PRENTICE-HALL INTERNATIONAL SERIES IN MANAGEMENT

Baumol Economic Theory and Operations Analysis

Churchman Prediction and Optimal Decision: Philosophical
Issues of a Science of Values
Hadley and Whitin Analysis of Inventory Systems
Herron, Rawdon, and Greenlaw Business Simulation
Holt, Muth, Modigliani, and Planning Production, Inventories, and the Work
Simon Force
Mass6 Optimal Investment Decisions: Pules for Action and
Criteria for Choice
Miller and Starr Executive Decisions and Operations Research
Pfiffner and Sherwood Administrative Organization

Vqm
PRENTICE-HALL QUANTITATIVE METHODS SERIES
Dr. W. Allen Spivey, Editor

Cyert and Davidson Statistical Sampling for Accounting Information

Hadley and Whitin Analysis of Inventory Systems

Kemeny, Schleifer, Snell, and Finite Mathematics with Business Applications

Thompson

Mass® Optimal Investment Decisions: Rules for Action and

Criteria for Choice

PRENTICE-HALL, INC.

PRENTICE-HALL INTERNATIONAL, INC., UNITED KINGDOM AND EIRE

PRENTICE-HALL OP CANADA, LTD., CANADA

J. H. DE BUSST, LTD., HOLLAND AND FLEMISH-SPEAKING BELGIUM

DTJNOD PRESS, FRANCE

MARUZEN COMPANY, LTD., FAR EAST

C. BERTELSMANN VERLAG, WEST GERMANY AND AUSTRIA

HERRERO HERMAN OS, SUCS., SPAIN AND LATIN AMERICA

ANALYSIS
OF INVENTORY SYSTEMS

G. Hadley
University of Chicago

T. M. Whitin
University of California, Berkeley

Prentice-Hall, Inc, Englewood Cliffs, N. J. 1963

 PRENTICE-HALL INTERNATIONAL, INC., London
PRENTICE-HALL OF AUSTRALIA, PTY., LTD., Sydney
PRENTICE-HALL OF CANADA, LTD., TOTOntO

PRENTICE-HALL FRANCE, S.A.R.L., Paris

PRENTICE-HALL OF JAPAN, INC., Tokyo

PRENTICE-HALL DE MEXICO, S.A., Mexico City

© 1963 by Prentice-Hall, Inc., Englewood Cliffs, N. J. All

rights reserved. No part of this book may be reproduced in
any form, by mimeograph or any other means, without
permission in writing from the publishers.

Library of Congress Catalog Card No. 62-7406

Printed in the United States of America

03295-C
 PREFACE

During the last fifteen years there has been a rapid growth of interest in
what is often referred to as scientific inventory control. Scientific inventory
control is generally understood to be the use of mathematical models to
obtain rules for operating inventory systems. The subject has attracted
such wide interest that today every serious student in the management
science or industrial engineering areas is expected to have had some experi¬
ence working with inventory models. Originally, the development of in¬
ventory models had practical application as an immediate objective. To
a large extent this is still true, but as the subject becomes older, better
developed, and more thoroughly explored, an increasing number of indi¬
viduals are working with inventory models because they present interesting
theoretical problems in mathematics. For such individuals, practical appli¬
cation is not a major objective, although there is the possibility that their
theoretical work may be helpful in practice at some future time. Thus,
today work is being done with inventory models at many different levels,
ranging from a concern only for practical problems to a concern only for
the abstract mathematical properties of the model.
The purpose of this text is to introduce the reader to the techniques of
constructing and analyzing mathematical models of inventory systems.
In doing so, it cuts across many of the levels of work being done with
inventory models. Thus, by reading the entire text it should be possible
for the reader to gain an understanding of the sort of work that is being
done and the kinds of problems that are encountered when studying inven¬
tory models at all levels—from the purely practical to the purely theoretical.
It was recognized that in many cases readers would be only interested
in one facet of the subject, such as practical applications. An effort has
been made to write the text in such a way that it can be read by a broad
group of readers with widely varying mathematical backgrounds. To ac¬
complish this, the material presented first in each chapter is of practical
§ interest, and involves the most elementary mathematics. The more de¬
tailed and exact developments, which are also more advanced mathe-
PC V
<
£ THE fillin' USMHs'
tmmi INSTITUTE OF TECHH0L06*
VI PREFACE

matically, come later. Topics of theoretical interest are treated in the

discussion of these more advanced models. An exception to this rule is
Chapter 8, which is devoted almost exclusively to topics that are more
advanced mathematically. Thus someone who was mainly interested in
practical applications could read Chapter 1, the first few sections of
Chapters 2 through 7, none of Chapter 8, and all of Chapter 9.
Because of its flexibility this text could be used for a variety of courses
including a full semester course in inventory theory, part of a course in
production, or part of a course in operations research. Both of the authors
have used essentially all of the material in this text in an inventory control
course. This course was one quarter in length at the University of Chicago
(requiring about fifteen hours per week of work) and one semester in length
at the University of California (requiring about nine hours per week of
work). One of the authors (G.H.) has also used the material for slightly
less than one half of a one quarter course in the production area. This
course covered all of Chapter 1, Chapter 2 through Sec. 2-9, Chapter 4
through Sec. 4-4, Chapter 5 through Sec. 5-2, Chapter 6 through Sec. 6-4,
and all of Chapter 9.
The material treated in this book is concerned almost exclusively with
the determination of optimal operating doctrines for systems consisting of
a single stocking point and a single source of supply. The reasons for doing
this are: 1) Many practical problems fall into this category. 2) Many
interesting mathematical problems arise even when attention is restricted
to these relatively simple systems. 3) It is extremely difficult to determine
optimal operating doctrines for more complex systems. Indeed, little has
been done in this area. Frequently, when one desires to examine a complex
multi-echelon system as a whole, he is, for a variety of reasons, far more
interested in the dynamic response and the stability of the system than in
determining an operating doctrine which will minimize some cost expres¬
sion for a specified stochastic input. The analysis of the dynamic response
and stability of a system by analytical methods and simulation techniques
will be treated in a separate volume, and hence these topics are not con¬
sidered in the present text either.
An effort has been made to provide a large number of original and inter¬
esting problems. The authors consider the problems to be very important,
and any serious reader should at least look them over and attempt to work
out a fair number.
The Graduate School of Business, University of Chicago, very generously
provided the secretarial services for having the manuscript typed. Jackson
E. Morris once again did an excellent job of providing the quotations which
appear at the beginning of the chapters. The authors are indebted to
Paul Teicholz and B. Lundh of Stanford University, who used their digital
computer programs to compute one of the examples for Chapter 4 and
 PREFACE VII

one for Chapter 5. Also the authors are indebted to a number of their
students who pointed up errors and misprints in the manuscript. Finally,
the authors express their appreciation to the editor of Operations Research
for permission to reprint a number of the Poisson properties which they
first published there.
G.H.
T.M.W.
 CONTENTS

-CHAPTER 1. The Nature of Inventory Systems 1

1-1. INVENTORY PROBLEMS 1
1-2. BRIEF HISTORICAL SKETCH 2
1-3. INVENTORY SYSTEMS 4
1-4. THE ECHELON STRUCTURE OF INVENTORY SYSTEMS 4
1-5. THE NATURE OF THE ITEMS 7
1-6. THE STOCHASTIC PROCESSES ASSOCIATED WITH THE
INVENTORY SYSTEM 8

1-7. THE RELEVANT COSTS 10

1-8. PROCUREMENT COSTS 1 1
1- 9. INVENTORY CARRYING COSTS 13
1-10. COSTS OF FILUNG CUSTOMERS’ ORDERS 17
1-11. STOCKOUT COSTS 18
1-12. COSTS OF OPERATING THE INFORMATION PROCESSING
SYSTEM 21
1- 13. SELECTION OF AN OPERATING DOCTRINE 21

•CHAPTER 2. Deterministic Lot Size Models and Their

Extensions 29

2- 1. INTRODUCTION 29
2-2. THE SIMPLEST LOT SIZE MODEL—NO STOCKOUTS 29
2-3. ADDITIONAL PROPERTIES OF THE MODEL; AN EXAMPLE 34
2-4. ACCOUNTING FOR INTEGRALITY OF DEMAND 40
2-5. CASE WHERE BACKORDERS ARE PERMITTED 42
2-6. THE LOST SALES CASE 47

2-7. THE CASE OF A FINITE PRODUCTION RATE 50

2-8. CONSTRAINTS 54

2-9. CONSTRAINTS—AN EXAMPLE 59

2- 10. PERIODIC REVIEW FORMULATION 52

2-11. QUANTITY DISCOUNTS—"ALL UNITS” DISCOUNTS 62
2-12. INCREMENTAL QUANTITY DISCOUNTS 66

vin
IX CONTENTS

CHAPTER 3- Probability Theory and Stochastic Processes 83

3-1. INTRODUCTION ' 83

3-2. BASIC LAWS OF PROBABILITIES 83
3-3. DISCRETE RANDOM VARIABLES 86
3-4. CONTINUOUS RANDOM VARIABLES 91
3-5. EXPECTED VALUES 98
3-6. TIME AVERAGES AND ENSEMBLE AVERAGES 104
3-7. PROBABILISTIC DESCRIPTION OF DEMANDS 106
3-8. JOINT DISTRIBUTIONS 11 4
3- 9. CONVOLUTIONS 1 18
3-10. MARKOV PROCESSES DISCRETE IN SPACE AND TIME 12/
3-11. MARKOV PROCESSES DISCRETE IN SPACE AND
CONTINUOUS IN TIME—QUEUING 129
3-12. OTHER TYPES OF MARKOV PROCESSES 135
3-13. PROPERTIES OF THE POISSON DISTRIBUTION 136
3-14. THE NORMAL DISTRIBUTION 140
3- 15. PROPERTIES OF THE NORMAL DISTRIBUTION 143

CHAPTER 4. Lot Size—Reorder Point Models with Stochastic

Demands 159

*4-1. INTRODUCTION 159

* 4-2. HEURISTIC APPROXIMATE TREATMENT OF THE
BACKORDERS CASE 162
> 4-3. HEURISTIC APPROXIMATE TREATMENT FOR THE
LOST SALES CASE 168
‘ 4-4. DISCUSSION OF THE SIMPLE MODELS AND A
NUMERICAL EXAMPLE 169
4- 5. THE EXACT DEVELOPMENT—INTRODUCTORY DISCUSSION 175
4-6. COMPUTATION OF TIME AVERAGES 177
4-7. EXACT FORMULAS FOR THE BACKORDERS CASE WITH
POISSON DEMANDS AND CONSTANT PROCUREMENT
LEAD TIME 181
4-8. AN IMPORTANT SPECIAL CASE 188
4-9. THE NORMAL APPROXIMATION 191
4- 10. AN EXAMPLE INVOLVING THE USE OF THE
EXACT FORM OF 3C 195
4-11. THE LOST SALES CASE FOR CONSTANT LEAD TIMES 197
4-12. STOCHASTIC LEAD TIMES 200
4-13. MODELS WITH Q = 1 204
4-14. QUANTITY DISCOUNTS 21 2
4-15. CONSTRAINTS 213
4-16. DETERMINATION OF OPERATING DOCTRINES WITHOUT
SPECIFYING STOCKOUT COSTS 217
 CONTENTS X

CHAPTER 5. Periodic Review Models with Stochastic Demands 235

5-1. INTRODUCTION 235

5-2. SIMPLE, APPROXIMATE < R, 7 > Models 237
5-3. THE EXACT FORMULATION OF THE < nQ, r, T > MODEL
FOR THE BACKORDERS CASE WITH POISSON DEMANDS
AND CONSTANT LEAD TIMES 245
5-4. APPROXIMATE FORM OF THE < nQ, r, 7 > MODEL FOR
LARGE Q 255
5-5. THE < nQ, r, 7 > MODEL FOR NORMALLY DISTRIBUTED
DEMANDS 256
5-6. EXACT EQUATIONS FOR < R, 7 > MODELS 260
5-7. THE < Q,r > MODEL AS THE LIMIT AS 7—► 0
OF THE < nQ, r, 7 > MODEL 263
5-8. MODELS OF THE < R,r,7> TYPE 265
5- 9. DERIVATION OF AVERAGE ANNUAL COST FOR
< R, r, 7 > MODEL 268
5-10. THE < R, r, 7 > MODEL WHEN DEMAND IS TREATED
AS A CONTINUOUS VARIABLE 275
5-11. COMPARISON OF THE PERIODIC REVIEW OPERATING
DOCTRINES 280
5-12. COMPARISON OF PERIODIC REVIEW AND TRANSACTIONS
REPORTING SYSTEMS 281
5-13. THE LOST SALES CASE 282
5-14. STOCHASTIC LEAD TIMES 285
5-15. QUANTITY DISCOUNTS 288

CHAPTER 6. Single Period Models 297

* 6-1. INTRODUCTION 297

* 6-2. THE GENERAL SINGLE PERIOD MODEL WITH TIME

INDEPENDENT COSTS 297
* 6-3. EXAMPLES 299
* 6-4. CONSTRAINED MULTIPLE ITEM PROBLEMS 304
6- 5. SINGLE PERIOD MODELS WITH TIME DEPENDENT COSTS 308
6- 6. MARGINAL ANALYSIS 312

CHAPTER 7. Dynamic Inventory Models 323

7- 1. INTRODUCTION 323
7-2. DYNAMIC PROGRAMMING 323
7-3. DYNAMIC PROGRAMMING FORMULATION OF OTHER
PROBLEMS 330
7-4. DYNAMIC-DETERMINISTIC LOT SIZE MODEL 336
xi CONTENTS

7-5. EXAMPLE OF THE DYNAMIC LOT SIZE MODEL 343

7-6. DYNAMIC MODELS WITH STOCHASTIC DEMANDS AND A
FIXED HORIZON 345
7-7. DYNAMIC MODELS WITH STOCHASTIC DEMANDS AND A
VARIABLE HORIZON 349
7- 8. THE PREDICTION PROBLEM 350

CHAPTER 8. Uses of Dynamic Programming in the Analysis of

Steady State Models 361

8- 1. INTRODUCTION 361
8-2. THE BASIC FUNCTIONAL EQUATION FOR PERIODIC
REVIEW SYSTEMS 361
8-3. OPTIMALITY OF Rr POLICIES FOR PERIODIC REVIEW
SYSTEMS—QUALITATIVE DISCUSSION 364
8-4. PROOF THAT AN Rr POLICY IS OPTIMAL WHEN f(y; T) IS
CONVEX AND THE UNIT COST IS CONSTANT 367
8-5. USE OF DYNAMIC PROGRAMMING TO COMPUTE OPTIMAL
VALUES OF R AND r 374
8-6. THE FUNCTIONAL EQUATION FOR TRANSACTIONS
REPORTING SYSTEMS 376
8-7. OPTIMALITY OF Rr POLICIES FOR TRANSACTIONS
REPORTING SYSTEMS 381
8-8. EXPLICIT SOLUTION OF THE FUNCTIONAL EQUATION
WHEN A POISSON PROCESS GENERATES DEMANDS
AND UNITS ARE DEMANDED ONE AT A TIME 382
8- 9. EXPLICIT SOLUTION OF THE FUNCTIONAL EQUATION IN
THE GENERAL CASE 388
8-10. EXPLICIT SOLUTION OF THE FUNCTIONAL EQUATION FOR
THE STEADY STATE PERIODIC REVIEW MODEL 391
8- 11. LOST SALES CASE 394

CHAPTER 9. Problems of Practical Application 401

9- 1. INTRODUCTION 401
9-2. RELEVANCE OF THE MODEL 401
9-3. DATA PROBLEMS 405
9-4. THE DEMAND DISTRIBUTION 406
9-5. DEMAND PREDICTION 413
9-6. THE LEAD TIME DISTRIBUTION 419
9-7. DETERMINATION OF COSTS 420
9-8. MULTI-ITEM PROBLEMS 423
9-9. PERSONNEL AND PROCEDURAL PROBLEMS 425
9- 10. THE EVALUATION PROBLEM 427
9-11. SUMMARY 429
THE NATURE

OF INVENTORY SYSTEMS

“How can it be that mathematics, being after all a product

of human thought independent of experience, is so admirably

adapted to the objects of reality?”

Albert Einstein
1-1 Inventory Problems

The control and maintenance of inventories of physical goods is a prob¬

lem common to all enterprises in any sector of a given economy. For
example, inventories must be maintained in agriculture, industry, retail
establishments, and the military. In the United States the total dollar
investment in inventories at any one time is immense. The sum runs to
more than 50 billion dollars for defense projects alone and more than 95
billion for private enterprise sectors of the economy. There are many
reasons why organizations should maintain inventories of goods. The
fundamental reason for doing so is that it is either physically impossible
or economically unsound to have goods arrive in a given system precisely
when demands for them occur. Without inventories customers would have
to wait until their orders were filled from a source or were manufactured.
In general, however, customers will not or cannot be allowed to wait for
long periods of time. For this reason alone the carrying of inventories
is necessary to almost all organizations that supply physical goods to
“customers”. There are, nonetheless, other reasons for holding inventories.
For example, the price of some raw material used by a manufacturer may
exhibit considerable seasonal fluctuation. When the price is low, it is profit-
able for him to procure a sufficient quantity of it to last through the high
priced season and to keep it in inventory to be used as needed in production.
Another reason for maintaining inventories, a reason particularly important
to retail establishments, is that sales and profits can be increased if one has
an inventory of goods to display to customers.
Two fundamental questions that must be answered in controlling the
inventory of any physical good are when to replenish the inventory and
how much to order for replenishment. In this book we shall attempt to
show how these questions can be answered under a variety of circum¬
stances. Essentially every decision which is made in controlling inventories
in any organization, regardless of how complicated the inventory supply
system may be, is in one way or another associated with the questions of
when to order and how much to order. There are certain types of inventory
problems, such as those concerned with the storage of water within dams,
in which one has no control over the replenishment of the inventory. (I|\
1
2 THE NATURE OF INVENTORY SYSTEMS CHAP. 1

other words, the resupply of the inventory of water within the dam depends
on the rainfall, and the organization operating the dam has no control over
this.) We shall not consider this type of problem here. The only problems
with which we shall concern ourselves are those in which the organization
controlling the inventory has some freedom in determining when, and in
what quantity, the inventory should be replaced. On the other hand, we
shall assume that, in general, the inventory system has no control over the
demands which occur for the item, or items, which it stocks. Again, this is
just the opposite of what one encounters in dealing with inventory problems
such as storage of water within dams, since the efflux of water through
the dam is completely within the control of the organization operating the
dam. In short, we are going to consider the type of inventory problem
encountered in business, industry, and the military.
We shall concentrate on showing how mathematical analysis can be used
to help develop operating rules for controlling inventory systems. When
mathematics is applied to the solution of inventory problems, it is necessary
to describe mathematically the system to be studied. Such a description is
often referred to as a mathematical model. The procedure is to construct
a mathematical model of the system of interest and then to study the
properties of the model. Because it is never possible to represent the real
world with complete accuracy, certain approximations and simplifications
must be made when constructing a mathematical model. There are many
reasons for this. One is that it is essentially impossible to find out what the
real world is really like. Another is that a very accurate model of the real
world can become impossibly difficult to work with mathematically. A
final reason is that accurate models often cannot be justified on economic
grounds. Simple approximate ones will yield results which are good enough
so that the additional improvement obtained from a better model is not
sufficient to justify its additional cost.
In this book we shall study a variety of mathematical models of in¬
ventory systems. Many of these are intended for practical application.
Others, however, have no immediate practical application because of the
restrictive nature of the assumptions. They are interesting and relevant,
however, because they exhibit some theoretical properties which are
important in understanding the nature of inventory systems.

1-2 Brief Historical Sketch

Although inventory problems are as old as history itself, it has only been
since the turn of the century that any attempts have been mdc to employ
analytical techniques in studying these problems. The initial impetus for
the use of mathematical methods in inventory analysis seems to have been
SEC 1-2 BRIEF HISTORICAL SKETCH 3

supplied by the simultaneous growth of the manufacturing industries and

the various branches of engineering—especially industrial engineering. The
real need for analysis was first recognized in industries that had a combi¬
nation of production scheduling problems and inventory problems, i.e.,
in situations in which items were produced in lots—the cost of set up
being fairly high—and then stored at a factory warehouse.
The earliest derivation of what is often called the simple lot size formula
was obtained by Ford Harris of the Westinghouse Corporation in 1915 [5].
This same formula has been developed, apparently independently, by many
individuals since then; it is often referred to as the Wilson formula since it
was also derived by R. H. Wilson as an integral part of the inventory
control scheme which he sold to many organizations. The first full length
book to deal with inventory problems was that of F. E. Raymond [6],
written while he was at M.I.T. It contains no theory or derivations, and
only attempts to explain how various extensions of the simple lot size model
can be used in practice.
It was not until after World War II, when the management sciences and
operations research emerged, that detailed attention was focussed on the
stochastic nature of inventory problems. Prior to that the systems had
been treated as if they were deterministic, except for a few isolated cases,
such as the work of Wilson, where some attempts were made to include
probabilistic considerations. During the war, a useful stochastic model
was developed which we shall refer to in Chapter 6 as the Christmas tree
model. Shortly thereafter, a stochastic version of the simple lot size model
was developed by Whitin, whose book [8], published in 1953, was the first
book in English which dealt in any detail with stochastic inventory models.
As has been noted above, the original interest in using analytical tech¬
niques to solve inventory problems arose in industry where engineers were
seeking solutions to practical problems. It is interesting to observe that
economists were not the first to take an active interest in inventory prob¬
lems even though inventories play a crucial role in the study of dynamic
economic behavior. The reason for this lack of interest probably lies in the
fact that economists were concentrating their attention mainly on static
equilibrium models. Recently, however, some economists and mathema¬
ticians have taken an interest in inventory models. They have not been
especially concerned with immediate practical applications; instead, they
have been interested in the models because of their mathematical properties
and economic interpretations. The paper by the economists Arrow, Harris,
and Marschak [1] was one of the first to provide a rigorous mathematical
analysis of a simple type of inventory model. It was followed by the often
quoted and rather abstract papers by the mathematicians Dvoretzky,
Kiefer, and Wolfowitz [3, 4]. Since then a number of papers by mathema¬
ticians have appeared. A recent full length book devoted to the mathe-
4 THE NATURE OF INVENTORY SYSTEMS CHAP. 1

matical properties of inventory systems is that of Arrow, Karlin, and

Scarf [2]. At the present time, work on inventory problems is being carried
on at many different levels. At one extreme a considerable amount of work
is concerned strictly with practical applications, while, at the other extreme,
work is being done on the abstract mathematical properties of inventory
models without regard to possible practical applications. The material
presented in this text will, in similar fashion, cover a fairly broad area.
Some material will be directly concerned with practical applications while
other material will be concerned with the mathematical structure of inven¬
tory systems. In this way, the reader will be introduced to the methods of
analysis used and the problems involved in carrying out investigations of
inventory systems at these various levels.

1-3 Inventory Systems

There are great differences between existing inventory systems. They

differ in size and complexity, in the types of items they carry, in the costs
associated with operating the system, in the nature of the stochastic
processes associated with the system, and in the nature of the information
available to decision makers at any given point in time. All these differ¬
ences can be considered to reflect variations in the structure of the in¬
ventory system. These variations can have an important bearing on the
type of operating doctrine that should be used in controlling the system.
By an operating doctrine we simply mean the rule which tells us when to
order and how much to order.
It is desirable to spell out in somewhat more detail the differences which
can exist in inventory systems—either real world systems or mathematical
models. In the following sections we shall make explicit some of these
differences.

1-4 The Echelon Structure of Inventory Systems

An item may be stocked in an inventory system at only a single physical

location, or it may be stocked at many locations. For example, if the organ¬
ization under consideration is the U.S. Air Force supply system, a spare
part for a certain type of aircraft may be stocked at over 100 bases and
repair facilities all over the world. If the organization under study is a
single privately owned lumber yard, the entire stocks of the organization
will be held at this lumber yard.
When there is more than a single stocking point, there exists the possi¬
bility for many forms of interaction between the stocking points. One of
SEC 1-4 THE ECHELON STRUCTURE OF INVENTORY SYSTEMS 5

the simplest forms of interaction involves one stocking point which serves
as a warehouse for one or more other stocking points. This leads to what
is referred to as a multiechelon inventory system. One possible type of
multiechelon system is illustrated in Fig. 1-1. The arrows indicate the
normal pattern for the flow of goods through the system. This might be
referred to as a four echelon system since there are four levels. Each level
is called an echelon. In the system shown, customer demands occur only
at the stocking points in level 1. These stocking points have their stocks

TTTTT 1
FIGURE 1-1.

replenished by shipments from warehouses at level 2, which in turn receive

replenishments for their stock from level 3, etc. Figure 1-1 represents only
one type of multiechelon system. In other cases, customer demands might
occur at all levels, or stocking points at any level might not only receive
shipments from the next highest level but might also get replenishments
from any higher level or from the source. Also, it might be allowable, on
occasion, to permit redistribution of stocks among various stocking points
at a given level.
Most inventory systems encountered in the real world are multiechelon
in nature. However, it is often true that one need not or cannot consider
6 THE NATURE OF INVENTORY SYSTEMS CHAP. 1

the multiechelon system in its entirety. The reason for this is that different
organizations operate different parts of the system. For example, Fig. 1-1
might refer to a production distribution system in which the source is a
plant where the item is manufactured, level 4 is a factory warehouse, level
3 represents regional warehouses, level 2 represents warehouses in various
cities, and level 1 represents the retail establishments which sell the item to
the public. In such a system the manufacturer might control only the
plant and the factory warehouse, while different organizations operate the
regional warehouses and still different organizations operate the city ware¬
houses and the retail establishments. Note that even at a given level many
different organizations may be involved. For example, each of the ware¬
houses in different cities may be under different ownership. In such a
system, each organization has the freedom to choose the operating doctrine
for controlling the inventories under its jurisdiction. One could not, in
general, attempt to analyze the system as a whole and dictate what oper¬
ating doctrine should be used by each stocking point at each level. Instead,
one might be concerned with the best way for one of the warehouses at
level 2 to control its inventories. In making the analysis, the customers
would be the retailers at level 1 and the source from which replenishments
are obtained would be the appropriate warehouse at level 3.
Frequently, there will be just a single source from which an inventory
system replenishes its stocks when it is desirable to do so. This source may
be the plant where the item is made, a factory warehouse, or simply a ware¬
house at a higher echelon. Sometimes, however, the system has two or
more alternative sources of supply available. For example, one of the
retailers in Fig. 1-1 might be able to order from several different warehouses
at level 2. A special case which can occur is that where the system under
study also controls the source of its supplies, i.e., the plant where the item
or items are made. In this case the problem is not strictly an inventory
control problem but also involves production scheduling. In this book we
shall not consider the general problem of combined production scheduling
and inventory control except in some cases where production is carried out
in lots rather than being continuous.
The basic inventory system that will be studied in this text will be much
simpler than the general sort of multiechelon system shown in Fig. 1-1. It
will consist of just one stocking point with a single source for resupply.
Customer demands arrive at the single stocking point, and at appropriate
times orders are placed with the source for replenishing the inventory. The
operation of this system is illustrated schematically in Fig. 1-2.
There are good reasons for restricting our attention to the structure illus¬
trated in Fig. 1-2. Perhaps the most important reason is that it is very
difficult to study analytically multiechelon systems of the type shown in
Fig. 1-1. In fact, very little work has been done in this area. It will be seen
SEC. 1-5 THE NATURE OF THE ITEMS 7

Customer Orders
demands Single Single
stocking
source
Goods to point Goods
customers

FIGURE 1-2.

that even the relatively simple structure that will be studied can become
very complex to analyze. The other reason is that for practical applications
the simple structure of Fig. 1-2 is often (although by no means always)
adequate. This is true because, as was noted above, even though real
world systems are usually of the multiechelon variety, it is often necessary
to consider the various stocking points individually because different
organizations control them. Even when a single organization controls a
number of stocking points, however, the interactions between them are
frequently sufficiently small to allow each to be studied independently of
the others.

1-5 The Nature of the Items

A large military supply system stocks over 500,000 different items, while
a typical department store may carry as many as 150,000 items. Other
inventory systems stock only one or two items. The items stocked can
differ from each other in many ways, They differ in cost, and in their
physical properties, such as weight and volume. Some items are perishable
and cannot be stored for long periods of time; others can be stored in¬
definitely without deterioration; others are subject to rapid obsolescence.
Often, items can be stored only under specially controlled conditions of
temperature, humidity, etc., and require special types of packaging for
storage. When more than a single item is stocked there can be interactions
between the items. For example, the items may be substitutes for each
other so that if the system is out of stock on one item a customer will accept
another. On the other hand, they may be complements so that usually one
will not be sold without the other. Frequently, the interactions will take
the form of having items compete for limited warehouse floor space or for
investment dollars, which are also limited.
Another different form of interaction can exist between the items carried
by an inventory system. This type of interaction is represented in what
might be called a multistage inventory system associated with a production
process. A typical block diagram for such a system is illustrated in Fig. 1-3.
8 THE NATURE OF INVENTORY SYSTEMS CHAP. 1

A characteristic of such a system is that the product or products being

manufactured can be inventoried at various stages of completion. (For
example, as raw materials, rough castings, finished castings, partially as¬
sembled units, etc.). The problem is to determine what inventories, if any,
should be maintained at the various stages, and what the operating doctrine
should be for controlling the stocks at all the stages. Very little work has
been done on problems of this sort, and they will not be considered in this
text, except for a few very simple situations which are treated in the
problems.
In this book the main characteristic of an item that will be of interest to
us is its cost. For some of the models developed, we shall assume that the
item can be stored indefinitely; in others, the item is assumed to become

FIGURE 1-3.

obsolete ultimately. Only one of the models will deal specifically with
perishable items that can be stored only for short periods of time. We shall
never need to concern ourselves directly with any of the special, controlled
conditions required for storage. Only a limited amount can be done in
treating interactions between items. Simple cases, where items are com-
peting for warehouse floor space or investment dollars, can be examined.
In a case where there is competition for floor space we shall also make use
of the unit volume of the items.

1-6 The Stochastic Processes Associated with the Inventory System

Generally speaking, it is almost never true that enough is known about the
process which generates demands for items carried by an inventory system
 SIC l-d STOCHASTIC PROCESSES ASSOCIATED WITH THE INVENTORY SYSTEM 9

to h« tilth* to prt'diot with wrtamty thu time pattern of demands. In

general, the Ixwt that ean la* dtme is to describe the demand in probabilistic
t^nns. We shall study ways in which this can be done. The process for
describing the demand pattern in probabilistic terms will he part of the
mathematical model of the system. When the demand pattern is described
probabilistically we shall say that a stochastic process generates the de«
mand pattern. In certain caws, the demand pattern will show enough
regularity so that, for an adequate approximation, it can Ik1 treated as
deterministic. When this is possible, the task of analyzing the model
mathematically is considerably simplified. Generally, however, it will not
Ik* possible to make such an approximation.
In the real world, the stochastic process associated with the demand pat¬
tern will always lie changing wit h time. However, in some cases, the change
with time may Iks so slow that in the mathematical model one can assume
that it d<K*s not change with time. In other eases, the change is sufficiently
rapid that a meaningful model must explicitly account, for it. We shah
consider both types of models those in which the stochast ic process re-
mimat constant over time and those in which it changes with time*
In general, if the inventory system under consideration has a number of
stocking points at which demands aim occur, it may or may not ba true that
tha demands at tha various stocking points ara independent of aaoh other*
It thay ara not independent tin* analysts becomes extremely complex*
Hinoi^ wa ara limiting our study to systems having just, a single stocking
pointt wa ara implicitly assuming that if the actual system contains morn
than a single stocking point, than tha demands at the different locations
ara independent. (Hherwise, it would not ba possible to study each stocking
point by itself independently of the others.
It is seldom economieally sound for a system to carry enough inventory
so that there will always be stock on hand whim a demand occurs* Because
of the stochastic nature of the demand pattern, there can he times whim
demands oceur and the system is out of stock. An important characteristic
of the process generating demands is what happens when a demand occurs
and the system is out of stock. Basically, there are two possibilities*
hither the demand is lost (as it might be in a department store whim the
customer goes to another storeb or it is backordered and the customer waits
until the inventory system obtains sufficient stock to meet his demand,
!?mhe text * we shall examine the two eases where fa) all requisitions oc¬
curring wh*m the system is out of stock are backordered, and (b) all
requisitions occurring when the system in out of stock are lost, These will
be referred to ns tin* backorders case and the lost sales case, respectively*
tu the backorders ease, all orders are ultimately filled, We shall aim that it
is much easier to treat the backorders case than it k the lost sales cam*.
In tact, results tor the lost sales case can be obtained only under very
10 THE NATURE OF INVENTORY SYSTEMS CHAP. 1

restrictive assumptions. The backorders case and lost sales case represent
fundamentally different stochastic processes.
For any particular inventory system it can happen that some demands
occurring when the system is out of stock are backordered while others are
lost. If it is possible to develop models of the system for the backorders and
the lost sales cases respectively, it is not difficult to develop a single model
in which some sales are lost and some are backordered.
We shall define the procurement lead time (or simply the lead time) for
an inventory system as the interval between the time when the stocking
point decides that an order for a replenishment should be made and the
time that the order arrives and is on the shelves, available to customers.
Often the procurement lead time will not be constant, since the time to fill
the order at the source, the shipping time, and the time required to carry
out the paper work, etc., can vary from one order to another. It is seldom
possible to predict in advance precisely what the lead time will be. Some¬
times the variations in the lead time will be small enough so that in the
mathematical model the lead tune can be assumed to be absolutely con¬
stant. In other situations, however, it will exhibit sufficient unpredictable
variability that it is necessary to assume that a stochastic process generates
the lead times. We shall consider both models in which the lead times are
constant and models in which lead times are stochastic variables.

1-7 The Relevant Costs

The costs incurred in operating an inventory system play a major role

in determining what the operating doctrine should be. The costs which
influence the operating doctrine are clearly only those costs which vary as
the operating doctrine is changed) Costs that are independent of the oper¬
ating doctrine used need not be included in any analysis where costs are
used as an aid in determining an operating doctrine. Fundamentally, there
are five types of costs which may be important in determining what the
operating doctrine should be. These are: 1) The costs associated with pro¬
curing the units stocked, 2) The costs of carrying the items in inventory,
3) The costs of filling customers' orders, 4) The costs associated with
demands occurring when the system is out of stock, 5) The cost of operating
the data gathering and control procedures for the inventory system.
In the following sections, we shall examine in more detail each of these
five types of costs. When doing so, it will quickly become apparent that it
is quite difficult to represent mathematically all the cost components with
complete accuracy. Consequently, for reasons outlined earlier, it is de¬
sirable to make some approximations when representing these costs in the
 SEC 1-8 PROCUREMENT COSTS 11

mathematical models to be developed. The nature of these approximations

will also be considered when discussing the costs.

1-8 Procurement Costs

We shall begin by examining the procurement costs. At the outset these

costs can be divided into two parts. First there is the amount which must
be paid to the source from which the procurement is made. The sum paid
to this source simply represents the cost of the units procured. Then there
are the costs incurred by the inventory system itself in making a procure-
ment. These costs can arise from many different factors and they can differ
considerably in nature from one inventory system to another. For example,
there are the costs of processing an order through the purchasing and
accounting departments. These include paper and postage costs, labor
costs, perhaps the cost of a telephone call to the source, or the cost of
computer time needed to make any necessary computations or to update
accounting records. Sometimes the cost of transporting the order from the
source of supply to the stocking point will be paid by the source and hence
will be included in the cost of the units. In other cases the inventory
system will pay the transportation costs. These transportation costs will,
of course, depend on the mode of transportation used. Also, there are
usually receiving costs incurred when the stock arrives at the warehouse.
It may be necessary to uncrate the goods, perform an inspection of them,
or perhaps even carry out detailed testing. Furthermore, additional ac¬
counting and control records must be prepared.
The costs incurred by the inventory system itself in placing an order,
which were outlined above, can be divided into two classes—those which
depend on the quantity ordered and those which are independent of the
quantity ordered. Transportation costs, part of the receiving costs, and
part of the inspection costs will depend on the quantity procured. We shall
find it convenient to include these costs in with the cost of the units them¬
selves. If Q units are procured, we shall denote those costs of the procure-
ment such as the cost of the units, transportation costs, etc., which depend
on Q by C(Q). The average unit cost when Q units are procured will be
C(Q)/Q. A case that will be of particular interest is that where the unit
cost is a constant (7, independent of the quantity ordered. > Then the cost of
the Q units will be simply CQ. It need not be true, of course, that thjg
simple linear dependence on Q will always be valid even though frequently
it is a satisfactory approximation. For example, it will not be correct if
quantity discounts are available either on the cost of the units themselves
or on transportation costs.
12 THE NATURE OF INVENTORY SYSTEMS CHAP. 1

Consider next the costs which are independent of the quantity ordered.
These costs include paper, postage, telephone charges etc., as well as the
labor costs incurred in processing the order. They also include those parts
of receiving and inspection costs which are independent of the order size.
If the inventory system controls the plant where the item under consider¬
ation is made, then assuming that the item is made in lots, the set up costs
for a production run will fall into this category. These costs which do not
depend on the quantity ordered are incurred each time an order is placed.
They will be referred to as the “fixed” costs of placing an order. We shall
usually denote the fixed cost of placing an order by A. The total cost of
placing an order for Q units will then be A + C(Q).

Average number of orders per unit time

FIGURE 1-4.

It is important to note that the fixed ordering costs have the property
that, usually, to a good approximation, the total fixed costs incurred on
placing N orders is simply AN. In other words, the total fixed costs in¬
curred are proportional to the number of orders placed. This proportion¬
ality of total fixed costs to the number of orders placed need not be exactly
true because of phenomena like the following: The number of people in the
office staff, and hence the annual cost of the office staff as a function of the
average number of orders placed per unit time might look something like
that shown in Fig. 1-4. As more and more orders are placed, the size of
the office staff must be increased. Each step in Fig. 1-4 might indicate an
increase in the office staff by one person. However, for a given size of office
staff, there is a range of ordering rates which requires this size of staff and
hence a range over which the annual cost of the office staff does not change.
 SEC 1-9 INVENTORY CARRYING COSTS 13

This is indicated by the flat portion of Fig. 1-4. Often, though, in the range
of interest, the curve shown in Fig. 1-4 can be approximated well enough
by a straight line through the origin so that a linear dependence of annual
cost on the average ordering rate is obtained.
We shall assume in the mathematical models developed in this book that
the fixed cost of placing N orders is AN, where A is the fixed cost of placing
a single order, so that the total fixed costs incurred are directly proportional
to the number of orders placed. As seen above, this will not in general be
Precisely true in any real world situation. Usually, it is a quite satisfactory
approximation. However, more importantly, models which make this
assumption can be used to handle situations in which there exist step
functions in the costs such as those shown in Fig. 1-4 which must be
explicitly accounted for. The technique for doing this will be presented in
Chapter 9.

1-9 Inventory Carrying Costs

Let us now examine the costs of holding inventory. Included in these

costs are the real out of pocket costs such as costs of insurance, taxes,
breakage and pilferage at the storage site, warehouse rental if the warehouse
is not owned by the inventory system, and the costs of operating the ware¬
house such as light, heat, night watchmen, etc. A cost which is frequently
the most important cost is not a direct out of pocket cost but rather an
opportunity cost which would never appear on an accounting statements
This is the cost incurred by having capital tied up in inventory rather than
having it invested elsewhere, and it is equal to the largest rate of return
which the system could obtain from alternative investments. By having
funds invested in inventory, one forgoes this rate of return, and hence it
represents a cost of carrying inventory.
The rate at which opportunity costs are incurred will, at any instant of
time, be proportional to the total investment in inventory. Similarly, the
rate at which breakage and pilferage costs are incurred will normally be
roughly proportional to the total investment in inventory at any point in
time. The rate at which insurance costs are incurred will not usually be
strictly proportional to the investment in inventory. Instead, a certain
amount of insurance will be carried and so long as the policy*is in effect the
amount of insurance will be constant, independent of fluctuations in the
inventory level. However, the policy will be revised periodically and hence
insurance costs can be made to vary to a certain extent with changes in the
inventory level. The precise way in which insurance costs will change with
the inventory level can vary greatly from one system to another.
The nature of tax assessments is not entirely uniform, although often

re mn L5^y/ ■ -

8«£tiE imimi iff

14 THE NATURE OF INVENTORY SYSTEMS CHAP. 1

they are levied on the on hand inventory in the warehouse on a specified

day in the year. In such cases, taxes are not incurred continuously through
the year but instead are incurred at a discrete point in time. If the inven¬
tory is low on the critical day the year’s taxes will also be relatively low,
and conversely, if inventory is high the year’s taxes will be relatively high.
When warehouse space is rented, it will normally be contracted for, and
the contract will be in force for a specified length of time. The amount of
space rented will then be based on the maximum amount needed for the
period of the contract. Thus the rate of incurring warehouse rental charges
will not fluctuate day to day with changes in the inventory level, although
the rental rates can be varied from month to month or year to year when
a new contract is negotiated. The costs of operating the warehouse may be
essentially independent of the inventory level, or part of these costs will
be independent of the inventory level, while the remainder will fluctuate
more or less proportionately with the inventory level.
The above discussion shows that not all the costs of carrying inventory
vary with the inventory level in the same way. Indeed, it is very difficult
to represent all of these costs with great accuracy in any mathematical
model. It is usually necessary to introduce some simplifying approxi¬
mations. In this text we shall assume that the instantaneous rate at which
inventory carrying costs are incurred are proportional to the investment in
inventory at that point in time. We shall use the symbol I to denote the
constant of proportionality; I will be caUecfthe carrying charge, and the .
physical dimensions^ J are cost per unit timeper monetary unit invested
in inventory. The particular dimensions that we shall use in measuring I
are dollars per year per dollar of inventory investment, and in these
J dimensions, it is usually true that 0 < I < 1. The instantaneous rate of
incurring carrying charges in the units of dollars per year can then be
written ICx where C is the unit cost of the item in dollars (or sometimes
the average unit cost if the unit cost is not constant) and x is the on hand
'inventory level.
From the above it is clear that we are assuming that the cost of carrying
a unit in inventory is directly proportional to the length of time for which
the unit remains in inventory. Suppose that the inventory level of the
system behaved as shown in Fig. 1-5 for a period of one year. Then the
carrying costs for the year are

IC £ x(t) dt

Note that the integral is the area under the curve between t = 0 and t = 1
year. Since the interval of integration is unity it is also true that the
integral is simply the average inventory (averaged over the period of one
year). Thus it follows that our definition of the way carrying costs are
SEC 1-9 INVENTORY CARRYING COSTS 15

incurred implies that the annual cost of carrying inventory is proportional

to the average inventory computed for that year. With this interpretation
the inventory carrying charge I is the fraction of the average investment
in inventory for a year which is incurred as carrying charges for the year.
The method chosen to compute the inventory carrying costs does not
represent with absolute accuracy any of the costs discussed above. It was
chosen because it does represent quite well the opportunity cost of carrying
inventory, which is normally the most important of all the costs. The rea¬
son it does not exactly represent the opportunity cost is because the largest
rate of return which can be obtained from other investments depends on
the length of time for which the investment is made. Other lesser problems

concern the fact that the maximum rate of return will also depend on the
sum that is available for investment and the general state of business. Thus
the opportunity cost represented in I can represent only an average rate of
return.
As has been noted previously, the rate of incurring breakage and pilferage
costs is roughly proportional to the investment in inventory, and hence
unless one has information to suggest otherwise, the method of costing we
are using seems reasonable for representing these costs. On the other hand,
our previous discussion has shown that the instantaneous rate of incurring
insurance costs will not in general fluctuate with movements in the inven¬
tory level. However, we have shown above that making the instantaneous
16 THE NATURE OF INVENTORY SYSTEMS CHAP. 1

rate of incurring inventory carrying charges proportional to the inventory

level yields an annual carrying cost which is proportional to the average
inventory. Now in many situations, especially those in which the average
demand rate remains relatively constant, it will be true that the annual
insurance costs will be essentially proportional to the average inventory,
since the insurance policy will be drawn up using the expected average
inventory as a basis for the value of the policy. Hence, in such cases, the
method to be used for incurring inventory carrying charges can be made to
represent the insurance costs quite accurately. In other situations, how¬
ever, such as those where the average rate of demand is continually chang¬
ing significantly with time, it no longer follows that annual insurance costs
can be made proportional to the average inventory (since it will be ex¬
ceptionally difficult to predict ahead of time what the average inventory
will be). For these cases, the procedure to be used for computing carrying
costs will only approximately represent the way in which insurance costs
are incurred.
The representation of tax costs by the method of computing carrying
charges which we are going to use is even more crude than the represen¬
tation of insurance costs, since if taxes depend only on the inventory in the
warehouse on a particular day in the year, they need have no relation what¬
ever to the average inventory for the year. On the other hand, the method
of computing carrying costs to be used here does often represent the way
in which the system sets aside funds for taxes. A tax fund will be set up,
and each unit contributes to the tax fund an amount proportional to the
length of time for which it remains in inventory. If the inventory system
uses this sort of procedure for accumulating money for payment of taxes,
then the costing procedure to be used here is the appropriate one to use.
In other situations where this is not done, the attempt to include taxes in
the carrying charge I can be considered to be only a rough approximation.
In general, it seems realistic to attempt to include a provision for taxes in
the carrying charge only when they are of such a magnitude that no special
effort is made to reduce inventories to a low level on the day when the tax
assessment is incurred. If special efforts are made to reduce inventories to
a low level at the date when taxes are incurred, then it is appropriate to
use a mathematical model which accounts for the fact that taxes are in¬
curred at a given point in time and not to try to include a provision for them
in the carrying charge. We shall later present models which could be used
for this purpose.
There is one other cost not mentioned above, for which an attempt is
often made to include it in the carrying charge. This is the cost of ob¬
solescence. The cost for each unit which must be disposed of at a loss be¬
cause it becomes obsolete is, at the time of obsolescence, the difference
between the original cost of the unit (plus any return which could have been
SEC 1-10 COSTS OF FILLING CUSTOMERS’ ORDERS 17

earned from the purchase date to the obsolescence date if the funds devoted
to the procurement of the item had been invested elsewhere) and its salvage
value. Obsolescence costs are always incurred at a fixed point in time, and
this obsolescence date often cannot be predicted with certainty in advance.
It is much more unrealistic to attempt to include obsolescence costs in the
carrying charge than it is to try to make a provision for taxes in the carry-
ing charge. The usual argument given is that by making a charge against
each item in proportion to the length of time it has remained in inventory,
one is setting up a fund to insure against obsolescence. However, such a
procedure can seriously distort the situation and lead one to using an
operating doctrine which may be far from being a good one. When obso¬
lescence is an important consideration, one should use a mathematical
model that explicitly takes into account the fact that obsolescence costs
are incurred only at a single point in time. We shall consider such models
in Chapter 7.
The inventory carrying charge I can then be thought of as the sum of
several terms and can be written

I = Ii + J2 + Iz + . . . (1-1)

where Ii might be the carrying charge arising from opportunity costs, I2

the carrying charge arising from pilferage and breakage, J3 the carrying
charge due to insurance costs, I4 the carrying charge due to taxes (if taxes
are included in carrying charge), etc. Thus I is obtained by evaluating
each of the components 7/ of the carrying charge and then summing the
results. Some of the details of making such computations are illustrated
in Chapter 9. It should be emphasized that in any particular application,
it is not necessarily true that all the cost components referred to above will
be relevant. On the other hand, there may be some costs that were not
referred to above.
Certain costs, such as floor space rental, will be essentially proportional
to the maximum inventory level. Although we do not in the text include
any inventory carrying charge terms which are proportional to the maxi¬
mum inventory level, it is very easy to include such terms in the mathe¬
matical models and some of the problems ask the reader to show how this
can be done.

1-10 Costs of Filling Customers’ Orders

In order to fill each customer’s order, a requisition must frequently be

processed through some sort of accounting operations where, among other
things, a shipping invoice is prepared and sent to the warehouse. In the
 THE NATURE OF INVENTORY SYSTEMS CHAP. 1
18

warehouse, someone must go to the proper bin and obtain the unit or units.
Next, it may be necessary to package the order for shipment. Finally, the
order is shipped to the customer. After the order has been shipped, a
record of this transaction is usually sent from the warehouse to accounting
where appropriate additional records are made. If a customer’s order ar¬
rives when the system is out of stock, it will usually be necessary to go
through special procedures to inform the customer of the existing situation.
The costs of the accounting operations referred to above, the salaries of
those in the warehouse who are concerned with filling orders, the costs of
packing, and the shipping costs, if paid by the inventory system, are all part
of the normal costs of filling customers’ orders. The important thing to note
about all these costs is that, while they will vary with the demand rate, they
will in general not depend on the operating doctrine used to control the
inventory system. Hence they need not be considered when studying costs
that vary as the operating doctrine is changed. On the other hand, the
costs arising from the special action required if a customer’s demand arrives
when the system is out of stock will depend on the operating doctrine, since
the fraction of the time that the system is out of stock will depend on the
operating doctrine. We shall find it convenient to include these latter costs
as a part of the stockout costs. With this convention, we shall not need
to consider in the future the costs of filling customers’ orders.

1-11 Stockout Costs

Let us now turn our attention to the costs incurred by having demands
occur when the system is out of stock. In doing so, we must distinguish
between the backorders and lost sales cases. Consider first the case where
all demands occurring when the system is out of stock are backordered.
In any practical situation, it is very difficult to determine accurately the
nature of the backorder costs. Backorder costs are inherently extremely
difficult to measure since they can include such factors as loss of customers’
goodwill (i.e., in the future, he may take his business elsewhere), or in
military supply systems, the cost of having part of some first line weapon
system inoperative because of lack of parts. Other parts of the backorder
cost can be somewhat easier to measure; however, these are usually a small
part of the total backorder cost. Such costs include the cost of notifying a
customer that an item is not in stock and will be backordered plus the cost
of attempting to find out when the customer’s order can be filled and giving
him this information. If the system itself uses the part, the backorder cost
may simply be the cost of keeping a machine idle for lack of parts. In such
a case one can obtain a relatively good measure of the backorder cost.
When units are demanded one at a time, then there will in general be a
 SEC 1-11 STOCKOUT COSTS 19

backorder cost associated with each unit backordered. When more than a
single unit can be demanded when a demand occurs, the backorder cost
could depend in a complicated way on the size of the demand. However,
we shall in this text assume that there is a backorder cost associated with
each unit backordered, even when units need not be demanded one at a
time. This cost will in general depend on the length of time for which the
unit remains backordered. It will be assumed here that this cost depends
only on the length of time for which the backorder exists. In principle,
the cost of each unit backordered as a function of the time t for which
the backorder remained on the books might be described by a nondecreas¬
ing function x(f), and this function might change from system to system.
In the real world it is essentially impossible to determine the precise shape
of 7r(f). About the most general function that can be used in practice is
ir(t) = x + fit. In other words, there is a fixed cost for each unit.back-
ordered plus a variable cost which is linear in the length of time for which
the backorder remains on the books, tin the mathematical models developed
m this text, when it is necessary to specify a backorder cost, the most
general function used will be x + ft. When r(t) involves t in a more compli-
cated way than x + M, it becomes considerably more difficult to work with
the mathematical models. Since these more complicated possibilities are
not useful in practice we shall not treat them in detail. However, some of
the problems ask the reader to formulate models involving more compli¬
cated functions for x(f).
If n units are backordered in a one year period, then the cost incurred in
that year attributable to the fixed cost x per unit backordered will be im.
If the number of backorders b(t) on the books as a function of time in the
year under consideration looks like that shown in Fig. 1-6, then the cost
incurred in that year attributable to the variable cost x£ per unit, back¬
ordered will be

x[fl + 4 + 4 •+•••+ 4]

since the height of each rectangle is unity, Now note that if t is measured
in years, then

will have the dimensions of units times years, i.e., unit years. Then if we
want the cost for the year to come out in dollars, x must have the di¬
mensions of dollars per unit year of shortage. Note that since the interval
of integration is unity,

I b(t) dt
Jo
is simply the average number of backorders which existed for the year, i.e.,
20 THE NATURE OF INVENTORY SYSTEMS CHAP. 1

FIGURE 1-6.

the number of unit years of shortage incurred for the year under consider¬
ation is numerically equal to the average number of backorders which
existed for the year.
Let us next consider the lost sales case. Since demands are lost if they
occur when the system is out of stock, the cost of a lost sale cannot depend
on time, for a lost sale is lost and there is nothing which corresponds to
the length of time for which a unit remains backordered. Thus when units
are demanded one at a time the cost of each lost sale will simply be a con¬
stant. If units need not be demanded one at a time, then the cost of a lost
sale could depend in a complicated way on the size of the demand. In the
text we shall assume that when there are costs incurred from lost sales,
then there is a fixed cost associated with each unit demanded which cannot
be met from inventory, even if units need not be demanded one at a time.
The cost of a lost sale includes a number of different factors. As in the
backorders case, most of these are not direct out of pocket costs that would
ever appear on a balance sheet. Perhaps the most important component
of the cost of a lost sale is the somewhat intangible goodwill loss. This can
include lost profits on sales of other items or on future sales of the given
item due to the fact that the customer temporarily or permanently takes
his business elsewhere or because he discourages other potential customers
by telling them that he received unsatisfactory service. The cost of a lost
sale also includes the costs of any special procedures used to inform the
customer that his demand cannot be supplied, and the profit lost in not
making the sale. We shall denote by 7r0 the cost, exclusive of the lost
 SEC M3 SELECTION OF AN OPERATING DOCTRINE 21

profit on the unit, of a unit having been demanded when the system is out
o stock. Then x will be used to denote the cost, including the lost profit
on the unit, of having a unit demanded when the system is out of stock.
Thus x is the sum of x0 and the unit profit on the item.
Sometimes we shall refer to the cost of having a unit backordered as the
cost of a backorder. Similarly, the cost of having a sale lost because a unit
is demanded when the system is out of stock will be referred to as the cost
of a lost sale. When we don’t care to differentiate between the costs of
backorders or of lost sales, we shall simply refer to stockout costs.

1-12 Costs of Operating the information Processing System

In order to use any given operating doctrine, an inventory system must

gather the information required for its use. The cost of obtaining the neces¬
sary information for decision making will clearly depend on the type of
operating doctrine used. These costs may include such things as the cost
associated with having a computer continuously update the inventory rec¬
ords, or the cost of making an actual inventory count, or the cost of
making demand predictions. The relevance of such costs to the mathe¬
matical models to be developed will be treated in the discussion of the
models.

1-13 Selection of an Operating Doctrine

The purpose in constructing a mathematical model of an inventory system

is to use it as an aid in developing a suitable operating doctrine for the
system. Usually one attempts to arrive at an operating doctrine that
will make profits as large as possible or costs as small as possible. In other
words, the criterion for selecting the operating doctrine is that of profit
maximization or cost minimization. In some cases the task of determining
the optimal operating doctrine is so difficult that it is either impossible or
uneconomical to determine the optimal doctrine, and instead, one opti-
mizes with respect to some subset of operating doctrines. In still other
cases, the mathematical model may be so complicated that it is extremely
difficult to do anything analytically. In such situations, one can use a
computer and simulation to study various operating doctrines. In general,
it is not possible to determine an optimal operating doctrine by use of
simulation, and it can be exceptionally difficult to find the optimum with
respect to some subclass of operating doctrines. About the best that can
be done using simulation is to study a small number of operating doctrines
and to select the one which seems to be the best. In this text we shall only
22 THE NATURE OF INVENTORY SYSTEMS CHAP. 1

be concerned with analytical techniques for working with mathematical

models, and hence we shall not deal with simulation as a tool for deter¬
mining operating doctrines.
In the above, we loosely referred to determining operating doctrines by
maximizing profits or minimizing costs. We have not yet specified precisely
what profits or costs we are referring to. We must indicate the various
cost components to be included, the time period over which they are
computed, and the technique to be used for handling the stochastic ele¬
ments in the problem. Of course, the profit will always be the revenue
minus the costs. However, the profit or cost need not be the same as that
which would be obtained if we were computing the strict accounting profit
or costjsmce for purposes of computing optimal operating doctrines it is
only necessary to include those costs which vary with the operating doctrine
(the variable costsJTj Costs which are independent of the operating doctrine
need not be included. The costs which must be included are those listed
in Sec. 1-7, which are the procurement, carrying, and stockout costs, and
the cost of operating the information processing system (recall that the
cost of processing customers’ orders need not be included since the part of
this cost that depends on the operating doctrine is included in the stockout
costs). There is another reason why the profit or cost will differ from what
would be computed from accounting records. This is because the stockout
and carrying costs include components which are not out of pocket costs,
but instead represent goodwill costs or opportunity costs. In many of the
models that we shall study, the revenues received will not depend on the
operating doctrine, and in this case the maximization of profits will be
equivalent to the minimization of the variable costs. In such cases we shall
formulate the problem as a cost minimization problem.
Consider now the question of the time period over which the profit or
cost is to be computed. For models which assume that an item will be
carried in the inventory system for only a limited time (perhaps because
it will become obsolete or is perishable), the profit that is to be maximized
or the cost to be minimized is the profit or cost for the entire length of time
for which the item is carried by the inventory system. However, in deter¬
mining the profit or cost one must, if the time period is long enough, take
account of the time value of money, so that what should be used in such
situations is the present worth of the profit or cost (i.e., the discounted
profit or cost). The maximization of profits or minimization of costs over
a time period of finite length will also cover those models in which the mean
rate of demand is allowed to change with time, since in such situations only
a finite time period (planning horizon) can be treated.
For models in which the nature of the processes generating the demand
pattern and lead times does not change with time (this implies, in par¬
ticular, that the mean rate of demand is constant over time) and for which
 SEC 1-13 SELECTION OF AN OPERATING DOCTRINE 23

it can be assumed that the item can be inventoried indefinitely, it is

usually convenient to imagine that the inventory system will continue to
operate for all future time. One logical procedure then is to maximize the
discounted profits or minimize the discounted costs over all future time.
Note that discounting must be used here if the profit or cost is to be finite.
Now>_ however, an alternative criterion suggests itself. This alternative
criterion is to determine the operating doctrine by maximising the average
annual profit or minimizing the average annual cost. [By definition, the
average annual profit <P and the average annual cost X are the long run
average annual profit and cost, respectively, and are defined as follows.
Let ‘W(f), $(f) be the total profits received and costs incurred (not dis¬
counted), respectively, for a time period of length f years. Then

<P = X = lim m (1-3)

r
For models of the type under discussion we shall in general find it more
convenient to maximize the average annual profit or minimize the average
annual cost, rather than maximizing the present worth of jail future profits
or minimizing the present worth of all future costs.
It would be very disconcerting if the optimization of the average annual
rates of incurring profits or costs led to an operating doctrine which was
quite different from that obtained by optimizing the present worth of all
future profits or costs. Fortunately, this is not the case. For the range of
interest rates and average times between the placement of orders which
are of practical interest, both methods will give either identical results or
almost identical results. It is easy to see, in fact, that if the annual profit or
cost was the same each year, so that (P, X would also represent the profit
and cost, respectively, for each and every year, then the two procedures
would give identical results for any value of the interest rate. For example,
let 3 be the present worth of all future costs, and X the annual (and also
average annual) cost. Then if i is the interest rate and a = (1 -f- i)~\
it follows that if we discount on a year by year basis

d = aX + a?X + a*X + ... = aX( 1 + a + a2 + . . .)

a
(1-4)
1 — a

Thus 3 and 3C differ only by the constant factor 1/i, and hence minimizing
either one also minimizes the other.
Usually, however, the actual profit or cost will fluctuate from one year
to the next, sometimes being above or below the average annual values as
defined above. Because of this it is possible, by making the interest rate
sufficiently high, to obtain different results by minimizing the average
24 THE NATURE OF INVENTORY SYSTEMS CHAP. 1

annual cost and minimizing the present worth of all future costs, or maxi¬
mizing the average annual profit and maximizing the present worth of all
future profits. £As has been noted above, however, the differences will be
negligible in almost all cases of practical value when the interest rate re¬
mains within the limits appropriate to the real world. It might be noted,
however, that the effect of the interest rate becomes more important as the
average time between the placement of orders is increased. In any case
where the interest rate was large enough to make a difference, then the
present worth would be the function to be optimized}
There remains the question of how any stochastic elements in the
problem are to be handled. For models in which the mean rate of demand
is imagined to remain constant over all future times, we shall again maxi¬
mize the average annual profit or minimize the average annual cost. When
stochastic elements are present, however, it will be necessary to introduce
probabilities and compute expected values in order to determine the aver¬
age annual profit or cost. These expected values will arise in a natural way
as time averages.
For situations where the mean rate of demand changes in some fashion
with time, we shall use the criterion of maximizing expected profits or
minimizing expected costs over some relevant time interval. It is not com¬
pletely evident in such cases that expected values are what should be used.
It would be quite reasonable to use expected values if it was true that the
system would repeatedly encounter the same sorts of conditions and that
the expected profit or cost could be interpreted as the average profit or
cost averaged over all times that the system faced a particular set of
circumstances. However, it will frequently be true that the system will
encounter a particular set of conditions only once, and they wifi never be
repeated agam. The rationale for using expected values in such cases lies
m the modem theory of utility introduced by von Neumann and Morgen-
stem [7], We shall not attempt to examine this theory in detail but
instead shall merely point out that it roughly implies the following: If it
is possible for an individual to express consistent preferences between
various situations whose outcomes must be described probabilistically
then there exists a function having a numerical value associated with
evepr possible outcome such that if he selects among the alternatives in
such a way that the expected value of this function is maximized, then he
wifi be acting m a way that is truly representative of his preferences The
numerical function referred to above is often caUed a utility function. We
shall assume m this text that the system under consideration does have a
utility function defined over the possible outcomes in any inventory
problem. Thus the system should always behave in a way that maximizes
its expected utility. We shaU further assume that the utility can be
measured in monetary units and is, in fact, with the proper definition of
 REFERENCES 25

the relevant components, simply the expected profit over some appropriate
time period. It is in this way that we justify the use of the expected profit
even in situations where the given set of conditions can never be expected
to be encountered again. We shall also see that in a number of cases the
maximization of expected profit is eequivalent to the minimization of ex¬
pected cost. In these cases then, one can equally well minimize the expected
cost over some relevant time interval.
We have noted previously that it is often very difficult in practice to
determine the stockout cost functions. To avoid this problem, an alterna¬
tive procedure might be to maximize the profit or minimize the cost, each
exclusive of the stockout cost, subject to a constraint that the average
fraction of the time for which the system is out of stock is not greater than
a specified value. Here, instead of specifying the nature of the stockout
cost, one instead specifies an upper limit to the average fraction of the t.imp
for which the system is out of stock. We shall also examine criteria of this
sort for the determination of operating doctrines and shall also investigate
their relation to the cases where a stockout cost function is specified.
Before closing this introductory chapter, we might say something about
the physical dimensions that will be employed in formulating the profit or
cost expressions in the remainder of the text. The equations will hold for
any consistent set of physical dimensions. However, we shall always
imagine that the monetary unit is the dollar and the time unit is a year.
Of course, physical quantities of any good will be measured simply in units,
where the unit need not be defined more precisely. A unit may, for example,
be a case of screwdrivers, a dozen screwdrivers, or a single screwdriver'
depending on the application. The only important thing to remember is
that no matter how the unit is defined, the definition must be used con¬
sistently throughout.
As a final remark it is worth observing that while we shall be concerned
with optimizing the mathematical models to be studied, it does not follow
that the real world system represented by the mathematical model will
also be optimized. Since a number of simplifying assumptions and approxi¬
mations must be made to obtain the mathematical model, the most we can
expect is that by using the optimal solution for the model in the real world
we shall have a “good” operating doctrine or one that is an improvement
over an existing operating doctrine. Indeed, in the real world situations,
there are usually such a large number of complications that it is exception¬
ally difficult to define what an optimal operating doctrine mpa.ns

REFERENCES
1. Arrow, K. J., T. Harris, and J. Marschak, “Optimal Inventory Policy,”
Econometrica, XIX (1951), pp. 250-272. ’
26 THE NATURE OF INVENTORY SYSTEMS CHAP. 1

2. Arrow, K. J., S. Karlin, and H. Scarf, Studies in the Mathematical Theory

of Inventory and Production. Stanford, Calif.: Stanford University
Press, 1958.

