t) := total demand up to time t

and T = [0, ∞) or T = N ∪ {0}. It is called stationary if

D(t + s) − D(t) = D(s) (2.1)

for every t, s ∈ T. For a periodic review model or equivalently T = N ∪ {0} it follows by

relation (2.1) that D(n) = nµ for some 0 < µ < ∞, while for a continuous review model
or T = [0, ∞) one can show that D(t) = µt for some 0 < µ < ∞ (see Theorem 1.4 of
[11]). Hence for a continuous review model it follows that the state space E is necessarily
[0, ∞) or equivalently the item under consideration is called indivisible1 . If the state space
E = N ∪ {0} the item is called divisible2 . Hence a stationary deterministic demand process
must be a linear function and without loss of generality we may assume now that the lead
time equals zero (why?). To set up a cost structure we introduce the so-called on-hand
stock process I := {I(t) : t ∈ T } on state space E given by

I(t) := actual stock on the shelves at time t.

Moreover, if the complete back ordering regime holds the process B := {B(t) : t ∈ T }
represents the back ordering process on state space E defined by

B(t) := amount of items back ordered at time t.

Combining both processes we introduce the net stock process IN := {IN (t) : t ∈ T }
given by
IN (t) := I(t) − B(t).
Since the complete backordering regime holds, it follows that

I(t) = max{IN (t), 0} and B(t) = − min{IN (t), 0}

and this implies that the on-hand stock and backordering process can be replaced by the
net stock process. The inventory holding and penalty costs of the inventory system are now
cost functionals defined on this process. To model these costs we introduce a so-called cost
1 indivisible item
2 divisible item

8
rate function f : R → [0, ∞) and this means for T = [0, ∞) that the inventory holding
and penalty costs during the interval [t, t + ∆t) and ∆t small are approximately given by
f (IN (t))∆t, while for T = N ∪ {0} the same cost on the interval [t, t + 1), t ∈ N ∪ {0}
is given by f (IN (t + 1)). To complete our discussion of the cost structure we additionally
introduce the next costs and rewards:
K := fixed cost per order
p := selling price per item (2.2)
c := ordering cost per item
It is assumed that the selling price per item is higher than the ordering cost per item and so
the profit
pprofit := p − c
per item is positive. As already observed we may assume without loss of generality that for
any control rule the lead time equals 0. The control rule we now introduce is the so-called
(R, S) control rule and this means that every R time units with R belonging to T \{0} a
replenishment order is issued and the size of the order is such that the net stock inventory
level is raised up to level S ∈ E. Since for any (R, S) control rule with backordering
the revenues are the same it is sufficient in this case to consider cost minimization. It can
be shown relatively easy (cf.[25]) for a deterministic stationary demand process and the
above cost structure that the optimal policy over all policies with respect to the average
cost over an infinite horizon is a member of the class of (R, S) control rules. To determine
the average inventory/holding and ordering cost Φb (R, S) for such a (R, S) control rule
and zero lead time under the backlog regime it is assumed without loss of generality that
at time 0 a replenishment order is issued and so the net stock level at time 0 equals S. By
definition it follows that the average cost equals (if it exists!)
Φb (R, S) := limt∈T ↑∞ t−1 Φb (t, R, S) (2.3)
with Φb (t, R, S), t ∈ T denoting the inventory holding/penalty and ordering costs up to
time t. To compute for a continuous review model or T = [0, ∞) the value Φb (t, R, S) we
introduce the deterministic counting process NR := {NR (t) : t ≥ 0} given by
NR (t) := sup{n ∈ N ∪ {0} : nR ≤ t}
representing the total number of replenishment events occurring after time 0. Within each
so-called replenishment cycle [nR, (n + 1)R), n ∈ N ∪ {0} it follows by the complete
backordering assumption that the total cost Φn+1 of the (n + 1)th replenishment cycle is
given by
∫ (n+1)R
Φn+1 = K + f (IN (y))dy.
nR
Since the demand function is linear and a (R, S) inventory control rule is used we obtain
for every nR ≤ y < (n + 1)R that IN (y) = S − µ(y − nR) and this implies
∫ R
Φn+1 = K + f (S − µy)dy. (2.4)
0

Hence it follows for t > R that

∑NR (t)−1 ∫ t
Φb (t, R, S) = Φn+1 + K + f (IN (y))dy (2.5)
n=0 NR (t)R

and using limt↑∞ t−1 NR (t) = R−1 and relations (2.3), (2.4) and (2.5) the average inven-
tory/holding and penalty costs Φ(R, S) are given by
( ∫ R )
1
Φb (R, S) = K+ f (S − µy)dy . (2.6)
R 0

9
Since µpprofit denotes the average income it is therefore advantageous to operate a given
(R, S) control rule if Φb (R, S) − µpprofit > 0. Using exactly the same approach we obtain
under the complete backordering regime for a periodic review model and R necessarily
belonging to N that
( ∑R−1 )
1
Φb (R, S) = K+ f (S − µn) . (2.7)
R n=0

To determine the optimal (R, S) control policy we therefore need to solve the cost mini-
mization problem
v(PBacklog ) := inf{Φb (R, S) : R ∈ T \{0}, S ∈ E} (PBacklog )
with Φ(R, S) either given by relation (2.6) or (2.7). It is now natural to consider the fol-
lowing class of cost rate functions.

Condition 1 The costrate function f : R → [0, ∞) is nondecreasing on (0, ∞), nonin-

creasing on (−∞, 0) and satisfies f (0) = 0.

By the separability of the decision variables R and S within the objective function it
follows that the optimization problem (PBacklog ) can be rewritten as
K + ϕ(R)
inf{ : R > 0}
R
with the function ϕ : (0, ∞) → R given by
∫ R
ϕ(R) := inf{ f (S − µy)dy : S ≥ 0}. (Ppartial )
0

Since Condition 1 holds this yields for every S > µR with R > 0 fixed that
∫ R ∫ R
f (S − µy)dy ≥ f (µR − µy)dy
0 0

and by this observation we obtain that the feasible set [0, ∞) of the optimization problem
(Ppartial ) can be replaced by the compact set 0 ≤ S ≤ µR. Therefore it follows that
∫ R
ϕ(R) = inf{ f (S − µy)dy : 0 ≤ S ≤ µR}
0

and to show that this optimization problem has an optimal solution we observe that
∫ R
1
f (S − µy)dy = (F (S) − F (S − µR)) (2.8)
0 µ
with the function F : R → R given by
∫ x ∫ 0
F (x) := f (z)dz for x ≥ 0 and F (x) := − f (z)dz for x < 0. (2.9)
0 x

Since the function F is clearly continuous we obtain by Weierstrass theorem (cf.[29]) that
an optimal solution exists and so we have already verified part of the following lemma.
Lemma 2 If the complete backordering regime and Condition 1 holds and ϕ : (0, ∞) →
[0, ∞) is the convex function representing the optimal objective value of optimization prob-
lem (Ppartial ) then for T = [0, ∞) it follows that
1
v(PBacklog ) = inf{ (K + ϕ(R)) : R > 0}.
R

10
 Proof. We only check that the function ϕ is convex since the representation has already
been verified. By Condition 1 it follows that the function F given by relation (2.9) is
concave on (−∞, 0) and convex on [0, ∞) (cf.[1]). Also the set K := {(S, R) : 0 ≤ S ≤
µR, R > 0} is a nonempty convex set and so it is easy to check that the function I : K → R
given by
∫ R
1
I(S, R) := f (S − µy)dy = (F (S) − F (S − µR))
0 µ
is convex on K. Since the optimal solutions S(Ri ), i = 1, 2 of the optimization prob-
lem (Ppartial ) associated with ϕ(Ri ), Ri > 0 belong to the convex cone K and hence also
αS(R1 ) + (1 − α)S(R2 ) ∈ K this implies

ϕ(αR1 + (1 − α)R2 ) ≤ I(αR1 + (1 − α)R2 , αS(R1 ) + (1 − α)S(R2 ))

≤ αI(R, S(R1 )) + (1 − α)I(R2 , S(R2 ))
= αϕ(R1 ) + (1 − α)ϕ(R2 )

and so the desired result is verified.

Intuitively the above formula is clear since for S > µR we will issue a replenishment
order at a positive net stock level and by Condition 1 it is always advantageous to wait
until the net stock level hits zero or equivalently S = µR. Clearly for T = N ∪ {0}
a similar simplification can be carried out. To solve the above optimization problem we
first observe that ϕ : (0, ∞) → R is convex if and only if the function ϕ0 : (0, ∞) → R
given by ϕ0 (R) := Rϕ(R−1 ) is convex (cf.[24]). This shows that the function R →
R(K+ϕ(R−1 )) is convex on (0, ∞) and so optimization problem (PBacklog ) can be replaced
by the convex minimization problem

inf{R(K + ϕ(R−1 )) : R > 0}.

This problem is easy to solve and its optimal solution (if it exists!) can be easily trans-
formed to the optimal solution of optimization problem (PBacklog ). A particular instance of
the above optimization problem is given by the following classical example.

Example 3 Consider for T = [0, ∞) and complete backordering the cost rate function
f : R → [0, ∞) given by
cH x x ≥ 0
f (x) = .
−bB x x < 0
Observe the value cH denotes the inventory holding cost per item in inventory per unit of
time and the value bB the backordering cost per backordered item per unit of time. To
compute the optimal (R, S) policy with respect cost minimization we only need to compute
by Lemma 2 the value of the objective function for 0 ≤ S ≤ µR. For this choice of
parameters it follows that
∫ R
cH S 2 bB (µR − S)2
f (S − µy)dy = + .
0 2µ 2µ
The above function is clearly convex with respect to (S, R) and by using the first order
conditions for fixed R > 0 we obtain that S(R) = µbB R(cH + bB )−1 and
∫ R
µbB R
= arg min{ f (S − µy)dy : 0 ≤ S ≤ µR}. (2.10)
cH + bB 0

Substituting the above solution of optimization problem (Ppartial ) into its objective function
we obtain after some computations that
µcH bB R2
ϕ(R) = .
2(cH + bB )

11
This implies that the cost associated with a (R, S) policy with S = µbB R(cH + bB )−1 is
given by
µbB R K µcH bB
Φb (R, )= + R.
cH + bB R 2(cH + bB )
Minimizing now the above convex function over R > 0 by evaluating the first order condi-
tion and applying relation (2.10) implies that the optimal (Ropt , Sopt ) control policy equals
√ √
2(cH + bB )K 2µbB K
Ropt = and Sopt = .
µcH bB µcH (cH + bB )

Moreover, it is easy to verify that the optimal objective value v(PBacklog ) with respect to
ordering and inventory/penalty costs is given by
√
2µcH bB K
v(PBacklog ) =
cH + bB
and so the optimal average profit (if one decides to operate the inventory system) is given
by √
2µcH bB K
optimal average profit = µpprofit − .
cH + bB
In case the cost approach is used and we consider the lost sales regime with continuous
review or T = [0, ∞) we need to assign penalty costs due to lost sales. The most natural
way to compute those penalty costs is to assume that for each demanded item not in stock a
penalty cost bL > 0 is paid. It is not difficult to verify using the linear demand process that
this can be modeled using a cost rate function f : R → [0, ∞) satisfying f (x) = bL µ for
every x < 0 and so this function also satisfies Condition 1. Also, since demand and hence
revenu is lost when the system is out of stock, it is no longer true that for every (R, S) rule
the reveues are the same and so we also need to take into account the received revenues
when operating a given (R, S) control rule. Under the lost sales regime it is easy to see
using Condition 1 that for any (R, S) control rule with S > µR it is as good to use the
control rule (R, µR) and so it is sufficient to minimize over all (R, S) control rules with
0 ≤ S ≤ µR. The average cost including revenues for a (R, S) control rule satisfying
0 ≤ S ≤ µR is now given by the optimization problem

v(PLostsales ) := inf{ΦL (R, S) : 0 ≤ S ≤ µR, R > 0} (PLostsales )

with the objective function given by

( ∫ Sµ−1 )
1
ΦL (R, S) = K+ f (S − µy)dy + (µR − S)bL − pprofit S .
R 0

Simplyfying the above objective function and using relation (2.9) we obtain the following
result.

Lemma 4 If the lost sales regime and condition 1 holds then for T = [0, ∞) it follows that
1 ( )
v(PLostsales ) = inf{ K + µ−1 F (S) − (pprofit + bL )S : 0 ≤ S ≤ µR, R > 0} + µbL .
R
In general the above optimization problem might not have an optimal solution and to
show this we consider the lost sales variant of Example 3.

Example 5 Consider for T = [0, ∞) the cost rate function f : R → [0, ∞) given by

cH x x ≥ 0
f (x) =
bL µ x < 0

12
The objective function listed within optimization problem (PLostsales ) reduces to

K+ 1
2µ cH S
2
− (pprofit + bL )S
h(R, S) :=
R
and so we need to solve the optimization problem

inf{h(R, S) : 0 ≤ S ≤ µR, R > 0}.

To solve the above optimization problem we need to consider different cases and to start
we assume that our data satisfy

2cH K ≤ µ(pprofit + bL )2 .

Fixing now R > 0 and optimizing over 0 ≤ S ≤ µR we first consider the optimization
problem
cH S 2
inf{ − (pprofit + bL )S : 0 ≤ S ≤ µR}. (PR )
2µ
By standard calculus it follows with v(PR ) denoting the optimal objective value of op-
timization problem (PR ) that for 0 < R ≤ c−1 H (pprofit + bL ) the optimal solution is
given by Sopt = µR, while for R > c−1 H (pprofit + bL ) the optimal solution is given by
Sopt = c−1
H µ(pprofit + b L ) and this implies

1
2 cH µR
2
− (pprofit + bL )µR if 0 < R ≤ c−1
H (pprofit + bL )
v(PR ) = (2.11)
− 2cµH (pprofit + bL )2 if R > c−1
H (pprofit + bL )

Dividing our original optimization problem into the optimization problems

pprofit + bL
inf{h(R, S) : 0 ≤ S ≤ µR, 0 < R ≤ } (PLostsales1 )
cH
and
pprofit + bL
inf{h(R, S) : 0 ≤ S ≤ µR, R > } (PLostsales2 )
cH
we will first analyze optimization problem PLostsales1 . By relation (3.11) we obtain

1 K pprofit + bL
v(PLostsales1 ) = −(pprofit + bL )µ + inf{ cH µR + :0<R≤ }.
2 R cH

Since we assume that 2cH K ≤ µ(pprofit + bL )2 it follows by standard calculus that the
optimal solution of optimization problem

1 K pprofit + bL
inf{ cH µR + :0<R≤ }
2 R cH
√
2K
is given by Ropt := 2
µcH and this implies

1 K pprofit + bL √
inf{ cH µR + :0<R≤ } = 2 2KcH µ
2 R cH
√
2K
Hence it follows with optimal solutions Sopt = µRopt and Ropt := 2 µcH
that
√
v(PLostsales1 ) = −(pprofit + bL )µ + 2
2KcH µ ≤ 0.

13
Also for optimization problem PLostsales2 we obtain

2cH K − µ(pprofit + bL )2 pprofit + bL

v(PLostsales2 ) = inf{ :R> }
2cH R cH

and this shows since 2cH K ≤ µ(pprof it + bL )2 that

cH K µ
v(PLostsales2 ) = − (pprofit + bL ) ≤ 0 (2.12)
(pprofit + bL ) 2

with optimal solutions Sopt = c−1H µ(pprofit + bL ) and Sopt = µRopt . Using now relations
(2.11) and (2.12) we finally obtain that v(PLostsales ) equals
√ cH K µ
min{ 2 2KcH µ − (pprofit + bL )µ, − (pprofit + bL )}
(pprofit + bL ) 2

and for both optimization problems PLostsales2 and PLostsales1 we obtain that the optimal
solution (Ropt , Sopt ) satisfies Sopt = µRopt . Clearly by Lemma 13 the optimal average
profit (if one decides to operate the inventory system) is given by
√ µ cH K
max{µpprofit − 2
2KcH µ, (pprofit − bL ) − }.
2 (pprofit + bL )

In case the data satisfies 2cH K > µ(pprof it + bL )2 the analysis of the above optimization
problem is left for the reader.

If the complete back ordering regime holds and the service level approach is used we do
not know the shortage costs. This implies that the cost rate function f vanishes on (−∞, 0)
and the same objective function as shown in relation (2.6) or (2.7) can be derived. Clearly
if the service level approach is used Condition 1reduces to the following one.

Condition 6 The cost rate function f : R → [0, ∞) is nondecreasing on (0, ∞) and

f (x) = 0 for every x ≤ 0.

As observed in the previous section one mostly uses in practice a service measure as a
restriction. For the above deterministic model the only suitable service measure is the fill
rate service measure. It is easy to verify for any R > 0 and S ≥ 0 that the fraction of
demand directly delivered from stock equals

S
Pf illrate (R, S) = .
µR

If the restriction Pf illrate (R, S) ≥ β with β a chosen number belonging to (0, 1] is intro-
duced, then the associated cost minimization problem ( Observe for any (R, S) control rule
satisfying the fill rate assumption the revenues are the same due to backordering regime) is
given by
v(PFillrate ) := inf{Φb (R, S) : S ≥ µβR, R > 0} (PFillrate )
and for this problem one can show the following result.

Lemma 7 If the complete backordering regime and Condition 6 holds and the fill rate
measure is used then for T = [0, ∞) it follows that

v(PFillrate ) = inf{Φb (R, µβR) : R > 0}.

14
 Proof. By Condition 6 it is easy to verify for any fixed R > 0 that the restriction S ≥
µβR is binding and so the optimization problem (PFillrate ) reduces to the one-dimensional
optimization problem
inf{Φb (R, µβR) : R > 0}.
This shows the desired result.
A well-known special case of the above problem and the first inventory control model
is the so-called Economic Order quantity model (cf.[14]).
Example 8 Let us consider the service level approach with β = 1 and the cost rate func-
tion f given by f (x) = cH x for every x ≥ 0 and zero otherwise. This corresponds to the
classical Economic Order Quantity model where no shortage is allowed. In this case we
know by the previous observations that S = µR and
K cH µR
c(R, µR) = + − µpprofit .
R 2
It is now easy to verify by taking the first order conditions that the optimal (Ropt (EOQ), Sopt (EOQ))
control rule is given by
√ √
2K 2µK
Ropt (EOQ) = and Sopt (EOQ) =
µcH cH
Moreover, the optimal objective value with respect to costs and profits is given by
√
v(PEOQ ) = 2µcH K − µpprofit
and the optimal average profit (if one decides to operate the inventory system) equals
√
optimal average profit := µpprofit − 2µcH K.
The results of the EOQ model can be made intuitively clear from the result in Example
3. Since in the Economic Order Quantity model no shortage is allowed one can only achieve
this in a cost approach by setting bB equal to infinity. Hence the results of Example 8
should be rediscovered by letting bB tend to infinity in Example 3 and as one can verify
this indeed holds. This concludes our discussion of deterministic inventory models with a
linear demand process and in the next section we consider the easiest possible stochastic
inventory models.

Exercise 9
1. Consider the EOQ model with K denoting the fixed ordering cost per order and
cH > 0 de inventory holding cost per item per unit of time. It is now well-known
that √
2K
Ropt (EOQ) =
µcH
and the costfunction f : (0, ∞) → R equals (do not include profit from sales!)
K cH µR
f (R) = + .
R 2
A disadvantage of this value Ropt might be that it is not practical to apply (why?).
To avoid this problem one might choose a fixed planning horizon RB ≤ Ropt and
consider the class of power of two policies given by RB 2k , k ∈ N∪{0}. To determine
within this class of power of two policies the optimal power of two policy one has to
solve the optimization problem
min{f (RB 2k ) : k ∈ N ∪ {0}}.
Give a procedure to compute the optimal power of two policy.

15
 2. If the optimal power of two policy √ is given by
√ RB 2kopt and RB ≤ Ropt show that
RB 2kopt belongs to the interval [( 2)−1 Ropt , 2Ropt ].
2 2

3. Show for an optimal power of two policy RB 2kopt satisfying RB ≤ Ropt that

f (RB 2kopt ) 1 1 √
≤ (√ + 2) ≈ 1.06
2
2
f (Ropt ) 2 2
√ √
(hint: Use f (( 2 2)−1 Ropt ) = f ( 2 2Ropt )!)

4. Analyze the effect of a optimal power of two policy on the model discussed in Ex-
ample 3.

Exercise 10

Consider different items i, i = 1, .., n and suppose that the demand for item i has a
linear demand process with rate µi , Ki is the fixed order cost per order for item i and ci
are the inventory holding cost per item i per unit of time. If the fill rate measure for all
items equals 1 determine then the optimal planning period for each item and the optimal
order-up-to-level. These items are stored in a distribution centre with a finite floor capacity
of f m2 and each product i has a floor capacity of fi m2 . If we do not take into account
that the orders can be ordered in such a way that it never happens that all optimal order-up-
to-levels are stored in the distribution centre construct then a optimization problem which
determines the optimal order-up-to-levels for each item given the finite capacity of f m2
and solve this model. (Hint: Use the Lagrangian relaxation technique involving Lagrangian
multipliers!). Do the same for the general single item inventory control model discussed in
this section.

Exercise 11

Analyze the lost sales model in Example 5 in case the data satisfy 2cH K > µ(pprof it +
bL )2 .

16
Chapter 3

The Newsboy Problem.

In this chapter we consider a single item inventory system and suppose that at the end of
every period a replenishment order is issued if after inspection the inventory level is below
the value S. The size of the replenishment is such that the inventory level at the beginning
of the next period is again equal to the so called order-up-to-level S. This means that we
consider a periodic review system with zero lead time and a so-called (1, S) control rule.
Without loss of generality we may assume that at time 0 the level of inventory equals S.
Moreover, to model the costs and profits of the above system we introduce the following
expressions:

K := ordering cost per order.

cH := inventory cost per item in stock at end of period.
bIS := penalty per demanded item not in stock.
p := selling price per item.
c0 := ordering cost per item.
For notational convenience the following related expressions are also considered:

pprofit := p − c0
a1 := cH + bIS
a2 := bIS − pprofit
In the above model the item short penalty cost is used and so the costs are charged at
end of period inventory levels. It is also assumed that the profit pprofit per sold item is
positive. To describe the demand process of the inventory model we introduce the sequence
{Yn , n ∈ N} of random variables with state space E either [0, ∞) or N ∪ {0} given by

Yn := total demand occurring in period n.

For the moment we do not impose any conditions on this sequence {Yn : n ∈ N}. Before
discussing whether we are considering the back ordering or lost sales regime or a combi-
nation of both we also need to define how costs are measured over time. In all the cases the
expected average cost over an infinite horizon is used and to define these costs we introduce
for every order-up-to-level S the sequence {Cn (S) : n ∈ N} of random variables given by

Cn (S) := total random cost minus profit due

to sales over the first n periods.

This implies that E(Cn (S)) denotes the expected costs (including profit due to sales!) over
the first n periods and the expected average costs are now represented by
E(Cn (S))
c(S) := lim sup .
n↑∞ n

17
The optimal order-up-to-level S policy is then an optimal solution of the optimization prob-
lem
min{C(S) : S ∈ E} (P)
with E := N ∪ {0} in case the demand process consists of discrete sizes and E := [0, ∞) in
case the demand process consists of continuous sizes. We also have to distinguish whether
the complete back ordering or the lost sales regime or a combination of both hold. For
simplicity only the back ordering and lost sales regime is considered and a combination of
both regimes is discussed in exercise 16. The above inventory model without fixed ordering
costs and restricted to only one period is known in the literature as the newsboy problem
and this problem serves as the easiest inventory model with a stochastic demand process.
Also, the one period version is mostly used to model the inventory of perishable items (cf.
[5]).

3.1 The newsboy problem with complete back ordering.

To analyse this model we first need to determine an expression for the random cost Cn (S)
for every n ∈ N. Since by the definition of the order-up-to-level policy it always holds
that the physical inventory level equals S at the beginning of each period we obtain that the
random number of items left in stock at the end of period k equals 1 (S − Yk )+ . Also, the
random number of back ordered items at the end of period k equals 2 −(S − Yk )− and so
the random inventory holding, penalty and ordering costs of the first n periods are given by
∑n ∑n ∑n
cH (S − Yk )+ − bIS (S − Yk )− + K 1{Yk >0} (3.1)
k=1 k=1 k=1

Since it is easy to verify that

−(S − Yk )− = Yk − S + (S − Yk )+
∑n
this implies by formula (4.1) and Sn := k=1 Yk the total demand up to period n that the
random inventory holding and penalty and ordering costs of the first n periods equal
∑n ∑n
bIS Sn − nbIS S + a1 (S − Yk )+ + K 1{Yk >0} . (3.2)
k=1 k=1

To compute the total profit of the first n periods we observe by the complete back ordering
assumption that we have to replenish at the end of period k the total demand occurring in
period k and this implies with pprofit the profit per sold item that the total profit up to period
n is given by
pprofit Sn (3.3)
Subtracting the profit in relation (4.3) from the costs in relation (4.2) it follows that the total
random costs Cn (S) up to period n equal
∑n ∑n
a2 Sn − nbIS S + a1 (S − Yk )+ + K 1{Yk >0}. (3.4)
k=1 k=1

To calculate the expectation of the above random costs the following assumption is intro-
duced. Observe it is not imposed that the random demands Yn , n ∈ N are independent and
this implies that these demands can be correlated.

Condition 4.1.
The random variables Yn , n ∈ N are identically distributed with right continuous cu-
mulative distribution function FY : [0, ∞) → [0, 1] satisfying 0 ≤ FY (0) < 1 and finite
1 (x)+ :=max{x,0}
2 (x)− :=min{x,0}

18
first moment µY . Moreover, if the demand process has state space [0, ∞) the cumulative
distribution function FY is continuous on [0, ∞).

In case the demand process has state space N ∪ {0} the cumulative distribution function
is represented by
∑n
FY (x) = P {Y1 = k}, n ≤ x < n + 1
k=0

and this implies that the cumulative distribution function is a right continuous step function.
By Condition 4.1 and formula (4.4) we obtain that the total expected costs E(Cn (S)) up
to period n is given by
na2 µY − nbIS S + na1 E((S − Y1 )+ ) + K(1−FY (0))
and this yields with
0 < γ := bIS (cH + bIS )−1 < 1
that the expected average cost c(S) (including the profit due to sales!) equals
a2 µY + K(1 − FY (0)) + bIS (γ −1 E((S − Y1 )+ ) − S) (3.5)
Since for an indivisible item the function
S → (S − x)+
is clearly convex on [0, ∞) for every x ≥ 0 this implies that the function
S → E((S − Y1 )+ )
is also convex on the same set and so it follows that the expected average cost function
c : [0, ∞) → R is convex. To rewrite this expected average cost function we first observe
for any random variable X that the identity
∫ ∞
+
(X) = 1{X≥x} dx
0

holds with 1A denoting the indicator function of the event A and so it follows with the
random variable X replaced by the random variable S − Y1 that
∫ s ∫ s
(S − Y1 ) =
+
1{Y1 ≤S−x} dx = 1{Y1 ≤x} dx. (3.6)
0 0

By this alternative representation we obtain that E((S − Y1 )+ ) equals

∫ S ∫ S ∫ S
E( 1{Y1 ≤x} dx) = E(1{Y1 ≤x} )dx = FY (x)dx (3.7)
0 0 0

and hence the corresponding cost minimization problem is given by

∫ S
−1
a3 + bIS min{γ FY (x)dx − S : S ∈ E} (PB )
0

with a3 := (bIS − pprofit )µY + K(1 − FY (0)). To solve this convex optimization problem
we introduce the inverse function FY← : [0, 1] → [0, ∞) of the cumulative distribution
function FY and this function is given by
FY← (y) := inf{x ≥ 0 : FY (x) ≥ y}.
Since the function FY is by definition right continuous it can be shown (cf.[37]) that the
infimum is attained and hence we can replace inf by min. Moreover, if the cumulative
distribution function FY is continuous and strictly increasing then it is easy to verify (make
a picture!) that FY (F (←
Y (y)) = y. It is now possible to show the following result.