3. Dvoretzky, A., J. Kiefer, and J. Wolfowitz, “The Inventory Problem:

I, Case of Known Distributions of Demand; II, Case of Unknown
Distributions of Demand,” Econometrica, XX (1952), pp. 187-222
and 450-466.

4. Dvoretzky, A., J. Kiefer, and J. Wolfowitz, “On the Optimal Character

of the ($, S) Policy in Inventory Theory,” Econometrica, XXT
(1953), pp. 586-596.

5. Harris, F., Operations and Cost (Factory Management Series.) Chi¬

cago: A. W. Shaw Co., 1915, pp. 48-52.

6. Raymond, F. E., Quantity and Economy in Manufacture. New York:

McGraw-Hill Book Co., 1931.

7. Von Neumann, J., and 0. Morgenstem, Theory of Games and Economic

Behavior. Princeton, N.J.: Princeton University Press, 1953.
8. Whitin, T. M., The Theory of Inventory Management. Princeton, N. J.:
Princeton University Press, 1953.
DETERMINISTIC

LOT SIZE MODELS

AND THEIR EXTENSIONS

Order is a lovely thing;

On disarray it lays its wing,

Teaching simplicity to sing.

Anna Hamsted Branch, The Monk in the Kitchen

 2-1 Introduction

o begin our discussion of inventory systems, an especially simple col¬

lection of models will be studied for which the rate of demand for units
stocked by the system will be assumed to be known with certainty and to
be constant over time. This chapter will be devoted to a study of such
models. As we mentioned before, in the real world demands can almost
never be predicted with certainty; instead they must be described in proba¬
bilistic terms. However, the deterministic models to be discussed are still
o interest because they provide a simple framework for introducing the
methods of analysis that will be used in more complicated systems and
ecause, sometimes, they are useful in examining certain aspects of real
world problems. Furthermore, the results obtained from these models
yield, qualitatively, the proper sort of behavior-^ven when the deter¬
ministic demand assumption is removed.

2-2 The Simplest Lot Size Model—No Stockouts

Let us consider the problem of controlling the inventory of a given item

at a single location (a retail store, for example) when the assumption is
made that the rate of demand for the item is deterministic and is a constant
Xjumts_per yearjndependent at time. The fundamental prohle~fn7Th^
system, andindeed for any inventory"system, as was pointed out in Chapter
1, is to determine when an order should be placed and how much should
be ordered. We shall suppose that the procurement lead time r is a constant
independent of X anJTBe quantity ordered. Furthermore, it wilfbe
thaFthe entire"quant^ofasi^l^eTTvered as a single package, i.e., it
never happens that an order is split so that part of it arrives at one time
and part at another time. We shall imagine that the item can be inven¬
toried indefinitely, and that it will never become obsolete. Then, as sug¬
gested in Chapter 1, it is convenient toimagine that the system will
continue to operate for all future time.} Since X and r are constant and
deterministic, it is immediately clear that when the system is operated
optimally, the same quantity will be ordered each time an order is placed
29 ’
30 DETERMINISTIC LOT SIZE MODELS CHAP. 2

and the on hand inventory at the point in time when an order arrives will
always have the same valued"!
In this section we wish to study a case where the system is never out
of stock when a demand occurs. This limitation can legitimately be im¬
posed since the demand is deterministic and the procurement lead time is
a constant. Initially it will be convenient to imagine that the quantity
demanded in a time £ is a continuous function of t, and to ignore the fact
that an integral number of units must be demanded. Later, account will
be taken of the integrality of demand. jjSince the system is never out of
stock when a demand occurs, and since X is a constant, the annual revenues
received from the sale of the item are a constant, independent of the oper¬
ating doctrine, and hence, the minimization ofcosts will yield the same
operating doctrine as the maximization of profit&J Following the suggestion
of Chapter 1, we shall determine the optimal Operating doctrine by mini¬
mizing the average annual cost. Recall that only those costs that depend
on the operating doctrine need be included. These costs have been dis¬
cussed in Chapter l.^The costs appropriate to this model are the cost of
the units purchased, the fixed cost of placing an order (the ordering cost),
and the inventory carrying costs^ It will be assumed that the costs of
operating the information processing activity are independent of the order
size and the reorder rule, and therefore do not need to be included in the
cost expression.
[To compute the average annu'al cost, we must compute the total cost
for an arbitrary time period of length f, then divide by f to obtain the aver¬
age annual cost for the period of length and finally allow f to approach
infinity, giving the desired average annual cost| The actual cost of operat¬
ing the system can vary from one year to another because the number of
orders actually placed can vary. Only when a year is an integral multiple of
the time between placing orders can the actual system cost be the same
each year. The differences between average yearly cost and the actual
system cost for a given year will be illustrated later by a concrete example.
The procedure we are about to use to determine the average annual cost
may, at first reading, seem unnecessarily complicated for the simple model
being studied. It is introduced now because it provides a rigorous way of
obtaining the average annual cost and because it allows us to introduce, in
the context of a very simple model, a technique which will later prove quite
useful in dealing with more complex models involving uncertainty. Further,
the technique used to derive the average annual cost is of course the
definition of the average annual cost, i.e., it is by definition the limit as f
approaches infinity of the average cost for a time period of length f.

* A rigorous proof that the optimal operating doctrine has this form can be given
using the methods to be introduced later in Chapter 8.
 SEC 2-2 SIMPLEST LOT SIZE MODEL—NO STOCKOUTS 31

If a quantity Q is ordered each time the system orders replenishment

stock, then after every Q demands an order for Q units is placed. Thus the
time T between the placement of orders is T = Q/\. Similarly, the time
between the arrival of successive procurements is T. We shall say that the
system goes through one cycle of operation in the time between the place¬
ment of two successive orders or the receipt of two successive procurements,
or, more generally, between any two points in time separated by an interval
of length T. During each cycle, the system repeats exactly its behavior
during the previous cycle. The length of a cycle is T.
Let v be the largest integer less than or equal to f/27. Then in time f
there are v complete cycles, and perhaps a fraction of another cycle. This
is independent of where we select the time origin. Since precisely one order
is placed per cycle, the number of orders placed in time t will be vary 4- 1;
it can be v + 1 if the included fraction of a cycle is long enough. The
number of orders placed can be written (f/77) + €, where |c| < 1 (|€| is the
absolute value of e), or (Af/Q) + e. If A is the cost of placing an order, then
the ordering costs incurred in the time f are [(f/T) + e]A. The cost of the
units ordered in the time f will be [(Af/Q) + e\QC if C is the unit cost of
^he item when ordered in lots of Q units. '
^e^^bothe^ESH~mventory lhTthe system at the time'of arrival of a
procurement. The on hand inventory immediately after the procurement
arrives will be s + Q. Consequently,
are

ICjJ 0Q + s - Xt) dt = IC [(Q + s)T - ^P] = /Ct[| + a] (24)

This follows since at any time t from the beginning of the cycle (which is
taken for convenience to be the time of arrival of an order), the on hand
inventory is Q + $ — A£, and the cost incurred between t and t + dt is
IC(Q + s — At) dt Summation over the length of the cycle gives (2-1)
when the relation T = Q/X is used.
In the time f there are

v = (f/77) - f, 0 < S < 1

full cycles. The inventory carrying costs for the time period of length f are

icr[| -$] ^l + sJ + ,7
where y is the inventory carrying cost for the fractional cycle of length
$ — vT; note that y is less than the inventory carrying cost for a single
cycle. The total variable cost for the time period of length f is then

+ «] QC + r + «] A + ICT [| - *] [| + ,] + „ (2-2)
32 DETERMINISTIC LOT SIZE MODELS CHAP. 2

The average annual cost for the time period of length f, X?, is obtained by
dividing the total cost by This yields

3Cr = XC + ^ ^ A + j + IC [| + s] - + s] + *

To obtain the average annual cost of the system, we allow f to go to

infinity. Therefore the average annual variable cost is

3C = XC + ^ A + IC [| + s]; (2-3)

The formula for X can also be derived using the following simple argument:
Since there are X demands per year and sinceall demands are met^ then
on thejjiverage X units per year must be procured at a cost of XC. Similarly,
if the order quantity is Q, then the number of orders placed per year must
average to X/Q, and the fixed procurement costs per year average to \A/Q.
Furthermore, by assumption, the inventory carrying cost per unit is* pro¬
portional to the length of time it remains in inventory. Consequently, the
inventory carrying cost per year must then be IC times the average inven¬
tory. The average inventory is one half the sum of the maximum inventory
Q + s and the minimum inventory s, i.e., (Q/2) + s. Summation of these
three terms gives (2-3).
For the moment (indeed, in all discussions through Sec. 2-10), we shall
restrict our attention to situations for which the unit cost of the item is
independent of the quantity ordered. Then XC is independent of Q and the
reordering rule and need not be included in the variable cost. Hence the
relevant average annual variable cost in this case, which is the sum of
ordering and inventory carrying costs, is

X = qA + IC ^ J (2-4)

We use the same symbol X in (2-4) as in (2-3) since by convention we shall

use X to represent the average annual variable costs; X will be assumed
to include all variable costs and no costs which are not a functi on of the
system variables Q, s.
Examination of (2-4) shows that the only term which depends on the
reorder rule is ICs. This term is minimized by having s = 0, so that the
system just runs out of stock as the new procurement arrives. The require¬
ment that s = 0 in order to minimize (2-4) allows us to determine im¬
mediately the optimal reordering rule for any given Q value. Let m be
the largest integer less than or equal to r/T, where r is the procurement
lead time. ’ Then, if we place an order when the on hand inventory reaches
the level
Tk = X(t — mT) — Xr — mQ = /x — mQ (2-5)
SEC 2-2 SIMPLEST LOT SIZE MODEL—NO STOCKOUTS 33

where m = Xt is the lead time demand (i.e., the number of units demanded
from the time an order is placed until it arrives), the on hand inventory will
be zero at the time the order arrives. The number rh is called the reorder
point; each time the on handfinventory in the system reaches rh an order
for Q units is placed. This is illustrated graphically in Fig. 2-1.
It remains to determine the optimal value of Q. We have seen that the
optimal value of s is zero for any Q. Thus the average annual cost really
depends only on the single variable Q; it can be written--

K = ±A+IC\ (2-6)

FIGURE 2-1.

(the value of 3C optimized with respect to s will always be written K). It

is desired to find that value of Q > 0 which minimizes (2-6). Recall that
demand is being treated as a continuous variable; thus Q can be treated as
continuous also. The calculus tells us that if the optimal Q (call it Q*—in
fact, all optimal quantities will hereafter be indicated by asterisks) lies in
the interval 0 < Q < °o, then it is necessary that Q* satisfy the equation

^ = o =-Aa , rc
dQ U Q*A+ 2
or
34 DETERMINISTIC LOT SIZE MODELS CHAP. 2

It will be noted that K = co when Q — 0 or oo and that K is finite for any

other Q > 0. Furthermore K is differentiable for all Q > 0. Thus the opti¬
mal Q must satisfy (2-7) and hence (2-8). However, (2-7) has only one
solution for Q > 0. Consequently, (2-8) gives the value of Q which yields
the unique absolute minimum of K for Q > 0.
The fact that (2-8) gives the Q which yields the absolute minimum of
K can be seen in another way. Recall that the calculus tells us that if Q*
satisfies dK/dQ = 0 and if d2K/dQ2 > 0 when evaluated at Q*, then Q*
yields a relative minimum of K. If, however, d2K/dQ2 > 0 for all Q in the
region of interest, not simply at Q*, then Q* yields the unique absolute
minimum of K in the region of interest. Now

d2K _ 2\A
(2-9)
dQ2 Qz

for all Q > 0 and hence the Q determined from (2-7) yields the absolute
minimum value of K.
The equation (2-8) is referred to in the literature under a variety of
names. It is sometimes called the lot size formula, the economic order
quantity, the square root formula, or the Wilson formula. We shall oc¬
casionally use several of the above names. Equation (2-8) might also be
called the Harris formula since, as indicated in Chapter 1, Harris seems to
have been the first to have derived it. Because the square root expression
in (2-8) appears frequently in later work, we use a special symbol Qw to
identify it; whenever Qw appears, it will always be defined by (2-8).
The problem of determining how to operate the system has now been
solved. The reorder point, given by (2-5) (with Q* replacing Q), tells us
when an order should be placed. The quantity to be ordered is given by
(2-8). Specification of the reorder point and the order quantity determines
all other quantities of interest.

2-3 Additional Properties of the Model; An Example

It is instructive to examine in somewhat greater detail the nature of the

model described in the previous section and the optimal operating doctrine
for it. If the inventory carrying costs, ordering costs, and K are plotted as
a function of Q on the same graph, one obtains curves something like those
shown in Fig. 2-2. The figure shows clearly that the Q which yields the
minimum is unique and that d2K/dQ2 > 0 for all Q > 0. It might also be
noted that the optimal Q occurs at the point where the slope of the ordering
cost curve is the negative of the slope of the inventory carrying cost curve.
It will be left to Problem 2-2 to demonstrate that, in this special case, the
two curves intersect at this point.
SEC 2-3 ADDITIONAL PROPERTIES OF THE MODEL 35

When an optimal policy is used, the average amount of inventory in the

system will be

Q*
2 (2-10)

The on hand inventory fluctuates between Q* and 0 and averages to Q*/2.

Equation (2-10) points out the interesting result that the average inventory
(and the maximum inventory, and Q*) should increase as the square root

FIGURE 2-2.

of the sales rate and not proportionally with the sales rate. Similarly, the
average inventory varies inversely as the square root of the cost so that the
average inventory for high cost items should be lower than for low cost
items, all other things being equals The above observations merely serve
to point out that in formulating an inventory policy, items with different
characteristics should be treated differently. This implies, for example,
that if a system was stocking a number of items with widely different rates
of demand and costs, it would not be optimal to require that the average
inventory for each item should be k weeks of stock where the same k was
used for all items.
36 DETERMINISTIC LOT SIZE MODELS CHAP. 2

The average annual cost of procurement and holding inventory K* under

the optimal procurement policy will be

K* -\a( IC \m . IC /2XAV/2 /I , . T J\V* /l \ 1/2

K -u(,2uj +t(,7c) ~{2^ic) +(|mc)
= V2XAIC = Kw (2-11)

The value of K* is also proportional to X1'2. The square root expression

appearing in (2-11) will appear frequently in later work and a special
symboi Kw will be used to represent it; whenever Kw appears it will be
defined by (2-11).
It is of interest to compare K* with the cost K of operating the system
w en a Q different from Q* is used. To make this comparison, we shall
compute K/K* as a function of Q/Q*. Now

X* = qA(2\AIC)-^ + /c| (2\AIC)~112 = + qI] (2-12)

Note that this equation is completely independent of the system pa¬

rameters. A plot of Eq. (2-12) yields the universal curve shown in Fig. 2-3.
lhe mterestmg thing about this curve is that it is rather flat in the neigh-
 SEC 2-3 ADDITIONAL PROPERTIES OF THE MODEL
37

borhood of the optimal Q. If the actual Q is off from the optimal Q m either
direction by a factor of two, costs are increased by only 25%. The same
sort of analysis can be used to study how the actual cost will vary from the
optimum cost tf one of the parameters such as X or A is not measured
proWems Exammatlons of sensitivity analyses of this sort are left to the

In the event that the procurement lead time is less than one cycle there
will never be more than a single order outstanding. Furthermore, there
will be no orders outstanding at the time immediately prior to placing an
order (i.e., just as the reorder point is reached). On the other hand, if the
procurement lead time is longer than one cycle, there will always be at least
one order outstandmg. As an aid in visualizing the situation where the

placed received

FIGURE 2-4.

procurement lead time is longer than one cycle, it is helpful to draw a

chart such as that shown in Fig. 2-4 which illustrates a case where the
procurement lead time is 3§ times the length of a cycle. It will be observed
that (after period 4 in Fig. 2-4) there are always either 3 or 4 orders out-
s ndmg For half the period three orders are outstanding and for half
the period four orders are outstanding. The average number of orders
outstandmg is therefore 3|. Hence the average amount on order is 3* Q
which is precisely the lead time demand We have thus shown in this
special case that the average amount on order is equal to the lead time
demand. Problem 2-10 asks the reader to prove that the average amount
on order is equal to the lead time demand for any arbitrary lead time r.
An 2j~4 the q,uantlty on order Just Prior to reaching the reorder point
is 3Q and immediately after hitting the reorder point is 4Q. In general
the amount on order just prior to reaching the reorder point is mQ where
38 DETERMINISTIC LOT SIZE MODELS CHAP. 2

m is the largest integer less than r/T, and (to + 1) Q immediately after
hitting the reorder point (or, in the razor’s edge case where r/T is an
integer to, there are always precisely to orders outstanding).
It is often useful to consider the quantity on hand plus on order. At the
reorder point the on hand inventory is p - mQ, and the quantity on hand
plus on order increases by Q units when the order is placed. Thus the
quantity on hand plus on order fluctuates between p and p + Q. The re¬
lationship between the on hand inventory and the on hand plus on order
inventory is shown in Fig. 2-5. The on hand inventory reaches its minimum
value 0 just prior to receipt of a procurement, and its maximum value

Q immediately after receipt of a procurement. The on hand plus on

order inventory reaches its minimum p just prior to reordering and its
maximum p + Q immediately after placing an order.
For the system under consideration, specification of the on hand inven¬
tory uniquely determines the quantity on hand plus on order, and con¬
versely, the specification of the quantity on hand plus on order uniquely
determines the amount on hand.* One can therefore specify a reorder point

—are me reorder point when the quantity on hand plus on order

lumps discontmuously, and the points in time corresponding to arrivals of procurements
when the on hand inventory jumps discontinuously. P
 SEC. 2-3 ADDITIONAL PROPERTIES OF THE MODEL
39

in terms of the on hand plus on order inventory. This reorder point will
be denoted by r, and r = (ji being the lead time demand); thus when the
on hand plus on order inventory reaches a level M an order is placed Later
when uncertainty in the demand is introduced, we shall see that the reorder
point cannot always be legitimately determined in terms of an on hand
inventory. Instead, it must be specified in terms of the on hand plus on
order inventory (or on a more general level defined to be the quantity on
hand plus on order minus backorders). It is only for certain special cases
that the reorder point can be specified in terms of an on hand level.

EXAMPLE Consider a system of the type described above which inven¬

tories an item whose parameters have the following values:

X = 600 units/yr., I = 0.20 "L~~ ,

A = $8.00, ip r = 1 year

C = $0.30 £

To be more specific, it can be imagined that this is a low cost item carried
by a department store. The high fixed cost of ordering arises because the
item is procured in Europe. This also accounts for the long lead time.
The optimal order quantity is

Q* = ftP 12(600) (8)

= 400
\IC V 0.20(0.30)

The time between placement of orders (the length of a cycle) is

r* - 91 _ 400 2
X 600 3F'

The lead time demand is

H = Xr = 600 units.

f^ie reor<^er P°in^ based on the on hand plus on order inventory level is then
r* = 600. To compute the reorder point based on the on hand inventory
level, we first note that t/T = f. The largest integer less than r/T is 1
Thus
rt = /x — Q* = 200 units.

The minimum average yearly cost of ordering and holding inventory is

K* = V2XAIC = V2(600) (8) (0.06) = V576 = $24

The yearly cost of the units themselves is
40 DETERMINISTIC LOT SIZE MODELS CHAP. 2

Thus the ordering and holding costs are a small part of the total cost
which includes the cost of the units.
The actual yearly procurement and holding costs vary from one year to
another. This can be seen from Fig. 2-6. Assume that we start measuring
tune immediately after the placing of an order. Then in the first year
shown, one order is placed, while in the second year two orders are placed.
After the second year a repetition of the above is obtained. For the first
year shown in Fig. 2-6, the actual system costs are $8 for ordering and $10
for holding, or a total of $18. In the second year, the actual costs are $16

for ordering and $14 for holding or a total of $30. The average of these
two values gives $24, which is that obtained above for K*.

2-4 Accounting for Integrality of Demand

In the previous sections demand has been treated as a continuous vari¬

able. For most practical applications of the simple model studied, Q* is
sufficiently large that it is quite satisfactory to treat the demand as con¬
tinuous and to round off the value of Q* to the nearest integer, since an
integral number of units must be ordered. In reality, of course, each de¬
mand must be for an integral number of units, and in addition Q* must be
an integer. It is instructive to examine a ease where the integrality of
demand is accounted for, and the Q values are restricted to being positive
integers. To be specific, it will be assumed that units are demanded one at
 SEC 2-4 ACCOUNTING FOR INTEGRALITY OF DEMAND
41

a time and the time between demands ts is known with certainty. Then
X = 1 /t„.
As in the continuous case, the number of units s on hand when an order
arrives should be zero. Now, however, this requirement does not uniquely
specify the time when an order should be placed since there is a time inter¬
val of length t, between demands. It should be noted, however, that there
is no need to have the order arrive until precisely the moment when a
demand occurs for otherwise costs will be increased. Thus the order should
arrive at a time t„ after the demand which reduces the on hand inventory
to zero, i.e., the system will have a zero on hand stock level for a time t,
each cycle. Immediately after the procurement arrives one unit is de¬
manded. Hence, the maximum inventory level is Q-1 and the minimum is
zero.
The inventory carrying costs per cycle are then*

ICt3[Q — 1 + <2 — 2 + .. . + 1+0]= JC ^ (~2 ~) (2-13)

The average number of cycles per year remains X/Q so that the average
mventory carrying costs per year are } 1C (Q - 1). The average yearly
procurement costs remain unchanged at XA/Q. Thus the average annual
cost of holding and procuring inventory is for a given Q > 1

K(Q) = ±A+\lC(Q- 1) (2.i4)

If Q* is the smallest Q value which minimizes K(Q), Q* being an integer

then if Q* > 2, it is necessary that 7

K(Q*) - K(Q* - 1) < 0; K(Q* + 1) - K(Q*) > 0 (2-15)

The optimal Q is then the largest Q for which

AK(Q) = K(Q) - K(Q - 1) < 0

or
Q* = 1
However, for Q > 1

= 2Q(Q - 1) C-2XA + ICQ(Q - 1)] < 0

*The reader should recall that

1 +2 +. . n(n + 1)
+n -
2
42 DETERMINISTIC LOT SIZE MODELS CHAP. 2

Thus Q* is the largest positive integer Q for which

Q(Q - 1) < (2-16)

When unity is negligible with respect to Q and it is assumed that Q can

vary continuously, then one obtains (2-8). In the discrete case just studied,
Q need not be unique. If the strict equality holds in the second condition
of (2-15), then either Q* or Q* + 1 is optimal.
The determination of the reordering rule is a little more complicated
than in the continuous case. If T is the length of a cycle, r is the procure¬
ment lead time, and if m is the largest integer less than or equal to r/T,
then t = r mT gives the time before the next order arrives when an
order should be placed. If m is the largest integer less than or equal to
t/ts, then a reorder should be placed at a time (1 + tii)t9 - t after a demand
reduces the on hand inventory to m (illustrate this graphically). A similar
analysis shows that in terms of the on hand plus on order inventory level
an order is placed at a time (m + 1 )t8 — r after the demand which reduces
the on hand plus on order inventory to M, where m is the largest integer
less than or equal to r/t8.

2-5 Case Where Backorders Are Permitted

The model presented in Sec. 2-2 required that all demands could be met
from stock, i.e., the system was never out of stock when a demand occurred.
We shall now study the more general case in which all demands must be
met ultimately, but it is permissible for the system to be out of stock when
a demand occurs. In such a case the demands occurring when the system
is out of stock are backordered until a procurement arrives. When a pro¬
curement does arrive it is assumed that all backorders are met before the
procurement can be used to meet any other demands.!
Clearly, if there were no costs associated with incurring backorders, then
it would be optimal never to have any inventory on hand. On the other
hand, if backorders are sufficiently expensive, then one should never incur
any. However, for an intermediate range of backorder costs, it will be opti¬
mal to incur some backorders towards the end of a cycle. In accordance
with the discussion of Chapter 1, we shall assume that the cost of a back- ]
order has the form w + M where t is the length of time for which the back-!
order exists. The cost of a backorder thus includes a fixed cost t and a
cost M which is proportional to the length of time for which the backorder!
exists. _/'
Let s be the number of backorders on the books when a procurement of
 SEC 2-5 CASE WHERE BACKORDERS ARE PERMITTED 43

Q units arrives (s is a non-negative number). Then, after satisfying the

backorders, Q-s units will be on hand, since s units were used to fill the back¬
orders. The time required for the Q-s units to be demanded will be Ti =
(Q ~ «) A- The length of time during one cycle over which backorders will
be incurred will then be T2 = T - Ti. The behavior of the system is
illustrated graphically in Fig. 2-7.*
The inventory carrying costs per cycle aref

IC JoTl (Q - s - M) dt = g (Q - ,)*

This result is also immediately evident from Fig- 2-7, since (Q — s)2/2X

FIGURE 2-7.

is the area of triangle (1). As before, there are on the average X/Q
cycles per year, so that the average yearly cost of carrying inventory is
IC{Q - s)2/2Q.
The backorder cost per cycle is

ITS Hr
f
X f
Jo
Xt dt = 7TS "j” X
z
TrXTl = TTS + ~-
zx

* Note that the minimum value of Q — s is zero. Costs could never be reduced
.by having backorders on the books immediately after the arrival of a procurement.
Thus, for a given Q, s must be in the interval 0 < $ < Q.
t Here and in the remainder of the chapter, we revert to treating the demand and
Q and s as continuous variables.
44 DETERMINISTIC LOT SIZE MODELS CHAP. 2

since
rr = s/x
2

The average annual cost of backorders is then

§['>*+ **’] 1
The average annual variable cost X, which includes the cost of ordering,
holding inventory, and backorders then becomes

« = IC(Q - 5)2 + £ [tXs + \ *«2] (2-17)

Here, X is a function of Q, s. We wish to find the absolute minimum of X

in the region 0<Q<oo?0<s. It is clear that for any finite s, X is in¬
finite when Q — 0 or oo. Thus the optimal Q, i.e., Q*, must indeed satisfy
0 < Q* < oo. If the optimal s, i.e., $*, satisfies 0 < $* < oo, then the differ¬
ential calculus requires that since X is differentiable everywhere in the
region of interest, Q*, s* must satisfy the equations

ax = ax
dQ ds (2-18)

As we shall see, the optimal solution need not always satisfy (2-18). In
certain cases, the optimal $ may be on the boundaries.
Consider now the problem of finding the solution to (2-18). From (2-17)
we see that
dx i r i -j
dQ ~ Q2 ^ 2 s)2 2

jn
+ ~q (Q — s) =0 (2-19)
or

2 s)2 “f” Q(Q ®) = Yq XA + 7rXs + 2 vs2l (2-20)

L. J
so

- Q2 = — j~\A + 7rXs + - x$2J + | s2 ■ (2-21)

Next observe that

or
f («--)+irX + ^.-0 (2-22)

Q = + (* + Ic)s (2-23)
sec 2-5 CASE WHIM BACKORDERS ARE PERMITTED 45

To solve explicitly for Q and a in terms of the system parameters, first

substitute <2 21) f«»r Q in i2-2!t. This yields after regrouping; the terms

l*1 i #/<v i i ux)a - 2\.i/r o (2-24)

We note immediately that if $ 0, (2-TO reduces to (n-XV- 2\AK\
wh.eh is not true in general. This simply means that, in general, when
t (I, there is no solution a sueh that 0 s **, In other words, the
solution is on the boutularies so that s 0 or a •«, In the event that.
(uX)5 2X.l/t', any s value in the interval satisfies (2-24).
When * tt, it is easy to determine whether a 0 or a ■ no is optimal.
It is only neee-sary to eompare the eorresponding eosts optimized over Q.
When a* 0, we see that (2 l7i reduees to (2-0! so that the optimal Q is
Qw given by 12-S> and the minimum cost is A‘„. When «* «, an order is
never planed and the minimum yearly oust is *X. I’his means that, in
reality the inventory system should nut Ik- uja-mted at all; it is better
simply to incur the backorder costs year after year. We see that if

»■ > \l“ ^ ft ‘»r r\ > A"«. (2-25)

then tin1 optimal solution has a* 0, so that Q* is given hy (2-H); if rr <5,

there should not be any inventory system. When v ■ 5, any value of #,
0 ‘ » * ' ti optimal (of course, the optimal Q value depends on the a
chosen). In the ease where * X, we have shown that the costs are the
same when a tt, and a ' . We leave it for Problem 2-IK to show that
the same cost is obtained for any other value of a.
Next consider the ease where f / 0. Then solution of the quadratic
equation (2-21/ yields the result

«* [* b IV\ ‘ | nX } [(2X/1/0 (l } *

fr / 0 (2-20)
To determine Q explicitly, (2-2.1) is used to eliminate a in (2-21). This
yields

2 y i vkt( _
tv kA f t I IVQ i n
tv »x
tv ! 1 )l. + tv,Q
* * + IV I
Thus Q*, the optimal value of Q, m
46 DETERMINISTIC LOT SIZE MODELS CHAP. 2

In the event that the s value computed from (2-26) is negative, then the
optimal s lies on the boundary, i.e., s* = 0. We leave it for the reader to
demonstrate in Problem 2-62 that if xX > Kw, s* = 0. When it ^ 0, s*
cannot be infinite (why?). Equation (2-27) holds only when the s computed
from (2-26) is positive; otherwise, Q* = Qw. Furthermore, if the s com¬
puted from (2-26) is positive, it is the optimal s, i.e., s* does not lie on the
boundary s = 0. To prove this recall that when s = 0, Q* = Qw; however
dX/ds < 0 at s = 0, when Q = Qw and Kw > xX, so that s = 0 cannot be
optimal if Kw > xX.
When, x = 0, (2-26), (2-27) become respectively

y 0* _ r 2XAIC 1^/2
[_#(x + /C)J~ + IC)] v* (2-28)
and

* + IC l1'2
^ Q* = : (2-29)

It will be noted from (2-28) that when x = 0, then s* > 0 unless x = oo,
i.e., under optimal operating conditions, some backorders will always be
incurred. Give an intuitive explanation for this result. —
The computation of the reorder point for this model is, in principle, the
same as that presented in Sec. 2-3. Now, however, the inventory levels
used must be redefined. The on hand inventory is no longer appropriate
since there may be no inventory on hand, but instead there may be back¬
orders at the time when an order should be placed. The appropriate level
to replace the on hand inventory is the amount on hand minus the back-
prders, which will often be referred to as the net inventory. If there is
inventory on hand, there will be no backorders and this level will be
positive. If there are backorders, there wifi be no inventory on hand and
this level will be negative. In .terms of the net inventory level, the reorder
pomt is then rt - „ - mQ* - s* where, as before, m is the largest integer
less than or equal to r/T. [It is possible for r*h to be negative. This
that an order is placed when the backorders reach a level jr£[.. The level
to replace the on hand plus on order inventory of Sec. 2-3 is the amount
on hand plus on order minus baekorders. This will often be referred to as
the inventory position of the system. The reorder point in terms of the
inventory position is r* = M - s*; r* can also be negative!

dkaS^W^v^T he °f d mVent0ry 10 be negative, a negative value in-

tat f-that+baC Ord“. With this convention it is unnecessary to introduce
the net inventory level. However, we prefer to follow the convention ^ways used in
industry and the rmhtaxy of having the on hand inventory be a non-negative variable.
SEC 2-6 THE LOST SALES CASE 47

EXAMPLE Consider an item with the following characteristics:

X = 200 units/year; C = $25

I = 0.20; A = $5.00

7T = $0.20 per unit; f = $10 per unit per year

r = 9 months

From (2-26)

s* = [10 + 5]_1{—40 + [2(200)(25) (1 + *) - ^(lfiOO)]1/2} = 5.27 » 5

Then, from (2-27)

The time between procurements is

94
r* = 200 - 0-12^

The lead time demand is \i = 0.75 (200) = 150 units. Thus the reorder
point based on the inventory position of the system is

r* = M - s* = 150 — 5 = 145

To determine the reorder point rh based on the net inventory level, we note
that m, the largest integer less than r/T is 6. Thus

rt = m ~ - s* = 150 - 144 - 5 = 1

2-6 The Lost Sales Case

In the previous section it was assumed that all demands incurred when
the system was out of stock were backordered. We shall now examine the
lost sales case, i.e., the case where a demand which occurs when the system
is out of stock is lost forever. If demands occurring when the system
is out of stock are lost, it is no longer true that the annual revenues received
will be independent of the operating doctrine! They will depend on the
length of time for which the system is out of stock, and hence on the oper¬
ating doctrine. Thus we cannot immediately conclude that maximization
of the average annual profit will yield the same operating doctrine as the!
minimization of the average annual cost. We shall now show, howeverj
that with the proper definition of the stockout cost, the minimization of
48 DETERMINISTIC LOT SIZE MODELS CHAP. 2

the average annual cost will yield the same result as the maximization of
the average annual profit. Let S be the unit selling price of the item, <P be
the average annual profit, tt0 the cost of.a lost sale exclusive of the lost profit,
and C the unit cost of the item] Then if jo is the fraction of time during
which the system is out of stock, the average annual profit is !

(P = X(S — C)( 1 — /0) — ttoX/o — (ordering and carrying costs)

= X(8 — C) — + S — C)\f0 — (ordering and carrying costs)
It is clear that X(S — C) is the annual profit that would be obtained if the
system was never out of stock and is independent of the operating doctrine.
Thus if we write t = t0 + 8 — C, so that n is the cost of a lost sale in¬
cluding the lost profit, and if in defining the average annual cost, 7r is
taken to be the cost of a lost salefthen the minimization of the average
annual cost will yield the same operating doctrine as the maximization
of the average annual profit, (since the two expressions differ only by
X(S — C) which is independent of the operating doctrine^ This is what we
shall do, and therefore we can again proceed to determine the optimal
values of Q and r by minimizing the average annual cost, which is the sum
of the ordering, carrying, and stockout costs.
Let f be the length of time per cycle during which sales are lost. For
any procurement quantity Q, the length of a cycle is T = (Q/X) + T.
Thus the average annual cost is

:_A + ~ Xf
XT 2 Q + XT Q + XT
since on the average there are X/(Q + XT) cycles per year, the inventory
carrying cost per cycle is ICQ2/2X, and the cost of lost sales per cycle is
tXT. Geometrically, the behavior of the system can be represented as in
Fig. 2-8.
A necessary condition that f *, Q* be optimal is that they satisfy
P \ T C1 ~
— = 0 = - (Q + Xf)~* \*A + ~ Q2 + rX3f + (Q + \f)~W = 0 I
or
. XA , IC n
Xx = ~Q + Y Q (2'31)
and

4 = 0 = + x?]~2 + Y Q2 + xX2f] + ICQ[Q + xf]-1

or

-\A +yQ2- t\2T + ICQXf = 0 (2-32)

 SEC. 2-6 THE LOST SALES CASE 49

provided that
0 < T* < oo, 0 < Q* < oo

Solving (2-31) for Q, we obtain

2\A11/2
/ (2-33)
IC _

i; If (ttX)2 < 2XAIC, there is no real value of Q which satisfies (2-31). If

(ttX)2 = 2XAIC there is a unique positive Q value which satisfies (2-31).
*? When (xX)2 > 2XAIC there are two positive values of Q which satisfy (2-31)
since in this case

7rX / 7rX\2 2\A 1/2

(2-34)
IC > L\ic) “ IC .
In the event that there is no real Q satisfying (2-31), there is no f, 0 < T
< oo, which will yield a minimum of 3C; hence the optimal value of T must <
be zero or oo. The optimal value is oo since the condition (7rX)2 < 2\AIC
implies that incurring the cost of lost sales all the time is cheaper than
operating a system where lost sales are never incurred. Thus in this case
there should not be any inventory system. |
Consider now the case where either one or two positive Q values satisfy
(2-31). Substitution of (2-33) into (2-32) yields after a little manipulation

- - w t [(0 - <^5>
However, because of (2-34), it follows that T < 0 for both signs. Hence,
the optimal T again does not lie in the interval 0 < T < 00. In this case
50 DETERMINISTIC LOT SIZE MODELS CHAP. 2

the optimal value is f = 0, since (xX)2 > 2XAIC, i.e., the cost of running
the system with no lost sales is at least as cheap as that of running it with
any positive quantity of lost sales. In the special case where (xX)2 =
2XAIC, any value of f is optimal.
What we have shown above is that if the inventory system should be
operated at all, then it is never optimal to incur any stockouts. Even if
lost sales are allowed, it follows that when (xX)2 > 2XAIC, the optimal
solution is precisely the same as the optimal solution to the model studied
in Sec. 2-2. The results of this section can be seen intuitively as follows.
We rearrange the time sequence of events in Fig. 2-8 so that there are no
lost sales for a long time (i.e., in this region we have a situation like Fig.
2-1) and for a long time there is nothing but lost sales. This does not change
the average yearly cost. However, if (xX)2 > 2XAIC, and the Q in Fig. 2-8
is that given by (2-8), then costs can be reduced by stocking during the
period of lost sales and ordering in lots of size Qw.

2-7 The Case of a Finite Production Rate

In the previous sections it has been assumed that an order for Q units
will arrive in the inventory system as a lot of size Q units, i.e., all Q units
are received at the same time. We shall now consider a situation in which
the inventory system is the factory warehouse. It win be imagined that
t e item is produced in lots at the factory, and goes directly from the
actory to the factory warehouse. Once the factory is set up to produce a
lot it will be imagined that the production rate is * units per year (inde¬
pendent of the size of the lot). Itwill be supposed that demands are
deterministic and are incurred at the factory warehouse at a rate X units
per year. Both the number demanded and produced will be treated as
continuous variables. ^Clearly, the system cannot operate unless \[/ > X/ .
Let us compute the average annual variable costs when Q is the size'of
the lot produced. It will be imagined that there is a fixed setup cost A
for each lot produced. It will also be assumed that there is an inventory
carrying cost of 1C dollars per unit year where C is the unit cost of the item.
Ihe unit cost of the item will be assumed to be independent of the lot size.
irst the case will be studied where the requirement is made that all de¬
mands must be met from inventory, i.e., no backorders or lost sales are
permitted.
Durrng the periods when the item is being produced in the factory, there
will be a net rate of inflow ^ - X of units into the warehouse. During the
periods when the factory is not producing the item there is a net rate of
outflow X of units from the factory warehouse. If s is the quantity on hand
m the factory warehouse when the factory starts to turn out units, it is
 SEC 2-7 THE CASE OF A FINITE PRODUCTION RATE 51

clear from the analysis of Sec. 2-2 that the optimal value of 5 is zero. It
remains to determine Q, the lot size, in such a way that the costs are
minimized.
The situation is illustrated geometrically in Fig. 2-9. The length of time
required to produce a lot is Tp = Q/ip. The on hand inventory in the
factory warehouse reaches its maximum value just as production is cut off
at the factory. The maximum on hand inventory is

- X) = Q ^1 - ^ (2-36)

The time required to deplete the on hand inventory at the warehouse is

-!(>-;) (2-37)
The length of a cycle is then T = Td + Tp = Q/\. •
The average annual costs of setup and holding inventory are

+ (MS)

since the inventory carrying costs per cycle are

10 (l
(2-39)
■reK(1-f)+£(1'
 52 DETERMINISTIC LOT SIZE MODELS CHAP. 2

*f Q*>° <Q* <°°, minimizes the K of (2-38), then it is necessary that

Q* satisfy

^ = 0 = 2 A -r 0 (2-40)
dQ U Q H)
which has the imique positive solution

n* _ f2 XA $ -|V2 r ^ “]i/2

The Q* of (2-41) does yield the absolute minimum value of K for Q in the
range 0 < Q < co.jlf & time r is required from the time the warehouse
submits an order tolhe factory until the first unit comes off the production
line then to compute the reorder point rh based on the on hand inventory
level, let m be the greatest integer less than or equal to r/T. If r — mT
< Td, the reorder point is rh = ^ — rnQ. However, if T — mT > Td the
reorder point is ’

Th
(2-42)

where i) =jfeT;We leave for Problems 2-30, 2-31, the task of sketching the
behavior of the on hand plus on order inventory and the determination of
tte reorder point in terms of the on hand plus on order inventory level.
^ t might be pointed out that it is sometimes difficult to decide what cost
C to use in computing inventory carrying charges at the factory warehouse.
The cost should be the variable production cost and should not include any
faxed production costs"]
Consider now the case where backorders are allowed. The cost of a
backorder will be assumed to have the form ir + fit where t is the length
of time for which the backorder exists. Let Q* be the optimal lot size and
s_ be the optimal value of the maximum number of backorders incurred
Rer cycle. Then it can be shown that if fi = Q s* =~lT8Foo andffs* = 0
Q is given by (2-41). When it 5* 0 ’

s* = (1- ^ (a- + ic)-1 |-tx

(2-43)
and

0* = (—+ lc-\m r_ f (2\A\ (XX)2 1

v * / U'-xyic) ic(t (2-44)
7 IC(t +10
+ IC) J

!ie t0 l1,01316?1 2-2\the detailed derivation of the equations which

have been written down above for the backorders case, as well as a more
 SEC 2-7 THE CASE OF A FINITE PRODUCTION RATE 53

detailed discussion of when the optimal s lies on the boundaries of the re¬
gion. We also leave for Problem 2-29 the proof of the fact that if demands
occurring when the system is out of stock are lost instead of backordered,
then provided that the system should be operated at all, it is never optimal
to incur any lost sales.

EXAMPLE A certain company makes a complete line of valves. These

valves are supplied directly to customers from the factory warehouse. The
valves are made in lots and the same production facilities are used to make
all the valves. One particular valve has the following properties: The
demand rate can be assumed to be known with certainty and to be constant
at 2500 units per year. The fixed cost of setup for each production run is
$50, and the unit variable cost of production is $3. The inventory carrying
charge I = 0.20. The production rate is 10,000 units per year. A period of
2 months is required from the time that a production requisition is received
at the factory until finished units begin to come off the production line. It
is desired to determine the optimal lot size and the warehouse reorder
point based on the assumption that stockouts are not permitted.
From (2-41), the optimal lot size is

A* = [2(2500X50) (10,000) T* [25 X 104 n ^T72

V L (0.20)(3) (10,000-2500) J“ l 0.60 U
_ (2-45)
= V55.6 X 102 = 745

This lot size is 15.5% greater than the lot size that would have been ob¬
tained assuming the production rate to be infinite, i.e., by use of (2-8).
The reader should answer for himself why the Q* for the finite production
rate is always greater than that for infinite production rates. Note that
this approach has ignored any holding costs on units in production. These
costs are independent of Q and hence do not appear in the variable cost
provided that the amount of time that each unit spends in production is
independent of Q. |
According to the above, the time between orders Q*/X = 0.298 years.
The on hand inventory in the warehouse reaches a maximum value of

= 745 (x ~ il^) = 559 (2_46)

in a time

m 745
= 0.0745 years = 27.2 days (2-47)
v 10,000

after the first unit comes off the production line. Since the lead time is
 54
DETERMINISTIC LOT SIZE MODELS
CHAP. 2

r 2/12 - 0.167 years, it follows that the reorder point based on the on
nana inventory is

ft = Xr = 2500 (0.167) = 417 (2-48)

2-8 Constraints

Most real world inventory systems stock many items, not merely a single
item. It is permissible to study each item individually only as long as there
are no interactions among the items. There can be many sorts of inter-

fofeTch nfw11 lt6r F°r eXa“Ple> the it6mS “ay be Partial substitutes
netirT?offlh caPacity “ay be hmited and the items are com-
peting for floor space; there may be an upper limit to the number of orders
and the items are competing for these; or there may be an upper limit on
the maximum investment m mventory and the items are competing for in-

on^heffl om-0 arS' TT "T C°nSider CaS6S where there are constraints
nlaced ^ f***’^0* 0n the number of orders per year which may be
placed, and/or on the maximum dollar investment in inventory

of wSSfolff ^ °aSe t1616 there is “ Upper limit /t0 tbe square feet
onJun? J + °0?fCer ®UPP°Se that n items are beinS stocked and that
case whirl ST T UP/5' lqUare f6et °f fl0OT space- We sbaU study the
orU T V“Smet fr0m inventoly 80 no backorders
O lost sales are allowed. If Q,- is the order quantity for item j, then if the
floor space constraint is not to be violated at any time, it must be true that
n

= fiQi + . . . -f- f„Qn < / (2-49)

^eltn8
phased tS certamty case
m the t0 so
T*that it** P0SsibiIitythat
will never be necessary0rders canthe
to have be
maximum quantity 0f each item on hand at the same time.
T be,the yearly demand rate (assumed to be deterministic), A, be
LdSl 6rm? C°Su Ci the Unit C0St (assumed to be independent of QA
L*an fcr the

2. A-i + ljCj
ilQs 2j

n is desired to find the absolute minimum of if in the region 0 < Q, < oo

j i,* * ., rij subject to the constraint (2-49). ’
+H?1 We review the mathematical background needed to solve
fc above problem. The procedure is as f„Ho™: First we ^e
 SEC 2-8 CONSTRAINTS 55

problem ignoring the constraint (2-49), i.e., we minimize over each Q,

separately. This yields
n (2XjA j . -
'X r ; 3 1 }•••, 71 (2-51)

If the Qj of (2-51) satisfy (2-49), then these Qj are optimal. In such a case
the constraint is not active, i.e., sufficient floor space is available so that
average yearly costs could not be reduced by increasing the amount of
floor space available.
On the other hand, if the Qj of (2-51) do not satisfy (2-49), then the
constraint is active and the Qs of (2-51) are not optimal. To find the opti¬
mal Qj, the Lagrange multiplier technique is used. We form the function

J = %. [& Ai + IiCi f] + 9 ( 4 IjQi ~ f) (2-52)

where the parameter 8 is a Lagrange multiplier. Then the set of Qj, j = 1,
. . . ,n, which yield the absolute minimum of K subject to (2-49) are solu-
tions to the set of equations

dQj U - Qj2 Ai + ~2~ + 3 = 1, • • -, n (2-53)

dJ yt
dO = ® ^ ~~ f (2-54)
i=i

These have the unique and hence optimal solution /

Qj* =
vljCj + 2d*fj i j l,...,n< (2-55)

where 6* is the value of 6 such that the Q* of (2-55) satisfy (2-54) The
function

^ fj [2 XyA j (IjCj + 26fj) x]1/2 — /

y=i

is a monotone decreasing function of 6; consequently, there is a unique

0 > 0 such that (2-54) is satisfied.
As is shown in Appendix 1, 6* = -dK*/df, where K* is the minimum ^
value of K when optimized with respect to the Qj for a given /. Thus, in an
intuitive sense, 8* gives the decrease in the minimum cost if the floor space
were increased by one square foot. The Lagrange multiplier is often called
an imputed cost or shadow price of floor space. Looked at another way, iL_)
there were no constraint on the floor space available, but if floor space cost
 56 DETERMINISTIC LOT SIZE MODELS CHAP. 2

6* dollars per square foot per year, then the average annual variable cost
would be

K = + IiCi f ] + 8* £ flQi (2-56)

The set of Q} which minimize this cost expression are precisely the same as
those which minimize (2-50) subject to (2-49). The problem described by
{2-56) which assigns a cost but no upper limit to the amount of floor space
is called a dual of the problem described by (2-50) and (2-49), which does
not charge for floor space but has an upper limit on the amount of floor
space available.
Consider next the case where there is a constraint on the total number
of orders which can be placed.) Assume that no more than h orders can be
placed per year. This requires that

(2-57)
3 = 1 't*3

Here we ignore the fact that in any given year the number of orders placed
must be an integer and can differ from the average value X,/Q, by as much
as unity. It is assumed that this influence is small. We assume now that
there is no fixed cost per order. (In Problem 2-35, the reader is asked to
generalize the theory to be developed to the case where there is also a fixed
cost associated with each order.) The only costs are then the inventory
carrying charges. Thus the average annual variable cost is

(2-58)
3=1 *

It is desired to find the absolute minimum of (2-58) subject to (2-57)

Inasmuch as only carrying costs appear in (2-58), it is immediately clear
that the constraint (2-57) must always be active, i.e., as many orders will
e processed as possible, since inventory carrying costs can be reduced in
tins way.
To determine the optimal Qj we form the function

2™e,^tof,a8raBSe multip“er' Then **“ oi*™111 0* — the

dQj 2 -qT J = (2-60)

 SEC 2-8 CONSTRAINTS 57

n \

= 0-2a
= 1 **
7

The unique optimal solution is

_ 2V%-
, j = 1,... ,n

where on substitution of (2-62) into (2-61)

-£ ataT (2-63)
J
In this case it is easy to solve explicitly for the optimal value of the La¬
grange multiplier. The value tj* can be interpreted as the imputed cost of
placing an order.
Finally consider the case where there is an upper limit D to the dollar
investment in inventory at any one time. This constraint requires that

E CjQj < D (2-64)

(here again we omit consideration of the possibility of time phasing orders

in such a way that every level would not reach its maximum value simul¬
taneously). It is desired to minimize (2-50), subject to (2-64). The con¬
straint (2-64) is formally equivalent to the floor space constraint (2-49).
Hence it is unnecessary to repeat the analysis; it is only necessary to
substitute Cj for /,• and I) for /. There are a number of minor variations on
the above three types of constraints. Some of these are pointed out in the
problems.
It is also possible to have two or more of the constraints imposed simul¬
taneously. Suppose, for example, that there is a constraint on the number
of orders placed per year and a constraint on the maximum dollar invest¬
ment in inventory at any time. Thus we wish to minimize (2-58) subject to
(2-57) and (2-64).* We know that the constraint (2-57) will be active.
However, (2-64) may or may not be active. Thus we first solve the problem
ignoring (2-64). If the optimal solution satisfies (2-64) we are finished. If
(2-64) is not satisfied, we introduce two Lagrange multipliers 6, <f> and form
the function

+- -») (2-65)

* It is conceivable that the constraints (2-57), (2-64) are inconsistent so that there
is no solution. Here we assume that they are consistent.
58 DETERMINISTIC LOT SIZE MODELS CHAP. 2

Then the optimal Q,- must be solutions to

IjCi §h,.r . . (2-66)

2 “ Q* + *'• J “ I>

0 = _A t2-r.7)
&{Qi

= 0 = J C/h - D (2-tlS)
j-i

From (2-66)

VCi(Ij + 24,*y J
<2-691

Then substitution of (2-69) into (2-67) yields

yhE AWT+ vi]'

Finally, substitution of (2-70) into (2-69) and (2-69) into (2

24iHI ytHd^

v r__xA ;E [VM -b 2**) )*■»].

f=i \_h + 2<t>* i 2 71)

The numerical procedure is as follows: 1) Determine ** from S2-711 2i

Then determine 6* from (2-70). 3) Finally determine the </; from ,-jW
he numerical value of <t>* can be determined from (2-71) using n trial and
techm9ue-However, a more efficient procedure is to use Newton's
method discussed in Appendix 2.
iD1nUmerically equating the Q* increase considerably
^ntThtfJw ^ C?8traint Whi<!h iH »» system, In the
the floorl!c A COnstr!mtS imp08(‘d °n th“ wr" » fount mint on
Lnrnmnt * investment in inventory,
cL b^LSe The0"6 <*ithw"r »"»“« '* *»»•• constraints
mL Z n T oN0mpUtafclH“al is as follows; Fir,., deter-
, the Qi from (2-8) ignoring; both constraints. If these 0 -mli-iv the

ns :i r optiT1'......!i t
oSTie otr t* cr“'“' ”>th" .....
SLl When Z ““V.. r"rt ..
2SLZTw *r ' ■',IV“ thc ““S'nliiiK M... ,l„„r ,.
but ignoring the investment eonstmiut. It Ihe (J, „,Mv tlie
 SEC. 2-9 CONSTRAINTS—AN EXAMPLE 59

investment constraint they are optimal. If they do not, we are then sure
that both constraints are active. Thus we introduce two multipliers and
solve the problem treating both constraints as being active. It is important
to note that it is necessary to examine the cases where one or both Lagrange
multipliers are zero, i.e., where one or both constraints hold as strict
inequalities.

2-9 Constraints—An Example

Consider a shop which produces and stocks three items. The manage¬
ment desires never to have an investment in inventory of more than
$14,000. The items are produced in lots. The demand rate for each item is
constant and can be assumed to be deterministic. No backorders are to
be allowed. The pertinent data for the items are given in Table 2-1. The
carrying charge on each item is I = 0.20. Determine the optimal lot size
for each item. '

TABLE 2-1
Data for Example
Item 1 2 3
Demand rate (units per year) Ay 1000 500 2000
Variable cost (dollars per unit) C3 20 100 50
Set up cost per lot (dollars) Aj 50 75 100

The optimal lot sizes in the.absence of the constraint are:

Qi = ^2(10(|K50) = 15g; & = = 61.

If these Q/s were used, the maximum investment in inventory would be

D = 20(158) + 100(61) + 200(50) = 3160 + 6100 + 10,000 = $19,260

This is greatfir_tha,n the_maximum allowable investment in inventory.
Hence the constraint is active, and on introduction of a Lagrange multiplier
p, we see by analogy with (2-55) that the optimal Q/s are given by

= ylcs(i + 2p*j’ j= lj 2’ 3
ou
DETERMINISTIC LOT SIZE MODELS CHAP. 2

where p* is the solution of the equation

tiJrffi- D
\/ + 2p*
- 14,000 - Jj X 10*'
\0.10 + P!

4- . /g-75 X IQ6 _j_ flO X 106

Thus
•Vo.io + p* ^ Volo + P*

Vo.10 + P* = ^ [1 + 1.935 + 3.16] = ^ = 0,436

14
or
P* = 0.091
Consequently, the optimal Q} are

FIGURE 2-10.
 SEC. 2-9 CONSTRAINTS—AN EXAMPLE 61

Substitution of these Q* values into the constraint shows that it does

indeed hold as a strict equality (to the accuracy with which the calculations
can be made when the Qj* are rounded to be integers).
The minimum cost of setups and holding inventory for the three items
m the absence of any constraint on investment in inventory is
3 _
K= Z = 632 + 1225 + 2000 = $3857/yr.
3 =1

The corresponding minimum cost in the presence of the constraint IS

K* = JE Ai + 1Ci = 667 + 1292 + 2105 = $4064/yr.

The cost m the presence of the constraint on inventory investment is thus

$207 per year higher than in the absence of such a constraint.
It is of interest to observe the way in which the optimal Q, and the
minimum average yearly variable cost change with D, the maximum allow¬
able investment in inventory. This is shown in Figs. 2-10, 2-11. When
D > $19,260, the optimal Q/s are simply those obtained in the absence of
any constramt. Similarly, when D > $19,260, K* = $3857 per year.

FIGURE 2-11.
 62 DETERMINISTIC LOT SIZE MODELS CHAP. 2

2-10 Periodic Review Formulation

Consider an inventory system in which an order is placed every T units of

time. It is desired to determine the optimal value of T by minimizing
the average annual variable cost, this cost being the sum of the ordering
and inventory carrying costs. It will be required that all demands be met
from inventory so that there are never any backorders or lost sales. We
assume as before that the demand rate X is known with certainty and does
not change with time. It is immediately clear that if the on hand inventory
is not to continually increase or decrease with each period, then the quantity
ordered each time will be Q = XT. Furthermore, in order to minimize
carrying charges, the on hand inventory when a procurement arrives should
be zero. If A is the fixed cost of placing an order, C the cost of one unit,
and I the inventory carrying charge, then the average annual variable cost
is
v A . Tri XT
K — Ji + IC (2-72)
and the optimal value of T is

<2-ra>
If (2-73) is multiplied by X we obtain the quantity ordered each period, i.e.,

Q* = XT* = = qw

which is precisely (2-8).

What we have shown is that for models in which the demand is deter¬
ministic, there is no difference between models which order a quantity Q
each time the inventory reaches the reorder point (these will be referred to
as (Q, r) or lot size-reorder point models) and periodic review models in
which an order is placed only at times separated by an interval T. When
uncertainty is introduced into the demand we shall later see that periodic
review models and (Q, r) models have a somewhat different structure.

2-11 Quantity Discounts—"All Units” Discounts

In the real world, it is not always true that the unit cost of an item is inde¬
pendent of the quantity procured. Often, discounts are offered for the pur¬
chase of large quantities. These discounts sometimes take the form of
price breaks of the following type: There are given quantities qo = 0, qi,
<b> ■ ■ ■ ,qm, qj < qj+i, j = 1, . . ., m and qm+i = <®, such that if a quantity
Q is purchased, q; < Q < qj+1, then the unit cost of each of the Q units is
 SEC 2-11 QUANTITY DISCOUNTS—"ALL UNITS" DISCOUNTS 63

Cj, i.e., the cost of Q units is CjQ, and Cy+i < Ct-. The total cost of purchas¬
ing Q units then has the form shown in Fig. 2-12. We shall refer to quantity
discounts of this sort as “all units” quantity discounts since the discount
applies to every unit purchased.
The introduction of “all units” quantity discounts increases the difficulty
of determining the optimal order quantity for an item. Let us now examine
how it is done. Let us consider the case where backorders are allowed, and
v = 0,7T ^ 0. For any Q value, the optimal s is given by (2-23). Hence, if
is the unit cost of the item, the average annual variable cost optimized
with respect to s is

Kj(Q) = + qA+tc ^ + ICj) 2’ J — 0, 1, ... ,m (2-74)

For q, < Q < qj+h Kj is the average annual variable cost. Note that it is
now necessary to include the cost of the units themselves in the variable
cost since the unit cost of the item depends on the quantity procured.
We can think of (2-74) as defining a cost curve for all Q, not simply for Q
in the interval q, < Q < qj+1. Jin this way one can obtain m + 1 cost
curves, one for each C/.' The important thing to notice is that these cost
curves do not intersect and that K}+1(Q) < Kj(Q) for all Q (proof?). A
typical set of these curves is shown in Fig. 2-13. The actual cost curve for
the system is represented by the solid portions of the curves shown in Fig.
2-13. The dashed portions are not physically realizable.
The problem is to determine the lowest point on the solid (broken) curve.
This can be done as follows: '.Firstjwe compute the Q value, call it Qim),
 64 DETERMINISTIC LOT SIZE MODELS CHAP. 2

which yields the minimum cost on the Km curve. If Q(m) satisfies qm < Q<m),
then Q(m> is optimal since it yields the minimum cost on the Km curve and
no cost on any of the other Kcurves, j < m, can be lower than this cost.
If Q(m) < qm, then Q("> is not physically realizable. In this case compute
Km(qm) which is the cost on the Km curve at the with price break; Km(qm)
is to be used in the next stage.
To begin the second stage of the computation (if it is needed), deter¬
mine which is the point of minimum cost on the Km^ curve. If
Qm-i < Q^~l) < qm, compute and compare it with Km{qm). If
■Km—i^ then Q^m ^ is optimal since no cost on the curves
Kj,j<m — l can be lower and no realizable cost on the Km curve is lower.
When Km(qm) < ») then qm is optimal (if Km(qm) = Hm_1(Qtm_1))
then either qm or is optimal). In the event that lies outside the
interval gw < < qm, compute Hm_1(g„_1) and = min
LKm{qm), K^iq^)]. Let gm_i be the q,- appropriate to Km-U i.e., in this
case either qm or qm~i.
If the optimal Q was not determined during the second stage, the third
stage is begun by computing », the Q value at which Hm_2 takes on its
minimum cost. If Q<"-» lies in the interval gm_2 < Q<—» < qm_, compute ■
2>) and compare it with lL_i. If Km^2(Q(m~2)) < Km^ then
Q 18 °PtimaL If K^i > Km^2(Q(m~2)) then w is optimal (if Km_,
 SEC 2-11
QUANTITY DISCOUNTS—“ALL UNITS” DISCOUNTS
65

~K^(Q(m 2)) either is optimal). When is not in the physically

TWema regl0n+COmPUte = and record
The remaining stages are merely a repetition of the above type of compu-
a ions. Unless one has some prior knowledge as to where the optimal Q
ibnvP +tSt t0 begin the comPutational Procedure as discussed
ove Clearly, if there were a large number of price breaks, it could be
quite time consuming to compute the optimal Q.

f: ^'onsider again the example presented on p. 37. Recall that

X = 600 units per yr„ A = $8.00, I = 0.20. Let us now assume that
quantity discounts of the type discussed above are available. It will be
imagined that there are two price breaks at Ql = 500 and q2 = 1000*
furthermore C0 = $0.30, C, = $0.29, C2 = $0.28. ’
To determine the optimal Q we first compute = Qv> where K =
K2 takes on its minimum value. Thus 1 - m

n(2, = E [2(600)8
V ICt \ 0.20(0.28) ~ 414

We observe that does not satisfy g® > 1000;>ence is not

physically realizable. Thus we next compute-J ~-~ *

GT/cJ
K2(q2) \C2 + ^A+IC2^ $200.80
y2 A

The second stage is begun by computing

Q(1) = 406

However, doesjfl^atisf£500 < Qm < iqqq and Q«> is not realizable -

ilence we next find “—----—. * **

•Ki(ei) = XCi + J A + ICt $ = $198.10

(±l A
Then ^ = min [K^qJ, K2(q2)] = K,(qi) = -10 and §! = 500.
For stage three

3 , °U 1 *2^ 1 c
- i'-m -400 " a/ *-
and Q(0) is allowable since 0 < Q(0) < 500. Then
JTr v.a

Ko(Qo) = ACo + V2XAIC0 ^$204

/However, Thus Q* = gi = 500 is optimal and the optimal

y value occurs at a price break. ’ -c-
66 DETERMINISTIC LOT SIZE MODELS CHAP. 2

2. By slightly modifying the nature of the discounts in the above example,

it is possible to illustrate a case where one of the is optimal. Assume
that qi = 300, §2 = 400 with Co, Ci, C2, X, A, and I being defined as in the
previous example. Now Q(2> = 414 satisfies q2 < Q(2), and hence Q«> is
optimal. The minimum average annual cost is

K2 = XC2 + V2 XAIC2 = $191.19

2-12 Incremental Quantity Discounts

Another type of quantity discount, which we shall call the incremental type
discount, charges C0 per unit for units 1, . . ., qh Ci per unit on units
Qi + 1, • • •, q2, etc. Geometrically, the total cost of Q units can then be
represented graphically as in Fig. 2-14. The total cost of Q units, C(Q),
when qj < Q < qj+1 can be written

C(Q) = Rj + Cj{Q - q,), j = 0,1,.. ., m

where Rj = C(g,-), Rq = 0, = 0, and qm+i = «. The average cost per
unit is

==^J'rCi~ C]q’ 3 = 0,1, ...,m (2-75)

Thus when no stockouts are allowed, the average annual variable cost if
Qi< Q < Qi+1 is
Kj = XCj + q (A+ R} - Cjqj) H ^ + ICj 2 — ICj ^ (2-76)

FIGURE 2-14.
 SEC 2-12
INCREMENTAL QUANTITY DISCOUNTS
67

(The corresponding cost equation for the case where backorders are allowed
becomes somewhat complicated to write down. We leave the analysis of
this case to Problem 2-66. The general results described here hold, however
w en backorders are allowed.) We can imagine K,- in (2-76) to be defined'
for all positive Q even though K, is physically realizable only if
& < y < Qj+1. If there are m price breaks we then obtain m-f- 1 curves-
as shown in Fig. 2-15. The actual total cost curve is the solid portion of
these curves.
The computation of the optimal Q for incremental discounts is somewhat

FIGURE 2-15. -

different than for "all units” diseountspfhe important thing to notice

m the present case is that the minimum of the average annual variable cost
cannot occur at one of the price breaks] To see this note first of all that the
total cost curve is continuous, i.e., K^tq,) = K}(q3),j = 1,. . ., m. Also,
the slope of K, at q, is less than the slope of at q3- (we ask the reader to
prove this in Problem 2-45). Thus the average annual cost does not have
a relative minimum at qjt and therefore the absolute minimum cannot occur
at q3. The computational procedure to determine the optimal Q is then as
follows: Compute Q<>\ the value of Q which minimizes K, for j = 0
1, . . ., m. From (2-76) we see that ?
68 DETERMINISTIC LOT SIZE MODELS CHAP. 2 i

/ go, _ [3 xtA+fi-wy (M7)

For those Q(]) which are physically realizable, i.e., gy < Q® < qj+1, deter¬
mine Kj{QW). The Q(f> corresponding to the smallest of these costs is the
optimal Q. Note that for this case it does not follow that if Q(m) is physi¬
cally realizable, then is optimal (the following example illustrates a
case where Q(m) is allowable, but not optimal).

EXAMPLE Consider an item on whiclj incremental quantity discounts §re

available. The first 100 units cost $10(5 each, and additional -units cost $98
each. For this item: X = 500 units per yr., I = 0.20, A = $50. To deter¬
mine the optimal Q value we first compute

-M-60 t
Since Qco) is in the allowable range, we determine

i£o(Qco)) « ACo + V2XAICo = $51,000

Next we compute the point where K± takes on its minimum value.

qu) = p(4 + &- Ciqjy = 113

Q(1) is also in the allowable range and

Ki(Q(1)) =^C1 + ~(A+R1~ CiaO + | IRi + \ IC1Q^

~ \ ICiqi = $51,234

Hence the optimal Q is Q* = Q® = 50. It is unnecessary to check the price

break points since they can never be optima.)

I
REFERENCES

1. Bowman, E. H., and R. B. Fetter, Analysis for Production Management,

Revised Ed. Homewood, Illinois: Richard D. Irwin, Inc., 1961.
2. Churchman, C. W., R. L. Ackoff, and E. L. Arnoff, Introduction to
Operations Research. New York: John Wiley and Sons, Inc., 1957.
Chapters 8, 9, 10 are devoted to inventory management. Chapter
8 considers the simplest lot size model with no stockouts and the case
 PROBLEMS 69

where backorders are allowed under the assumption that ir = 0.

Some simple probabilistic models are also treated in this chapter.
Chapter 9 gives a very detailed discussion of “all units” quantity
discounts and suggests essentially the same sort of computational
procedure which we have used. Chapter 10 gives an elementary
treatment of constraints and the Lagrange multiplier technique.
3. Fetter, R. B., and W. C. Dalleck, Decision Models for Inventory Manage¬
ment. Homewood, Illinois: Richard D. Irwin, Inc., 1961.
In Chapters 1, 2 the authors present the simplest lot size formula
with no stockouts and the finite production rate case with no stock¬
outs. Some probabilistic models are also treated. All units quantity
discounts are discussed; however, the computational procedure sug¬
gested for determining the optimal Q seems to be incorrect.
4. Magee, J. F., Production Planning and Inventory Control. New York:
McGraw-Hill Book Co., Inc., 1958.
5. Morris, W. T., Engineering Economy. Homewood, Illinois: Richard D.
Irwin, Inc., 1960.
6. Raymond, F. E., Quantity and Economy in Manufacture. New York:
McGraw-Hill Book Co., Inc., 1931.
7. Sasieni, M., A. Yaspan, and L. Friedman, Operations Research, Methods
and Problems. New York: John Wiley and Sons, Inc., 1959.
Chapter 4 discusses inventory models. The simplest lot size model
with no stockouts is treated as well as the backorders case for % — 0.
The finite production rate model with no stockouts is also developed.
Over half of the chapter is devoted to simple probabilistic models.
8. Welsh, W. E., Scientific Inventory Control. Greenwich, Conn.: Manage¬
ment Publishing Corp., 1956.
A very elementary treatment intended for those in industry with
no background in mathematics who are trying to apply the subject.
9. Whitin, T. M., The Theory of Inventory Management, Revised Ed.
Princeton, N.J.: Princeton University Press, 1957.

PROBLEMS

2-1. The soft goods department of a large department store sells 500 units
per month of a certain large bath towel. The unit cost of a towel to
the store is $0.50 and the cost of placing an order has been estimated
to be $2.00. The store uses an inventory carrying charge oil = 0.17.
Assuming that the demand is deterministic and continuous, and that
70 DETERMINISTIC LOT SIZE MODELS CHAP. 2

no stockouts are allowed, determine the optimal order quantity.

What is the time between the placing of orders? The procurement
lead tune for the towels is one month. What is the reorder point
based on the on hand inventory level?
2-2. Prove that for the lot size model discussed in Sec. 2-2 it is true that
the average annual inventory carrying charges are equal to the aver¬
age annual fixed procurement costs when Q = Q* i.e., show that in
Fig. 2-2 the ICQ/2 and the \A/Q curves cross at Q = Q*.
2-3. A soap manufacturer makes several different detergents using the
same equipment for all of them. It costs $1000 to clean the equip¬
ment and to prepare it for a run of a given detergent. A particular
detergent has a demand rate of 100 tons per month. This can be
considered to be deterministic. The variable production cost of one
ton is $200. The firm uses an inventory carrying charge of I = 0.17.
What is the optimal quantity to produce per run if no stockouts are
to be allowed? What is the time between runs?
2-4. Derive an economic lot size formula which expresses the optimal
order quantity m terms of its dollar value. Assume that no stockouts
are allowed. What practical advantages might there be in using this
type of lot size formula?
2-5. Derive a formula for the ratio (K - K*)/K* in terms of y when lots
oi size yQ , y> 0, are ordered rather than Q*. Assume that no
stockouts are allowed. Show that if yQ* is replaced by 1 /yQ* the
same relation is obtained.
2-6. A function/(z) is said to be convex in some interval a < x < b if for
any,two different values xh x2 in the interval, and all a, 0 < a < 1

/[oci + (1 - a)x2] < af(x/) + (1 - a)f(x2)

The function is called strictly convex if the strict inequality holds for
a < 1. Geometrically, a function is convex if the line joining
two pomts on the curve y = /(*) lies on or above the curve. Illus¬
trate a convex function graphically. Show that the sum of convex
functions is also convex. Prove that a strictly convex function can
only have a single relative minimum in any interval, and that this is
also the absolute minimum in the interval.
2-7. Stow that UQ) = XA/Qisa strictly convex function, and that/,(Q)
- /CQ/2 is convex, for 0 < Q < =o. Thus show that the function K
defined by Eq. (2-6) is strictly convex. By use of the results of Prob¬
lem 2-6 show that dK/dQ = 0 can have only one solution in the
interval 0 < Q < «, and that the solution will yield the absolute
 PROBLEMS 71

minimum value of K in this interval. Hint: Note that the proof

requires that one show that K cannot have a relative maximum.
2-8. In the real world it is often difficult to estimate precisely the values
of such parameters as A or I. Consider a case where X, C, I have been
estimated correctly, but A has not. Let Aa be the true value of the
ordering cost and A our estimate of it. Let Q be our estimate of Q*,
Q being obtained from Eq. (2-8) using A instead of Aa. Then denote
by K the average annual cost that would be incurred by using Q as
the order quantity in the actual system. Determine an equation
relating K/K* to A/Aa, where K* is the minimum average annual
cost. What useful information does this curve provide? Is it better
to have an estimate of Aa which is too high or too low? Also deter¬
mine the relation between Q/Q* and A[Aa. Derive a similar set of
results for incorrect estimates of the carrying charge.
2-9. Determine the first and second derivatives of Eq. (2-6). How does
the shape of the K curve depend on the values of A, I, C, X? For
fixed X, I, C sketch the nature of a family of K curves with A as a
parameter.

2-10. Prove in general for the case of deterministic demand and constant
lead times that the average amount on order is equal to the lead time
demand.

2-11. At a large automobile repair shop a certain part has a very low de¬
mand. The demand is 8 units per year. This can be assumed to be
deterministic and constant over time. The cost of placing an order
for this part is $1.00. The unit cost of the part is $30, and the shop
uses an inventory carrying charge of I = 0.20. Use the theory
developed in Sec. 2-4 to take account of the integrality of demand
and determine the optimal order quantity. What is the time between
placing of orders? What Q value would be obtained using Eq. (2-8)?
2-12. For the example solved on p. 37, what would be the optimal Q value
if the approach of Sec. 2-4 was used?

2-13. From the theory of the backorders case discussed in Sec. 2-5, show
that the average on hand inventory can be written iQ — s + B(Q, s)
where B(Q, s) is the average number of backorders (this average is
to be taken over all time not merely for the time over which back¬
orders exist).

2-14. From the theory of the backorders case discussed in Sec. 2-5, show
that the average annual cost of backorders arising from the M term
is simply f times the average unit years of shortage incurred per year.
 72 DETERMINISTIC LOT SIZE MODELS CHAP. 2

Furthermore, show that the average unit years of shortage per year
is numerically equal to the average number of backorders.
2-15. Introduce the reorder point r explicitly into the model of Sec. 2-5 and
eliminate the variable s.
2-16. For the material presented in Sec. 2-5 show that the optimal values
of Q and s will be independent of whether one bases the inventory
carrying costs on the quantity on hand or the quantity on hand plus
on order.
2-17. Carry out in detail the derivation of Eqs. (2-26) and (2-27).
2-18. Show that if the inequality sign is replaced by an equality in Eq.
(2-25) then any s > 0 is optimal, i.e., show that the same cost will be
obtained for any s provided that one optimizes over Q.
2-19. Consider the case studied in Sec. 2-5 where backorders are allowed.
Instead of taking the cost of a backorder to be ir -f- fit, assume in¬
stead that the cost of a backorder has the form aebt, a, b > 0, where
t is the length of time for which the backorder exists. Derive the
equations which yield the optimal values of Q, s.
2-20. For the model discussed in Sec. 2-5 determine the fraction of the time
that the system is out of stock when the optimal values of Q and s
are used. Assume that i = 0, r ^ 0.
2-21. A chemical company produces a certain organic chemical in batches.
The annual demand rate for this chemical is 100,000 pounds. The
demand can be considered to be known with certainty and the rate
does not change with time. The fixed cost of producing a batch is
$500. The variable cost of production is $2.00 per pound. There is
a backorder cost of $5.00 per pound per year, and no fixed cost of
a backorder, i.e., t = 0. Determine the optimal batch size and the
optimal number of backorders to incur.
2-22. Show that if Eqs. (2-28), (2-29) for the case of x = 0, * * 0, are
substituted into Eq. (2-17), the minimum cost is

no* _ 77"
r /rni/2
— JS-u 1 + j- - 1C\*(* + IC)]- 1/2

2-23. The management policy of a certain company is to never run out of

stock. The sales department carried out an analysis of a particular
item to evaluate this policy. The demand is deterministic and con¬
stant over time at 625 units per year. The unit cost of the item is $50
independent of the quantity ordered. The cost of placing an order is
$5.00 and the inventory carrying charge is I = 0.20. Units can be
 PROBLEMS 73

backordered at a cost of $0.20 per unit per week. Calculate the

optimal operating doctrine under the assumption that no stockouts
are allowed, and under the assumption that units can be backordered
at the cost indicated above. What is the dollar loss per year caused
by the no stockout policy if the sales department has correctly esti¬
mated the pertinent parameters?
2-24. For the situation described in the previous problem, the company
president insisted that an additional backorder cost of $0.50 per unit
be included in the cost expression. What effect does this have on Q*
and on X*?

2-25. Consider an item with the characteristics listed below. Solve for the
optimal values of Q and s. Compute the average annual variable cost
using Eq. (2-17) and also the equation of Problem 2-22.

X = 900 units/yr., I = 0.20, C = $90

A = $2.00, 7r == 0, it == $16 per unit per year

2-26. Solve the preceding problem for the case where w = oo. Compute
the average annual variable cost and compare with the results of the
previous problem.
2-27. The inventory turnover rate is a variable frequently referred to,
especially in retailing. The turnover rate is defined to be the annual
demand divided by the average inventory. Give an intuitive inter¬
pretation of the turnover rate. Compute the optimal turnover rate
for the case where no stockouts are allowed and the case where
backorders are permitted. Hint: What is the average inventory in
the backorders case?
2-28. Derive in detail Eqs. (2-43) and (2-44). When will s* = 0 if # ^ 0?
2-29. Examine the situation in which the production rate is finite and lost
sales are allowed. Show that, as in the case where the production
rate is infinite, it is never optimal to have any lost sales.
2-30. Let r be the production lead time for the finite production rate case.
Obtain the reorder points, both for the no stockout case and the back¬
orders case, in terms of the various inventory levels of interest.
2-31. Sketch on the same graph for the case of a finite production rate and
no stockouts the behavior of the on hand inventory and the quantity
on hand plus on order.
2-32. Re-solve Problem 2-3 assuming that the production rate of detergent
is 400 tons per month.
74 DETERMINISTIC LOT SIZE MODELS CHAP. 2

2-33. The J electronics company produces a certain type of magnetic core

in lots. These cores can be produced at a rate of 1600 cores per day.
The daily demand for cores of this type is 250 per day. The cost of
setting up for a production run is $700. The inventory carrying
charge is I = 0.30 and the unit variable cost of a core is $0.10. Under
the assumption that no stockouts are allowed, calculate the optimal
size of a production run. How long will a production run take?
What will be the time between runs?
2-34. In the preceding problem, assume that cores can be backordered and
that x = $0.01 and x = $0.04 per unit year. Now determine the
optimal values of Q and s.
2-35. An inventory system stocks n items. The cost of placing an order for
items of typej is Aj. The carrying charge I is the same for all items.
There is also a constraint on the number of orders which can be pro¬
cessed, the maximum allowable number of orders per year being h.
Derive formulas for the optimal order quantity for each item such
that the average annual variable cost of procurement and holding
inventories is minimized subject to the limitation on the number of
orders which can be processed.
2-36. Show how one can treat a constraint on the number of man years
per year that are available for making setups in some shop which
makes a number of different items in lots. Derive formulas for the
optimal lot size for each item. Assume that there is also a setup cost
associated with each setup.

2-37. Derive in detail the equations which determine the optimal order
quantities for a number of different items when the items are coupled
together through a constraint on the available floor space and a
constraint on the maximum investment in inventory. Assume that
no stockouts are allowed.

2-38. Derive in detail the equations which determine the optimal order
quantities for a number of items when the items are coupled together
through a constraint on the number of orders, a constraint on the
available floor space, and a constraint on the maximum investment
in inventory. Assume that no stockouts are allowed.

2-39. Show what modifications are needed in the theory developed in the
chapter to handle a constraint on the maximum investment in inven¬
tory if the constraint is on the average investment in inventory.
2-40. A small shop produces three machined parts 1, 2, 3 in lots. The shop
has only 700 sq. ft. of storage space. The appropriate data for the
three items are presented in the following table.
 PROBLEMS 75

item ® © ©
X (units/yr.) 5000 2000 10,000
A (dollars) 100 200 75
C (dollars) 10 15 5
fi (sq. ft./unit) 0.70 0.80 0.40

The shop uses an inventory carrying charge of I = 0.20. If no stock¬

outs are allowed, determine the optimal lot size for each item.
2-41. For models of the type discussed in this chapter, is it possible to place
a constraint on the total dollars spent on purchasing items for
inventory? Discuss the circumstances under which this type of
constraint can be implemented.
2-42. The E. E. Automotive Parts Co. produces crankshafts in lots. The
company has decided to use an economical lot size formula to mini¬
mize the sum of setup costs and inventory carrying charges. Eight
thousand crankshafts are demanded every year. It is estimated that
each setup costs $245 (based on standard hours). Inventory carrying
charges are $2.00 per unit per year. Assume that requirements for
setups on other parts have limited the number of setups per year on
crankshafts to four. What is the optimal lot quantity to produce?
If the setup crew has time available for ten setups, what is the opti¬
mal lot size?
2-43. The A. C. Trash Co. stocks containers, each of which occupies four
square feet of rack space. The available rack space is limited to 600
square feet. The appropriate data on these containers are as follows:
X = 2000 units per year, A = $5.00, C = $2.00, I = 0.25. How
much would the first additional foot of storage space be worth? How
much would 100 additional square feet of rack space be worth?
A purchasing agent is concerned with the procurement of three types

item ® © ©

annual demand 1000 3000 2000

ordering cost (dollars) 10 10 10

unit cost (dollars) 30 10 20

space requirement | ) 5 10 8
\ umt J
 76 DETERMINISTIC LOT SIZE MODELS CHAP. 2

of items. The comptroller of the company has placed a $2400 con¬

straint on the value of average inventory, and there is a 2500 square
foot limitation on storage space. The data for the items are given
in the preceding table. The inventory carrying charge is I = 0.25.
Determine the optimal purchase quantity for each of the items.
2-45. Prove that when incremental quantity discounts are given, the opti¬
mal Q value can never occur at one of the price breaks. Do this by
evaluating the derivative of K, and Kj+1 at qj+1 and showing that the
derivative of Kj+1 is less than the derivative of Kj.

2-46. The purchasing agent for the G. W. Dog Food Company can buy
horsemeat from one source for $0.06 per pound for the first 1000
pounds and $0,058 per pound for each additional pound. The
company requires 50,000 pounds per year. The cost of placing an
order is $1.00. The inventory carrying charge is I = 0.25. Compute
the optimal purchase quantity.
2-47. Derive in detail Eq. (2-74).

2-48. A supplier of made to order metal fixtures quotes the cost of Q units
as 400 + 25Q dollars. The purchaser of these fixtures requires 2000
units per year. He uses them at a uniform rate. His direct costs for
making out purchase orders, receiving, inspection, and other costs
that are incurred on each shipment amount to $20 per order. The
carrying charge is 7 = 0.20. The time elapsing between the placing
of an order and the delivery of the units to the stockroom is three
months. Calculate the economic purchase quantity and the reorder
point (based on the on hand inventory level).

2-49. The XYZ Corporation has a special petty cash fund it uses for a
known steady state flow of contributions to charity. The corporation
periodically replenishes this fund by withdrawing cash from its
savings bank account (which pays 4 percent interest on deposits).
The fund is never allowed to run out. A bank messenger is paid $5 00
per delivery of cash from the bank. If the optimal withdrawal sched¬
ule involves nine withdrawals per year, what is the dollar volume of
annual contributions to charity?

2-50. The fee which a particular brokerage firm charges its clients is 20
+ 0.01 V dollars, where v is the market value of the stocks purchased
or sold on any given day. A client of the firm spends $10,000 annually
over and above the 8 percent interest he receives from his invest¬
ments. Assuming that the client accumulates the funds for invest¬
ment at a uniform rate, how much should he accumulate before
investing, i.e., what is the optimal value of v?
 PROBLEMS
77

2-51. A firm is committed to a capital expansion program that requires a

rowed at0a0W1I1lS P '^S?00 P6r year' °n any &ven quantity bor-

Sion nnn « ^g e e’ 7 lnterest rate is 4 percent for the first
$100,000 and 5 percent for any additional amounts. The costs of
negotiating eaeh loan are 1500. The loans will not be r^aM unffl

2’62‘ wlSXViT?r°blem Then the bant Charges 4 Percent the

stme W “ Pment additi0nal arnounts Arrowed at the

2’53' oSrXo m?'kSt Pr°blem’ the int6reSt charSed on amounts

over $50,000 is 6f percent. Solve the revised problem.

2"64’ 7 Pd+iCy ? °rdfmg tW0 Particular items at the same

time. The characteristics of the items are: item (1)- X = 5408
units per year, C = $10, A = $3.00, I = 0.30: item (2)- X = 845
units per year, C = $1.00, A = $3.00, J = 0.30. Maimgement’s
of a wei° ar7 0rders can be Pla«ed only at the start
of a week Calculate the best policy for ordering the two items if
7b mU,st be°rdered at the same time. What are the average an-
nh+a - ?l°f tblS, P0|lcy? Calculate what additional savings can be
itemsned ^ eStabllshlng a dlfferent ordering policy for each of the

2-55. A purchasing agent for consumable aircraft spare parts is to decide

between three sources. Source A will sell a particular component for
$10 each regardless of the quantity ordered. Source B will not accept
order for less than 600 umts, but sells them for $9.50 each if an
order for 600 or more units is placed. Source C will not accept an
order for less than 800 units, but charges $9.00 each if an order for
800 or more units is placed. Annual demand for the component is
2500 units and I = 0.25. A fixed cost of $300 is incurred each time
that an onier is placed. Which source should the purchasing agent
select. What quantity should he purchase? What annual costs will

components? ^ inventory’ and Purchasing the

2-56. Suppose in thn preceding problem that source B requires a purchase

of at least 3000 units and source C requires a purchase of at least
5500 units. Also change the ordering cost to $100 per order. Re-solve
the problem.

2-57. An item in inventory is sold at a unit price of p, the rate of sales being
X umts per year. If a quantity Q is ordered the unit cost is 5 + (a/Q)
The cost of placing an order is A and the inventory carrying charge
78 DETERMINISTIC LOT SIZE MODELS CHAP. 2

is I. What procurement quantity will maximize the average annual

profits if it is assumed that no stockouts are allowed?
2-58. Consider an item with an annual demand of X units per year. If a
quantity Q < of the item is ordered, the unit cost is g0e“aQ and if a
quantity Q > is ordered, the unit cost is b = g0e_affi. The fixed
cost of placing an order is A and the inventory carrying charge is I.
For the case where no stockouts are allowed, show how to determine
the optimal order quantity Q.
2-59. Consider the curves defined by Eq. (2-74). Imagine that X, A, I are
fixed, but that Cy is a variable parameter. For a given Cj let Q* be
the value of Q which minimizes K. Let the minimum of K be K*.
Determine the curve which gives the locus of the points (Q* K*) as
Cj varies from zero to infinity.
2-60. Derive formulas which yield the optimal values of Q and $ in the
case where backorders are allowed when account is taken of the
integrality of demand of Q and of $. Assume that x = 0 and x ^ 0.
Follow the same sort of procedure that is used in Sec. 2-4.
2-61. Give an intuitive explanation as to why in the backorders case when
x = 0, x ^ 0, it must always be true that s* = 0 or oo. Hint: Use
a modification of the type of argument introduced in Sec. 2-6.
2-62. In Eq. (2-26) show that if s* = 0 then xX = Kw, Also show that in
Eq. (2-27), if we set xX = Kw, then Q* = Qw.
2-63. Show that
dQ^ldA dK_sldA
Q* 2 A* K* 2 A

for Eqs. (2-8), (2-11) respectively. What does it say about the per¬
cent change in Q* for small percentage changes in A. For the ex¬
ample on p. 37 determine the approximate change in Q* if A in¬
creases to $8.50. What is the approximate change in X*? What are
the exact changes in Q* and A*? Derive formulas similar to the
above for changes in X, C, and /.
2-64. An inventory system stocks n items. It has been found necessary to
review and place orders for all items at the same time. Let T be the
time between reviews, A the cost of review and of placing orders for
all n items, I the inventory carrying charge, Xy the annual rate of
demand for item j and Cy its unit cost. Determine a formula for the
optimal value of T by minimizing the average annual variable costs
of review and holding inventory for all n items under the assumption
that no stockouts are allowed. Find the optimal T for a system which
carries items for which the appropriate parameters are: A = $75,
 PROBLEMS 79

1 Xl = 5°0, Xs = 300, Xs = 1000> Cl = ^2-00> Ci = 525-00,

C3 — $10.00. Compute the average annual cost and the order
quantity for each item. Suppose that it was possible to use a (Q, r)
model on each item separately with the ordering costs being a[ =
$15, A2 = $40, A3 = $20. Determine the optimal Q values and the
average annual variable cost for the three items. How much can be
saved by using a (Q, r) system for each item?
2-65. Except for Sec. 2-7 it has always been assumed in this chapter that
orders were never split, i.e., the entire Q units ordered arrived in a
single batch. Now suppose that each order is always delivered in
two parts, a fraction/x being delivered at a time n after the order is
placed and a fraction 1 — /i being delivered at a time r2 after the
order is placed. Derive formulas for the optimal Q, and reorder
point for this case under the assumption that backorders are not
allowed. How does one make sure that stockouts do not occur?
2-66. Study the problem of determining the optimal Q in the presence of
incremental quantity discounts when backorders are allowed. Show
that the results of Sec. 2-12 hold in this more general case.
2-67. From Eq. 2-11, what can we deduce about economies of scale in the
cost of operating an inventory system, i.e., all other things being
equal, would there be an advantage to having a large system with a
relatively high rate of demand rather than a smaller system with a
relatively low rate of demand?

2-68. Consider a firm which has several different warehouses, and let the
demand rate at warehouse i be A,-. It is possible to operate this
collection of warehouses on a decentralized basis, by having each
warehouse order for its own needs using its own order quantity and
reorder point, or the system can operate on a centralized basis, with
a system reorder point and an order quantity for the entire system
which is allocated in the appropriate amounts. Prove that when
the unit cost of the item is constant, the carrying charge is the
same at each warehouse, the ordering cost and lead time are inde¬
pendent of whether the system is operated on centralized or de¬
centralized basis, and when transportation costs from the source to
the warehouses do not depend on the mode of operation, then the
centralized mode of operation never leads to higher costs than the
decentralized mode of operation. Would the conclusion be altered if
quantity discounts were offered? Assume in the analysis that stock-
outs are not permitted.

2-69. Consider the simplest deterministic (Q, r) model treated in this

chapter, i.e., the one in which no stockouts are allowed and the unit
80 DETERMINISTIC IOT SIZE MODELS CHAP. 2

cost is a constant independent of the quantity ordered. Suppose that

instead of determining Q* and r* by minimizing the average annual
cost, we attempt to determine these quantities by minim^ng the
discounted cost over all future time. Different answers for the dis¬
counted cost can he obtained depending on what point in a cycle one
selects as the time origin. For simplicity, select the time origin as a
point just prior to the arrival of an order so that nothing is on hand
at the time origin. Imagine now that continuous discounting is used
(the reader not familiar with this will find it described on p. 377) so
that if i is the annual interest rate, the present worth of a cost H
incurred at time t in the future is Her'1. Consider now any cycle, the
beginning of the cycle being taken as the point when an order arrives.
The cost of placing an order and of the Q units ordered, discounted
to the beginning of the cycle is (A + CQ)e~^ where tr is the time
from the arrival of an order until the reorder point is reached. In
computing the discounted cost one must include the cost of the units
even though C is constant. Why? The carrying cost discounted to
the beginning of the cycle is
rT
IoC / (Q - \t)e~u dt
Jo
When introducing discounting, then, in general, the rate of return
need not be included in the carrying cost since it is expressed in i. We
have denoted by I0 the carrying charge exclusive of the rate of return.
(The reader should study this point in detail to see why the rate of
return should not be included.) Let H be the total cost per cycle
discounted to the beginning of the cycle, i.e., discounted to the time
when an order arrives. Then the discounted costs over all future time
are

d = H e~inT
n—0
Determine the value of Q which minimizes 3. Expand the equation
which determines Q* in powers of i and show that to a first approxi-
mation

n* = / ~~ __
W \(J0 + i)C~Qv>

How does the frequency with which orders are placed affect the good¬
ness of the above approximation? As a specific example consider the
case where A = $15.00, C = $20.00, J. = 0.10, i = 0.10, X = 1000
units per year. Determine the exact value of Q which minimizes 3
and the value Qw. Pick another point other than the point when an
 problems
81

(‘.U .0 UM,‘K <‘<m(imumi‘s <h.seountinK, the yearly interest rate is i

and it is compounded once nor ovelo i c in,, *■

lot" 5?*?+7Z :lv*j**

fr<,m the ,1 obtained usinK continuous di^mnttttWh^kQ* not?

r* r:!.. ^-ssesesk
, 1 , an ord< r- U hieh 'vas used in the above? Hint- Tt ran
difficult, to wilve explicitly the equation db/dQ » 0 for ^'lI oZit
mg the limiting form, determine the coefficients of A C hC and

r ('.«ml “ »
r rcr r? “s •***£ &
'■ "my '* w"r,“ "““'y *» ‘h»t
PROBABILITY THEORY

AND STOCHASTIC PROCESSES

“Lest men suspect your tale untrue,

Keep probability in view."

John Gay
 3-1 Introduction

We have noted previously that the demands for units stocked by an inven-
ory system can seldom be predicted with certainty. Instead, they must be
described in probabilistic terms. Realistic inventory models must account
for this uncertainty in demand. Often, it is also true that procurement
lead times must be described probabilistically. It is the purpose of this
chapter to present the background material from the theory of probability
and stochastic processes that will be needed in the future chapters. It will
e assumed, however, that the reader has previously had at least a brief
introduction to the theory of probability.

3-2 Basic Laws of Probabilities

It will be recalled that the foundations of probability can be developed

from several pomts of view, such as: (1) the a priori method, (2) the fre¬
quency approach, (3) the degree of rational belief, (4) the axiomatic
approach. In practical application one will frequently use several defini¬
tions of probability for example, the a priori and frequency definitions—
simultaneously in the solution of some problem. Here we shall not attempt
to deve op the theory from any one of the above points of view. Instead
we shall sunpiy set down the basic laws for operating with probabilities,
which should be familiar to the reader, and explain their operational
meaning.
Consider an experiment whose outcome can lead to one or more of the
following finite number of events, which will be denoted symbolically by
l’ Imaguie that the outcome of the experiment is not deter-
mimstic, but can only be described probabilistically. Let p(AA be the
probability that As occurs. We shall use the notation At D A,- i ^ j to
mean that both At and A,- occur, the notation A, U As, i ^ j, to mean that
™ i10Lb \°CCUr,S’ End the Symbo1 Ai t0 mean that Ai does not occur,
inen the three basic laws of probability are:
(1) 0 < V{A,) < 1
(3-1)
83
84 PROBABILITY THEORY AND STOCHASTIC PROCESSES CHAP. 3

(2) p(Ai U Aj) = p(Ai) + p(A}) — p(Ai O Aj), i ^j (3-2)

(3) p(Ai H Ai) = p(Ai\Aj)p(A3) = p(Ai\Ai)p(A%)} i

P(Aj), p(Ai) ^ 0
n A3) = 0 if p(A,) = 0 or p(A{) = 0 (3-3)
Equation (3-1) simply indicates that p{A3) is a non-negative number
which is not greater than unity. If Aj does not occur, then p(Aj) = 0, and
if Aj always occurs, then p(Aj) = 1. The second law states that the
probability of Ai or Aj occurring is equal to the probability that Ai occurs
plus the probability that Aj occurs, minus the probability that Ai and Aj
occur. Two events, Ai} Aj, i ^ j, are called mutually exclusive if they
cannot occur together, i.e., Ai and Aj cannot occur; then p(Aj O Aj) = 0.
For mutually exclusive events (3-2) reduces to the simple addition rule

p(Ai U Aj) = p(Ai) + p(Aj)t i 7* j (3-4)

By the definition of Aj it follows that Aj and Aj are mutually exclusive.
Furthermore, either Aj or Aj must occur. Hence

p{Aj U Aj) = 1 = p(Aj) + p(Aj)

More generally, suppose that the finite set of events Ai, . . . , Ar are mutu¬
ally exclusive. This means that no combinations of two or more of these
events can occur simultaneously. Then

p(Ai U i-2 U . . • U Ar) = p(Ai) + p{A%) + . . . + p(Ar) (3-5)

To prove this we introduce the new event A2* = A2 U Az U • * • U Ar
and note that p(Ai O ^-2*) == 0 since Ai cannot occur simultaneously with
any one of the events A2, . . . , Ar, i.e., with A2*. Thus

p(Ax U A2 U .. . U Ar) = p(Ai U A2*) = p(Ai) + p(A2*)

The same argument is now repeated with p(A2*), etc. until (3-5) is obtained.
The third law (3-3) serves to define the conditional probabilities p(Ai\A3)
and p(Aj\Ai). These conditional probabilities have the following intuitive
meaning: The quantity p(Ai\A3) is the probability that Ai has occurred
(or will occur) if we know that event Aj has occurred. In practical ap¬
plications, this intuitive explanation of the meaning of p(Ai\Aj) often
allows us to compute this probability directly without the use of (3-3).
In the event that
V(Ai\Aj) = j>(At)
so that the conditional probability is independent of A,-, then the events
Ai, A j are said to be independent and (3-3) reduces to
?(Ai pi Ay) = p(Ai)p(Aj), i j (3-6)
 SEC 3-2 BASIC LAWS OF PROBABILITIES 85

Intuitively, Ait Aj are independent if the occurrence of A{has no influence

on whether A, will occur or not, and vice versa.

EXAMPLE If the foundations of probability are developed from the a priori

approach, then it is possible to derive the laws (3-1) through (3-3). The
a priori approach assumes that the results of the experiment under con¬
sideration can be described by a finite number n of equally likely and mutu¬
ally exclusive outcomes. If n{ of these outcomes have the property A{
then the probability of the event A{ is defined to be

P(A,.) = *

Since n > 0, n{ > 0 and »< < n, (3-1) follows immediately from this
definition. If m, outcomes have the properties At and Ah then the number
of outcomes which have the property At or A, is m + % - nij and hence

p(Ai U A,) = ^ ^ ^ = p(Ai) + p(Ai) - p(Ai n Aj)

which is (3-2). Finally, note

'fl'ij 'fl'ij Wtf

p(Ai n a,) PiMAiMAj), % ^ 0
n Uj n

where n{j/nj is the probability of At given that A, has occurred, i.e.,

Inj = p(Ai\Aj)

This is (3-3). Thus the three basic laws have been derived in this case.
Often it is convenient to imagine that the n outcomes can be represented
symbolically by points in a space. The resulting space is called the sample
space. If we think of the outcomes as being represented by points in a
plane, then it is possible to draw diagrams such as Fig. 3-1. The sample
space will be assumed to consist of the points (represented by crosses)
inside the rectangle. The set of outcomes with property At is the set of all
points inside the crosshatched area marked A{. Similarly, the set of out¬
comes with property A} is the set of all points inside the crosshatched re¬
gion marked As. The set of points with the properties A> and As are those
in the doubly ruled area, i.e., these points represent the set of points with
the property At D Aj. The set of points with the property At (j Aj is
the set of all points in the shaded areas. Fig. 3-1 makes it especially easy
to derive the formula (3-2). If Ai} As are mutually exclusive, then in Fig.
3-1, the areas representing A{ and As do not intersect.
86 PROBABILITY THEORY AND STOCHASTIC PROCESSES CHAP. 3

FIGURE 3-1.

3-3 Discrete Random Variables

Many of the events which will be of interest to us in this book will be

capable of being represented by non-negative integers. For example, the
event might be the number of units demanded in one month. This event
can be represented by a non-negative integer x which gives the number of
units demanded. Consider an experiment the totality of whose mutually
exclusive outcomes can be described by the n + 1 non-negative integers
0, 1, 2,. . ., n. Let p(x) be the probability that the outcome of the experi¬
ment can be described by the non-negative integer x. Then since one of the
integers 0, 1, . . ., n must describe the outcome, and since no more than one
can describe the outcome (because the events are mutually exclusive), we
must have

p(0 UlU2U.«-Uw) = l = p( 0) + p(l) + . . . + p(ri)

n
= T ?(*) (3-7)
x—0

If the experiment is performed many times, then the fraction of the time
that the outcome can be described by the integer x will approach p(x).
The outcome of the experiment can then be thought of as being described
by the variable x, &n.dp(x) can be looked upon as a function of x which gives
the probability of x. The variable x which can only take on the non¬
negative integral values 0, 1, . . ., n is called a random variable since it
describes the outcome of an experiment, the outcome of which is deter-
 SEC 3-3
DISCRETE RANDOM VARIABLES
87

d bf h?- aT °f fWe' The function ?(*) is called the probability

0 i6 rand°m variable x> PC*) must satisfy (3-7) and
nnTsiW “* L N°te the above v('x) must be defined for every
LJiy ZZf: <>ntC°meS ml,St 16 defined 80 tllat *»y are
Consider now the probability p(0 U 1 U . . . U r), r < n. This is the
probabxhty that« takes on the value 0 or 1 or ... or r, the probability

X byX)°and th“° 01 e,Ual *° r' We Sha11 den0te 11118 proba-

J>M - P(0 U 1 U • . . U ,) - p(0) + ?(!) + ...+ p(r) M

Note that r cao take on the values 0,1.», the same values which a
n assume.. Thus we can define a new function P(x), which will be called
e—^ Probafiility of x or the cumulative function for x- it is the
i’mftoJ r'T™ experimerit -™lds a value less than or
£ r °n W*“Cl‘ ^ gently be of use to us is some-
times called the complementary cumulative function. It is defined by

P(x) = 1 - p(x-l) = p(x) + p(x + 1) + . . . + p(n) (3_9)

P(x) is the probability that the outcome of the experiment will yield a
value greater than or equal to *. We shall find it convenient to use the

cZStoctST f”ncti<’“ m“ch ,nore treque,i*ly th“ ““

The random variable introduced above, which could only assume one of
+ integral values, will be called a discrete random variable. More gen-
erally any random variable that can only take on integral values will be
ca led discrete The probability law that describes a discrete random vari¬
able is uniquely defined by specifying either the density function, or the
a *yen unctl0n) or the complementary cumulative function The
probability law is also referred to as the distribution of the random variable.
The distribution of a discrete random variable is called a discrete distribu-

An important special case of the type of probability density function

which we have been discussing above is called the binomial density. One
way this distribution arises is in the study of repeated independent “trials”
which can have only two possible outcomes. These two outcomes can
without loss of generality be referred to as success or failure. Let p be the
probability that a trial will be a success and 1 - P the probability that it
will be a failure. Let us imagine that we perform n trials in sequence and
we desire to determine the probability that the first z trials 0 <x<n
are successes and the last n-x are failures. Since the trials are inde¬
pendent, the desired probability is simply the product of the probabilities
 88 PROBABILITY THEORY AND STOCHASTIC PROCESSES CHAP. 3

for each event, i.e., p*(l — p)n~x. This same probability is obtained if we
ask for the probability of x successes and n — x failures in any specified
order.
Next, let us ask what is the probability of having x successes in n trials
without regard to the order in which the successes are obtained. Note that,
events corresponding to obtaining x successes in different orders are mutu¬
ally exclusive. Hence, the probability of having x successes without regard
to the order in which they are obtained is simply the sum of the proba¬
bilities for each possible order. The probability for x successes in any
given order is px( 1 — p)n~x. Thus, all that remains is to compute the
number of possible orders in which x successes can occur. This is simply the
number of ways in which n objects (the trials) can be separated into t wo
groups (without regard to order within a group) such that x objects are in
one group and n - x in the other. Of course, this is the number of combi¬
nations of n things taken x at a time, i.e., n!/x!(n - x)!. Sometimes we

shall use the simplified notation (n) to represent, a! x!(«. ■ x)!. We con¬

clude, therefore, that the probability of x successes in n trials, without

regard to order, is

p(x) = b(x;n,p)

We can think of x, the number of successes, as being a random variable

which can take on the values 0, 1, . . . , a and b(x; n, p) as being tin- density
function for this random variable. We note that fctx; n,,,) is a legitimate
density function since b(x; n, p) > 0 and

50- [p + (i - p) I" i

by the binomial expansion. Note that, b(xis the xth term in tin*
binomial expansion of [p + (1 - p)]» We shall always use the notation
b(x; n, p) for the binomial density function. If the random variable x has
the density function b(x; n, p), then x is said to have a binomia l dist ribut ion.
Note: We shall use frequently the binomial expansion and hence it might
be helpful to review briefly its derivation. The binomial expansion of
(x + b)* for any real number n is a special case of the Maelaurin expansion

/(*) = E h/o,(0)xt

when f(x) = (x + &)». In the above expansion/<;>(()) Ls the jth derivative

of / evaluated at x = 0. When / (x) * (x +

/O)(0) - n(n — 1) . . . (n — j -)- j > l-

/“"(()) /(())
 SEC 3-3
DISCRETE RANDOM VARIABLES
89

and Hence

m = 6- + £ _V * • • V/* .7 T I) _
Jf-2 £3?>n“>
y=i

By ‘he r,ti° test> the **» converges if

]/ ' 5 ' 1 the event that n is a positive integer, /«(0) =0 j > w and
the infinite senes becomes a finite sum of n + 1 terms which can be written

(x
+ ~ (j) xihn n a positive integer.

When « is a negative integer, we can write

n(n - 1) . . . (» - j + 1) = (-ij/dnl +j - 1)(|ft, +j _ 2) . . . ,n|

= ( ^/(M+j-l)!
^ ; (Ini-11!
(M - 1)!
and so

(x

n a positive integer, and \z/b\ < 1.

EXAMPLES 1 A given base has n missiles. These missiles are inspected

during^™tr T? T.1116 ?ro,bability tha* “y single missile mil not faU
d ?°th; If a11 n misslles are operating at the beginning of a given

Tg" t fcend rfthfrontt “““ * S,ffl be ^

2‘ t m is *°fed n times- The Probability of a head is p. Then the

probability of obtaining * heads is given by (3-10). To check (3-10) in a
e y simple caseimagine that a coin is tossed three times and we wish to

HTH rim T’ h +u W° headS Can be obtained in *be following ways HHT,

.2/1 J. ''e,’ ,t,hree The Probability of each of these orders is
P ( p), so that the probability of two heads is 3p2(l — p). Now

3!
= 3
211!

so that (3-10) also gives 3p2(l — p).

In the above paragraphs we have discussed random variables which are

defined over a finite set of non-negative integers. We shall also have use
for random variables which are defined over all the non-negative integers
90 PROBABILITY THEORY AND STOCHASTIC PROCESSES CHAP. 3

so that a p(x) can be defined for every non-negative integer x. As before,

we assume that the outcome of the experiment will yield one and only one
non-negative integer x. Thus the various x values are mutually exclusive
and we must have
00

p(*) = l> P(x) > 0 (3-12)

X—0

Although for the case of finite n we could prove (3-7) from the fundamental
laws, we cannot prove (3-12) directly. Instead we take (3-12) as a postu¬
late.
A particularly interesting density function which is defined for all non¬
negative x is called the Poisson density. It is defined by

p(x) = p(x; m) = e-», x = 0, 1, 2, . . . (3-13)

where p is a constant whose physical interpretation will be made clear later.

We shall make frequent use of the Poisson density or distribution in future
chapters. The Poisson density will always be denoted by p(x; p). The
random variable x whose probabilities are given by (3-13) is said to be
Poisson distributed. The cumulative and complementary cumulative
Poisson functions will be denoted by P(x; p) and P(x; p) respectively,
where

?(x'’ J») = J fi p(.x> m) = fj e~» (3-14)

i—0 J ' y=j J*
Observe that

p(x‘,n) = e~»

e*
n\
o-?1

Another distribution of a random variable defined over all the non¬

negative integers, which we shall use occasionally, is called the negative
binomial distribution.
The density function for the negative binomial distribution of order n is

p(x) = bN(x-, n, p) = (* + n_ x ^ P.(1 _ p).

0 < p < 1, * = 0, 1, 2, . . . (3-15)

where n is a positive integer and p is a constant. We shall always use

 SEC. 3-4
CONTINUOUS RANDOM VARIABLES
91

btfix] n, p) to represent the negative binomial density

From the binomial
expansion of [1 - (1 - p)]-» = p- we obtain

’-set-T1)*1-*
Multiplication by P» shows that the sum of 6v(x; n, p) from * = 0 to « is
unity as desired.
A random variable having a negative binomial density can arise in a
variety of ways. One simple probability problem which gives rise to such
a distribution is as follows: Suppose that we toss a coin which has a
probability P of getting a head until we obtain n heads. We then ask:
„tb? Probablb1f that exactly * + n tosses are required to obtain
n heads. This means that n — 1 heads are obtained in the first x + n - 1
tosses and a head is obtained on the (x + n)th toss. The probabUity of
bilitv nf l mA + * t0SS6S is6(n“1^ + n-l.P) and the proba-
bihty o a head on the last toss is p. Hence, the desired probability is
pb(n - 1; x + n - 1, p), which is (3-15), and the random variable x (not n)
nas a negative binomial distribution.
When n = 1 in (3-15) we obtain

p(x) - bm{x) 1, p) = P(1 — p)* x = 0, 1, 2,. . . (3-16)

This density function is assigned a special name—it is called the geometric

density function, and if the probability of z is given by (3-16), x is said to
have a geometric distribution.
We have noted previously that the probabUity of an impossible event is
zero. Conversely, if x is a random variable defined over a finite set of non-
negative integers, or over aU the non-negative integers, the expression
PW - 0 for a particular x is usuaUy interpreted in practice to mean that
the integer x will never occur. Sometimes, however, since it is often neces¬
sary to estimate probabilities from historical or experimental data the
expression p{x) = 0 may be interpreted to mean that x can occur' but
only very rarely.

3-4 Continuous Random Variables

In the previous section we discussed random variables which were defined

only for a finite number of non-negative integers or for all non-negative
integers. Now we would like to consider variables which can take on any
value between 0 and ® or-» and «>, and which cannot be predicted with
certainty, but instead can only be described in probabUistic terms. Such
variables wUl be called continuous random or stochastic variables. It is
92 PROBABILITY THEORY AND STOCHASTIC PROCESSES CHAP. 3

necessary to change slightly the probabilistic description of random

variables when one moves from the discrete to the continuous case. The
change required is similar to that needed in mechanics in moving from the
study of point masses to the study of continuous mass distributions. From
the intuitive point of view, if every x value (or even every x value in some
interval a < x < b) had a positive probability it would not be possible to
have the sum over all x values equal to unity. Therefore, the probability
of any specific value of x must be zero. The notion of a continuous random
variable is really a mathematical abstraction of reality, since nothing is
continuous in the real world, i.e., there are certain discrete building blocks
which are not infinitely divisible. Even time, as we measure it, is not in¬
finitely divisible. Thus in dealing with continuous random variables, what
we are actually interested in is the probability that x will lie in some in¬
terval a<x<b. We shall never be concerned with the probability that
*takes on a S1™11 value but rather with the probability that x lies in some
interval.
To describe the probability that a continuous random variable x lies in
some interval we introduce a function f(x) > 0, called the density function
for x, with the property that the probability that x lies in the interval
0 — x be., p(a < x < b) is given by

p(a < x < b) = i: /(*) dx (3-17)

It is important to note that f(x) is not the probability of x. Instead we can

say that the probability that * lies in the infinitesimal interval xtox + dx
is f(x) dx. We can imagine / (x) describes a curve in the plane, and the
probability that * lies in the interval a < x < b is the area between the
curve and the * axis from x = atox = b. This is illustrated in Fig. 3-2.

FIGURE 3-2.
 SEC 3-4 CONTINUOUS RANDOM VARIABLES 93

Inasmuch as x must lie in the interval —« to oo in any experiment, we must

have

The probabilistic description of a continuous random variable just in¬

troduced does require that the probability of a specific value of x be zero
since p(x = a) is found by setting b = a in (3-17), which yields a value of o’
For continuous random variables, the fact that p(x) = 0 does not imply
that the number x cannot occur, since whenever an event involving *
occurs, a specific value of x will be observed.
We shall be dealing with continuous random variables which are re¬
stricted to non-negative values, and with those which can take on any real
values. In discussing the theory, we shall usually treat the random vari¬
ables as if they can assume any real values, as we did in (3-18) and (3-19).
There is no loss of generality in doing this, for a continuous random
variable x which is restricted to be non-negative can be allowed to take on
any real values if its density function is defined to be zero for all x < 0
since this definition will insure that the probability that x lies in an interval
Xi<x< x2, Xi, x2 < 0 will be zero as desired. To convert integrals that
involve random variables which can take on any real values to a form suit¬
able for random variables that must be non-negative, the lower limit of
is replaced by 0. Any function f(x) > 0, such that (3-18) holds, can
be considered to be a legitimate probability density function. I
The function

= (3.19)

is called the cumulative probability function or the cumulative function

?! x- It gives the probability that the random variable takes on a value less
than or equal to x. The complementary cumulative function Fix), defined
by Fix) = 1 — P(x) is the probability that the random variable takes on
a value greater than or equal to x. Note that

dF(x) dF(x)
dx = /(*), dx
= -f(x)

The probability law for a continuous random variable is uniquely defined

by specifying either the density function, the cumulative function or the
complementary cumulative function. The probability law is also referred
to as the distribution of the random variable. The distribution of a con¬
tinuous random variable is called a continuous distribution.
In dealing with continuous random variables, it is often of interest to
study some function 8(x) of the random variable x. Then v = 9(x) is
also a random variable. Let us study how to determine the density function
 94 PROBABILITY THEORY AND STOCHASTIC PROCESSES CHAP. 3

for v from the known density function for x. We shall assume that v = 8(x)
is invertible, and that the inverse has a continuous derivative so that we
can write x = f(v), (possibly with different f functions needed for different
ranges of x), and dx = ^'(w) dv where = df/dv. Then

f(x) dx = dv

where the limits on the second integral are those corresponding to x = —oo
and x = oo respectively. The range of values for the variable v may not
be from -« to ». Suppose that the range is from wmin to iw Then after
we have converted the second integral in (3-20) to the form
Cv max

itata h^ dV = lj = 0 f°r V> Vm*x> V < ymin (3-21)

Hv) > 0, h(v) will be the density function for the random variable v.
In the event that the transformation from a: tow is one to one, and
t 7u°’,S0 that Wmitt = 0(-co) and W = «(«), then h{v) = f[f(v)W(y).
If the transformation is one to one and < 0 so that vmin = fl(oo)
and w - «(-»), then in going from (3-20) to (3-21) the limits of inte¬
ntion must be interchanged so that h(v) = In general, if
me transformation from v to x is one to one, h(v) = fbP(v)]\^'(v)\
When *(») is not unique so that different * functions are needed for
. ferent 1’anges of x values, then the right-hand side of (3-20) must be
m erpreted as the sum of two or more integrals, one for each integrated
over the appropriate range of v values. An example involving a non-unique
v function will be presented below.
We shall make use of several continuous distributions in our later work
Ihe one to be used most frequently will be the well-known normal distri¬
bution for which the density function is

n(x;iji, ff) = • e-(V2<r2)(x-/*)2

where n, <r are specified constants whose physical interpretation will be

given later. The coefficient 1/V2x <r appears in order that (3-18) hold To
prove that (3-18) holds when »(*; <r) is given by (3-22), we wish to show

V2^X„ dx = 1 (3-23)

To do this it is convenient to introduce the new random variable

 SEC 3-4 CONTINUOUS RANDOM VARIABLES 95

so that dx — adw, i.e., = c. Furthermore when x = —oo7w= — oo and

when x = w = Hence the density function for w, which we shall
designate by <f>(w), is a normal density with n = 0, <t = 1, i.e.,

^ = vfc 6~W2/2 = nC^(w);* ^'(w) (3-24)

so that to prove that (3-23) holds is equivalent to showing that

To do this note that

“2/2 dw J_ e~“2/2 du

2tt i:s. e~(l/2)[w*+U*] fly]

and the product of two single integrals has been converted into a double
integral. Then write w2 + u2 = r2, w = r cos 6, u = r sin 0, so that on
changing to polar coordinates

P= Jo r6~T,/2 dr d6 = / Te~rV2 dr = jj #= 1
This proves that (3-23) is correct.
Another continuous distribution which we shall use occasionally is called
the gamma distribution. The gamma density function is defined by

«! ’ -0 (3-25)
[o , x < 0
where a is a non-negative integer and p is any positive number. Note that
^Fis case the random variable is restricted to being a non-negative
number it can never take on negative values. The function y(x; a, p) is
called a gamma density of order a + 1.
To show (3-18) holds we observe that

y(x; a, p) dx PifixYe-e* dx = uae du

 96 PROBABILITY THEORY AND STOCHASTIC PROCESSES CHAP. 3

where u = /3x. However, if we integrate by parts

/-» 7(a:; “> ^dx = ~i + ^4iyi jf du

1
(a - 1)! JQ yCt-le-U

We repeat the integration by parts on the right-hand side of the above

equation. After a steps we are reduced to

t y(x; a, P)dx = / e-“ du = -e-“|? = 1

which is what we desired to show.

It is possible to prove, although we shall not do so, that the integral

r(« + 1) = fi(j3x)ae~Px dx (3-26)

converges to a finite positive value for any real value of oc > —1 The
value of the integral depends only on a and is denoted by r(« + 1) The
SS "r11^ ^ fmma function of « + L Thus it is possible to
generahze the gamma density (3-25) to any real a > -1 if a! is replaced
ylt“ + 1-) When “ 1S a non-negative integer, then r(<* + 1) = a\.

EXAMPLES 1. If m (3-25), a = 0, we are reduced to the density function

{fterfe, x > 0
[0 , x < 0

SlSd+h1 °ft+he,g!mma density’ Le-> a Samma density of order one,

Bution? + Tlty said to have an exponential distri¬
bution^ We shall denote the exponential density by e(x:p)The cumu¬
lative function for e(x; 8) is

$(x) = P [ = 1 — e~&x
Jo ' /
The complementary cumulative is

E(x) = 1 - = e-Bx

The tmctiom «(«;». Eiz), £(*) Me plotted in Pig, 3-3 to the ca* of
 SEC 3-4
CONTINUOUS RANDOM VARIABLES 9

FIGURE 3-3.