19
Lemma 12 If 0 < γ := bIS (cH + bIS )−1 < 1 it follows that FY← (γ) is an optimal
solution of the cost minimization problem (PB ). Moreover, the optimal cost objective value
v(PB ) (including the profit due to sales!) is given by
∫ γ
v(PB ) = (bIS − pprofit )µY + K(1 − FY (0)) − (cH + bIS ) FY← (x)dx
0

Proof. Consider the objective function f : E → R of the optimization problem (PB )

given by
∫ S
f (S) = γ −1 FY (x)dx − S
0

If the demand has state space [0, ∞) and hence FY is by assumption a continuous function
we have to minimize over [0, ∞) and so by differentiating the above convex objective
function it follows for every S > 0 that

f (1) (S) = γ −1 FY (S) − 1

with f (1) (S) denoting the derivative of the function f at the point S. In case FY (0) ≥ γ
an optimal solution is given by 0 and for FY (0) < γ an optimal solution Sopt satisfies

γ −1 FY (Sopt ) − 1 = 0

Moreover, if the demand has state space N ∪ {0} and hence we have to optimize over
N ∪ {0} we obtain since the cumulative distribution function FY is a step function and so
FY (x) = FY (S) for every S ≤ x < S + 1 that the increment f (S + 1) − f (S) is given by
∫ S+1
f (S + 1) − f (S) = γ −1 FY (x)dx − 1 = γ −1 FY (S) − 1
S

Again for FY (0) ≥ γ it follows that zero is an optimal solution and for FY (0) < γ an
optimal solution Sopt is given by

Sopt = min{S : γ −1 FY (S) ≥ 1}.

For both divisible and indivisible demand this implies that FY← (γ) is an optimal solution
and so we have verified the first part. To prove the second part we observe that Y1 has
the same distribution as FY← (U1 ) with U1 a uniform on (0, 1) distributed random variable.
This implies

E((FY← (γ) − Y1 )+ ) = E((FY← (γ) − F ← (U1 ))+ ) (3.8)

∫ γ
= (FY← (γ) − FY← (x))dx
0

and so we obtain
∫ γ
γ −1 E((FY← (γ) − Y1 )+ ) − FY← (γ) = −γ −1 FY← (x)dx
0

Applying now relation (4.5) it follows that v(PB ) is given by

∫ γ
−1
(bIS − pprofit )µY + K(1 − FY (0)) − bIS γ FY← (x)dx
0

and by the definition of γ the desired result is proved.

First we observe that it seems reasonable to assume that

bIS > cH and bIS > pprofit .

20
Also by the above computation it is clear that the profit pprofit per sold item does not play any
role in the computation of the optimal order-up-to level quantity FY← (γ). This is intuitively
clear since under the complete backorder regime one will always sell on average µY in
each period and this is independent of the way the inventory is controlled. By a similar
argument the fixed ordering cost K > 0 does not play any role in the determination of the
optimal order-up-to-level S. In the next subsection we will analyze the newsboy problem
with lost sales.

3.2 The newsboy problem with lost sales.

If excess demand is lost rather than backordered and we have the same cost components as
in the previous model we again try to determine an expression for the random cost Cn (S)
(including the profit due to sales!). As before we assume that Condition 4.1 holds. The
only difference is now that the number of sold items is not the same as in the previous
model due to the lost sales regime. By this observation it follows using formula (4.1) that
the random inventory holding and penalty and ordering cost of the first n periods are given
by ∑n ∑n
bIS Sn − nbIS S + a1 (S − Yk )+ + K 1{Yk >0} . (3.9)
k=1 k=1
To compute the total ordering and selling costs of the first n periods we observe that we
have to replenish at the end of period k the total demand directly delivered from stock in
period k and this quantity equals min{S, Yk }. This implies that the total profit due to sales
up to period n are given by
∑n
pprofit min{S, Yk }.
k=1

Since the identity

min{S, Yk } = S − (S − Yk )+
holds this profit equals
∑n
npprofit S − pprofit (S − Yk )+ . (3.10)
k=1

Subtracting now the profit in relation (4.10) from the costs in relation (4.9) it follows that
the random costs up to period n is given by
∑n ∑n
bIS Sn − n(pprofit + bIS )S + (a1 + pprofit ) (S − Yk )+ + K 1{Yk >0} .
k=1 k=1

By this observation we obtain that the expected cost E(Cn (S)) up to period n has the form
nbIS µY − n(pprofit + bIS )S + K(1 − FY (0)) + (a1 + pprofit )E((S − Y1 )+ )
and this implies with
0 < γ := (bIS + pprofit )(bIS + cH + pprofit )−1 < 1
that the expected average cost c(S) equals
∫ S
c(S) = bIS µY + K(1 − FY (0)) + (bIS + pprofit )(γ −1 FY (x)dx−S).
0

By the above expression the convex cost minimization problem for the newsboy problem
with lost sales is given by
∫ S
a4 + (bIS + pprofit ) min{γ −1 FY (x)dx−S : S ∈ E} (PL )
0

with a4 := bIS µY + K(1 − FY (0)). Similarly as in Lemma 25 one can show the following
result.

21
Lemma 13 If 0 < γ = (bIS + pprofit )(b2 + cH + pprofit )−1 < 1 then it follows that FY← (γ)
is an optimal solution of the cost minimization problem (PL ). Moreover, the optimal cost
objective value v(PL ) (including the profit due to sales!) is given by
∫ γ
v(PL ) = bIS µY + K(1 − FY (0)) − (bIS + cH + pprofit ) FY← (x)dx.
0

Although the above optimization problem seems to be solved there are still some prob-
lems occurring in practice. In the newsboy model it is assumed that the exact holding and
penalty costs can be determined. However, in practice, most of the time one can only give
estimates of these costs. Based on these estimates one now uses in practice the constant γ
and so this creates an error in the optimal order-up-to-level quantity even if the cumulative
distribution function of the demand is exactly known. Moreover, this also yields a deviation
from the optimal objective value and the error involving this is even more important. In the
next subsection we will discuss the associated sensitivity problem together with the special
case that the demand has a cut-off normal distribution. If this holds one can simplify the
above formulas for the optimal order-up-to-level and the optimal objective value. Before
starting the next subsection we first list the following exercises.
Exercise 14 (exponentially distributed demand) If the distribution of the demand Y1 is
given by FY (x) = 1−exp(−λx) derive then an analytical expression for the optimal order-
up-to-level quantity of a newsboy problem with complete back ordering and lost sales.
Moreover, give an analytical expression for the optimal objective value of the newsboy
problem with complete back ordering and lost sales.
Exercise 15 (geometrically distributed demand) If the distribution of the demand Y1 is
given by FY (x) = 1 − (1 − q)xxy+1 with ⌊x⌋ denoting the lower entier of x, i.e the biggest
integer smaller than or equal to x give then an analytical expression for the optimal order-
up-to-level quantity of the newsboy problem with complete back ordering and lost sales.
Exercise 16 (combination of lost sales and back ordering) If in the newsboy problem each
unit of excess demand is lost with probability 0 < β LS < 1 and back ordered with prob-
ability 1 − β LS compute then the expected average cost and derive an expression for the
optimal order quantity.

3.3 Computation of the optimal order size and sensitivity

analysis.
Clearly it is very important to calculate for given instances the optimal order-up-to-level
and so we also need to pay attention to the computational aspects of the newsboy problem.
To compute this optimal order-up-to-level it is necessary that we are able to evaluate using
a black box the cumulative distribution function FY in any given point. It might be that
this black box itself is an algorithm. An example of such a black box is given by the Fast
Fourier Transform method which can be applied if the state space is given by N ∪ {0} and
it is easy to evaluate the generating function of the one period demand Y1 in the so-called
Fourier frequencies. Using simulation it is also possible by means of the Monte Carlo
method to compute FY← (γ) in case FY is a continuous distribution. We now impose the
following condition.

Condition 4.2.
There exists some black box or oracle to evaluate the cumulative distribution function
FY in any given point.

To calculate the optimal order-up-to-level FY← (γ) we like to apply a standard bisection
procedure (cf.[29]) and to start this procedure we first need to identify an interval containing

22
the optimal value. This means that one needs to know beforehand an upper bound on the
optimal order-up-to-level FY← (γ) and to compute such an upper bound we list the well-
know Markov inequality. Observe this Markov inequality also bounds the number of terms
in the sequence {pk : k ∈ N} with pk := P {Y1 = k} to be computed by the FFT method.

Lemma 17 If the function g : [0, ∞) → R is increasing on [0, ∞) and positive on (0, ∞)

with inverse function g ← : R → [0, 1] given by

g ← (y) := inf{x ≥ 0 : g(x) ≥ y}

and Eg(Y1 ) < ∞ then it follows for every 0 < γ < 1 that
( )
Eg(Y1 )
FY← (γ) ≤ g ←
1−γ

Proof. Since the function g is increasing on [0, ∞) and strictly positive on (0, ∞) and
the random variable Y1 representing the demand is nonnegative we obtain for every x > 0
that ( )
g(x)P {Y1 > x} ≤ E g(Y1 )1{Y1 >x} ≤ Eg(Y1 ).
By this inequality it follows that

Eg(Y1 )
FY (x) = 1 − P {Y1 > x} ≥ 1 − . (3.11)
g(x)

Observe now for x = g ← ( Eg(Y1)

1−γ ) that

Eg(Y1 )
g(x) ≥
1−γ
and this shows by relation (4.11) that

Eg(Y1 )
FY (g ← ( )) ≥ γ.
1−γ
Hence it follows that ( )
Eg(Y1 )
F ← (γ) ≤ g ←
1−γ
and the desired result is verified.
For the specific values of γ associated with the back ordering and lost sales regime we
can construct by the above inequality an upper bound on the optimal order-up-to-level. As
an example we mention in case

E(exp(cY1 )) < ∞ for some c > 0

that the function g : [0, ∞) → R is given by g(x) = exp(cx) and this implies by Lemma
41 that
FY← (γ) ≤ c−1 (log (E (exp(cY1 ))) − log(1 − γ))
Another application of the above lemma is given by E(Y1α ) < ∞ for some α ≥ 1. If this
holds then ( )1
← E(Y1α ) α
FY (γ) ≤ .
1−γ
Observe the distribution of the demand in each of the periods can in some cases be a direct
consequence of the physical interpretation of the demand process. As an example we take
as a demand process a compound Poisson process. By the stationary and independent
increments of this process it follows that the demands in each period are independent and

23
 50

–50

–100

–150

Figure 3.1: Average cost of newsboy problem with lost sales

identically distributed and their distribution is given by a compound Poisson distribution.

In case the individual demand Yn of the nth arriving customer has state space N we may
use Adelson’s recursion scheme (cf.[17]) as a black box. Moreover, since it is well known
(cf. [17]) that the probability generating function P of the total demand in one period is
given by P(z) = exp(−λ(1 − PY (z))) with PY (z) the probability generating function of
the random variable Y1 we may also apply the FFT method to compute the distribution and
use this algorithm as a black box. In the above figure we list the cost function of a newsboy
problem with lost sales in case the demand Y1 has generating function exp(−10(1− 13 (z +
z 2 + z 3 )). Moreover, the order cost c0 equals $2 and the selling price p is given by $5 and
this implies that the holding cost per newspaper at the end of the day is given by $2 and
the penalty cost for each demanded newspaper not available by 3. In the past the so-called
inversion algorithms for the Fourier and Laplace transform were not available and so one
tried to compute the optimal order-up-to-level by assuming that the cumulative distribution
function FY is actually a cut-off normal distribution. This distribution was justified by
referring to the central limit theorem. For this distribution one may obtain more detailed
formulas as shown by the following example.

Example 18 If the random variable Y1 has a cut off normal distribution with parameters
µ > 0 and σ then it follows that

N (x) − N (0)
FY (x) =
1 − N (0)

with N denoting the normal distribution with expectation µ and variance σ 2 . This means
that ∫ x
1 (z − µ)2
N (x) = √ exp(− )dz
σ 2π −∞ 2σ 2

24
It is easy to verify for any 0 < y < 1 that

FY← (y) = N ← ((1 − N (0))y + N (0)) (3.12)

Moreover, if the function Φ denotes the unit normal distribution with zero expectation and
variance 1 we obtain that N (x) = Φ( x−µσ ) and this yields for every 0 < y < 1 that

N ← (y) = σΦ← (y) + µ (3.13)

Substituting formula (4.13) into formula (4.12) it follows that

FY← (y) = σΦ← ((1 − N (0))y + N (0)) + µ (3.14)

and so we obtain that

−µ −µ
Sopt = FY← (γ) = σΦ← ((1 − Φ( ))γ + Φ( )) + µ
σ σ
Since in most practical cases µ is much bigger than σ it follows that Φ( −µ
σ ) is almost zero
and so
Sopt ≈ σΦ← (γ) + µ
In Section 1.5.5 of (cf.[17]) an easy algorithm to compute the inverse of the unit normal
distribution is listed and by applying this we obtain for given numerical values for γ the
numerical values for Sopt . To compute the optimal objective value as mentioned in Lemma
25 and 26 we observe by formula (4.14) that
∫ γ ∫ γ
← −µ −µ
FY (x)dx = σ Φ← ((1 − Φ( ))x + Φ( ))dx + µγ
0 0 σ σ

Moreover, since µ is bigger then σ implying Φ( −µ

σ ) is negligible we obtain that
∫ γ ∫ γ
FY← (x)dx ≈ σ Φ← (x)dx + µγ and µY ≈ µ
0 0

It is now easy to verify that

∫ ∫ Φ← (γ)
γ
1 y2 − exp( 21 Φ← (γ)2 )
Φ← (x)dx = √ y exp(− )dy = √
0 2π −∞ 2 2π
and this finally implies for both the optimization problem (PB ) and (PL ) that the optimal
objective value is given by
k1 µ + k2 σ + K
with k1 and k2 properly chosen.

We will now discuss the sensitivity of the optimal objective solution value of the news-
boy problem with respect to the holding cost cH and the penalty cost bitemshort . As in the
Economic Order Quantity model one is interested in the flatness of the objective function
around the optimal solution FY← (γ) of the optimization problem

min{γ −1 E((S − Y1 )+ ) − S : S ∈ E}. (GP )

with respect to the parameter γ. Suppose we do not have exact information on the value of
γ and so we use as an estimation the value γ est . Using this value and Lemma 25 or 26 the
decision maker uses as the optimal order-up-to-level quantity the value FY← (γ est ) instead
of the correct but unknown optimal value FY← (γ). In the next lemma we will evaluate the
difference of the objective function of the optimization problem (GP ) in both values.

25
Lemma 19 If we consider the optimization problem (GP ) with objective function f : E →
R given by
f (S) := γ −1 (E((S − Y1 )+ ) − S
then it follows that the value FY← (γ) is an optimal solution of optimization problem (GP )
and
γ
0 ≤ f (FY← (γ est )) − v(GP ) ≤ (1 − est )(FY← (γ) − FY← (γ est )).
γ
Proof. The first part is already verified in the first part of Lemma 25 and so we only
need to check the upper bound in the above inequality. To verify this upper bound we
observe by relation (4.8) that
∫ β
E((FY← (β) − Y1 )+ ) = (FY← (β) − FY← (x))dx
0

for every 0 < β < 1. By the above formula we obtain that f (FY← (γ est )) equals
(∫ γ est )
−1 γ
γ ← ← ←
(FY (γ est ) − FY (x))dx − γ est FY (γ est ) + ( est − 1)FY← (γ est )
0 γ
∫ γ est
γ
= −γ −1 FY← (x))dx + ( est − 1)FY← (γ est )
0 γ
and this shows that the nonnegative difference f (FY← (γ est )) − v(GP ) is given by
(∫ γ ∫ γ est )
γ
γ −1 FY← (z)dz − FY← (z)dz + ( est − 1)FY← (γ est )
0 0 γ
Since the inverse cumulative distribution function FY← : [0, 1] → [0, ∞) is increasing we
obtain by considering separately the cases γ est > γ and γ est ≤ γ in the previous formula
the desired result.
By the above result it follows for the newsboy problem under the complete backo-
rdering or lost sales regime that by Lemma 25 an upper bound on the absolute difference
between the true optimal objective value and the function value evaluated at the estimated
optimal value FY← (γ est ) is given by
γ est
bIS (1 − )(FY← (γ) − FY← (γ est ))
γ
In case we assume additionally that γ est = (1 + e)γ the upper bound on the absolute
difference reduces to
ebIS (FY← (γ) − FY← ((1 + e)γ))
Unfortunately the above upper bound cannot be computed since the decision maker does
not know the real value of γ but the above from suggests for reasonable cumulative distri-
bution functions with a strictly positive density that an conservative estimate of the absolute
difference is of the order e2 in case γ est = (1 + e)γ (see exercise 20) In the next section we
will discuss the service measure approach applied to the newsboy problem with complete
backordering.

Exercise 20
1. Show for a cumulative distribution function FY with a strict positive density f that
∫ β
FY← (β) − FY← (0) = (f (FY← (z))−1 dz
0

for every 0 < β < 1 (Hint:Use the relation FY (FY← (β)) = β to compute the
derivative of FY← (β))).

26
 ←
2. If γ est := (1+e)γ and FY has a strictly positive density show that (1− γ est
γ )(FY (γ)−
FY← (γ est )) is of the order e2 .

Exercise 21

1. Suppose we consider a newsboy problem under the lost sales regime with K = 0, the
order cost per newspaper equals $2 and the selling price $5. Moreover, the random
variable Y1 has a compound Poisson distribution with generating function given by
1
exp(−10(1 − (z + z 2 + z 3 )).
3
Give an interpretation of the random variable Y1 and determine the penalty and
holding cost per newspaper.

2. Write a computer program in Maple using both Adelson’s recursion scheme and
the FFT method to determine the optimal-order-up-to-level S and give a plot of the
cumulative distribution function and the cost function (including profit due to sales)
as a function of the order-up-to-level S.

3.4 Service measures and the newsboy problem.

In case we apply a service measure approach to the newsboy problem with complete back
ordering and order-up-to-level S and satisfying Condition 4.1 the penalty cost bitemshort
is replaced by some service measure. This approach is mostly applied in practice since it
is in general impossible to estimate the penalty cost. Using this approach there are two
different service measures to be used. If we impose the no-stock-out service measure it is
immediately clear that the random number of no stockouts during the first n periods are
given by ∑n
1{Yk ≤S}
k=1

This implies that the no-stock-out probability represented by Pno−stockout (S) is given by
∑n
E( k=1 1{Yk ≤S} )
Pno−stockout (S) = lim = FY (S)
n↑∞ n

Introducing now the service level constraint Pno−stockout (S) ≥ α with α specified by the
management we need to solve by relation (4.5) the problem

−pprofit µY + K(1 − FY (0)) + cH v(PBS )

with v(PBS ) given by

min{E((S − Y1 )+ ) : FY (S) ≥ α, S ∈ E} (PB S)

Since the objective function in optimization problem (PB S) is clearly an increasing func-
tion of S the optimal solution of the above optimization problem is given FY← (α). This
implies by Lemma 25 that we choose the same optimal order-up-to-level as in the newsboy
problem with complete back ordering and penalty cost bIS given by cH α(1 − α)−1 . By
taking the more appropriate fillrate service measure it follows by the definition of the con-
trol rule that the random number of items directly delivered from stock during the first n
periods is given by ∑n
min{S, Yk }
k=1

27
This implies that the fraction of demand directly delivered from stock represented by
Pf illrate (S) equals
∑n
E( k=1 min{S, Yk })
lim = µ−1Y E(min{S, Y1 })
n↑∞ E(Sn )
Introducing now the service level constraint Pf illrate (S) ≥ β with β specified by the
management we need to solve by relation (4.5) the problem

−pprofit µY + K(1 − FY (0)) + cH v(PBF )

with v(PBF ) given by

min{E((S − Y1 )+ ) : µ−1
Y E(min{S, Y1 }) ≥ β, S ∈ E} (PB F )

For the optimization problem (PB F ) one can show the following result.
←
Lemma 22 It follows that FeY (β) is an optimal solution of the optimization problem
←
(PB F ) with FeY denoting the inverse function of the cumulative distribution function
∫ x
FeY (x) := µ−1
Y (1 − FY (z))dz
0

Moreover, if the state space is given by [0, ∞) the optimal objective value v(PB F ) is given
by
←
v(PB F ) = −(pprofit + cH β)µY + K(1 − FY (0)) + cH FeY (β)

Proof. Since
min{S, Y1 } = S − (S − Y1 )+
we obtain
∫ S
E(min{S, Y1 }) = S − E((S − Y1 )+ ) = (1 − FY (z))dz. (3.15)
0

Since the objective function of the optimization problem (PB F ) is increasing we obtain
that the optimal solution Sopt must satisfy
∫ S
Sopt := min{S : (1 − FY (z))dz > µY β}
0

and this shows the first part. Moreover, in case the state space is given by [0, ∞) we obtain
that the inequality in the optimization problem (PB F ) is binding at the optimal solution
and so
←
E(min{FeY (β), Y1 }) = µY β
By this observation and relation (4.15) it follows that
← ←
E((FeY (β) − Y1 )+ ) = FeY (β) − µY β

and this yields the second part.

The cumulative distribution function FeY is called the equilibrium distribution associ-
ated with the demand distribution FY . This concludes our discussion on the use of ser-
vice measures in the newsboy problem. In the next section we will introduce a general
framework for analyzing the most important single-item inventory models. Observe this
framework is based on the theory of regenerative processes and enables us to derive ex-
act formulas for the different cost structures instead of the approximations used in most of
the literature on inventory control models. Examples of those approximations are given in
Chapters 7, 8, 9 of (cf.[5]). Before presenting these inventory models we review the basic
results of regenerative processes.

28
Chapter 4

The Theory of Regenerative

Processes.

The theory of regenerative processes plays a major role within the analysis of single item
inventory control models. Since for our purpose it is sufficient to consider only pure re-
generative processes we will not discuss delayed regenerative processes. For a complete
survey on regenerative processes the reader is referred to [35]. Observe the set T either de-
notes [0, ∞) or N∪{0}. Applied to inventory control models this distinction means that for
T = [0, ∞) the inventory/penalty costs continuously change according to changes in the
inventory level while for T = N ∪ {0} these costs are charged at period ending inventory
levels. A simple example of the last case is given by the newsboy problem of the previous
section.

4.1 Regenerative processes and the renewal argument.

Let {Xn : n ∈ N} be a sequence of nonnegative independent random variables and suppose
the length Xn denotes the inter arrival time between the (n − 1)th and nth event in some
specific probability problem. Observe the state space S of the random variables Xn , n ∈ N
is either N ∪ {0} or [0, ∞). As an example we mention the arrival process of customers and
an event represents the arrival of a customer. Letting S0 := 0 and Sn = X1 + .. + Xn it is
now clear that Sn denotes the random time the nth event occurs.

Definition 23 The stochastic process S := {Sn : n ∈ N ∪ {0}} is called a renewal

process1 if the nonnegative random variables {Xn : n ∈ N} are independent and the
random variables {Xn : n = 2, 3, ..} (not necessarily X1 !) are identically distributed with
(right-continuous) cumulative distribution function F. The renewal process is called pure2
if the random variable X1 has the same distribution as the random variables Xn , n =
2, 3, ..and it is called delayed 3 if the (right-continuous) cumulative distribution function G
of the random variable X1 is different from the cumulative distribution function F.

Observe the cumulative distribution function G in Definition 23 is called the delay dis-
tribution4 and the random times Sn , n ∈ N are called renewal points.5 . To avoid pathologi-
cal cases we always assume that the cumulative distribution function F satisfies F (0) < 1.
1 renewal process
2 pure renewal process
3 delayed renewal process
4 delay cumulative distribution function
5 renewal points

29
However, it might occur that the cumulative distribution function F is defective6 or equiv-
alently F (∞) < 1 and this corresponds to the case that P {Xn = ∞} > 0 for every
n ∈ N\{1}. If this happens we call the associated pure or delayed renewal process termi-
nating7 or defective and in this case it follows immediately that F (0) < F (∞) < 1. If
F (∞) = 1 the pure or delayed renewal process is called proper8 . An example of a termi-
nating renewal process occurs in case a person is making a phone call to a company and
every time the line is busy with probability b > 0. In case the time between two consec-
utive attempts to phone the company is given by some random variable and these random
variables are independent and identically distributed we can model the times at which the
company is called as a defective renewal process. Clearly if the person has success in his
attempt to get connected to the company the renewal process stops and this means that the
next ”inter arrival” time has the value ∞. By our assumption this happens with probabil-
ity 1 − b. An important stochastic process related to the renewal process is given by the
stochastic counting process9 N = {N(t) : t ≥ 0} with state space N ∪ {0} and defined by

N(t) := largest integer n satisfying Sn ≤ t.

Clearly for every t ≥ 0 the random N(t) denotes the number of renewals occurring before
time t and formally this is given by

N(t) = max{n ∈ N ∪ {0} : Sn ≤ t}.

Since F (0) < 1 the sample paths of the stochastic process N are step functions and these
step functions are right continuous with left hand limits. A stochastic process of which
the sample paths satisfy these conditions is called cadlag10 (“continu à droite,limites à
gauche”). Moreover, by the nonnegativity of the inter arrival times Xn , n ∈ N an alterna-
tive representation of the random variable N(t) is given by (make a drawing!)
∑∞
N(t) = 1{Sn ≤t} (4.1)
n=1

with 1A denoting the indicator function of the event A. This means that the random variable
1A equals one if the event A occurs and equals zero otherwise. To determine the cumulative
distribution function of the random variable N(t) we observe that the event {N(t) ≥ n} is
the same as the event that the nth renewal occurs before time t and so we obtain

{N(t) ≥ n} = {Sn ≤ t}. (4.2)

By relation (4.2) it follows that

P{N(t) ≥ n} = P{Sn ≤ t}

and since
P{N(t) = n} = P{N(t) ≥ n} − P{N(t) ≥ n + 1}
we obtain
P{N(t) = n} = P{Sn ≤ t} − P{Sn+1 ≤ t}. (4.3)
In the next definition the so-called renewal function associated with the counting process
N = {N(t) : t ≥ 0} is introduced.