2. Consider the change of variables v = w2/2 or w = ±(2v)1i*in the normal

distribution with M = 0 and a- = 1, i.e., in (3-24). When w = «, v = «, ;
^ T °j ^ =,°[ W = _c°’ v = Here we have a case where ^ is double
valued, and the range of v is from 0 to «>. We can write (3-20) as

W, /__«-« *»+^ j‘«-««&,

Note that when w ranges from -« to °c, v moves over the interval 0 to oo
twice. On changing to the variable v this becomes

1
V 0-/2)Q—0 fly V (l/2)e-V fly y-( 1/2) e-v fly
2vV i:

Hence the density function for the random variable v is

y/~ v (1/2>e ”, v > 0, and 0 for v < 0

This is simply a gamma density function with 0 = 1, « = -§. We leave

it for Problem 3-4 to show that r(§) = Vir.
 98 PROBABILITY THEORY AND STOCHASTIC PROCESSES CHAP. 3

3-5 Expected Values

In Chapter 1 it was suggested that the minimization of average annual

costs was a useful criterion for the determination of operating doctrines for
inventory systems. In Chapter 2 we saw how to average over time for de¬
terministic systems to determine average annual costs. When certain
inputs to an inventory system are not deterministic but instead are random
variables, then in order to determine average annual costs, we shall see
later that it also becomes necessary to average over the possible values
which the random variables can take on.
Consider an experiment whose outcome can be described by a random
variable x which can assume only the non-negative integral values 0,
1, . . ., n. Let p(x) be the probability of x. Let us imagine this experiment
is performed a very large number of times. We add up the x values ob¬
tained on each experiment and then divide by the number of experiments to
obtain the average value of x which we shall denote by x. We now ask:
To what limit will x tend as the number of times the experiment is re¬
peated is increased to infinity?” The definition of p(x) implies that as the
number of times the experiment is repeated approaches infinity, then the
fraction of the time when we obtain x approaches p(x). Thus x should
approach the value
n

V- = X) ^0) (3-27)
x=Q

The number n defined by (3-27) will be called the mean or expected value
of x. It is also called the mean of the probability distribution of x.
Let us now consider some function of the random variable *, say 6{x)
( or any inventory model x might be the demand in a given year and 6(x)
tiie cost of operating the system). Note that 6{x) is also a random variable.
Then if we find the average value of 6(x) by repeating the experiment a
large number of times, as the number of times which the experiment is
repeated approaches infinity, we expect the average value of 6(x) to tend to
the number

Y^e(x)p(x) (3-28)

which will be called the expected value of 6(x). Note that in general, the
expected value of 0(x) will not be 8(n), ju being the expected value of x.
. In event that the random variable x can take on any non-negative
integral value, it is only necessary to replace the upper limit n on the sum¬
mation signs m (3-27) and (3-28) by » to obtain the appropriate expected
values for this case. If the expected value is to have any meaning when x
 SEC 3-5 EXPECTED VALUES 99

is allowed to range over all non-negative integers, the sums (3-27), (3-28)
must converge. In all cases of interest to us they will.
When 6(x) — (x — p)2, p being the mean of x, then (3-28) is called the
variance of x and is denoted by <r2. The variance of # is a rough measure of
the sort of spread in x values about the mean that can be expected. If <r2 is
large we can expect a greater spread in the x values than if a2 is small. The
number <r = V<r2 is often called the standard deviation of x. The number
<r2 is also called the variance of the distribution of x.
By its definition

a2 = (x - p)2p(x)
2=0

This can be written

°"2 = X) (x2 “ x + v2)p(%) = ^2 x2p(x) — 2p X) xp{x) + M2 X) Pte)

x=0 2=0 2=0 2=0

= Z
2=0
z2p(a:) - 2m2 + m2 Z x2pOO - p2
2=0
(3-29)

and we have obtained a useful result which will be of assistance below in

computing the variances of specific distributions. Equation (3-29) also
holds if x is defined for all non-negative integers, provided that the sums
converge.
Let us now evaluate the means and variances of the discrete distributions
introduced in Sec. 3-3. For the binomial distribution, we see from (3-27)
and (3-10) that
n\
x! (n — x)!
p*(i — p)n

- t - <-)*-

= np 5 (n ~ *) P"(l - = np[p + (1 - p)]"-1 = np

In the first step we cancelled the x in xl to yield (x — 1)1 and factored out
np. In the second step we made the substitution u — x — 1. The resulting
summation is simply the binomial expansion of [p + (1 — p)]n_1. Hence
the mean of the binomial distribution is np. This is to be expected intui¬
tively, since if p is the probability of a success on a single trial, one would
expect the average number of successes in n trials to be np.
 100 PROBABILITY THEORY AND STOCHASTIC PROCESSES CHAP. 3

To compute <r\ we use (3-29). Thus

^ = S^^Tn-x)\pX(1 " A''* ~ M2

= 5 *(* ~ 1} x\(n- X)lpX(1 ~ ')*"*

+ S *xT(n — a;)!pX(1 ~ ^ ~ M* (3-31)

- n(n - l)p2 S ^ _ 2 - M)! p“^ “ p)b_2_“ + np - (np)2

= n(n — l)p2 + np — (np)2 = np( 1 — p)

Note that the mean and variance of the binomial distribution depend on
born n and p. *
Consider next the Poisson distribution. We first note that the mean is

(3-32)

pros the parameter p appearing in the definition (3-13) of the Poisson

“on is simply the mean of the distribution. The variance is, by

172=-p2= tx(x ~ -p2

12 (3_33)

Thus ^ = p and the variance of the Poisson distribution is numerically

equal to its mean. ^
For the negative binomial distribution

(3-34)
-nJLfAt(uV‘)^-^
r «ls=n A
 SEC 3-5 EXPECTED VALUES 101

Furthermore
oo

<r2 = n, p) - p2
x—0

— ^2 x(x — l)6iv(^; n, p) + /t — /x2

z**2
(3-35)
n(n + 1)(1 — pV
]C n + 2, p) + n — ju2
u-0

n(n + 1)(1 - pY = P?
~~ M
M P2
The geometric distribution is a special case of the negative binomial for
which n = 1. Thus for the geometric distribution, as defined by (3-16)

1 — P
(3-36)
P '

Consider now the computation of expected values for continuous random

variables. If f(x) is the density function for the random variable x, then
the fraction of the time that x lies in the interval x to x + dx should ap¬
proach j(x) dx as the number of times the experiment is repeated
approaches infinity. Hence if 6(x) is any function of x, the average value
of d(x) in the limit as the number of experiments over which 6{x) is averaged
approaches infinity will tend to

/: 6(x)f(x) dx

provided that the integral converges. The number computed according to

(3-37)

(3-37) is called the expected value of 6(x). The mean p (when 0(x) = x)
and variance <r2 [when 0(x) = (x — jt)s] of x and of the distribution of x
are then defined to be

xf(x) dx; <r‘

-L (x — p)2f(x) dx (3-38)

Let us now compute the mean and variance for the normal and gamma
distributions defined in Sec. 3-4. For the normal distribution given by
(3-22) the mean is

i r
=— / xe~ (1/2*’) Or-*)’ dx _1 f (X-p[ e-(l/2*>)(x-
dx + p
2ir <r J—<*
V/2jt *v 2x J — ® O'

+ + (M8)
 102 PROBABILITY THEORY AND STOCHASTIC PROCESSES CHAP. 3

where
1 (x - ixV
V 2\ <r )
Thus the parameter m in the definition (3-22) of the normal distribution is
its mean.
The variance of the normal distribution is

(x - dx = ~= / w2g-(”!/2) dw
a v 2t ./-»
Now integrate by parts writing u = w,dv = we~wi/2 dw so that v =
—e~w2/2. This yields

7= / w2e~~w2/2 dw

-we-™1!2 + / e-»V2 dw V = 0-2 (3.40)

Thus the parameter cr appearing in the definition (3-22) of the normal

distribution is its standard deviation.
A plot of the normal density function looks something like that shown
in Fig. 3-4. It is symmetric about x — y.) and the inflection points occur
at one standard deviation on either side of the mean.
Let us finally examine the gamma distribution. Before turning to the
computation of the mean and variance, we shall first show that for any
real a > 0, T(a + 1) = aT(a). To prove this, recall that from (3-26)

T(a+ 1) = / dx wae~w dw

Integration by parts yields

. T(a + 1) = —wae~w |°° * a j w' le~w dw = aT(a), a > 0 (3-41)

Jo

We shall now compute the mean and variance for the gamma distribu¬
tion in the case where a can be any real number greater than — 1 and 0 is
any real positive number. From the generalization of (3-25) with T(a+ 1)
replacing a!, we see that the mean is

M = rJ*(Ma e-?xdx
Jo T(a +1) “
(3-42)
_ T(q + 2) 1 f00 f3((3x)a+1 a 4 1
~
e~Px dx =
T(a + 1) 0 Jo T(a + 2)
 SEC 3-5 EXPECTED VALUES 103

by (3-41). Similarly, the variance is given by (we leave for Problem 3-22
the proof that the equivalent of (3-29) holds for continuous random
variables)

(a + 2)(a+ 1) a -f- 1
R2
m= ;S2
To express a, /3 in terms of n, a note that on using (3-42) in (3-43)

P=£ (3-44)
SO

a = (;)' - 1 (3-45)

A qualitative plot of the gamma densities for several integral values of a

and a fixed /3 is shown in Fig. 3-5.
 104 PROBABILITY THEORY AND STOCHASTIC PROCESSES CHAP. 3

3-6 Time Averages and Ensemble Averages

In the previous section we have given the mathematical definition of the

expected (or mean) value of a random variable and have indicated how
the expected value can be computed once the probability density func¬
tion for the random variable is known. In dealing with .inventory models
we shall have frequent use for expected values. Physically, the expected
value of a random variable will be interpretable as a. time average of the
random variable, an ensemble average, or both. We shall now introduce the
notions of the time average and ensemble average of ,a random variable,
and show their relation to the expected value of the random variable.
The notion of a time average is quite straightforward. We imagine that
we have a system and we continue to allow it to operate for all future time,
with demands and lead times being generated by the appropriate stochastic
processes for the system. Then the time average of any random variable x
associated with the system will be either

limy
t—+ oo t i: x(t)dt

if u; is defined at each point in time (such as the on hand inventory or the

number of backorders) or
 SEC 3-6 TIME AVERAGES AND ENSEMBLE AVERAGES 105

if z is defined only with respect to a time period (such as the number of

orders placed per year). Then n is the number of time periods used in
computing the sum. The above definitions hold for either discrete or con¬
tinuous random variables. To be specific, it will be assumed in the follow¬
ing discussion that x is discrete, although precisely the same results hold
true if £ is continuous.
In general, time averages will be of interest only when the stochastic
processes associated with the system are not changing with time. The
notion of a time average will be of most importance in computing the aver¬
age annual cost introduced in Chapter 1. One cannot usually compute
analytically the average values of a random variable in the time average
sense using only the above definitions, since the average will depend on
the nature of the stochastic processes involved. To relate a random var¬
iable to these stochastic processes, one determines a probability density
function p(x) which has the property that, as the length of the time interval
taken for consideration becomes sufficiently long, the fraction of the time
that the random variable has a given value x approaches p(x)> Then the
expected value of x can be computed as described in the previous section,
and because of the way p(x) is defined it will be the average value in the
time average sense.
The other and more important type of average is called an ensemble
average. The notion of ensemble averages has been used a great deal by
physicists in the study of statistical mechanics, and the same notion is quite
helpful in studying inventory systems. Suppose that we are studying an
inventory system for which the stochastic process generating demands is
changing with time. We then ask the question, “What do we mean by
saying that the probability that x units will be demanded in the next week
is p(x)?)} Clearly we cannot mean that if we observed the system for a long
time the fraction of the time that x units were demanded will approach
p(x). The reason is that the mean rate of demand may be changing with
time and hence a time average would be meaningless.
A clear physical interpretation can be given to p(x) in the following way:
Instead of considering just a single system, suppose that instead, we con¬
sider a large number N .of identical systems. Imagine that we start them
all off at the beginning of the week in identical states and allow the demands
for each system to be generated by the appropriate stochastic process in
such a way that the demand for each system is generated independently of
the others. Since the systems are identical, the same stochastic process
will be associated with each system. However, the week’s demand will vary
from one system to another, since we assumed that they operate indepen-
 106 PROBABILITY THEORY AND STOCHASTIC PROCESSES CHAP. 3

dently, i.e., the stochastic processes for the various systems operate inde¬
pendently of each other. With this picture in mind, the fraction of the
systems which have demands of precisely x will approach p(x) as iV —> oo.
Thus if one computes the expected demand in accordance with the rules
given in the previous section, this expected demand will be the ensemble
average demand, i.e., the average over the ensemble of systems of the
week’s demand.
Let us now imagine that the stochastic processes associated with a sys¬
tem do not change with time. Again we can imagine that we have an
ensemble of N of these (identical) systems operating. After the ensemble
of systems has been operating for a long time, it will reach a condition of
statistical equilibrium. This means that the ensemble average of any
relevant variable will become independent of time. When statistical equi¬
librium is reached, then, for any random variable, an ensemble average
over the ensemble of systems will be precisely the same as a time average
over all time for a single system,* so that p\x) can be interpreted as the
fraction of the time that a single system will have the value x, or the frac¬
tion of the systems in the ensemble that will have the value x at any given
point in time.
The notion of an ensemble average is sufficiently general so that every
expected value used in this book can be thought of as an ensemble average,
and every probability as the limit as the number of systems in the ensemble
approaches infinity, of the fraction of the systems in the ensemble which
yield the given value of the random variable under the conditions specified.
When the stochastic process under consideration does not change with
time, an expected value can be interpreted both as a time average and an
ensemble average. Furthermore, in this case, the expected value will be
the same regardless of which way it is computed. Similarly, when the
stochastic process does not change with time, a probability statement can
be interpreted as referring to the long run fraction of the time that the
random variables will have a specific value, or the fraction of the systems in
the ensemble that will have the value.

3-7 Probabilistic Description of Demands

In this section, we would like to investigate the way in which the proba¬
bilistic nature of the demands received by an inventory system can be
described in the case where the nature of the process generating the de-

The equivalence of time averages and ensemble averages in this case is, roughly
speaking, a statement of the celebrated ergodic theorem of statistical mechanics. No
mathematically satisfactory proofs of the ergodic theorem were available until 1931
when proofs were provided by G. D. Birkhoff and John von Neumann.
 SEC 3-7 PROBABILISTIC DESCRIPTION OF DEMANDS 107

mantis does not change with time. Let us begin by noting that the number
of units demanded in any time period will depend on the time between
demands and on the number of units demanded when a demand occurs.
In the real world, both the time elapsed between demands and the quantity
demanded can be random variables. The time t between demands will be a
continuous random variable and the number of units demanded when a
demand occurs will be a discrete random variable.
Consider first the random variable t representing the time between the
occurrence of demands. Let G(t) be the probability that a time greater
than or equal to t elapses from the time of a given demand until the time of

the next demand. In general G(t) could be a function of the tirmw of occur¬
rence of all previous demands, of the quantities demanded at each of the
previous demands, of calendar time, and perhaps many other things. We
shall here restrict ourselves to the case where G(t) depends only on the tima
since the last demand, and not on the times of any other demands, quan¬
tities demanded, or calendar time.
If we were to plot G(t) as a function of t, it might have one of the shapes
shown in Fig. 3-6. Then G(t) must be unity at t = 0, since the probability
is unity that the demand under consideration will occur after the previous
demand. Furthermore, we assume that it is certain that another Hamanrl
will occur at some future time, so that G(t) must approach 0 as t approaches
oo. In the event that G(t) can be represented by curve 1 in Fig. 3-6, then
the time between demands is deterministic, i.e., the time between demands
 108 PROBABILITY THEORY AND STOCHASTIC PROCESSES CHAP. 3 S

is always t0. Curves such as those shown in Fig. 3-6 can often be represented
quite well by the complementary cumulative function of one of the gamma
densities for an appropriate choice of a, 0. The distribution of the time
between successive demands is often referred to as the interarrival dis¬
tribution. h
A form of G(t) which will be of greatest interest to us will be that where
G(t) is the complementary cumulative function for the exponential dis¬ V
tribution (see example on p. 94), i.e.,
T
Git) = <rXf (3-46)
When G(t) has the form (3-46), we say that the times between demands
are exponentially distributed. The density function for t is S.
dC
= Xe~x< *e(Jt’x) (3-47)
T
and the probability that the next demand arrives between t and t + dt
after the present demand is e(t) X) dt. The average time t between demands
is

t = Jq \te~u dt = i (3-48) T
th
so that the average number of demands per unit time is X. tr
If G(t) is given by (3-46), it is not difficult to compute the probability (3
Vn(t) that precisely n demands occur in a time period of length t imme¬
diately following the occurrence of a demand. This is most easily done by
obtaining a recursive relation for the Vn(t). Note first of all that

V0(t) = G(t) = (3-49)

When n > 1, let us imagine that the first demand occurring in the time
interval of length t occurs between r and r + dr after the demand which an
serves as the time origin for the interval under consideration. If precisely
u demands occur in the interval of length t, and if the first one occurs be¬ m
tween r and r + dr, then precisely n — 1 demands must occur in the in¬ dh
terval of length t — r. The probability that n — 1 demands occur in an
interval of length t - r following a demand is simply Vn~i (t — r). Thus nu
the joint probability that the first demand occurs between r and r + dr th<
and that n — 1 demands occur in the interval of length t — r following will
this is
Xe-^Fn-i^ - r)dr (3-50) at
int
Now events corresponding to different times of occurrence of the first als
demand are mutually exclusive. Hence, the probability Vn(t) that there
occ
are precisely u demands in a time interval of length t is found by summing
at
(3-50) over all r, 0 < r < % i.e.,
In
 SEC. 3-7 PROBABILISTIC DESCRIPTION OF DEMANDS 109

This is true ami can easily lie proved by induction. Wo have already shown
that (d-o2) is true for n 0 and « 1. It remains to show that if it is
true for « m 1, it is also true for « m. Tills follows, since if w<i use
(2-f>2) evaluated fur n m 1 in Cl-ol.), we obtain

*,!.*>' 11 |r: c ^ll rl dr

{tn 11!

A"c K‘ (A t)m
{t — r)w 1 dr hi
1)! «; ml *'

and hence, by induction, (Jt-;72) holds for all non-negative interns n.

We have* obtained the interesting result that the number of demands
m occurring in a time* interval of length t following a demand Ls Poisson
distributed, with the mean of the Poisson distribution being A/, i,e., Pm(£)
pirn; A/b Note that since A is the mean rate of demand, At! is the mean
number of demands whieh will oeeur in the time interval of length L Thus
the mean of the* Poisson distribution is the average number of demands that
will oeeur in the interval.
In the preceding derivation, the time interval was not allowed to begin
at an arbitrary point in time. Instead, it was required that the time
interval begin at. a point in time immediately following a demand. It is
also of interest to compute tin* probability that precisely m demands will
occur in a time interval of length t when observation of the system is begun
at a random point in time rather than immediately following a demand,
in order to do this we must first establish another interesting property
 110 PROBABILITY THEORY AND STOCHASTIC PROCESSES CHAP. 3

which follows from having the times between demands exponentially

distributed.
Let us compute the conditional probability that the next demand after
a specified demand occurs between t and t -f- dt after the specified demand,
given that it must occur after time t. If A is the event that no demand
occurs up to time t, and B is the event that the demand occurs between t
and t + dt, then p(A) = e~u and p(B) = Xe~X! dt. We wish to com¬
pute p{B\A). From (3-3)
p(A\B)p(B)
P(B\A) =
P(A)

However, p(A\B) = 1, since if the demand occurs between t and t + dt it

cannot have occurred before t. Thus

p{B\A) = = \dt

The important thing to note is that p (A\B) is independent of t. This implies

that if we start observing the system at a random point in time and ask
what is the probability that a demand will occur in the next infinitesimal
time interval of length dt, the probability is X dt, independently of when we
begin our observation, since this probability does not depend on how much
time has elapsed since the last demand.
By use of (3-53) we can now compute the probability U0(t) that no de¬
mand occurs in a time period of length t from the time when we begin ob¬
serving the system, if we start observation at a random point of time. To
compute U0(t), observe that U0(t + dt) is simply the probability that no
demands occurred up to time t multiplied by the conditional probability
that no demand occurs between t and t + dt given that none occurred up
to t. This latter probability is 1 — X dt. Thus

U0(t + dt) = (1 — X dt) U0(t)

or
dU0
Uo
Hence
U0 = ce~M

where c is a constant of integration. Now at t = 0, U0 = 1 since we begin

observation at t — 0. Hence we obtain the unique solution

U0(t) = (3-54)
Recall that we began our discussion by assuming that the probability that
the time from one demand to the next is greater than or equal to t is
 SEC 3-7 PROBABILISTIC DESCRIPTION OF DEMANDS 111

G(t) = e u. We have now proved the interesting result that, if we begin

observation of the system at a random point in time rather than at the time
a demand occurs, the probability that the time until the next demand will
be greater than or equal to t is again e~M. In other words, the probability
that the next demand will not occur for a time t or longer is independent of
how long it has been since the last demand occurred.
Having computed U0(t), we can easily determine the probability UJt)
that exactly n(n > 1) demands occur in a time period of length t when
observation of the system is begun at a random point in time. If the first
demand occurs between r and r + dr (0 < r < t) after observation is
begun (time 0), then to have n demands occur by time t, it must be true
that precisely n — 1 demands occur in the period from r to t. The proba¬
bility that the first demand occurs between r and r + dr and that n — 1
demands occur from r to t is

_ dXJ^rl v^(t _r)dr = Xe-XrVn_i(t _ r) dr

Note that Tn-i rather than Un-1 appears, since the time interval beginning
at time r and extending to time t begins immediately after the occurrence
of a demand. Thus on averaging over r

= ^ J0 e~XTV*-i(t ~ r) dr, n > 1 (3-55)

Hence, from (3-52) we see that

Un(t) = Vn(t) = p(n; \t) = e~™ n = 0,1,2, ... (3-56)

and Un(t) also has a Poisson distribution with mean \t. We have proved
then that the probability of having n units demanded in a time interval of
length t is p(n; Xi), and this is independent of whether we begin observation
immediately after the occurrence of a demand or at a random point in time.
The equality of the Un and Vn is a peculiar feature of the exponential inter-
arrival distribution and does not hold for an arbitrary interarrival distri-
bution. When the interarrival distribution is exponential, we sometimes
say that the demands are generated by a Poisson process. An important
characteristic of the Poisson process is that it has no memory, i.e., the
probability that a demand occurs in a time period of length dt is X dt inde¬
pendently of when any previous demands occurred.
Let us suppose that a Poisson process is generating demands and that
units are always demanded one at a time, so that the quantity demanded is
always unity. Then Un(t) is also the probability that n units are demanded
in a time period of length t. Thus, in this case, the probability of having
 112 PROBABILITY THEORY AND STOCHASTIC PROCESSES CHAP. 3

precisely n units demanded in a time period of length t is the Poisson proba¬

bility p(n; \t). Now X is the average rate at which units are demanded,
and \t is the expected number of units demanded in the time period. In
later chapters we shall use the terminology the mean or average rate of
demand (or demand rate) to mean the average rate at which units are
demanded. Note that when units are not demanded one at a this will
be different from the mean rate at which demands occur.
The Poisson distribution is especially easy to work with analytically.
Furthermore it turns out that, in practice, it is frequently true that the
process generating demands can be approximated fairly well by a Poisson
process. Almost all of the future work in this book which treats demands
discretely will assume that they are generated by a Poisson process.
In the event that the time between demands has a gamma distribution
(3-25) with a an integer, then we say that an Erlang process of order a + 1
is generating the demands. An Erlang process of order one is a Poisson
process. We leave the task of working out the probabilities Vn(t) for an
Erlang process of order <x + 1 to the problems. These probabilities
are quite complicated when a > 1 and are rarely used in the analysis of
inventory systems.
By the use of what we have learned above, let us compute for a Poisson
process the probability that the Qth demand occurs between time t and
t + dt after we begin observing the system. This probability can be
thought of as the probability of A and B where A is the event that Q — 1
demands occur between 0 and t and B is the event that a demand occurs
between t and t + dt. Note that for a Poisson process B is independent of
A, and hence p(A (~) B) = p(A)p(B) where

p(A) = p(Q - 1; \t), p{B) = \dt

Thus the desired probability is

Hence the density function for the time until the Qth demand is a gamma
density with a = Q - 1, i.e., a gamma density of order Q. We have just
shown that if for some inventory system a Poisson process generates de¬
mands, and if units are demanded one at a time, then if an order is placed
each time Q units are demanded, the density function for the time between
the placing of orders is a gamma density of order Q.
In the event that the number of units requested when a demand occurs,
i.e., the quantity demanded, can vary randomly from demand to demand,
the task of determining the probability that any given number of units will
be demanded in a time period of length t becomes much more complicated.
We shall illustrate this by studying one of the simplest possible cases in
 SEC. 3-7 PROBABILISTIC DESCRIPTION OF DEMANDS 113

which the number of units demanded can be a random variable. We shall

assume that a Poisson process generates the demands, and that the number
of units demanded when a demand occurs has a geometric distribution, i.e.,
the probability that when a demand occurs x units will be demanded is

bN(x - 1; 1, 1 - v) = (1 - v)v^\ x = 1, 2,. . . (3-57)

In the original definition of the geometric distribution (3-16), x could take

on the value zero. Here we always want at least one unit demanded when
a demand occurs—hence the change from x to x — 1 in going from (3-16)
to (3-57).* It is also convenient to write v = 1 — p. Note that we assume
that the number of units demanded per demand is independent of how long
it has been since the last demand. The process formed by having a Poisson
process generate demands and the quantity demanded being distributed
according to a geometric distribution is sometimes referred to as a “stutter¬
ing Poisson” process.
For a stuttering Poisson process, let us now compute the probability
that precisely n units will be demanded in a time period of length t. Since
a Poisson process generates the demands, we know that the results are
independent of the time origin. If n units are demanded, then no more
than n demands could have occurred. However, it is possible to have n
units demanded if there are 1, 2, . . ., or n demands. Events representing
different numbers of demands in the period are mutually exclusive. The
probability of precisely j demands is p(j; fit) if 1/fi is the mean time be¬
tween demands. Thus if Un(t) is the probability that n units are demanded
n

Unit) = Pin\j)p(r>fit), n = 1, 2, . . . (3-58)

y=i

where p(n\j) is the probability that precisely n units are demanded when j
demands occur. Equation (3-58) does not hold for n = 0. The probability
that no units are demanded is the probability that no demands occur. Thus

U0(t) = p(0; fit) = e~& (3-59)

It remains to compute p{n\j). Let xt be the random variable representing

the number of units demanded when demand i occurs. Then p(n\j) is the
distribution of the random variable n = xi + x2 + . . . + x3-. We know
the distribution of each Xi (it is geometric). Later, when we discuss con¬
volutions, we shall show how to obtain the distribution of the sum 0f in¬
dependent random variables. We shall state and use here the fact that if
each Xi has the geometric distribution (3-57), then n — j, the demand in

* Stated differently, if y is the demand in excess of one unit, then y has the geometric
distribution (3-16).

THE HSU UBOT

%mmi sisiiiyiE of
114 PROBABILITY THEORY AND STOCHASTIC PROCESSES CHAP. 3

excess of one unit per demand, has a negative binomial distribution

— jj, 1 — v). Hence

p(n\j) = bN(n — j;j, 1 — v) = _ j)(l — p)j'p»-> (3-60)

Thus

Un(t) = v^tj{ (" I #]’« n = l,2,... (3-61)

The Un{t) given by (3-59), (3-61) are referred to as the stuttering Poisson
distribution. It is rather difficult to work with this distribution, which is
one of the simplest representing situations in which more than a single unit
can be demanded at a time. Fortunately, as we shall see shortly, there is a
simple approximation to all these distributions in terms of a continuous
random variable which lends itself quite easily to analytical work and which
often provides a sufficiently good approximation in practical applications.

3-8 Joint Distributions

Often in working with inventory systems it is necessary to deal with two or

more variables simultaneously. For example, the random variables may
refer to demands in different time periods, or one or more random variables
may refer to demand while one or more different random variables may
refer to the procurement lead time. We wish to study briefly here density
functions involving two random variables.
Consider two discrete random variables x, y, both defined over the non¬
negative integers, which always occur together. We can then introduce a
function p(x, y) defined for each set of values x, y which gives the proba¬
bility of the event x Pi y, i.e., the event that x and y occur. The function
p(x, y) is called the joint density function (or joint distribution) for the
random variables x> y. It is typical to use the usual functional notation
for a function of two variables, p(z, y), rather than the symbol p(x O y).
Since one set of values x, y must occur, and since we as usual assume that
two different values of either random variable are mutually exclusive, we
must have
QO 00

12 12 p(*» y) 1
x=0 y=0
= (3-62)

Let us now ask the question, “What is the probability of x without

regard to what value y takes on?” We shall denote this probability by
p(x); p(x) must simply be the summation of p(x, y) over all possible values
of y, i.e.,
 SEC 3-8 JOINT DISTRIBUTIONS 115

OO

p(x) = Z p(x> v) (3-63)

y—0

Similarly, p(y), the probability of y without regard to the value of x is

given by
oo

p(y) = Z p(x> y) (3-64)

x=0

The functions p(x), p(y) are called the marginal densities for the random
variables re and y respectively. Note that they are legitimate densities
smee p(x) > 0, p{y) > 0 and from (3-62)

Z p(?) = h Z p(y) = 1
x=0 y—0

If either or both of the above random variables are defined only over a
finite number of integral values, the summations are taken only over the
allowed values of the variable. When p(x) ^ 0 and p(y) ^ 0, the con¬
ditional probabilities p{x\y) and p{y\x) are defined by

p(x, y) = p(x\y)p(y) = p(y\x)p(x)

In terms of the conditional probabilities we can write

pfr) = Z p(x\y)p(y); v(.y) = Z p(.y\xM*) (3-65)

y~o x=o

We say that x and y are independent if p(x, y) = p(x)p(y).

EXAMPLE An important application of marginal distributions in the anal-

ysis of inventory systems can be stated in an especially simple form as
follows. Let the demands on the system be generated by a Poisson process
and imagine that units are demanded one at a time, so that the probability
that a; units are demanded in a time t is p(x; Xt), where X is the mean demand
rate. Assume also that the procurement lead time is a random variable
which can only take on one of the following finite number of values
h,. . ., tk, the probability of being l(U). We now ask what is the prob¬
ability that x units will be demanded in a lead time. In other words, we
want the marginal distribution of lead time demand. Note that the number
of units demanded in the lead time is not independent of the lead time U.
However, from (3-65) we see that the probability p(x) that x units are
demanded in a lead time is
h

p(x) = Z P(-X> (3-66)

i=l
116 PROBABILITY THEORY AND STOCHASTIC PROCESSES CHAP. 3

Consider next the case where 2 is a discrete random variable defined over
the non-negative integers and y is a continuous random variable defined for
all real y. We then take f(x, y) to be the joint density function for x and y;
f(%, y) dy is the probability that the discrete random variable has the value
x and that the continuous random variable lies between y and y + dy.
Then, since some pair of values must occur, we have

X) / /(*> y)dy = 1 (3-67)

x=QJ—«

The marginal densities p(x), v(y) of x, y respectively are defined by

P(x) = f f(*,y)dy; v(y) = f(x, y) (3-68)

Then p(x) is the probability that the discrete random variable has the value
x} and v(y) dy is the probability that the continuous random variable lies
between y and y + dy.
Next when v(y) ^ 0, p(x) 0, we define conditional density functions
g(y\x) and h(x\y) such that

/Or, y) dy = h(x\y)v(y) dy = [g(y\x) dy]p(x) (3-69)

Then h(x\y) is the conditional probability that the discrete variable has the
value x given that the continuous variable has the value y, and g(y\x) dy is
the conditional probability that the continuous random variable lies be¬
tween y and y + dy given that the discrete random variable has the value x.
Then (3-68) can be written

/ CO

h(x\y)»(y) dy,
go

v(y) = ^ g(y\x)p(x) (3-70)

-co x= 0

We say that x, y are independent random variables if

/Or, y) = p(x)v(y) (3-71)

If, for an inventory system, the procurement lead time is treated as a

continuous random variable, and the number of units demanded in any
given time period as a discrete random variable, then we have an example
of the situation just discussed. For the case where the procurement lead
time has a gamma distribution, and a Poisson process generates demands,
with units being demanded one at a time, let us compute the marginal
distribution p(x) of lead time demand. This result will be of use to us later.
By assumption, the probability density for the lead time t is y(t; a, 0)
where y(t; a, @) is given by (3-25), and the density function for the number
SEC 3-8 JOINT DISTRIBUTIONS 117

of units demanded in a time t is p(x; \t) where X is the mean rate of demand.
Thus

y(t; a, 0)
my e -fit.? p(x; \t) (3-72)
T(a + 1)

where for added generality we have replaced a\ in (3-25) by r(a + 1).

Note that p{x) \t) corresponds to h(x\y) above. Then using (3-70)

my
v(x\ — [ PO'O-—iMil—e-os+x)(^ = —ft--—— [ ix+ae-w+\)t
P{X) ~ Jo xl F(a + 1)
L)°6 ' M x\T(a + 1) Jo 1 6

/3a+1X* T(a + x + 1) jf‘

x\T(a +1) (/3 + X) a-j-x+1 T{a + x + 1) 6 at

(3-73>
by (3-26). Thus if a is a positive integer, the marginal distribu¬
tion of lead time demand p(x) has the negative binomial distribution
bxlx; a + 1, 0/(0 + X)]. There is no reason why a must be an integer in
the definition of the negative binomial distribution. We can consider (3-73)
to be a generalized definition of the negative binomial distribution which
includes cases where a is not an integer. We have shown, therefore, that if
the lead time is gamma distributed, and demands are generated by a
Poisson process, units being demanded one at a time, then the marginal
distribution of lead time demand is a negative binomial distribution.
It remains to consider the case where the two random variables x, y are
continuous. Now we define a density function/(x, y) such that f(x, y) dx dy
is the probability that x lies in the interval x to x + dx, and y lies in the
interval y to y + dy. Then it must be true that

f(x, y) dx dy = 1 (3-74)

The marginal densities are given by

u(x) = J f(x, y) dy; v(y) = J f(x, y) dx (3-75)

where u(x) dx is the probability that x lies in the interval x to x + dx and

v(y) dy is the probability that y lies in the interval y to y + dy.
Conditional density functions h(x\y) and g{y\x) are defined by

f(x, y) = h{x\y)v(y) = g(y\z)u(z), v(y) ^ 0, u(x) ^ 0 (3-76)

where h(x\y) dx is the probability that x lies in the interval x to x + dx

given that the other random variable has the value y, and g(y\x) dy is the
 118 PROBABILITY THEORY AND STOCHASTIC PROCESSES CHAP. 3

probability that y lies in the interval y to y + dy given that the other

random variable has the value x. Then

u(x~) = J_a K^lvHy) dy; v(y) = J g(y\x)u(x) dx (3-77)

If we can write f(x, y) = u(x)v(y), the random variables x, y are said to be

independent.
Consider now two discrete random variables xh x2 defined over the
non-negative integers and let p(Xl, x2) be their density function. If d(xh x2)
is any function of xh x2, then the expected value of 8(xx, x2) is
00 00

£ £ 0(zi, x2)p(xh x2) (3-78)

Xl —0

As a special case, imagine that 6(xi, x2) depends only on x%, i.e.,
6(xi, x2) = iP(Xl). Then from (3-78) the expected value of x//(xi) is
“ » »
£ £ Kxdpixi, X2) = £ *60 £ p{xh x2) = £ t(xi)p(xi) (3-79)
zi-u x% u XI = 0 XI = 0 X1 = 0

where p(xi) is the marginal density for xx. As another special case imagine
that 6(xi, x2) = Xi ± x2. Then the expected value of xx ± x2 is
00 oo ^

£ £ ± x2) = £ Xipfa) ± £ x2p(x2) = Ml ± /12 (3-80)

Xl — 0 X2 — 0

where mi, M2 are the expected values of xi, x2 respectively. More generally,
the expected value of a finite sum of random variables xx + . . . -|- xn is
the sum of the expected values, i.e., mi + . . . + m». This is'proved by
letting the event x% = x2 + . . . + Xn and applying (3-80) to x± + x%. Then
the argument is repeated on xt Relations like (3-78), (3-79), (3-80) also
hold if one random variable is discrete and the other continuous, or if both
are continuous, provided that the appropriate summation signs are replaced
by integrals. Similarly, the expected value of

Zl±X2d= . . .=tx„ is Ml ± M2 ± . . . ± nn

3-9 Convolutions

Consider the discrete random variable x, defined over the set of non¬
negative integers, whose density function is p(x). Then the generating
function or e-transform of x, written <P(s) or z{p{x)} is defined as
SEC. 3-9 CONVOLUTIONS 119

for those $ for which the series converges. Because the p(x) must sum to
unity, we have <P(1) = 1. Consequently, the series (3-81) will always
converge for s in the interval — 1 < s < 1. Note that a knowledge of (P(s)
is as good as a knowledge of p(x) since if <P(s) is expanded in a power series
as in (3-81) the p(x) values can be found, i.e.,

and, since the expansion of a function in a power series is unique, p{x) is

uniquely determined by (P($). Generating functions are useful in many
ways. They will be of use to us in this section in finding distributions of
sums of random variables.

EXAMPLES 1. The generating function for the Poisson distribution will

always be denoted by <P(s; p). It is

p) = = E xl
M! e-p = y x!
rc = 0 *=0
(3-83)

2. The generating function for the negative binomial distribution will be

written 6V(s; n, p). It is, for integral valued n,

n, p) = y ^ ^ j ^ pn(l — p)^1

1 - (1 - P)«
The same result is obtained for any positive n, even if it is not an integer,
if (3-73) is used. The proof is left for Problem 3-23.
3. The generating function for the geometric distribution will be denoted
by p). It is found by setting n = 1 in (3-84). Thus

P
@0(8} p) (3-85)
1 - (1 - p)s

Consider two independent, discrete random variables x1} x2 defined over

the non-negative integers, and let their density functions be Pi(%i) and
P2(x2). Often in applied work one is faced with the problem of finding the
probability distribution for the random variable y = xz + £2, i.e., of finding
the distribution of the sum of the random variables X\, x2. Let us study how
this can be done. The probability p(y) that the random variable Xi + x2
has the value y is the sum of the probabilities of the y + 1 mutually
 120
PROBABILITY THEORY AND STOCHASTIC PROCESSES CHAP. 3

exclusive events (x1 = j) n (x2 = y - j),j = 0,1,. •, y. However, since

the random variables xh x2 are independent

?(*i = j n x2 = y - j) = Pl(j)p2(y - j) (3-86)

Hence
y

p(y) = Pi(j)p»(y - j) (3-87)

3=0

The density function p(y) is called the convolution of the density functions
Pi(xi) and p2(x2).
Now denote by (Pi(s) and (P2(s) the generating functions for p1(x1) and
Pz(xz) respectively. Then by definition
” eo

^ = Z CP*(«) = £>20V (3-88)

However, a theorem on the multiplication of power series says that

<Pi(s)<P2(s) = g |”g Pi(k)p2(j - A;) j s> (3_89)

iVofe: A heuristic derivation of (3-89) can be given as follows:

(?l(s)6>2(s) =

bi(0) + Pi(l)s + Pi(2)«* + . . .][p2(0) + p2(l)s + P2(2)S2 + . . .]

= Pi(0)p2(0) + Pi(0)p2(l)s + Pi(0)p2(2)s2 + pi(0)p2(3)s3

+ Pi(l)ps(0)s + px(l)p2(l)s2 + Pi(l)p2(2)s3

+ Pi(2)p2(0)s2 + pi(2)p2(l)s3

Pi(S)p2(0)s3 + . . .

= Pi(0)p*(0) + fo(0)p2(l) + p1(l)p2(0)]s + [Vl{0)p2{2) + ^(1)^(1)

+ Pi(2)p2(0)>2 + [pi(0)p2(3) + pi(l)p2(2)

+ Pi(2)p2(l) + Pi(3)p2(0)]s3 + . . .

T P^k)v2(j - k)
L&=o
s3

SSJ dmonstrates the important result that the generating

function for the random variable y = Xl + z2 is the product of the generat-
ntg functions for x, and x,. If the modem variable x„ x, have th“e
 SEC. 3-9 CONVOLUTIONS 121

density function, and if <P(s) is the generating function for this distribution,
then the generating function for y = xi ■+■ x2 is (P2(s).
Consider the random variable y = xx + x2 + . . . + x„ which is the sum
of n independent random variables xi} the generating function for the
probability density of x< being (Pt<s). To compute the density function for
y, we note that the density function for yi = xi + x2 is the convolution of
the density functions for xi and x2. The generating function for the density
function of yi is then (Pi(s)(P2(s). The density function for y2 = yi + x3
= Xi + x2 + x3 is the convolution of the density functions for y1 and x3.
The generating function of the density function for y2 is <Pi (s) <P2 (s) CP3 (s).
Continuing this process, we see that the density function for y = yn_2 + Xn
is the convolution of the density function for yn_2 and xn. The generating
function for the density function of y is then (Pi(s)(P2(s) . . . (P„($), i.e., the
product of the generating functions corresponding to each of the random
variables Xi. In the event that each x» has the same density function p(x),
then the generating function for the probability density of y is <P»(s) where
<P{s) is the generating function for p{x), and the probability density of y
is called the n-fold convolution of p(x). The density function which is the
n-fold convolution of p(x) we shall denote by p«(x). Note that

pm(x) = p{x)
and
X

p<ri(x) = p^-'KyMx - y), n = 2,3,...

y=0

Let us suppose that the random variables xi} i = 1, . . . , n, are Poisson

distributed with means p<. The generating function for the Poisson distri¬
bution is given by (3-83). Then, if the xt are independent, the generating
function for the probability density of y = xt + . . . + xn is

(P(s; m)<P(s; y2) . . . <p(s; ju») = e~<*+ ■ ■ ■ = (p(s- w Mn)

(3-90)

which is the generating function for a Poisson distribution with mean

l*i + ••■ + Hn- We have thus proved that if the xf, i = 1,. . ., n, are inde¬
pendent random variables which are Poisson distributed with means ir¬
respectively, then y = x: + . . . + xn is also Poisson distributed with a
mean mi + • • . + m»-
Next let us suppose that we have n independent random variables x,-
each of which has a geometric distribution with mean (1 — p) /p. Then
the probability density of y = xj + . . . + xn is the n-fold convolution of
the geometric distribution with the parameter p. The generating function
for the geometric distribution is given by (3-85). Hence the generating
function of the probability distribution for y is given by
 122
PROBABILITY THEORY AND STOCHASTIC PROCESSES CHAP. 3

(P"(S; p) “ [j- (1-p) J = n> p) (3-91)

Thus 2/ has a negative binomial distribution bN(y; n, P). We have used this
result previously without proof when developing the stuttering Poisson
distribution. We have now provided the proof.
In the above paragraphs, we have shown how to determine the gener¬
ating function of the probability for a finite sum of independent random
variables xi + being given and specified. Let us now study a
somewhat different case where n can also be a random variable. We shall
imagine that each random variable Xi has the same probability density
Pi(x) whose generating function will be denoted by <*(«). The * are as-
sumed to be independent random variables. The generating function of the
p obabihty density function for n, p2(n), will be written (P2(s). The proba-
-Jl ^isV ~ h V ~ Xl + ‘ ’ ‘ + Xn’ where ■?' is a specified non¬
ruLIrV integer
negative

P 0) p(y j) — E p2(ri)p(xi + . . . + xn = j) (3-92)

71 = 0

If (P*(s) is the generating function for p*(j), then

<P*(s) = E P*(j)st = E[E P.(»)p(*i + ... + Xn=j)lsl-

J =° 3=0 Ln = 0 J
or on rearranging the summation signs

<P*^ = ^ P2^ [E PFi + •■• + !„= i)s3J (3-93)

However the quantity in brackets in (3-93) is precisely the generating

function for a* + . . . + for a fixed n> Le.; (p»(s). Thus S S

^(s) = E P2(n)W(s)
n=0
(3-94)
However, by definition
00
<*(«) = E pz(n)s"
n=0
Thus
<?*(«) = <Pt[a»i(*)] (3-95)
t Writt6n aS a function of a function, i.e., (P*(s) is <p2 of
L r' Justobtamed Provides a way for deriving the generating
obtlmthf f! ,StUttepn? Poisson distribution, and hence another way to
obtain the stuttering Poisson itself. For the stuttering Poisson, p2{n) is a
SEC 3-9 CONVOLUTIONS 123

Poisson distribution so that d\,s*; dO whore d is the mean rate at

which demands oeeur (set* 8-88), and is the generating; function for
the geometric distribution. However, the generating function (3-85) for
the geometric distribution must be moditied slightly since' for use in the
stuttering Poisson, the probabilities are stepped up by unity so that the
probability that one unit is demanded is the geometric probability of zero.
Thus (3-Sf>) must in* multiplied by .s' to obtain the correct generating func¬
tion, so that (recall that if is hen* convenient to write p • 1 — v)

Thus tlu* generating function for the stuttering Poisson distribution, which
will be written d)» is

i\ p) r ddi [«o ■ o/ti -wi)!} ; (8-97)

Led us now turn our attention to continuous random variable's. We do

not- expect to be able to deliue a generaling function for the probability
density fix) of some* eonliuuous rantlom variable x such that, when ex¬
panded in a power series, the coeiHcienfs of tin* various powders will yield
the density function of x for all possible* x values, since x is continuous,
instead consider the function

fm * p r« i s.y
««*) / L ( D1 J-‘f(x)dx (3-98)
J »» j« o LJ JJ*

In tlu* expansion (8 98), the' coeflieieut of & is { * lpju,//I where

ju/ • j xf(x) dx (3-99)

The numhe*r p, is e*alle*el theg/th mome*uf of fix) about, the origin. The func¬
tion ‘J(.v), if if e*xists, is <*alle*el the* momtmf g(*ne*rat ing function of f(x) or x,
and it taims the* plaea*, as we* shall se*e*, e>f the* generating functie>n for discrete
variables.'1’ Whe*n x is de*line*d only for non-negative x1 then fF(.v) is also
calle*el the* Laplace transform oi fix). If might be* noted, although we shall
not. prove* it, that a, density function is imi((U<*ly de*tc*rmincd by its moment
generating func*tiom
Le*t Xi, x-i he* two e'ontinuous independent, ranelom variable's with densities
fi(A),fnU'j) and moment generating functions 9a(s) and fhi(,s) respectively.
We*, shall uenv compute the demsity function g(y) and moment generating
function <;(s) for the random variable* y xi L x«. The probability that

* If is also possible* to define moment, generating functions for discrete random var¬
iables, We* leave an analysis of t heses momemt generating functions for discrete random
variables to the* problems.
 124
PROBABILITY THEORY AND STOCHASTIC PROCESSES CHAP. 3

l(T)av^erZ^ vt ffven that Zl lies between Xl and ^ + tel is

few f +v, 1 dXl dXl and dXi = dy^ Here the analysis is slightly dif-
tfan -tf°r tbe Case wh^e both variables are restricted to be non negative
than it is when one or both variables can take on all real values. Hence we
to blnTf r ati°nJ0r the CaSe where both variables are restricted
to be non-negative, since.this case is the more difficult to derive. Thus if
y xx is to be non-negative, xx can range from 0 to y so that

~ J0 ^Xi^2^y ~ ^i) dxi (3-100)

Let us next compute the moment generating function of g(y) in the case
where x1; x2 and hence y must be non-negative. By (3-98)
f“ [v
S(«) h(y ~ Xi)e~ys dxi dy (3-101)
= Jo Jo hiXl)h

^S)fCr?-xwid?d l°,be a double integraI> Wh0se inte^d ia

shownin Fiv ^7 ’ lntegrated °Jer the ^aded region of the sl2/-plane
“ Jlg; 3:7- f8 wntten 111 (3-101), the integral is found by summing
of a,rea
shown m Fig. 3-7. The integral canparallel
equally t0 thebe2:1evaluated
well the ™
axis suchbyassumming

FIGURE 3-7.
 SEC 3-9 CONVOLUTIONS 125

over Xi, infinitesimal strips of area parallel to the y axis such as the one
shown in Fig. 3-7. Thus (3-101) can be written

S(«) = Jq j MMMy — xi)e-y* dy dxi

(3-102)
= Jq fi(xi) j* f2(y - xi)e~y‘ dy dxi
However

/ My ~ xi)e~«s dy = e-*is / Mu)e-W du = e~nsS2(s)

Jx i JO
Thus

8(s) = Ms) Mxi)e-Xia dx! = JFi(s)9:2(s) (3-103)

This proves that the moment generating function for g(y) is the product of
the moment generating functions for fi(xi) and fzixf). Precisely the same
result holds if xi, X2 can take on any real values. The proof is left for
Problem 3-24.
The above discussion shows that, given n independent random variables
Xi with moment generating functions ^-(s), the moment generating function
g($) of the random variable y = xi + ... + «» is
S(«) = SitoftC*) . . . SF«(«) (3-104)
In the special case where each Xi has the same density function f(x) and
moment generating function ^(s), then g (s) = 3rn(s). In this case g(y) is
called the n-fold convolution of f(x), and when the random variables must
be non-negative

g(y) = f(y — yi)f(yi — ya).. . /(yn-i — yn~2)f(yn-2 — x)

fix) dx dyn—2. . . dy (3-105)

The moment generating function for the normal distribution, which will
be written 3ft (s; fi, <r) is

3ft (s; fi. 6 (l/2<r2)(x—/i)2 ^

(3-106)
Q~ (1/2) (mAO2 / g” (l/2tr2) [x2+2(or2a —ja)x]
v/2?r a

= --L e~ (1/2^ W*)*eQ./2a*) (0 J g— (l/2<r2) [x+ (<r2«—/x)]2

v2ir a
= g(l/2)(<r2a2—2^a)
 126 PROBABILITY THEORY AND STOCHASTIC PROCESSES CHAP. 3

If Xi, x2 are two independent random variables which are normally dis¬
tributed with means /u, M2 and standard deviations 01, <r2 respectively, then
the moment generating function of y = xi -f- is
ea/2)(<n!j!-2ws)e(l/2)(«*s!-2^M) _ e(l/2) [(»i«+«‘)»»-2(w+/o)«]

— 31 (s; mi + M2, VV? + of) (3-107)

Therefore, since the moment generating function uniquely defines the

density function, we see that y is normally distributed with moan mi + M2
and variance of + of. ^
“oment generating function or Laplace transform for the gamma
distribution, which will be written T(s; a, /S) is

T(s; oc, “ JMlL 6 &xe~sx dx ua

e-[!+(«/«]« gu
r(« +1) r(« +1)

(1 + l)u' e~ B+<*/£)]« du
r(a + 1)

sV (3-108)
1 +
0/
The above result can be used to provide another way to obtain the
probability density for the time required to incur Q demands when a
Jroisson process generates the demands. Recall that the time between
demands is exponentially distributed. The moment generating function
for^the exponential distribution is 1/[1 + (s/d)] as is seen by setting
“ ., J m • If foe Qth demand occurs at some time t after observation
on the system is begun, then t can be written t = 4 + 4 + . . . 4. tQ where
V if dfmaQ,d °CCUrs at the time ll>the sec°od demand occurs at a time t%
after the first demand, etc. Then each «,■ is exponentially distributed with
th ^ ,meaa- Furtlierinore, the U are independent random variables.
Thus the density function for t is the Q-fold convolution of the exponential
distribution, and its moment generating function is

= T(«; Q-1,0)
1 + (3-109)

Thus the random variable t has a gamma distribution of order Q. This

same result was obtained previously in a different way.
 SEC 3-10 MARKOV PROCESSES DISCRETE IN SPACE AND TIME 127

3-10 Markov Processes Discrete in Space and Time

Consider a system which is observed only at times 4, 4, 4, . . . (where

= 4 + At and A2 > 0 is a finite constant). At these times, the system
will be in one of n discrete states = A specific example of
the general situation just outlined will be that of an inventory system
whose condition is reviewed only at discrete equally spaced intervals of
time. The states of the system referred to above will be the inventory
position of the system. We shall assume that if the system is in state i at
time 4, then the probability that it will be in state j at time 4+i is a# > 0.
We shall suppose that the a# are independent of the past history of the
system and independent of time, so that the dij are constants. Since the
system must be in one of the states j at time 4+i we must have
n

X On = 1 (3-110)
l

The are called transition probabilities, and they are the conditional
probabilities of finding the system in state j at time 4+i if it was in state i at
time 4.
If the system is in state i at time 4, the probability that it is in state j
at time 4+2 is
n
a\f = aikakj (3-111)
k=i

since from 4 to 4+1 it can undergo a transition from i to any state h with
probability da (which may be 0) and from state Jc at time 4+1 to state j at
time 4+2 with probability a*y. The probability of transitions from i to h to j
is diicCLkj, and therefore the probability of moving from i to j from time 4 to
4+2 is found by summing over all h.
Let pr(i) be the probability that the system is in state i at time 4. Then
Pr+i(j) is given by
n
Pr+l(j) == ^ y Pr(i)dij (3-112)
i=l

and from (3-111)

n n n n
iPr+fcO) == ^ / Pr('Z')dij^ = ^ ^ ^ y Pr(i)dikdkj = V ^ pr+i(Ic)dkj (3-113)
i=l i = l &= 1 k=l

It follows thus that if we are given the probabilities p0(i) that the system
is in state i at time to, and the transition probabilities a#, then it is possible
to compute the probabilities that the system is in state j at time
4, t = 1, 2,.... The type of process we have been describing is called a
 128 PROBABILITY THEORY AND STOCHASTIC PROCESSES CHAP. 3

Markov process which is discrete in space and time (it is also referred to as
a Markov chain). It is said to be discrete in space and time because the
states of the system (referred to as the space variable) are discrete, i.e. can
be represented by integers, and because we are interested in the states of
the system only at discrete points in time. We cannot say anything about
the system at times other than to, h, k,. . .. The characteristic feature of a
Markov process is that the probability that the system is in a given state
at time tr+l depends only on the state that it was in at time tr, and the
transition probability, and is independent of the previous history of the
system (i.e., how it got to the state it is in at tr).
It might be expected that as time goes on, the probabilities pJj) will
become less and less influenced by the initial probabilities p0(j). We shah
say that the system is in a steady state if the probabilities pr(j) do not
change with r, i.e.,
Pr+l{JJ Pr\jj ior ail j

If a steady state exists, we shall denote by p{j) the probability that in

steady state the system is in state j. These probabilities can be found by
wn ing p(j) pr(j) — Pr+iO) in (3-112) to yield the following set of n
Homogeneous linear equations in n unknowns
n
Vid) = p(i)ciij, j = 1, ,n (3-114)
1=1

thT^J °f eqUaf?DSWiI1 n0t Uniquely determine the PU). However, since

to uSty^Te be m °ne and °nly °ne °f the StateS’ the p(j) must sum

J2 p(f) 1
y=i
= (3-115)

With this additional constraint, the p(j) will be uniquely determined for
cases of interest to us (this is proved below).

nr^abSwT laiCT :°r\0nly haVe an interest in the steady state

frntr,b t P9\’ andJnot m the transient behavior of the system starting
from some imtial condition. Furthermore, we shah only be interested in

anv otherTaT^Tfr 6Very St&te °an ultimately be reached from

reTnWf 1 W1 n0t necessarily be true that every state can be
cient numr ? fVen.fat.® “ a sin§le time step, but if one allows a suffi-
anv dven stat J 6 P°SSible t0 reach GVery °ther state from
m ZIZZT i uV pr0cess 0f this type is called an irreducible
uair nfL? J°r SUCh a pr°CeSS> there exists an integer N for each
S the CeS ' / !U that ^ > °’ Where is t^ probability that
the system is m state * at time tr it will be in state j at time tr+N.
For an irreducible Markov process with only a finite number of states
 SEC 3-n MARKOV PROCESSES CONTINUOUS IN TIME 129

it is a well known result from the theory of Markov processes that the p(j)
defined by (3-114), (3-115) exist, are unique, and each p(j) > 0. We shall
prove this without making use of the theory of Markov processes. How¬
ever, the proof does require some understanding of linear algebra, which
may not be familiar to the reader. An understanding of the proof is not
needed in later work. It is only necessary for the reader to recall that the
proof demonstrates the uniqueness of the p(j). To proceed with the proof,
let the matrix A = | |at-j| | be an nth order matrix containing the transition
probabilities, and p = [p(l),. . . , p(n)] be a row vector containing the
p(j). Then (3-114) can be written p(/„ - A) = 0 where /„ is the identity
matrix of order n. By (3-110), each row of A sums to unity and each row
of ln — A sums to zero. Thus the columns of — A are linearly dependent,
and \ln A[, the determinant of ln - A, vanishes. Thus there exist p(J),
not all zero, which satisfy p(Jn — A) = 0.
Let us now drop the nth equation in (3-114). The resulting set of n — 1
equations in n variables can be written p„_i(/»_! - An_a) = p(n)an, where
Pn-1 is a vector containing the first n — 1 components of p, /„_j is the iden¬
tity matrix of order n — 1, An_i is the submatrix formed from A by crossing
off the last row and column, and an = [a^, . . ., an,n-i]. Note that at
least one component of a, is positive, for otherwise it would be impossible
to ever leave state n. Thus at least one row of A„_i sums to a non-negative
number less than unity. From the theory of Leontief matrices, we know
that (In—i — An-i)-1 exists and can be written as a power series. Thus

Pn-i = p(A)a„[/„_! + A„_i + An_i + . . . ] (3-116)

For a given p(n), the vector p„_i is uniquely determined. If p(n) = 0 all
p(j) vanish and hence cannot sum to unity. Thus p(n) > 0. When
p(n) > 0, each component of p„_x will be positive, since for each i and j
there exists an N such that element aiP in AN is positive. Hence, on apply-
ing (3-115), we see that the p(j') exist, are uniquely determined, and each
p(j) is positive. This proves what we wanted to show.

3-11 Markov Processes Discrete in Space and

Continuous in Time—Queuing

In the previous section, we assumed that the state of the system was ob¬
served only at discrete points in time. We were interested in determining
the steady state probabilities that, at the time of observation, the system
was in one of a finite number of different discrete states. Now we wish to
study situations where the system is continually under observation and
the state of the system is known at each instant of time. Again we shall
assume that the states are discrete and can be represented by integers,-
 130 PROBABILITY THEORY AND STOCHASTIC PROCESSES CHAP. 3

however, we now permit an infinite number of discrete states. We wish to

be able to compute the probability that the system is in any one of the
states at any instant of time, and here also we shall usually be interested
only in the steady state probabilities. The assumptions in this section
differ from those of the previous section in that time is now a continuous
variable and we allow for an infinite number of discrete states. The mate¬
rial to be considered will later be applicable to what will be called trans¬
actions reporting inventory systems. Again the states will refer to
inventory levels in applications which will be of interest to us.
If the system is in state i at time t, we shall denote by f(i, t;j, r) the
probability that it is in state j at time t > t. As in the previous section
/(b b 3i T) is called a transition probability. When the process is a Markov
process, then the probability p(j; t) that the system is in state j at timp r
depends only on the probabilities p(i; t) that the system was in state i at
time t, and the transition probabilities f(i, t;j, r); it does not in any way
depend on what has happened at times previous to t except through the
Then
oo

P(j; 0=1] p(i; t;j, t) (3-117)

i=0

where in writing (3-117) we have allowed for an infinite number of discrete

states. If the system is in state i at time t, it must be in some state j at
time t. Thus it must be true that
00

X)/(b t;j, 0 = 1 (3-118)

j=0

Given the times t < n < r2 and the transition probabilities

2), t;j, ti), ri]j7 r2)

then
” 00
P(i; rO = 2] P(i; t;j, r2) = P(k; n)/(£, Tl;j, r2)
t==0 k=0

= X] P(y> 4X ^ T1 )f(k, Ti;j r2)~1

i=0 L/fc = 0 J

Thus we must have for any times t < n < t2

m, t;j, r2) = Kh t; k, n)/(*, n;j, r2) (3-119)

*=o

For problems of physical interest, it is usually true that as r approaches

h J\i) t, h r) approaches zero when j ^ % and unity when j = i, i.e., as
 SEC 3-11 MARKOV PROCESSES CONTINUOUS IN TIME 131

T t the probability of finding a change in the state of the system also

approaches zero, and the probability of finding it in the same state ap¬
proaches unity. Although/(f, t;j, r) approaches zero as r £ when i ^ j,
the ratio f(i, t;j, r)/(r — 2) can approach a finite value different from
zero, which we shall denote by a^t), i.e.,

lim—-i'7’^ = aij{t), i^j (3-120)

When r differs infinitesimally from t, i.e., r = t + dt, then

f(h t;j, t + dt) = —t’3’dtt + dt) dt = cm® dt, i ^ j

The aij(t) dt are called the infinitesimal transition probabilities, and the
aa(t) are called the transition densities. Furthermore, from (3-118)

f(i, t; i, t + dt) = 1 — X) (0 (3-121)

y=o
y
By use of the infinitesimal transition probabilities, it is possible to obtain
a set of differential-difference equations which determine the state proba¬
bilities p(i; t). By (3-117), (3-120), (3-121)

p(i; t + dt) = p(i; t) 1 dt ^ ^ dij (t) + X p0'; dt

i=0 y-o
J
or

t + dt) — p(i; Q
dt ~p(i; t) X a,ij(t) + X p(i; f)ay<(0
J =0 y=o
y

However, in the limit as approaches zero, the quantity on the left in the
above equation becomes dp(i; t) /dt, so that

^jr1 = ~v(i; t) X ai}(t) + X PUl tM), i = 0, 1, 2,. . . (3-122)

y~o y=o
y^ y 5^*

Here we have an infinite set of differential-difference equations (if there

are an infinite number of states). They are differential equations with
respect to time and difference equations with respect to the states. If we
are given the p(j; to) at some time to, we expect intuitively that the equa¬
tions (3-122) will determine uniquely the p(j; t) at all future times.
The cases which are the easiest to solve and which are also of practical
interest are those for which Poisson processes generate transitions between
 132 PROBABILITY THEORY AND STOCHASTIC PROCESSES CHAP. 3

states. By this it is meant that Poisson processes generate certain events

such as the occurrence of demands or arrivals of orders in inventory sys¬
tems. To move from state i to state j will require the occurrence of one or
more of these events. Thus all the probabilities f(i, t;j, r) will be Poisson
densities, or weighted sums and products of them. The important thing
to note is that/0, t; j, t + At)/At will involve At or a higher power of At and
hence approaches zero as At -»0 in every case except that where the transi-
tion from i to j involves only the occurrence of a single Poisson event. In
this case the ratio approaches a constant fii3 which is the mean rate at which
the Poisson event under consideration occurs. Thus all the aiS will be
zero except for transitions between states involving only the occurrence
of a single Poisson event. We leave the detailed proof for the general case
to Problem 3-25. The method of proof will be illustrated for a particular
example later. Therefore, we can write

^7 (a constant) if a transition from i to j involves the occurrence

_ _ J °f only a single Poisson event
0 if a transition from i to j requires the occurrence
- more than one Poisson event (3-123)
where 1/jty is the mean time between transitions from i to j. Note that the
ay do not depend on t. For most systems discussed in subsequent chapters
it is true that for any given, fixed i, there will be only one or two values of
3 tor which a{j ^ 0, i.e., transitions from state i can be made only to one or
two other states in an infinitesimal time interval of length dt.
The ®5uatl0ns (3-122) teH how the state probabilities p(i; t) change with
time. Usually we shall be interested in cases for which the p(*; t) tend
toward steady state values which do not change with time. These steady
state values will be denoted by p(j). Since the steady state probabilities
must satisfy (3-122), and since they are independent of time, i.e.
dp(i)/dt = 0, then we see from (3-122) that the steady state probabilities
must satisfy

ao- P0)d;i, i — 0, 1, 2,. . ., j; p(J) 1 (3-124)

1 3^' J=0

, 100,tr“sitions1are Seated by Poisson processes, the a*- are given by

(-123). Here we have a set of difference equations (perhaps an infinite set)
1 s°lve for the state probabilities
The equations (3-124) can be given a simple intuitive explanation. If the
probability that the system is in state i is to remain a constant, then the
probability that the system leaves state i in time dt must be exactly equal
to the probability that the system moves from some other state to state i.
 SEC 3-11 MARKOV PROCESSES CONTINUOUS—QUEUING 133

The probability that the system leaves state i in time dt is p(i) a^dt,
and the probability that it changes from some other state to i is
dt. On equating these two terms and cancelling dt, we obtain
(3-124) which may be called the balance equations. The left-hand side of
(3-124) can be interpreted as the rate of change of p(i) arising from transi¬
tions out of state i, and the right-hand side the rate of change of p(i) from
transitions into p(i). As noted above, these two rates must be equal.
Markov processes continuous in time and discrete in space which are
concerned with systems that service in some way randomly generated
demands are often referred to as queuing problems. Hence, when the
theory is applied to inventory systems, we often say that queuing theory
is being applied to study the inventory system.
It is important to note that the theory developed above can easily be
generalized to systems where the description of each state requires the
specification of two or more non-negative integers rather than merely a
single integer. For example, suppose that two non-negative integers are
required to describe each state. Then let p(i, j; t) be the probability that
the system is in state (£, j) at time t, and (kj;mn{t) dt be the probability that
if the system is in state (i, j) at time t, it will be in state (m, n) at time
t + dt. With these definitions, (3-122) becomes

dp(i, i; i0
dt ~"P(h 3) 0 EE <'Hj;mn(t)
m=0n=0
ij 7*mn

^ y ^ y P(P^} W) t) ^mn;ij (0 (3” 125)

m—0n=0
ij 7*mn

EXAMPLE Perhaps the simplest example of a queuing model is that of a

single service facility (perhaps a repair station) in which the time required
for service is a random variable with an exponential distribution, i.e., the
probability that service takes longer than t is e~^ (a Poisson process gen¬
erates the service times). The arrivals of customers at the service facility
are also assumed to be generated by a Poisson process, the mean arrival
rate being X. A customer arrives at the facility and goes into service im¬
mediately if the facility is not busy; otherwise the customer waits in line.
Customers are serviced in order of arrival. Let the states refer to the
number of customers in the system, i.e., in service or in line (note that only
one customer can be in service). Thus there are an infinite number of
states. We shall assume that customers arrive one at a time. If the system
is in state i at time t (i.e. i customers in the system), then it can be in state
j > i at time r if one of the following sets of events occurs: 1) there are
j — i arrivals in the time r — t and no units are serviced, 2) there are
 PROBABILITY THEORY AND STOCHASTIC PROCESSES CHAP. 3

j ~ i + 1 arrivaIs and one completes service, 3) there are j-i + 2

arrivals and 2 units complete service, etc. Thus
00

Kh t;j, r) = gp[j -i + k; X(r - 0M*|i - i + k, i, r - t], j> i

where
V\Mi ~~ i + k, i} n, r — t]

is the conditional probability that k customers are serviced in time r - t

given that there were j — t — k arrivals, that i were in the system origi-
nally, and that n is the mean servicing rate. However

]jm P(xj XAQ = fX, x = 1

At—>0 At \0, x = 2, 3 . . .