Definition 24 For any renewal process S = {Sn : n ∈ N ∪ {0}} and the associated
counting process N = {N(t) : t ≥ 0} the function R : [0, ∞) → [0, ∞] given by

R(t) := E(N(t))
6 defective cumulative distribution
7 terminating (pure or delayed) renewal process
8 proper (pure or delayed) renewal process
9 counting process
10 cadlag stochastic process

30
is called a renewal function11 . It is called a pure renewal function12 if the renewal process
S is pure and it is called a delayed renewal function13 if the renewal process S is delayed.

By the definition of the stochastic counting process N = {N(t) : t ≥ 0} given by

relation (4.3) we observe for every t ≥ 0 that
∑∞ ∑∞
R(t) = E(N(t)) = E( 1{Sn ≤t} ) = P{Sn ≤ t}. (4.4)
n=1 n=1

To write out the probability P {Sn ≤ t} it follows for a pure renewal process that

P{Sn ≤ t} = F n∗ (t)

with F n∗ denoting the n-fold convolution of the cumulative distribution function F and so
by relation (4.4) we obtain that the pure renewal function R equals
∑∞
R(t) = F n∗ (t). (4.5)
n=1

Moreover, in case we are dealing with a delayed renewal process with delay cumulative
distribution function G it follows that

P{Sn ≤ t} = (G ∗ F (n−1)∗ )(t)

for every n ∈ N (observe F 0∗ (t) = 1 for every t ≥ 0!) and so in this case the delayed
renewal function Rd (observe a delayed renewal function is sometimes denoted with a
subscript d) is given by
∑∞
Rd (t) = (G ∗ F (n−1)∗ )(t) = G(t) + (G ∗ R)(t) (4.6)
n=1

with R denoting the pure renewal function associated with the cumulative distribution func-
tion F . It is possible that the expected number R(t) of renewals up to time t equals infinity
or equivalently that the series in relation (4.4) diverges. This holds for example if F (0) = 1.
However, if F (0) < 1 one can show the following result.

Lemma 25 If F (0) < 1 then it follows that the delayed or pure renewal function R satisfies
R(t) < ∞ for every t ≥ 0.

Proof. We only give a proof for a pure renewal process since by relation (4.6) the
proof for a delayed renewal process is similar with some obvious modifications. If the pure
renewal process S is terminating we obtain by the independence of the random variables
Xn , n ∈ N that

F n∗ (t) ≤ P{X1 < ∞, ..., Xn < ∞} = (F (∞))n

for every t ≥ 0 and hence by relation (4.5) it follows that

∑∞ F (∞)
R(t) ≤ (F (∞))n = < ∞.
n=1 1 − F (∞)

To show the result for a proper and pure renewal process we observe since F (0) < 1 and
F is right continuous that

limx↓0 F (x) =: F (0+) = F (0) < 1

11 renewal function
12 pure renewal function
13 delayed renewal process

31
and so one can find some b > 0 satisfying F (b) < 1. Moreover, for any α > 0 we obtain

E(exp(−αX1 )) = E(exp(−αX1 )1{X1 ≤b} ) + E(exp(−αX1 )1{X1 >b} )

≤ P{X1 ≤ b} + exp(−αb))P{X1 > b}

and so it follows that

E(exp(−αX1 )) ≤ F (b) + exp(−αb)(1 − F (b)) < 1. (4.7)

Furthermore we obtain by Markov’s inequality and the random variables Xn are indepen-
dent and identically distributed that

exp(−αt)P{Sn ≤ t} ≤ E(exp(−αSn )) = (E(exp(−αX1 ))n (4.8)

and this shows by relations (4.4), (4.7), and (4.8) that the value R(t) is bounded from above
by
∑∞ n exp(αt)E(exp(−αX1 ))
exp(αt) (E(exp(−αX1 ))) = <∞
n=1 1 − E(exp(−αX1 ))
showing the desired result.
If the defective renewal process is delayed with delay cumulative distribution function
G then it follows for every n ∈ N that

P{Sn ≤ t} ≤ P{X1 < ∞, .., Xn < ∞} = G(∞)(F (∞))n−1

and so for the associated delayed and defective (increasing) renewal function we obtain for
every t > 0 that
∑∞
Rd (t) ≤ G(∞)(F (∞))n−1 = G(∞)(1 − F (∞))−1 .. (4.9)
n=1

Similarly for a defective pure renewal process it follows for every t > 0 that

R(t) ≤ F (∞)(1 − F (∞))−1 . (4.10)

By relations (4.9) and (4.10) it is clear for a defective (pure or delayed) renewal process that
the renewal function R is uniformly bounded. To give a more elementary expression than
relation (4.4) for the (proper or defective) renewal function is in general difficult and so we
have to use numerical techniques to compute the value of R(t) for some given t. However,
if the inter arrival times are Erlang distributed it is possible to derive a relatively easy
formula for the proper and pure renewal function. This is discussed on page 177 of [10]. Up
to now we only introduced the definition of a proper and defective renewal function and did
not use its specific probabilistic interpretation. This probabilistic interpretation is given by
the next result and is a direct consequence of the so-called renewal argument14 . Observe we
restrict ourselves for the moment to cumulative distribution functions F satisfying F (0) =
0 and having a continuous density f on (0, ∞). It is now possible to prove the following
result.

Lemma 26 If the (proper or defective) distribution F of the interarrival times has a con-
tinuous density f on (0, ∞) and satisfies F (0) = 0 then the pure renewal function R is a
solution of the so-called renewal equation given by
∫ t
R(t) = F (t) + R(t − x)f (x)dx.
0
14 renewal argument

32
 Proof. Clearly for every t > 0 it follows that

R(t) = E(N(t)1{X1 ≤t} ) + E(N(t)1{X1 >t} ). (4.11)

By the definition of the counting process N it is clear that N(t)1{X1 >t} = 0 and this yields
by relation (4.11) and conditioning on the time of occurrence of the first renewal point that
∫ t
R(t) = E(N(t)1{X1 ≤t} ) = E(N(t)|X1 = x)f (x)dx.
0

By the definition of the renewal counting process it follows for every x ≤ t that
e − x)|X1 = x)
E(N(t)|X1 = x) = E(1 + N(t (4.12)
e − x)) = 1 + R(t − x)
= E(1 + N(t

with Ne := {N(t)
e : t ≥ 0} a renewal counting process independent of X1 and having the
e is
same distribution as the original renewal counting process N (Actually the process N
the renewal counting process starting at time x!). This shows
∫ t ∫ t
R(t) = (1 + R(t − x))f (x)dx = F (t) + R(t − x)f (x)dx
0 0

and so we have verified that the renewal process satisfies the above integral equation.
The probabilistic argument leading to relation (4.12) is called the renewal argument. It
will be proved later in a more general setting that the above result holds for any (proper
or defective) cumulative distribution function F on [0, ∞) and that the renewal function is
the unique solution of the above integral equation.
∫ t However, if the cumulative distribution
function F has no density the Riemann integral 0 R(t − x)f (x)dx does not exist and so
we wonder what to do for all cumulative distribution functions F with no density. In this
case the integral equation of Lemma 26 is formally represented by
∫ t
µ(t) = F (t) + µ(t − x)dF (x)
0−

with the above improper integral defined by

∫ t ∫ t
µ(t − x)dF (x) = limϵ↑0 µ(t − x)dF (x) (4.13)
0− ϵ

and the distribution F extended to (−∞, ∞) by setting F (x) = 0 for every x ≤ 0. Without
going into details we only mention that for every ϵ < 0 the integral
∫ t
h(t − x)dF (x)
ϵ

with h some given function is called the Riemann-Stieltjes integral15 (cf.[4]). This integral
can be defined in a similar way as the Riemann integral by dividing the interval [ϵ, t] into
the finite set of equally spaced points

ϵ := a0 < a1 < .. < an := t

and considering the upper and lower sums of the form

∑n ∑n
Uk (F (ak+1 ) − F (ak )) and Lk (F (ak+1 ) − F (ak ))
k=0 k=0
15 Riemann-Stieltjes integral

33
with
Uk := supak ≤y≤ak+1 {h(t − y)} and Lk := inf ak ≤y≤ak+1 {h(t − y)}.
By increasing the number of equally spaced points and letting the width of the above parti-
tion go to zero one can show (cf.[39]) for any function h bounded on [0, t] with only a finite
number of discontinuities on this interval which do not coincide with the discontinuities of
the function F that the upper and lower sums converge to the same limit and this limit is
denoted by ∫ t
h(t − x)dF (x).
ϵ
Applying now relation (4.13) yields an interpretation of the so-called improper integral
∫ t
h(t − x)dF (x).
0−

In case the cumulative distribution function F has a density f on (0, ∞) and F (0) = 0 it
can be shown by the mean value theorem16 (cf.[39]) that
∫ t ∫ t
h(t − x)dF (x) = h(t − x)f (x)dx
0− 0

and so the Rieman-Stieltjes integral can be seen as a classical Riemann integral. Also, if
0 < F (0) < 1 and f is the density on (0, ∞) we obtain
∫ t ∫ t
h(t − x)dF (x) = h(t)F (0) + h(t − x)f (x)dx
0− 0+

with ∫ ∫
t t
h(t − x)f (x)dx := limϵ↓0 h(t − x)f (x)dx.
0+ ϵ
Moreover, if F is an arithmetic distribution given by
∑⌊x⌋
F (x) := pk
k=0

for every x ≥ 0 then by the above definition it follows that17

∫ t ∑⌊t⌋
h(t − x)dF (x) = h(t − n)pn
0− n=0

The definition of a Riemann-Stieltjes integral not only holds for an (increasing) cumulative
distribution function F but for any increasing function on (0, ∞) bounded on finite intervals
and so by Lemma 25 we can define the integral
∫ t
h(t − x)dR(x)
0−

with R the renewal function for any cdf F satisfying F (0) < 1. In case the cumulative
distribution function F associated with the renewal function R satisfies F (0) = 0 and has
a continuous density f on (0, ∞) then it follows that R(0) = 0 and R is differentiable with
continuous density r and as before we obtain that
∫ t ∫ t
h(t − x)dR(x) = h(t − x)r(x)dx
0− 0
16 interpretation Riemann-Stieltjes integral for continuous density
17 Interpretation Rieman-Stieltjes integral for arithmetic distribution

34
 ∑⌊x⌋ distribution function F the renewal function R is a
Moreover, for an arithmetic cumulative
step function given by R(x) = k=0 rk and in this case it follows that
∫ t ∑⌊t⌋
h(t − x)dR(x) = h(t − k)rk .
0− k=0

We can now introduce using the definition of the Riemann-Stieltjes integral the Laplace-
Stieltjes transform Lh of any finite valued increasing function h : [0, ∞) → R and this is
given by
∫ ∞ ∫ t
Lh (α) := exp(−αx)dh(x) := limt↑∞ exp(−αx)dh(x)
0− 0−

with the complex number α chosen in such a way that the above integral has a finite value.
This means in particular by relation (4.5) that for every α ∈ C with Re(α) > 0 the Laplace-
Stieltjes transform ∫ ∞
LR (α) := exp(−αx)dR(x)
0−

of the pure renewal function reduces to

∑∞ ∫ ∞ ∑∞
LR (α) = exp(−αx)dF n∗ (x) = E(exp(−αSn )).
n=1 0− n=1

Since it is well-known that

E(exp(−αSn )) = (E(exp(−αX1 )))n = (LF (α))n

this finally yields that LR (α) is given by the simple expression

LR (α) = LF (α)(1 − LF (α))−1 .

By relation (4.6) it follows similarly that the Laplace-Stieltjes transform LRd (α) of the
delayed renewal function Rd is given by

LRd (α) = LG (α)(1 − LF (α))−1

with LG (α) denoting the Laplace-Stieltjes transform of the delay cumulative distribution
function G. As for cumulative distribution functions the function
∫ t
t→ h(t − x)dF (x)
0−

is a so-called convolution of the functions F and h18 and a shorthand notation is given by
h ∗ F . Since integral equations of the above convolution-type occur very often in applied
probability models we introduce the following definition.

Definition 27 If the function F is a (proper or defective) cumulative distribution function

on [0, ∞) and h : [0, ∞) → R denotes some function then the integral equation
∫ t
µ(t) = h(t) + µ(t − x)dF (x), t ≥ 0
0−

or in shorthand notation
µ(t) = h(t) + (µ ∗ F )(t)
is called a renewal-type equation19 .
18 convolution of functions F and h
19 renewal type equation

35
 Although in Definition 27 the Riemann-Stieltjes integral is used and so also arithmetic
cumulative distributions are included we list for clarity also the definition of a discrete
renewal-type equation.

Definition 28 If the function F is a (proper or defective) arithmetic distribution on N∪{0}

and h : N ∪ {0} → R is some sequence then the integral equation
∑n
µ(n) = h(n) + µ(n − k)pk , n ∈ N ∪ {0}
k=0

or in shorthand notation
µ(n) = h(n) + (µ ∗ F )(n)
is called a discrete renewal type equation.20

The next result shows that every renewal-type equation has a unique solution and this
solution has an elementary form. Before discussing this result we list the following defini-
tion.

Definition 29 A function g : [0, ∞) → R is called bounded on finite intervals21 if for every

x > 0 it follows that sup 0≤y≤x {|g(y)|} < ∞.

The next result is of extreme importance since in almost all applications within applied
probability it is first verified using the so-called renewal argument that the function under
consideration satisfies a renewal-type equation with the functions F and h known. After
checking this the next result is used to write down the explicit solution of this equation in
terms of the pure renewal function R and the function h.

Theorem 30 If F is some (proper or defective) distribution on [0, ∞) satisfying F (0) < 1

and h : [0, ∞) → R is a function bounded on finite intervals then there exists a unique
solution v : [0, ∞) → R within the class of functions bounded on finite intervals satisfying
the renewal-type equation

µ(t) = h(t) + (µ ∗ F )(t), t ≥ 0.

Moreover, this solution is given by

v(t) = h(t) + (h ∗ R)(t)

with R denoting the pure renewal function associated with the cumulative distribution F .
In case the cumulative distribution function F has a density f on (0, ∞) and satisfies
F (0) = 0 then the unique solution v is given by
∫ t
v(t) = h(t) + h(t − x)r(x)dx
0

with r(x) the derivative of the pure renewal function R at the point x.

Proof. In order to prove the above result we first need to verify for any cumulative
distribution function F that the above function v is indeed bounded on finite intervals and
satisfies the renewal-type equation. After that we show that necessarily the difference func-
tion v1 − v2 with the functions vi , i = 1, 2 bounded on finite intervals and satisfying both
the renewal-type equation must be the zero function and this shows that the renewal-type
equation has a unique solution within the class of functions bounded on finite intervals. To
start with the proof that the function v is finite on bounded intervals we observe using h
20 discrete renewal-type equation
21 function bounded on finite intervals

36
is a function bounded on finite intervals that for every t ≥ 0 there exists some Ut < ∞
satisfying
sup0≤y≤t {|h(y)|} ≤ Ut .
Since the pure renewal function R is increasing this implies (remember the definition of
the Riemann-Stieltjes integral!) that sup0≤y≤t {|(h ∗ R)(y)|} is bounded from above by
∫ y ∫ t
sup0≤y≤t { |h(y − x)|dR(x)} ≤ Ut sup0≤y≤t { dR(x)} = Ut R(t).
0− 0−

By Lemma 25 it follows that R(t) is finite and so by the above inequality the function h ∗ R
is bounded on finite intervals. Using this observation it is clear that the function

v(t) = h(t) + (h ∗ R)(t)

is also bounded on finite intervals. To check that the function v : [0, ∞) → R given by

v(t) = h(t) + (h ∗ R)(t)

is a solution of the renewal-type equation µ(t) = h(t) + (µ ∗ F )(t) we observe by the

closed form expression for v that

(v ∗ F )(t) = (h ∗ F )(t) + ((h ∗ R) ∗ F )(t) (4.14)

= (h ∗ (F + (R ∗ F )))(t)

Observe now for every t ≥ 0 and relation (4.5) that

∑∞
F + (R ∗ F )(t) = F (t) + F k∗ (t) = R(t)
k=2

and this shows by relation (4.14) that

(v ∗ F )(t) = (h ∗ R)(t). (4.15)

Since v = h + h ∗ R we obtain by relation (4.15) that

(v ∗ F )(t) = v(t) − h(t)

and this verifies that the function v : [0, ∞) → R is a solution of the renewal-type equation
µ = h+µ∗F . Finally, if vi , i = 1, 2 are functions bounded on finite intervals and solutions
of the renewal-type equation µ = h + µ ∗ F then it follows that the difference function
w : [0, ∞) → R given by
w(t) = v2 (t) − v1 (t)
is bounded on finite intervals and satisfies for every t ≥ 0 the equation

w(t) = v2 (t) − v1 (t) = (v2 ∗ F )(t) − (v1 ∗ F )(t) = (w ∗ F )(t).

Iterating now we obtain for every n ∈ N that w(t) = (w ∗ F n∗ )(t) and since w is bounded
on finite intervals this yields for every t ≥ 0 that there exist some Ut such that
∫ t
|w(t)| ≤ |w(t − x)|dF n∗ (x) ≤ Ut F ∗ (t)
0−

Since n ∈ N is arbitrary and t is fixed it follows by Lemma 25 that F n∗ (t) ↓ 0 as n ↑ ∞ and

so we obtain that |w(t)| = 0 for every t > 0 or equivalently w is the zero function. In case
the distribution F has a density the last formula follows immediately from the observation
about the Riemann-Stieltjes integral before Definition 27 and this finishes the proof of the
above result.

37
 Clearly for every proper or defective arithmetic distribution p:= {pk : k ∈ N ∪ {0}}
satisfying p0 < 1 it follows for every |z| < 1 that |P(z)| < 1 with P denoting the
generating function of the distribution p. Consider now the discrete renewal-type equation

µ(n) = h(n) + (µ ∗ F )(n), n ∈ N ∪ {0} (4.16)

with h:= {h(n) : n ∈ N ∪ {0}} some real valued sequence satisfying

∥h∥∞ := sup {|h(n)|} < ∞.

n∈N∪{0}

By relation (4.16) it follows for every |z| < 1 that the generating function Rµ (z) of the
sequence µ satisfies
∑∞
Rµ (z) = Rh (z) + (µ ∗ F )(n)z n
n=0

with Rh (z) denoting∑the generating function of the sequence h. Reversing the summation
∞
signs we obtain that n=0 (µ ∗ F )(n)z n equals
∑∞ ∑∞
µ(n − k)z n−k pk z k = Rµ (z)P(z)
k=0 n=k

and this shows

Rµ (z) = Rh (z) + Rµ (z)P(z) (4.17)
for every z satisfying |z| < 1. Since |P(z)| < 1 for every |z| < 1 this yields by relation
(4.17) that
Rµ (z) = Rh (z)(1 − P(z))−1 (4.18)
and so it follows that the generating function of the solution
∑n
v(n) = h(n) + h(n − k)rk
k=0

of the discrete renewal-type equation given by relation (4.16) is easy to compute. By rela-
tion (4.1) the value rn has the probabilistic interpretation
∑n ∑n
rn = R(n) − R(n − 1) = E( 1{Sk =n} ) = P{Sk = n}
k=1 k=1

for every n ∈ N while by the independence of the identically distributed random variables
Xn , n ∈ N the value r0 equals
∑n ∑n p0
P {Sk = 0} = P {X1 = 0, .., Xk = 0} = .
k=1 k=1 1 − p0
By relation (4.18) it is now easy to compute part of the sequence µ by means of the F F T
method. In case the cumulative distribution function F with F (0) = 0 has a density f on
(0, ∞) and one likes to compute the value of the renewal function then one can use so-
called quadrature formulas from numerical integration (cf.[13], [33], [2]) to approximate
the integral (R ∗ F )(t). Clearly it can happen that a simple discretization method like the
trapezoidal rule does not work due to round off errors and heavily oscillating densities and
so one needs to stay critical. Observe there exist more sophisticated and accurate techniques
to compute the renewal function by means of inverting Laplace transforms numerically or
to solve the renewal equation by means of spline functions. However, these techniques
are beyond the scope of these notes. To introduce the spline approach we observe that the
renewal equation and more generally the renewal-type equation is a very special case of
the so-called linear Volterra equation of the first kind and in numerical mathematics there
are so-called spline methods to solve these equations numerically. The interested reader
is referred to [16] and [2]. By applying a technique from numerical mathematics one can

38
only approximate the renewal function on a interval of small length and so we do not know
anything about the tail of the renewal function. Fortunately it is possible to give asymptotic
results for the renewal function associated with a proper (pure or delayed) renewal process.
An example is the so-called weak renewal theorem22 given by the next result.

Theorem 31 For any proper pure renewal process S = {Sn : n ∈ N ∪ {0}} and 0 <
E(X1 ) < ∞ it follows that

R(t)
limt↑∞ = (E(X1 ))−1 .
t
Moreover, for any proper delayed renewal process S = {Sn : n ∈ N ∪ {0}} and 0 <
E(X2 ) < ∞ it follows that

Rd (t)
limt↑∞ = G(∞)(E(X2 ))−1 .
t
Proof. We only give a proof of the above result for a proper renewal process since it is
relatively easy to verify the result for a proper delayed renewal process using the first part
and the observation that the renewal function for a proper delayed renewal process is given
by (see relation (4.6))
Rd (t) = G(t) + (G ∗ R)(t).
Observe for a proper pure renewal process it follows that
∑N(t)+1 ∑∞
t ≥ E( Xn ) = E(Xn 1{N(t)+1≥n} )
n=1 n=1

and since {N(t) + 1 ≥ n} = {Sn−1 ≤ t} and Xn , n ∈ N are independent and identically

distributed we obtain
∑∞ ∑∞
E(Xn 1{N(t)+1≥n} ) = E(X1 ) E(1{N(t)+1≥n} )
n=1 n=1
= E(X1 )E(N(t) + 1)

(This is known as Wald’s identity!). Hence for every t ≥ 0 it holds that

t ≥ E(X1 )(R(t) + 1)

and this shows

R(t)
lim inf t↑∞ ≥ (E(X1 ))−1 .
t
Now we are done by showing that
R(t)
lim supt↑∞ ≤ (E(X1 ))−1 .
t
(a)
Consider now for every a > 0 the renewal process {Sn : n ∈ N} with
∑n (a) (a)
S(a)
n := Xk and Xk := min{Xk , a}.
k=1

If we introduce the associated counting process N(a) := {N(a) (t) : t ≥ 0} given by

N(a) (t) := sup{n ∈ N ∪ {0} : S(a)

n ≤ t}

then it follows for every realization (make a drawing!) that

N(t) ≤ N(a) (t)

22 weak renewal theorem

39
for every t ≥ 0 and hence
E(N(t)) ≤ E(N(a) (t)) (4.19)
for every t ≥ 0. Moreover, by the definition of the renewal process N(a) and Wald’s identity
we obtain that
(a)
E(min{X1 , a})(1 + E(N(a) (t)) = E(SN(a) (t)+1 ) ≤ t + a

and by this observation it follows that

E(N(a) (t))
lim supt↑∞ ≤ (E(min{X1 , a}))−1 .
t
Hence by relation (4.19) and a > 0 arbitrary this yields

E(N(t)) E(N(a) (t))

lim supt↑∞ ≤ lim supt↑∞ ≤ (E(min{X1 , a}))−1
t t
This shows
E(N(t))
lim supt↑∞ ≤ inf a>0 {(E(min{X1 , a}))−1 } = (E(X1 ))−1
t
and hence the desired inequality is verified.
In the above result we never used that Xn is concentrated on [0, ∞) or on N ∪ {0}
and so the above result holds for all cumulative distribution functions. We will now con-
sider in detail pure and delayed regenerative processes and their properties. To start with
these processes we first introduce the definition of a simplified version of a so-called pure
regenerative process and for convenience these stochastic processes are called simplified
pure regenerative processes. Observe the set T denotes the time axis and is either given by
T = [0, ∞) (continuous time) or T = N ∪ {0} (discrete time).

Definition 32 A stochastic process X = {X(t) : t ∈ T } with metric state space E is

called a simplified pure regenerative process if there exists some positive finite constant
σ ∈ T such that for every n ∈ N ∪{0} the distribution of the shifted stochastic process
Xnσ := {X(t + nσ) : t ∈ T } is independent of n.

Observe the points nσ, n ∈ N are called the regeneration points of the simplified pure
regenerative process X. Moreover, although we assume that the state space of the stochastic
process X is a metric space in most applications this state space is a subset of Rm and so
if the reader does not like to think of a metric state space he/she can always replace this
without loss of generality by a subset of the finite dimensional space Rm . In this case
the stochastic process X is actually a vector valued process. In the next example we will
discuss a well-known model from inventory control which is actually a simplified pure
regenerative process.

Example 33 An example of a simplified pure regenerative process is given by the so-called

(R, S) single item inventory control model with T = [0, ∞), zero leadtime and complete
backordering. To define this model we first introduce some stochastic input processes.
Clearly for any inventory control model we need to introduce an increasing stochastic
demand process D := {D(t) : t ≥ 0} given by

D(t) := total demand for item up to time t.

It is always assumed that the demand process D is cadlag and the state space of this
demand process is either N ∪ {0} (divisible item) or [0, ∞) (indivisible item). In particular

40
for the (R, S) model the demand process D is given by a compound Poisson process23 .
This means that the random variable D(t) is represented by
∑N(t)
D(t) = Yn
n=0

with Y0 := 0 and Yn , n ∈ N the stochastic demand of the nth arriving customer and N :=
{N(t) : t ≥ 0} a Poisson arrival process having rate λ > 0. Furthermore, the random
variables Yn , n ∈ N are independent and identically distributed and the Poisson arrival
process N is independent of the demands Yn , n ∈ N. We introduce now the stochastic
process I := {I(t) : t ≥ 0} with

I(t) := actual stock on the shelves at time t

This process is called the on-hand stock process24 (Dutch: aanwezige voorraad proces)
and since the demand process D is cadlag the on-hand stock process I is also cadlag.
Since the demand of every customer arriving during a zero stock is completely backordered
another important inventory process is given by the so-called stochastic backordering pro-
cess25 (Dutch: het naleveringen proces) B := {B(t) : t ≥ 0} given by

B(t) := amount of backordered items at time t.

Again this proces is cadlag and using the definition of the on-hand stock process and the
backordering proces we introduce the so-called netstock26 or net inventory process IN :=
{IN(t) : t ≥ 0} (Dutch: het netto voorraad process) given by

IN(t) := I(t) − B(t).

Also the netstock process is cadlag and due to the complete backordering assumption it
follows that
B(t) = − min{IN(t), 0} and I(t) = max{IN(t), 0}.
Hence we may replace the on-hand stock process I and the backordering process B by the
netstock process IN and clearly random costs up to time t related to keeping inventory and
not serving customers are determined by the behaviour of the netstock process IN up to
the same time t. Since the lead time is zero and so there exist no items ordered and not
yet delivered it also follows that the decision to order at some time t should be based on
the realization of the net stock process IN up to time t. The ordering cost depend on the
control rule and so also these costs are determined by the netstock process. Observe now
that the (R, S) inventory control rule is used and this means by definition that every R time
units the net stock process is inspected and a replenishment order is issued if the level of
the net stock process is below level S. Hence an order at time nR is triggered if and only
if
IN(nR−) := lim IN(t) < S.
t↑nR

Moreover, the size of the order is such that all excess demand up to that time will be back-
ordered and the net stock process is raised up to level S. By this observation it follows for
every sample path that IN(nR) = S (assume the stochastic process IN starts at level S
at time 0!) and using the stationary and independent increments of a compound Poisson
process we obtain (make a drawing of a sample path!) that the netstock process IN is a
simplified regenerative process with σ in Definition 32 equal to R if the system is controlled
by the (R, S) rule.
23 compound Poisson process
24 on-hand stock process
25 backordering process
26 net stock process

41
 As already observed in Example 33 the penalty costs for not serving customers imme-
diately and the inventory holding costs for keeping inventory are completely determined by
the netstock process and this shows in case there exist some costrate function f and

Cinvp (t) := inventory holding and penalty costs up to time t

that ∫ t
Cinvp (t) = f (IN(v))dv.
0
This proves for every t ≥ 0 that
∫ t
Cinvp (t + R) − Cinvp (t) = f (IN(v + R))dv.
0

and so the outcome of the random variable Cinvp (t + R) − Cinvp (t) depends completely
on the outcome of the process {IN(v + R) : 0 ≤ v ≤ t}. Due to the zero lead time the
ordering cost process Cord = {Cord (t) : t ≥ 0} given by

Cord (t) := ordering costs up to time t

is completety determined by IN and in particular for every t ≥ 0 the random variable

Cord (t + R)− Cord (R) completely depends on the process {IN(v + R) : 0 ≤ v ≤ t}.
Introducing now the combined cost process C := {C(t) : t ≥ 0} with

C(t) := Cinvp (t) + Cord (t) = total cost of system to time t

we obtain that the stochastic process C has increasing sample paths, C(0) = 0 and C(t +
R) − C(R) is completely determined by the stochastic process {IN(v + R) : 0 ≤ v ≤ t}.
Without loss of generality we may also assume that the sample paths of the cost process C
are right continuous (this is a regularity condition) and so the cost process C belongs to the
following class of stochastic processes (cf.[7]).

Definition 34 A stochastic process A := {A(t) : t ≥ 0} is called an additive functional of

the simplified pure regenerative process X27 with regeneration points nσ, n ∈ N if A(0) =
0 and the stochastic process A has increasing right continuous sample paths. Moreover,
for any t > 0 the random variable A(t) is completely determined by the stochastic process
{X(s) : 0 ≤ s ≤ t} while for every t ≥ 0 the stochastic increment A(t + σ) − A(σ) is
completely determined by the shifted stochastic process {X(s + σ) : 0 ≤ s ≤ t}.

Although the above definition is not very precise it means that there exists some real
valued and maybe very complicated deterministic function g with domain the sample paths
of the stochastic process X satisfying

A(t) = g({X(s) : 0 ≤ s ≤ t}) (4.20)

and
A(t + σ) − A(σ) = g({X(s + σ) : 0 ≤ s ≤ t}) (4.21)
for every t ≥ 0. Since the stochastic process X is a simplified regenerative process with
regeneration points nσ, n ∈ N this shows combining relations (4.20) and (4.21) that

A(t + σ) − A(σ) =d g({X(s) : 0 ≤ s ≤ t}) = A(t). (4.22)

In particular by relation (4.22) it follows for every n ∈ N that

A((n + 1)σ) − A(σ) =d A(nσ)

27 additive functional of a simplified pure regenerative process

42
and this shows in case E(A(σ)) < ∞ that by induction E(A(nσ)) < ∞ for every n ∈ N.
Since A is increasing we obtain that E(A(t)) < ∞ for every t ≥ 0 and applying now
relation (4.22) this yields
E(A(t + σ) − A(σ)) = E(A(t)) < ∞ (4.23)
Using relation (4.23) the next important result is easy to prove.
Theorem 35 If the stochastic process A = {A(t) : t ≥ 0} is an additive functional of a
simplified pure regenerative process X with regeneration points nσ, n ∈ N and E(A(σ)) <
∞ then it follows that E(A(t)) < ∞ for every t ≥ 0 and
E(A(t)) E(A(σ))
limt↑∞ =
t σ
Proof. For any t > σ it follows trivially that
E(A(t)) = E(A(σ)) + E(A(t) − A(σ))
and this shows by relation (4.23) that
E(A(t)) = E(A(σ)) + E(A(t − σ)
for every t > σ. Applying the above argument iteratively we obtain
E(A(t)) = nE(A(σ)) + E(A(t − nσ))
for every nσ ≤ t < (n + 1)σ and since A is increasing implying A(t − nσ) ≤ A(σ) for
every nσ ≤ t < (n + 1)σ this finally yields
nE(A(σ)) ≤ E(A(t)) ≤ (n + 1)E(A(σ)) (4.24)
for every nσ ≤ t < (n + 1)σ. Dividing now E(A(t)) by t and applying relation (4.24) we
obtain for t ↑ ∞ the desired result.
The above theorem is very important and is one of the most useful theorems for simpli-
fied regenerative processes.
Example 36 Considering Example 33 and using Theorem 35 we obtain that the average
cost of a (R, S) inventory control model with lead time 0 and complete backordering is
given by
E(C(R))
R
with C(R) denoting the random cost up to time R. It follows immediately that
∫ R
C(R) = K1{order issued at time R} + f (IN(v))dv
0

and since the event {order issued at time R} is the same as the event {a customer arrives
within [0, R]} we obtain
(∫ )
R
E(C(R)) = KP{N(R) ≥ 1} + E f (IN(v))dv
0
∫ R
= K(1 − exp(−λR)) + E(f (IN(v)))dv.
0

Moreover, since at time 0 the inventory level equals S and (make a picture!) IN(v) =
S − D(v) for every 0 < v < R this finally shows that the average cost of a (R, S) control
rule is given by
∫R
K(1 − exp(−λR)) + 0 E(f (S − D(v)))dv
R

43
 Up to now we only discussed simplified pure regenerative processes with deterministic
regeneration points. A more general definition of a regenerative process is now given by
the following. Observe this definition also includes Markov processes and as we will show
a lot of applied stochastic processes are processes satisfying the next definition.

Definition 37 A stochastic process X = {X(t) : t ∈ T } with metric state space E is

called a pure regenerative process if there exists some increasing pure renewal process
{σ n , n ∈ N ∪ {0}} with state space T and σ 0 := 0 such that for each n ∈ N ∪ {0} the
post σ n -process Xσn := {X(t + σ n ) : t ∈ T } is independent of the random variables
σ 0 , ..., σ n and the distribution of the post-σ n stochastic process Xσn is independent of n.

An important consequence of the definition of a pure regenerative process is given by

the following result. This result will be used to show that in all inventory control models
with complete backordering the so-called inventory position process is a pure regenerative
process.

Theorem 38 If the stochastic process X = {X(t) : t ∈ T } is a pure regenerative process

with metric state space E and increasing pure renewal process {σ n , n ∈ N ∪ {0}} and the
function Φ : E → R is Borel measurable then the stochastic process Φ ◦ X : = {Φ(X(t)) :
t ∈ T } is a pure regenerative process with the same increasing pure renewal process.
Moreover, if T = [0, ∞) and the function Φ : E → R is continuous then the stochastic
process Φ ◦ X is cadlag if the stochastic process X is cadlag.

Observe Borel measurable is a technical condition and if the reader does not feel com-
fortable with this condition he may think of a function with at most a finite number of
discontinuities. In the next example we consider an example within queueing theory of a
pure regenerative process.

Example 39 An example of a pure regenerative process is given by a so-called G/G/1

queueing model. In this case at time point 0 a customer enters an empty system and the
arrival process of customers is given by a counting process N associated with a renewal
process. Also the random variables Bn , n ∈ N with Bn denoting the service time of the
nth arriving customer are independent and identically distributed and these service times
are independent of the renewal process. Finally there is one server to serve customers with
first-come first-served queueing discipline and a waiting room with infinite capacity. If we
consider for this model the stochastic process X = {X(t) : t ≥ 0} given by

X(t) := number of customers in system at time t.

then this process is a pure regenerative process with regeneration points σ n , n ∈ N rep-
resenting the random times that for the nth time a customer enters an empty system. In
order to guarantee that the renewal process {σ n : n ∈ N ∪ {0}} is pure one needs to
impose that the system will not explode or equivalently that λE(B1 ) < 1 with λ denoting
the arrival rate of the counting process. Remember λ equals the reciprocal of the expected
interarrival time. It is hopefully clear that the value λE(B1 ) denotes the expected amount
of work which enters the system per unit of time and to avoid exploding this should be lower
than the capacity of the server. For more details on special cases of the above queueing
model the reader is referred to Chapter 8 of [38] and for a detailed overview of regenerative
processes occurring within single server queueing theory one should consult [26].

As for simplified pure regenerative proceses we now introduce the definition of an

additive functional for pure regenerative processes.

Definition 40 A stochastic process A := {A(t) : t ≥ 0} is called an additive functional

of the pure regenerative process X28 with regeneration points σ n , n ∈ N ∪{0} if A(0) = 0
28 additive functional of a pure regenerative process

44
and the stochastic process A has increasing right continuous sample paths. Moreover, for
every t ≥ 0 the stochastic increment A(t + σ n ) − A(σ n ), n ∈ N ∪ {0} is completely
determined by the shifted stochastic process Xσn := {X(s + σ n ) : 0 ≤ s ≤ t}.

Although the above definition is not very precise it means that there exists some real
valued and maybe very complicated deterministic function g with domain the sample paths
of the stochastic process X satisfying

A(t) = g({X(s) : 0 ≤ s ≤ t}) (4.25)

and
A(t + σ n ) − A(σ n ) = g({X(s + σ n ) : 0 ≤ s ≤ t}) (4.26)
for every t ≥ 0 and n ∈ N ∪ {0}. Since {σ n : n ∈ N ∪ {0}} is a renewal process and the
shifted stochastic process Xσk has the same distribution as X it follows by conditioning
on the random variable σ k − σ k−1 that

A(σ k ) − A(σ k−1 ) = g({X(s + σ k−1 ) : 0 ≤ s ≤ σ k − σ k−1 }) (4.27)

=d g({X(s) : 0 ≤ s ≤ σ 1 }) = A(σ 1 )

for every k ∈ N and by a similar argument using Xσk−1 is independent of σ 0 , ..., σ k−1
that
A(σ k ) − A(σ k−1 ) independent of σ 0 , ..., σ k−1 . (4.28)
As for additive functionals for a simplified pure regenerative process one can show the
following result.

Lemma 41 If the stochastic process A = {A(t) : t ≥ 0} is an additive functional of

the pure regenerative process X with regeneration points σ n , n ∈ N ∪{0} satisfying 0 <
E(σ 1 ) < ∞ and E(A(σ 1 )) < ∞ then it follows that E(A(t)) < ∞ for every t > 0.

Proof. Introducing the counting process N = {N(t) : t ≥ 0} given by

N(t) := sup{n ∈ N : σ n ≤ t}

it follows by the monotonicity of the additive functional A that

(∑ )
N(t)+1
E(A(t)) ≤ E A(σ k ) − A(σ k−1 ) (4.29)
k=1
(∑∞ )
=E (A(σ k ) − A(σ k−1 ))1{N(t)+1≥k}
k=1

By the definition of N(t) we obtain

1{N(t)+1≥k} = 1{N(t)≥k−1} = 1{σk−1 ≤t}

and this implies using relations (4.27),(4.28) and (4.29) that

∑∞
E(A(t)) ≤ E(A(σ k ) − A(σ k−1 ))E(1{N(t)+1≥k} )
k=1
= E(A(σ 1 ))(1 + E(N(t))).

Since by assumption 0 < E(σ 1 ) < ∞ we obtain by Lemma 25 that E(N(t)) is finite for
every t ≥ 0 and substituting this in the above relation yields the desired result.
It is now easy to show one of the most important results in applications for pure regener-
ative processes and this result generalizes the result for additive functionals of a simplified
regenerative process.

45
Theorem 42 If the stochastic process A = {A(t) : t ≥ 0} is an additive functional
of the pure regenerative process X with regeneration points σ n , n ∈ N ∪{0} satisfying
0 < E(σ 1 ) < ∞ and 0 < E(A(σ 1 )) < ∞ then it follows that E(A(t)) < ∞ for every
t ≥ 0 and
E(A(t)) E(A(σ 1 ))
limt↑∞ = .
t E(σ 1 )
Proof. The first part is already verified in Lemma 25 and so we only verify the second
part. It follows for every t > 0 that

E(A(t)) = E(A(min{t, σ 1 })) + E(A0 (t)) (4.30)

with the increasing stochastic process A0 given by A0 := {A(t) − A(min{t, σ 1 }) : t ≥

0}. To analyse the second term of relation (4.30) we observe that

(A(t) − A(min {t, σ 1 })1{σ1 >t} = 0

and this shows

∫ t
E(A0 (t)) = E(A0 (t)1{σ1 ≤t} ) = E(A0 (t)|σ 1 = x)dFσ (x). (4.31)
0−

Moreover, by the definition of the stochastic process A0 we obtain for every x ≤ t that

E(A0 (t)|σ 1 = x) = E(A(t − x + σ 1 ) − A(σ 1 )|σ 1 = x) (4.32)

and since A(t − x + σ 1 ) − A(σ 1 ) is independent of σ 1 and has the same distribution as
A(t − x) this yields
E(A0 (t)|σ 1 = x) = E(A(t − x))
for every x ≤ t. Introducing now the function h : [0, ∞) → R given by

h(t) := E(A(min{t, σ 1 }))

we have shown for every t ≥ 0 that

∫ t
E(A(t)) = h(t) + E(A(t − x))dFσ (x) (4.33)
0

Applying now Lemma 25 and Theorem 30 yields

∫ t
E(A(t)) = E(A(min{t, σ 1 })) + E(A(min{t − x, σ 1 }))dR(x). (4.34)
0

and to analyze the behaviour of the above expectation for t ↑ ∞ we introduce the function
G : [0, ∞) → [0, 1]) given by

E(A(min{t, σ 1 })) h(t)

G(t) = := . (4.35)
E(A(σ 1 )) E(A(σ 1 ))
Since the sample paths of the (increasing) additive functional process A are right continu-
ous and 0 < E(A(σ 1 )) < ∞ the function G is a proper cumulative distribution function
and this shows by relations (4.34) and (4.35) that

E(A(t)) = E(A(σ 1 ))Rd (t) (4.36)

with Rd satisfying ∫ t
Rd (t) = G(t) + G(t − x)dR(x).
0

46
Hence Rd is a delayed renewal function with delay distribution G and so we obtain by the
weak renewal theorem and relation (4.36) that

E(A(t)) Rd (t) E(A(σ 1 ))

limt↑∞ = E(A(σ 1 )) lim = .
t t↑∞ t E(σ 1 )

This verifies the desired result and we are done.

Since mostly a cost process is an additive functional it is clear that the average cost of
a pure regenerative process29 equals the ratio of the expected costs in the first cycle [0, σ 1 ]
and the expected cycle length E(σ 1 ). This result is of extreme importance and is used every
time to compute the average costs of a specific system. Sometimes we are not interested in
the average cost but in the so-called discounted cost. The expected discounted cost over an
infinite horizon of a stochastic process X with a so-called nonnegative cost rate function
f 30 is now given by
∫ ∞
cf (α) := E( exp(−αt)f (X(t))dt), α > 0.
0

For pure regenerative processes one can now show very easily the following result for the
discounted costs.

Theorem 43 For any pure regenerative process X = {X(t) : t ≥ 0} with regeneration

points σ n , n ∈ N ∪ {0} and cf (α) < ∞ it follows that
∫σ
E( 0 1 exp(−αt)f (X(t))dt)
cf (α) = .
1 − E(exp(−ασ 1 )

Proof. By the definition of an integral it follows immediately that

∫ σ1 ∫ ∞
cf (α) = E( exp(−αt)f (X(t))dt) + E( exp(−αt)f (X(t))dt)
0 σ1

Looking
∫ ∞ at the last term and reversing the expectation with the integral sign we obtain that
E( σ1 exp(−αt)f (X(t))dt) equals
∫ ∞
exp(−αt)E(exp(−ασ 1 )f (X(t + σ 1 )))dt.
0

Since the stochastic process X is a pure regenerative process it follows that the random
variable σ 1 is independent of the random variable X(t + σ 1 ) for any t > 0 and this implies

E(exp(−ασ 1 )f (X(t + σ 1 ))) = E(exp(−ασ 1 ))E(f (X(t + σ 1 )))

Moreover, by the regenerative character of the stochastic process X the stochastic variable
X(t + σ 1 ) has the same distribution as X(t) and this yields

E(f (X(t + σ 1 ))) = E(f (X(t))

for every t > 0. By the previous observations it follows that

∫ ∞ ∫ ∞
E( exp(−αt)f (X(t))dt) = E(exp(−ασ 1 )) exp(−αt)E(f (X(t)))dt
σ1 0
∫ ∞
= E(exp(−ασ 1 ))E( exp(−αt)f (X(t))dt)
0
29 average cost pure regenerative process
30 cost rate function

47
and this finally implies
∫ σ1
cf (α) = E( exp(−αt)f (X(t))dt) + E(exp(−ασ 1 ))cf (α)
0

showing the desired result.

This concludes our section on regenerative processes. For a more general approach to
the discounted cost function and its relation with the average cost the reader is referred to
[23]. In the next section we will apply the above results to single item inventory control
models.

48
Chapter 5

Basics of Stochastic Single Item

Inventory Control.

In single item inventory control the decision maker is dealing with two objectives. First
of all, he likes to control the cost of keeping inventory, and secondly, he likes to maintain
a certain service level. In order to control the cost of keeping inventory and maintaining
a certain service level the decision maker faces two main questions. These questions are
when to order and how much to order (cf.[5]). Clearly the question of how much to order
depends on the demand process the decision maker is expecting in the future. Moreover,
if at some time in the future the inventory system is out of stock and during that period
demand is arriving the decision maker needs to know whether this demand is lost or can
be backordered. To model this we need to distinguish between lost sales1 or complete
backordering 2 . In the lost sales case it is assumed that any demand arising when the
system is out of stock is lost while in the complete backordering case it is assumed that
this demand will be backordered and is filled as soon as a new replenishment arrives. In
this section we will only consider the complete backordering case. Clearly the amount
of backordered items depends on the time it takes before a new order arrives and so we
introduce the following assumption with respect to the arrival of orders. As before the set
T either denotes [0, ∞) or N ∪ {0}

Assumption:
If a replenishment order is issued at some time t ∈ T this order arrives at the facility at
time t + L with L a given fixed constant belonging to T .

As already mentioned the constant L given by the above assumption is called the lead
time3 and when L equals 0 this corresponds to an instantaneous replenishment.4 Observe
this occurred in the newsboy problem. Moreover, since the lead time L is a given constant
for all replenishment orders it follows that an order placed earlier than another order will
arrive sooner at the facility and so no overtaking of orders will take place. This property
plays an important role in the mathematical analysis of the basic single item inventory
models. A generalization to random lead times is given by the next one.

Assumption:
If the nth replenishment order is issued at time t ∈ T then this order arrives at the
facility at time t + Ln with Ln a nonnegative random variable concentrated on T and
1 lost sales
2 complete backordering
3 lead time
4 instantaneous replenishment

49
independent of the order size and the order moment. The random variables Ln , n ∈ N are
identically distributed with right continuous cumulative distribution function FL .

If we consider stochastic lead times then it is possible that a replenishment order issued
earlier will arrive later at the facility and so overtaking of orders can take place. With
respect to stochastic lead times we always assume within the mathematical analysis of
inventory control models that the following approximation assumption holds.

Approximation assumption:
If the lead times are random variables then it is assumed that no overtaking of orders
takes place.

To show that overtaking for only two orders can take place with positive probability we
list the following exercise.

Exercise 44 In case replenishment order i = 1, 2 is issued at time t1 with t1 < t2 and the
random lead times Li are exponentially distributed with parameter λ > 0 compute then
the probability that order 2 will arrive sooner at the facility as order 1.

To describe the behaviour of the inventory level we also need to introduce the demand
process for a single item. In this section we will consider a stochastic demand process
D = {D(t) : t ∈ T } with state space [0, ∞) or N ∪ {0} and the random variable D(t)
represents the total demand up to time t. By its definition the process D is increasing and
it is assumed for T = [0, ∞) that this process is cadlag. In case the state space of the
stochastic demand process D is [0, ∞) the item is called indivisible5 (example: gasoline)
and in case the state space is N ∪ {0} the item is called divisible6 . Moreover, to define the
cost structure of an inventory model governed by some inventory control rule to be specified
later we also need to introduce the following different inventory processes defined on the
same state space as the demand process (cf.[17]). We first consider the stochastic process
I : = {I(t) : t ∈ T } with

I(t) := actual stock on the shelves at time t.

This stochastic process is called the on-hand stock process7 (Dutch: het aanwezige voor-
raad proces) and since we assume for T = [0, ∞) that the demand process is cadlag this
process is also cadlag for T = [0, ∞). Another inventory process is given by the stochas-
tic backordering process8 B : = {B(t) : t ∈ T } (Dutch: het naleveringen proces) with
B(0) := 0 and
B(t) := amount of items backordered at time t.
Again for T = [0, ∞) this process is cadlag. Clearly this represents at time t the amount of
demand not yet delivered. Using now the definition of the on-hand stock process and the
backordering process we obtain the so-called net stock9 or net inventory stochastic process
IN : = {IN(t) : t ∈ T } (Dutch: het netto voorraad proces) given by

IN(t) := I(t) − B(t)

for every t ∈ T . Again the stochastic process IN is cadlag for T = [0, ∞) and since we
only consider single item inventory control models with complete backordering it follows
that
B(t) = − min{IN(t), 0} and I(t) = max{IN(t), 0}. (5.1)
5 indivisibleitem
6 divisibleitem
7 on-hand stock process
8 backordering process
9 net stock process

50
By the definition of the net stock process it is clear that the cost of an inventory system
governed by some inventory control rule depends solely on the stochastic process IN. It
will be shown for all classical single-item inventory control models with properly defined
stochastic demand processes that a shifted version of the stochastic process IN is a (sim-
plified) pure regenerative process and so we can apply the results of the previous section.
To prove such a result we first need to relate the net stock process to the so-called inventory
position process. To define this process we first introduce the cadlag stochastic process
O : = {O(t) : t ∈ T } given by

O(t) := total amount ordered and not yet delivered at time t.

The inventory position process10 IP = {IP(t) : t ∈ T } (Dutch: het voorraadpositie

proces) is now given by
IP(t) := IN(t) + O(t).
In case the lead time L is zero it is clear that the stochastic process O is identically zero
and so the net stock process coincides with the inventory position process. Moreover, it is
also clear that our decision to order should depend on the level of the inventory position
process. The next result relates the inventory position process to the net stock process in
case of complete backordering with a fixed lead time L. Its proof is easy and is listed for
completeness (cf.[18]).

Theorem 45 If complete backordering occurs with a fixed lead time L and the stochastic
process D is cadlag then it follows for every t ≥ 0 that

IN(t + L) = IP(t) − (D(t + L) − D(t)) P-almost surely

Proof. Since all orders have a fixed lead time L ∈ T it follows that in the interval
(t, t + L] all orders O(t) outstanding at time t did arrive. Moreover, replenishment orders
issued after time t did not arrive at the facility and by this observation we obtain that

O(t) = addition to net stock in (t, t + L] (5.2)

Moreover, since the demand process D is cadlag it follows that D(t + L)−D(t) represents
the total demand occurring in the interval (t, t + L] and this implies by relation (5.2) and
the complete backordering assumption that

IN(t + L) = IN(t) + O(t) − (D(t + L) − D(t)) (5.3)

By the definition of the inventory position process we know that

IP(t) = IN(t) + O(t)

and by relation (5.3) this shows the desired result.

By the above proof it is also easy to verify that the relation between the inventory
position process and the net stock process also holds for a stochastic lead time L satisfying
the approximation assumption and so we obtain the relation

IN(t + L) = IP(t) − (D(t + L) − D(t)) P-almost surely

for every t > 0. Observe at this point we need the no overtaking assumption. Also by the
above result we obtain that the netstock process is properly on [L, ∞). The next result is
very important and shows the connection between the theory of regenerative processes and
the different inventory processes. Before mentioning this result we introduce the following
basic condition.(cf.[9]).

10 inventory position process

51
Condition 46 The vector-valued stochastic process
X = {(IP(t), D(t + L) − D(t)) : t ∈ T }
with some initial distribution at time 0 and deterministic lead time L and complete back-
ordering is a (simplified) pure regenerative process with the increasing pure renewal se-
quence {σ n : n ∈ N ∪ {0}} ⊆ T N∪{0} satisfying 0 < E(σ 1 ) < ∞ and this renewal
sequence contains as a subset the times that a replenishment order is issued.

Using the above condition it is easy to show the following result.

Theorem 47 If Condition 46 holds then it follows that the shifted net stock process INL :=
{IN(t + L) : t ∈ T }, the shifted backordering process BL := {B(t + L) : t ∈ T } and the
shifted on-hand stock process IL := {I(t + L) : t ∈ T } are (simplified) pure regenerative
processes with the same increasing renewal sequence.
Proof. To verify that the stochastic process INL is regenerative consider the continuous
function Φ : R2 → R given by Φ(x, y) = x − y. This shows by Theorem 45 that
Φ(X(t)) = IP(t) − (D(t + L) − D(t)) = IN(t + L)
for every t ∈ T and so by Theorem 38 and condition 46 the desired result follows. To show
that the shifted backordering process BL is also a (simplified) pure regenerative process
we observe by relation (5.1) that
B(t + L) = − min{IN(t + L), 0}
and applying now the first part and Theorem 38 proves the desired result. A similar proof
applies to the shifted on-hand stock process IL and we are done.
A similar result also holds for a random lead time L if the no overtaking assumption
holds. Consider now a single item inventory control model with complete backordering,
T = [0, ∞), a deterministic lead time and a given control rule. If in this model the cost
of a replenishment order equals K > 0 and the inventory/penalty costs are represented by
some cost rate function f : E → [0, ∞) then we introduce the following stochastic cost
processes. First introduce the increasing cost process Cord := {Cord (t) : t ≥ 0} given by
Cord (t) := total ordering cost up to time t. (5.4)
Since by Theorem 45 the netstock process is properly defined on the time axis [L, ∞) we
will only consider the penalty and inventory holding costs after time L and so we introduce
the increasing cost process Cinvp := {Cinvp (t) : t ≥ 0} given by
Cinvp (t) := inventory/penalty cost on (L, t + L]. (5.5)
Since the inventory/penalty costs are represented by a cost rate function f it follows that
∫ t+L ∫ t
Cinvp (t) = f (IN(v))dv = f (IN(v + L))dv. (5.6)
L 0

Adding the two stochastic cost processes we finally introduce the increasing stochastic cost
process C = {C(t) : t ≥ 0} given by
C(t) := Cinvp (t) + Cord (t) (5.7)
and so the average cost of the system (if it exists) has the form
Φaverage := limt↑∞ t−1 E(C(t)). (5.8)
To simplify the notation in the next result we introduce for convenience
pr := P{replenishment occurs in a cycle}. (5.9)
Observe the next result is a key result in this chapter.