Thus, when j > i, ai} = 0 unless j = i + 1. Furthermore

= Xp[0|l, i, yU, 0]
Tf probability that no units are serviced in a time interval
l Sf 1Sn1’Slnce eyen/ a customer is in service, the probability that
' *° C°mplete Se"i0e “ «-"■ Consequently, we

3 =i+ 1
all other j > i
Similarly, if j < i

all other j < %

StioLTo stltef V t 01 ^ P1 that the aii are Zer° eXCept for tran-
i e ^ arrtfr ^ invo ve the <*™*ce of just a single Poisson event,
i ;a nT1Val °J tbe completion of service, so that transitions in the time
d can occur only to adjacent states. The states of the system and the
allowable transitions can be represented in a diagram such as Fig. 3-8.

FIGURE 3-8.
 SEC 3-12 OTHER TYPES OF MARKOV PROCESSES 135

We can immediately write down the balance equations (3-124) for the
steady state probabilities p(j) in this special case; they are

= Xp(0); np(n + 1) + \p(n - 1) = (X + f*)p(n), n = 1, 2, . . .

Then if p = \/p we obtain, on sequential computation beginning with
P( 1) = PP(0)> PM = Pnp(0), n = 1, 2, . . . . Now it must be true that the
state probabilities sum to unity, i.e.,

*0) |> = KO)^] = 1, P < 1

Thus if p < 1, the steady state probabilities are distributed according to

the geometric distribution, i.e.,

P(j) = p’( 1 — p), 3 = 0, 1, 2, . ..

If p > 1, i.e., the mean servicing rate is not greater than the mean arrival
rate, there does not exist any steady state and the expected length of the
line is continually increasing.

3-12 Other Types of Markov Processes

There are two other types of Markov processes which are of interest. One
is that which is discrete in time and continuous in the space (state) variable,
and the other is that which is continuous in both space and time. We shall
discuss very briefly Markov processes which are discrete in time and con¬
tinuous in space, but shall do no more than merely mention something
about the nature of those continuous in space and time.
To introduce Markov processes which are discrete in time and continuous
in space, we imagine a system which is observed only at times U,
r = 0, 1, 2,..., where tr = 4-i + At, At > 0 being a finite constant, the
state of which can be described by a continuous non-negative variable x.
Let gT(x) dx be the probability that the state of the system lies between x
and x + dx at time tr, andf(x\y) dx be the probability that the state of the
system lies between x and x + dx at time tr+1 > tr given that the state of
the system was y at time tr. Then

gr+i(x) = f(x\y)gr(y) dy (3-126)

If a steady state density g(x) exists such that g(x) dx is the steady state
probability that the system lies between x and x + dx, then g(x) is given by

g(x) = JQ f(x\y)g(y) dy (3-127)

 136 PROBABILITY THEORY AND STOCHASTIC PROCESSES CHAP. 3

Equation (3-127) is a homogeneous Fredholm integral equation. In gen¬

eral, the solution of such an equation can be quite difficult. Later, in our
discussion of periodic review inventory systems, we shall have occasion to
solve (3-127) in an especially simple case.
Markov processes, which are continuous in space and time, are often
referred to as diffusion processes. The mathematical analysis of such
systems is more complicated than those we have studied above, since the
determination of the state probabilities involves the solution of partial
differential equations. By use of Markov processes which are continuous
in space and time, it is possible to develop a model for generating demands,
in the case where demand is assumed to be a continuous random variable,
along the same lines as that of the Poisson process for discrete demands.
The result is a normal distribution with mean Xf and variance Dt, where X
is the mean rate of demand and D is a constant which can be specified
arbitrarily. We shall not go through the analysis, since the notion of a
continuous demand variable will be introduced in another way.

3-13 Properties of the Poisson Distribution

We shall make use of the Poisson distribution quite frequently in later

chapters. In Appendix 3 are tabulated a group of properties of the Poisson
distribution which we shall have need for. This section will be devoted to
showing how a number of these properties are derived. The remaining
derivations will be left to the problems. Recall that the definition of the
Poisson probability density is given by (3-13). Many of the relations will be
expressed in terms of the complementary cumulative function

P(x; m) = p(j] /i) (3-128)

j=x

rather than the cumulative function. This is done because the more com-
piete tables of the Poisson distribution [2,7] tabulate the complementary
cumulative function rather than the cumulative, and also because, if compu¬
tations are being made with a digital computer, it is often easier to evaluate
rather than the cumulative function. The purpose of the deriva¬
tions to follow is to express all relations involving the Poisson distribution
directly in terms of p(x; ft) and P(x; ft) so that numerical computations
can be made easily using the tables in [2,7], Even when the computations
are to be made on a digital computer, it will usually be desirable to make
hese transformations, since the resulting expressions will be easier to
handle on the computer than the expressions in their original form [which
may involve sums of the P(j; ft) or integrals over time of the p(j; n) or
P(j; wJ- The relations derived in the following paragraphs will hold for all
 SEC. 3-13 PROPERTIES OF THE POISSON DISTRIBUTION 137

non-negative integers r provided that we use the convention p(j; y) = 0

and P(j; y) = 1 when j is a negative integer. Even though r is to be non¬
negative, it is possible in some of the expressions to encounter negative
arguments for p(j; y) and P(j; y) if r is small enough. The expressions
will be correct in such cases if the above convention is used.
From (3-13)

xp(x;*) = ^7 c—
(3-129)
= - l;*t), X = 1,2,

This is property 1 of Appendix 3. In the future we shall refer to the proper¬

ties as, for example, A3-5, which means property 5 of Appendix 3.
Sums of the following sort are often needed

Mm(r) = ^2jmp(j; m), m = 0, 1, 2, . . . (3-130)

3~T

An especially easy way to reduce such sums to a simple form is through the
use of a recurrence relation. Note that

Mm(r) = Y^jmp{j-, p.) = - i;aO

3-r j=r

00 00 m—1 / i \

= T (* +
k = r— 1
/*) = /* E E (m r ) knp(k> /*)
k — r— 1 n=0 \ n /

171— 1 / -j \

= ^E(m~ ) Mn(r - 1), m=l, 2,... (3-131)

This is A3-2. In particular

so

wW = E-7'pU; m) = Mo(r - 1) = /iP(r - 1; /i) (3-132)

j=T

which is A3-3.
In determining the expected number of backorders for some of the future
models to be discussed, it is necessary to evaluate numerically the ex¬
pression

E O' ~ f)p(i; aO

This can be converted to a form which is easy to evaluate by noting that

138 PROBABILITY THEORY AND STOCHASTIC PROCESSES CHAP. 3

CO CO

Z 0 — r)p(i; aO = ^2jp(j; aO - rP(r; m)

'=r y=r (3-133)
= nP(r - 1; n) - rP(r; n)
by (3-132). This is A3-10.
We also have need for sums of the form

Qm(r) = YJjmP(i)ti
y=r

These are also most easily reduced to computable form through the de¬
velopment of a recurrence relation. To obtain such a relation note that
00

Z [0- i)m+l-jm+1]P(j;»)
j-r

= (r - l)”*+ip(r; /£) - rm+^p{r-, /*) - (r + l)“+ip(r + 1; y) - . . .

= (r - l)m+lP(r; n) - ^n+1(r) (3-134)

where Am+i(r) is given by (3-130). Also, expanding by the binomial the¬

orem, we obtain
oo
Z tO - i)m+1 - im+1]P(i; m)
j=r

= Zz
y=ri = 0
(-l)m+1“{m t X) j4P(i; /i)
\ 1 /
(3-135)

Equating (3-134), (3-135), we find that

(-i)—‘("+ 1) aw

+ fim+i(r) - (r - l)m+1P(r; fi) J, m> 1 (3-136)

which is A3-7.
In order to use (3-136) we must evaluate

O0(r)=ZPO» (3-137)
i—r

This is easily done by writing Qo(r) as follows

 SEC 3-13 PROPERTIES OF THE POISSON DISTRIBUTION 139

Qo(V) = p(r; ti) + p(r + 1; m) + p(r + 2; p) + . . .

p(r + 1; /x) + p(r + 2; m) + • . .
p(r + 2; n) + . . .

= P(r', u) + 2p(r + 1; m) + 3p(r + 2; ju) + • - .

oo
= E0'-r + 1 Mj; *) = juP(r - 1; + (1 - r)P(r; m) (3-138)
y=r

from (3-133). This is A3-6.

By use of (3-138), (3-136), (3-131), and (3-132) we see that

fii(r) = EiPtf; #*) = I [^o(r) + At2(r) - (r — l)2P(r; aO]

y=r •

= |P(r- l^)+^P(r;M)+^P(r- l;/i) + ^P(r-2;M)

(r ~ l)2
P(r;/*)

= ^ P(r - 2; aO + »P(r - 1; a*) - P(r; m) (3-139)

This is A3-8.
Several integrals over time involving the Poisson distribution will also
be needed. The simplest of these is

Jo Jo rl A Jo rl

= 1i r" ^r+i
I XT PXT
(3-140)
A _ (r + 1)! 6 o +Jo (r + 1)

= -P(r+ 1;XT)

which follows from repeated integration by parts. This is A3-16. It imme¬

diately follows that

e~z dx
(n + r)!

p^p^ + r+ijxr), « = 0,1,2, ... (3-141)

from (3-140). This is A3-17.

 140 PROBABILITY THEORY AND STOCHASTIC PROCESSES CHAP. 3

Finally

f”P(r; \t) dt

fn+l ” f* fn+1
——- P(r; Xf) — X / —— 'p(r — 1; \t) dt (3-142)
n+ 1 v ’ o yow+1

STT Pfr; XT> ~ r- D!P(,‘ + ' + ‘; XT>

on integration by parts setting du = t* dt, along with the use of A3-15
and (3-141). This is A3-19.

3-14 The Normal Distribution

The normal density function n(x; p, <r) is defined by (3-22). The proba¬
bility that the random variable x lies in (he interval ,r, < .r < j-» is given
by the integral of n(x; p, or) from xx to x«. In (3-22), p is t he expected value
of x and or2 is the variance of x. A normal distribution with p 0 and
v = 1 is called the standardized normal distribution, and its density func¬
tion given by (3-24) will always he denoted by <f>{w). The complementary
cumulative of <t>(w) is

<M//) dt/ (3-143)

Almost all handbooks of mathematical tables [2, for example | provide

tables of <j>(w) and <f>(yj) (or a function from which <b{tr) can be easily
computed).
Given a table of <t>(to)} it is easy to mako computations with the
normal distribution. Note that

n(x-, p, a) =
(3-144)

s fm /,. \
n(x; p, a) dx = J cj>(w) dw = <I> f ^^ J
* (^7^) (3-145)

where w = (z - p)/a.
It is often easier to work analytically with models of inventorv systems
if all variables can be treated as continuous. This allows one to eliminate
he problems caused by discreteness and to take derivatives instead of
dealing with differences, etc. In many cases the demand rate for an inven-
 SEC. 3-14 THE NORMAL DISTRIBUTION 141

tory system will be sufficiently high that the problems caused by discrete¬
ness can be ignored and all variables can be treated as continuous. When
demand is treated as continuous, the most frequently used distribution to
describe the quantity demanded in a given time is the normal distribution.
There are several reasons why the normal distribution holds a central
place in working with continuous random variables representing the de¬
mands on the system. One is that the normal distribution is especially
easy to work with and is well tabulated. Another and more important
reason is that empirical studies have shown that, quite often, the normal
distribution seems to approximate very well the demand distributions
over the relevant time intervals which are encountered in practice. A final
theoretical reason for the importance of the normal distribution is that all
the distributions studied in this chapter, the binomial, negative binomial,
Poisson, gamma, and stuttering Poisson approach, as the mean of each of
the distributions approaches infinity, a normal distribution whose mean
and variance are the mean and variance of the appropriate distribution
under consideration. We shall not prove this theorem since in itself it is
not of great interest to us. The more important questions are how fast
each of the distributions approaches the normal distribution and how
much error can be involved if the normal distribution is used to approxi¬
mate one of these distributions with a given mean and variance. These
questions are in general quite difficult to answer (in fact they have not
been answered completely), and we shall not attempt to present any gen¬
eral discussion of them.
It is important, however, to recognize that for large means, the Poisson
distribution can be approximated by the normal, since tables of the Poisson
distribution are available only up to ft = 100. Por larger ft values, the
normal approximation must be used. It may be desirable to say a little
more about the normal approximation to the Poisson, because at first
glance it may not be clear what is meant, inasmuch as the Poisson distribu¬
tion is discrete and the normal distribution is continuous.
Recall that p(x; ft) is the Poisson probability of exactly x when ft is the
mean. For a given ft, a plot of p(x; ft) could be represented by a series of
vertical fines as shown in Fig. 3-9. Now it is possible to represent proba¬
bility as areas in Fig. 3-9 also if we draw in rectangles as shown. The length
of the base of each rectangle is unity and it extends from x — § to x +
The probability of x is then the area of the rectangle which includes x since
the length of the base of the rectangle is unity and its height is p(x; ft).
Thus we can write
00

p{x] ft) = p(x; ft) Ax; P(x;») = p(y- M) ay

y=x

where Ax = Ay = 1 is the length of the base of the rectangle in Fig. 3-9.

142 PROBABILITY THEORY AND STOCHASTIC PROCESSES CHAP. 3

FIGURE 3-9.

Note that p(x; p) is the area under the curve from x — § to x + and
P(x; jj) is the area under the curve from x — § to <*>.
The theorem stated above says that as fi —> co for the Poisson distribu-
tion, the broken curve outlining the area of Fig. 3*-9 approaches a normal
curve and

p(x;/i) V/*); P(X; p) ->$ (3-146)

Note that the mean and variance of the normal distribution are the
as the mean and variance of p(x; p).
There is one other point related to approximating the Poisson distribu-
tion with the normal which deserves attention. Recall that P(0; p) = 1.
Hence if we are approximating P(x; p) by $(a: — p/^p), then we are saying
that to the accuracy required in the problem, we can assume that

i.e., the area under the normal curve to the left of x = 0 is negligible. This
will of course become closer and closer to being correct as p increases. This
point will arise a number of times in later work.
When approximating p(x; p) with the normal distribution for finite p,
increased accuracy can be obtained over using (3-146) if one notes that in
Fig. 3-9, p(x; p) is the area from » - } to * + | and P(x; p) is the area
from x — % to infinity. Thus for finite p, more suitable approximations are

1 /•*+( 1/2)
p(x; p)
Jx-(l/2)
(3-147)
= $
)
 SEC. 3-15 PROPERTIES OF THE NORMAL DISTRIBUTION

x i M
(3-148)
^ .

Often it is suggested that the normal approximation to the Poisson will

be sufficiently accurate whenever n > 25. One must be somewhat careful
of what one means by sufficiently accurate, since the error depends on the
range of x which is of interest. The approximation is best for x near the
mean and becomes progressively worse as x increase towards infinity (i.e.,
as one moves out on the tail of the distribution). However, for practical
work, one usually need not be overly concerned with precisely how well
the normal distribution approximates the Poisson, since the Poisson dis¬
tribution itself will only be an approximate representation of the real world
demand distribution. Probably, for most practical applications, the normal
approximation to the Poisson will give sufficient accuracy if /i > 25. How¬
ever, since Poisson tables [7] are available which go up to ^ = 100, there
is no necessity for using the normal approximation until ft > 100.
Because of (3-146) we see that as t —»°o

-(l/2)(i-Xf)VXf
(3-149)
V2xX tl'2

In the event that the process generating demands is an Erlang of order n

instead of Poison, or if the process generating demands is Poisson but the
number of units demanded per demand is a random variable, or if the lead
time is a random variable and we are dealing with the marginal distribution
of lead time demand, then it will still be true that the density function for
the number of demands in a time t approaches a normal distribution with
mean Xt where X is the mean rate at which demands occur. For this more
general case, though, the variance need not be Xt. It can, however, be
written Dt where D is a number which may differ from X. The proportion¬
ality of the variance to time remains unchanged. In general, then, when
representing the demand in a time ( by a normal distribution, we shall
write the density function as

n(x; Xi, VWt) = e-a/8)C*-#>vz>( (3.150)

and if the normal density is an approximation to the Poisson, D = X.

3-15 Properties of the Normal Distribution

We shall need certain properties of the normal distribution, just as the

corresponding properties of the Poisson distribution were needed, to con-
from (3-152). This is A4-7.
 SEC 3-15 PROPERTIES OF THE NORMAL DISTRIBUTION
145

We shall also have need for integrals of the form

Mx, Th T2) = )V-d.

= Jti CDO1/2 ^ [(Di)1/2J dt’ Tl’ T* > 0 ^3'156)

where n is an integer. Consider first J0(x, Th T2). This integral is some-

what more difficult to evaluate than those studied above. If
X — \t X
(* ~ A01
^3

y(t) =
II

(D01/2’ (DO1/2 2D1/2i3/2j ai (3-157)

Then
CTi
1 Vx — \t
Jo(x, Tx, T2)
-L CDtyi*4,, imi!\

1 /ct* *r*
r-
=jnr X (x - A01
X,M "L(D01/2JL (DO1/2 22)1/2^3/2J

i f* (x ~ AO (3-158)
'a; - XH
r, 2D1/2^2 0 .(DO1/2J ®

1 fvW , , 1 (T.
= -- / <j>(w) dw - — <t>[y(t)] dt
A Jv(.Ti) XK J jri <
We now observe that

dJa(x, Tlt T2) = _ l fCTt y(t)

dx D JTl t 4>[y(t)'] dt (3-159)

Substitution of (3-159) into (3-158) yields

Jo
««»)

To determine Jo we must solve the differential equation (3-160) Multi¬

ply both sides of (3-160) by e-<*»/*>. Then we see that

i &-»•»■*] - (3.161)

On integrating both sides from x to a> and noting that

lim e-^mjoix, Th T2) = 0

146 PROBABILITY THEORY AND STOCHASTIC PROCESSES CHAP. 3

we obtain

Jo(x} Tij T2) —

9
— p2\xlD
De f <3-i62>
Now

/" - jf [w<*>* «»*»

can be considered to be a double integral integrated over the shaded area
of the p£-plane shown in Fig. 3-10. As written in (3-163), the integral is
found by summing elementary strips parallel to the p axis. The integral
can be evaluated by adding up elementary strips parallel to the £ axis, like
the one shown in Fig. 3-10. The area of one of these strips is
f(DTyi*p+\T
<t>(p) dp I *-<**/») di

e-(2\x/D)
= ^(p)dp[|— __ -±:
D e-(2\*T/D)e- (2\Tll2p/Dlli)
2X __
so

I"
D r(2\x/D)$ |"__ J2 e- (2\2T/D) e-(2X!Ti/2p/-D1/2)^(p)
2X L(DT)“!] 2X /'
7(*- \T)/(DTy 2
(3-164)

FIGURE 3-10.
 SEC. 3-15 PROPERTIES OF THE NORMAL DISTRIBUTION 147

/(i-XT)/(Z)T)Ui

y/2~ &>*+ (4xri'VD‘'s)p+ (4x*r/D)]e2x»r/2)

]{x-\T)/(.DT)M V2x

- g2\sr/D ^== e_W/2)[o+(2xr/(Dr)i«)]i

/(»-xn/@T)»

e2X>T/D$ fa: + XZH

L(dt)i/2J (3-165)

On substitution of (3-165) into (3-164) and (3-164) into (3-162), we obtain

finally

Jo(x, Th TO = JT0(z, TO - W9(x, TO

where

W0(x, T) = i $ T^ Xrl _ I *»,/** f^+ Xr-1

^ ; XWLPT)1/2J X6 #L(^1/2J (3_166)
which is A4-9.
Consider next the evaluation of Jx(x, Th TO. As a first step, we integrate
Jo(x, Th TO by parts writing do = t~112 dt. This yields

Jo(x, Th TO =

* [(09“]n + D Tl’ Ti> ~ f> X, Tyui * [<SJk] *

(3-167)
Next observe from (3-159) that

dJp(x, Tu TQ _x_ [Tt t-vi , fa; - \H , X

D Jti D1/2 [(D^)1/2J ^ ^ D J°(X) ^ (3-168)

After substitution of (3-168) into (3-167) and solving for Jt(x, Tx T2) we
obtain ’

T (T m rp \ 2(Dt)112 rx — xni-
Ux,Ti,T0-—

dJo(x, 7i, T2)

^[l + ^]jo(x,ThTO (3-169)
X2 dx
148 PROBABILITY THEORY AND STOCHASTIC PROCESSES CHAP. 3

We have now expressed Ji(x} Th T2) in terms of known quantities. It

remains to give an explicit form for Ji(x, Th T2). From (3-166)

= —- * _ 2 fx + XTl
dx \(DT)u*vl(DTy/*] D liDT)1'*]

+ \(DT)112 eAxlD(t> [(£>r)i/2] (3-170)

Substitution of (3-166) and (3-170) into (3-169) yields

JiQc, Th T2) = Wi(x, T2) - W1(x, Tx)

where

2(DT)U> fx - XT'
Wx(x, T)
X2

_ e2\x/D^ x + XT
,{DT) v* ]}
, pu*x , Fx - XT 1 . 2z fa; + X7H
+ x2rw2^ [(Dr)1'2] + x2® $Lcdt)i/2J

x3ri/2 e *L(-D7y/2J
or

+ 17!(^-25,i'’)*[^] (3-m)
+ e,,»(-L) (. _ 5) . [|d_-] - gs ^ [^]
which on using A4-18, is A4-10.
By use of the integration by parts trick, one can express Jn+1(2, Ti, T2)
in terms of Jn(x, Th T2) and JTly T2), thus obtaining a recurrence
relation for the J’s. Integration of Jn(x, Th T2) by parts with dv = tn~w dt
yields

Jn+i(x, Tx, Ti)

2(ptyiH* x - xnT*
X2 ’’L(Z)01/2JrI
C2w 4- 11D t2
+ --T~-
X2 Jn(x, Tx, Ti) + ^ Jn-i(x, Tx, Ti), n = 0, 1, 2,. . . (3-172)

which is A4-11.
 REFERENCES 149

Integrals of the form

Rn(x, Ti, T2) = « [^y^] dt, n = 0,1,2,. .. (3-173)

will also be needed. On integrating (3-173) by parts with do = tndt, we see

that

Rn(x, T1} T2)

1
tn+1$
x — xn T*
n+ 1

2(n + 1) ^2) 2(72, -j- i) ^ ^2) (3-174)

which is A4-14. Hence, once the J}s are known, the R’s can be evaluated
using (3-174). For example, from A4-9, A4-10 (which were derived above)

Ro(x, Ti, T2) = V0(z, T2) - V0(x, Ti)

where

which is A4-16.

REFERENCES

1. Bharucha-Reid, A. T., Elements of Markov Processes and Their Applica¬

tions. New York: McGraw-Hill Book Co., Inc., 1960.
2. Burrington, R. S. and D. C. May, Handbook of Probability and Statistics
with Tables. Sandusky, Ohio: Handbook Publishers, Inc., 1953.
3. Feller, W., An Introduction to Probability Theory and Its Applications
2nd ed. New York: John Wiley and Sons, Inc., 1957.
 150 PROBABILITY THEORY AND STOCHASTIC PROCESSES CHAP. 3

This well-known work presents a good treatment of discrete ran¬

dom variables. However, it does not cover continuous random
variables.
4. Jewell, W. S., “The Properties of Recurrent-Event Processes,”
Operations Research, Yol. 8, No. 4, 1960, pp. 446-472.
6. Kemeny, J. G., and J. L. Snell, Finite Markov Chains. Princeton, N. J.:
D. Van Nostrand Co., Inc., 1960.
6. Kemeny, J. G., H. Mirkil, J. L. Snell, and G. L. Thompson, Finite
Mathematical Structures. Englewood Cliffs, N. J.: Prentice-Hall
Inc., 1959. ’
Gives a very elementary but clear treatment of discrete and con¬
tinuous random variables, including Markov processes discrete in
space and time.
7. Molina, E. C., Poisson’s Exponential Binomial Limit. Princeton, N. J.:
D. Van Nostrand Co., Inc., 1942.
Gives the most complete tables of p(x; M) and P(x; M) generally
available. However, a somewhat more complete set of tables has
been computed recently by R. Pelletier of the Defense Electronics
Division of the General Electric Co. (Report R 60 DSD 13). These
may be published in the near future.
8. Morse, P. M., Queues, Inventories, and Maintenance. New York: John
Wiley and Sons, Inc., 1958.
Provides an introduction to queuing theory and some of its appli¬
cations.
9. Takacs, L., Stochastic Processes. London: Methuen and Co., Ltd., 1960.
Presents about one hundred problems and their solutions. The
problems emphasize various physical applications of stochastic
processes. A concise introduction to Markov processes is also given.
10. Wadsworth, G. P., and J. Bryan, Introduction to Probability and Ran¬
dom Variables. New York: McGraw-Hill Book Co., Inc., 1960.
Gives a fairly elementary discussion of continuous random varia¬
bles as well as some discussion of discrete random variables.

PROBLEMS

3-1. Let a; be a discrete random variable which can only take on the n
values j + 1, j + n. If p(j + i) = l/n, i = 1,. . ., n, then x is
said to have a uniform or rectangular distribution. Similarly, if a: is a
 PROBLEMS 151

continuous random variable, and its density function/^) is given by

0 , x < a

/(*) = * a<x<b,b>a

-0 , x > b

then x is said to have a uniform or rectangular distribution. Compute

the mean and variance for the discrete and continuous uniform
distributions referred to above.
3-2. Suppose that demands are being generated by an Erlang process of
order n. Compute the probability that a demand occurs between t
and t -f- dt, given that no demand has occurred for a time t since the
previous demand.
3-3. A measure sometimes used in describing processes which generate
demands is the coefficient of variation k for the process; k is defined
to be crX where 1/X is the mean time between demands and c2 is the
variance of the time between demands. What is k for the Poisson
process and the Erlang process of order n?
3-4. Show that T(l/2) = vV. Also show that

Hint: To evaluate r(§), let x = u2 in the definition of r(|).

3-5. Compare the variance of the Poisson and negative binomial distribu¬
tions when they have the same mean. Express the variance of the
negative binomial distribution in terms of its mean and n, the order
of the negative binomial distribution. Recall that if a Poisson process
is generating demands, and lead times have a gamma distribution,
then the marginal distribution of lead time demand is a negative
binomial distribution. What can be said then about the effects of the
introduction of uncertainty on the variance of the marginal distribu¬
tion of lead time demand?
3-6. Let <P(s) be the generating function for the density function p(x).
Show that n, the mean of x, is given by m = (P'(l) where (P'(l) is the
derivative of (P(s) evaluated at s = 1. Also show that

a-2 = (P"(l) + (P'(l) - [(P'Cl)]2

where (P"(l) is the second derivative of (P(s) evaluated at s = 1. By

use of these results, compute the mean and variance of the Poisson
and negative binomial distributions.
 152 PROBABILITY THEORY AND STOCHASTIC PROCESSES CHAP. 3

3-7. Compute the generating function for the binomial distribution, and
use the results of Problem 3-6 to compute the mean and variance of
the binomial distribution. Let xi,. . ., xm be independent random
variables having binomial distributions with the same value of p,
but perhaps different n values. What is the distribution of
y = xi + ... +XJ!

3-8. Consider the random variable y which is the sum of a random number
of independent, identically distributed, discrete random variables
Xi, i.e., y = xi + . . . + xn where n is a random variable and the xf
are independent and all have the same distribution. Let py, px, pn be
the expected values of y, xi} n respectively and of, <rl, al the corre¬
sponding variances. By use of the results of Problem 3-6 and the
generating function for y show that

= MxM» and <ty = pn& + plan

Hint: Recall that (3-95) holds and (P(l) = 1.

3-9. Derive the results of Problem 3-8 directly without the use of gener¬
ating functions. Hint: If p(x) is the density function for the Xi and
r(n) the density function for n, then the probability of any set
(zi,. . ., xn, n) is
P(zi)pfe) . . . p(xn)r(ri)

Mv 23 23 yp(xi) ■ • ■ p(xn)r(ri) = V V V x{p(xi)r(n)

n i&inn*' n **• given
6 n
V'»* < = 1

ay = 23 X) (*i + • • . +
n XI, ,Xn
Xn)2p(xi). . . p(xn)r(n) — pi
given n

Xi + . • • + xl +x XiXj p(xi) . . . p(xn)r(n) - i4

n Xl,...,Zn
hJ

~ wMz + n(n — l)pl]r(n) — pi

3-10. Note that the results of Problem 3-8 allow us to compute the mean
and variance of the distribution of demand in a time t when the time
between demands and the number of units demanded per demand
are random variables, provided that we know the means and var¬
iances of the distributions of the times between demands and the
number of units demanded per demand. By use of the results of
Problem 3-8 compute the mean and variance of the stuttering Poisson
distribution. Note also that the results of Problem 3-8 allow us to
 PROBLEMS 153

compute the mean and variance of the marginal distribution of lead

time demand in the case where the lead time is a discrete variable.
Here x{ might be interpreted as the demand in day i.
3-11. Consider a situation in which demands are treated as discrete, and
the procurement lead time is a continuous random variable with
density function g(t). Let p(x\i) be the probability that a; units are
demanded in time t and let <P(s; t) be the generating function for this
density. Let y be the random variable representing the lead time
demand, and let (P*(s) be its generating function. Show that

*•« = JQ <?0; t)g(t) dt

Find <?*(s) in the case where a Poisson process generates demands,

units being demanded one at a time, and the lead time has a gamma
distribution. In this way obtain another derivation of the fact that
the marginal distribution of lead time demand has a negative bino¬
mial distribution.
3-12. Let p(x\t) be the probability that x units are demanded in a time t.
Assume that the mean of this distribution is \t where X is the mean
rate of demand and its variance is o%t. Let g(t) be the density func¬
tion for the procurement lead tune with mean fxt and variance of.
Assume that x is either discrete or continuous and that t is
continuous. If y is the random variable representing the lead time
demand, and fiy, a% are the mean and variance of y respectively,
show that ’
Vv = \iit and a\ = ntal + XV?
Hint:

+ <r? = y2p(y\t)g(t) dy dt = (alt + \H*)g(t) dt

3-13. Suppose that demands are generated by a Poisson process and that
the number of units demanded per demand is described by a binomial
distribution. Determine the density function for the number of
units demanded in a time t. Find the generating function for this
density, and the mean and variance of the density function.
3-14. Suppose that the procurement lead time is a continuous random
variable with mean \it and variance of. Suppose also that the process
generating demands is such that the mean time between demands is
and the variance of the number of demands occurring in time t is
cfdt. Finally imagine that the number of units demanded per demand
is a random variable with mean \xx and variance By use of the
results of Problems 3-8 and 3-12 determine the mean and variance
 154 PROBABILITY THEORY AND STOCHASTIC PROCESSES CHAP. 3

of the lead time demand in terms of the other means and variances.

3-15. Derive the mean and variance of the normal and gamma distributions
using their moment generating functions.

3-16. Let x be a discrete random variable with density function p (x). Then
the moment generating function for x is defined to be

3K(«) = X) «""?(*)
x=0

Given 9TC(s), how can one determine the mean and variance of x?
Compute 971(5) for the binomial, Poisson, and negative binomial
distributions.

3-17. By use of the results of Problem 3-16 compute the mean and variance
of the binomial, Poisson, and negative binomial distributions.

3-18. Let xly x2 be two independent discrete random variables with moment
generating functions m(s) and 9Jl2(s) respectively. Determine the
moment generating function for y = xx + x2.

3-19. Consider a process generating demands for which the density func¬
tion for the time between arrivals is g(t), i.e., the probability that a
demand occurs between t and t + dt after the last demand is g(t) dt,
and this depends only on t. Suppose that we start observing a system
at a time r after the last demand and ask what is the probability
g(t\T) dt that the next demand will occur between t and t + dt after
we begin observation. Show that

g= = Jr
Evaluate for the case where a Poisson process is generating
demands. Now suppose that we begin observation of the system at a
random point in time. This means that when observation is begun
we have no knowledge of the time when the last demand occurred.
The probability that the last demand occurred between r and r + dr
in the past is equal to the joint probability that the next demand
after a given demand does not occur in a time r, i.e., (?(r), and that
we started observation between r and r + dr. This latter proba¬
bility is independent of r if we begin at random and is proportional
to dr, i.e., adr. Thus the desired probability is aG{r) dr where a is
determined so that the integral over r is equal to unity. By use of
these results, show that if u(t) dt is the probability that a demand
 PROBLEMS 155

occurs between t and t -f- dt when observation is begun at random,

then

U(t) = X jT g(t + r) dr = \G(t)

where X is the mean rate at which demands occur. Compute u(t) for
the Poisson distribution and show that g(t) = u(t).
3-20. By use of the results of Problem 3-19, show how to compute the
probabilities Un(t) that n demands occur in a time t when observa¬
tions are begun at random, given the probabilities Vn(t) that n units
are demanded in a time t when t is measured from a particular
demand.
3-21. Compute u(t) when an Erlang process of order m is generating de¬
mands. Attempt also to compute the Un(t) and Vn(f) for this process.
Note that u(t) is defined in Problem 3-19 and Un(t), Vn(t) in Problem
3-20.
3-22. Show that the equivalent of Eq. (3-29) holds for continuous random
variables.
3-23. Show that Eq. (3-84) is the generating function for the negative
binomial distribution even if n is not an integer.
3-24. Let xi, x2 be continuous, independent random variables which can
take on any real values. Show that the moment generating function
for y = xi + x2 is the product of the moment generating functions
for x\, x2.
3-25. Consider a Markov process discrete in space and continuous in time.
Prove that if Poisson processes generate transitions between states,
then the only an, i ^ j, which are different from zero are those which
correspond to the occurrence of a single Poisson event.
3-26. Provide a probabilistic interpretation of Eq. (3-140).
3-27. Compute the Poisson probability p(30; 25) = p(x; M) and the normal
approximation to it. Also compute P(30; 25) and the normal approx¬
imation to it. Next compute P(30; 25) — P(40; 25) and the
normal approximation to it. Compute p(95; 75), P(95; 75), and
P(95; 75) — P(105; 75) and the corresponding normal approxima¬
tions. In making the above computations, it is suggested that
Molina’s tables be used.
3-28. Differentiate Eq. (3-166) with respect to T and use this result to
demonstrate that Wa(x, T) is correct.
3-29. Determine the limiting forms of the Poisson properties derived in
Sec. 3-12 obtained by replacing the Poisson terms with their normal
156 PROBABILITY THEORY AND STOCHASTIC PROCESSES CHAP. 3

approximations. Compare these results with the corresponding ones

obtained directly, using the normal distribution.
3-30. Derive all properties in Appendix 3 not derived in the text.
3-31. Derive all properties in Appendix 4 not derived in the text.
3-32. Compute the mean and variance of the stuttering Poisson distribu¬
tion from its generating function.
3-33. Consider an inventory situation in which units are demanded one at
a time, and the number of units demanded in a timp. period of length
t has a Poisson distribution, with the mean demand rate being 40
units per year. The operating doctrine is to place an order when the
on hand inventory reaches a level r. If the procurement lead time is
two months, determine the value of r such that the probability of
running out of stock in the lead tune and having one or more demands
occur when the system is out of stock is less than or equal to 0.02.
LOT SIZE

REORDER POINT MODELS

WITH STOCHASTIC DEMANDS

“When your pills get down to four,

Order more.”

Anonymous
 4-1 Introduction

Beginning with this chapter, we wish to study inventory models in which

explicit account is taken of the fact that demands on the system cannot be
predicted with certainty but instead must be described probabilistically.
The introduction of randomness into the nature of the demand pattern
brings to the fore at the outset several new considerations which did not
enter into the analysis when studying the deterministic models of Chapter
2.
One of these new considerations concerns how much is known about the
state of the system at any point in time. For the deterministic models of
Chapter 2, it is possible to determine for all future times precisely what the
state of the system will be if the state is known at a given time and if the
quantity to be ordered and the reorder point are specified. However, when
randomness is introduced into the demand pattern, it is no longer possible
to make such predictions, since the times of occurrence of the demands
(and perhaps also the number of units demanded per demand) are random
variables. One cannot know the state of the system at each point in time
unless each transaction (demand, placement of order, receipt of shipment,
etc.) is recorded and reported as it occurs. We shall say that an inventory
system is using transactions reporting if all transactions of interest are
recorded as they occur, and the information is immediately made known
to the decision maker. When transactions reporting is used, then it is
possible to make decisions concerning the operation of the system, such as
the decision as to whether or not to place an order each time a demand
occurs. The effort required to operate a transactions reporting system can
vary widely with the circumstances. In some cases, it involves nothing
more than placing a card in the stock bin indicating that it is time to re¬
order when the stock gets down to the card. Here no actual recording or
reporting is done, since in this simple situation the recording and reporting
are automatic through the appearance of empty spaces in the bin. In other
cases, transactions reporting may require the operation of a large scale data
processing system in which all transactions are fed into a computer that
automatically updates all records, prints out orders to procure when
necessary, etc.
159
 LOT SIZE—REORDER POINT MODELS CHAP. 4

It is not always desirable to have inventory systems use transactions

reporting, since it may be too expensive to do so. The other procedure
commonly used in the real world will be referred to as periodic review. When
a periodic review procedure is used, the state of the inventory system is
examined only at discrete, usually equally spaced points in time. Decisions
concerning the operation of the system, such as whether or not to place an
order, are made only at these review times. In fact, the decision mn.W
knows nothing about the state of the system at times other than the review
times.
In this chapter we shall confine our attention to transactions reporting
systems, and m the next chapter periodic review systems will be studied.
We noted m Chapter 2 that there was no difference between what we have
ere referred to as transactions reporting systems (there they were called
7’ r)A m«dels) and Periodic review systems. When randomness is intro¬
duced, however, we shah see that considerable differences exist between
the two systems.
Another new consideration concerns the criterion to be used for deter-
t11°P^ma*_°Perating doctrine for the system under consideration,
in ^ chapter, as in Chapter 2, we shall only be concerned with cases in
which the process generating the demands does not change with time. In
par lc ar, this implies that the mean rate of demand remains constant
over ime. As in Chapter 2, the criterion that will be used to determine the
optimal policy is the minimization of the average annual variable cost.
. de average annual variable cost is still defined as the limit as f approaches
infinity of the average annual cost for a time period of length f. When the
demand over time is a random variable (and perhaps the procurement lead
tme is also a random variable), the average annual cost cannot be evalu-
ated quite so simply as in Chapter 2. We shall show how to do this later.
A final new consideration concerns the state of the system at the time a
procurement arrives. For the deterministic models of Chapter 2, it was
possible to predict exactly the on hand inventory at the time of arrival of a
procurement. However, when the lead time demand is a random variable,
i isi no onger possible to predict exactly how much inventory will be on
an w en a procurement arrives. This on hand inventory level will also
e a ran om variable. We shall define the expected value of the net inven-
?r7 “ e backorders case or the expected value of the on hand inventory
sales case at the time a procurement arrives to be the safety
s oc wi e denoted by s. The value of s can be either positive, nega-
T/f'-°,r zero , backorders case, but must be non-negative for the
ost sales case. I For deterministic models, we noted in Chapter 2 that it is
never optimal to have any stock on hand when an order arrives. However,
when demands are described probabilistically, it will frequently be true
 SEC 4-1 INTRODUCTION 161

ihat the optimal safety stock should be positive. The reason for this is that
if the safety stock was zero, then because of the random nature of the lead
time demand, the system would very frequently run out of stock before
the arrival of the order, thus incurring stockout costs. When it is expensive
to incur backorders or lose sales, then on the average it is cheaper to carry
some additional stock to avoid these stockout costs.
In this chapter, as in Chapter 2, we shall study lot size-reorder point
models for a single installation. Recall that a lot size-reorder point model
is one such that a quantity Q is ordered each time the appropriate inventory
level (the on hand inventory, the net inventory, the on hand plus on order
inventory, or the inventory position) reaches the reorder level. The purpose
jpf the analysis is to determine the optimal value of the order quantity Q
and the reorder point.
It is important to observe that by its definition, a lot size-reorder point
modeb assumes that an order is placed when the inventory level reaches the
reorder point, i.e., there is no overshoot of the reorder point.; In order for
this to be true the state of the system must be examined after every
demand. Thus the use of a lot size-reorder point model (a (Q, r) model)
implicitly requires that a transactions reporting system be usedl
The ability to place an order precisely as the reorder point is reached also
implies that, when the integrality of demand is taken into account, the
number of units demanded per demand cannot be a random variable.
When the number of units demanded per demand is a random variable,
then it is possible to overshoot (unavoidably) the reorder point. For such
situations, it may no longer be appropriate to order a fixed quantity each
time an order is placed. An alternative procedure is to set two levels
r, R(R > r) such that if the inventory level falls to x1 x < r on some
demand, we order up to the level R, i.e., a quantity R — x is ordered. We
shall refer to such an operating doctrine as an Rr doctrine, and a transac¬
tions reporting model which uses such a doctrine as an {R, r) model. The
task of working with (R, r) models when the number of units demanded per
demand is a random variable is much more difficult than for the models to
be discussed in this chapter. We shall discuss such models in Chapter 8.
It should be noted that a (Q, r) model is a special case of an (R, r) model,
with R = r + Q, which is appropriate when units are demanded one at a
time and there is never any overshoot of the reorder point.
In the real world, even if the number of units demanded per demand is
a random variable, then provided that the probability of a large overshoot
is very small, one will often operate the system using a (Q, r) model. To
d° this one might compute a Q* and r* using the model described in the
next section, and then either order Q* each time an order is placed, or
order up to Q* + r* each time the reorder level is crossed!
 162 LOT SIZE—REORDER POINT MODELS CHAP. 4

4-2 Heuristic Approximate Treatment of the Backorders Case

Before turning to the presentation of the detailed, exact formulations of

{Q, r) models for some special cases, we shall present a heuristic approxi¬
mate treatment of (Q, r) models. In this section, the backorders case will
be examined, and in the following section the lost sales case will be studied.)'
It is these simple approximate treatments that are almost always to be
found m the texts which discuss (Q, r) models [2, 6]. ■ Such developments
require a large number of assumptions and approximations, j Frequently,
m the presentations given of these models, only a small fraction of the
assumptions and approximations needed are explicitly stated. Indeed it is
hard to be able to appreciate all of them until the exact treatment is cov¬
ered. For this reason no attempt will be made to list every approximation
and assumption at this time. A number of assumptions and approximations
will be listed here, however. We shall see later what additional implicit
assumptions were made. In addition, we shall later examine the ways in
which models to be derived here are more general than might be apparent
from the discussion to follow. Even though many approximations and
assumptions are made in this and the following section, the resulting
models are especially useful for practical applications because of their
simplicity, and because, often, the special cases for which the exact equa¬
tions are available do not represent the real world situation much more
accurately than the approximate models, i.e., the additional assumptions
needed to reduce the exact model to the simple model are frequently
warranted m practice.
. sha11 assume that the system under consideration consists of a single
mstaUation which uses transactions reporting. We wish to determine the
optimal order quantity Q and reorder point r for a given item. If the system
s oc s more than a single item, it will be assumed that there are no inter¬
actions between the items. The optimal values of Q and r will be found by
minimizing the average annual variable cost. At the outset, we shall make
the i olio wing assumptions:

(1) The unit cost C of the item is a constant independent of Q.

(2) The backorder cost is tt per unit backordered. There is no cost -M
which depends on the length of time t for which the backorder exists,
i.e., it = 0.

(3) There is never more than a single order outstanding.*

“ ^ The ^ of operating the information processing system is independ¬
ent of Q and r.

arbiJ™^6^ s'mPie model developed can also be applied when an

arbitrary number of orders can be outstanding.
SEC 4-2 HEURISTIC TREATMENT OF THE BACKORDERS CASE 163

Assumption (3) implies that at the time the reorder point is reached there
are no orders outstanding, so that the inventory position (the amount on
hand plus on order minus backorders) is equal to the net inventory (on
hand minus backorders). Thus the order point will be the same regardless
of whether it is based on the inventory position or net inventory. The
additional assumption will now be made that:
v (5) The reorder point r (based on the inventory position or net inven¬
tory) is positive.
This is almost always true in practice, since one will not normally wait
until there are backorders on the books before placing an order. ’ Because
of (5), there will be no backorders outstanding at the reorder point. In

placed received

FIGURE 4-1.

fact, at the reorder point, the inventory position is equal to the on hand
inventory. The behavior of the on hand inventory and the inventory
position for a system of the type being described is illustrated in Fig. 4-1.
For the model being examined, any one of the inventory levels, on hand,
net, or inventory position can be used to define the reorder point, and the
reorder point has the same value for any one of them. jNote that to use the
on hand level we must assume that after an order arrives, it is sufficient to
meet all backorders and raise the on hand inventory level above the reorder
point. If this ever failed to happen, the reorder point would never be
reached again and the system would proceed to accumulate backorders.
When the reorder point is thought of in terms of the inventory position of
the system, then assumption (3) guarantees that the on hand inventory
will always be raised above the reorder point when an order arrives, for
 164 LOT SIZE—REORDER POINT MODELS CHAP. 4

otherwise it would not be possible to have only a single order outstanding.

_ A® Chapter 2, we say that the system goes through one cycle in the
time between the placing of two successive orders or the receipt of two
successive procurements. Unlike the case for deterministic models, it is no
longer true that the system repeats itself exactly each cycle. Even the
length of the cycle is now a random variable.' However, the system does
repeat itself in the sense that the inventory position varies between r and
r + Q during each cycle.
We shall now proceed to evaluate the various terms in the average annual
cost. These terms include the ordering costs, the cost of carrying inventory,
and the cost of backorders. Because of assumption (1), it is unnecessary to
include the cost of the units. The average annual cost of units procured is
independent of Q and r, since the unit cost C is independent of Q. We shall
here treat all variables as continuous. It is a trivial task to modify the
analysis to the case where Q, r and the demand variable are treated as
discrete; this is left to Problems 4-1 through 4-3. Let f(x; t) dx be the
probability that the number of units demanded in a time t lies between x
and x + dx, and suppose that the mean rate of demand, which is constant
over time, is A.
:^'^e co®t of placing an order will be taken to be A. Since the average
annual demand is X and since an order is placed after every Q demands,
the average annual cost of placing orders is \A/Q. The inventory carrying
charge will be denoted as usual by I. |The average annual cost* of carrying
inventory will be IC times the average number of unit years of stock held
per year. Some variation of the following argument is often given to obtain
the average number of unit years of stock held per year, tirst observe
that de&iition the net inventory is the on hand inventory minus the
backorders, i.e., the net inventory is the difference of two random variables.
Then, by the results of Sec. 3-8, the expected value of the net inventory at
any time is the expected value of the on hand inventory minus the expected
value of the backorders. Thus the expected on hand inventory is equal to
the expected net inventory plus the expected number of backorders. Now
when backorders are expensive, then if they are incurred at all during a
cycle, they will be incurred only close to the time when a procurement is
due m. However, to determine the average unit years of stock held per
year it is necessary to integrate the expected on hand inventory over time,
or equivalently integrate the expected net inventory plus the expected
backorders over time. The expected backorders term will be significantly
different from zero only a very small fraction of the time if the backorder

n s c apter and the next, we shall employ the convention that the expected value
o any ran om variable computed on an annual basis will be referred to as the average
annua va ue. of the variable, while expected values for all other time intervals, or at
given pomts in time, will be referred to as expected values.
SEC 4-2 HEURISTIC TREATMENT OF THE BACKORDERS CASE 1 65

cost is high and will therefore make only a slight contribution"to the inte¬
gral. Thus it will be assumed that the expected number of backorders is
negligible, and for computing the inventory carrying charges, the expected
on hand inventory can be taken to be equal to the net inventory!
By definition, the expected net inventory at the time of arrival of a
procurement is the safety stock s. The expected net inventory immediately
after the arrival of a procurement is then Q + s. Thus, if the arrival of an
order initiates a cycle, the expected net inventory at the beginning of a cycle
is Q + s and is s at the end of a cycle. These will also be the expected
values of the on hand inventory at the corresponding times, when the ex¬
pected number of backorders can be neglected. In this case, since the mean
rate of demand is constant, the expected on hand inventory will decrease
linearly from Q + s at the beginning of the cycle to $ at the end of the cycle
and will average to

- (Q + s) + | s = ^ + s (4-f)

which is the average number of unit years of stock held per year (illustrate
this graphically).
In Chapter 2 we introduced the equivalent of s directly as one of the
variables. When dealing with probabilistic models, it is convenient to
eliminate the variable $ and replace it by r (Problem 2-15 showed that this
was easy to do in the deterministic case). Since the reorder point in terms
of the on hand or net inventory is r\ and since nothing arrives between the
time an order is placed on reaching r and the time this order arrives (i.e.,
assumption (3) holds), then if the order requires a time r to arrive and if x
units are demanded in this time,, the net inventory £(rr, r) will be £(x} r) =
r — x at the time of arrival of the order, and the expected value of the net
inventory averaged over all x for a given r is

f £0, r)f(x-, r)dx = f (r - x)f(x; r) dx (4-2)

Jo Jo
If the procurement lead time is constant, then (4-2) is the safety stock.
Suppose, however, that the procurement lead time is a random variable
such that g(r) dr is the probability that the procurement lead time lies
between r and r + dr. Then by (3-70), the expected value of %(x, r) aver¬
aged over x and r is

/ / r)f(x; r)g(r) dxdr = (r — x)h(x) dx (4-3)

Jo Jo Jo
where

h(x) = f(%; r)g(r) dr

 166 LOT SIZE—REORDER POINT MODELS CHAP. 4

is the marginal distribution of lead time demand.

We can therefore write

s=J (r — x)h(x) dx = r — jj, (4-5)

where fx is the expected lead time demand, i.e.,

xh(x) dx (4-6)

and where h(x) f(x, t) if the lead time is a constant r or h(x) is given by
(4-4) when the procurement lead time is a random variable with density
Trtr'rr\ averaSe annual inventory carrying costs are therefore
f k|_(Q/2) + r — ji].
It remains to evaluate the average annual cost of backorders. The
argument used to obtain this cost usually runs as follows: The average
number of backorders incurred per year is simply the expected number of
backorders incurred per cycle times the average number of cycles per
year, i.e., \/Q times the expected number of backorders incurred per cycle.
(This argument is correct, as we shall see later.) Now the number of back-
orders i?(z, r) incurred in a cycle will simply be the number of backorders
on the books when a procurement arrives. If the lead time demand is
X> ^he number of backorders will be

, v fo if x - r < 0
v(x, r) = (4-7)
[x - r if £ — r > 0
Thus the expected number of backorders per period 7j(r) is

^(r) = Jo v(x, r)h(x) dx = £ (* - r)h(x) dx

-i: xh(x) dx - rH(r)

where, as above, h(x) is the marginal distribution of lead time demand and
(4-8)

jf “ T complementary cumulative of h(x)\ The average annual cost

of backorders is then
7rX r f00
Tq[J xh(x) dx - rtf(r)J

All the terms in the average annual variable cost X have now been found;

X = qA+IC
[1 + r ~ M] + ^ **(*) dx ~ rff(r) J (4-9)
SEC 4-2 HEURISTIC TREATMENT OF THE BACKORDERS CASE 167

since by assumptions (1) and (4) it is unnecessary to include the cost of the
units or the cost of operating the information processing system. We wish
to determine the values of Q and r which minimize the 3C of (4-9). If the
optimal values Q*, r* satisfy 0<Q*<oo?0<r*<oo? then Q* r* must
satisfy the equations

XA IC 7r\ , x
(4-10)

f-o = rc + |[- ■rh(r) + rh(r) — H(r)] (4-11)

Here we have two equations to be solved for Q and r. It is convenient to

write (4-10), (4-11) as

q _ + 7ri?(r)]
(4-12)

and
QIC
H(r) = (4-13)
7r\ \/

Recall that 77 (r) is given by (4-8). A procedure for numerically solving

for Q* and r* will be given in Sec. 4-4.
- Qften it is desirable to evaluate 5Z numerically for given values of Q
and r. If we assume that h(x) is a normal distribution with mean fi (the
expected lead time demand) and standard deviation a, i.e., h(x) =
n(x; /i, cr), then \

^ xh(x) dx = j xn(x; n, o) dx = J^ ^ <j> dx (4-14)

r00 /*«
= cr / v<j>(v) dv + j* <l>(v) dv (4-15)
J (r—fi)/cr J(r—pi)/<r

= ^(Lvif) + ^(Lr^)
Thus

K -1 * +IC [i+' - <•] + f [ <* - (!—“)

3C

+ <r<t> (^)] (4-16)

In this form, 3C is easily evaluated with the aid of normal tables.

 168 LOT SIZE—REORDER POINT MODELS CHAP. 4

4-3 Heuristic Approximate Treatment for the Lost Sales Case

The simple approximate treatment of the lost sales case differs very little
from the backorders case studied in the previous section. From our study
of the lost sales case in Sec. 2-6,| it is clear that the minimization of the
average annual cost is equivalent to the maximization of the average annual
profit if in the cost expression the cost of a lost sale includes the lost profit.)
It is also clear that, in general, the average number of cycles per year is no
longer X/Q but is instead X/(Q + XT), where f is the average length of timo
per cycle for which the system is out of stock. In the real world T is usually
a very small fraction of the total length of the cycle. Since it is very incon¬
venient to include f in the analysis, the following assumption is usually
made in the simple treatments.
(a) The value of T is small enough to be neglected, so that the average
number of cycles per year is X/Q.
We shall suppose that assumptions (1), (3) and (4) of Sec. 4-2 also apply
here. Assumption (2) of Sec. 4-2 is not needed here, since for the lost sales
case the cost of a lost sale will always be x; there is never a term propor¬
tional to time, as it does not make any sense to talk about the time for
which a lost sale exists. We do assume, however, that x includes the lost
profit.
Xbe only difference between the lost sales and backorders models
comes in evaluating the safety stock expression. The expected on hand'
inventory when a procurement arrives will be s, the safety stock, and the
expected on hand inventory immediately after a procurement arrives will
be Q -fi- s. Thus the expected on hand inventory varies between Q + s
and s in a cycle and averages to (Q/2) + s. Let e(x, r) be the on Wnd
inventory when the procurement arrives if the lead time Hp™^ is x
Then
r — x, r — x > 0
t{x, r) =
[0, r - x < 0
and the expected amount on hand when a procurement arrives is

S = Jo t('X’ ^ = Jo ^ ~ ^ (4-17)

where h(x) represents the marginal distribution of lead time demand.

However

s = Jo (r~ *)Kx) dx - j (r- x)h(x) dx

(4-18)
= r — M + j xh(x) dx — rH(r)
SEC 4-4 DISCUSSION OF THE SIMPLE MODELS 169

Thus the average annual cost of carrying inventory is

xh(x) dx — rH(r) (4-19)

It will be noted that the expected number of lost sales per cycle is pre¬
cisely the same as the expected number of backorders per cycle in Sec. 4-2,
i.e., is given by (4-8). This follows, since if the lead time demand is x, the
number of lost sales will berr-rif# — r > 0 and 0 otherwise. Averaging
this over x, we obtain (4-8). It follows that the average annual variable
cost in the lost sales case is

3C = | A + IC | + r - m

+ (iC + ^) Jr xh(x) dx - rff(r) ] (4-20)

Again we wish to determine the values of Q and r which minimize 3C. If

0<Q*<°°,0<r*<ooJ then Q*, r* must satisfy d3Z/dQ = d3Z/dr = 0.
The equivalents of (4-12), (4-13) then become

Q 4 2\\A -f- 7rq(r)]

IC
(4-21)

H(r) = ..- (4-22)

K) Xtr + QIC

A numerical example illustrating the solution of these equations will be

given in the next section. If h(x) is a normal distribution, then the equiva¬
lent of (4-16) for the lost sales case is

x-^ + rc[f + r-„]

+ (rc + §)[&,-0*(^) + ^(^)] (4-23)

4-4 Discussion of the Simple Models and a Numerical Example

In this section we shall discuss some interesting properties of the two models
developed in the previous two sections, present a numerical technique for
determining the optimal Q and r values, and give a numerical example to
illustrate the use of the models.
First note from (4-12), (4-21) and (4-8) that in both the backorders and
 170 LOT SIZE—REORDER POINT MODELS CHAP. 4

lost sales cases Q* > Qw, i.e.; the optimal Q value is never less than the
Wilson Q. In fact, for all normal cases, Q* > Qw for any r (not simply the
optimal r). The intuitive reason for this is that the expected number of
backorders or lost sales per cycle depends on r but is independent of Q.
oweyer, the average annual cost of backorders or lost sales is proportional
to 1 /Q. Therefore, for a fixed, specified value of r, when w > 0, it always
pays to increase Q somewhat thus incurring more carrying charges per
year m order to gain the benefits of correspondingly reducing the average
i
annual cost of backorders or lost sales. The value of Q should not be in¬
creased indefinitely because after a point the incremental savings in
expected stockout costs becomes less than the incremental increases in
expected carrying costs.
Secondly it will be observed that (4-13) does not make any sense if
; 7" > L ThlS WlU never occur for cases for which this model is in-
en e to apply, i.e., to high backorder costs. This anomaly arises because
the backorders term was omitted in determining the inventory carrying i
costs Note that a similar problem does not arise in the corresponding
equation (4-22) for the lost sales case.
Before going on to a third observation concerning the models, it will be
convenient to introduce a numerical procedure for solving the pairs of
equations (4-12), (4-13) or (4-21), (4-22). This procedure is as follows:
fltial estlmate of & i-e-> write Qx = Qw. Then use Q, in
j or v*“22) to compute n. The n so obtained is used in (4-12) or (4-217
o compute Q2. This Q2 is used in (4-13) or (4-22) to compute r2, etc. This
iterative procedure is continued until Q and r are obtained with sufficient
accuracy. If the equations have a solution at all, then the above iterative
scheme must converge to a minimum cost solution^
A proof of the convergence of the iterative scheme will be carried out
using a graphical argument for the backorders case. The corresponding
argument for the lost sales case can easily be made and is left for Problem
f'4^EnUat,10nS t12 and (4'13) can be thought of as describing two curves
whe^Or"Pnne' F°r th! CTe described by equation (4-13), we note that
££ Lkiis"a,ld when 8 ■A/IC-m -1—- «•
dQ t\ dr = IC 1
~dr ~ -fch(r)
or (4-24)
dQ 7rX h(r)
For the curve described by (4-12) note that when r = 0

'2\[A + Tji]
Q =
IC
and when r = =c, Q = Qw. Furthermore dr/dQ < 0 If
one plots the two *
curves something like Fig. 4-2 will be obtained.
.*
SEC 4-4 DISCUSSION OF THE SIMPLE MODELS 171

To initiate the iterative scheme, (4-13) is used to compute r when Q = Qw,

i.e., we start on the curve defined by (4-13) at the point where Q = Qw>
The r value so obtained is used in (4-12) to compute a new Q value Q2,
i.e., we move from the point (Qw, rx) on the curve defined by (4-13) to a
point on the curve defined by (4-12) having the ordinate rx. The Q2 value
is used in (4-13) to compute a new r, i.e., we move from the curve defined
by (4-12) to the curve defined by (4-13) at constant Q. Thus a series of
steps is obtained as shown in Fig. 4-2. It is clear that the iterative scheme

must converge to Q* and r*. In practice, the convergence is usually quite

rapid.
It turns out that the solution Q*, r* to (4-12), (4-13) or (4-21), (4-22) is
always unique. The proof of this follows from the fact that CfC(Q, r) is
convex. The details of this proof are to be developed in Problems 4-5
through 4-8. In the backorders case, it is not necessarily true that a solution
will always exist. An investigation of when a solution does not exist is
left to Problem 4-9. For the lost sales case, however, it is easy to show that
there will always be a solution. The proof is left to Problem 4-10.
We are now able to make a third observation concerning the backorders
and lost sales models. If we examine the optimal values of Q and r com-
 172 LOT SIZE—REORDER POINT MODELS CHAP. 4

puted from the lost sales model and the backorders model when the same
values of the parameters, i.e., the same values of X, A, I, C, and x are used
in both, it will be found Q* for the backorders case is larger than Q* for the
lost sales case while r* for the backorders case is smaller than r* for the lost
sales case. This can be seen immediately if one follows through the iterative
computational procedure. At the first step with Qx = Qw> a larger n is
obtained in the lost sales case than the backorders case. This in turn gives
a smaller Q2 in the lost sales case than the backorders case. Consequently,
the ratio on the right of (4-22) will again be smaller than that of (4-13)
and r2 for the lost sales case will be larger than r2 for the backorders case.
Continuing in this way we see that the result stated above holds. In prac¬
tice the difference between the Q*, r* values for the lost sales and back¬
orders cases will generally be very small.
The intuitive explanation of the above behavior is roughly as follows:
Note that if the same values of Q and r are used in both models, the average
annual inventory carrying cost will be less for the backorders model than
for the lost sales case (since in the backorders model for those cycles when
backorders occur, the initial inventory in the next cycle is reduced by the
backorders to being less than Q, while it is never less than Q in the lost sales
case). Consequently, other things being equal, it pays to have a larger Q
in the backorders case than in the lost sales case. On the other hand, an
increase in r causes less of an increase in average annual inventory carrying
charges for the lost sales case than for the backorders case, while yielding
the same change in the expected stockout cost in both cases (provided the
Q’s are the same). Hence, other things being equal, it will be advantageous
to have a higher r in the lost sales case than in the backorders case.
Sometimes it is of interest to compare the average annual cost of opera¬
tion obtained from one of the two models discussed above with that for the
corresponding deterministic model using the same values of the parameters.
The average annual cost for the deterministic model will never be greater
than the average annual cost for the model which treats demand as a
stochastic variable. The difference in average annual costs of the prob¬
abilistic model and the deterministic model will be called the average
annual cost of uncertainty. This indicates what annual savings one could
obtain, on the average, if all uncertainties in the demand and lead time
variables could be eliminated, the mean values remaining unchanged.
We now give an example illustrating the use of the simplified models.