52
 (∫ σ )
Theorem 48 If Condition 46 holds and E 0 1 f (IP(0) − D(v + L))dv is finite then it
follows that the average cost Φaverage exists and
( (∫ σ1 ))
Φaverage = E(σ 1 )−1 Kpr + E f (IP(0) − D(v + L))dv .
0

Proof. Since the ordering cost are completely determined by the inventory position
process IP and by Condition 46 the process IP is a (simplified) pure regenerative process
it follows that the increasing process Cord given by relation (5.4) is an additive functional
of the inventory position process IP. Moreover, since the sequence {σ n : n ∈ N ∪ {0}}
of regeneration points contains as a subsequence the points in time that a replenishment
order is issued it follows that in the (n + 1)th cycle [σ n , σ n+1 ), n ∈ N ∪ {0} at most one
replenishment order is issued and if so this happened at time σ n . Therefore the ordering
costs in each cycle are bounded above by K and this shows using Theorem 42 and 0 <
E(σ 1 ) < ∞ that
limt↑∞ t−1 E(Cord (t)) = E(σ 1 )−1 Kpr .
Also by Condition 46 and Theorem 45 it follows that the shifted net stock process INL is a
(simplified) pure regenerative process and this shows by relation (5.6) that the process Cinvp
is an additive functional of INL . To analyse the process Cinvp we observe by Condition 46
that the first replenishment order after time 0 is issued at time σ 1 or later and so no order
is issued within the interval (0, σ 1 ). This shows for every 0 < v < σ 1 that the random
variable IP(v) is completely determined by IP(0) and the demand up to time v and in
particular it follows
IP(v) = IP(0) − D(v)
for every 0 ≤ v < σ 1 . This implies by Theorem 45 that
IN(v + L) = IP(v) − (D(v + L) − D(v)) = IP(0) − D(v + L) (5.10)
for every 0 ≤ v < σ 1 and this yields
∫ σ1 ∫ σ1
f (IN(v + L))dv = f (IP(0) − D(v + L))dv.
0 0

Applying now relation (5.9) and Theorem 42 yields

(∫ σ1 )
limt↑∞ t−1 E(Cinvp (t)) = E(σ 1 )−1 E f (IP(0) − D(t + L))dt . (5.11)
0

and this implies using relations (5.11) and (5.8) the desired result.
For T = N ∪ {0} the next result can be verified by means of a similar proof.
(∑ )
σ 1 +L−1
Theorem 49 If Condition 46 holds and E n=L f (IP(0) − D(n)) is finite then it
follows that the average cost Φaverage exists and
( (σ +L−1 ))
1∑
−1
Φaverage = E(σ 1 ) Kpr + E f (IP(0) − D(n)) .
n=L

Every time we like to derive the average cost of a given single item inventory control
model with complete backordering Theorem 48 yields an expression for this average cost
if this model satisfies Condition 46.
In inventory control one is also interested in other measures. The first measure to be
discussed is the average number of backorders per unit of time given by (if it exists!)
∫ t+L
−1
Bbackorder/time := limt↑∞ t E( B(v)dv) (5.12)
L

where B = {B(t) : t ≥ 0} denotes the backordering process. Applying now Theorem 48

one can show the following result for this measure.

53
 (∫ σ1 )
Lemma 50 If Condition 46 holds and E 0
max{D(t + L) − IP(0), 0}dt is finite then
it follows that
(∫ σ1 )
−1
Bbackorder/time = E(σ 1 ) E max{D(t + L) − IP(0), 0}dt .
0

Proof. If the function f : R → R is given by f (x) = − min{x, 0} = max{−x, 0}

then it follows by relation (5.1) that
∫ t+L (∫ t )
E( B(v)dv) = E f (IN(v + L))dv
L 0

for every t > 0. This shows applying relation (5.12) and Theorem 48 with K = 0 that
(∫ σ1 )
Bbackorder/time = E(σ 1 )−1 E max{D(t + L) − IP(0), 0}dt
0

and so we have verified the desired result.

Another measure is the average physical stocklevel per unit of time given by (if it
exists!)
∫ t+L
−1
Bstocklevel/time := limt↑∞ t E( I(v)dv) (5.13)
L
where I = {I(t) : t ≥ 0} denotes the on-hand stock process. Applying again relation (5.1)
one can show in a similar way as in Lemma 50 the following result for this measure.
(∫ σ )
Lemma 51 If Condition 46 holds and E 0 1 max{D(t + L) − IP(0), 0}dt is finite then
it follows that
(∫ σ1 )
−1
Bstocklevel/time = E(σ 1 ) E max{IP(0) − D(t + L), 0}dt .
0

Proof. Apply a similar proof as in Lemma 50 with f : R → R given by f (x) =

max{x, 0}.
The popularity of the the average physical stocklevel per unit of time and the number of
backorders per unit of time is due to the following easy observation. If we introduce linear
penalty and linear holding costs then the cost rate function f has the form

f (x) = cH max{x, 0} − p min{x, 0}.

This shows by Theorem 48 and Lemma 50 and 51 that the average cost equals

E(σ 1 )−1 Kpr + cH Bstocklevel/time + pBbackorder/time

and so we obtain a linear combination of these measures.

Another performance measure used within inventory control is the ready rate measure
representing the fraction of time that the stocklevel is positive and so this means (if it
exists!) (∫ )
t+L
−1
Preadyrate := limt↑∞ t E 1{I(v)>0} dv . (5.14)
L

Again by Theorem 48 and relation (5.1) one can show the following result for the ready
rate measure.

Lemma 52 If Condition 46 holds then it follows that

(∫ σ1 )
Preadyrate = E(σ 1 )−1 E 1{IP(0)−D(t+L)>0} dt .
0

54
 Proof. Since by Condition 46 we know 0 < E(σ 1 ) < ∞ it follows that
(∫ σ1 )
E 1{IP(0)−D(t+L)>0} dt ≤ E(σ 1 ) < ∞.
0

Applying now relation (5.1) and Theorem 48 with K = 0 and f (x) = 1 if x > 0 and 0
otherwise the desired result follows.
We now introduce the average number Bstockout of stockout occasions given by (if it
exists!)
Bstockout := limt↑∞ t−1 E(Cstockout (t)) (5.15)
with Cstockout = {Cstockout (t) : t ≥ 0} the stockout process defined by

Cstockout (t) = number of stockouts within (L, t + L]. (5.16)

Before mentioning the next result let D((σ 1 + L)−) := limt↑L D(σ 1 + t). Since the
demand process D is increasing we obtain by Theorem 10.2 of [31] that

P{D((σ 1 + L)−) ≤ IP(0)} = lim P{D(σ 1 + L − n−1 ) ≤ IP(0)}.

n↑∞

Observe the idea for the proof of the next results for the average number of stockout can
already be found in [27].

Lemma 53 If Condition 46 holds then it follows that

Bstockout = E(σ 1 )−1 (P{D(L) ≤ IP(0)} − P{D((σ 1 + L)−) ≤ IP(0)}) .

Proof. Since the increasing cost process Cstockout in relation (5.16) is completely
determined by the shifted net stock process INL it follows by Condition 46 and Theorem
47 that the cost process Cstockout is an additive functional of the process INL . Hence by
the average cost result for additive functionals and relation (5.15) we obtain that

Bstockout = E(σ 1 )−1 E(Cstockout (σ 1 )). (5.17)

To compute the value E(Cstockout (σ 1 )) we need to compute the average number of stock-
out occasions of the shifted netstock process INL up to time σ 1 . By our basic assumption
we know that the renewal sequence σ n , n ∈ N ∪ {0} contains as a subset the random
points in time that a replenishment order is issued and this replenishment order arrives L
time units later. By this observation at most one replenishment order arrives in the interval
[L, σ 1 + L) and if so it arrives at time L. This means that the cadlag net stock process
{IN(t + L) : t ≥ 0} is monotone decreasing on (0, σ 1 ) due to the arrival of demands and
so at most one stockout occasion happens in this interval. By this observation we obtain
that
E(Cstockout (σ 1 )) = E(1{stockout occurs in (L,σ1 +L)} ). (5.18)
By the definition of a stockout it follows that

1{stockout occurs in (L,σ1 +L)} = 1{IN(L)≥0,IN((σ1 +L)−)<0} . (5.19)

To compute this expectation we first observe that

1{IN(L)≥0,IN((σ1 +L)−)<0} = 1{IN(L)≥0} − 1{IN(L)≥0,IN((σ1 +L)−)≥0}. (5.20)

Since the shifted cadlag net stock process INL is monotone decreasing on (0, σ 1 ) it follows
that
1{IN(L)≥0,IN((σ1 +L)−)≥0} = 1{IN((σ1 +L)−)≥0} (5.21)

55
and this shows by relations (5.19),5.20) and (5.21) that
1{stockout occurs in (L,σ1 +L)} = 1{IN(L)≥0} ) − 1{IN((σ1 +L)−)≥0} )
Hence by relation (5.18) we obtain that the expected cost within interval (L, σ 1 + L) is
given by
E(Cstockout (σ 1 )) = P{IN(L) ≥ 0} − P{IN((σ 1 + L)−) ≥ 0}
and by relation (5.17) the desired result follows.
In the above proof it is clear that the probability Pstockout of a stockout during a cycle
is given by
Pstockout = P{IN(L) ≥ 0} − P{IN((σ 1 + L)−) ≥ 0}.
Before considering the computation of the average number Bitemshort of items short and
the fill rate measure Pf illrate we first prove the next limit result for the cadlag demand
process D.
Lemma 54 If Condition 46 holds and E(D(t)) < ∞ for every t ≥ 0 then it follows that
the average demand exists and this is given by
(∫ σ1 )
−1 −1
lim t E(D(t)) = (LE(σ 1 )) E (D(t + L) − D(t))dt .
t↑∞ 0

Proof. Since Condition 46 holds we obtain that the stochastic process {D(t + L) −
D(t) : t ≥ 0} is a regenerative process with the increasing pure renewal sequence {σ n :
n ∈ N ∪ {0}}. Introducing now the stochastic process A = {A(t) : t ≥ 0} with
∫ t
A(t) = D(v + L) − D(v)dv.
0

it follows by the monotonicity of the demand process D that the stochastic process A is
increasing. Moreover, we obtain
∫ t+σn
A(t + σ n ) − A(σ n ) = D(v + L) − D(v)dt.
σn

and this shows that the process A is an additive functional of the pure regenerative process
{D(t + L) − D(t) : t ≥ 0}. Hence by the average cost result for an additive functional of
a pure regenerative process it follows that
(∫ σ1 )
limt↑∞ t−1 E(A(t)) = E D(t + L) − D(t)dt . (5.22)
0

At the same time we obtain by the definition of A(t) that

(∫ ) (∫ )
t+L L
A(t) = E D(v)dv −E D(v)dv . (5.23)
t 0

Since by relation 5.22 it follows that limt↑∞ t−1 E(A(t)) exists we obtain by relation (5.23)
and the monotonicity of D together with E(D(L)) < ∞ that
(∫ )
t+L
limt↑∞ t−1 E(A(t)) = limt↑∞ t−1 E D(v)dv . (5.24)
t

Again by the monotonicity of D it follows that

(∫ )
t+L
LE(D(t)) ≤ E D(t)dt ≤ LE(D(t + L)
t

56
and by relations (5.24) and (5.22) this shows the desired result.
In the next result we need the itemshort process Citemshort = {Citemshort (t) : t ≥ 0}
defined by

Citemshort (t) = number of items short occurring within [L, t + L) (5.25)

and compute the average number of items short given by (if it exists!)

Bitemshort := limt↑∞ t−1 E(Citemshort (t)). (5.26)

As before the idea for the proof of the next result for the average number of items short can
already be found in [27].

Lemma 55 If Condition 46 holds and E(min{IP(0) − D(((σ 1 + L)−), 0}) is finite then
it follows that Bitemshort equals

E(σ 1 )−1 (E(min{IP(0) − D(L), 0}) − E(min{IP(0) − D((σ 1 + L)−), 0})) .

Proof. Since the increasing cost process Citemshort in relation (5.26) is completely de-
termined by the shifted backordering process BL it follows by Condition 46 and Theorem
47 that the cost process Citemshort is an additive functional of the process INL . Hence by
the average cost result for additive functionals and relation (5.26) we obtain

Bitemshort = E(σ 1 )−1 E(Citemshort (σ 1 )).

given that E(Citemshort (σ 1 )) is finite.To compute this expression and to show it is finite
we observe by Condition 46 that at most one replenishment order arrives in the interval
[L, σ 1 + L) and if so it arrives at time L. Therefore the cadlag backordering process B
has a possible jump at L and this process is monotone increasing on [L, σ 1 + L). Since
Citemshort (σ 1 ) represents the number of items short occurring within interval [L, σ 1 + L)
this implies
Citemshort (σ 1 ) = B((σ 1 + L)−) − B(L).
Using now relation (5.1) we obtain

Citemshort (σ 1 ) = min{IN(L), 0} − min{IN((σ 1 + L)−), 0}

and this implies by relation (5.10) and our assumption the desired result.
Since the fillrate measure Pf illrate is defined as the fraction of demand directly deliv-
ered we obtain by Lemma 54 that
Bitemshort
Pf illrate = 1 − and a = limt↑∞ t−1 E(D(t)) (5.27)
a
By the above analysis it is sufficient to verify that any inventory control model (with com-
plete backordering!) to be discussed satisfies the basic assumption together with an iden-
tification of the renewal or regeneration point σ 1 . This observation concludes our general
discussion of single item inventory control rules. As already observed the decision to order
depends on the inventory position process and in the inventory literature the following basic
control rules are studied (cf.[5]). We will first introduce the control rules used in periodic
review systems.

Control rules for periodic review systems.

57
 • The (R,S) inventory control rule11 . In this rule every R time units the inventory po-
sition process is inspected and a replenishment order is issued if the level of the
inventory position process at the inspection time is below S. Moreover, the size of
the order is such that all excess demand up to that time will be backordered and
the inventory position process is raised to order-up-to level S. In the most general
optimization problem the variables R > 0 and S > 0 are decision variables and
need to be chosen in an optimal way dependent upon the chosen cost structure. A
straightforward generalization of the (R, S) inventory control rule is given by the
following.
• The (R,s,S) inventory control rule12 . In this rule every R time units the inventory
position process is inspected and a replenishment order is issued if the level of the
inventory position process is below s ≤ S. If this is not the case no order is issued
and one waits until the next inspection time. The order size is now similarly de-
termined as in the (R, S) rule and so in the most general optimization problem the
decision variables are given by R, s and S
• The (R,s,nQ) inventory control rule13 . In this rule every R time units the inventory
position process is inspected and a replenishment order is issued if the level of the
inventory position process is below or at level s. If this is not the case no order is
issued and one waits until the next inspection time. The order size is chosen to be
an integer multiple of Q such that after ordering the inventory position process will
be above level s and smaller than or equal to level s + Q. Observe as in the (R, S)
inventory control model all excess demand will be backordered. In the most general
optimization problem the decision variables are given by R, s and Q.

It will be shown in the next subsections that all the cost functions of periodic review
models can be reduced to simple transformations of corresponding cost functions of
the (R, S) model and so this yields despite the large amount of different formula and
the completely different control rules an easy interpretation.of all these models. As
observed in Section 7.3 of Silver, Pyke and Peterson (cf.[5]) the advantages of periodic
review systems are that a periodic review system is simpler to administer, a better prediction
of the workload of the staff involved is possible, it is more effective in detecting spoilage
and the review costs are less then for ”electronic” continuous review systems. Finally it is
easier to combine orders for different items into a single order by using the same review
interval for all the included items. Of course the main disadvantage of periodic review
systems are the higher safety stocks14 and so higher holding costs are required to maintain
the same service level. Observe the safety or buffer stock is defined as the average level of
the net stock process just before a replenishment order arrives. In case of continuous review
systems one continuously monitors the inventory position process and this gives rise to the
following rules.

Control rules for continuous review systems.

• The (s,S) inventory control rule.15 In this rule a replenishment order is issued in case
the level of the inventory position drops below s ≤ S and the size of the order is such
that both the excess demand is backordered and the level of the inventory position
process is raised to order-up-to level S. In the most general optimization problem
the decision variables are given by s and S. A very special and important subcase is
11 (R,S) control rule
12 (R,s,S) control rule
13 (R,s,S) control rule
14 safety stock
15 (s,S) control rule

58
 given by the (S, S) control rule. If this rule is used every arriving customer generates
a replenishment order and the order size is given by the demand of this customer.
• The (s,nQ) inventory control rule16 . In this rule a replenishment order is issued in
case the level of the inventory position is below or at level s. The size of the order
is chosen to be an integer multiple of Q and is such that both the excess demand is
backordered and the level of the inventory position process will be between s and
s + Q. In the most general optimization problem the decision variables are s and Q.

It will be shown in the next subsections that all the cost functions of continuous re-
view models can be reduced to simple transformations of corresponding cost functions
of the (S, S) model and so this yields despite the large amount of different formula and
the completely different control rules an easy interpretation of all these models. Fi-
nally we assume that the demand process for the different periodic review models is given
by a compound Poisson process17 while for the continuous review models it is given by a
compound renewal process18 . In the next sections we will show that the different inventory
control models all satisfy the basic assumption and this yields the possibility to identify the
average costs. At the same time we will discuss for a divisible item how to compute by
means of the F F T method the different cost functions.

16 (s,nQ)
control rule
17 compound Poisson process
18 compound renewal process

59
Chapter 6

Regenerative Processes and

Single Item Inventory Control.

In this chapter we consider the different single-item inventory control models with com-
plete backordering and a deterministic lead time and verify that these models satisfy Con-
dition 46.

6.1 Analysis of the (R,S) model.

Under this rule every R time units the inventory position process is inspected and an or-
der is placed if the level of the inventory position process at the inspection time is below
S. Moreover, the size of the order is such that all excess demand up to that time will be
backordered and the inventory position process is raised to order-up-to level S. Hence, the
variables R > 0 and S > 0 are decision variables and need to be chosen optimally depend-
ing upon the chosen cost structure. It is assumed that the demand process D associated
with the (R, S) inventory control model is a compound Poisson process with arrival rate
λ > 0 and the associated cumulative demand distribution FY of the individual demands
Yn , n ∈ N satisfies FY (0) = 0 with E(Y1 ) < ∞. This means that
∑N(t)
D(t) = Yn , Y0 = 0
n=0

with N := {N(t) : t ≥ 0} denoting the Poisson arrival process of customers and the
independent and identically distributed individual demands Yn , n ∈ N are independent of
N. Without loss of generality, the initial inventory position level is set at S or equivalently,
IP(0) = S. In the next theorem it is shown that the (R, S) model with IP(0) = S satisfies
Condition 46. Observe the proof includes both divisible and indivisible items.
Theorem 56 For any (R, S) model with a compound Poisson demand process, lead time
L and complete backordering the stochastic process
X : = {(IP(t), D(t + L) − D(t)) : t ≥ 0}
and IP(0) = S is a simplified pure regenerative process with the increasing sequence of
regeneration points given by nR, n ∈ N ∪ {0} and the points nR contain as a subset the
points in time that a replenishment order is issued.
Proof. To show that the stochastic process X is a simplified pure regenerative process
we first observe for every n ∈ N that the incremental compound Poisson process DnR :=
{DnR (t) : t ≥ 0} given by
DnR (t) := D(nR + t) − D(nR)

60
has the same distribution as the original compound Poisson process D. This means that
the stochastic process DnR is again a compound Poisson process. By the definition of the
(R, S) rule we obtain that IP(nR) = S and since the process {IP(t + nR) : t ≥ 0} is
completely determined by IP(nR) = S and the stochastic process {DnR (s) : 0 ≤ s ≤ t}
this yields that there exists some function g such that

IP(t + nR) = g({DnR (s) : 0 ≤ s ≤ t}) (6.1)

for every t ≥ 0. By relation (6.1) we obtain that the shifted stochastic process

XnR := {(IP(t + nR), D(t + nR + L) − D(t + nR)) : t ≥ 0}

has the representation

{(g({DnR (s) : 0 ≤ s ≤ t}), DnR (t + L) − DnR (t)) : t ≥ 0}

and by the previous observation its distribution is independent of n. Since the inspection
times nR contain as a subset the replenishment moments this shows that the stochastic
process X satisfies Condition 46.
By the definition of the (R, S) policy we already observed that the inventory position
process IP up to time s is completely determined by the demand process up to time s
given by {D(t) : t ≤ s}. This implies by the stationary and independent increments1 of
a compound Poisson process (cf.[34]) that the random variables IP(t + nR) and D(t +
nR + L) − D(t + nR) are independent for any t ≥ 0 and n ∈ N ∪ {0}. Moreover, by
relation (5.10) we obtain for every 0 ≤ t < R that the relation

IN(t + L) = S − D(t + L) (6.2)

holds. Applying Theorem 56 it is now easy to identify the average cost of a (R, S) model.

Theorem 57 For any (R, S) model satisfying the conditions of Theorem 56 and
∫ R
E( f (S − D(t + L))dt) < ∞
0

it follows that the average cost Φ(R, S) of a (R, S) control policy is given by
( ∫ )
R
Φ(R, S) = R−1 K(1 − exp(−λR)) + E(f (S − D(t + L))dt .
0

Proof. By the definition of the (R, S) control rule it follows that {replenishment order
in a cycle} = {D(R) > 0} and this shows using FY (0) = 0 that

pr = P{D(R) > 0} = P{N(R) ≥ 1} = 1 − exp(−λR).

Applying now Theorem 56 and 48 with σ 1 replaced by R and IP(0) by S yields the desired
result.
The expression for the average number of backorders per unit of time, the average
physical inventory level per unit of time and the ready rate measure associated with the
(R, S) model is given by the next result.

Theorem 58 For any (R, S) model satisfying the conditions of Theorem 56 it follows that
∫ R
Bbackorder/time (R, S) = R−1 E(max{D(t + L) − S, 0})dt.
0
1 stationairy and independent increments

61
Also the average physical stock level per unit of time is given by
∫ R
Bstocklevel/time (R, S) = R−1 E(max{S − D(t + L), 0})dt
0
while the ready rate measure equals
∫ R
Preadyrate (R, S) = R−1 P{S − D(t + L) > 0}dt.
0
Proof. By the monotonicity of the compound Poisson demand process D it is obvious
that ∫ R
E(max{D(t + L) − S, 0})dt ≤ R(E(D(R + L)) − S)) < ∞
0
and so the first formula for the average number of backorders per unit of time is an im-
mediate consequence of Theorem 56 and Lemma 50 with σ 1 replaced by R and IP(0) by
S. The second formula for the average physical stock level per unit of time follows by a
similar approach replacing Lemma 50 by Lemma 51, while the third formula for the ready
rate measure can be shown replacing Lemma 50 by Lemma 52.
Since for any random variable Y it follows
max{S − Y, 0} − max{Y − S, 0} = S − Y (6.3)
we obtain since D is a compound Poisson demand process that
E(max{S − D(v), 0} − E(max{D(v) − S, 0}) = S − λvE(Y1 ) (6.4)
for every v ≥ 0. This shows by Theorem 58 that
Bstocklevel/time − Bbackorder/time = S − (2R)−1 λE(Y1 )((R + L)2 − L2 ) (6.5)
and so the physical stock level per unit of time and the average number of backorders per
unit of time are related. Moreover, since for any nonnegative random variable Z it follows
that ∫ S
E(max{S − Z, 0}) = P{Z ≤ z}dz (6.6)
0
and P{D(v) < y} = P{D(v) ≤ y} for any continuous cumulative demand distribution
FY we obtain for any continuous distribution FY and Theorem 58 that
∫ S
Bstocklevel/time (R, S) = Preadyrate (R, y)dy (6.7)
0

with Preadyrate (R, y) denoting the fraction of time the inventory level is positive for a
(R, y) control rule. To derive the different expressions for the average number of items
short and the average number of stockout occasions and the associated service measures
we can use Lemma 53 and 54 applied to the (R, S) model.
Theorem 59 For any (R, S) model satisfying the conditions of Theorem 56 it follows that
the average number of stockout occasions is given by
Bstockout (R, S) = R−1 (P{D(L) ≤ S} − P{D(R + L) ≤ S})
and the stockout service measure equals
Pstockout (R, S) = P{D(L) ≤ S} − P{D(R + L) ≤ S}.
Moreover, the average number of items short is given by
Bitemshort (R, S) = R−1 (E(min{S − D(L), 0}) − E(min{S − D(L + R), 0}))
and the fill rate measure equals
Pf illrate (R, S) = 1 − (λE(Y1 ))−1 Bitemshort (R, S).

62
 Proof. Since D is a compound Poisson process the random variable D((R + L)−)
has the same distribution as D(R + L) and this shows by Theorem 56 and Lemma 53
with σ 1 replaced by R and IP(0) by S the formula for the average number of stockout
occasions. A similar proof applies to the stockout service measure replacing Lemma 53 by
the observation after Lemma 53. The formulas for the average number of items short and
the ready rate measure can be shown similarly using Lemma 55 and the observation after
Lemma 55. For the last formula also observe that E(D(t)) = λE(Y1 )t for every t ≥ 0.
By the probabilistic interpretation of the average number of stockouts it follows for
R > 0 fixed that the function S → Bstockout (R, S) is decreasing and by Theorem 59 we
obtain limS↑∞ Bstockout (R, S) = 0 and

limS↓0 Bstockout (R, S) = R−1 exp(−λL)(1 − exp(−λR)).

Also it is easy to see that

limS↑∞ Bitemsshort (R, S) = 0 and limS↓0 Bitemshort (R, S) = λE(Y1 ).

Moreover, since for any random variable Z it follows that

min{Z, 0} + max{Z, 0} = Z

we obtain by Theorem 59 and E(D(v)) = λvE(Y1 ) for every v ≥ 0 that the average
number of items short equals

λE(Y1 ) + R−1 (E(max{S − D(L + R), 0}) − E(max{S − D(L), 0})). (6.8)

This shows applying relation (6.6) and Theorem 59 that

∫ S
Bitemshort (R, S) = λE(Y1 ) − Bstockout (R, y)dy. (6.9)
0

with Bstockout (R, y) the average number of stockout occasions for a (R, y) control rule.
In the next subsection we discuss how to compute the different objectives associated with
the (R, S) model in case we are dealing with a divisible item.

6.1.1 Computation of the objectives in a (R,S) model.

We will now consider the problem of computing the different cost functions in case we are
dealing with a divisible item. This means that the state space of the individual demands
Yn is given
∑by N and looking at Theorem 57 we need to derive the generating function
∞
Rg (z) := n=0 z n g(n) of the sequence g:= {L(f )(n) : n ∈ N ∪ {0}} given by
∫ R
L(f )(n) := E(f (n − D(t + L))dt, n ∈ N ∪ {0}. (6.10)
0

To do this we first start with a cost rate sequence f:= {f (n) : n ∈ Z} satisfying f (n) = 0
for every n ≤ −1. Its generating function Rf is given by
∑∞
Rf (z) := f (n)z n
n=0

and we always assume that Rf (z) is finite valued for every complex valued z with |z| < 1.
Moreover, the generating function of the individual demand Y1 is given by

PY1 (z) := E(z Y1 )

and it is clear that |PY1 (z)| < 1 for every complex valued z with |z| < 1. It is now possible
to show the following result.

63
Theorem 60 If f : Z → R is a cost rate sequence satisfying f (n) = 0 for every n ≤ −1
and Rf (z) is finite valued for every complex valued z satisfying |z| < 1 then it follows
that Rg (z) is finite valued for every complex valued z satisfying |z| < 1. In particular we
obtain that
Rf (z) exp(−λL(1 − PY1 (z)))
Rg (z) = (1 − exp(−λR(1 − PY1 (z))))
λ(1 − PY1 (z))

Proof. By relation (6.10) and f nonnegative we obtain by the monotone convergence

theorem that ∫ R (∑
∞ )
Rg (z) = z n E(f (n − D(t + L)) dt (6.11)
0 n=0

Since the sequence f satisfies f (n) = 0 for every n ≤ −1 it follows

∑n
E(f (n − D(t + L)) = f (n − k)P {D(t + L) = k}
k=0

and this implies reversing the order of summation that

∑∞ ∑∞ ∑n
z n E(f (n − D(t + L)) = zn f (n − k)P {D(t + L) = k}
n=0 n=0 k=0
∑∞ ∑∞
= zk f (n − k)z n−k P {D(t + L) = k}
k=0 n=k

for every 0 ≤ t ≤ R. Hence we obtain

∑∞
z n E(f (n − D(t + L)) = Rf (z)E(z D(t+L) ) (6.12)
n=0

and since it is well known that

E(z D(t+L) ) = exp(−λ(t + L)(1 − PY1 (z)))

it follows by relation (6.11) that

∫ R
Rg (z) = Rf (z) exp(−λ(t + L)(1 − PY1 (z)))dt.
0

Since |PY1 (z)| < 1 for every |z| < 1 the desired result follows by standard calculus.
In case penalty costs are introduced within the (R, S) model these are mostly linear
and this means for a selected p > 0 that the used cost rate sequence satisfies f (n) = −pn
for every n ≤ −1. To reduce the case of linear penalty costs to the case of no penalty
costs introduce the related cost rate sequence f1 := {f1 (n) : n ∈ Z} given by f1 (n) :=
f (n) + pn. It is clear that the sequence f1 vanishes on {−1, −2, ..} and the generating
function of the sequence g1 = {L(f1 )(n) : n ∈ N} is given by Theorem 60 replacing
Rf (z) by
∑∞
Rf1 (z) = f (n)z n + pz(1 − z)−1 .
n=0

Moreover, we obtain
∫ R
L(f ) − L(f1 ) = −pnR + p E(D(t + L))dt.
0
∫R
and since D is a compound Poisson process and so 0 E(D(t + L))dt = λE(Y1 )(t + L)
it follows that ∫ R
λE(Y1 )
E(D(t + L))dt = ((R + L)2 − L2 )
0 2

64
This shows that the sequence g:= {L(f )(n) : n ∈ N ∪ {0}} can be written as

λE(Y1 )p
L(f )(n) = L(f1 )(n) + ((R + L)2 − L2 ). (6.13)
2
Evaluating now the generating function of the first expression in relation (6.13) by using
Theorem 60 we can apply the F F T method (cf.[21]) to compute the value L(f )(n) and
so it is possible by means of relation (6.13) to evaluate the average cost of a (R, S) control
policy.

Algorithm 61 Computation average cost for the (R,S) model by means of the FFT method.

1. Compute for fixed R > 0 the term L(f1 )(S) by applying the F F T method. Go to
step 2.
2. Evaluate
λE(Y1 )p
I(R, S) := L(f1 )(S) − pSR + ((R + L)2 − L2 )
2
and go to step 3.
3. Evaluate
Φ(R, S) = R−1 (K(1 − exp(λR)) + I(R, S)).

In the next example we consider a numerical instance of the (R, S) model and show the
graph of the average cost of a (R, S) model for some given fixed R > 0 using the build-in
FFT procedure available in MATLAB.

Example 62 In this numerical we assume that L = 54 days, K = $30 and the linear
holding costs per day are given by 0.15 times $39. The linear penalty costs are 0.3 times
$39 per day with $39 representing the value of the item under consideration . Moreover, it is
assumed that the demand process is a compound Poisson process with expected interarrival
times equal to 36.9 days. We know that the the lead time is given by 54 days and from past
data the empirical distribution function of the individual demand is given

k empirical probabilities
1 9/63
2 14/63
3 5/63
4 6/63
5 7/63
6 7/63
7 1/63
8 1/63
9 1/63
10 8/63
12 2/63
20 2/63

by the next table. For these data the average cost for R = 1/2 is computed as a function
of the order-up-to level S by using the FFT procedure of MATLAB and this function is
shown in Figure 1.

65
100

90

80

70

60

50
Figure 6.1: Average cost S −→ Φ( 12 , S).
40

30 To compute the average number Bstockout (R, S) of stockout occasions as given by

Theorem 59 we observe for any nonnegative random variable D with state space {0, 1, 2, ..}
20
and finite first moment that
∑∞ ∑∞ ∑∞
10
P{D >n}z n = P{D = k}z n (6.14)
n=0 n=0 k=n+1
0 ∑
25 ∞
0
for 5every complex
10
z with |z| ≤
valued 15 20
1. Since n=0 P{D30>n} = E(D) 35
we may reverse
the double summation and this yields that the last term in relation (6.14) equals
∑∞ ∑k−1 ∑∞
(1 − z k )P{D = k}
z P{D = k} = k=1
n
k=1 n=0 1−z
Substituting this expression into relation (6.14) and denoting by PD (z) the generating func-
tion of the random variable D this shows
∑∞ 1 − PD (z)
P{D > n}z n = <∞ (6.15)
n=0 1−z
for any complex valued z with |z| ≤ 1. Moreover, for any compound Poisson process it is
well known that
E(z D(t) ) = exp(−λt(1 − PY1 (z))) (6.16)
for every t ≥ 0 and by relations (6.15) and (6.16) it is easy to verify the following result.

Theorem 63 For fixed R > 0 the generating function of the sequence {Bstockout (R, n) :
n ∈ N ∪ {0}} defined by
∑∞
P(z) := Bstockout (R, n)z n
n=0

is finite valued for every complex valued z with |z| ≤ 1 and this generating function is
given by
exp(−λL(1 − PY1 (z))) − exp(−λ(R + L)(1 − PY1 (z)))
R(1 − z)

66
 Proof. By Theorem 59 it follows that

Bstockout (R, n) = R−1 (P{D(R + L) > n} − P{D(L) > n})

Applying now relations (6.15) and (6.16) yields the desired result.
To compute the average number Bitemshort (R, S) of items short as given by Theorem
59 we observe for any nonnegative discrete random variable D and sequence f0 := {f0 (n) :
n ∈ Z} given by f0 (n) = n for every n ∈ N ∪ {0} and zero otherwise that

E(max{n − D, 0}) = E(f0 (n − D))

Moreover, by relation (6.12) it follows that

∑∞
E(f0 (n − D))z n = Rf0 (z)E(z D ) (6.17)
n=0

and since Rf0 (z) = z(1 − z)−2 we obtain

∑∞
E(f0 (n − D))z n = z(1 − z)−2 E(z D ). (6.18)
n=0

Using now relations (6.16) and (6.18) it is easy to verify the following result and this en-
ables us in combination with relation (6.8) to compute the average number Bitemshort (R, S)
of items short in a (R, S) model.

Theorem 64 If for fixed R > 0 the sequence {g(R, n) : n ∈ N ∪ {0}} is defined by

g(R, n) := R−1 (E(max{n − D(R + L), 0}) − E(max{n − D(L), 0}).

then its generating function ∑∞

g(R, n)z n
n=0

is finite valued for every complex valued z with |z| < 1 and it is given by

z (exp(−λ(R + L)(1 − PY1 (z))) − exp(−λL(1 − PY1 (z)))

R(1 − z)2

Proof. By the definition of the sequence {g(R, n) : n ∈ N ∪ {0}} we obtain

g(R, n) = R−1 (E(f0 (n − D(R + L)) − E(f0 (n − D(L))

and this shows by relations (6.16) and (6.18) the desired result.
To conclude this subsection on the (R, S) model we list the following numerical exam-
ple.

Example 65 Suppose the individual demand Y1 has the distribution

1
P{Y1 = 1} = P{Y1 = 2} = P{Y1 = 3} =
3
and the arrival rate of the Poisson arrival process equals 5. Hence the generasting function
of the individual demand Y1 equals
∑3 1 k
PY1 (z) = z
k=1 3
and the generating function of the random variable D(t) is given by
∑3 1 k
exp(−λt(1 − PY1 (z))) = exp(−5t(1 − z ))
k=1 3

67
Moreover, if the cost rate function f is given by

f (x) = 6n, n ∈ N ∪ {0} and f (x) = −10n, otherwise

this means that the inventory holding costs per item per unit of time are given by 6 Dutch
guilders and the penalty cost per unit of time per backordered item are given by 10 guilders
The time is measured in weeks and by the above observation we obtain that

f1 (n) = 16n, n ∈ N ∪ {0}

It now follows that

∑∞ 16z
Lf1 (z) = 16 nz n =
n=0 (1 − z)2
and so by Theorem 60 the generating function of the sequence
∫ R
L(f1 )(n) := E(f1 (n − D(t + L))dt
0

with L = 3 is given by
∑3
16z exp(−15(1 − k=1 31 z k )) ∑3 1
∑ 3 (1 − exp(−5R(1 − z k ))
5(1 − z) (1 − k=1 3 z ))
2 1 k k=1 3

Moreover, since E(Y1 ) = 2 we obtain

λE(Y1 )
((R + L)2 − L2 ) = 5((R + 3)2 − 9)
2
and so we can compute the average cost for an arbitrary (R, S) policy.

This concludes the section on the analysis of the (R, S) model and the computation of
the different cost functions by means of the F F T method (cf. [21]). In the next section we
will consider the (R, s, S) model. Observe we follow in the rest of these notes the approach
discussed in [19]. For a different approach the reader is referred to [6].

6.2 Analysis of the (R,s,S) model.

Under this rule every R units the inventory position process is inspected and a replenish-
ment order is placed if the cadlag inventory position process just before time nR satisfies

IP((nR)−) := limt↑nR IP(t) < s ≤ S.

Observe for s = S that the previous (R, S) model is reobtained. If the level IP((nR)−) is
not below s no order is triggered and one waits until the next inspection time. The order size
is now similarly determined as for the (R, S) model (including the excess demand!) and the
decision variables are given by S, s and R. As for the (R, S) model the arrival process is a
compound Poisson process with arrival rate λ > 0 and the cumulative demand distribution
FY of the individual demands Yn , n ∈ N satisfies FY (0) = 0 with E(Y1 ) < ∞. If we
like to show (cf.[19]) that the process

X = {(IP(t), D(t + L) − D(t)) : t ≥ 0}

is a simplified pure regenerative process with regeneration points nR it follows in case

IP(0) is set equal to S that at some inspection points a replenishment order is issued and
hence the inventory position process IP at those points equals S, while at other inspection
points which do not trigger a replenishment order this is not the case. Therefore with this

68
initial distribution of the process X at time 0 this is never a simplified pure regenerative
process with regeneration points nR and so we first need to find some initial cumulative
distribution function
F (x, y) := P{IP(0) ≤ x, D(L) ≤ y}
of the process X at time 0 for which it holds that

F (x, y) = P{IP(nR) ≤ x, DnR (L) ≤ y}

for every n ∈ N and DnR (L) := D(nR + L) − D(nR). To achieve this we first observe
that the random variables IP(t) and D(t + L) − D(t) are independent for every t ≥ 0.
This is due to the fact that the random variable IP(t) is completely determined by the
stochastic process {D(s) : 0 ≤ s ≤ t} and the compound Poisson demand process D has
independent and stationairy increments. Hence it follows that the probability

P{IP(nR) ≤ x, DnR (L) ≤ y}

equals

P{IP(nR) ≤ x}P{DnR (L) ≤ y} = P{IP(nR) ≤ x}P {D(L) ≤ y}

and so we are looking for an invariant distribution2 of the cadlag process {IP(nR) : n ∈
N∪{0}}. By the definition of the control rule it follows immediately that s ≤ IP(nR) ≤ S
and to find the invariant distribution of the embeddded process {IP(nR) : n ∈ N ∪ {0}} it
is more convenient to consider the process (cf.[9])

Vn := S − IP(nR).

with state space [0, S − s] for an indivisible item and {0, 1, ..., S − s} for a divisible item.
By the definition of the control rule it is easy to verify (make a drawing!) that

Vn+1 = (Vn + Bn )1{Vn +Bn ≤S−s} (6.19)

for every n ∈ N with Bn denoting the random demand in period [nR, (n + 1)R). Since
we deal with a compound Poisson process these random demands are independent and
identically distributed with cumulative distribution function

FB1 (x) := P{B1 ≤ x} = P{D(R) ≤ x} := FD(R) (x)

and Bn independent of Vn . Hence the process {Vn : n ∈ N ∪ {0}} is a discrete time

Markov process and for this Markov process one can show the following result. Before
showing this result we give the definition of an invariant distribution of a discrete time
Markov process.

Definition 66 An initial cumulative distribution function Finv of the discrete time Markov
process V = {Vn : n ∈ N ∪ {0}} given by

Finv (x) := P{V0 ≤ x}

is called an invariant distribution of the stochastic process V if Finv (x) = P{Vn ≤ x} for
every n ∈ N ∪ {0}.

The next result identifies the invariant distribution of the discrete time Markov process
V given by relation (6.19).
2 invariant distribution

69
Theorem 67 The invariant distribution of the discrete time Markov process V = {Vn :
n ∈ N ∪ {0}} given by relation (6.19) has the form

1 + RFD(R) (v)
Finv (v) = ,0 ≤ v ≤ S − s (6.20)
1 + RFD(R) (S − s)

with RFD(R) the pure renewal function associated with the cumulative distribution function
FD(R) of the random variable D(R).

Proof. By relation (6.19) the equilibrium random level Veq ∈ [0, S − s] of the Markov
process V must satisfy the relation

Veq =d (Veq + B)1{Veq +B≤S−s} (6.21)

with the random variable B independent of Veq . By relation (6.21) it follows with W :=(Veq +
B)1{Veq +B≤S−s} that the distribution Finv of the random variable Veq satisfies

Finv (v) = E(1{Veq ≤v} ) = E(1{W≤v} ) (6.22)

= E(1{W≤v} 1{Veq +B>S−s} ) + E(1{W≤v} 1{Veq +B≤S−s} )

for every 0 ≤ v ≤ S − s. Analyzing the first term in relation (6.22) we obtain

1{W≤v} 1{Veq +B>S−s} ) = 1{Veq +B>S−s}

and this shows that

E(1{W≤v} 1{Veq +B>S−s} ) = P{Veq + B > S − s} := a.

To analyze the second term in relation (6.22) we first observe that

1{W≤v} 1{Veq +B≤S−s} = 1{Veq +B≤v}

and by the independence of the random variables Veq and B this yields

E(1{Veq +B≤v} ) = P{Veq + B ≤ v} = (FD(R) ∗ Finv )(v)

for every 0 ≤ v ≤ S − s. Hence relation (6.22) can be rewritten as

Finv (v) = a + (FD(R) ∗ Finv )(v)

for every 0 ≤ v ≤ S − s and substituting the above relation for Finv an arbitrary number of
times in the above equation it follows that
∑n (n+1)∗
Finv (v) = a(1 + k∗
FD(R) (v)) + (FD(R) ∗ Finv )(v).
k=1

(n+1)∗
for every 0 ≤ v ≤ S −s and n ∈ N. Letting n ↑ ∞ and using limn↑∞ (FD(R) ∗Finv )(v) =
0 shows
Finv (v) = a(1 + RFD(R) (v))
with RFD(R) denoting the pure renewal function associated with the cumulative distribution
function FD(R) . Since Finv (S − s) = 1 it must hold that

a = (1 + RFD(R) (S − s))−1

and so the desired result is verified.

By Theorem 67 and using a similar proof as in Theorem 56 with some obvious modifi-
cations one can show the next result.

70
Theorem 68 For any (R, s, S) model with a compound Poisson demand process, lead time
L and complete backordering the associated stochastic process

X : = {(IP(t), D(t + L) − D(t)) : t ≥ 0}

with IP(0) =d S−Veq , Veq independent of D and the distribution of Veq given by relation
(6.20) is a simplified pure regenerative process. The regeneration points of this process are
given by nR, n ∈ N ∪ {0} and contain as a subset the replenishment moments.

To simplify the notation in the remainder in this subsection we introduce with Veq as
in Theorem 68 the stochastic process D1 = {D1 (t) : t ≥ 0} given by

D1 (t) = D(t) + Veq

for every t ≥ 0. As for the (R, S) model it follows that

IN(t + L) = IP(0) − D(t + L) =d S − Veq − D(t + L) (6.23)

for every 0 ≤ t ≤ R and by the stationairy and independent increments of a compound

Poisson process D and Veq independent of D we obtain that the random variable IP(t +
nR) is independent of the random variable D(t + nR + L) − D(t + nR). It is now possible
to identify the average cost Φ(R, s, S) of a (R, s, S) model. Observe it is always assumed
that the used costrate function f is nonnegative.

Theorem 69 For any (R, s, S) model satisfying the assumptions of Theorem 68 and
∫ R
E(f (S − D1 (t + L))dt < ∞
0

it follows that the average cost Φ(R, s, S) of a (R, s, S) control policy is given by
( ∫ )
R
Φ(R, s, S) = R−1 KP{D1 (R) > S − s} + E(f (S − D1 (t + L))dt .
0

Proof. By Theorem 68 we can apply Theorem 48 with σ 1 replaced by R and IP(0) by

S − Veq . To write out the formula in Theorem 48 we only need to determine the probability
of a replenishment order at time R. Since the event {replenishment order at time R} is the
same as the event {IP(0) − D(R) < s} and IP(0) equals S − Veq we obtain

pr = P{D(R) + Veq > S − s} = P{D1 (R) > S − s}

showing the desired result.

By relation (6.10), Theorem 69 and Veq is independent of D we obtain
∫ R
E(f (S − Veq − D(t + L))dt = E(L(f )(S − Veq )) (6.24)
0

with L(f )(y) representing the inventory/penalty of a (R, y) control rule and so the in-
ventory/penalty costs for a (R, s, S) model are related to the inventory/penalty costs of a
(R, S) model. The expression for the average number of backorders per unit of time, the
average physical inventory level per unit of time and the ready rate measure associated with
the (R, s, S) model is given by the next result.

Theorem 70 For any (R, s, S) model satisfying the assumptions of Theorem 68 it follows
that ∫ R
Bbackorder/time (R, s, S) = R−1 E(max{D1 (t + L) − S, 0})dt.
0

71
Also the average physical stock level per unit of time is given by
∫ R
Bstocklevel/time (R, s, S) = R−1 E(max{S − D1 (t + L), 0})dt
0

while the ready rate measure equals

∫ R
Preadyrate (R, s, S) = R−1 P{S − D1 (t + L) > 0}dt.
0

Proof. Since 0 ≤ Veq ≤ S − s it is obvious that

∫ R
E(max{D(t + L) + Veq − S, 0})dt < ∞
0

and so the above formuals are an immediate consequence of Theorem 68 and Lemma 50,
51 and 52 with σ 1 replaced by R and IP(0) by S − Veq .
By Theorem 70 and 58 it is clear by the definition of the process D1 := {D1 (t) : t ≥ 0}
that
Preadyrate (R, s, S) = E(Preadyrate (R, S − Veq )) (6.25)
with Preadyrate (R, y) the ready rate measure associated with the (R, y) rule. Similarly we
obtain
Bbackorders (R, s, S) = E(Bbackorders (R, S − Veq )) (6.26)
and
Bstocklevel (R, s, S) = E(Bstocklevel (R, S − Veq )) (6.27)
with Bbackorders (R, y), respectively Bstocklevel (R, y) denoting the average number of
backorders per unit of time and the average physical inventory level per unit of time in
case a (R, y) control rule is applied. Similarly one can derive expressions for the average
number of items short and the average number of stockout occasions together with the as-
sociated fill rate and the stockout service measure using Lemma 53 and 55 applied to the
(R, s, S) model with the above initial distribution. This yields the following result.

Theorem 71 For any (R, s, S) model satisfying the assumptions of Theorem 68 it follows
that the average number of stockout occasions is given by

Bstockout (R, s, S) = R−1 (P{D1 (L) ≤ S} − P{D1 (R + L) ≤ S})

and the stockout service measure equals

Pstockout (R, s, S) = P{D1 (L) ≤ S} − P{D1 (R + L) ≤ S}.

Moreover, the average number of items short is given by

Bitemshort (R, s, S) = R−1 (E(min{S − D1 (L), 0}) − E(min{S − D1 (L + R), 0}))

and the fill rate measure equals

Pf illrate (R, s, S) = 1 − (λE(Y1 ))−1 Bitemshort (R, s, S).

Proof. Since 0 ≤ Veq ≤ S − s it is obvious that

E(min{D1 (R + L) + Veq − S, 0})dt < ∞

and so by Theorem 68 in combination with Lemma 53 and 55 with σ 1 replaced by R and

IP(0) by S − Veq we obtain the desired result.

72
 By Theorem 71 and 59 it follows immediately that

Bstockout (R, s, S) = E(Bstockout (R, S − Veq )) (6.28)

and
Bitemshort (R, s, S) = E(Bitemshort (R, S − Veq )) (6.29)
with the random variable Veq independent of the functions y → Bstockout (R, y) and y →
Bitemshort (R, y) representing respectively the average number of stockout occasions and
the average number of items short for the (R, y) rule. Hence also the service measures
of a (R, s, S) model are related to the service measures of a (R, S) model. A similar
observation also holds for the no-stockout service measure and the fill rate service measure.
For these measures we obtain

Pstockout (R, s, S) = E(Pstockout (R, S − Veq )) (6.30)

and
Pf illrate (R, s, S) = E(Pf illrate (R, S − Veq )) (6.31)
In the next subsection we will discuss how to compute the different objectives of the
(R, s, S) model for a divisible item using the FFT method.(cf.[21])

6.2.1 Computation of the objectives in a (R,s,S) model.

We will now focus on the computation of the different cost functions for the average cost in
case we are dealing with a divisible item. If this holds it follows that the decision variables
S and s are nonnegative integers and the random variable Veq has state space {0, 1, .., S −
s}. We now need to evaluate the value
∫ R
g1 (S) = E(f (S − Veq − D(t + L))dt. (6.32)
0

By the independence of Veq and the demand process D it follows for every 0 ≤ t ≤ R that
∑S−s
E(f (S − Veq − D(t + L)) = pm E(f (S − n − D(t + L)) (6.33)
n=0

with
pn = P {Veq = n}, n = 0, .., S − s. (6.34)
Hence with g := {L(f )(n) : n ∈ N ∪ {0}} denoting the sequence belonging to the (R, S)
model given by
∫ R
L(f )(n) := E(f (n − D(t + L))dt
0
(see relation (6.10)!) we obtain by relations (6.32) and (6.33) that
∑S−s
g1 (S) = g(S − n)pn . (6.35)
n=0

and since we already showed in the previous section how to compute the sequence g by
means of the F F T method we only need to give a procedure to compute the distribution
of Veq . Observe now for every n ∈ N ∪ {0} and S0 := 0 that
∑∞ ∑n
1 + RFD(R) (n) = P{Sk ≤ n} = r(m)
k=0 m=0

with ∑∞
r(m) := P{Sk = m}.
k=0

73
Hence it follows for every n ∈ N ∪ {0} and R(−1) := 0 that
(1 + RFD(R) (n)) − (1 + RFD(R) (n − 1)) = r(n)
and ∑S−s
1 + RFD(R) (S − s) = r(k).
k=0
By these observations, relation (6.34) and Theorem 67 we obtain
(1 + RFD(R) (n)) − (1 + RFD(R) (n − 1)) r(n)
pn = = ∑S−s (6.36)
1 + RFD(R) (S − s) k=0 r(k)

for every n = 0, 1, ...S − s and this yields by relation (6.35) that

∑S−s
g(S − n)r(n)
g1 (S) = n=0 ∑S−s (6.37)
n=0 r(n)

By relation (6.37) it is now sufficient to compute the generating function of the sequence
{r(n) : n ∈ N ∪ {0}} and this will be done in the next lemma.
Lemma 72 The generating function R(z) of the sequence r := {r(m) : m ∈ N ∪ {0}}
given by relation (6.36) equals
1
Rr (z) = <∞
1 − PD(R) (z)
for every z satisfying |z| < 1 with
PD(R) (z) := E(z D(R) ) = exp(−λR(1 − PY1 (z))).
Proof. By relation (6.36) we obtain
∑∞ ∑∞ ∑∞
R(z) = r(n)z n = P {Sk = n}z n (6.38)
n=0 n=0 k=0
∑∞ ∑∞
= P {Sk = n}z n .
k=0 n=0

For every k ∈ N it follows that

∑∞
P{Sk = n}z n = E(z Sk ) = (E(z X1 ))k = (PD(R) (z))k
n=0

while it is clear using S0 := 0 that

∑∞
P{S0 = n}z n = 1
n=0

Substituting the above expression into relation (6.38) yields

∑∞ 1
Rr (z) = (PD(R) (z))k =
k=0 1 − PD(R) (z)
and this shows the desired result.
To compute the average number of stockouts and the average number of items short we
can immediately apply relations (6.28) and (6.29). Using these relations and relation (6.36)
we obtain ∑S−s
Bstockout (R, S − n)r(n)
Bstockout (R, s, S) = n=0 ∑S−s (6.39)
n=0 r(n)
and ∑S−s
Bitemshort (R, S − n)r(n)
Bitemshort (R, s, S) = n=0 ∑S−s (6.40)
n=0 r(n)
and so by the previous lemma and the F F T method (cf.[21]) one can compute the above
service measures. In the next section we will discuss the last periodic review model.

74
6.3 Analysis of the (R,s,nQ) model.
According to this inventory rule the inventory position process is inspected at every R > 0
time units and a replenishment order is issued if the level IP((nR)−) := limt↑nR IP(nR)
satisfies
IP((nR)−) ≤ s.
If this is not the case no order is triggered and one waits untill the next inspection time. The
order size is chosen to be a multiple of Q such that after ordering all backordered demand
is fulfilled and the inventory position process at time point nR, n ∈ N ∪ {0} satisfies

s < IP(nR) ≤ s + Q.