EXAMPLE A large military installation stocks a special purpose vacuum

for use m radar sets- Th.e average annual demand for this tube is
1600 units. Each tube costs $50. The tube must be made to order, and
hence each time an order is placed it is necessary to go through a process
of accepting bids and negotiating a contract. It is estimated that the cost
SEC 4-4 DISCUSSION OF THE SIMPLE MODELS 173

of placing an order is $4000. The installation uses an inventory carrying

charge of I = 0.20. It has been found that if a demand occurs when the
system is out of stock, it is possible to obtain such a tube from a small
stock carried at one of the manufacturers. However, the cost of sending a
plane there to obtain it and the other concomitant expenses amount to
$2000 over the cost of the unit. An empirical investigation has shown
that the marginal distribution of lead time demand is essentially normally
distributed with mean 750 units and standard deviation 50 units. It is
desired to compute the optimal order quantity, the reorder point, and the
safety stock.
The description of the operation of the system presented above indicates
that we here need the model for th< krt sales case, since if a demand occurs
when the system is out of stock it is not backordered, but is instead pro¬
cured from outside the system. The cost of a lost sale ir to be used in the
model is $2000, for by definition, t includes the lost profit which, in this
case, since the selling price is zero, means the stockout cost net of the item’s
cost. The optimal values of Q and r are found using (4-21), (4-22). The
iterative procedure discussed above will be used to solve the equations.
We note that X = 1600 units per year, C = $50, I = 0.20, A = $4000.
The initial estimate for Q is

e, = e. - - V2(1Jf(5Cj = -1130
Then rx is computed from (4-22). In this case )

since the lead time demand is normally distributed with mean 750 and
standard deviation 50. Thus

M - 750\ QJC 1130(10) = 1.130

= 0.00352
V 50 ) \* + QJC ~ 1600(2000) + 1130(10) 321.1

From the normal tables we find that

n - 750
= 2.695

n = 750 + 134.7 = 884.7

To compute Qi we use (4-21). First we compute ijfa) from

n — (i
iKn) = (m - ri)4>
 174 LOT SIZE—REORDER POINT MODELS CHAP. 4

From the normal tables, <£(2.695) = 0.01057, so

v(n) = -134.7(0.00352) + 50(0.01057)

= -0.47414 + 0.52850 0.05436

Thus

Q2 = , ^/2(1600)|"4000 + 20001.054361]

Then for n

* O
\
~ 750 \
50 )
L147
321.1 “
,
and

r2 - 750 0 fid
50 “ 2'69
or
ri = 750 + 134.5 = 884.5

There has been essentially no change in the safety stock. Additional itera¬
tions are not needed since the changes will be negligible. The optimal
values are Q* = 1147, r* = 884, s* « 134. Because of the high cost of
s ockouts a safety stock of 134 tubes is carried. The expected time between
the placing of orders is 8.60 months.
The average annual cost of ordering, carrying inventory, and stockouts
is easily computed from (4-23). It is

x («H») + M + 884 - 750]

Tin , 2000(1600)1, .
+ L10 + 1147 J[- 134(0.00357) + 50(0.01071)]

= 5580 + 7075 + 160 = $12,815/yr.

The corresponding cost for the deterministic case is

Ku = V2XAJC = V2(1600)(4000) (10) = Vk28 X 10* = $ll,305/yr.

since it is never optnnal to have lost sales in the certainty case Thus the
average annual cost of uncertainty is

x ~ K*> = 12,815 — 11,305 = S1510

which is $1/0 more than the average annual cost of $1340 incurred to carry
the 134 units of safety stock Of thp in 70 <kiaa .* ^
of Stock™,wh,l0 +n • „ *1/0’ $160 ls the average annual cost
of stockouts while the remaining $10 is incurred because Q* differs from
6 a W en usmg the °Ptimal Q and r values, the average annual
SEC 4-5 EXACT DEVELOPMENT—INTRODUCTION 175

stockout cost is only $160. This means that on the average a lost sale will
be incurred only once every 12.5 years.
Often it is desirable to know how sensitive Q* and r* are to changes in
the parameters such as A, /, or x. In Fig. 4-3, it is shown how Q* and r*
vary with x. It will be observed that Q* is very insensitive to x while r*
is slightly more sensitive but not really too sensitive, since a change in x
from $500 to $5000 increases r* by only about 40 units, i.e., increases the
safety stock by 40 units. Since the mean lead time demand is 750 units,
the safety stock changes from about 109 units to 149 units or a change of
less than 50 percent for a factor of 10 change in the cost of a lost sale. This

Backorder cost tv (dollars)

FIGURE 4-3.

points out that even if one cannot measure x too accurately, then provided
it is of the right order of magnitude, the Q and r values computed will not
be very far from the optimal ones that would be obtained on using the
correct value of x.

4-5 The Exact Development—Introductory Discussion

After having presented the elementary approximate treatment of lot size-

reorder point models in the previous three sections, we would now like to
 176 LOT SIZE—REORDER POINT MODELS CHAP. 4

turn to a consideration of the exact formulation. Recall from the introduc¬

tion to this chapter that a lot size-reorder point type of operating doctrine
can be optimal only if transactions reporting is used, and if, in the discrete
case, units are demanded one at a time. These assumptions will be made
here.
The main emphasis will be concentrated on developing the exact average
annual cost expression for the case where a Poisson process generates
demands, the procurement lead time is a constant, and units demanded
when the system is out of stock are backordered. We shall also consider
extensions of this case to situations where the lead time is allowed to be a
random variable. It will be seen, however, that it is not at all an easy
matter to develop an exact treatment of this case, except under very re¬
strictive or unrealistic assumptions. The lost sales case will be studied also.
The lost sales case is much more difficult to treat than the backorders case,
and, indeed it is only when the restriction is made that never more than one
order is outstanding that it is possible to obtain exactly the cost expression.
Except for the material to be presented in the next section, we shall in
this chapter only consider cases where a Poisson process generates demands,
and units are demanded one at a time. The discussion of cases where the
order size can be a random variable, or where the interarrival distribution
is something other than exponential will be deferred to Chapter 8. These
latter eases lead to much more complicated expressions than the cases to be
considered here. We shall, however, also consider in this chapter the normal
approximation to the Poisson distribution. This normal approximation is,
m addition, the continuous approximation to some of the more complicated
cases to be examined in Chapter 8.
me procedure used to determine the average annual cost for the simple
approximate models developed in Secs. 4-2 and 4-3, was to compute the
expected cost per cycle and then to multiply by the average number of
cycles per year. This is a very powerful technique for developing the
average annual cost expression. It can be used under a very wide variety
of circumstances as we shall see in this chapter, the next chapter, and in
the problems of Chapter 8. We shall use this procedure to treat the lost
sa es case m this chapter. To illustrate a different and interesting approach
we shall treat the backorders case with constant lead times and a Poisson
process generatmg demands as a Markov process discrete in space and
continuous m time. Here, one first computes the state probabilities, and
then uses them to obtain the average annual cost. While the Markov type
of analysis is useful in a variety of cases, it does not seem to have nearly the
general applicability of the procedure which first computes the excepted
cost Per cycle and then multiplies by the average number of cycles per
year. The Markov form of analysis, however, does have the advantage that
SEC 4-6 COMPUTATION OF TIME AVERAGES 177

it yields directly the state probabilities, which are not obtained directly
from the other method of generating the average annual cost.
Before going on to develop the exact annual cost expression for the
backorders case with constant lead times, we shall first compute some time
averages that will be needed, and also say a word about the possible inter¬
pretations of the state probabilities and their use in computing the time
averages.

4-6 Computation of Time Averages

The criterion to be used in determining the optimal values of Q and r will

be the minimization of the average annual cost, where the average annual
cost is defined to be the limit as f —> <» of the average cost for a time period
of length £. Assume that we select arbitrary values for Q and r, and that
we start out the system in operation using these values. At any arbitrary
point in time after the system starts to operate we begin observing the
system. Suppose that in a time £ since observation was begun the system
goes through n complete cycles and perhaps part of another one. We can
use either the placement or receipt of an order as convenient events to
define the beginning of a cycle. Assume that demands occurring when the
system is out of stock are backordered. In this section, we shall not make
any special assumptions concerning the nature of the stochastic processes
generating demands and lead times except to assume that they do not
change with time, and that units are demanded one at a time.
Let Ti be the length of cycle i, TI be the length of time during which the
system has stock on hand in cycle i, T" be the length of time for which the
system is out of stock in cycle i (Ti = Ti+ T"), Qt- be th&jmit years of
inventory held during cycle. CAt» be the unit yearsof shortage incurred
during cycle i, and i.
The number of orders placed in the time period of length f is n + e where
€ = 0 or 1, depending on whether or not an order is placed in the fraction
of a cycle which may be included. The average number of orders per year
placed over the time span £ is (n + c)/£. As £ oo, it must be true that
n —> oo 1 and the average number of orders placed per year approaches*

* The limits of random variables studied in this section cannot rigorously be considered
to be limits in the strict mathematical sense. The reason for this is that one cannot
guarantee that there exists a T0(e) so large that for all T > T0(e), the difference between
the random variable and its limiting value is less in absolute value than e, for any e > 0,
simply because of the stochastic nature of the process. The sort of statement which can
be made is that the random variable approaches a given limit with probability one, i.e.,
the probability that the random variable differs in absolute value from its limiting value
by any specified positive quantity approaches zero as T —» ». The limits employed in
this section will be interpreted in this way.
 LOT SIZE—REORDER POINT MODELS CHAP. 4

lim? = * (4-25)
f—no f Q f-no f Q

since «=/f-»0, nQ approaches the total demand in the period f, and by

defimtion of the mean demand rate X,

iS f~K 0-26)
Hence we have shown that in the backorders case, the avm^e-numher^
qrders per year, i.e., the average number of cycles per year is.X/Q,. This is
independent of the specific nature of the demand or lead time distributions
provided that an order is placed after every Q demands.
The fraction of the time that the system is out of stock which will be
written Pout, is
n

Pout = (4-27)

This then is, by definition, the probability that the system is out of stock if
we observe the system at a random point in time. Also

n 71
Do* 2>
B = lim lim (4-28)
f—► OO

are the average unit years of stock held per year, the average unit years of
shortage incurred per year, and the average number of backorders incurred
per year. If the on hand inventory were always a constant D, then the unit
years of stock held per year would be D. Thus, Hk.the expected value of
th£xm hand inventory at any instant of time. Similarly, B is the wtperrtpH
number of backorders on the books at any point of time.
It is desirable to rewrite the expression for E. Note that since the limit
of a product is the product of the limits

2 T" 2 «<
E = lim ~ = lim J —-*=1—
f ^ L
r £T
i=i
*

n n (4-29)
I>" Dir
Dr"
SEC 4-6 COMPUTATION OF TIME AVERAGES 179

The second limit is X, since £?« 1 £* is the total number of units demanded in
the time i T". The ratio, in the limit as f —> °°, must be X by the defini¬
tion of the mean demand rate. We have just shown that the average
number of backorders incurred per year is the mean rate of demand times
the probability that the system has no stock on hand.
Another interesting result can be obtained using the fact that the limit
of a product is the product of the limits. We note that

D = lim lim ^ lim (4-30)

£—>oo % £—>oo 7i Q >00 £ Q
n

where by definition

Q = lim (4-31)

is the average number of unit years of stock held per cycle. Furthermore,
by definition of X, limf_>» nQ/f = X. Thus (4-30) follows. The above
shows that the average number of unit years of stock held per year is \/Q
times the average number of unit years held per cycle.
The values of D, B, Ey and P0ut will depend on the nature of the sto¬
chastic processes generating demands and lead times. They can be com¬
puted by first computing the corresponding expected value per cycle and
then multiplying by \/Q. The expected values per cycle are computed by
using the information given about the demand and lead time distributions.
However, if we can determine the steady state probabilities $(x) that the
net inventory has the value x, then the values of D, B, E, and PQut may be
computed directly. Thus

Imax ~1 0

B — ; B = — x\f/(x); Pout == ^(®) (4-32)

X—0 x= — 00 X = — 00

It is important to note that the steady state probability ^(x) can be given
several interpretations. It is the probability that the inventory position
is x if we observe the system at a random point in time (after the system is
in a steady state mode of operation). It is also the long run fraction of the
time that the net inventory will have the value x. Finally, it is the limit
as N —»co of the fraction of an ensemble of N systems in statistical equi¬
librium that will have the inventory position x at any given point in time.
The analysis for the lost sales case is precisely the same as for the back-"
orders case examined above. If the same definitions are used as in the
backorders case, it is immediately clear that the average number of lost
 180 LOT SIZE—REORDER POINT MODELS CHAP. 4

sales per year is \Pout, and D, the average unit years of storage per year, is
given by (4-28) and (4-32). The average number of cycles per year is
different in the lost sales case than in the backorders case, however. We
shall now compute the average number of cycles per year in the lost sales
case. For this case, Q units are demanded in Tl, not in time T{ as in the
backorders case. The average number of cycles is, by definition

However, nQ is the demand in time £?= i Tl and hence the ratio of these two
quantities must approach X. Next observe that

(4-34)
i=l t=l i=l
and hence

I
lim = 1 - lim 1 - Po (4-35)
T—►« S f—f
Thus

!im 7 = ^ (i - Pout) (4-36)

and the average number of cycles per year is (\/Q)(l - Pout). It is also
true that

where T is the average length of time per cycle for which the system is out
of stock.
 SEC. 4-7 EXACT FORMULAS 181

4-7 Exact Formulas for the Backorders Case with Poisson Demands
and Constant Procurement Lead Time

We shall now determine an exact expression for the average annual cost
in the case where a Poisson process generates the times between demands,
units are demanded one at a time, the mean rate of demand being X units
per year, and the procurement lead time is a constant r. In addition to
treating the demand variable as being discrete, the order quantity Q, the
reorder point r, and all the inventory levels will also be treated as discrete
variables.
As yet,- we have not discussed which inventory level (or levels) can be
used to define the reorder point. We can see at once that the on hand
inventory (or net inventory) cannot be used to rigorously define the reorder
point, since, if there was a very heavy demand during some cycle and a
huge number of backorders were incurred, then the arrival of whatever
outstanding orders there were might never bring the on hand inventory
back up to the reorder point again, and hence another order would never be
placed. This leaves only the inventory position (the amount on hand plus
on order minus backorders) as a suitable level for defining the reorder point.
The inventory position does provide a suitable level, since the difficulty
referred to above with the on hand level cannot occur. If during some cycle
there is very heavy demand and a considerable number of backorders are
incurred, it will also be true that a correspondingly large number of orders
will be placed, for the reorder point in terms of the inventory position will
be crossed a number of times.
If r is the reorder point in terms of the inventory position, then imme¬
diately after an order is placed the inventory position is Q + r. Thus the
inventory position must have one of the values r + 1,. . . , r + Q. It is
never in a state r for a finite length of time, since as soon as a demand occurs
which reduces the inventory position to state r an order is placed bringing
the state to r + Q. Note that specification of the inventory position tells
us nothing about the on hand inventory or net inventory. If the inventory
position is r + j, there may be no orders outstanding with the net inventory
being r + j, one order outstanding with the net inventory being r + j - Q,
etc. For Poisson demands, where there is a positive probability for an
arbitrarily large quantity being demanded in any time interval, it is the¬
oretically possible to have any number of orders outstanding at a particular
instant of time. (Of course the probability of a large number of orders
outstanding will often be very small.) In a case where there was a finite
upper limit to the amount that could be demanded in any time period,
then there could never be more orders outstanding than total to an amount
equal to the maximum lead time demand. On the other hand, specification
of the net inventory does uniquely determine the number of orders out-
 182 LOT SIZE—REORDER POINT MODELS CHAP. 4

standing and the inventory position, since if x is the net inventory and n
orders are outstanding, the inventory position is x + nQ, and this can be
written r + j,j — . fQ. This restriction determines both n and
j. The value of n is the largest non-negative integer which makes
r<x + nQ<r + Q.
In the simple approximate model discussed in Sec. 4-2, we assumed that
there was never more than a single order outstanding, and that the reorder
point could be based on the on hand inventory. We have just seen that, in
general such restrictions cannot be made rigorously. It may be justifiable
to stipulate, however, that the probability that more than one order is
outstanding is very small.
In order to compute P0ut, the expected number of backorders on the
books at any point in time, and the expected on hand inventory at any
point in time we need the state probabilities for the net inventory. The
straightforward way to compute these would be to attempt to write down

the difference equations (3-124) which describe transitions between the net
inventory states. Difficulties are encountered with this procedure, however,
since the lead times are constant and are not generated by a Poisson proc¬
ess. We shall therefore use a different approach, and instead we begin by
hnding the state probabilities p(r + j) that the inventory position of the
system is r+j,j = 1,. . ., Q. A knowledge of these permits the compu¬
tation of the state probabilities for the net inventory. The advantage of
doing this is that the nature of the procurement, lead times dot's not, enter
into the computation of the p(r + j).
In a time dt, the system inventory position moves from state r j■ j to
r+J ~ 1,3 > 2, if a demand occurs. This happens when a Poisson process
generates demands and units are demanded one at a time with probability
X di. If the system is in state r + 1 and a demand occurs, it, names to state
1 e SmCe the demand triggers the Pigment of an order. Thus we have
the diagram representing the transitions shown in Fig. 4-4
The balance equations (3-124) then read for this particular case
SEC 4-7 EXACT FORMULAS 183

XP(r +i + 1) = \p(r + j), j = 1, . . ., Q - 1

(4-38)
Xp(r + Q) = XP(r + 1)
Thus,
p(r + Q) = P(r + Q - 1) = . . . = p(r + 1)

and since £?-i p(? + j) = 1, it follows that the unique solution for p(r + j)
is

p(r + j) = j = l,.. .,Q (4-39)

We have shown that, in the steady state, each of the inventory position
states r + j has the same probability 1/Q which is independent of j. We
say that the states are uniformly distributed and that p(r + j) is a uniform
distribution.
To compute the state probabilities for the on hand inventory and the
number of backorders, consider the system at any instant of time t Also
consider the time t — r, r being the procurement lead time. Note that
everything on order at time t — r will have arrived in the system by time t
and nothing not on order at time t — r can have arrived in the system by
time t Thus, if the inventory position of the system was r + j at time
t — r, the probability that there are x units on hand at time t is the
probability that r + j — x units were demanded in the lead time r if
r + j — x > 0 and is zero otherwise. For Poisson demands, the probability
of having this number of units demanded is p(r + j — x; Xr) when
r + j — x > 0. However, the probability that the inventory position is
r + j at time t — r is 1/Q. Then if \f/i(x) is the state probability that x
units are on hand at any time t, to find \pi(x) we multiply p(r + j — x; Xr)
by 1/Q and sum over the j for which r + j — x > 0. Thus

1 Q
fifr) = q P(r + j - x; \t)
V j =1

1 r+Q^x
= q 22
u = r-\- 1—x
Xt)
(4-40)

= k Z)
^ Lu = r+1—x
Xr) ~ Z
u=r+Q+l—x
pfaXr)
q [P(r + 1 — X) Xr) — P(r + Q + 1 — x; Xr)], 0 < x < r+ 1

1 A \ r+Q—x
*&) = q Z p(r + 3 - Xt) = q Z
ti = 0
p(u> Xt)
(4-41)
= ^ [1 - P(r + Q + 1 - *; Xr)], r+l<x<r+Q
 184 LOT SIZE—REORDER POINT MODELS CHAP. 4

Here we have the state probabilities for the on hand inventory.

In the same way, if the system is in state r + j at time t — r, the prob¬
ability that there are y backorders at time t is p(y + r + j; \r),y > 0.
Thus the state probability fo(y) that there are y backorders at time t is

. . l -A i »+r+Q
&(y) = q 22 p(y + r + j- Xt) = - ^ p(U; Xt)
y= l ^ u = i/-f-r+l
l (4-42)
= Q lp(y + r + 1; Xt) - P{y + r + Q + 1; Xt)], y >0

These same state probabilities have been obtained previously in a different

form in [3].
With the help of Appendix 3 we can now easily compute the expected
values needed. First of all PQUt, the probability that there is no stock on
hand at time t, is

°° ^ i r*00 co
Pont = Z ^ = |_Z P(y + r+ 1; Xt) -J^P(y + r +
Q Q + l; Xt) J
If-" 00 00 (4-43)
=o Z L«=r+1
J2
t* = r+Q+l
^(«;Xt)1
J
and from A3-6

Pont = q [a(r) — a(r + Q)] (4-44)

where

a(p) = W=»+l
Z P(u’ Xt) = XtF(v; Xr) - vP(v + 1; Xt) (4-45)

Thus E(Q, r), the average number of backorders incurred per year is
by (4-29) ’

E(Q, r) - XPout = ^ [a(r) - a(r + Q)] (4-46)

The expected number of backorders on the books B(Q, r) at any time t

is by definition

B(Q> r) = J2 yi2(y)
y= 0

= Q^ Z0 yip(y + r + 1; Xt)
y=
- P(y + r + Q + 1; Xt)]

j r co ^

= Q LSi ('U~r - 1)P(M> Xt) - Z

u = r+Q+1
(u-r-Q- 1 )P(U; Xt)1
J

(4-47)
SEC 4-7 EXACT FORMULAS 185

E (u-v - 1 )P(u; Xt)

U = V+l

= 2 «P(«; Xr) - (v + 1) ^ P(u;\r) (4-48)

and by A3-8, A3-6

oo

E uP(u; Xt) = ■ P(v — 1; Xr) + \tP(v; Xt)

U =V+ 1

»(p + 1)
P(» + 1;Xt) (4-49)

(» +1 E )
li —»+l
p(«; M = (Xt)(» + 1)P(»; Xt)

- »(» + l)P(p+1;Xt) (4-50)

Hence if

m = E (« - » - Xr) = ^P(i> - 1; Xt)

- (Xr)yP(u; Xt) + 1} P(v + 1; Xt) (4-51)

P(Q, »•) = g [/3(r) - /3(r + Q)] (4-52)

We have now evaluated the expected number of backorders at any point

in time. As we observed in Sec. 4-6, B(Q, r) is also the average unit years
of shortage incurred per year.
It remains to evaluate the expected on hand inventory at any time. By
definition, the expected on hand inventory D(Q, r) is
r+Q
m, r) = e xxPi(x)
x=0

= Q +1— Xr) — P(r + Q + 1 — x; Xr)]

H-Q
+ E *[1 - P(r + Q + 1 - Xt)] l (4-53)
x —r+1 J

1 r r+Q r

= p,-! E 31 + E xp(r +1 ~ x> Xt)

^ lx = r+l x=Q

E xP(r + Q + l — z; Xt)
 186 LOT SIZE—REORDER POINT MODELS CHAP. 4

Now

1 r+Q
r* rv ! Q „ _

+ r («4>

Also

v v

^ ^ + 1 ~ *5 = £ zP(i; + 1 - x; Xr) = (v + 1 _ «)P(«; Xr)

X—1 14=1

«> w

= S (" + 1 ~ u^u> Xr) + £ (« - v - 1 )P(u; Xr) (4-55)

n=»4-l
Thus

m, r) = ^ + r

If" 00

+ (r + 1 - «)**(«; £ (r + Q + 1 - u)P(«; Xr)

1 14=1

oo

+ Z (« - r - l)P(tt; Xr) - £ (u-r-Q- 1 )P(„; Xr)l

“=r+1 «-r+Q+l J

However,
oo ^

1] (r + 1 - tt)P(«; Xr) - ^ (r + Q + 1 - «)P(u; \T)

W=1 14=1

— Q ^ P(w,\t) = —Q/x (4-56)

14=1

when ^ = Xr is the mean lead time demand. The last step follows from
A3-6. Then from (4-47), we see that

D(Q, r) = 1} + r - „ + B(Q, r) (4-57)

There is another simple procedure that one can use to compute the
expected value of the on hand inventory. By comparison of the result so
obtained with (4-57), an interesting result can be obtained. Note that the
expected value of the inventory position is the expected value of the on
and inventory plus the expected amount on order minus the expected
number of backorders, since the expected value of the sum of several ran¬
dom variables is the sum of the expected values. Thus the expected on hand
 SEC 4-7 EXACT FORMULAS 187

inventory is equal to the expected inventory position minus the expected

amount on order plus the expected number of backorders. Now the ex¬
pected value of the inventory position is

2> + j)p(r + j) = ~^2 (r + 3) = + r (4-58)

3=1 A

and so

D(Qj r) = S-i— -j- r — J -f- B(Q, r) (4-59)

where J is the expected amount on order. Comparison of (4-59) with

(4-57) shows that
J = m (4-60)

i.e., the expected amount on order is equal to the mean lead time demand.
This result, which we have just proved rigorously for the case under con¬
sideration, holds under much more general assumptions. The fact that the
expected amount on order should be the mean lead time demand can be
seen intuitively as follows. Imagine that orders flow into one end of a
pipeline and that procurements flow out the other end. Since all demands
are ultimately met, the mean rate of flow of units ordered into the pipeline
must be X. Since an order remains in the pipeline for a time r, the expected
number in the pipeline should be Xr = jjl.
All the terms in the average annual variable cost 3C have been evaluated;
3C consists of the expected ordering costs, holding costs, and backorders
costs. It is

3C = ^4 + /C | + r - M + B(Q, r)] + r£(«, r) + #B(Q, r)

(4-61)
= |4 + Jc[| + | + r-M] + tE(Q, r) + (* + IQBiQ, r)

where E(Q} r) and B(Q, r) are defined by (4-46) and (4-52) respectively.
We are here assuming that the unit cost C of the item is constant and is
independent of Q, so that the cost of the units themselves need not be in¬
cluded. We are also assuming that the cost of operating the information
processing system is independent of Q and r and hence need not be included
in 3C.
We have in the above derivations carried out the operations acting as if
r was positive. In actuality, with the proper interpretations, the above
formulas also hold when r is negative or zero. The proper interpretation
is to take P(v; Xr) = 1 if v < 0. We ask the reader to prove these state¬
ments in Problem 4-56. For example, if v is negative
 188 LOT SIZE—REORDER POINT MODELS CHAP. 4

a{p) = Xr + |w| (4-62)

Kv) = + Xr|»| + ^ (|»| — 1) (4-63)

In general, it is not very easy to determine the optimal values of Q and r

by use of the exact expression (4-61) for X, since E and B depend on Q and
r in a rather complicated fashion. Indeed, when the a(r + Q) and /3(r + Q)
terms are included, it does not seem easy to prove that there cannot exist
local minima different from the absolute minimum. To determine Q* and
r* in the case where the a(r + Q) and /3(r + Q) terms are included, it seems
necessary to use a digital computer along with an appropriate search
routine to find Q* and r*. A digital computer code to determine the Q and r
values which minimize the 3C of (4-61) has been developed by P. Teicholz
and B. Lundh of Stanford University. This program determines a relative
minimum but does not provide any guarantee that the minimum so ob¬
tained is the absolute minimum if there exist local minima, different from
the absolute minimum. In Sec. 4-10 we shall present an example, some of
the calculations for which were made with this code. It is worthwhile to
point out that although it seems difficult to rule out theoretically the
existence of local minima, the authors have been unable to detect the
existence of such in making numerical computations with specific problems.
Fortunately, it turns out that in practice it is seldom necessary to use
the exact formulation developed above. In the next section, it will be
pointed out that for almost all cases of practical interest, X reduces to a
form which makes it fairly easy to determine Q* and r*. About the only
time that the exact formulation is required is when it costs very little to
incur backorders. Such cases do not occur too frequently in the real world.

4-8 An Important Special Case

In problems of practical interest, it is usually true that the terms a(r + Q),
P(r + Q) in E(Q, r) and B(Q, r) respectively, are negligible. These terms
are important only if there is a significant probability that the lead time
demand will be greater than r + Q. If this happens, then there wfil stiff
be backorders on the books after the arrival of a procurement, i.e., the
procurement will not be sufficient to remove all the backorders. It will
never be optimal to have a sizeable probability that so many backorders
will be incurred that the arrival of an order will not be sufficient to meet all
of them unless it costs very little to incur backorders. This is seldom the
case for most real world problems, and hence in practice it is usually a very
good approximation to neglect a(r + Q) and /3(r + Q). When air + Q),
SEC 4-8 AN IMPORTANT SPECIAL CASE 189

P(r + Q) are negligible, a considerable simplification of X occurs. It can

be written

*-^ + Jc[§ + | + r-M] + ^«(r) + ^(^ + ICDPir) (4-64)

Then, if Q*, r* are the smallest values of Q, r which minimize X(Q, r) it is

necessary that

AqX(Q*, r*) = X(Q*, r*) - X(Q* - 1 ,r*) < 0)

> or Q* = 1 (4-65)
AaX(Q* + 1, r*) = X(Q* + 1, r*) - X(Q*, r*) > OJ

ArX(Q*, r*) = x(Q*, r*) - X(Q*, r* - 1) < 0}

) (4-66)
ArX(Q*, r* + 1) = X(Q*, r* + 1) - X(Q*, r*) > OJ

Thus Q* is the largest Q for which AqX(Q, r*) < 0, or Q* = 1, and r* is the
largest r for which ArX(Q*, r) < 0.
When X is given by (4-64)

AqX{Q, r) = [XA + Xx«(r) + (x + /C)/3(r)]^ - + y (4-67)

If AqX(Q, r) < 0, this implies that

Q(Q - 1) < J| [A + Mr) + (# +x 1C) d(r)] (4-68)

To determine ArX (Q, r) note first from (4-45) that

Ara(r) = —P(r; Xr) (4-69)

Also, from (4-51)

Ar/3(r) = ~~ p(r - 2; Xr) + r(\r)p(r - 1; Xr) - XrP(r - 1; Xr)

—^-p(r; Xr) + | [P(r + 1; Xr) + P(r; Xr)]

= P(r — 2; Xr) + rp(r; Xr) - XrP(r; Xr) — Xrp(r — 1; Xr)

p(r; Xr) + rP(r; Xr) - |p(r; Xr)

= (r — Xr)P(r; Xr) — rp(r; Xr) (4-70)

Thus

ArX(Q, r) = 7C + ^ {[—Xtt + (t + IC)(r - Xr)]P(r; Xr)

— (v + IC)rp(r; Xr)} (4-71)

 190 LOT SIZE—REORDER POINT MODELS CHAP. 4

and if ArJC (Q, r) < 0 this implies that

[‘ - ^ fr - *>] *> > ^ (4-72)

A computational procedure for finding the optimal Q and r is to begin

with Qw in (4-72) and determine the largest r satisfying (4-72). Use this r,
°al1 * rVn ('4’68-) t0 determine & by finding the largest Q which satisfies
(4-68). Then Q2 is used in (4-72) to determine r2j etc. Continue until there
is no change in r and Q.
Often it is possible to treat Q as continuous even though it is desirable
to treat r as discrete. For example, when Q gets to be of the order of 10 or
12, it is normally satisfactory to treat Q as continuous. However, r will not
usually be treated as continuous unless p is at least 25. For continuous Q
(4-68) becomes: ’

Q = Vlt [A + 7r“W + T'IC (4-73)

When Q is treated as a continuous variable and rasa discrete one, then

(4-73) is used with (4-72) to determine Q* and r* Although the task of
etermining Q* and r* for the case being considered here is considerably
more time consuming than for the simple approximate model discussed in
bee. 4-2, it can easily be carried out by hand in a short time

AMPLE The Milex Company sells reflector lenses at an average rate of

50 per year. Units are almost always demanded one at a time, and it is felt
that to an adequate approximation it can be assumed that a Poisson process
generates demands. Ordering and receiving costs amount to $100 for each
order The cost of each lens is $50, independent of the quantity ordered,
and the company uses an inventory carrying charge oil = 0.20. Demands
occurring when there is no stock on hand are backordered and the cost of a
backorder has been estimated to be $50 per backorder plus $500 per unit
year of shortage. The procurement lead time is always very close to 0 4
years. The company in the past has been using a periodic review system,
but has recently changed to using transactions reporting. Therefore the
company would like to install a lot size-reorder point system for controlling
e reflector inventory. Compute the optimal order quantity and reorder
point.

ab°7e’ we see that x = 50 units per year, A = $100,

it $10, x - $50, x = $500 per unit year, r = 0.4 years. To begin the
iterative procedure we compute

I2\A ,-
Qw —
-Jq = V1000 « 31.6
SEC. 4-9 THE NORMAL APPROXIMATION 191

It is seen that Qw is large enough that we should be able to treat Q as

continuous. However, the mean lead time demand p = Xr = 20 is small
enough that r should be treated as discrete. Thus equations (4-72), (4-73)
will be used in the iterative procedure.
Now observe that when Q = Qw = 31.6, and r = 31, (4-72) becomes

b - liwif]<si ■ ^w-013475'+(aoo5383)
(31-6) (10)
= 0.152 > 0.126
50(50)

However, if r = 32, the direction of the inequality is reversed. Thus the

largest r satisfying (4-72) when Q = Qw is 31, i.e., rx = 31. This value of r
was determined after trying out several different values.
Substituting rx into (4-73), we find that

Q2 = 10 jlOO + 50(0.0186) + ^ (0.0227) jj/2 = 31.8

On using Q2 in (4-72) we find that r2 = 31 so that r is unchanged. Hence

the iteration process terminates. If we round Q to 32 units and use this in
(4-72), r again turns out to be 31. Thus the optimal order quantity is 32
and the optimal reorder point based on the inventory position is 31. Since
the expected lead time demand is 20 units, the safety stock is 11 units.

4-9 The Normal Approximation

For large lead time demands the Poisson probabilities appearing in the
model developed in the last sections can be conveniently approximated by
a normal distribution. In such cases, it is usually a good approximation to
treat Q and r as continuous variables also. Here we shall develop the
equations corresponding to those in the previous section for the case where
the lead time demand can be considered to be normally distributed. Recall
from Sec. 3-14 that as v(xiu) approaches a normal distribution
with mean p and with variance a2 = p. In the development here, we shall
not specifically set cr2 = p in the normal distribution. The reason for this
is that later we shall want to consider the case of variable lead times. It
will be found that it is often desirable to represent the marginal distribution
of lead time demand by a normal distribution. However, in such cases, the
variance a2 will be greater than p because of the influence of variable lead
times. In this section, we shall imagine the lead times to be constant.
However, by not specifically setting a2 = p, and instead leaving <r as an
independent parameter, we shall simultaneously obtain the appropriate
 192 LOT SIZE—REORDER POINT MODELS CHAP. 4

equations that will be useful when considering the lead times to be random
variables also.
Since r and Q are now being treated as continuous the inventory position
can take on any value between r and r + Q rather than only the values
r + 1,. . . , r + Q. The probability that the inventory position lies be¬
tween x and x + dx, r < x < r + Q is simply dx/Q, because in Sec. 4-7
we showed that p(r + j), the probability that the system was in the state
r+j was l/Q. This follows by using the same trick of drawing in rec¬
tangles used in Sec. 3-14 to show how the Poisson distribution approached
the normal. Note that if the same procedure was used as in Sec. 3-14 the
inventory position would range from r + §tor + Q + i This gives a
slightly better approximation than using the range r to r + Q, but since
for practical purposes the difference is negligible, and since in the physical
system the inventory position cannot get above r + Q, we prefer to use the
range r to r + Q. Note that

dx _
Q ~ (4-74)

as desired. In place of the Poisson density p(x; M) for the probability that
x units are demanded in a lead time, we shall now use

(4-75)

as the probability density for lead time demand, where p. is the expected
lead time demand. In the following, we shall make use of the fact that when
it is valid to use the normal approximation, then one can assume that
$(-M/<r) = 1 (see Sec. 3-14).
Then, if &(a0 is the probability density for the quantity on hand at any
time t, it follows that

<ir~-)) 0<x<r (4-76)

Wl) “ i r-* (~^) * > ^ [l - * (L±-g - ^ - >)],

r < x < r+ Q (4-77)

Similarly, the density function My) for the number of backorders on hand
at any tune t is
SEC 4-9 THE NORMAL APPROXIMATION 193

Then

F°- ■ l My) iy - i [l * (!1±^-M) iy

Now by A4-6
-r*(-t<ity~,i)»] <«•)

Thus the expected number of backorders incurred per year is

£(<2, r) = XPout = ^ [a(r) - a(r + Q)] (4-80)

Where

«(») = tr0 ^ ^ ^ (v m) $ ( ^ ^ (4-81)

Next, the expected number of backorders at any time is

r) = JQ yMy) iy = ^ (r--^ ^—-) %

From A4-6 and A4-7

V>~l‘)i«} (4'82)

f »* (E±^-i;) * = I [«■ + <' - >w* (Vs)

- | (v - /t)0 (4-83)

Therefore

r) = ~~ 08(r) - /S(r + Q)] (4-84)

 194 LOT SIZE—REORDER POINT MODELS CHAP. 4

where

m = | [«* + (» - *)*]* - | (» - (4-85)

Finally
fr+Q Q
D(Q, r) = Jo a^i(x) dx = | + r- iL1 + £(Q, r) (4-86)

All the terms needed in the average annual cost X have been evaluated;
X is

~ QA + IC\_ 2 + r-/1J + r) + (* + IC)B(Q, r) (4-87)

The values of Q and r which minimize X will be solutions to

dX dX
dQ~ dr (4-88)

Problem 4-16 asks the reader to obtain all the first and second derivatives
of X.
In the usual case where the a(r + Q) and /9(r + Q) terms are negligible
and can be ignored, the equations of (4-88) reduce to

+ Mr)] + 2(x + IOftr)

Q= (4-89)

[x\ - (fr + IC)(r - M)]$ + (x + IC)c-<t> (^~~j = QIC (4-90)

The same sort of iterative scheme used to determine Q* and r* for the
simple model of Sec. 4-2 can also be applied here. That is, we first set
Q = in (4-90) and determine rh the value of r which satisfies this
equation. Then rx is used in (4-89) to yield Qi, etc. Now, of course, it will
be somewhat harder to solve (4-90) than it was to solve the corresponding
equation for the simple case.
At this point we might note that when x = 0, a(r + Q) is negligible, and
S
the contribution of B(Q, r) to the inventory carrying charge is negligible,
(4-87), (4-89), and (4-90) reduce to (4-16), (4-12), (4-13) respectively. This
s ows us that the model of Sec. 4-2 is applicable, when the assumptions
stated above apply, for any number of orders outstanding, and is not
restricted merely to the case of never more than a single order outstanding.
It is not simple to make numerical computations using the exact form
for X obtained in this section. When the a(r + Q) and /3(r + Q) terms are
included, 3C is not generally convex. This nonconvexity makes it difficult
 SEC. 4-10 AN EXAMPLE USING EXACT & 195

to rule out the existence of local minima which differ from the absolute
minimum. The nonconvexity has also caused convergence difficulties with
Newton’s method and the method of steepest descents in examples which
the authors have attempted to work out in which the a(r + Q) and
0(r + Q) terms were important. When the a(r + Q) and p(r + Q) terms
must be included, it would seem necessary to use a computer and an
appropriate search routine to determine Q* and r*. It is fortunate that in
practice it is seldom necessary to include these terms.

4-10 An Example Involving the Use of the Exact Form of X

Consider an item for which the demand over any time interval can be
considered to be Poisson distributed with X = 400 units per year. The
procurement lead time is a constant, and r = 0.25 yr. The values of the
other relevant parameters are A = $0.16, C = $10.00, I = 0.20, x =
$0.10, X = $0.30 per unit year. It is desired to determine Q* and r* when
a Qr operating doctrine is used to control the system, and Q*, r* are
determined by minimizing the exact expression for X given by (4-61).
This problem was solved on the Burroughs 220 computer using the code
referred to in Sec. 4-7. The results obtained were

Q* = 19, r* = 96, X* = $31.75 per year (4-91)

As was noted in Sec. 4-7, the code determines a relative minimum, but does
not guarantee that it is the absolute minimum, if there exists more than one
relative minimum. The authors could find no indication that there exist
other relative minima, and hence it seems very likely that the above values
are indeed the optimal values of Q and r. One finds that in X*, the various
cost components have the following values

^ A = 3.368 (4-92)

IC [If + r* “ M + B(Q*’ r+)] = 15-386 (4'93)

wE(Q*, r*) = 12.341 (4-94)

*B(Q*, r*) = 0.658 (4-95)

Observe that the backorder costs account for slightly over 40 percent of
the average annual cost. To show that the above answer does yield a
relative minimum, we present the values of X for neighboring (Q} r) points
in the following table
 19 6 LOT SIZE—REORDER POINT MODELS CHAP. 4

TABLE 4-1
Values of X (Q, r)

Q 18 19 20
r

95 31.94 31.91 31.95

96 31.90 31.75 31.91
97 31.99 32.11 32.27

Note that Q* is much larger than Qw which is

(4.96)

For this example, the a(r + Q), p(r + Q) terms are not completely negli¬
gible, but are not really large as compared to the a(r), fi(r) terms, when
evaluated at Q* and r*9 i.e.,

air*) = 6.204 a(r* + Q*) = 0.3241 (4-97)

0(r*) = 42.86 /3(r* + Q*) = 1.18 (4-98)

Even though the values of tt and f are absurdly small, the values of the
a(r "t" Q)j “b Q) terms evaluated at Q*, r* are not very large compared
to the a(r), p(r) terms. This suggests that the approximate model studied
in Sec. 4-8 which neglects the a(r -(- Q), fi(r + Q) terms should converge
and yield Q and r values close to Q* and r*. This is indeed correct. In fact,
the r value obtained from the approximate model is r*, and the Q value
obtained treating Q as continuous comes within one unit of Q*. The se¬
quence of Q and r values obtained starting with Qx = Qw = 8 in (4-72) and
then using (4-73) to obtain the new Q is: Qx = 8, rx = 105; Q2 = 12.9,
r2 = 101; Qz = 15, r3 = 99; Q4 = 17.5, r4 = 98; Q5 = 18.4, r5 = 97; Q6
= 19.4, r6 = 96; Qt = 20.3, 7*7 = 96. It should be pointed out that the
manual computations with the approximate model are quite arduous and
required about three hours of computation, whereas, working with the exact
cost expression, Q* and r* were obtained in only about 15 seconds on the
Burroughs computer (which is not a really high speed computer).
It is interesting that the approximate model did so well in this case
where t and f are exceptionally small. At the outset one might have
expected that it would do rather poorly. It would appear that the approxi¬
mate model of Sec. 4-8 is applicable under very general circumstances. It
might be noted that if we attempted to use the simple, approximate model
of Sec. 4-2, which neglects the backorders term in the carrying cost and
SEC. 4-11 THE LOST SALES CASE FOR CONSTANT LEAD TIMES 197

which flRsnmps tt = 0, we do not obtain convergence. The sequence of

iterations is: Qi = Qu, = 8, ?i = 102; Qz = 14, r2 = 95; Qz = 18, r3 = 87;
Qi = 24 and n should be the largest integer satisfying

P(r; 100) > ^ = | > 1 (4-99)

which is impossible.

4-11 The Lost Sales Case for Constant Lead Times

Interestingly enough, the exact equations for the lost sales case are much
more difficult to develop than the corresponding equations for the back¬
orders case. In fact, the exact equations have not been developed for the
lost sales case when more than one order is allowed to be outstanding
except for the case where Q = 1. In this section we wish to examine what
makes the lost sales case so difficult to treat, and to derive the exact
formulas for the case where only a single order can be outstanding. In
Sec. 4-13 we shall consider the case where units are ordered one at a time
(.Q = 1), but any number of orders are allowed to be outstanding.
Consider a system in which demands occurring when the system is out
of stock are lost. Again, as with the backorders case, we imagine that the
stochastic processes generating the demands and lead times are such that
the system can be described by a Markov process for which there exists
a steady state. To be specific, imagine that the procurement lead time is
a constant r and that a Poisson process generates demands with units being
demanded one at a time. As usual, assume that we desire to determine the
optimal values of Q and r when the system is operated using a lot size-
reorder point control policy.
If the inventory on hand plus on order is used to define the reorder point
r, then the quantity on hand plus on order fluctuates between Q + r and
r’+ i during each cycle. The states are not necessarily, however, uniformly
distributed. The reason for this is that the amount on hand plus on order
will not move from t -f- j to r -|- j — 1 when a demand occurs if the system
is out of stock. When the system is out of stock the amount on hand plus
on order does not change when a demand occurs. Unlike the inventory
position in the backorders case, it is not possible in the lost sales case to
treat the changes in the amount on hand plus on order independently of
the amount on hand. This is in part what makes it so difficult to treat the
lost sales case. It is necessary to take explicit account of the number of
orders outstanding and the times at which they were placed. However,
even if it was possible to obtain the distribution of the amount on hand
plus on order, additional difficulties would be encountered. The procedure
 198 LOT SIZE—REORDER POINT MODELS CHAP. 4

used to compute the distribution of the on hand inventory from the inven-
tory position in the backorders case will not work here. It is still true that
everything on order at time t — r will arrive in the system by timp t and
nothing not on order at time t — r can arrive by t. However, it no longer
follows that if the amount on hand plus on order is r+j at time t - r, the
probability that x > 0 units are on hand at time t is the probability that
»• + 3 — x units were demanded in the time r. The reason is that it is possi¬
ble for the system to run out of stock and have one or more demands occur
while the system is out of stock in the period t - T to t. Those demands
occurring when the system is out of stock are lost, and hence x units can be
on hand at time t even if more than r + j — x demands occurred in the lead
time.
In the backorders case there could be any number of orders outstanding
when a Poisson process generated demands. Since there cannot be back¬
orders m the lost sales case, the number of orders outstanding cannot be
greater than the largest; integer less than or equal to (Q + r)/Q = 1 -)-
(r/Q). The maximum number of orders which can be outstanding is there¬
fore determined by the values of r and Q. If r<Q then there can never
be more than a smgle order outstanding. In the lost sales case then, it is
possible to stipulate that there is only a single order outstanding if one
requires that r < Q.
We shall now derive the exact equations for the lost sales case when the
stipulation is made that there is never more than a single order outstanding
I o obtain the average annual cost we shall here first compute the expected
cost per cycle and then multiply by the average number of cycles per year
rather than attempting to find the state probabilities and compute the
costs directly as was done in the backorders case. The reason for changing
the approach is that it is not too easy to compute the state probabilities

Let us first compute the expected length of time per cycle during which
the system is out of stock. Since there is never more than one order out¬
standing, nothing is on order when the reorder point is reached, i.e., at the
reorder point r units are on hand. If the system reaches an out of stock
condition in the time interval t to t + dt after the reorder point is hit, this
means that m the time 0 to t, r - 1 units have been demanded and the rth
one is demanded between t and t + dt. This probability is \p(r - 1; \t) dt.
If the system does reach an out of stock position between t and t + dt it
will be out of stock for a length of time r - t during the cycle. Hence the
expected length of time out of stock per cycle is

-r X(T $ (r- 1)1 e X< dt = tP(T> Xr) -rP(r + l; Xr) (4-100)

SEC 4-11 THE LOST SALES CASE FOR CONSTANT LEAD TIMES 199

From (4-37) the average number of cycles per year is X/(Q + XT) where
T is given by (4-100).
It is very easy to compute the expected number of lost sales per cycle.
If the demand in the lead time r is x > r, the number of lost sales is x — r.
Thus the expected number of lost sales is

7; (x — r)p(x; Xr) = XtP(t — 1; Xr) — rP(r; Xr) (4-101)

x = r+1

by A3-10. However, by A3-13 we see that the expected time out of stock
(4-100) is the expected number of lost sales times the expected time between
demands 1/X. This correspondence is what one would expect intuitively.
It remains to compute the expected unit years of stock held per cycle.
First note that the probability that w units are on hand when the order
arrives is P(r; Xr), for w = 0, and p(r — w; Xr), 1 < w < r. These are also
the probabilities that w + Q units are on hand after the arrival of a
procurement. Thus the probability that v units are on hand after the
arrival of a procurement is

j P(r; Xr), v = Q
(4-102)
{p(Q + r — v; Xr), Q < v < Q + r

The expected unit years of stock held per cycle will be computed in two
parts. First, for the time period up to the time the reorder point is reached
(this time is a random variable), and second for the period of fixed length
t from the time the reorder point is reached until the next order arrives.
Given that v units are on hand after an order arrives, there will be v units
in stock until the first demand occurs, and there will be v — 1 in stock
from the time that the first demand occurs until the second occurs, etc.,
and r + 1 in stock from the time the (v — r — l)th demand occurs until
the (v — r)th demand occurs which reduces the on hand inventory to the
reorder point. The mean time between demands is 1/X. Thus, given that
the on hand inventory is v, the expected number of unit years of stock held
until the reorder point is reached is

— 1-h. ..H-t? — — r — 1)]

X

= ^ (v - r) - ^ (v - r - l)(t> - r) = ^ [v(v + 1) -r(r + 1)]

Averaging over the initial inventory, i.e., over v, it is seen that the ex¬
pected unit years of stock held until the reorder point is reached is
 200 LOT SIZE—REORDER POINT MODELS CHAP. 4

2 Q+r -j

2x 2 + !) “ r(r + 1)M(») = ^ [Q(Q + 1) - r(r + l)]P(r; Xt)

1 r_1
+ o7 X) t(r + C ~ w)(r + Q + 1 - «) — r(r + l)]p(w; Xr) (4-103)
A u= 0

The expected unit years of stock held from the time the reorder point is
reached until the next order arrives is the integral from 0 to r of the
expected amount on hand at time t, i.e.,
fr r— 1

/ J2 (r - x)p(x; \t) dt (4-104)

x—0

On expanding out (4-103) and (4-104) and summing the results, we find
that the.expected number of unit years of stock held per cycle is'(to~Be'
worked -out-in Problem 4-62) . -— -^

^ Q(Q + 1) + - Y + ^ p(r - - Y P(r; Xr) (4-105)

Therefore the average annual cost becomes . ’

x-«f^{A + rc[seM + « + T-t]

+ ^ [l*P(r - 1; Xr) - £ P(r; Xr) jj (4-106)

where x is the cost of a lost sale including the lost profit and T is given bjr
(4-100). We ask the reader to prove in detail in Problem 4-74 that mini¬
mization of the average annual cost with the cost of a lost sale being defined
to include the lost profit will yield the same Q* and r* as the maximization
of the average annual profit.
It is interesting to note that when T is negligible, (4-106) becomes simply
the discrete version of the simple lost sales model studied in Sec, 4-3. In'
practice it is rather difficult to determine the optimal values of Q and r
by use °f (4-106). Fortunately, for most real world computations, it is
almost always true that the simple model will suffice, and it is unnecessary
to use (4-106). The reader should recall, of course, that (4-106) holds rigor¬
ously only if never more than one order is outstanding, and that the theory
for more than a single order outstanding has not been worked out.

4-12 Stochastic Lead Times

In the simple approximate models we allowed the procurement lead time

to be a random variable with density function g(j). The exact formulations
SEC 4-12 STOCHASTIC LEAD TIMES 201

of the (Q, r) models presented in the last several sections have assumed
that the lead time was constant. We would now like to investigate the
problems involved in attempting to properly account for lead time vari¬
ations in the exact models. As long as there is never more than a single
order outstanding, no theoretical difficulties are encountered. If it is
imagined that the procurement lead time can be described by a random
variable with density function g(r), it is only necessary to average the ex¬
pected annual cost for a given r over r to obtain the appropriate average
annual cost. Thus if 3C(Q, r, r) is the average annual cost for a given r,
the appropriate average annual cost averaged over r is

X(Q, r) = jf “ X(Q, r, r)g{r) dr (4-107)

and 3C(Q, r) is the expression to be minimized. It is easy to see that 3C(Q, r)

will be obtained directly if in computing the various expected values one
uses the marginal distribution of lead time demand rather than the lead
time demand for a fixed r. The detailed proof of this for the various cases
studied is left for Problem 4-58.
Unfortunately, however, for (Q, r) models in which the demand is Poisson
distributed, and demands occurring when the system is out of stock are
backordered, it is not possible to specify rigorously that never more than a
single order is outstanding. For any time interval of length t > 0, there is
a positive probability that an arbitrarily large number of demands will
occur in this interval, and hence a positive probability that anyt given
number of orders will be placed in the time interval. This in turn implies
that there is a positive probability that any given number of orders will
be outstanding at any point in time. The only sort of statement that one
can legitimately make is that the probability that more than a single order
is outstanding is very small.
Let us now examine the situation where more than a single order is
allowed to be outstanding at any point of time. In the general derivation
of the backorders case we allowed an arbitrary number of orders to be
outstanding. When more than a single order can be outstanding, however,
difficulties are encountered in properly representing the lead times as ran¬
dom variables. We would like to treat the procurement lead times as inde¬
pendent random variables when more than a single order is outstanding,
i.e., we would like to assume that the lead time for a given order is inde¬
pendent of the lead times of the other orders which are outstanding.
However, if this is to be rigorously correct, then we must allow orders to
cross, i.e., they need not be received in the same order in which they were
placed. This is illustrated geometrically in Fig. 4-5. There is a positive
probability that order 2 will arrive before order 1. Order 2 can arrive be¬
fore order 1 if order 1 arrives between ir and t".
202 LOT SIZE—REORDER POINT MODELS CHAP. 4

In practice, it is almost always true that orders are received in the same
sequence in which they were placed, so that orders cannot cross. If this is
true, then lead times cannot be considered to be independent random
variables, i.e., the time of an arrival of an order placed at time t can depend
on the times of arrival of the other orders on the books when the order is
placed at time t For example, if orders are backed up at the source for
some reason, new orders will also have to wait in line until the previous
orders have been processed. Unfortunately, there seems no easy way to
handle situations where the lead times are not independent. In general, it is
very difficult to describe the dependence between lead times without
developing a detailed model of the source. No work has appeared in the
literature as yet which deals with models having the lead times being
dependent random variables.

We see therefore that it is very difficult to give a precise treatment of

variable lead times when more than a single order can be outstanding.
It might be pointed out that even if we imagined that orders could cross
and that the lead times were independent variables, it is not easy to develop
the exact equations in general. The reason is that it is hard to rigorously
allow for orders to cross. The reader should note that the procedure used
for the constant lead time case to obtain the state probabilities for the on
hand inventory assumes that all orders on the books when an order is
placed arrive before that order. It would not be correct, for example, if we
are going to allow orders to cross, to say that for the case of Poisson de¬
mands that the equivalent of (4-40), when the lead times are independent
random variables with density g(r), is

1 f* JL i Q
= QJ0 Z-/ P(r + i “ x'> Wfftf dr = - h(r + j - x) (4-108)
SEC. 4-12 STOCHASTIC LEAD TIMES 203

where h(r + j — x) is the marginal distribution of lead time demand. The

reason is that this approach assumes that orders arrive in the sequence
placed and do not cross. There is one special case which can be solved
exactly when the lead times are independent random variables and orders
can cross. This is when a Poisson process generates the lead times, i.e., the
lead time density is
g(r) = de

Then, for Poisson demands, the system can be solved by using Markov
analysis for a system continuous in time and discrete in space, since Poisson
processes generate all transitions. This has been carried out in [3]. The
details will not be presented here since the state probabilities become fairly
complicated. Furthermore the specific lead time distribution is not very
realistic, usually, from a practical point of view, and in addition, since
orders do not normally cross, the generalization is not particularly helpful
in the real world. We shall, however, illustrate the method of analysis for
a particularly simple case in the following section.
Often in the real world it is true that even although two or more orders
are outstanding at any point in time, the interval between the placing of
orders is usually large enough that there is essentially no interaction between
orders, and to a good approximation, it can be assumed that the lead times
are independent as well as assuming that orders do not cross. In this case
(4-108) is correct, i.e., to develop the appropriate expected values for the
cost expression it is only necessary to replace in the discrete case p(x; At)
by h(x).
When treating lead times as independent while simultaneously making
the assumption that orders do not cross, it is often convenient when demand
is Poisson distributed to assume that the lead time density can be described
by a gamma distribution. In practice, the lead time distribution, when
known, can often be fitted fairly well by a gamma distribution. From
Sec. 3-7 we know that if the random variable representing demand in any
time period is Poisson distributed and the lead time has a gamma distri¬
bution, then the marginal distribution of lead time demand has a negative
binomial distribution. Thus to compute the expected costs it is only
necessary to replace p(x; Ar) by bN[x; a + 1, + A)] in the expressions
for \pi(x) and (y)- For hand computations, of course, it is much more
difficult to work with the negative binomial distribution rather than the
Poisson since the negative binomial is not well tabulated. Intuitively, as
the variance of the procurement lead time distribution increases, the
variance of the distribution of states for the net inventory increases. This
in turn means that the safety stock and r must be increased. The compu¬
tation of Q and r treating the lead time as a constant equal to the mean
procurement lead time, when in actuality the procurement lead time is a
 204 LOT SIZE—REORDER POINT MODELS CHAP. 4

random variable, can lead to carrying a safety stock which is much too low.
The amount of the error increases as the variance of the lead time distri¬
bution increases. When the model is being used which treats Q and r as
continuous and assumes the demand variable to be normally distributed,
one usually assumes that the marginal distribution of demand is also
normally distributed and the variance of the normal distribution is chosen
to give the proper variance. In this case the formulas of Sec. 4-8 can be
used for variable lead times without modification.