This means for a divisible item that IP(nR) belongs to {s + 1, ..., s + Q} while for an
indivisible item we obtain that IP(nR) belongs to (s, s + Q]. In this inventory control
problem the variables R, s and sometimes Q are decision variables and need to be chosen
in an optimal way depending upon the chosen cost structure. It is assumed that the demand
process D associated with the (R, s, nQ) inventory control model is a compound Poisson
process with arrival rate λ > 0 and the cumulative demand distribution FY of the individual
demands Yn , n ∈ N satisfies FY (0) = 0 with E(Y1 ) < ∞. If we want to show that the
cadlag process
X = {(IP(t), D(t + L) − D(t)) : t ≥ 0}
is a simplified pure regenerative process with regeneration points nR we need as for the
(R, s, S) model choose the initial distribution

F (x, y) := P{IP(0) ≤ x, D(L) ≤ y}

in such a way that

F (x, y) = P{IP(nR) ≤ x, DnR (L) ≤ y}.
with DnR (L) := D(nR + L) − D(nR). By a similar argument as for the (R, s, S) model
it holds that

P{IP(nR) ≤ x, DnR (L) ≤ y} = P{IP(nR) ≤ x}P {D(L) ≤ y}

and so we need to determine the invariant distribution of the stochastic process {IP(nR), n ∈
N ∪ {0}}. By the definition of the control rule we obtain for every n ∈ N ∪ {0} that

IP((n + 1)R) = IP(nR) − Bn + τ n Q (6.41)

with Bn denoting the random demand in period [nR, (n + 1)R) and

τ n := min{k ∈ N ∪ {0} : s − (IP(nR) − Bn ) ≥ (k − 1)Q} (6.42)

Introducing now the stochastic process V := {Vn : n ∈ N ∪ {0}} given by

Vn := s + Q − IP(nR) (6.43)

it follows for every n ∈ N ∪ {0} that 0 ≤ Vn < Q . Moreover, substituting the definition
of the stochastic process V into relation (6.42) yields

τ n = min{k ∈ N ∪ {0} : s + Q − IP(nR) + Bn ≥ kQ}

= min{k ∈ N ∪ {0} : Vn + Bn ≥ kQ}

and so it follows by relation (6.41) that

Vn+1 = s + Q − IP(nR) + Bn − τ n Q (6.44)

= Vn + Bn − τ n Q = (Vn + Bn )(mod Q).

75
Since we deal with a compound Poisson process the random demands Bn , n ∈ N ∪ {0} are
independent and identically distributed and Bn is independent of Vn and this shows that
the process V is a discrete time Markov process. The next result identifies the invariant
distribution of the discrete time Markov process V.

Theorem 73 The invariant distribution of the discrete time Markov process V = {Vn :
n ∈ N ∪ {0}} with state space [0, Q) and given by relation (6.44) has the form
x
Finv (x) = ,0 ≤ x < Q
Q

while for state space {0, ..., Q − 1} it is given by

x+1
Finv (x) = , x ∈ {0, ..., Q − 1}
Q

Proof. To verify the first part we introduce the sequence mod Bn , n ∈ N ∪ {0} of
independent and identically distributed random variables given by
mod
Bn := Bn (mod Q).

Clearly mod Bn belongs to [0, Q) and by relation (6.44) we obtain that

Vn+1 = (Vn + Bn )(mod Q)(Vn +mod Bn )(mod Q) (6.45)

for every n ∈ N. By this equality we obtain with

Fmod (x) := P {mod Bn ≤ x}, 0 ≤ x < Q

that ∫ Q
P{Vn+1 ≤ x} = P{Vn+1 ≤ x|mod Bn = y}dFmod (y) (6.46)
0
To calculate the above conditional probability we observe by the independence of Vn and
mod
Bn and relation (6.45) that

P{Vn+1 ≤ x|mod Bn = y} = P{(Vn + y)(mod Q) ≤ x}

for every 0 ≤ y < Q and this shows for every 0 ≤ y < x < Q (make a picture !) that

P{Vn+1 ≤ x|mod Bn = y} = P{Vn ≤ x − y or Vn > Q − y} (6.47)

Similarly we obtain for every x ≤ y < Q that (make a picture!}

P{Vn+1 ≤ x|mod Bn = y} = P{Q − y < Vn ≤ x + Q − y} (6.48)

If we assume that the initial distribution is given by

x
P{V0 ≤ x} =
Q
then by relations (6.47) and (6.48) we obtain that
x
P{V1 ≤ x|mod B0 = y} =
Q
for every 0 ≤ y < Q. This shows by relation (6.46) that
∫ Q
x x
P {V1 ≤ x} = dFmod (y) =
Q 0 Q

76
and so this chosen initial cumulative distribution function is an invariant distribution of the
stochastic process V. This shows the first part and since the second part can be verified
similarly we omit the proof of this.
By Theorem 73 it follows for an indivisible item and U a uniform on [0, 1) distributed
random variable that
V0 =d QU =⇒ Vn =d QU
Moreover by the same result we obtain for a divisible item and U a discrete random variable
uniformly distributed on {0, 1, .., Q−1
Q } that

V0 =d QU =⇒ Vn =d QU

and this last result is also proved on page 246 of [15]. By Theorem 73 and using a similar
proof as in Theorem 56 with some obvious modifications one can show the following result.
In this subsection we always denote by U a random variable uniformly distributed on [0, 1]
for an indivisible item and a discrete uniformly distributed random variable on {0, .., Q−1
Q }
for a divisible item.

Theorem 74 For any (R, s, nQ) model with a compound Poisson demand process, lead
time L and complete backordering the associated stochastic process

X : = {(IP(t), D(t + L) − D(t)) : t ≥ 0}

with IP(0) =d s + Q − QU and U independent of D is a simplified pure regenerative

process. The regeneration points of this process are given by nR, n ∈ N ∪ {0} and contain
as a subset the replenishment moments.

To simplify the notation in the remainder in this subsection we introduce with U as in

Theorem 74 the stochastic process D2 = {D2 (t) : t ≥ 0} given by

D2 (t) = D(t) + QU

for every t ≥ 0. As for the (R, S) and (R, s, S) model it follows that

IN(t + L) = s + Q − QU − D(t + L) (6.49)

for every 0 ≤ t < R. Moreover, by the stationary and independent increments of the
compound Poisson process D we obtain that IP(t+nR) is independent of D(t+nR+L)−
D(t+nR) for every n ∈ N∪{0}. It is now possible to identify the average cost Φ(R, s, Q)
of a (R, s, S) model. Observe it is always assumed that the used costrate function f is
nonnegative.

Theorem 75 For any (R, s, nQ) model satisfying the assumptions of Theorem 74 and
∫ R
E(f (s + Q − D2 (t + L)))dt < ∞
0

it follows that the average cost Φ(R, s, Q) of a (R, s, nQ) control policy is given by
( ∫ )
R
Φ(R, s, Q) = R−1 KP{D2 (R) ≥ Q} + E(f (s + Q − D2 (t + L)))dt .
0

Proof. By Theorem 74 we can apply Theorem 48 with σ 1 replaced by R and IP(0)

by s + Q − QU. To write out the formula in Theorem 48 we only need to determine the
probability of a replenishment order at R. Since the event {replenishment order at time R}
is the same as {IP(0) − D(R) ≤ s} and IP(0) equals s + Q − QU it follows that

pr = P{s + Q − QU − D(R) ≤ s} = P{D2 (R) ≥ Q}

77
and this shows the desired result.
As for the (R, s, S) model it follows by relation (6.10), Theorem 75 and U independent
of D that
∫ R
E(f (s + Q − QU − D(t + L)))dt = E(L(f )(s + Q − QU)) (6.50)
0

with L(f )(y) representing the inventory/penalty cost of a (R, y) control rule. Hence also
for this model the inventory/penalty cost is related to the inventory penalty cots of a (R, S)
model. The expression for the average number of backorders per unit of time, the average
physical inventory level per unit of time and the ready rate measure associated with the
(R, s, nQ) model is given by the next result.

Theorem 76 For any (R, s, nQ) model satisfying the assumptions of Theorem 74 it follows
that
∫ R
Bbackorder/time (R, s, Q) = R−1 E(max{D2 (t + L) − (s + Q), 0})dt.
0

Also the average physical stock level per unit of time is given by
∫ R
Bstocklevel/time (R, s, Q) = R−1 E(max{s + Q − D2 (t + L), 0})dt,
0

while the ready rate measure equals

∫ R
Preadyrate (R, s, Q) = R−1 P{s + Q − D2 (t + L) > 0}dt..
0
∫R
Proof. Since 0 E(max{s + Q − D2 (t + L), 0})dt < ∞ the above formulas are an
immediate consequence of Theorem 74 and Lemma 50, 51 and 52 with σ 1 replaced by R
and IP(0) by s + Q − QU.
By Theorem 76 and 58 it is clear that

Preadyrate (R, s, Q) = E(Preadyrate (R, s + Q − QU)) (6.51)

with Preadyrate (R, y) the ready rate measure associated with the (R, y) rule. Similarly we
obtain

Bbackorder/time (R, s, Q) = E(Bbackorder/time (R, s + Q − QU)) (6.52)

and
Bstocklevel/time (R, s, Q) = E(Bstocklevel/time (R, s + Q − QU)) (6.53)
with Bbackorder/time (R, y),respectively Bstocklevel/time (R, y) denoting the average num-
ber of backorders per unit of time and the average physical stock per unit of time in case a
(R, y) control rule is used. If we now want to derive the expressions for the average num-
ber of items short and the average number of stockout occasions and the related service
measures we can use Lemma 53 and 55 applied to the (R, s, nQ) model with the above
initial distribution and this yields the following result.

Theorem 77 For any (R, s, nQ) model satisfying the assumptions of Theorem 74 it follows
that the average number of stockout occasions is given by

Bstockout (R, s, Q) = R−1 (P{D2 (L) ≤ s + Q} − P{D2 (R + L) ≤ s + Q})

78
and the stockout service measure equals
Pstockout (R, s, Q) = P{D2 (L) ≤ s + Q} − P{D2 (R + L) ≤ s + Q}
Moreover, the average number of items short is given by
Bitemshort (R, s, S) = R−1 (E(min{s+Q−D2 (L), 0})−E(min{s+Q−D2 (L+R), 0}))
and the fill rate measure equals
Pf illrate (R, s, Q) = 1 − (λE(Y1 ))−1 Bitemshort (R, s, S).
Proof. Since E(min{s + Q − D2 (L + R), 0} is finite and D2 ((R + L)−) has the
same disrtribution as D2 (R + L) we obtain by Theorem 74 and Lemma 53 and 55 with σ 1
replaced by R and IP(0) by s + Q − QU the desired result.
Aagain it follows by Theorem 77 and 59 that
Bstockout (R, s, Q) = E(Bstockout (R, s + Q − QU)) (6.54)
and
Bitemshort (R, s, Q) = E(Bitemshort (R, s + Q − QU)). (6.55)
with U independent of the functions y → Bstockout (R, y) and y → Bitemshort (R, y)
denoting respectively the average number of stockout occasions and the average number of
items short if a a (R, y) control rule is applied. Again the service measures of this model
are related to the service measures of the (R, S) model. Since the random variable U is
uniformly distributed on {0, .., Q−1
Q } for a divisible item it follows that the random variable
Q − 1 − QU has the same distribution as QU and this shows for a divisible item that
Bstockout (R, s, Q) = E(Bstockout (R, s + 1 + QU)) (6.56)
and
Bitemshort (R, s, Q) = E(Bitemshort (R, s + 1 + QU)) (6.57)
Moreover, for an indivisible item it follows that the random variable U is uniformly dis-
tributed on (0, 1] and so the random variable Q − QU has the same distribution as QU.
This shows for an indivisible item that
Bstockout (R, s, Q) = E(Bstockout (R, s + QU)) (6.58)
and
Bitemshort (R, s, Q) = E(Bitemshort (R, s + QU)) (6.59)
Again in the next section we will consider how to compute the different objectives for the
(R, s, nQ) model.

6.3.1 Computation of the objectives in a (R,s,nQ) model.

We will now focus on the computation of the different cost functions for the average cost in
case we are dealing with a divisible item. If this holds it follows that the decision variables
s and Q are nonnegative integers and we need to evaluate (see Theorem 75) the sequence
g2 := {g2 (n) : n ∈ N ∪ {0}} given by
∫ R
g2 (n) := E(f (n − QU − D(t + L))dt (6.60)
0

Since
g2 (n) = E(L(f )(n − QU)) (6.61)
with Lf given by relation (6.10) and U independent of the sequence {Lf (n) : n ∈ N ∪
{0}} it is easy to show the following result whenever the sequence {f (n) : n ∈ N ∪ {0}}
satisfies f (n) = 0 for every n ≤ −1.

79
Theorem 78 If f : Z → R is a cost rate sequence satisfying f (n) = 0 for every n ≤ −1
then it follows that the generating function Rg2 (z) of the sequence g2 equals

(1 − z Q )Rf (z) exp(−λL(1 − PY1 (z)))

(1 − exp(−λR(1 − PY1 (z))))
Q(1 − z)λ(1 − PY1 (z))

Proof. By relation (6.50) and relation (6.12) we obtain that the generating function of
the sequence {g2 (n) : n ∈ N ∪ {0}} is given by
∑∞
Eg(n − QU)z n = Rg (z)E(z QU ) (6.62)
n=0

Since the random variable U is uniformly distributed on {0, 1, .., Q−1

Q } it is easy to verify
that
1 ∑Q−1 k 1 − zQ
E(z QU ) = z =
Q k=0 Q(1 − z)
and using relation (6.62) and Theorem 60 the desired result follows.
In case penalty cost are introduced within the (R, s, nQ) model we already know that
these penalty costs are mostly linear and this means for a selected value p > 0 that the used
cost rate sequence f (n) satisfies

f (n) = −pn, n = −1, −2, ..

Introduce now as done for the (R, S) model the cost rate sequence

f1 (n) := f (n) + pn, n ∈ Z

This sequence vanishes on {−1, −2, ..} and for f2 := f − f1 it follows that the expectation
∫ R
E(f2 (s + Q − QU − D(t + L))dt
0

equals
∫ R
−pR(s + Q) + p E(QU + D(t + L))dt = A
0
with
Q+1 pλE(Y1 )
A := −pR(s + )+ ((R + L)2 − L2 )
2 2
and so the expectation
∫ R
E(f (s + Q − QU − D(t + L))dt
0

is given by
∫ R
E(f1 (s + Q − QU − D(t + L))dt + A.
0
Computing now the first expression by the F F T method is possible and this completes
our discussion of the computation of the average costs. Using relations (6.56) and (6.58)
one can also evaluate the different service measures as discussed in Theorem 77 and so
our discussion of the different periodic review models is finished. In the next subsection
we start with the continuous review models. The first model to be considered is the (s, S)
inventory control model.

80
6.4 Analysis of the (s,S) model.
Under this rule an order is triggered at the moment the level of the inventory position pro-
cess drops strictly below level s ≤ S and the size of the order is such that all excess demand
is backordered and the level of the inventory position process is raised to order-up-to-level
S. Hence the variables s and S are decision variables and they need to be chosen in an
optimal way depending upon the chosen cost structure. It is assumed that the demand pro-
cess D associated with the (s, S) inventory control model is a compound renewal process
with arrival rate λ > 0 and the associated cumulative demand function FY of the indepen-
dent and identically distributed individual demands Yn , n ∈ N satisfies FY (0) = 0 and
E(Y1 ) < ∞. This means that
∑N(t)
D(t) = Yn , Y0 := 0
n=0

with the arrival process N := {N(t) : t ≥ 0} given by

N(t) := max{n ∈ N ∪ {0} : σ n ≤ t}.

∑n
with σ n := k=0 Tk , T0 := 0 and n ∈ N ∪ {0}. The interarrival times Tk , k ∈ N
are independent and identically distributed with right continuous cumulative distribution
function FT satisfying FT (0) = 0 and 0 < E(T1 ) = λ−1 < ∞. Also it is assumed that
the demands Yn , n ∈ N are independent of the arrival process N. If we like to show that
the stochastic process

X := {(IP(t), D(t + L) − D(t)) : t ≥ 0}

is a pure regenerative process with regeneration points σ n , n ∈ N ∪ {0} we need to deter-

mine as for the (R, s, S) model the distribution function

F (x, y) := P{IP(0) ≤ x, D(L) ≤ y} (6.63)

satisfying
F (x, y) = P{IP(σ n ) ≤ x, D(σ n + L) − D(σ n ) ≤ y} (6.64)
for every n ∈ N. Since the random time σ n denotes the arrival moment of the nth arriving
customer and

D(σ n + L) − D(σ n ) = total demand in interval (σ n , σ n + L]

we obtain by our assumption on the demand process that D(σ n + L) − D(σ n ) is inde-
pendent of {D(t) : t ≤ σ n } and has the same distribution as D(L). Moreover it follows
that the random variable IP(σ n ) is completely determined by its initial value at time 0
and the realization of the demand process {D(t) : t ≤ σ n }. This shows by our previous
observation that for every n ∈ N the random variable D(σ n + L) − D(σ n ) is independent
of IP(σ n ) and it has the same distribution as D(L). Hence we obtain for every n ∈ N that

P{IP(σ n ) ≤ x, D(σ n + L) − D(σ n ) ≤ y} = P{IP(σ n ) ≤ x}P{D(L) ≤ y}

and this implies by relations (6.63) and (6.64) that we need to find the invariant distribution
of the stochastic process {IP(σ n ) : n ∈ N ∪ {0}}. To find this invariant cumulative
distribution function it is as for the related (R, s, S) model more convenient to consider the
stochastic process V = {Vn : n ∈ N ∪ {0}} given by

Vn := S − IP(σ n )

As for the (R, s, S) model it is easy to verify by the definition of the control rule that

Vn+1 = (Vn + Yn+1 )1{Vn +Yn+1 ≤S−s} (6.65)

81
with Yn+1 denoting the demand of the (n + 1)th arriving customer. By our assumption
on the demand process the random variable Yn+1 is independent of the random variable
Vn and since the random variables Yk , k ∈ N are independent and identically distributed
it follows that the process V is a discrete time Markov process. Copying now the proof of
Theorem 67 one can show the following result.

Theorem 79 The invariant distribution of the discrete time Markov process V = {Vn :
n ∈ N ∪ {0}} given by relation (6.65) has the form

1 + RFY (v)
Finv (v) = ,0 ≤ v ≤ S − s (6.66)
1 + RFY (S − s)

with RFY the pure renewal function associated with the cumulative distribution function of
the random variable Y1 .

Applying now Theorem 79 one can show the following key result.

Theorem 80 For any (s, S) model (s ≤ S) with a compound renewal demand process
having arrival rate 0 < λ < ∞, lead time L and complete backordering the associated
stochastic process
X : = {(IP(t), D(t + L) − D(t)) : t ≥ 0}
with IP(0) =d S − Veq , Veq independent of the demand process D and the distribution
of Veq given by relation (6.66) is a∑pure regenerative process. The regeneration points
n
of this process are given by σ n = k=0 Tk , n ∈ N ∪ {0} and contain as a subset the
replenishment moments.

Proof. Clearly the random variables σ n+1 − σ n , n ∈ N ∪ {0} form a pure renewal pro-
cess. Moreover, the inventory position IP(t + σ n ) at time t + σ n is completely determined
by the random variable IP(σ n ) and the shifted stochastic process Dσn = {Dσn (s) : s ≥
0} with
Dσn (s) := D(s + σ n ) − D(σ n )
up to time t. This implies that there exists some maybe complicated deterministic function
g satisfying
IP(t + σ n ) = g(IP(σ n ), {Dσn (s) : 0 ≤ s ≤ t}) (6.67)
and this yields for every n ∈ N that

Xσn = {(IP(t + σ n ), D(t + σ n + L) − D(t + σ n )) : t ≥ 0}

= {(g(IP(σ n ), {Dσn (s) : s ≤ t}), Dσn (t + L) − Dσn (t)) : t ≥ 0}

By the chosen initial distribution it follows using Theorem 79 that

IP(σ n ) =d S − Veq

and together with the observation that the shifted stochastic process Dσn has the same
distribution as D we obtain by relation (6.67) that the distribution of the process Xσn is
the same as the distribution of the process X. Finally the shifted demand process {Dσn (s) :
0 ≤ s ≤ t} is independent of T1 , ..., Tn and since the inventory position process at σ n
is completely determined by the demands Y1 , ..., Yn and IP(0) = S − Veq with Veq
independent of D it also holds that the the process g(IP(σ n ), {Dσn (s) : s ≥ 0}) is
independent of T1 , .., Tn . Hence it follows that the shifted stochastic process Xσn defined
above is independent of σ 0 , .., σ∑
n and this shows that the process X is a pure regenerative
n
process with regeneration points k=0 Tk . The last part of the above result is immediately
clear and so the desired result is verified.

82
 In case s = S it follows by Theorem 80 that the random variable Veq is degenerate at
0. Again we obtain

IN(t + L) = IP(t) − (D(t + L) − D(t)) (6.68)

=d S − Veq − D(t + L)

for every 0 ≤ t < σ 1 = T1 . Before discussing the average cost of a (s, S) model with
a compound renewal demand process we first consider the special case of the (S, S) in-
ventory control model. In this model each arrival of a customer triggers a replenishment
order and such a control rule is often used in spare parts inventory management. More-
over, in the approximative M ET RIC approach3 for multi-echelon systems (cf. [36],[32]
) this model also plays a prominent rule since for a compound Poisson demand process the
departure process of replenishment orders is the same as the compound Poisson demand
process. Using now Theorem 80 one can show the following average cost result for the
(S, S) model.

Theorem 81 For any (S, S) model satisfying the assumptions of Theorem 80 the average
cost Φ(S) of a (S, S) control policy is given by
(∫ )
T1
Φ(S) = λK + λE f (S − D(t + L))dt . (6.69)
0

Proof. By Theorem 80 we can apply Theorem 48 with IP(0) replaced by S (Use

Veq = 0 with probability one!) and σ 1 by T1 .
In case we consider a (S, S) model with a compound Poisson demand process and
satisfying the assumptions of Theorem 80 we can simplify the expression for the average
cost in relation (6.69). This is shown by the following result.

Theorem 82 For any (S, S) model with a compound Poisson demand process and satisfy-
ing the assumptions of Theorem 80 it follows that the average cost Φ(S) of a (S, S) control
policy is given by
Φ(S) = λK + E(f (S − D(L)).

Proof. Since T(1 denotes the arrival epoch

) of the first customer after time 0 we obtain
∫ T1
that the integral E 0 f (S − D(t + L))dt equals
∫ ∞
E(f (S − (D(t + L) − D(t)))1{D(t)=0} )dt. (6.70)
0

By the compound Poisson process assumption the random variable D(t + L) − D(t) is
independent of the random variable D(t) and has the same distribution as D(L). This
shows

E(f (S − (D(t + L) − D(t)))1{D(t)=0} ) = E(f (S − D(L))E(1{D(t)=0} )

(∫ )
T
and so E 0 1 f (S − D(t + L))dt reduces to
(∫ ∞ )
E(f (S − D(L))E 1{D(t)=0} dt = E(f (S − D(L))E(T1 ).
0

Using now Theorem 81 yields the desired result.

3 Metric approach for multi-echelon systems.

83
 Looking at relation (6.12) it follows for any sequence f := {f (n) : n ∈ Z} with
f (n) = 0 for every n ∈ {−1, −2...} that
∑∞
E(f (n − D(L))z n = Rf (z) exp(−λL(1 − PY1 (z)))
k=0

in case the demand process is a compound Poisson process. Applying now the same argu-
ment for sequences f satisfying f (n) = −pn for every n ≤ −1 (see section (R, S) model!)
one can evaluate relatively easy the average cost of a (S, S) model. We will now discuss
the same numerical example as for the (R, S) model and show in the next example the
graph of the average cost of a (S, S) control policy as a function of S.

Example 83 In this numerical we assume that L = 54 days, K = $30 and the linear
holding costs per day are given by 0.15 times $39. The linear penalty costs are 0.3 times
$39 per day with $39 representing the value of the item under consideration . Moreover, it is
assumed that the demand process is a compound Poisson process with expected interarrival
times equal to 36.9 days. We know that the the lead time is given by 54 days and from past
data the empirical distribution function of the individual demand is given

k empirical probabilities
1 9/63
2 14/63
3 5/63
4 6/63
5 7/63
6 7/63
7 1/63
8 1/63
9 1/63
10 8/63
12 2/63
20 2/63

by the next table. For these data the average cost for the (S, S) model is computed
as a function of the order-up-to level S by using the FFT procedure of MATLAB and this
function is shown in Figure 2 at the top of this page.

We will continue with the determination of the average cost function of a (s, S) model.
As for the (R, s, S) model we introduce now the stochastic process D1 = {D1 (t) : t ≥ 0}
given by
D1 (t) := D(t) + Veq
with Veq independent of the demand process D and the distribution of Veq given by rela-
tion (6.66).

Theorem 84 For any (s, S) model (s ≤ S) satisfying the assumptions of Theorem 80 and
(∫ )
T1
E f (S − D1 (t + L))dt <∞
0

it follows that the average cost Φ(s, S) of a (s, S) control rule is given by
(∫ )
T1
Φ(s, S) = λKP{Y1 + Veq > S − s} + λE f (S − D1 (t + L)dt .
0

84
300

250

200

150

Figure 6.2: Average cost S −→ Φ(S).

100

Proof. By Theorem 80 we can use Theorem 48 with σ 1 replaced by T1 and IP(0) by

50
S − Veq . To write out the formula in Theorem 48 we only need to determine the probability
of a replenishment order at T1 . Since the event {replenishment order at time T1 } is the
same as {IP(0) − D(T1 ) < s}, IP(0) equals S − Veq and D(T1 ) equals Y1 it follows
0 that
pr = P{IP(0) S− D(T1 ) ≤ s} = P{S − Veq − Y1 < s}
0 10 20 30 40 50 60 70

and this shows the desired result.

As in Theorem 82 one can simplify the average cost of the (s, S) model in case the
demand process is a compound Poisson process.

Theorem 85 For any (s, S) model (s ≤ S) with a compound Poisson demand process
and satisfying the assumptions of Theorem 80 it follows that the average cost Φ(s, S) of a
(s, S) control rule is given by

Φ(s, S) = λKP{Y1 + Veq > S − s} + E(f (S − D1 (L)).

The expression for the average number of backorders per unit of time, the average
physical inventory level per unit of time and the ready rate measure associated with the
(s, S) model are given by the next result.