4-13 Models with Q = 1

In certain real world situations it is optimal to order units one at a time as

demanded. This can be true, for example, if the demand for the item is
very low or the item is very expensive. The solution to the (Q, r) model of
Sec. 4-7 will indicate whether or not a Q = 1 is optimal. In this section we
shall study models in which it is required that Q = 1. These simple models
provide an excellent means for illustrating the effects of various ways of
handling stochastic lead times.
The backorders case will be considered first. The state probabilities and
cost expression can be readily obtained from the results of Sec. 4-7 by
setting Q = 1. However, it is instructive and also very easy to derive the
results directly, and hence we shall do this. Since Q = 1, i.e., an order is
placed each time there is a demand, then the inventory position must
remain constant. Denote the inventory position by R; the problem is to
determine the optimal value of R. For situations where this model is of
interest, it will be desirable to treat demands as discrete and R as an integer.
It will be assumed then that the number of units demanded in any time
interval has a Poisson distribution and that the mean rate of demand is X.
Suppose that the procurement lead time is a constant r. Then to com-
pute the probability Mx) that x > 0 units are on hand at time t, note that
everything on order at time t — r will have arrived in the system by time
t and nothing not on order at time t — r can have arrived in the system.
The inventory position at time t - r is R. Thus fc(%) is the probability
that R — x units were demanded in the lead time, i.e.,

fp(R — x) Xr), 0 < x < R

Mx) = (4-109)
|P(jB;Xt), x = 0

Similarly ^2 (y), the probability that there are y > 0 backorders at time t is

My) = p(R + y;\r), y> 0 (4-110)

SEC 4-13 MODELS WITH Q = 1 205

Thus Pout and the expected number of backorders B(R) are respectively

Po«t = X>(y) = P(fl;Xr) (4-ni)

y=0

oo oo

£(£) = y #*&/) = X) (y - Xr)

y=0 w = iE

= XrP(P - 1; Xr) - RP(R; Xr) (4-112)

The expected value of the on hand inventory at time t is

y x\pi{x) = y (P — u)p{u; Xr) = P — m + P(P) (4-113)

x=0 u=0

The average annual cost of backorders and holding inventory is therefore

3C(P) = IC(P - m) + ^rXP(P; Xr) + (IC + *) [XrP(P - 1; Xr)

- PP(P;Xr)] (4-114)

Note that the average annual ordering costs \A are independent of R and
hence need not be included in 5C.
If R* is the smallest P which minimizes 3Z(R), then it is necessary that
R* satisfy
A3C(R*) < 0; AX(R* + 1) > 0 (4-115)
Now
A3C(R) — IC — xXp(P — 1 ,* Xr) — (IC + f)P(R; Xr)

Thus R* is the largest R for which

P(P; xr) + p(r- i; Xt) > ^r+lc (4'116)

If A3C (P* + 1) = 0, then P* and P* + 1 both minimize 3C. When w = 0,

(4-116) reduces to finding the largest P such that

P(P;Xr)>T^ (4-117)

When t = 0, (4-116) reduces to finding the largest P such that

^ P(P; Xr) + p(P - 1; Xr) > ^ (4-118)

Having obtained the average annual cost and the optimality condition
for the ease of constant lead times, we now turn to the case where we allow
the lead times to be random variables. To begin, we shall imagine that lead
times are generated by a Poisson process, i.e., the lead time density is expo-
 206 LOT SIZE—REORDER POINT MODELS CHAP. 4

nential, with mean r = 1/8, and that the lead times are independent random
variables. This implies that orders can cross and need not be received in the
same sequence in which they were placed. The problem can be solved in
this case as a Markov process. Let the states be the net inventory (there
are an infinite number of these). The probability that the system moves
from state v to v — 1 in time dt is the probability that a demand occurs,
i.e., X dt. The probability that the system moves from state v to v + 1 in

FIGURE 4-6.

time dt is the probability that an order arrives. To compute this proba¬

bility note that if the system is in state v (net inventory = v) then R — v or-
ders are outstanding. The probability that any one of these orders arrives is
5 dt. The probability that some one of them arrives is therefore (R — v)8dt.
Only transitions to adjacent states correspond to the occurrence of a
single Poisson event, and hence only such transitions can occur in time dt.
The diagram representing the transitions then looks like that in Fig. 4-6.
If \p(v) is the steady state probability that the net inventory is v, the
balance equations are
 SEC 4-13 MODELS WITH Q = 1 207

\^(p + 1) + (R — v + l)8\//(v — 1) = [X + (R — v)5]ip(v),

v = R - 1, R — 2,. . .
X^CB) = - 1)
Then if

M = g = Xr

which is the mean lead time demand, we find on successive substitution

beginning with $(R — 1) = ^(R) that

\f/(R — &) = — ^(R), k = 1, 2,. . . (4-119)

Since

- k) = = *■*(«) = 1
fc = 0 jfc = 0 /c*

it follows that
<K-R) = e-* (4-120)
and from (4-119)

\p(v) = p(R — t;; Xr), y = i? — 1, R — 2,. . . (4-121)

Comparison with (4-109) and (4-110) leads to the interesting result that
precisely the same state probabilities are obtained as for the case of con¬
stant lead times, i.e., it makes no difference whether the lead time is a
constant r or whether the lead time is exponentially distributed with mean
r, provided that when the lead time is a random variable the lead times
corresponding to different orders are independent random variables. The
same optimal R will be obtained in either case. This result is peculiar to the
case where Q = 1. If Q > 1 it is not true that the state probabilities for
the constant lead time case are the same as where the lead times are
exponentially distributed, independent random variables. This is shown
in [3], referred to above, where the state probabilities are computed for
arbitrary Q when the lead times are exponential, independent random
variables.
An even more surprising result can be proved for the case where Q = 1.
It says that if demand is Poisson distributed, then when Q = 1, and the
lead times are independent random variables (i.e., orders can cross), (4-121)
gives the state probabilities for any lead time density g(r) with mean f.
In other words, the state probabilities and the optimal value of R are
independent of the nature of the lead time distribution if the lead times are
independent. The proof which we shall present is patterned after that of
Takacs [5].
 208 LOT SIZE—REORDER POINT MODELS CHAP. 4

Before presenting the proof we shall demonstrate an interesting property

of the Poisson distribution which will be needed. Suppose that we know
that n > 0 Poisson events have occurred in the interval 0 to t. Let us
compute the conditional probability that the first occurs between k and
k + dk, the second occurs between k and k + dk, etc., and the n-th occurs
between tn and tn + dtn, k < k < . . . < tn. The probability that n events
occur in the time interval and the ith one occurs between U and k + dk,
i = 1,.. . , n, must be

dk] dk]. . . dtn] [e-x«-fc>] (4-122)

since no event must occur for times £ in the intervals

0 <Z <k, k<£<k, . . ., tn<$<t

while event i does occur between k and k + dk. Equation (4-122) can be
condensed to
\ne~u dk ■ ■ . dtn (4-123)

Now the probability that n events occur in the interval 0 to t is simply

pin, \t). Hence the conditional probability that they occur as outlined
above, given that n events occur in the interval, is obtained by dividing
(4-123) by p(n; \t) which yields

dk dk. . . dtn (4-124)

We leave to Problem 4-63 the proof that n\/tn is a legitimate density func¬
tion. We note from (4-124) that since the density function is independent
of the k, the occurrences of the different events are independent random
variables. Furthermore, the probability that any one occurs between
k and k + dk is simply dk/t. To see how the n\ comes about in the joint
density it can be imagined that the time interval from 0 to i is divided up
to yield 2n + 1 boxes, n boxes corresponding to the time intervals k to
k + dk and the remaining n + 1 corresponding to the remaining n + 1
time intervals. The Poisson events can be imagined to be balls which are
tossed into the boxes. Now suppose that the n balls are tossed into the
oxes. All balls must go into one of the boxes, and for boxes corresponding
to the tune intervals k to k + dk only one ball can fit into a box. The
probability that one of the balls goes into the box corresponding to the
interval k to k + dk is then n dk/t. Given that one of the balls goes into the
box corresponding to the interval k to k + dk, the probability that one of
the n — 1 remaining balls goes into the box corresponding to the interval
k to k + dk is (n - 1) dk/t, etc. Thus the probability that one of the balls
goes into each of the n boxes corresponding to the intervals k to k + dk
SEC 4-13 MODELS WITH Q = 1 209

is (4-124). What we have shown is that if we are given that at least one
Poisson event occurs in the time interval 0 to t, then the probability that
any one occurs between £ and £ + d£ is d£/1, and this is independent of how
many events have occurred in the interval or of their times of occurrence.
Let us now return to the proof that the state probabilities are inde¬
pendent of the lead time distribution. The probability that a unit ordered
at time U will arrive by time t > U is
rt-u
S(t- ti) = I g(r) dr (4-125)

Then if we know that at least one demand has occurred in the time interval
0 to t, the probability that any one occurred between U and U + dti and
that the unit ordered arrives by time t is

Consequently, the probability that any particular order placed in the

interval 0 to t has arrived by time t is

J Jo S(t - ti) dti = } Jo S(£) d£ (4-126)

Suppose now that the system has been operating for a length of time t.
We assume that R units were on hand initially. Let us compute the proba¬
bility that the net inventory is x at time t if there have been in total n de¬
mands since the system began operation. If the net inventory is x, then
n + x — R of the orders have arrived and R — x have not arrived.
The probability of this, which will be written q(x] n, t) is simply the
binomial probability of n + x — R successes in n trials when the proba¬
bility of a success is given by (4-126), i.e.,
q(x; n, t) =

(B-x)[fjfs®*r',[rjfii-s®^r ^
On weighting q(x; n, t) by the probability of having n units demanded, we
find that Wx(t) the probability that the net inventory is x at time t is

^*(0 = X)
n = R~x
n> 0

[1 - £(£)] d$x-*
e-*Jo (4-128)
(R - a) I
210 LOT SIZE—REORDER POINT MODELS CHAP. 4

To determine lim*-** Wx(t) we note that

lim /" [1 - S(I)] df = f [1 - 5(f)] df = f -f d[l - 5(f)]

Jo Jo Jo
f00 JO /*«

“X *5{«-X (4-129)

Use of (4-129) in (4-128) shows that

lim Wx(t) = p(R - x; Xf) (4-130)

t—>oo

i*e., the limiting state probabilities are Poisson and they depend only on the
mean procurement lead time f.
We have demonstrated above that when Q = 1, demand is Poisson dis¬
tributed, and lead times are independent (and orders can cross), then the
state probabilities and the optimal R are independent of the nature of the
procurement lead time distribution. Let us now examine what happens
if we do not allow orders to cross but instead require that they arrive in
the same sequence as placed. The procedure suggested in Sec. 4-12 will be
used in which it is imagined that lead times can be treated as independent
and that they arrive in the same sequence in which they were placed. Then,
if g(r) is the lead time density, the average annual cost is given by (4-107).
Suppose now that g(r) is a gamma distribution, i.e., g{r) = y(r; a, 0).
Then, from (3-73), we know that the marginal distribution of lead time
demand has the negative binomial distribution bN[x; a + 1, p/(p + X)].
Hence the average annual cost has the form
oo

3Z(R) = IC[R — n] + \t a + 1, p/(J3 + X)]

x—R

00

+ (x + IC) (x - R)bN[x; a + 1, 0/(0 + X)] (4-131)

x—R

and by analogy with (4-116), R* is the largest R for which

* bN^R ~ 1; a + -1’ 0/(0 + x)]

+ ^2 bN[x] a + 1; 0/(0 + X)] > ^ p (4-132)

The optimal R computed from (4-132) will in general be larger than the R
computed from (4-116). It can be considerably larger if the variance of
the marginal distribution of lead time demand is much greater than the
variance of the Poisson distribution for the mean lead time. Thus, for the
case of Q = 1, the procedure suggested in Sec. 4-12 yields a higher variance
SEC 4-13 MODELS WITH Q = 1 211
of the net inventory distribution, and will in general yield a larger R* than
would have obtained under the assumption that orders can cross.
To close this section let us examine briefly the lost sales case. Even when
Q = 1, it is not easy to treat rigorously the case where more than a single
order can be outstanding (here R would have to be one in order to specify
rigorously that no more than one order could be outstanding—this case is
of no interest). However, the problem can be solved if it is assumed that
the procurement lead time has an exponential distribution with mean
f = 1/8 and that lead times are independent random variables (and orders
can cross). Markov analysis can be used to solve the problem just as it

could for the backorders case under similar assumptions. If x is the value
of the on hand inventory at any point in time, x can only have one of the
values 0, 1, 2, . . ., R. Fig. 4-7 shows the transition diagram from which
the balance equations can immediately be written down.
If \f/(x) is the steady state probability that x units are on hand at time t
then

+ 1 ) + (R-v + 1 )fy(p - 1) = [X + (R - v)f]f(v),

v = 1, . . . , R - 1 (4-133)
and
AiKR) = - 1), W(i) = R&KO)
212 LOT SIZE—REORDER POINT MODELS CHAP. 4

Thus

(4-134)
*(o) - (RT-
and since

M"

II
17 = 0

it follows that

Hence
*=[§(r;ri^r (4-135)

*(V) - (Xr) R > v-0,l,...,R (4-136)

17 = 0 U'

Multiplication of the numerator and denominator by e~XT shows that

p(R — v; Xf)
#(v) = v = 0, 1,. . ., R (4-137)
1 — P(R + 1; Xr)’

It has been shown in [1] that the are given by (4r-137) if lead times are
independent and orders can cross, for any lead time distribution g(r) such
that the cumulative distribution is continuous (the proof did not include
the case of constant lead times, however). Conny Palm [4] has also
presented a proof that the V'(j) are independent of the lead time distri¬
bution if lead times are independent and orders can cross. We shall not
present either of these proofs.
When the state probabilities are given by (4-137), the average annual cost
is
l-P(R;\t) ~ p CR;Xf)
3C = IC *4" Xt (4-138)
1 — P(R + 1; Xt) _ l-P(R + 1; Xf)

provided that the unit cost of the item is constant. We leave a discussion
of the determination of the optimal R to Problem 4-64.

4-14 Quantity Discounts

The preceding analysis has assumed that the unit price of the item was a
constant independent of the size of the order quantity Q. In the real world
quantity discounts are frequently available on the purchase of large quanti¬
ties. Two types of discounts sometimes encountered are all units and
incremental quantity discounts discussed in Secs. 2-11, 2-12 respectively.
Precisely the same computational procedures that were valid in the de-
SEC 4-15 CONSTRAINTS 213

terministic case can be applied to the case of stochastic demands discussed

in this chapter. The average annual cost curves optimized with respect to
r will again behave like those shown in Figs. 2-13, 2-15. It then follows
that the computational procedures given in Secs. 2-11, 2-12 also apply to
the case of stochastic demands. The only modification in wording necessary
is where it reads compute the optimal Q value, it should now read compute
the optimal Q and r values, and in the all units discount case, when the
cost is computed at a break point, the optimal value of r corresponding to
the order quantity at the break point should be used.
For the lost sales case, it will be recalled that the cost of the unit appears
in the stockout cost and hence the stockout cost will change as the unit
cost is changed. Usually, however, the effect of quantity discounts on the
stockout cost is negligible and can be ignored.

4-15 Constraints

The previous discussion in this chapter has assumed that the item under
consideration can be treated independently of any other items carried by
the system. When more than a single item is carried there will almost al¬
ways be some sort of interaction between the items. These interactions
may or may not have an important influence on determining the optimal
Q and r values. When they are not important, the models already discussed
may be used to determine Q* and r*. When the interactions are of conse¬
quence, things become more complicated, and some attempt must be made
to account for the interactions. Typical simple interactions are of the form
discussed in Sec. 2-8, i.e., competition for warehouse space, for the limited
number of orders that can be placed, or for the allowable total investment
in inventory. Formally, these constraints can be handled in much the same
way as in Sec. 2-8. However, as we shall see, this formal generalization
may not always be satisfactory. It will also become apparent that it is not
a simple matter to develop a satisfactory means for treating these con¬
straints when demand is a stochastic variable.
Assume that the system stocks n items. Consider first the case where
there is a limitation on the total number of orders (or setups) which can be
handled per year. Immediately a problem of interpretation arises. Because
of the stochastic nature of the demands, it is theoretically possible (if the
demand variable is being described by a Poisson or normal distribution) to
have any arbitrarily large number of orders placed in one year. Thus one
cannot guarantee that there will never be more than a given number of
orders placed per year. One can, however, talk about the probability that
no more than a given number of orders will be placed, or we can require
 214 LOT SIZE—REORDER POINT MODELS CHAP. 4

that the expected number of orders placed be less than or equal to a given
value.
The situation where one places a limitation on the expected number of
orders placed is the easiest to handle. The expected number of orders
placed per year for all n items is

(4-139)

The computational procedure is to first solve the problem ignoring the

constraint, i.e., compute the Qj and r;* values using one of the models given
in the previous sections. Then if N < h, h being the limitation on the ex¬
pected number of orders placed per year, the solutions so obtained are
optimal. If N > A, the constraint is active. Then from the theory of
Lagrange multipliers, we know that we should consider the function

= E rj) + n (jt ^ - a) (4-140)

where n is a Lagrange multiplier. Then the set of Qj and rs which minimize

the average annual cost E?-i X,<Qy, r,) subject to the constraint N = h is
the same set which minimize when n is adjusted so that N — h. Now since
Qi, and r} appear in JF only in the terms

fj) — r,) + ij j = 1,... ,n (4-141)

it follows that for any v, J will be minimized when each of the n expressions
rj) is minimized. The computational procedure then is to minimize
each 3j(Qj, Tj) for a given 17 and thus determine a set of Qj and rj; compute
N and compare it with h, select a new i) and repeat the process until N = h.
The set of Qj, rj so obtained will be optimal.
It must be recognized that the above procedure only sets an upper limit
on the expected number of orders placed per year. In any given year the
number of orders placed could be greater than h or less than h. The compu¬
tation of the probability that precisely M orders will be placed is not
especially easy, in general, when there are n items, although it is easy
for a single item in certain cases. Let us then first consider the case of a
smgle item when a Poisson process generates demands, units are demanded
one at a time, and all demands occurring when the system is out of stock are
backordered. Thus if the inventory position is r + j at the beginning of
the year, precisely m orders will be placed (to > 1) if the demand lies in
the interval (to - 1 )Q + j to mQ+j- 1. The probability of this is

P[(to — 1)Q -(-j; X] —P(mQ -\~j;\)

 SEC 4-15 CONSTRAINTS 215

No orders will be placed if the demand is less than j. The probability of

this is
1 -PU;\)
Now the probability that the system will be in state j at the beginning of
the year is 1/Q. Hence 0(m), the probability that precisely m orders will
be placed, is

1 Q
0W = 0 E {p[(m - 1 )Q + j; X] -P(mQ + j; X)}, m > 1
^ y=i

q1 E P(u;\)-2 P(u; X)
^ ^u=(m-l)Q+l u = mQ+l

+ E P(u> X)1 (4-142)

t£=(m+l)0+l J
and from A3-6

Q(m) = q (XP[(m - 1 )Q; X] -(m - l)QP[(m - 1)Q + 1; X]

- 2XP(mQ; X) + 2mQP(mQ + 1; X) + XP[(m + 1)Q; X]
— (m + 1 )QP[(m 1)Q -|- 1; X]}, m > 1 (4-143)
When m = 0, we see that

0W=xt[i-P(i;x)]
(4-144).
{X - XP(Q; X) + QP(Q + 1; X)}

Equations (4-143) and (4-144) thus give the probability that precisely m
orders are placed in a one year period for a single item.
When there are n items it is much more complicated to find the prob¬
ability that precisely M orders are placed. Let be the probability
that precisely m,* orders are placed for item j in one year. Assuming that
demands for the various items are independent, we see that the probability
that mi orders are placed for item 1, m2 for item 2, etc., is

II 0y(m3) = 0i(mi) 02(m2) . . . 0„(m„) (4-145)

y=i

where n is used to denote a product. Thus the probability Q(M) that

= M is the sum of (4-145) over all integers m, > 0 such that
x wv = M. Clearly, the resulting probability distribution is not es¬
pecially simple in form. The easiest way to generate the probabilities is by
 216 LOT SIZE—REORDER POINT MODELS CHAP. 4

use of generating functions. We leave a discussion of this to Problem 4-59.

There does not seem to be any simple explicit form for the distribution
unless n is large, in which case Q(M) should be approximately normally
distributed by the central limit theorem.
As suggested above, another approach to dealing with a restriction on
the number of orders placed is to make explicit use of the probability
density Q(M) that M orders are placed. We might for example assume
that there is an additional cost G for each order placed over a given num¬
ber h. The expected extra ordering cost is then
00

(M -h) 6(M) (4-146)

M~h

Where Q(M) is a function of Qi, . . ., Qn but not of the 77. In this case
there is no constraint. The cost function to be minimized is
n 00

= E rs) + G Y, (.M - h) 6(M) (4-147)

i=i M=h

Still another procedure which makes use of Q(M) would be to minimize

rj) subject to the constraint that the probability that more
than h orders are placed is less than or equal to a, i.e.,
00

X) Q(M)< a (4-148)
M = k+1

This constraint, if it is active, could be handled by the Lagrange multiplier

technique.
The last two procedures suggested cannot easily be used analytically
even when n is large enough to assume that 0(M) is normally distributed.
The reason for this is that B(M) will depend on all n of the Qy. Hence it
does not become possible to solve for one set of Qy, 77 independently of the
others. They will all be coupled together and this makes the computa¬
tional problem much more difficult. From a practical point of view then,
about the only thing that one can do is to use the constraint on the expected
number of orders placed. This is not too satisfactory unless one knows
about the sort of fluctuations there can be about the average. This cannot
be easily determined analytically. Thus it is difficult to handle this con¬
straint in an entirely satisfactory way.
The same sorts of problems are encountered in handling the other types
of constraints discussed in Sec. 2-8, i.e., constraints on floor space and
investment in inventory. We leave a discussion of these to Problems
4-60 and 4-61. Fortunately, in practical situations, the interactions between
items giving rise to the types of constraints considered here often turn out
 SEC. 4-16 AVOIDING SPECIFICATION OF STOCKOUT COSTS 217

to be not too important and can be neglected. Hence each item can be
studied independently of the others and the models developed in the
previous sections can be applied.

4-16 Determination of Operating Doctrines without

Specifying Stockout Costs

Inasmuch as it can be very difficult to assign numerical values to the

stockout costs in the real world, it is interesting to examine procedures for
determining operating doctrines which do not require that one assign
explicit values to the stockout costs. A procedure which immediately
comes to mind is that of minimizing the average annual costs of ordering
and carrying inventory, subject to the constraint that the average fraction
of the time out of stock is not greater than a fixed value. This criterion
has been suggested in Chapter 1. It can be used either in the backorders
or lost sales cases. Another criterion which is applicable to the backorders
case only is to minimize the average annual costs of ordering and carrying
inventory subject to the constraint that the expected number of back¬
orders on the books at any point in time is not greater than a specified
value. We would now like to investigate in somewhat more detail these
two alternative criteria for determining values of Q and r.
Consider first the case where there is a constraint on the average fraction
of time that the system can be out of stock. The average fraction of time
which the system is out of stock is what we previously defined as Pout.
It is also the expected number of backorders or lost sales incurred per
year divided by the average rate of demand, i.e., E(Q, r)/X. Thus we wish
to minimize (in the backorders case)

|A+/c[|+r-M + B(Q, r)] (4-149)

subject to the constraint

E(Q, r)
X </ (4-150)

where f is the upper limit to the average fraction of time for which the
system is out of stock. Now it is clear that if the costs (4-149) are to be
minimized, the average fraction of time out of stock should be as large as
possible. Hence the constraint (4-150) will be active, i.e., E(Q, r) = X/
(or when Q and r are treated as discrete, E(Q, r) should be as close to X/ as
possible while not being greater than X/).
To minimize (4-149) subject to E(Q, r) = X/, we know from the theory
of Lagrange multipliers (see Appendix 1) that we form the function
 218 LOT SIZE—REORDER POINT MODELS CHAP. 4

F{Q, r, 8) =

+ ?•-;* + B(Q, r) J + 8 [E(Q, r) - X/] (4-151)

where 8 is a Lagrange multiplier, and Q*, r* and 8* will be solutions to

§l = §l = dF_
dQ dr dd (4-152)

since we do not expect the optimal values of any of the variables to lie on
the boundaries.
Note, however, that minimizing F(Q, r, 6) for a given d will yield the
same Q*{6) and r*(6) as minimizing

3C-^A+/c[| + r- M + B(Q, r) J + 8E{Q, r) (4-153)

since 8\f is independent of Q and r. Hence, to determine Q* and r* we can

first determine the functions Q*(8) and r*(6) by minimizing (4-153) and
then selecting that value of 6, i.e., B*, for which E[Q*(d), r*(8)] = X/.
The values of Q*(B) and r*(8) evaluated at 8* are Q* and r* respectively.
We can now make the interesting observation that 3C is simply the
average^ annual cost, including backorder costs if for the backorder cost
x = 8, x = 0. Thus we can use the models developed in this chapter to
handle the present case. Indeed specification of the average fraction of
time out of stock is equivalent to having a backorder cost with # = 0 and
x — 8*. Note that the imputed value of x will be uniquely determined by/.
The computational procedure to determine Q* and r* is, however, more
complicated when/ is specified than when x and x are given. The compu¬
tational procedure is as follows. Set x = 0 and select an initial estimate
of x, say 80. Compute Q*(60), r*(80) using whatever model developed in
this chapter may be appropriate. Also compute E[Q*(80), r*(80)]/X = f0.
If /o > / select a new value of x, say 8h such that 81 > 80. If /„ < / select
a 01 < 00 p* f° = f> the Q and r values are optimal and no additional
computations are needed). Then determine

Q*(eO,r*(di) and #[Q*(0i), r*(^)]/X = /t

and repeat the above procedure. As additional computations are made,
it becomes possible to quickly zero in by interpolation on the value of 8
such that the expected fraction of the time out of stock is the desired value.
It should be clear that the procedure of assigning a value to / in reality
assigns a value to the backorder cost also, so that in the end one has
implicitly assigned a numerical value to the backorder cost. The method
of handling the lost sales case is precisely the same as for the backorders
case examined above.
 SEC 4-16 REFERENCES 219

Consider next the case where the constraint requires that the expected
number of backorders at any point in time be less than or equal to a
specified value. Again it is clear that the constraint will be active, and
will have the form B(Q, r) = 8. Thus we wish to minimize (4-149) subject
to this constraint. From the theory of Lagrange multipliers and what has
been shown above, this is equivalent to minimizing

X = ^A + + r — m] + (v + IC) B(Q, r) (4-154)

with respect to Q and r for a given 77, thus yielding Q*(r)) and r*(j)), and
then determining the optimal value 77* of the Lagrange multiplier 77 such
that £[$*(77), t**(t7)] = 8. Then $*(77*) and r*(7j*) are the optimal values
of Q and r. The numerical procedure for making the computations is the
same as that outlined above. It follows that specification of the expected
number of backorders is equivalent to setting x = 0 and uniquely deter-
mining x. Thus by specifying 8 we implicitly determine a unique value
of x when x = 0.
A more general procedure than the two possibilities discussed above
would be to specify both the average fraction of the time which the system
could be out of stock and the expected number of backorders at any point
in time. The theory of Lagrange multipliers immediately shows that this
would impute nonzero values to both x and x, since in this case two
Lagrange multipliers would be needed.

REFERENCES

1. Arrow, K. J., S. Karlin, and H. Scarf, Studies in The Mathematical

Theory of Inventory and Production. Stanford, California: Stanford
University Press, 1958.
2. Fetter, R. B., and W. C. Dalleck, Decision Models for Inventory Man¬
agement. Homewood, Illinois: Richard D. Irwin, Inc., 1961.

3. Galliher, H. P., P. M. Morse, and M. Simond, “Dynamics of Two

Classes of Continuous Review Inventory Systems,” Operations Re¬
search, Vol. 7, No. 3, June 1959, pp. 362-384.

4. Palm, Conny, “Analysis of the Erlang Formula for Busy-Signal Arrange¬

ments,” Ericsson Technics, Vol. 6, 1938, p. 39.
5. Takacs, L., “On the Generalization of Erlang’s Formula,” Acta Mathe-
matica, Academiae Scientiarum Hungericae, Tomus VII, 1956, pp.
419-432.
220 LOT SIZE—REORDER POINT MODELS CHAP. 4

6. Whitin, T. M., The Theory of Inventory Management, Rev. Ed. Princeton,

New Jersey: Princeton University Press, 1957.
7. Whitin, T. M., and J. W. T. Youngs, “A Method for Calculating
Optimal Inventory Levels and Delivery Times,” Naval Research
Logistics Quarterly, September 1955, pp. 157-173.

PROBLEMS

4-1. Derive the equivalent of the model presented in Sec. 4-2 when a
Poisson process generates demands, units are demanded one at a
time, and the procurement lead time is a constant r. Treat Q and
r as discrete variables. Show that the following numerical procedure
can be used to determine Q* and r* To begin, set Q = Qw in the
following expression and determine the largest integer r, call it r±
which satisfies this inequality

P(r;Xr)>^
7TA

Then substitute into

2X
Q(Q !) < J£f + 7r[XrP(r; Xr) — rP(r + 1; Xt)]}

and determine the largest Q, call it Q2, which satisfies this expression.
Use Q2 in the previous relation and determine r2 by finding the
largest r which satisfies the inequality, etc.
4-2. Derive the equivalent of the lost sales model presented in Sec. 4-3
when a Poisson process generates demands and the procurement lead
time is a constant r. Develop a numerical procedure for determining
Q* and r*.

4-3. Modify the results of Problems 4-1, 4-2 for the case where Q can be
treated as continuous, but r is to be treated as discrete.
4-4. Draw the equivalent of Fig. 4-2 for the lost sales case and prove that
the iterative scheme suggested in Sec. 4-4 converges in this case.
4-5. A function of two variables X(Q, r) is said to be convex over some
region of the Qr-plane if for any two distinct points (Qlt n) (Q2, r2)
m the region and for any a, 0 < a < 1,

3C[>Qi + (1 — ol)Q2, an + (1 — a)r2]

< aX.(Qh n) + (1 - a)X(Q2, r2) (4-155)

 PROBLEMS

If we think of X as being a surface in three dimensions, then X is

convex if the line joining any two points on the surface lies on or
above the surface. Intuitively, a convex function will then have a
shape of a bowl. X is said to be strictly convex if the strict inequality
holds in Eq. (4-155) when 0 < a < 1. The convex function X is
said to be strictly convex with respect to one of the variables, say Q,
if for any points (Qh r), (Q2, r), Qi ^ Q2 and any r, the strict in¬
equality holds in Eq. (4-155) for 0 < a < 1. Prove that the sum of
convex functions is also a convex function. Prove that the absolute
minimum of a strictly convex function over some region in the
Qr-plane is unique. Show that if X is convex and if the absolute
minimum of X over some region is taken on at two distinct points
(Qi> r0, (Qi, r2), then it is also taken on at any point [aQ1 + (1 — a)Q2,
an + (1 — a)r2], 0 < a < 1. Prove that any relative minimum of
a convex function is also the absolute minimum
4-6. Prove that
_1
(x — r) h(x) dx
Q
is a convex function of Q and r for Q > 0, r > 0 when h(x) is a prob¬
ability density function. Prove that it is strictly convex if h(x) is
everywhere positive over the allowable range of x. Note that the
result still holds for all non-negative r < rm if h(x) = 0, x > rm and
the upper limit in the integral is replaced by rm (i.e., in the context
of an inventory problem, there is an upper limit to the lead time
demand). Note that the above results depend only on the non¬
negativity of h(x) and not on its shape.
4-7. By making use of the results of Problems 4-5, 4-6, prove that the cost
functions for the backorders and lost sales models presented in Secs.
4-2, 4-3 are convex. Also show that they are strictly convex if h(x)
is a normal distribution, and in this instance show that Q* and r*
are unique. Show that in any case X(Q, r) is a strictly convex func¬
tion of Q, so that there cannot be two or more different optimal
solutions with the same r* but different Q values.
4-8. The results of Problems 4-6, 4-7 show that Q* and r* are unique
when h(x) >0 for all a; in the range of interest. However, Q* and
r* are unique even if h(x) =0 over some interval. Suppose for
example that h(x) has the form shown in Fig. 4-8 and h(x) = 0
between r0 and n- Sketch the shape of H(x) and the shape of the
curve defined by Eq. (4-13) or (4-22). What is the slope of the curve
defined by Eq. (4-13) or (4-22) between r$ and rj? Now prove that
Q* and r* must be unique. Hint: For r in the interval 0 to r0,
 222
LOT SIZE—REORDER POINT MODELS CHAP. 4

3C(Q, r) is strictly convex. Similarly X(Q, r) is strictly convex for r

m the interval rx to rm. This shows (why?) that if the minimum is
not unique, two or more r values in the interval r0 to rx must yield
the minimum value. Show by finding the slope of the curve corre¬
sponding to Eq. (4-12) or (4-21) that this cannot happen.
4-9. Under what conditions is there no solution to Eqs. (4-12), (4-13)?
How can this be interpreted on Fig. (4-2)?
4-10. Prove that there always exists a solution to Eqs. (4-21), (4-22).
4-11. Modify the derivations presented in Secs. 4-2, 4-3 for the backorders
and lost sales models to show that they hold for any number of
orders outstanding, provided that r is based on the inventory posi¬
tion in the backorders case and the quantity on hand plus on order

in the lost sales case. Are any additional assumptions required in

either case?

4-12. A warehouse which deals in hardware products services a large

number of hardware stores in a given city. Solder of a certain type
IS sold by the warehouse to the stores in a carton containing two
dozen rolls of solder. It is almost always true that a store will not
place an order for more than a single carton. The warehouse will
not ship less than a carton and hence it almost always ships a single
carton per order. The inventory manager at the warehouse is plan¬
ning to institute a lot size-reorder point system to control the
inventory of this particular type of solder. An analysis of historical
records shows that the procurement lead time demand for the solder
is approximately normally distributed with mean of 125 cartons
and standard deviation of 25. The average yearly rate of demand
 PROBLEMS 223

is 500 cartons per year. He estimates that the paper work cost in
preparing an order invoice is $1.50 and the cost of receiving an
order and placing it in the storage bin, along with the accompanying
paper work on receipt of procurement, is $2.00 plus $0.15 times the
size of the order. A carton of solder costs $12.00. The warehouse
uses an inventory carrying charge of I = 0.18. All demands occurring
when the warehouse is out of stock are backordered. It is estimated
that the goodwill cost plus the cost of writing a special letter to the
hardware store if a demand occurs when the system is out of stock
amounts to $25.00. Compute the optimal order quantity and re¬
order point for the warehouse. What is the cost of uncertainty?
4-13. For the example worked out in Sec. 4-4, determine howQ* and r*
vary with A. Allow A to vary from $,500 to $50,000 and plot the
results.
4-14. A large discount house is planning to install a (Q, r) system to con¬
trol the inventory of a particular model of an AM-FM radio. The
number of units demanded has been found to be essentially Poisson
distributed with a mean of 5 radios per week (this does not include
the Christmas season demand which will be handled separately).
The procurement lead time is essentially constant and requires 3
weeks. Each radio costs the store $40.00. An inventory carrying
charge of I — 0.20 is used. If a demand occurs when the model is
not in stock, the customer will almost always go to another of the
nearby discount stores rather than selecting a different model or
waiting until a new shipment arrives. The store has decided that
each such demand occurring when the model is not in stock costs
the store a goodwill loss of $20.00 plus a loss in gross profits of
$25.00. The total cost of placing an order is estimated to be $3.00.
Determine Q* and r*. What is the cost of uncertainty? On the
average, how many lost sales will be incurred per year?
4-16. A military supply center stocks air bearing gyros for a ballistic
missile. The mean rate of demand for these gyros has been 100 per
year over the past three years and the number of units demanded
appears to be described quite well by a Poisson distribution. The
procurement lead time is essentially constant and has the value
6 months. The gyros cost $2000 each. The cost of placing an order,
incoming inspection, etc., is estimated to be $100.00. An inventory
carrying charge of / = 0.20 is used. All units demanded when the
system is out of stock are backordered. It is difficult to estimate
the cost of being out of stock. Instead, the requirement has been
made that the probability of being out of stock must not be greater
than 0.0005. If the system is to operate using a (Q, r) model, de-
 224 LOT SIZE—REORDER POINT MODELS CHAP. 4

termine Q* and r*. What is the imputed cost of a backorder? What

is the cost of uncertainty?
4-16. Compute all the first and second derivatives of the cost function
3C(Q, r) given in Eq. (4-87).
4-17. Assume that X(Q,r) is approximated by the nonhomogeneous
quadratic form

3C(Q, r) « X(Q0, r„) + (Q ~ Qo) + ^ (r - r0)

1 3f3C ,q d2x , , 1 d2X .

+2 W Qo)2 + (Q ~ Qo)(r To)2
dQodr0 ro)+2^(r
where the derivatives are evaluated at Q0, r„. Show that the set of
equations whose solution minimizes the quadratic form are precisely
the same set of equations obtained using Newton’s method to solve
the equations dX/dQ = dx/dr = 0 when Q0, r0 are the initial guess
at the solution.
4-18. Show that X(Q,r) in Eq. (4-87) is not a convex function of Q and r
in general. Hint: Is it a convex function of r for a given Q?
4-19. Solve Problem 4-12 under the assumption that the cost of a back¬
order is $5.00 plus $10.00 per carton per week rather than a fixed
charge of $25.00 per backorder.
4-20. Solve Problem 4-15 under the assumption that the cost of a back¬
order is $1000 plus $5000 per unit week of shortage.
4-21. Consider an item for which the demand can be treated as deter¬
ministic, the demand rate being 500 units per year. The item’s cost
is $20.00 per unit and the system which stocks the item uses an
inventory carrying charge of I = 0.20. The cost of placing an order
is $40.00. The procurement lead time is not a constant. However,
it is always either 3 weeks or 6 weeks, the probability that it is 3
weeks being 0.6. All demands occurring when the system is out of
stock are backordered, and the cost is $250 per backorder. The
inventory is controlled using a {Q, r) system. Determine Q* and r*.
What is the cost of uncertainty?
4-22. The housewares department of a large retail store is installing a
(Q, r) system to control certain of the items it stocks. The procedure
is to have the sales people count these critical items at the end of
each day, and if any level has dropped to the reorder point an order
is placed the next day. Occasionally, there is an overshoot of the
reorder point, but it has been decided that its effect on the results
will be small and can be ignored. An analysis of sales data over the
 PROBLEMS 225

past year (excluding the Christmas season) has yielded the following
for weekly demand on a particular electric coffeemaker:

Number of units demanded 0 1 2 3 4 5 6 7

in one week

Frequency 2 6 3 8 2 4 2 1

There were never more than 7 units demanded in one week. The
unit cost of the coffee pot is $8.00 and the store uses an inventory
carrying charge of I = 0.13. The procurement lead time is one
week. The cost of placing an order is estimated to be $0.50. The
store feels that because of goodwill loss, the cost of a lost sale includ¬
ing lost profit is $10.00. Compute Q*, r* under the assumption that
the above data represent the true probability distribution for de¬
mand. What are the values of Q*, r* if one assumes that the
number of units demanded per week is represented by a Poisson
distribution whose mean is that obtained from the data given above?
4-23. An automotive repair shop installs new mufflers on autos. Past
history indicates that the number of units demanded for a certain
model of muffler is Poisson distributed with a mean of 1 per day.
The procurement lead time is always either 8 or 15 days, the prob¬
ability of 15 days being 0.7. A muffler costs the shop $6.00 and it
uses an inventory carrying charge of I = 0.20. The cost of placing
an order is estimated to be $1.00. Requests for muffler changes
which occur when the dealer is out of stock are taken elsewhere,
and the goodwill loss plus lost profits is estimated to be $25.00. If a
(Q} r) system is used to control the inventory of this muffler, what
are Q*, r*? What is the average annual cost of uncertainty? What
assumption was made about the nature of the lead times? Is it valid?
4-24. Work out the equations which determine Q* and r* for the simple
model discussed in Sec. 4-2 when

h(x) = \
0

i.e., h(x) has an exponential distribution. Is there anything unusual

about the resulting equations?
4-25. Solve the example presented in Sec. 4-4 when the marginal distribu¬
tion of lead time demand has an exponential distribution rather
than a normal distribution. Assume that the mean of the exponen¬
tial is the same as the mean of the normal, and that all cost data
 226 LOT SIZE—REORDER POINT MODELS CHAP. 4

given in the example apply here. Use the results of Problem 4-24
in making the computations.
4-26. Work out the equations which determine Q* and r* for the simple
model discussed in Sec. 4-2 when the marginal distribution of lead
time demand is gamma distributed with mean y, and standard
deviation a.
4-27. Solve the example presented in Sec. 4-4 when the marginal distribu¬
tion of lead time demand has a gamma distribution with the same
mean and standard deviation as the normal distribution in the
example.
4-28. Derive the exact cost equation and the equations which determine
Q* and r* under the assumption that the marginal distribution of
lead time demand is exponential. Assume that lead times are inde¬
pendent and that orders do not cross.
4-29. Derive the exact cost expression and the equations which determine
Q* and r* when the marginal distribution of lead time demand is a
gamma distribution with mean ju, and standard deviation <r. Assume
that lead times are independent and that orders do not cross.
4-30. Solve Problem 4-12 under the assumption that the marginal dis¬
tribution of lead time demand has an exponential distribution whose
mean is the same as that of the normal distribution referred to in
the problem.
4-31. Examine Molina’s tables and note that often by adding merely a
single unit of safety stock, the probability of incurring a backorder
or lost sale during the lead time can be reduced from a sizable value
to an exceptionally small value. Would one normally expect such
an increment in protection by the addition of a single unit in the
real world? Why or why not?
4-32. Is it possible when the demand variable is treated as continuous to
have the demand in any time period of length t described by a
gamma distribution with mean \ty instead of by a normal distribu-
tion, while still being able to describe the (Q, r) system as a Mark ov
process continuous in space and time? Why or why not?
4-33. A low demand, expensive spare part for an aircraft is stocked at a
large military depot. The average rate of demand is 3 per year.
The demands behave as if they were generated by a Poisson process.
Units are ordered one at a time as demanded. Each unit costs $2000
and the system uses a carrying charge of I = 0.20. Demands
occurring when the system is out of stock are backordered. There
is no fixed cost for a backorder, but the grounding of an aircraft for
 PROBLEMS 227

lack of this part costs 110,000 per week. The procurement lead
time can be assumed to be a constant equal to 6 months. Determine
the optimal value of the inventory position which is to be main¬
tained at a constant value through time.
4-34. Solve Problem 4-33 under the assumption that the procurement
lead time has a gamma distribution with mean equal to 6 months
and standard deviation equal to 1 month. Imagine that lead times
are independent random variables and that orders are received in
the same sequence in which they were placed.
4-35. Solve Problem 4-33 assuming that in addition to the backorder cost
of $10,000 per week, there is a fixed cost per backorder of $5000.
4-36. Solve Problem 4-34 assuming that in addition to the backorder cost
of $10,000 per week, there is a fixed cost per backorder of $5000.
4-37. Plot a curve showing how in Problem 4-33 the optimal inventory
position varies with the backorder cost f.
4-38. Solve Problem 4-12 under the assumption that there is no fixed cost
for incurring a backorder but that instead a backorder costs $20.00
per week out of stock.
4-39. Solve Problem 4-15 under the assumption that instead of specifying
the probability of being out of stock, there is a cost of $10,000 per
unit week of shortage.
4-40. A paint store orders a type of one pint cans of red paint by the
carton. There are 24 pints in a carton and after each twenty-four
demands, a carton is ordered. The order size is always one carton.
One carton costs $12.00, and the store uses an inventory carrying
charge of I = 0.20. Demands occurring when the store is out of
stock are lost, and the owner estimates the cost of a lost sale to be
$1.00. Three weeks are required from the time that an order is
placed until the shipment is received. What is the optimal reorder
point?
4-41. Solve the example given in Sec. 4-4 under the assumption that
“all units” quantity discounts are available of the following sort:
the unit cost is $50.00 if 0 < Q < 1200; the unit cost is $49.50 if
1200 < Q < 2500; and the unit cost is $49.00 if 2500 < Q < °o.
4-42. Solve the example given in Sec. 4-4 under the assumption that
incremental quantity discounts are available of the following sort:
for all units between 0 and 1200 the unit cost is $50.00. For addi¬
tional units up to 3000 units the unit cost is $45.00. For all units
above 3000, the unit cost is $40.00.
 228 LOT SIZE—REORDER POINT MODELS CHAP. A

4-43. Solve Problem 4-12 under the assumption that “all units” quantity
discounts of the following sort are available: the unit cost is 112.00
if 0 < Q < 50 and is $11.00 if 50 < Q < «.
4-44. Solve Problem 4-14 under the assumption that “all units” quantity
discounts of the following sort are available: the unit cost of a radio
is $40.00 if 0 < Q < 39, and is $37.00 if 40 < Q < 99 and is $33 00
if 100 < Q < oo.

4-45. Consider the problem of developing a (Q, r) system for an item when
the order quantity must be an integral multiple of a fixed package
quantity. Modify the simplified models developed in Secs. 4-2,
4-3 to take account of this, and obtain the equations which deter¬
mine Q* and r*.

4-46. The cosmetics department of a large department store has recently

introduced a (Q, r) system to control many items in the department.
A particular type of expensive perfume must be ordered in multiples
of one dozen bottles since the standard package contains 12 bottles.
The demand for this type of perfume averages 3 per week and is
Poisson distributed. A standard package costs the store $70.00.
An inventory carrying charge of 1 = 0.20 is used. The cost of placing
an order amounts to $0.50. This particular perfume is not easy to
obtain elsewhere, and hence demands occurring when the store is
out of stock are backordered. However, only the store’s most
wealthy customers purchase this perfume and hence the store con-
siders it bad policy to be out very frequently. The cost of a back¬
order is taken to be $100.00. Determine Q* and r* if the procure-
ment lead time is 5 weeks. How much would the average annual
savings be if no restrictions were placed on the size of the order
quantity?

4-47. For the model presented in Sec. 4-7 compute the probability that
precisely n orders are outstanding at any point in time. Also com¬
pute the expected amount on order and show that it is equal to the
expected lead time demand. Hint: Specification of the net inven¬
tory uniquely determines the number of orders outstanding.
4-48. For the model presented in Sec. 4-9 compute the probability that
precisely n orders are outstanding at any point in time. Also com-
pute the expected amount on order and show that it is equal to the
expected lead time demand. Hint: Same as for Problem 4-47.
4-49. Compute the mean of the distribution 6(m) defined by Eqs. (4-143),
(4-144) and show that it is equal to \/Q. ’
 PROBLEMS 229

4-50. Consider an item with the following characteristics

X = 700 units/yr. A = $15.00
C = $50.00 7T = $1.00
I = 0.20 7t = $15.00 per unit year

The lead time demand can be considered to be normally distributed

with a mean of 300 and a standard deviation of 50. The lead time
can be assumed to be constant. Try to determine Q* and r* using
Eqs. (4-89) and (4-90). Are any difficulties encountered?
4-51. What changes are needed in the equations used to determine Q* and
r* for the simple models developed in Secs. 4-2 and 4-3 if there is
also an annual carrying cost which is proportional to the maximum
on hand inventory which can exist. Answer the same question for
the model presented in Sec. 4-9.
4-52. In the text we have always assumed that the cost of a backorder
never had a more complicated form than ir + itt, where t is the time
for which the backorder exists. In general, of course, the cost of a
backorder can be an arbitrary function of the time for which the
backorder exists. When the cost of a backorder depends on t in a
more complicated way than x + itt, then the procedures used in the
text cannot be used to compute the expected annual cost of back¬
orders. One must use a procedure which explicitly takes account of
the length to time for which each backorder exists. To do this it is
convenient to imagine that each unit ordered is tagged to meet a
particular demand, so that no backorder costs are incurred for any
given unit ordered, unless the demand it is to satisfy occurs before
the order arrives. Imagine that the Q units ordered at a particular
point in time are to satisfy demands r + 1, . . . , r + Q occurring
after the order is placed. Show that when the lead time is a constant
r, then for the case where demand is Poisson distributed, and x(Z) is
the cost of a backorder which lasts for a length of time t, the expected
backorder costs associated with the order (i.e., per cycle) are

X) f Xir(T ~~ +3 - 1; K) d£
3-1 Jo

What then are the expected annual backorder costs? For the case
where ir(t) = w + ft show that the above formula leads to the same
results as those obtained in the text. Compute the average annual
backorder costs when

(a) ir(t) = 7To ~b TTit + 7T212 (b) 7f(t) == TTQGht, 6 > 0

230 LOT SIZE—REORDER POINT MODELS CHAP. 4

What sort of modifications are needed when the lead time is a

stochastic variable?
4-53. Solve Problem 4-52 under the assumption that the demand in any
time period is represented by a normal distribution.
4-54. Compute the variance of the distribution of the on hand inventory
for the model developed in Sec. 4-7.
4-65. Compute the variance of the distribution of the on hand inventory
for the model developed in Sec. 4-9.
4-56. Derive the expression for X corresponding to Eq. (4-61) under the
assumption that r is negative. Show that Eq. (4-61) is correct
provided that a(v)f ft(v) are given by Eqs. (4-62) and (4-63)
respectively.

4-57. Derive in detail Eq. (4-83).

4-58. Show that the X(Q, r) defined by Eq. (4-107) will be obtained
directly if the marginal distribution of lead time demand is used in
computing the state probabilities rather than the distribution of
lead time demand for a fixed r.

4-69. Attempt to compute the generating function for 0(m) defined by

Eqs. (4-143) and (4-144). How would one obtain the generating
function for Q(M) where M is the total number of orders placed for
all n items in one year. Attempt to obtain an explicit expression
for 0(M). Why is Q(M) approximately normally distributed when
n is large?

4-60. Discuss the use of floor space constraints in (Q, r) models with
stochastic demands when there are n items. Show that the same
sort of problems are encountered as those encountered in placing
a limitation on the number of procurements made.

4-61. Discuss the problems involved in applying a constraint on invest-

ment in inventory for (Q,r) models with stochastic demands when
there are n items. Note that in this case there is an upper limit to
the inventory that can be on hand. What is it? How relevant is
the upper limit to a constraint on the average investment in inven¬
tory? How difficult is it to compute the probability that the total
investment in inventory will go above a specified value?
4-62. Derive in detail Eq. (4-105).
4-63. Prove that nl/tn obtained from Eq. (4-124) is a legitimate density
function for the situation discussed there. Hint: What is the
allowable range of variation for £t?
 PROBLEMS 231

4-64. How can R* be determined when the average annual cost is given
by Eq. (4-138)?
4-66. Obtain the E(Q, r) and B(Q, r) terms for the model of Sec. 4-7 by
first computing the expected number of backorders or unit years of
shortage incurred per cycle and then convert to an annual basis.
Hint: Take a cycle to be the time between the arrival of two suc¬
cessive procurements. Note that it is necessary to take explicit
account of the fact that the length of the cycle is a random variable.
4-66. Note that when units are demanded one at a time, there are pre¬
cisely Q demands in every cycle. Imagine here that a cycle refers to
the time between the placing of two successive orders. Then the
inventory position is r + Q until the first demand occurs, r + Q — 1
until the second demand occurs, etc., and r + 1 until the Qth demand
occurs. Compute the expected costs (displaced by a lead time) of
carrying inventory and of backorders incurred between the oeur-
rence of 'the jth and the 0’ + l)st demands. Then the corresponding
costs per cycle are the sum of these costs from j = 0 to Q — 1. In
this way compute the costs per cycle of carrying inventory and of
backorders when a Poisson process generates demands and the lead
times are constant. Hint: The time between demands is not con¬
stant but is exponentially distributed.
4-67. For any real world system the ultimate use of any item demanded
may vary widely from one demand to the next. It is conceivable
that a different backorder cost might be associated with each end
use. This is especially true in military supply systems where the
item may be used in a variety of ways. Suppose that there are N
different end uses and for end use i, the cost of a backorder is + iut.
Let fi be the fraction of the demands that will be for end use i (or
the probability that any given demand will be for end use i). Show
how to account for this sort of behavior within the framework of
the models developed in this chapter. Do they really implicitly
include this case also?
4-68. Combine the models developed in Secs. 4-2 and 4-3 to obtain a model
for the case where a fraction / of the demands occurring when the
system is out of stock are backordered while the remainder are lost.
Assume that the cost of a lost sale is different from the cost of a
backorder.
4-69. Solve Problem 4-12 under the assumption that instead of specifying
the cost of a backorder, it is required that the average fraction of
the time that the system is out of stock not be greater than 0.01.
What is the imputed cost of a backorder?
232 LOT SIZE—REORDER POINT MODELS CHAP. 4

4-70. Solve Problem 4-14 under the assumption that instead of specifying
the cost of a lost sale, it is required that the average fraction of the
time that the system is out of stock not be greater than 0.05. What
is the imputed cost of a lost sale?
4-71. Solve Problem 4-22 under the assumption that instead of specifying
the cost of a lost sale, it is required that the average fraction of the
time that the system is out of stock not be greater than 0.02. What
is the imputed cost of a lost sale?
4-72. Solve Problem 4-33 under the assumption that instead of specifying
the cost of a backorder, it is required that the average fraction of
the time that the system is out of stock not be greater than 0.005.
What is the imputed cost of a backorder?
4-73. Consider a system for which the mean rate of demand is constant
over time, and the number of units demanded in any given time
interval is Poisson distributed. The procurement lead time is always
a constant r, and demands occurring when the system is out of stock
are backordered. Imagine that there is no fixed cost of incurring a
backorder, and if there are any backorders on the books at time t,
the backorder cost incurred between t and t + dt is tt0 dt Note that
the backorder cost is independent of the number of backorders out¬
standing. Furthermore, assume that annual carrying costs are pro¬
portional to the maximum value of the inventory position. The
cost of placing an order is A, and the unit cost of the item is C
independent of the size of an order. Develop the average annual
cost expression for a (Q,r) model for the above item, and obtain
formulas for computing Q* and r*.
4-74. In the lost sales case studied in Sec. 4-11, prove that the minimization
of the average annual cost with the cost of a lost sale defined so as
to include the lost profit will yield the same Q* and r* as the maxi¬
mization of the average annual profit.
4-75. In order to treat r and Q as continuous and to differentiate 3C with
respect to these variables, what restrictions must be placed on h(x)
in the model of Sec. 4-2 if the derivatives are to exist for all Q > 0,
r > 0?
PERIODIC REVIEW MODELS

WITH STOCHASTIC DEMANDS

A decision has to be made once each orbit. Each go minutes the

astronaut — and Mercury control — have to decide over the Pacific

whether he is going to be brought down in the Atlantic. This is

a recurring, tough decision they have to make.’

Newspaper Report
5-1 Introduction

Although there is no difference between (Q, r) and periodic review models

for deterministic systems, the nature of the two types of models becomes
somewhat different for stochastic demands. This chapter will be devoted
to studying different types of periodic review models. In the real world,
one can find both transactions reporting and periodic review systems in
operation. However, at the present time, the number of systems using
periodic review appears to be much greater than the number using trans¬
actions reporting. The reason for this, of course, is that transactions re¬
porting systems can be costly and difficult to operate in practice. It might
be pointed out, though, that with the increased use of high speed data
processing equipment, the trend is more and more in the direction of
changing over to transactions reporting. This is especially true in military
supply systems. There are benefits to be gained from using transactions
reporting if it is not too costly because, among other things, it is possible to
cut down on the average investment in inventory by doing so. This will
be examined in more detail later in the chapter.
What is involved in making a “review” when a periodic review procedure
is used can vary widely from one system to another. In some situations,
transactions may have been recorded as they occurred but were not trans¬
mitted to the decision maker in a form that could be used. In this case it
may only be necessary to aggregate, either by hand or on a digital com¬
puter, all the transactions over the past period, determine the relevant
variables of interest, and send them to the decision maker. In other cases
it may be necessary to make an actual physical count of the inventory
before any decisions are made. This will often be done even when there is
some form of transactions recording, because the recorded data may not
be accurate enough for use. A typical example of this latter situation is a
department store where the buyer will make a stock count, even though
information from sales slips or tickets is available, simply because the
information taken from sales slips or tickets will, for a variety of reasons,
often be considerably in error. Of course, it should be noted that even
when transactions reporting systems are used, actual inventory counts will
235
 236 PERIODIC REVIEW MODELS WITH STOCHASTIC DEMANDS CHAP. 5

be needed occasionally to eliminate errors arising from breakage, spoilage,

pilferage, mistakes, etc.
For transactions reporting systems in which the mean rate of demand is
constant, units are demanded one at a time, and the process generating
lead times does not change with time, it is clear that if the operating doc¬
trine is also to remain time invariant, then a Qr doctrine is the only one
that can be thought of which makes full use of the ability to make a decision
after every transaction, and hence is an optimal operating doctrine. This
can be proved rigorously using the methods introduced in Chapter 8. In
the case of periodic review systems, however, things are somewhat more-
complicated. Several possible operating policies suggest themselves.
One operating doctrine for periodic review systems which is frequently
used in practice requires that an order be placed at each review time if
there have been any demands at all in the past period (a periodic review
system will be said to have gone through one period’s operation in the time
between two successive reviews). A sufficient quantity is ordered to bring
the inventory position or the amount on hand plus on order up to a level R.
Note that with this system, the quantity ordered can vary from one review
period to the next. This operating doctrine will be called an “order up
to R” doctrine.
. -^n alternative operating doctrine is to make a procurement at a review
time only if the inventory position or the amount on hand plus on order is
less than or equal to r. In such a case a sufficient quantity is ordered to
bring the appropriate inventory level up to R(R > r). This operating
doctrine will be referred to as an Rr rule. An order up to R rule is a special
case of an Rr rule in which r = R — 1 when the inventory levels are treated
as discrete variables, and r = R when they are treated as continuous
variables.
An intermediate type of operating doctrine is that in which, as above, a
procurement is made at a review period only if the inventory position or
the amount on hand plus on order at the review time is less than or equal
to r. However, the quantity ordered is chosen to be an integral multiple
of some fundamental quantity Q, i.e., the quantity ordered is nQ where
n — 1, 2, 3, . . . . The value of n is chosen to be the largest integer such
that alter the order is placed the appropriate inventory level is less than or
equal to R = r + Q. Such an operating doctrine will be referred to simply
as an nQ doctrine. It will be observed that when the inventory levels are
treated as discrete variables, then an order up to R rule is a special case of
the nQ rule for which Q = 1 and R = r + 1. When the inventory levels
are treated as continuous variables, it is still true that the order up to R
rule is a special case of the nQ doctrine in the limit as Q —> 0.
The three operating doctrines described above are essentially the only
ones ever used in practice for periodic review systems. This chapter will
SEC 5-2 SIMPLE, APPROXIMATE (R, T> MODELS 237

develop and compare a set of models which use these three operating doc¬
trines in the case where the stochastic processes generating demands and
lead times are time invariant. The purpose of these models will be not
only to determine the appropriate values of r, R for a given T, but also to
determine the optimal value of T, the time between reviews, in cases where
this time is also a variable. Models which use an order up to R policy will
be referred to as (R, T) models, those which use an Rr policy as (R, r, T)
models, and those which use an nQ policy as (nQ, r, T) models. Of course,
any model which can be used to determine the optimal value of T as well
as R, r can also be used to determine the optimal values of R, r for
a given T.
In Chapter 4, the annual cost of operating the transactions reporting
system was not included in the average annual cost. The reason was that
it was assumed that this cost was independent of Q and r and hence did not
need to be included.
In this chapter, however, in order to determine the optimal value of T, it
is necessary to introduce the cost of making a review. The cost of making
a review will be assumed to be independent of the model variables. The
review cost will not include the cost of placing an order since, in general, an
order may not be placed at each review.
As in Chapter 4 we shall begin with two simple approximate models
which are also the most useful ones for practical applications.

5-2 Simple, Approximate (R, T) Models

In the real world the most widely used operating doctrine for periodic
review systems is the order up to R doctrine. We shall here present the
analogues of the backorders and lost sales models of Secs. 4-2 and 4-3
respectively, for periodic review systems using an order up to R policy.
Again all variables will be treated as continuous.
Consider first the backorders case. The time between reviews will be
denoted by T, and at each review time a sufficient quantity is ordered to
bring the inventory position of the system up to a level R. The problem
is to determine the optimal values of R and T. At the outset, the following
assumptions will be made:
1. The cost J of making a review is independent of the variables R and T.
2. The unit cost C of the item is constant independent of the quantity
ordered.
3. Backorders are incurred only in very small quantities. This will be
taken to imply that when an order arrives, it is almost always sufficient to
meet any outstanding backorders.
 238 PERIODIC REVIEW MODELS WITH STOCHASTIC DEMANDS CHAP. 5

4. The cost of each backorder is t, and the cost is independent of the

length of time for which the backorder exists.
5. When the procurement lead time is a random variable, it is assumed
that orders are received in the same sequence in which they were placed,
and furthermore, the lead times for different orders can be treated as
independent random variables. It might be noted that for (Q, r) models,
the dual assumptions that orders were received in the sequence placed and
that lead times for different orders were independent random variables
could not both hold rigorously, since there existed a positive probability
that two successive orders could be separated by an arbitrarily short time
interval. Here, however, orders can never be more closely spaced than by
an interval of length T, and hence if T is great enough, it is possible, pro¬
vided that there is a sufficiently small range of variation in the lead time,
to have both assumptions hold simultaneously.
As in Chapter 4, let A be the cost of placing an order and I be the inven¬
tory carrying charge. The density function for the demand x in a time
interval of length t will be written f(x; t), and X will denote the average
demand rate. The variable costs which must be included in average annual
costare: (a) review, (b) ordering, (c) carrying and (d) backorder costs.
Since the time between reviews is T, the average annual review cost is
J/T. Furthermore, since an order will be placed at each review (when
demand is treated as continuous, the probability of no demands occurring
in the period is zero), the average annual ordering cost is A/T. If we write
L = A + J, then the average annual costs of review and ordering are L/T.
In this case the ordering and review costs can be combined since an-.order
is placed at each review time.
The average annual cost of holding inventory will be found by com¬
puting the expected holding cost per period and then multiplying by 1/T to
obtain the average annual cost. It is convenient to use as a period the
time between the arrival of two successive orders rather than between the
placement of two successive orders in making this computation. Inasmuch
as the inventory position of the system is R immediately after reviewing
and placing an order, and since everything on order will have arrived in a
procurement lead time and nothing which is not on order can arrive in this
time, the expected net inventory immediately after the arrival of a procure¬
ment must be R — ix, where jjl is the expected lead time demand. Because
the mean rate of demand remains constant over time, the expected net
inventory must decrease linearly with time and have the value R — — XT
just prior to the arrival of the next order, since the expected demand per
period must be the expected amount ordered, i.e., XT, so that if the expected
net inventory immediately after the arrival of a procurement is R — fx, it
is therefore R — fi — XT just prior to the arrival of a procurement. Now
SEC 5-2 SIMPLE, APPROXIMATE <R, T> MODELS 239

because of assumption 3 it must be true that the integral over time of the
net inventory must very closely approximate the integral over time of the
on hand inventory. Hence the expected unit years of storage incurred per
period is to a good approximation

T [| (R - M) + | (R - m - XT)] = T [i2 - „ - y]

so that the average annual cost of carrying inventory is

IC [i? - M - y] (5-i)

It will be shown later when the precise expressions are derived that the
simple argument given above is correct in the case where the expected unit
years of backorders incurred per year is negligible with respect to the
expected unit years of storage. Note that the above argument applies even
if the procurement lead time is a random variable.
In order to compute the average annual cost of backorders, it is necessary
to compute the average number of backorders incurred per year. This will
be done by computing the expected number of backorders incurred per
period and multiplying by 1 /T. Consider first the case where the procure¬
ment lead time is a constant r. Then an order placed at time t will arrive
in the system at time t + r, and the next procurement will arrive in the
system at time t + r + T. After the order is placed at time £, the inventory
position of the system is R. We wish to compute the expected number of
backorders occurring between t + r and t + r + T. A backorder will
occur in this period under assumption 3 if and only if the demand in the
time period r + T exceeds 22. Assumption 3 assures us that after the
arrival of the order placed at time t there will be no backorders on the books,
and hence they must all be incurred between times t + r and t + r + T.
Hence the expected number of backorders incurred per period is(

(x - R)f(x; r+T)dx (5-2)

Jr
Note that it is no longer the lead time demand but instead the lead time
plus one period's demand which is relevant.,! Intuitively, the reason for this
is that once the order is placed at time t, another order cannot be placed
until time t + T regardless of what happens, i.e., protection is needed for
a lead time plus T. We found above that the expected value of the net
inventory at the time of arrival of a procurement is R — /z — XT. This is
by definition the safety stock, and as expected, it does depend on the ex¬
pected demand in time r + T.
Suppose now that the procurement lead time r is a random variable
240 PERIODIC REVIEW MODELS WITH STOCHASTIC DEMANDS CHAP. 5

with density g(r)1 and let rmm, rma3: be the lower and upper limits respec¬
tively to the possible range of lead time values. Then if n, r2 are the lead
times for the orders placed at times t and i + T respectively, the expected
number of backorders incurred per period must be
f Tnias CTnrni

/
/* <*>

/ (x - R)f(x; T2 + T)g(r2)g(n) dx dr2 dn

(x — R)h(x; T) dx (5-3)

where

h(x] T) = f(x;r2 + T)g{r2) dr2

This follows since

g(n) dn = 1

In order for (5-3) to hold rigorously, it is necessary that there be no overlap

of the lead time distributions, i.e.,

However, the result will be approximately correct if the overlap is small.

Note that h(x; T) is not the marginal distribution of lead time demand.
If we define h(x; T) to be f(x; r + T) when the procurement lead time
is a constant (and to be (5-4) when the procurement lead time is a random
variable, then in all cases, the expected number of backorders incurred per
period is given by (5-3). The average number of backorders incurred per
year is

E(R, (x- R)h(x; T) dx (5-5)

and the average annual cost of backorders is tE(R7 T).