Theorem 86 For any (s, S) model (s ≤ S) satisfying the assumptions of Theorem 80 it

follows that
(∫ )
T1
Bbackorder/time = λE max{D1 (t + L) − S, 0}dt .
0

Also the average physical stock level per unit of time has is given by
(∫ )
T1
Bstocklevel/time (s, S) = λE max{S − D1 (t + L), 0}dt ,
0

85
while the ready rate measure equals
(∫ )
T1
Preadyrate (s, S) = λE 1{S−D1 (t+L)>0} dt .
0

Proof. The above formuals are an immediate consequence of Theorem 80, relation
(6.68) and Lemma 50, 51 and 52 with σ 1 replaced by T1 and IP(0) by S − Veq .
By Theorem 86 and the observations on the (S, S) inventory model it is easy to verify
that
Bbackorder/time (s, S) = E(Bbackorder/time (S − Veq )) (6.71)
and
Bstocklevel/time (s, S) = E(Bstocklevel/time (S − Veq )) (6.72)
with Bbackorders (y), respectively Bstocklevel (y) denoting for the (y, y) control policy the
average number of backorders per unit of time and the average physical inventory per unit
of time. Similarly it follows that

Preadyrate (s, S) = E(Preadyrate (S − Veq )) (6.73)

with Preadyrate (y) for the (y, y) control policy the fraction of time the inventory level
is strictly positive. As in the proof of Theorem 85 we obtain for a (s, S) model with a
compound Poisson demand process that the expressions can be simplified and in that case
these expressions are given by

Bbackorder/time (s, S) = E(max{D(L) − (S − Veq ), 0}) (6.74)

and
Bstocklevel/time (s, S) = E(max{S − Veq − D(L), 0}). (6.75)
Moreover, the fraction of time Preadyrate (s, S) the inventory level is positive is given by

Preadyrate (s, S) = P{S − Veq − D(L) > 0}. (6.76)

If we want to derive the different expression for the average number of items short and
the average number of stockout occasions and the associated service measures then we can
use Lemma 53 and 55 applied to the (s, S) model with the above initial distribution and
so we obtain the following result. Remember the random variables D(L), Y1 and Veq are
independent.

Theorem 87 For any (s, S) model satisfying the assumptions of Theorem 80 it follows that
the average number of stockout occasions is given by

Bstockout (s, S) = λ (P{D1 (L) ≤ S} − P{D1 (L) + Y1 ≤ S})

and the stockout service measure equals

Pstockout (s, S) = P{D(L) + Veq ≤ S} − P{D(L) + Y1 + Veq ≤ S}.

Moreover, the average number of items short is given by

Bitemshort (s, S) = λ(E(min{S − D1 (L), 0}) − E(min{S − D1 (L) − Y1 , 0})))

and the fill rate measure equals

Pf illrate (s, S) = 1 − (λE(Y1 ))−1 Bitemshort (s, S).

86
 Proof. By Theorems 80, Lemma 53 and 55 we obtain with IP(0) replaced by S − Veq ,
σ 1 by T1 and D((T1 + L)−) has the same distribution as D(T1 + L) for every L > 0
that the average number of stockout is given by

E(T1 )−1 (P{D(L) ≤ S − Veq } − P{D(T1 + L) ≤ S − Veq })

and the average number of items short by

Bitemshort (s, S) = E(T1 )−1 (E(min{S−Veq −D(L), 0})−E(min{S−Veq −D(T1 +L), 0}).

By the definition of T1 and the compound renewal demand process it follows that

D(T1 + L) = D(T1 + L) − D(T1 ) + D(T1 )

= D(T1 + L) − D(T1 ) + Y1
=d D̃(L) + Y1

with D̃(L) independent of Y1 and D̃(L) has the same distribution as D(L). Substituting
this into the above expressions yields the desired result for the average number of stockouts
and the average number of items short. The formulas for Pstockout (s, S) and Pf illrate (s, S)
are an immediate consequence
Again by the above result the service measures of a (s, S) model are related to the
service measures of a continuous review model where the demand of every customer gen-
erates a replenishment order. In particular we obtain for such a model that Veq = 0 with
probability one and this yields by Theorem 87 that

Bstockout (s, S) = E(Bstockout (S − Veq )) (6.77)

with
Bstockout (y) := λ(P{D(L) ≤ y} − P{D(L) + Y1 ≤ y}) (6.78)
the average number of stockout occasions for a (y, y) control rule. Moreover, it follows
that
Bitemshort (s, S) = E(Bitemshort (S − Veq )) (6.79)
with Bitemshort (y) the average number of items short for a (y, y) control rule. In the next
subsection we will show how to compute the different objectives of the (s, S) model.

6.4.1 Computation of the objectives in a (s,S) model.

Similarly as for the (R, s, S) models and a compound Poisson demand process one can
compute for a divisible item the average cost by observing that
∑S−s
E(f (S − Veq − D(L)) = Ef (n − k − D(L))pk
k=0

(For more details the reader is referred to the section on the (R, s, S) model!). At the
same time this shows that the inventory/penalty costs of a (s, S) model is related to the
inventory/penalty costs of a continuous review model in which every arriving customer
generates a replenishment order. In the next example we will again show the graph of the
average cost as as function of the parameters s and S.

Example 88 In this numerical we assume that L = 54 days, K = $30 and the linear
holding costs per day are given by 0.15 times $39. The linear penalty costs are 0.3 times
$39 per day with $39 representing the value of the item under consideration . Moreover, it is
assumed that the demand process is a compound Poisson process with expected interarrival

87
600

500

400

300

200
Figure 6.3: Average cost (s, S) −→ Φ(s, S).

100
times equal to 36.9 days. We know that the the lead time is given by 54 days and from past
0 data the empirical distribution function of the individual demand is given
80

60 k empirical probabilities 60
70

40 1 9/63 40
50

2 30
14/63
20 20
10
0 0 3 5/63
4 6/63
5 7/63
6 7/63
7 1/63
8 1/63
9 1/63
10 8/63
12 2/63
20 2/63

by the next table. For these data a two-dimensional plot is given of the average cost for
the (s, S) model is computed as a function of the two decision variables by using the FFT
procedure of MATLAB and this function is shown in Figure 3 at the next page.

This concludes our analysis of the (s, S) model and in the last subsection we will con-
sider the (s, nQ) model.

6.5 Analysis of the (s,nQ) model.

According to this inventory rule an order is triggered at the moment the inventory position
drops below or equals s. The order size is a multiple of Q such that after ordering all
excess demand is backordered and the inventory position process is between s and s + Q.
Hence the variables s and possibly Q are decision variables and they need to be chosen in

88
an optimal way depending upon the chosen coststructure. It is assumed that the demand
process D associated with the (s, nQ) inventory control model is a compound renewal
process with positive arrival rate λ < ∞ and the associated cumulative demand function
FY of the independent and identically distributed individual demands Yn , n ∈ N satisfies
FY (0) = 0 and E(Y1 ) < ∞. This means that
∑N(t)
D(t) = Yn , Y0 := 0
n=0

with the arrival process N := {N(t) : t ≥ 0} given by

N(t) := max{n ∈ N ∪ {0} : σ n ≤ t}

∑n
with σ n := k=0 Tk , T0 := 0 and n ∈ N∪{0}. Observe the interarrival times Tk , k ∈ N
are independent and identically distributed with right continuous cumulative distribution
function FT satisfying FT (0) = 0 and E(T1 ) = λ−1 < ∞. Also it is assumed that the
demands Yn , n ∈ N are independent of the arrival process N. Since σ n represents the
arrival time of the nth customer it follows by the above definition of the ordering policy
that a replenishment order is issued at time σ n if and only if

IP(σ n −) − Yn ≤ s

At the same time it follows that

s < IP(σ n ) ≤ s + Q

and this means for a divisible item that IP(σ n ) belongs to the state space {s + 1, .., s + Q}.
If we want to show that the cadlag process

X = {(IP(t), D(t + L) − D(t) : t ≥ 0}

is a pure regenerative process with regeneration points σ n we need similarly as for the
(R, s, S) model choose the initial distribution

F (x, y) := P {IP(0) ≤ x, D(L) ≤ y} (6.80)

in such a way that

F (x, y) = P {IP(σ n ) ≤ x, DnR (L) ≤ y} (6.81)
for every n ∈ N with DnR (L) := D(nR + L) − D(nR). Since the random time σ n
denotes the arrival time of the nth arriving customer and

D(σ n + L) − D(σ n ) := total demand in interval (σ n , σ n + L]

we obtain since D is a compound renewal process that the random variable D(σ n + L) −
D(σ n ) is independent of {D(t) : t ≤ σ n } and this random variable has the same distri-
bution as D(L). Moreover, the random variable IP(σ n ) is completely determined by its
initial value at 0 and the demand process {D(t) : 0 ≤ t} up to time σ n and this shows that
random variable D(σ n + L) − D(σ n ) is independent of IP(σ n ). Hence it follows that

P{IP(σ n ) ≤ x, DnR (L) ≤ y} = P{IP(σ n ) ≤ x}P{D(L) ≤ y} (6.82)

for every n ∈ N and this yields by relations (6.80) up to (6.82) that we like to find the
invariant distribution of the stochastic process {IP(σ n ) : n ∈ N∪{0}}. As in the (R, s, S)
model it follows by the definition of the control rule that

IP(σ n+1 ) = IP(σ n ) − Yn+1 + τ n Q (6.83)

89
with Yn+1 the demand of the (n + 1)th arriving customer and

τ n := min{k ∈ N ∪ {0} : s − (IP(σ n ) − Yn+1 ) ≥ (k − 1)Q}

Introducing now the stochastic process V := {Vn : n ∈ N ∪ {0}} given by

Vn := s + Q − IP(σ n )

it follows that 0 ≤ Vn < Q. Also by substituting the definition of the random variable Vn
into the random variable τ n we obtain that

τ n = min{k ∈ N ∪ {0} : s + Q − (IP(σ n ) − Yn+1 ) ≥ kQ}

= min{k ∈ N ∪ {0} : Vn + Yn+1 ≥ kQ}

and this shows by relation (6.83) that

Vn+1 = s + Q − IP(σ n ) + Yn+1 − τ n Q (6.84)

= Vn + Yn+1 − τ n Q = (Vn + Yn+1 )(mod Q)

The random demands Yn , n ∈ N are by assumption independent and identically distributed

and the random variable Yn+1 is independent of Vn . This shows that the process V is a
discrete time Markov process and this process has the same structure as the associated
process V for the (R, s, S) model. By a similar proof as used in Theorem 73 one can
identify the invariant distribution of the stochastic process V.

Theorem 89 The invariant distribution of the discrete time Markov process V = {Vn :
n ∈ N ∪ {0}} with state space [0, Q) and given by relation (6.84) has the form
x
Finv (x) = ,0 ≤ x < Q
Q
while for state space {0, ..., Q − 1} it is given by
x+1
Finv (x) = , x ∈ {0, .., Q − 1}
Q
By Theorem 89 it follows for an indivisible item and U a uniform on [0, 1) distributed
random variable that
V0 =d QU =⇒ Vn =d QU
Moreover by the same result we obtain for a divisible item and U a discrete random variable
uniformly distributed on {0, 1, .., Q−1
Q } that

V0 =d QU =⇒ Vn =d QU

and this last result is also proved on page 246 of [15]. By Theorem 89 and copying the
proof of Theorem 80 one can show the following result. Observe in this subsection we
always denote by U a random variable uniformly distributed on [0, 1] for an indivisible
item and a discrete uniformly distributed random variable on {0, .., Q−1
Q } for a divisible
item.

Theorem 90 For any (s, nQ) model with a compound renewal demand process, lead time
L and complete backordering the associated stochastic process

X : = {(IP(t), D(t + L) − D(t)) : t ≥ 0}

with IP(0) =d s + Q − QU and U independent of the demand process D is∑a pure regen-
n
erative process. The regeneration points of this process are given by σ n = k=0 Tk , n ∈
N ∪ {0} and contain as a subset the replenishment moments.

90
 Again it follows for a (s, nQ) model that the above stochastic process X satisfies

IN(t + L) = IP(t) − (D(t + L) − D(t)) (6.85)

=d s + Q − QU − D(t)

with the random variable U independent of the demand process D. Also we obtain by the
compound renewal demand process and the definition of the regeneration points σ n that
the random variable IP(t + σ n ) is independent of D(t + σ n + L) − D(t + σ n ) for every
t ≥ 0 and n ∈ N. To simplify the notation in this subsection we introduce as for the
(R, s, nQ) model the stochastic process D2 := {D2 (t) : t ≥ 0} given by

D2 (t) := D(t) + QU

with U as in Theorem 90. Applying now relation (6.85) and Theorem 90 and copying the
proof of Theorem 75 it is possible to identify the average cost Φ(s, Q) of a (s, nQ) control
rule.

Theorem 91 For any (s, nQ) model satisfying the assumptions of Theorem 90 and
(∫ )
T1
E f (s + Q − D2 (t + L))dt <∞
0

it follows that the average cost Φ(s, nQ) of a (s, nQ) control rule is given by
(∫ )
T1
Φ(s, nQ) = λKP{QU + Y1 ≥ Q} + λE f (s + Q − D2 (t + L))dt .
0

Proof. By Theorem 90 we can use Theorem 48 with σ 1 replaced by T1 and IP(0)

by s + Q − QU. To write out the formula in Theorem 48 we only need to determine the
probability of a replenishment order at T1 . It follows that the event {replenishment order
at time T1 } is the same as the event {IP(0) − D(T1 ) ≤ s} and since IP(0) equals
s + Q − QU and D(T1 ) equals Y1 this implies

pr = P{QU + Y1 ≥ Q}.

Applying now Theorem 48 and relation (6.85) yields the desired result.
Again for a compound Poisson demand process one can simplify the expression for the
average cost Φ(s, nQ) by exactly the same method as used in the proof of Theorem 85.

Theorem 92 For any (s, nQ) model with a compound Poisson process and satisfying the
assumptions of Theorem 90 it follows that the average cost Φ(s, nQ) of a (s, nQ) control
rule is given by

Φ(s, nQ) = λKP{QU + Y1 ≥ Q} + E(f (s + Q − D2 (L))).

Again the inventory/penalty costs of a (s, nQ) model is related to the inventory/penalty
costs of a continuous review model in which every arriving customer generates a demand.
The expression for the average number of backorders per unit of time, the average physical
inventory level per unit of time and the ready rate measure associated with the (s, nQ)
model are given by the next result.

Theorem 93 For any (s, nQ) model satisfying the assumptions of Theorem 90 it follows
that
(∫ )
T1
Bbackorder/time (s, Q) = λE max{D2 (t + L) − (s + Q), 0}dt .
0

91
Also the average physical stock level per unit of time is given by
(∫ )
T1
Bstocklevel/time (s, Q) = λE max{s + Q − D2 (t + L), 0}dt
0

while the ready rate measure equals

(∫ )
T1
Preadyrate (s, Q) = λE 1{s+Q−D2 (t+L)>0} dt .
0

Proof. The above formuals are an immediate consequence of Theorem 90 and Lemma
50, 51 and 52 with σ 1 replaced by T1 and IP(0) by s + Q − QU.
Again we observe that
Bbackorders (s, Q) = E(Bbackorders (s + Q − QU)) (6.86)
and
Bstocklevel (s, Q) = E(Bstocklevel (s + Q − QU)) (6.87)
with Bbackorders (y), respectively Bstocklevel (y) denoting the average number of backo-
rders per unit of time and the average physical inventory per unit of time when using a
(y, y) control policy. Similarly it follows that
Preadyrate (s, Q) = E(Preadyrate (s + Q − QU)) (6.88)
with Preadyrate (y) the fraction of time the inventory level is strictly positive when using
a (y, y) inventory rule. As in the proof of Theorem 82 we also obtain for a (s, Q) model
with a compound Poisson demand process that the expressions can be simplified and they
are given by
Bbackorders (s, Q) = E(max{D(L) − (s + Q − QU), 0}) (6.89)
and
Bbackorders (s, Q) = E(max{s + Q − QU − D(L), 0}). (6.90)
Moreover, the fraction of time Preadyrate (s, Q) the inventory level is positive is given by
Preadyrate (s, Q) = P{s + Q − QU − D(L) > 0}. (6.91)
If we want to derive the different expression for the average number of items short and the
average number of stockout occasions and the associated service measures then we can use
Lemma 53 and 55 applied to the (s, Q) model with the above initial distribution and so we
obtain the following result.
Theorem 94 For any (s, nQ) model satisfying the assumptions of Theorem 90 it follows
that the average number of stockout occasions is given by
Bstockout (s, Q) = λ(P{D2 (L) ≤ s + Q} − P{D2 (L) + Y1 ≤ s + Q})
and the stockout service measure equals
Pstockout (s, Q) = P{D2 (L) ≤ s + Q} − P{D2 (L) + Y1 ≤ s + Q}.
Moreover, the average number of items short is given by
Bitemshort (s, Q) = λE(min{s + Q − D2 (L), 0}) − E(min{s + Q − D2 (L) − Y1 , 0}))
and the fill rate measure equals
Pf illrate (s, S) = 1 − (λE(Y1 ))−1 Bitemshort (s, S).

92
 Again we obtain by the above result that

Bstockout (s, Q) = E(Bstockout (s + Q − QU))

with Bstockout (x) given by relation (6.78). Moreover, it follows that

Bitemshort (s, Q) = E(Bitemshort (s + Q − QU))

with Bitemshort (x) given by relation (6.79). In the next subsection we discuss the compu-
tation of the different objectives in the (s, Q) model.

6.5.1 Computation of the objectives in a (s,Q) model.

To compute the above average cost for a divisible item we first focus on the first term in the
average cost expression of Theorem 91 given by

P{QU + Y1 ≥ Q}

Since in this case it follows that Q ∈ N and U is a discrete uniform random variable on
{0, Q
1
Q } it is easy to check for every k ∈ {0, .., Q − 1} that
, .., Q−1

1
P{Q − 1 − QU = k} = P{QU = Q − 1 − k} =
Q
and so Q − 1 − QU has the same distribution as QU. By this observation and Y1 integer
valued we obtain that

P{QU + Y1 ≥ Q} = P {Y1 > Q − 1 − QU} = P {Y1 > QU}

1 ∑Q−1
= P{Y1 > k}
Q k=0

and so we can compute the above expression if we can evaluate the sequence
∑n
r := { P {Y1 > k} : n ∈ N ∪ {0}}. (6.92)
k=0

The generating function of the sequence r given by relation (6.92) can easily be derived
from relation (6.15) and so we mention without proof the following result..

Lemma 95 The generating function Rr (z) of the sequence r given by relation (6.92) has
the form
1 − PY (z)
Rr (z) =
(1 − z)2
with PY (z) := E(z Y1 ) denoting the generating function of the distribution of Y1 .

In case the demand process is a compound Poisson process we also need to evaluate by
Theorem 92 the sequence r:= {r(n) : n ∈ N ∪ {0}} given by

r(n) := E(f (n − QU − D(L)) = E(f (n − (QU + D(L))). (6.93)

For this sequence one can show the following result.

Lemma 96 If f : Z → R is a cost rate sequence satisfying f (n) = 0 for every n ≤ −1

then it follows that the generating function Rr (z) of the sequence r defined by relation
(6.93) is given by

Rf (z)(1 − z Q ) exp(−λL(1 − PY (z)))

Rr (z) =
Q(1 − z)

93
 Proof. By a similar argument as in relation (6.12) it follows that Rr (z) equals
∑∞
z n E(f (n − (QU + D(L))) = Rf (z)E(z QU+D(L) ) (6.94)
n=0

Since the random variables QU and D(L) are independent we obtain that

E(z QU+D(L) ) = E(z QU )E(z D(L) )

and since ∑Q−1

E(z QU ) = Q−1 z n = (Q(1 − z))−1 (1 − z Q )
n=0
we obtain that
(1 − z Q ) exp(−λL(1 − PY (z)))
E(z QU+D(L) ) =
Q(1 − z)
This shows by relation (6.94) the desired result.
As for the (R, S) model we observe in case penalty costs are introduced that these are
mostly linear and so we obtain that the used cost rate function f satisfies

f (n) = −pn, n = −1, −2, ..

To reduce the case of linear penalty cost to the case of no penalty costs we introduce as
before the cost rate sequence f1 := {f1 (n) : n ∈ Z} given by

f1 (n) = f (n) + pn.

This sequence vanishes on {−1, −2, ..} and it follows that the difference

E(f (n − (QU + D(L))) − E(f1 (n − (QU + D(L))) (6.95)

equals
Q−1
−pn + E(QU + D(L)) = −pn + + λE(Y1 )L
2
Using now relation (6.95) and lemma 96 it is easy to evaluate the expression r(n) given by
relation (6.93). Finally we will focus on the evaluation of the average number of items short
in a (s, nQ) model in case we deal with a compound Poisson demand process. It is now
possible to derive the following result and this result is an easy consequence of relations
(6.39) and (6.18).

Lemma 97 It follows that the the generating function Rr (z) of the sequence r:= {r(n) :
n ∈ N ∪ {0}} with

r(n) := E(max{n − (D(L) + Y1 + QU), 0}) − E(max{n − (D(L) + QU), 0})

and D a compound Poisson demand process is given by

−z exp(−λL(1 − PY (z)))(1 − PY (z))(1 − z Q )

Rr (z) =
Q(1 − z)3

Proof. It is clear that

r(n) = E(f0 (n − (D(L) + Y1 + QU)) − E(f0 (n − (D(L) + QU)))

with the cost rate sequence f0 := {f0 (n) : n ∈ Z} given by f0 (n) := n for every
n ∈ N ∪ {0} and zero otherwise. Also we know for any nonnegative random variable D
that ∑∞
E(f0 (n − D))z n = Rf0 (z)E(z D )
n=0

94
and since ∑∞ z
Rf0 (z) = nz n =
n=0 (1 − z)2
this implies
∑∞ zE(z D )
E(f0 (n − D))z n =
n=0 (1 − z)2
By the independence of the random variables D(L), Y1 and QU it follows by the the
previous formula that
∑∞ zE(z D(L) )E(z Y1 )E(z QU )
E(f0 (n − (D(L) + Y1 + QU))z n =
n=0 (1 − z)2

and
∑∞ zE(z D(L) )E(z QU )
E(f0 (n − (D(L) + QU))z n = .
n=0 (1 − z)2
Combining now the last two equalities and using the definition of r(n) we obtain
∑∞ zE(z D(L) )(E(z Y1 ) − 1)E(z QU )
r(n)z n = (6.96)
n=0 (1 − z)2

It is easy to check that

1 − zQ
E(z QU ) =
Q(1 − z)
and using this it follows by relation (6.96) that
∑∞ −z exp(−λL(1 − PY (z)))(1 − PY (z))(1 − z Q )
r(n)z n =
n=0 Q(1 − z)3

showing the desired result.

This last result concludes our discussion of the (s, nQ) model and our discussion of the
single item inventory control models.

95
Bibliography

[1] A.W.Roberts and D.E.Varberg. Convex Functions. Academic Press, New York, 1973.

[2] C.T.H.Baker. The Numerical Treatment of Integral Equations. Monographs on Nu-

merical Analysis. Clarendon Press, Oxford, 1977.

[3] D.Erlenkotter. An early classic misplaced:ford w.harris’s economic order quantity

model of 1915. Management Science, 35(7):898–900, 1989.

[4] D.V.Widder. The Laplace Transform. Princeton Mathematical Series. Princeton Uni-
versity Press, 1972.

[5] E.A.Silver, D.F.Pyke, and R.Peterson. Inventory management and Production Plan-
ning and Scheduling. Wiley, New York, 1998.

[6] E.Bazsa, J.B.G.Frenk, and P.Den Iseger. Inventory control and regenerative pro-
cesses:theory. Technical Report 9926, Ribes research papers, Econometric Insti-
tute,EUR, 1999.
[7] E.Cinlar. Introduction to Stochastic Processes. Prentice-Hall, 1975.

[8] E.L.Porteus. Stochastic Models, chapter Stochastic inventory theory, pages 605–652.
Handbooks in Operations Research and Management Science. North Holland, Ams-
terdam, 1990. editors:D.P.Heyman,M.J.Sobel.
[9] E.M.Bazsa, J.B.G.Frenk, and P.W. den Iseger. Modeling of inventory control with
regenerative processes. International Journal of Production Economics, 71:263–276,
2001.
[10] E.Parzen. Stochastic Processes. Holden-Day, 1962.

[11] E.Seneta. Regularly Varying Functions, volume 508 of Lecture Notes in Mathematics.
Springer Verlag, Berlin, 1976.

[12] F.B.S.L.P.Janssen. Inventory Management Systems (Control and Information Issues).

PhD thesis, Tilburg university, Center for Economic Research, 1998.

[13] F.Stummel and K.Hainer. Introduction to Numerical Analysis. Scottish Academic

Press, 1980.

[14] F.W.Harris. How many parts to make at once. Factory (The magazine of manage-
ment), 10:135–136, 1913.

[15] G.Hadley and T.M.Whitin. Analysis of Inventory Systems. Prentice Hall, 1963.
[16] H.Brunner and P.J. Van der Houwen. The Numerical Solution of Volterra Integral
Equations, volume 3. CWI, Amsterdam, 1986.
[17] H.C.Tijms. Stochastic Models: An Algortihmic Approach. Wiley, New York, 1995.

96
 [18] I.Sahin. Regenerative Inventory Systems. Wiley, New York, 1989.

[19] P.Den Iseger and E.Bazsa. A new approach to solving inventory models. Technical
report, Econometric Institute, EUR, Rotterdam, 1999.

[20] J.Abate and W.Whitt. The fourier-series method for inverting laplace transforms.
Queueing Systems, 10:5–88, 1992.
[21] J.Abate and W.Whitt. Numerical inversion of probability generating functions. Opa-
rations Research Letters, 12:245–251, 1992.
[22] J.B.G.Frenk. Stochastische Optimalisering. Stichting Syllabi, EUR Rotterdam, 2000.

[23] J.B.G.Frenk, R.Dekker, and M. Kleijn. A unified approach of single item component
replacement models. Mathematical Methods of Operations Research, 45:437–454,
1997.
[24] J.B.Hiriart-Urruty and C.Lemarechal. Convex Analysis and Minimization algorithms
I. Springer Verlag, New York, 1993.
[25] J.Bramel and D.Simchi-Levi. The Logic of Logistics. Springer Series in Operations
Research. Springer-Verlag, New York, 1997.
[26] J.W.Cohen. On Regenerative Processes in Queueing Theory, volume 121 of Lecture
Notes in Economics and Mathematical Systems. Springer Verlag, Berlin, 1976.
[27] A.G.de Kok. Basics of inventory management. Technical report, Econometric Insti-
tute ,Tilburg University, 1992.
[28] M.J.D.Powell. Approximation Theory and Methods. Cambridge University Press,
Cambridge, 1981.
[29] M.S.Bazaraa, H.D.Sherali, and C.M.Shetty. Nonlinear Programming (Theory and
Applications). Wiley, New York, 1993.
[30] M.T.Melo. Stochastic Lotsizing in Production Planning. PhD thesis, Econometric
Institute, Erasmus university, 1996.

[31] P.Billingsley. Probability and Measure. Wiley, New York, 1979.

[32] P.H.Zipkin. Foundations of Inventory Management. Mc Graw Hill, Boston, 2000.

[33] P.Linz. Analytical and Numerical Methods for Volterra Equations. SIAM, Philadel-
phia, 1985.

[34] P.Protter. Stochastic Integration and Differential Equations. Springer Verlag, 1992.
[35] S.Asmussen. Applied Probability and Queues. Wiley, New York, 1987.

[36] S.Axsater. Inventory Control. International Series in Operations Research and Man-
agement Science. Kluwer, New York, 2000.

[37] S.I.Resnick. Regular Variation and Point Processes. Springer Verlag, New York,
1987.

[38] S.M.Ross. Introduction to Probabilty Models. Academic Press, 6 edition, 1997.

[39] W.Rudin. Principles of Mathematical Analysis. McGraw-Hill, New York, 1976.

97

View publication stats