All the terms needed in the cost expression have now been evaluated.
The average annual variable cost is

^ — j, + IC^JR -- fx — —-J + 7rE(R, T) (5-6)

For a given T, the value of R which minimizes 3C must satisfy

|| = 0 = IC-Z&(R;T)
where

T) = h (x; T) dx (5-8)
SEC. 5-2 SIMPLE, APPROXIMATE (R, T) MODELS 241

is the complementary cumulative of h(x; T)\ Thus the optimal value

of R, is a solution to

fi(R-, T) = — (5-9)
7T

For a given T, (5-9) will normally yield a unique R although in certain

cases (to be examined in Problem 5-4) R may not be unique. There will
be no solution if ICT/ir > 1. This cannot happen, however, if the assump¬
tions originally made are to be valid, since it would imply that backorders
would be incurred very frequently.
To determine the optimal T value, one could attempt to solve
dX/dT = 0 simultaneously with (5-9), say by Newton’s method, thus
obtaining R* and T* at the same time. The method of steepest descents
could also be used. In this case, however, it is also quite simple merely
to tabulate X as a function of T, using the R* for the given T in computing
X, plot the results and in this way determine T*. The development of the
details of using Newton’s method or the method of steepest descents is
left to Problems 5-5 and 5-7.
The equations for the lost sales case differ little from those for the back¬
orders case. The review-ordering and stockout costs are precisely the same.
A slight correction is made in the inventory carrying cost. Since for the
lost sales case the safety stock is the expected value of the on hand inven¬
tory at the time of arrival of a procurement, we see using the same argu¬
ments as in Sec. 4-3 that the safety stock is

R — n — XT + J (x — R)h(x; T) dx (5-10)

For periodic review systems this expression is approximate even when

there is never more than a single order outstanding and holds only when
lost sales are incurred in very small quantities. Equation (5-10) ignores
the effects of lost sales which can occur between the time an order is placed
and the time it arrives. It considers only lost sales occurring between the
time the order arrives and the time the next order arrives. So long as the
system is out of stock on the average only a very small fraction of the time,
the correction to (5-10) itself will be small, and hence (5-10) is a sufficiently
good approximation. As long as the average fraction of time out of stock
is very small, the model also applies to cases where more than a single order
can be outstanding at any point in time. The average annual cost of
holding inventory in the lost sales case is then

ic [s - m - y+ &] (5-n)
 242 PERIODIC REVIEW MODELS WITH STOCHASTIC DEMANDS CHAP. 5

The average annual variable cost for the lost sales case is thus

3C = ^ + /C (x — R)h(x; T) dx (5-12)

and the equivalent of (5-9) becomes

ICT
S(R; T) (5-13)
7T + ICT

This completes the development of the simple order up to R models. With

the approximations made, they can be applied when lead times are random
variables and when more than a single order is outstanding.

EXAMPLE A large California warehouse follows a policy of reviewing all

items quarterly. It uses an order up to R policy for every item. Consider
one item it carries—say a particular type of tractor tire. The mean demand
rate has been constant over time at the value of 600 per year. The ware¬
house orders tires directly from the manufacturer and the lead time r is
nearly constant, and has the value 6 months. The demand in the time
r + T can be represented quite well by a normal distribution with mean
600 (r + T) and variance 900 (r + T). The variance is not equal to the
mean (as would be true if a Poisson process were generating demands),
since the number of units demanded per demand is also a random variable,
and this increases the variance. The cost to the warehouse of each tire is
$15.00; the warehouse uses an inventory carrying charge of / = 0.20. All
demands occurring when the warehouse is out of stock are backordered and
the cost of a backorder is estimated to be $25.00. It is desired to determine
the optimal value of R.
Note first of all that when T is specified, ig* can be found without a
knowledge of the review or ordering costd From the data given above
T + r = | years and the expected demand in time T + r is 600(f) = 450,
and the variance of the demand in this time is 900(f) = 675, or the stan¬
dard deviation is 25.981. Thus from (5-9), R* is the solution to

(R - 450\ ICT 0.20(15)

0.030
V 25.981 ) x 25(4)
From the normal tables it follows that

ig* - 450
1.881
25.981
and
ig* = 450 + 48.8 = 498.8 « 499

From (5-5) and A4-25 the average number of backorders incurred per year
is
SEC 5-2 SIMPLE, APPROXIMATE <R, 7} MODELS 243

4[25.98*(1.881) + (450 - 499)^(1.881)] = 4[25.98(0.0679) - 49(0.030)]

= 1.175

Thus the average fraction of time out of stock is --g-inr = 0.00196, which is
small as the use of the model requires. Note that since r = 6 months
and T = 3 months, there will always be precisely two orders outstanding.
Suppose now that the combined cost of ordering and reviewing is $25.00.
Let us then determine the optimal value for T, the time between reviews.
This will be done simply by tabulating 5C as a function of T, where for
each T the optimal R for that T is used. The cost expression for the case
under consideration is

25
X 3 [R - 300 300 T]
T'
~R - 600(0.5 4- T)'
+ 25T -f\/9oo(°- 5 + T) <j>
l - V900(0.5 + T) .
R - 600(0.5 + T)'
+ [600(0.5 + T) #i*rL V900(0.5 + T) .

The results are presented in Table 5-1, and X is plotted as a function of T

in Fig. 5-1. It is seen that the optimal value of T is about 1.9 months, and
the use of this T value would result in an average annual savings of about
$24.00 as compared to the T value of 3 months now in use. The value of T*
could be computed more accurately. The value given is within 0.05 months
of the exact value. For real world applications this accuracy is quite
sufficient, especially since it would not be possible to save more than a
fraction of a dollar per year by computing T* more accurately. The values

TABLE 5-1
Data for Example

Average
Annual Average Average
Safety Review and Annual Annual Average
T R*(T) Stock Ordering Carrying Backorder Annual
(months) (units) (units) Cost Cost Cost Cost X

0.5 382 57 $600 $208. $23 $831

1.0 403 53 300 234 23 557
1.5 426 51 200 265 27 492
1.7 436 51 177 281 25 483
1.8 440 50 167 285 27 479
1.9 445 50 158 292 27 477
2.0 449 49 150 297 31 478
3.0 499 49 100 372 29 501
 244 PERIODIC REVIEW MODELS WITH STOCHASTIC DEMANDS CHAP. 5

in Table 5-1 have been rounded off to the nearest dollar value. The back¬
order costs jump around somewhat because R is rounded to the nearest
integer.
The behavior of the safety stock in Table 5-1 deserves some comment.
At first it seems rather surprising that it should decrease with increasing T.
The reader can easily check that it continues to decrease even for higher T.
Intuitively, the reason for this is that the more frequently that the net
inventory gets down in the neighborhood of the safety stock, the greater

will be the required safety stock to prevent unduly large backorder costs.
However, m the periodic review model being studied, the net inventory
^ neighborhood of the safety stock once each period. Hence, the
,, TT T’ thf, fger the safety stock required. Another peculiar feature is
that the model has the safety stock approach infinity as T 0 We know
intuitively that this is not correct. The problem lies in the formulation of
the model. We have assumed that there are no backorders on the books at
the time an order arrives. This assumption is not valid when T is small
lne exact formulation given later will rectify this problem.
SEC. 5-3 EXACT FORMULATION OF THE <nQ, r, T) MODEL 245

An interesting comparison of the (R, T) model being studied and the

simple (Q, r) model of Sec. 4-2 can be made using the present example if
one assumes that the $25.00 combined ordering and review cost is entirely
an ordering cost, i.e., reviews are free. If the values of A, C, I, x are used in
the (Q, r) model of Sec. 4-2 to obtain an optimal Q and r, one finds on carry¬
ing out the iteration process that Qw = 100, Q* = 108, and r* — 343. The
average time between procurements for the (Q, r) model is T = Q/\ =
= 0.18 yr. = 2.16 mo.
The above results show that for the periodic review model T* is slightly
less than the average time between procurements for the (Q, r) model while
the safety stock of 50 is higher for the periodic review model than the value
43 for the (Q, r) model. This sort of behavior will always occur when
comparing (Q, r) models and periodic review models as was done above.
The reason for this behavior lies in the fact that in the periodic review
system, the safety stock must offer protection for a lead time plus one
period, while in the (Q, r) model protection is needed only for the lead time.
This means that the safety stock will be greater for the (R, T) model. Also,
it simultaneously pays to order somewhat more frequently, thus reducing
the average on hand inventory to partially offset the extra carrying
charges arising from the increased safety stock needed.

5-3 The Exact Formulation of the (nQ, r, T) Model for the Backorders
Case with Poisson Demands and Constant Lead Times

At this point, instead of obtaining the exact form of the {R, T) models in
certain cases, we shall instead move on to obtain the exact equations for the
{nQ, r, T) model in the backorders case where the number of units de¬
manded in any time period can be described by a Poisson distribution and
the procurement lead time is always a constant r. Recall that an (R, T)
model is a special case of the (nQ, r, T) model with Q = 1 and R = r + 1.
Thus once having obtained the exact equations for the (nQ, r, T) model
we can immediately obtain the exact equations for an {R, T) model under
the same assumptions which apply in deriving the (nQ, r, T) model.
We shall in the derivations treat the inventory levels as discrete variables
as well as the demand variable. Recall that an nQ operating doctrine is one
such that, at a review time, an order is placed if and only if the inventory
position y of the system at the review time is less than or equal to r. If
y < r, then a quantity nQ is ordered, n = 1, 2, 3,. . ., where n is chosen so
that r<y + nQ<r + Q. Thus, immediately after a review, the inven¬
tory position of the system will be in one of the Q states r + 1, r + 2,
. . •, t + Q.
To begin, we shall compute the steady state probability p(r + j) that the
 246 PERIODIC REVIEW MODELS WITH STOCHASTIC DEMANDS CHAP. 5

inventory position of the system is r + j immediately.after a review. To

do this, we note that if we concern ourselves only with the state of the
system immediately after a review, the process generating transitions be¬
tween states can be considered to be a Markov process discrete in space
and time, provided that the demands in different periods are independent
random variables. For a fixed T, these variables can be independent even
if a Poisson process does not generate demands. However, if the variables
representing demands in different periods are to be independent for any
T > 0, this will be true only if a Poisson process generates demands.
It need not be true, though, that units are demanded one at a time. The
quantity demanded per demand can be a random variable, and demands in
different periods will be independent for any T > 0, so long as a Poisson
process generates demands and the order size does not depend on past
history. Thus, for example, the stuttering Poisson distribution has the
property that demands in different periods are independent for any T > 0.
Let us now compute the transition probabilities aih where a{j is the
probability that if the inventory position is r + i immediately after agiven
review it will be r +j immediately after the next review. Denote by
p(x; T) the probability that x units are demanded in the time interval
between reviews.
We shall first compute Oy when j < i. Then the inventory position can
be r + j immediately after a review when it was r -f- i immediately after the
previous review if i - j, i-j + Q, i~j + 2Q, etc. units have been
demanded m the period between reviews, since if d is the demand then d
must satisfy the equation (see Fig. 5-2)

r + i - d + nQ = r + j, n = 0,1, 2,. . .
or

d = i-j + nQ, n = 0,1, 2,. . . (5-14)

The probability that the demand is i — j + nQ is p(i — j + nQ; T).

Since the probabilities for different n values are mutually exclusive, it
follows that

aa = X) ~ 3+ nQ;T), j <i (5-15)

n=0

However, when j > i, the system cannot be in state j unless the demand
has been at least i — j -j- Q. Thus

aa ~ Yj PO' - j + nQ; T), j > i (5-16)

n= l

From Sec. 3-10 we know that the p(r + j) must satisfy the equations
SEC 5-3 EXACT FORMULATION OF THE (nQ.r.T) MODEL 247

p(r + j) = E P(r + i)aih j = 1, . . ., Q (5-17)

1=1

Now note that

EE^-i + ng^+f^fi-i + nQ;!)

4=1n = 1 i=jn = 0

= E HE v(nQ+Q -v;T)+ £ P(nQ + v; T) 1

n=0 L® = 1 u=0 J

oo

= E P(fc; 20 = 1, 3 = 1, • • •, Q (5-18)
k=0

where, when j = 1, any summation whose upper limit is j — 1 is taken to

be 0. Then the sum of the a*y over i is equal to unity for all j. We see
immediately from (5-17) that p(r + j) = constant satisfies (5-17). Further¬
more, since Y,?=i p(r + j) = 1, it follows that

p(r +j) = jQ ^ 1 ’ ^ (5-19)

10 otherwise
 248 PERIODIC REVIEW MODELS WITH STOCHASTIC DEMANDS
CHAP. 5

Thus in the periodic review ease the distribution of the inventory position
immediately after a review is uniform. Recall that for (Q, r) models the
distribution was uniform at each instant of time. Note that (5-19) hoi(js
for any distribution ot demand as long as the demands in different periods
are independent. Also observe that, each <Ui > 0, so that from the results
of Sec. 3-10 we know that (5-19) is the unique solution to (5-17)
We are now in a position to develop the average annual erst expression “
At this point we shall restrict our attention to the ease when- units are
demanded one at a time, and a Poisson process generates the demands
[In principle, we could consider a case where the order size is a random 7
variable, say the stuttering Poisson distribution, but t he equations become '
exceptionally complex and difficult to work with) We shall here also restriefr
our attention to the case where the unit cost of the item is a constant ^
independent of the quantity ordered. The cost, of a backorder will I*
assumed to have the form tt + n when* t is t he length of time for which the
backorder exists. The costs of review and placing an order will be denoted
by J A respectively. As usual, the inventory carrying charge will he de¬
noted by I. The average annual cost, will be computed by determining the
expected cost per period and multiplying by 1 T to obtain the average
annual cost. Problem 5-12 asks the reader to go through the same sort, of
analysis used m bee. 4-0 to show that this is correct.
Since there are on the average l/T reviews per year, the average annual
cost of reviews is J/T. The average annual ordering cost is not A/T si me
an order need not be placed at each review. Instead, the
Sll hi Sr? 7 ™Ap"/T> whcrti P" * tin probability that an order
will be placed at any given review time. Let us now compute [f the
inventory position of the system is r+j immediately after a review the
probability that it will be less than or equal to r at the time of the next
review is the probability that; or more units are demanded in the time T

tfiTESEVf» *****
is r + j immediately after a review is 1 /Q, Tims

Vor itl-(rAT)
and from A3-6
\T
Vor
Q [1 - -P(Q;XT)3 + P(Q + l; XT’) (5-20)

Note that Vor is a function of T and Q l,ut is independent, of r

badcorders °f dctermininK average annual cost of

hp ZTji 7 avera«® number of backorders incurred per year will

periodPZposyeTat1afevfet:tlkjs
it,view taKos pH<-e
place 7 timeTr'^
at f'"1 ha<’k<,nicr‘S
t. The next P<ir
review takes
SEC 5-3 EXACT FORMULATION OF THE (nQt r, T) MODEL 249

plaee at time t + T. We shall compute the expected number of backorders

incurred between time t + r and t + r + T. The reason for doing this is
that everything on order immediately after the review at time t will arrive
in the system by t + r but nothing not on order can arrive before time
t + r+T^ The random variable representing the number of backorders
incurred between t + r and t + r + T can then be thought of as the differ¬
ence between two random variables, the first giving the number of backr
orders on the books at time t + r + T and the second being the number of1
backorders on the books at time t + r. If the inventory position of the.'
system is r + j immediately after the review at time t, the expected number
of backorders on the books at time t + r + T is

X) (x - r~ 3)p[x; X(r + T)]

x = r+j

and at time t + r

X (x-r- j)p(x; Xr)

x = r+j

Thus, if the inventory position is r + j immediately after the review at

time t, the expected number of backorders incurred between time t + r
and t + r + T is
ao

X (x-r-j) {p[x; X(r + D] -p(x; Xr)} *

x=rJrj

On averaging over the states j and dividing by T, we see that the average
number of backorders incurred per year is

E(Q, r} T) = -jyp ^2 ^2 (x — r — j) {p[x; X(r + T)] — p(x; Xr)}

—-V1 j = l x = r+j

= or E Y^(x~ u){plx; X(r + T)] -p(x; Xr)}

u = r+l x=u

= OT E (X(T + T)Piu - !; Hr + T)] (5-21)

uP[u; X(r + T)] —XtP(u - 1; Xr) + uP(u; Xr)}

Then on using A3-6 and A3-8 we can write

E(Q, r, T) = ^ [A(r, T) — A(r + Q, T)] (5-22)

 250 PERIODIC REVIEW MODELS WITH STOCHASTIC DEMANDS CHAP. 5

where

A(v, T) = ^ t+T) -00, r)3 (5-23)

and

t) = ~P(v - 1; Xt) - (Xi)vP(v; Xt) + °&-±XL p(„ + 1; Xt) (5-24)

The same sort of procedure is used to compute the average number of

unit years of shortage incurred per year. The expected number of unit years
of shortage incurred per period is the expected value of the integral of the
backorders from time i + r to i + r + !T with t being defined as above.
However, the expected value of the integral is the integral of the expected
number of backorders (since the expected value of a sum of random vari¬
ables is the sum of the expected values). For any time t -f~ £ between
t + r and t + r+ T, the expected number of backorders on the books when
the inventory position of the system was r + j immediately after the
review at time t is

T (x - r - j)p(x; X£)
x = r+j

and the expected unit years of shortage incurred from t + r to t + r + T is

r-f-T oo

T (x-r- j)p(x; X|)

x=r +3
Averaging over the states j and dividing by T, we find that the average
number of unit years of shortage incurred per year is

1 H~Q /*r+ T 00
B(Q, r,T)=—*r / y) (x - u)p(x; Xg) d£
^ u = r-\-l Jr x=u

or from A3-10

l HiQ rr+T
B(Q, r,T) = ~ 22 / - 1; X£) -uP(u] X?)] d| (5-25)
V u-r-fl Jr

However,

Jrr+T [XgP{u - 1; Xg) —uP(u; X?)] d£ = |\ {(T + T)*P[tt - 1; X(r + T)]

- t*P(u - 1J Xt)} + ^ + 1} {P[u + 1; X(t + T)] -P(u + 1; Xt)}

ui(T + TJPlw, X(t + T)] —tP(u; Xt)} (5-26)

 SEC. 5-3 EXACT FORMULATION OF THE <nQ, r, T) MODEL 251

By continued application of the properties in Appendix 3, B(Q, r, T) can

be written

B(Q, r, T) = i [T(r, T) — T(r + Q, T)] (5-27)

where
T(v, T) = H(», T + r) - B(®, r) (5-28)
and

s<». ‘1 - P(»; X‘) + H P(» - li xi)

+ + l; Xi) - * + + 2> P(» + 2; XI) (J-29)

The detailed derivation of (5-26) through (5-29) is left for Problem 5-13.
The average annual cost of backorders is then irE(Q, r, T) + fB(Q, r,T).-
It remains to evaluate the average unit years of storage incurred per
year. Again the procedure used above will be applied. The expected unit
years of storage incurred between t + t and t + r + T will be multiplied by
1 IT to obtain the desired answer. If the inventory position of the system
was r + j immediately after the review at time t, the expected value of the
net inventory at any time | between t + r and t + r + T is
oo

J2 (r + j ~ x)p(x; X£) = r + j - X£
x=0

The expected value of the on hand inventory at any time, is the expected
net inventory plus the expected number of backorders. Thus at time £, the
expected on hand inventory is
oo

r+j-\£+ ^2 (x r j)p(x; X£)

x**r+j

Integrating from r to T + r, averaging over the states j, and dividing by

T, we find that the average number of unit years of storage incurred per
year is
I « fr+T
D(Q, r,T) = Jt (r+j ~ X$) df + B(Q, r, T) (5-30)

0 1 XT
= ^ + 2 + r-r> (5-31)

where n is the expected lead time demand. Of course, (5-31) is also the
expected value of the on hand inventory at any point in time. The pro¬
cedure used to compute E, B, and D here differs from that used to compute
the corresponding quantities for the (Q, r) model in Chapter 4, because here
 252 PERIODIC REVIEW MODELS WITH STOCHASTIC DEMANDS CHAP. 5

we have available the probability distribution for the inventory position

only at points immediately following a review, while for the corresponding
(Q, r) model, the distribution was known at each point in time.
It is possible to compute the average unit years of storage incurred per
year in another way. Since the expected on hand inventory at any point
in time is equal to the expected inventory position minus the expected
quantity on order plus the expected number of backorders, it is dear that
B{Q, ri T) is equal to l/T times the integral over one period of the expected
inventory position as a function of time minus 1/7' times the integral over
one period of the expected quantity on order as a function of time plus
B(Q, r> T)- To compute the integral of the expected inventory position
over a period it is convenient to use the time between two reviews to define
the period. If the inventory position is r + j immediately after the review
at time t, the expected value of the inventory position at time t +
0 < ? < T is r + j — . Integrating this from 0 to T and averaging over
j, we have

^j(Ww*-r[S + i + ,-*] (M2)

Now l/T times the integral over one period of the expected quantity on
order as a function of time is simply the. expected amount, on order at any
point in time. Thus by comparison with (5-31), we have proved a,gain in
this case that the expected amount on order at any point in time is equal
to the expected lead time demand.
All the terms needed in the cost expression have now been evaluated.
The average annual cost is

x = ¥ + 1 Vm + IC [:2 + \ ~ 0 - y] + r, T)

+ (* + IC)B(Q, r, T) (5-33)

Just as with the exact form of the (Q, r) model of Chapter -1, it is difficult
to determine Q*, r*, and T* for the (nQ, r, T) model manually. Again it is
the terms in E(Q, r, T) and Ii(Q, r, T) involving r + Q which cause the
difficulty. Here, however, unlike the situation encountered for the (Q, r)
model, it is not in general a good approximation to drop these terms. It is
quite possible that the optimal Q will be small even though the average
quantity ordered is not; in such cases, the r + Q terms will not be negli¬
gible. To determine Q* r* and T*, a digital computer will usually be
required to perform the search needed to determine, the values of Q, r, and
T which minimize 3C.
SEC 5-3 EXACT FORMULATION OF THE <nQ, r, T> MODEL 253

A code to compute Q*(T) and r*(T) for the (nQ, r, T) model discussed
above has been written by P. Teicholz and B. Lundh of Stanford University
for the Burroughs 220 computer. The search procedure used in this code
determines a relative minimum, but does not necessarily determine the
absolute minimum if there exist a number of relative minima. The results
presented in the following example were computed using this code.

EXAMPLE Let us compute Q*, r* and T* when an nQ operating doctrine is

used to control an item for which the number of units demanded in any
time period is Poisson distributed with the mean rate of demand being 100
units per year. The lead time is a constant and has the value 0.25 yr. The
values of the other relevant parameters are: J = $2.00, A = $4.00, C —
$10.00,1 = 0.20, ir = $20.00, and f = $1.00. Note that the ordering cost
is twice the review cost. When the ordering cost is small with respect to
the review cost, then it should be optimal to order at essentially every
review, so that we would expect in such a case that Q* = 1 and the {nQ,
r, T) model is equivalent to an {R, T) model. It is only when the ordering
cost is high with respect to the review cost that one would expect Q* to be
greater than unity. This is the case to be studied here.
By use of the computer program referred to above, Q*{T) and r*(T) were
computed for various values of T. The results are presented in Fig. 5-3.
The results are somewhat surprising in that Q*(T) = 1 until T becomes
less than 0.15 years. Indeed Q* = 1, r* — 63, and T* 0.25 years. Thus
in a situation where one might expect Q* > 1, it turned out that Q* = 1.
Another interesting feature of the computation is that when T < 0.15,
Q*(T) changes very rapidly from unity to Qw = 20. The computer program
referred to above also provided for the possibility of setting Q = 1 and
optimizing over r for a given T, Thus the optimal order up to R policy and
the corresponding cost could be found. As has already been noted, the opti¬
mal nQ doctrine for a given T is an order up to R doctrine down to T =
0.15. The divergence between the two doctrines at T = 0.10 is also illus¬
trated in Fig. 5-3. The cost differences grow rapidly as T is reduced below
0.10.

The above example seems to suggest that an order.up toJg policy will
be
have guessedjndgmalLy. It is only when the probability of placing an order
at any review is not too high that an {nQ, r, T) model can yield significantly
lower costs than an {R, T) model. There do arise cases where the review
interval will arbitrarily be set to a very small value so that the periodic
review system becomes almost a transactions reporting system (viz., a
computer will be used to process all transactions, but instead of doing it on
254 PERIODIC REVIEW MODELS WITH STOCHASTIC DEMANDS CHAP. 5

a transactions basis, it will be done once per day or once every couple of
days). In such situations Q* may be quite large compared to the average
demand in the time between reviews, and here, of course, an nQ operating
doctrine would be much preferred to an order up to R doctrine. However,
for these cases one could almost use the results of a (Q, r) model for the
periodic review Q and r. It may not always be possible to do this, though,

and have r determined accurately enough. When Q* is of the order of Qw,

then normally, the r + Q terms in E and B can be ignored. In the next
section, we shall present a simplified version of the (nQ, r, T) model
obtained by dropping the r + Q terms in E and B, which will hold for large
Q values.
Suggestions for the use of an nQ type of operating doctrine have appeared
in several places [4, for example]. The model developed in this section was
first presented in [3].
SEC. 5-4 FORM OF THE (nQ,r,T> MODEL FOR LARGE Q 255

5-4 Approximate Form of the (nQ, r, T) Model for Large Q

We shall here consider the case where Q is sufficiently large to justify ignor¬
ing the r + Q terms in E(Q, r, T) and B(Q, r, T) in (5-33). Then (5-33)
reduces to
„ j , A . TnrQ .1 . XTl
X = f + y Vor + rc|_2 + 2 + r - M - ~2\
+ | A(r, T) + T(r, T) (5-34)

If Q*, r* are the smallest values of Q, r which minimize 3C for a given T, it is

necessary that

AqX(Q*, r*) = X(Q*, r*) - 3C(Q* - 1, r*) <0;

AqX(Q* + 1, r*) > 0 (5-35)

Ar3C(Q*, r*) = ac(Q* r*) - JC(Q* r* - 1) < 0;

Ar3C(Q*, r* + 1) > 0 (5-36)

Let us now compute AqX(Q, r). From (5-21) it follows that

XT
AflP* = - Q(Q^TYj [1 - P(Q - 1; XT7)] (5-37)
Hence

A«3C«J,r) - - {xA[1 - P(Q - 1 ;XT)] - ^<3(Q - D

+ 7rA(r, T) + (♦ + JC)T(r, T)| (5-38)

Next consider Ar3C(Q, r). The following are needed in performing the
differencing:

Ar{P(r - 1; X<)> = - r(r(~)21} P(r; Xt) (5-39)

AT{rP(r; Xt)} = - ^ p(r; \t) + P(r; Xt) (5-40)

Ar{r(r + l)P(r + 1; Xt)} = — r(r + l)p(r; X£) + 2rP(r; Xi) (5-41)

Ar{r(r + l)(r + 2)P(r + 2; \t)} = r(r + 1) ^3P(r; Xt)

X<(^f) + 3]p(r;Xi)| (5-42)

256 PERIODIC REVIEW MODELS WITH STOCHASTIC DEMANDS CHAP. 5

In each of the above, the differences were converted to a form which involve
only p(r; \t) and P(r; Xi).
It follows that

ArA(r, T) = | [Ar/3(r, r + T) - Ari3(r, r)] (5-43)

where
Ar|3(r, t) = (r — Xt)P(r; \t) — rp(r; X<) (5-44)
Similarly
ArT(r, T) = ArE(r, r + T) - ArE(r, r) (5-45)
where

ArS(r, f) = ^ [(XO2 - 2r\4 + r(r + l)]P(r; X«)

(5-46)
+ [Xi - (r + l)]p(r; Xi)
Then

Ar3C(Q, r) = IC + ^ ArA(r, 21) + ~^IC) ArT(r, T) (5-47)

In (5-47) it is in general not a good approximation to neglect the terms

involving r in comparison to the terms involving r + T. This is especially
true since we are assuming that Q is large, which will often imply that T
is relatively small.
To evaluate numerically Q*(T), r*(T) one can begin with an initial guess
for Q, call it Qi. Then using Qi for Q in (5-47) determine ri by finding the
largest r for which Ar3C(Q, r) < 0. Use n in (5-38) and determine Q2 by
finding the largest Q for which Aq3Z(Q, r) < 0. Then use Q-i in (5-47) to
determine t%, etc. The computational procedure is fairly time consuming,
hut not nearly so much as the one suggested in the previous section. The
example of the previous section suggests that one should begin with
Qi = Qw In fact, it may often be quite satisfactory to use Qw as Q*, and
simply determine the optimal value of r when Q = Qw using (5-47).

5-5 The (nQ, r, T) Model for Normally Distributed Demands

We shall now work out the equations for the (nQ, r, T) model for the case
where the demand density for a time interval of length t is normal with
mean Xi and variance Dt. The demand variable in any relevant time inter¬
val will therefore be treated as continuous. Furthermore, Q and r will also
be treated as continuous variables.
To begin we shall show that if v(x; T) is the density function for the
demand per period (x assumed continuous), then, immediately after a
SEC. 5-5 NORMALLY DISTRIBUTED DEMANDS 257

review, the distribution of the inventory position will be uniform, provided

that demands in different periods are independent random variables. Note
that this result depends only on the fact that demands in different periods
are independent, not that demands in any non-overlapping intervals are
independent. It is only when we wish to allow T to vary over any positive
values that this latter restriction is required. Let p(r + y) dy be the
steady state probability that the inventory position lies between r + y
and r + y + dy immediately following a review. The same sort of analysis
used in Sec. 5-3 shows that p(r + y) must satisfy

p(r + £)g(£; y) d% (5-48)

where
QO

?(l; v) = X) *>(! — y + nQ-, T) (5-49)

and « = 0ify<£, «~1, ify>£. Now it is only necessary to note that

+ nQ; T)d$= [ »(€; T) d£ = 1

J° »=« Jo
Therefore p(r + y) = constant 0 < y < Q, is a solution to (5-48). Since

p(r + y) dy = 1
it follows that

(5-50)
[0, otherwise

We are now ready to work out the various average annual costs for the
case where demand is normally distributed. The average annual review
cost is again J/T, and the average annual ordering cost is Apor/T where
the probability of placing an order at any review time is now

1i /■* /tf-xrw.
fQ^/y- \T\ , VDT f(Q-^/VoT
f .
" « j. 4 \VW) dV-—
*«)# (5-51)

Thus on using A4-6, and recalling that if the normal approximation is to

be valid, the value of the normal density at y = 0 must be essentially zero
and the value of the complementary cumulative essentially unity, we see
that
_^|"i T I ±(Q-\T\ VDT . /Q - XT\
Vor~ Q L \Vdt)\+ \Vdt) Q Kvdt) ( }
258 PERIODIC REVIEW MODELS WITH STOCHASTIC DEMANDS CHAP. 5

It is easily seen that the average annual cost of carrying inventory is

IC[^ + r~ n-Y + B(Q’r’T)] (5'53)

where B(Q, r, T) is, of course, the average number of unit years of shortage
incurred per year, i.e., the expected number of backorders at any point in
time.
It remains to obtain explicit formulas for the average number of back-
orders incurred per year and the average number of unit years of shortage
incurred per year. The same method of analysis is used as for the discrete
case studied in Sec. 5-3. It is clear that the average number of backorders
incurred per year is

E(Q,r, T) = £= fQ fm (I - r - y) -1 0p-X(r + T)1

QT JO Jr+y WD(t + T) L VD(t + T) J

~ f5'54)

However, by A4-25

(5_55)
Thus from A4-6, A4-26,

Dt
2
and

m **, T) = ^ [fc(r, T) - h(r + Q, T)] (5-57)

where

h(v, T) — y [l(v, T + t) — l(v, r)] (5-58)

SEC 5-5 NORMALLY DISTRIBUTED DEMANDS 259

Similarly
1 rQ rr+T
lj£-r-y)vsi*(tvm)did,d»
(5-59)
1 fr+Q fr+T r*
d£ dt du
qt L I L
We shall evaluate B(Q, r, T) in several steps. First, by use of the properties
in Appendix 4, we see that

~D2 + 2XV
U{a'r)- L L 4X3

_D3/2tI/2
X2

(5-60)

Next

Xr5/22)l/2 Z)l/2T3/2r ( £)mTmr2

_ ll _ * (L.ZL±l\ I fXr5/22)1/2 iwy/y ZWV'V
6X 8X4J V VDt / L 6 ~ 3 +
+ 6X

_l_ ^>3/2^/2 , Z)3/2r1/2r j}5/2ri/2~[ /r _ Xt\

+ 12X + 4X2 + 4X3 J*V )

+Si«iwiK^) <«»
Then

F(Q, r, T) = ^ [T(r, T) - T(r + Q, T)] (5-62)

where

T(«>, T) = | {S(v, t + T)- S(v, t )}

and S is defined by (5-61).
The average annual cost can now be written

3C = ^ + ^Por + /C^-t-r- — n — irE(Q, r, T)

+ (JC + w)B{Q, r, T) (5-63)

Unfortunately, 3C is not in general a convex function of Q and r. This
260 PERIODIC REVIEW MODELS WITH STOCHASTIC DEMANDS CHAP. 5

complicates the task of computing Q* and r*, since there is no guarantee

that there won’t be a number of local minima. Furthermore, it is very
difficult to obtain any solution to dX/dQ — dX/dr = 0. The task of de¬
termining Q* and r* is quite arduous. It appears to be necessary to use a
computer and to apply some appropriate search procedure just as in the
discrete case.

5-6 Exact Equations for (R, T) Models

It is easy to determine the exact ^equations for an (Rr T) model when the
humber_of units demanded in any time period is Poisson distributed and
the lead time is constant, since it is only necessary to set Q = 1 and R —
r + 1 in the results of Sec. 5-3. First, from (5-20), it is clear that
por = XTp(0; XT) + P(2; XT) - p(l; XT) + P(2; XT) = P( 1; XT) (5-64)
as expected. It is easy to evaluate E(l, R — 1, T), B{ 1, R — 1, T), since
in (5-21), (5-25), when Q = 1, it is unnecessary to sum over u from r + 1
to r + Q; it is only necessary to set u = r + 1 = R. Thus

E( 1, R - 1, T) = | {X(r + T)P[R - 1; X(r + T)] - RP[R; X(r + T)]

- \rP(R - 1; Xr) + RP(R; Xr)} (5-65)

and

J5(l, R - 1, T) = | [| {(r + T)>P[R - 1; X(r + T)] - PP{R - 1; Xr)}

+ RiR^ ^ {P[R + 1; x(r + T)] - P(R + 1; Xr)}

- R {(r + T)P[R; X(r + T)] - rP(R; Xr)} ] (5-66)

Thus, the average annual cost is

3C(R) = | + |P(1; XT) + 1C [e - n - y] + R — 1,T)

+ (JC + *)B( 1, R - 1, T) (5-67)

Then

A3C(R) =IC-Z {.P[R; X(r + T) - P(R;\r)} +

[- | {(t + T)*p[R - 2; X(r + T)] - Pp(R - 2; Xr)}

SEC. 5-6 EXACT EQUATIONS FOR (R, T) MODELS 261

+ ^ {2RP[R; X(r + T)] - 2RP(R] Xr)

- R(R + l)p[R;\(r + T)]

+ R(R + 1 )p(R; Xr)> - {(r + T)PIR; X(r + T)] - rP(R; Xr)

- (r + T)(R - 1 )p[E - l;X(r+ T)] + r(i? - 1 )p(R - l;Xr)}]

= IC - \ [x - f (IC + *)] {PIR; X(r + T)] - P(R; Xr)}

- {IC^— {(r + T)P[R; x(r + T)] - rF(E; Xt)}

- (7Cy- {?[«; x(r + D3 - P(B; Xt)} (5-68)

The smallest value of R which minimizes X is the largest R for which

AX(fi) < 0. The expression for A0C(B) is fairly simple, and it is not too
difficult to carry out numerical computations using it. In many instances,
however, the terms involving Xr are negligible with respect to those in¬
volving \(r T)j and in this case, A3C(fi) reduces to the simpler form

AK(R) = ic - \ {x - [f - (r + T)] (IC + #)} P[B; X(r + T)]

- {IC^ ® Rp[R;\(r + T)] (5-69)

Note that (5-68) holds for all T while (5-69) holds only if T is sufficiently
large. To determine the optimal T it is probably simplest to'tabulate X
optimized with respect to R as a function of T.
Let us next obtain the exact equations for the (R, T) model when R is
treated as continuous and the demand in any relevant time interval can
be assumed to be normally distributed with mean Xf and variance
To obtain the desired formulas from the results of Sec. 5-5, we must set
r = R and take the limit as Q -»0 rather than setting r + l = F, Q=1
as in the discrete case. For the continuous case it is clear that por should
be unity, since the probability that a positive quantity will be demanded
in any time period is unity. This is indeed the result obtained from (5-52)
when Q = 0, and the assumption is made that the normal density function
is essentially zero for non-positive arguments.
To compute the average number of backorders incurred per year and
the average unit years of shortage incurred per year, i.e., E{0, R, T),
B(0, R, T), note that for any function/(y) that
262 PERIODIC REVIEW MODELS WITH STOCHASTIC DEMANDS CHAP. 5

^p(V)dy-m) (5-70)

since by definition of the derivative, if F(y) = / f(y) dy then

I
lim^ f(y) dy = lim m)-F(O) _dF
= >o = /(0) (5-71)
Q-»0 y jo Q—*0 Q

It follows immediately from (5-54) that

E(o, 72, r) = ^{VD(r + T) t p.r X(r + D~

rl L Vd(t + T) .

+ (S_Xr)4(^i)} (w
(5-72)

Similarly, from (5-59) and (5-60)

B(0, R,T) = y [ff(fl, r + T) - U(R, r)] (5-73)

where U is defined by (5-60).

Thus the average annual cost is

3C = | + 7c[t2-m-^] + xl?(0, R, T)
+ (IC + *)B(0, R, T) (5-74)
where L = J + A. The optimal value of R for a given T can be found by
solving the equation

- n - tp -u ^ ^
- 0 - 7C + + (7C + x) rs (5-75)
where
57? _ 1
572 “ 7

5E 1 VdU(R,r+ T) dU(R, tY
dR T (5-77)

17(72,2) = (2

—— g2XR/n *
2X2 ® (^)
SEC. 5-7 LIMIT OF <nQ, r, T) MODEL AS T-»0 263

Probably the simplest way to determine T* is to compute 3C(E*) for a

sequence of T values and to determine T* graphically as described in Sec.
5-2.

5-7 The (Q, r) Model as the Limit as 7—>0 of the (nQ, r, 7) Model

The total cost expression (5-33) for the {nQ, r, T) model holds for all T > 0.
As T —»0, the average, annual cost approaches infinity if the review cost
J > 0. However, if the review costs are omitted, then the average annual
cost should approach a finite limit as T —»0. In fact, this limit should be
precisely the average annual cost for a (Q, r) model, and Q*(T), r*(T)
should approach Q*, r* for the {Q, r) model with the same values of the
parameters. We shall now prove this. Recall that in the derivation of
(5-33) it was assumed that the demand over any time interval was Poisson
distributed and that lead times were constant. Thus the limit as T—>0
should be a {Q, r) model with demand being Poisson distributed and con¬
stant lead times, i.e., the model developed in Sec. 4-7. The result also
holds true when the lead times must be considered to be random variables,
provided the approximations introduced in Chapter 4 that the lead times
can be simultaneously treated as independent random variables without
any crossing of orders are valid. The proof in this case is left for Problem
5-18.
The continuity of transition between (Q, r) models and {nQ, r, T) models
referred to above is important, because by including the review costs in the
{nQ, r, T) model, and adding the average annual cost of operating the
transactions reporting system to the average annual cost of the (Q, r)
model, it is theoretically possible to determine whether it is better to use a
transactions reporting system or a periodic review system with an nQ type
of operating doctrine.
Let us now show that as T —»0, X for the (nQ, r, T) model with demand
being Poisson distributed and constant lead times approaches X for the
(Q, r) model when the review costs are omitted. It follows from this that
Q*(T), r*(T), the optimal Q, r values for a given T, approach Q*, r* for the
(Q, r) model, since the review costs do not influence the determination of
Q* and r*.
First observe from (5-33) and (5-20)

lim 2* = £ - £ iim P(Q; XT) + lhn±P(Q + l; XT)

T—*0 I Q Q T—K) T—■*-

Inasmuch as Q > 1
lim P(Q; XT) = lim |p(Q + 1; XT) = 0
T—>0 T-K) 1
264 PERIODIC REVIEW MODELS WITH STOCHASTIC DEMANDS CHAP. 5

since p(j; XT) approaches zero as T—>0 for all j except j = 0 (which is
excluded in the above summations). Thus

Si If = Q (5'79)

and the ordering costs approach \A/Q, which is the correct term for the
(Q, r) model.
Consider next
-j r+Q «
lim E(Q, r,T) = lirn ^ Z (x - u) {p[x; X(r + T)] - p(x; Xr)}
u = r+l x=u

By definition of the derivative

[x; X(r + T)] — Xr)

T

= X[p(x — 1; Xr) — p(x; Xr)]

However

Z (* - «)[p(* - 1; Xr) - p(x; Xr)]

x=* u
°° oo
= Z (* — u)p(x — 1; Xr) —
X—U
y (x — w)p(x; Xr)
x—U

= Z (y — tt)p(j/;Xr) + Z
2/ = w—1 j/=w—1
p(t/; Xr)

00

- 2] (* - u)p(x; Xr) = P(w; Xr)

X—U

Therefore

lim P(Q, r, T) = — P(u; Xr) (5-80)

which from (4-45), (4-46) is F(Q, r).

Finally, consider

i HlQ /*r+r 00
hm B(Q, r, !T) = lim Wy J] / Z ~ u)p(x; X?) d£

By definition of the derivative

1 l*T+r 00

Jt Z - X?) d£= Y; (x ~ u)p(x; Xr)

SEC. 5-8 MODELS OF THE <R, r, T) TYPE 265

Thus
1 H-Q j*

lim B(Q, r, T) = q ^ (x ~ “Ms; Xr) (5-81)

which is equivalent to (4-47). Of course, the inventory carrying cost term

approaches the corresponding one for the (Q, r) model, since \T/2 goes to
zero with T.
We have shown that each term of the average annual cost for the
(nQ, r, T) model approaches the corresponding term for the (Q, r) model
as T —»0. Hence Q*(T) and r*(T) for the (nQ, r, T) model approach Q*
and r* for the (Q, r) model, and this demonstrates what we desired to show.

5-8 Models of the (R, r, 7) Type

It will be recalled that if a periodic review system uses an Rr operating

doctrine, then if at a review time the inventory position (in the backorders
case) or the quantity on hand plus on order (in the lost sales case) is less
than or equal to r, a quantity is ordered which is sufficient to bring the
inventory position or the quantity on hand plus on order up to R. For
the types of costs and stochastic demands used in this chapter it is usually
true that an Rr operating doctrine is the optimal one, if all demands occur¬
ring when the system is out of stock are backordered. This will be shown
in Chapter 8. The nQ and order up to R doctrines are only approximations
to the optimal Rr doctrine. We shall obtain below the exact expression
for the average annual cost in the backorders case when an Rr operating
doctrine is used, the number of units demanded in any time period is
Poisson distributed, and the lead times are constant. The derivation,
however, will not be quite so simple as for the (R} T) and (nQ, r, T) models.
Furthermore, it will be seen that the task of computing R* and r* is ex¬
ceptionally arduous and would be almost impossible to carry out manually,
although the computations could be made easily on a large scale digital
computer.
Before restricting our attention to the case where the number of units
demanded in any time period is Poisson distributed, let us begin as we did
for the (nQ, r, T) model by attempting to compute the steady state prob¬
ability p(r + j) that the inventory position immediately after a review has
the value r + j, for any distribution p(x; T) of the demand in a period.
The only assumption to be made will be that demands in different periods
are independent random variables. If we knew the p(r + j), then we could
compute the average annual cost just as we did for the (nQ, r, T) model.
Here also, transitions between these states can be considered to be a
266 PERIODIC REVIEW MODELS WITH STOCHASTIC DEMANDS CHAP. 5

Markov process discrete in space and time. Let us then set up the equations
whose solution yields the p(r + j). The inventory position of the system
immediately after a review can only have one of the R — r values
r + 1, r + 2,. . . , To compute the transition probabilities a#,
assume that the inventory position immediately after a review is
r + i, i = 1,. . . , R — r. We must compute the probability that it is
in state r + j immediately after the next review. If j > i and j < R — r
this probability is zero, since there is no way in which such a transition can
be made. If j < i, the probability is p(i — j; T). When j = R — r and
i < R — r, the probability is P(i; T), where P(x; T) is the complementary
cumulative of p(x; T), since any number of demands which reduce the
inventory position by the next review to or below r will result in having the
inventory position immediately after the next review be R. When
i = R — r, then j = R — r if there are no demands or if there are R — r
or more demands, i.e., the probability of this is p(0; T) + P(R ~ r; T).
f
Thus

(P(i ~ j; T), j <i, j i* R~r

(5-82)
” \0, j>i, j^R-r

a-i.R-r = P(i; T), i^R — r (5-83)

o-r-t.r-t = p(0; T) + P(R ~ i", T) (5-84)

These transition probabilities can be conveniently represented as shown in

Table 5-2.

TABLE 5-2

Transition Probabilities for Rr Doctrine

Inventory Position After Review at Time t + T

R R - l R - 2 R - 3 r+1

P(0; T) +
R p(l; T) P(2; T) p(3; T) p(R - r - 1; T)
P(R - r; T)
►-*
4

Inventory R — I P(0; T) p(R - r - 2; T)

-i

9(1; T) P(2; T)
1

Position
After R — 2 P(R - r - 2; T) 0 9(0; T) 9(1; T) p(R - r - 3; T)
Review
at Time t R — 3 P(R - r - 3; T) 0 0 9(0; T) p(R — r — 4; T)

r+ 1 P(l; T) 0 0 0 9(0; T)

The steady state probabilities p(r + j) that the inventory position

immediately after a review is r + j will be the unique solution to the set
of equations

i
SEC 5-8 MODELS OF THE (f?,r,T> TYPE 267

R-r

p(r + i) = X) p{r + i)aij, j = l, ... ,R — r

^R (5'85)
X) ^=1
/=i

However, it is not at all apparent from what we have studied previously

how we can solve (5-85) and obtain explicit values for the p(r + j). It is
immediately obvious, though, that the states are no longer uniformly
distributed, since it is intuitively clear that state R should have a higher
probability than the other states. We shall make no further attempt to
solve these equations directly. In the next section we shall obtain the
p(r + j). We shall not do so, however, by solving (5-85) directly. Instead,
we shall derive the exact expression for the average annual cost by com¬
puting the expected cost per cycle (as usual, the system is said to go through
one cycle’s operation in the time between the placement of two successive
orders), and then multiplying by the average number of cycles per year.
The p(r + j) will be obtained as a byproduct of this computation.
It might be pointed out that if we knew the distribution of the inventory
position at the review times prior to placing any order it would be a simple
matter to compute the p(r + j). Let 6(x) be the steady state probability
that the inventory position of the system has the value rata review time
prior to placing an order. Note that x can have any integral value from
R to —oo. Then

p(r + j) = 6(r+j), j = 1, 2, . . . , R - r - 1 (5-86)

p(R) = 6(R) + «(*) (5-87)

X — — 00

We can also characterize transitions from the inventory position at one

review time (before any order is placed) to those at the next review time as
a Markov process. This makes it possible to develop a set of equations
from which the 6(x) can in principle be determined directly. We shall now
obtain these equations and note that they cannot be easily solved explicitly
either. It should be observed that a state y can be reached from states x
such that x> y, r + 1 <x and states x for which x < r (since for these
states an order will be placed raising the inventory position to R at the
start of the period). Thus it follows that the equations relating the 0(y)
are

e(y) = p(R — y;T) ^ e(*) + E K1- V‘> T)e(.x)> V ^ r O88)

a;* — * a;=»r-l-l
268 PERIODIC REVIEW MODELS WITH STOCHASTIC DEMANDS CHAP. 5

d(y) = p(R - y;T) 6(x) + “ V’ T)6(x)>

3C= —00 Z=y

r + 1 < 2/ < R (5-89)

These balance equations can be written down directly. However, the

reader is asked to provide a step-by-step derivation in Problem 5-19.
Equations (5-88), (5-89) represent an infinite set of equations to be solved
for 6{y). Unfortunately, it is no more apparent how to solve these equa¬
tions explicitly than it was (5-85). Hence we shall abandon the attempt to
derive the average annual cost treating the operation of the system as a
Markov process, and instead go on to use the technique of computing the
expected cost per cycle and then multiply by the average number of cycles
per year.

5-9 Derivation of Average Annual Cost for (R, r, T) Model

In this section we shall derive the exact average annual cost expression for
an (R, r, T) model by computing the expected cost per cycle and then
multiplying by the average number of cycles per year, hirst, we shall
obtain the form of the average annual cost without specifying in detail the
nature of the stochastic process generating demands. The only assumptions
that will be made concerning the stochastic process generating demands is
that it does not change with time and that demands in different periods are
independent. We shall also assume that the lead times are constant. It
will not be possible to write down explicitly all the cost terms since no
special assumptions are being made about the nature of the demand dis¬
tribution. However, from this cost expression, we shall be able to obtain
the general solution to (5-85) for the p(r + j), which we were unable to do
in the previous section. After this we shall restrict our attention to the case
where a Poisson process generates demands and units are demanded one
at a time. In this case we shall obtain explicitly the average annual cost as
a function of R, r, and T.
When an Rr operating doctrine is used, an order need not be placed at
each review. The time between the placing of two successive orders (i.e.,
the length of a cycle) will always be an integral multiple of the time T
between reviews; however, the number of periods included in a cycle will
be a random variable. Hence the length of a cycle is also a random variable
which can only assume the values nT, n = 1, 2, 3, ... .
Let p(x; T) be the probability that x units will be demanded in one
period. Then, since the demands in different periods are assumed to be
independent pM(x; T) is the probability that precisely x units are
demanded in n periods, where pin)(x; T) is the n-fold convolution of p(x; T).
SEC 5-9 DERIVATION OF AVERAGE ANNUAL COST 269

If t is a review time, let the inventory position after the decision as to

whether or not to order has been made be r + j, and H(r + j, T) be the
expected cost of carrying inventory and of backorders incurred from t + r
to t + r + T, where r is the procurement lead time.
Suppose now that an order is placed at time to. Next imagine that at
review time to + nT, the inventory position of the system is r + j and
that no order has been placed since to. This means that precisely R — r — j
units have been demanded in the time period of length nT since ta. The
probability of this is pM(R — r — j;T). The expected cost of car¬
rying inventory and of backorders incurred from to + nT + r to
<o + (n + 1)T + r will be H(r + j, T). Now the time from to until the
next order will be placed may require more than one or more than two or
more than three, etc., periods. If more than n periods are required, then
there will be a contribution to the cost in period n + 1. The contribution
to the expected cost in period n + 1 will depend on the inventory position
at the beginning of period n + 1, and will be H(r + j, T) if the inventory
position is r + j at the beginning of the period. The probability that more
than n periods will be included in the cycle and that the inventory position
is r + j at the beginning of period n + 1 is simply pM(R — r — j; T).
Thus the contribution to the expected cost per cycle of carrying inventory
and backorders from period n + \(n > 1) is

E V^(R -t-j; T)H(r + j, T)

3 = 1

There will always be at least one period per cycle and the contribution to
the expected cost from the first period is simply H(R, T), since immediately
after placing an order the inventory position is R. The expected cost per
cycle of carrying inventory and of backorders is then found by summing
the expected contributions for period n over all n > 1. Thus the expected
cost per cycle is

H(R, T)+J2J2 pM(R - r - j; T)H(r + j, T)

n—1

which can be written

E E P(n) (R-r-j; T)H(r + j, T) (5-90)

n=0 3 — 1

if we use the definitions

p(o)(o; T) = 1, pCB)(x; T) = 0 for x ^ 0

Having found the expected cost per cycle of carrying inventory and of
backorders, we must next compute the average number of cycles per
270 PERIODIC REVIEW MODELS WITH STOCHASTIC DEMANDS CHAP. 5

year. This is simply the reciprocal of the expected length of a cycle. Since
the length of a period is T, the expected length of a cycle is T times the
expected number of periods per cycle. Let us then compute the expected
number of periods per cycle. A cycle will be precisely one period if the
demand in the first period is greater than R ~ r. The probability of this
is P(R — r; T), where P(x; T) is the complementary cumulative of p(x; T).
Consider now the probability that a cycle will contain precisely n(n > 2)
periods. If R — r — j units have been demanded in the first n — 1 periods
after the previous order was placed, then the cycle will contain precisely
n periods if the demand in period n is for j or more units. Thus the prob¬
ability that a cycle contains precisely n periods is

J2p*-»(R-r-j;T)P(j;T)
y-i

Hence the expected number of periods in a cycle is

P(R - r; T) + Z Z np<”~»(R - r - j; T)P(j; T)

n=2 j—1
which can be written

Z Z np^KR ~r-j; T)P{j; T)

n=lj=\
(5-91)

if we use the convention described above with respect to p(0)(x; T).

We are now able to write down the average annual cost. If, as usual, J
is the cost of a review and A is the cost of placing an order, then

, PM(R ~r-j; T)H(r + j, T)

X(R, r,T) = ± +-^£7- (5-92)
T Z Z nPln~l)(R -r-j; T)P{j; T)
n=1j—1

When the distribution of demand is specified, it then becomes possible to

evaluate p(n)(x; T) and H(r + j, T) explicitly. Before doing this for the
case where a Poisson process generates demands and units are demanded
one at a time, we shall obtain from (5-92) the steady state probabilities
p(? + j) that the inventory position immediately after a review is r + j,
and hence we shall obtain the solution to the equations (5-85).
Note that the average annual cost of carrying inventory and of back¬
orders is 1/T times

H(r + j,T) (5-93)

SEC. 5-9 DERIVATION OF AVERAGE ANNUAL COST 271

so that (5-93) is the corresponding expected cost per period. However,

from the definition of the p(r + j) and H(r 4-j, T) the expected cost per
period is £f-T p(r +j)H(r +j, T). This suggests that

E PM& ~ r ~ r>

p(r + j) - „ ~k'-T° ’
EE -r-y, T)P(j; T)
n= 1 y= 1

y = 1,. . . , B — r (5-94)

We shall now show that this is indeed true. All that is required is to show
that the p(r + j) of (5-94) satisfy (5-85). When the a# given in Table 5-2
are substituted into the equations p(r + j) = p(r + they
become

p(R) = *£ 1 P(r + 0-P(»; T7) + P(*)[P(0; T) + P(R- r; T)] (5-95)

i— 1

p(r + j) = E P(r + 3 = 1, • • • > R — r ~ 1 (5-96)

i—3

Let us next show that

P(r + j) = 5 E P(H>@ -r~j;T), j = 1,R — r (5-97)

n=0

satisfy (5-95), (5-96) for any value of S. Consider first the equations (5-96).
When p(r + j) is given by (5-97)

E
i-j
p(r + i)p(i — 3) T) = 5 E E P(n)CR — r — i; T)p(i — j; T)
n = 0 i=j

= 5 E E
n—1 w=0
Pin~1}(R — r - j — u; T)p(w; T) = 5 E
n=X
— r —j;T)

The last step follows from the definition of yM(R — r - j\T) in terms of
p(n_1) and p (seep. 119). However, when j = 1,. . ., R — r — 1, according
to the definition used above p(0)(i? — r — j; T) = 0 so that

E P(,l>(# — r — j;T) = E P(n>(^ ~ 7 — j; T)

tt=0 n=l

Thus the p(r + j) given by (5-97) do indeed satisfy (5-96).

We shall now prove that the p(r + j) of (5-97) also satisfy (5-95). This
requires that we demonstrate that
 272 PERIODIC REVIEW MODELS WITH STOCHASTIC DEMANDS CHAP. 5

« R—t—1
1 + ^ r) = £ E
E
n—1
£
=1 i
P{n)(R - r - i; T)P(i; T)

+ l1 + E P(,,)(0; r)][p(0; T) + P(R- r; T)] (5-98)

SfftP<r!(0-;P,r ^ p(0>(x; ^ =
P( ,(0; T), it follows1 that x ** Now because p(0;T) =

P(0; T) + p(0; T) ]T pc«)(0; T) = p(»(0; T) + V p(0; T)p<-»(0- T)

n=1
00

= I] P(B,(0; T)
n— 1

Thus the right hand side of (5-98) can be written

« R—r

E E PM(R -r-i; T)P(i; T) + £ pW(0; D

n= 1

^ Consequently, to show that (5-98) holds is equivalent to showing that

00 R—r

E E PM(R ~r- i; T)P(i; T) = 1 (5-99)

Cleariy (5-99) does sum to unity since it is merely the probability that one
nW rIqUired t0 W R~r™te demanded. The analytic
proof can be obtained as follows y

E E PM(R ~r~i; T)P(i; T)

71 = 0 X = 1

= SS pM(B ~ r ~ i; T)~ E E E PM(R - r - i; T)p(u; T)

n-0 x = 1 u=o

" 1 + S £ pM(B - r ~ - £ £“‘ Tipiu; T)

n—0 v—Q u—Q

Thus the proof of (5-98) reduces to showing that

 SEC 5-9 DERIVATION OF AVERAGE ANNUAL COST 273

However
.R—r—1 R—r—v — 1
S5
n-u »-u «-u
pM(v’T^u’T) = E E E Vw(i~u;T)p(u-,T)
ji = 0 i=o u=0

« fi-r-1 t
— E 1=1
2D + E E E P(n-1)(^ — «; T)p(u; T)
n=2 i=0 u=0

- S-r-1 - R-r

=E E 20 = E E P(n)(B — r — i\T)

which is what we desired to show. Thus the p(r + y) of (5-97) do satisfy

(5-95) and (5-96).
It remains to show that when
oo .fi — r

5-1 = E E »PG-1,(B - r - i; T) Ptf; 20

n=l j~l

that i p(r + i) = 1. This is equivalent to showing that

co R—r
E E On)CR - r - i; T)
oo .R — r

= E E np^\B - r - j; T) P(j; T) (5-100)

7i = i y=i

To show that (5-100) is valid, note that

» R—r

E E »P<*"U(B - r - j; T) P(j- T)
n=iy=i

= E E (» + !) P(n)(R ~ r ~
71 = 0 /= 1
r, D P(J; D

= EEV”>(fi-r- y; 2D PC;; T)
n=o y=i

+ E E ^’(B ~ r - r, D P(j; T)
n=oy=i

However, by (5-99), we are reduced to showing that

274
PERIODIC REVIEW MODELS WITH STOCHASTIC DEMANDS CHAP. 5

Now

” K—T

Z Z npw(R -r- ji T) P(j; T) = Z Z nPM(R ~ r ~ j; T)

n — 1j=l
n — 1j — l

* R—r j— l
~ Z Z Z nPM(.R - r - j; T) p(u; T)
n=l 3 — 1 u=Q

g Z PU)(R -r-j;T)+f^J2(ri- 1) p^(R - r - j; T)

"E En=l i—0 u—0

Z nP(n)(*';r)p(m;t7)
« JB — r ^ jj_r

= Z Z PCn)(-K — r — j; T) + Z Z wP(n+1)(jR — r ~ j) T)
n=1 £1 7=i
00 R—r w
- ZZ - r - 3; T) = £ £ PW(B - r - i; T)

which is what we needed to demonstrate. We have thus shown that the

probability p(r + j) that the inventory position of the system immediately
after a review is r+j is given by (5-94). This holds for any distribution
of demand, the only assumption required being that demands in different
periods are independent. Furthermore, the result is independent of the
nature of the stochastic process which generates the lead times.
Let us now concentrate our attention on the case where the number of
units demanded in any time period is Poisson distributed. Then p(x; T)
= p(x; XT) and pM(x; T) = p(x; nXT). We can next evaluate H(r + j, T)
explicitly. We have already done this in developing (nQ, r, T) model.
The expected number of backorders incurred from t + r to t + r+ Tii
the inventory position is r + j at time t is

e(r +j,T) = Z (*-*•“ 3) {P[x; \(r + T)] - p(x; Xr)}

X = T-{-3

= X(r + T) P[r +j-l; X(r + T)] - (r + j) P[r + j; X(r + T)}

- Xr P(r + j- 1; Xr) + (r + j) P(r+j; Xr) (5-101)

Similarly, the expected number of unit years of shortage incurred from

t + r to t + r+ Tis
SEC 5-10 THE (R,r,T> MODEL WHEN DEMAND IS CONTINUOUS
275

fr+T “

+j,T)= (x - r ~ j) P(.x; X£)d£

Jr x=*r-\-j

" I ((r + Pl> + 3 - 1; x(r + T)] - r2 P(r + j - 1; Xr)}

+ (l+^^+j + D {P[r +j + 1. X(T + T)] _ p(r +j + 1; Xt)}

~(r+ j) {(r + T) P[r +j; X(r + T)] - TP(r + i; Xr)} (5-102)

by (5-26). Finally, the expected unit years of storage incurred from t 4-

to t + r + T is ^

d(r+j, (r +j — X|) d% + b(r + j, T)

= T (r + j - M + b(r + j, T)
(5-103)

Then H(r + j, T) is simply

H(r +i> ?) = r(r+j - h~y) + *<r + i, T)

+ (ir + IC) b(r + j, T) (5-104)
The average annual cost is then
« R—r
r A+ P(R ~ r - T, n\T) H(r + j, T)
X(R,r, T) = — + — V-V"1-—_
(5-105)
P Z) Z)
n = 1 y-1
- r - r, (« - 1)XT] P(j-XT)
It does not seem easy to reduce the summations appearing in (5-105) to a
more simple form. Clearly, it would be an impossibly laborious task to try
to determine R*, r* and T* manually using (5-105). However, the com¬
putation could be carried out quite easily on a large computer such as the
IBM 7090 and should not take over one minute on this computer.

5-10 The (R, r, T) Model When Demand is Treated as a

Continuous Variable

In this section we shall develop the analog of the (R, r> T) model studied
in the previous section for the case where R, r, and the demand variable
are treated as continuous.
 276 PERIODIC REVIEW MODELS WITH STOCHASTIC DEMANDS CHAP. 5

Let v(x; T) be the density function for the demand in a period. Demands
in different periods will be assumed to be independent random variables.
Again we shall assume that the lead time r is a constant. If the inventory
position immediately after a review occurring at time t is r + x, let
H(r + x, T) be the expected costs of carrying inventory and of backorders
incurred from t + r to t + r+T. Suppose that an order is placed at
time t0. The probability that at time to -f- uT) u > 1, the inventory posi¬
tion of the system lies between x and x + dx given that no order has been
placed since time to is v(n)(R — r — x; T)dx, where vin)(x; T) is the 71-fold
convolution of v(x; T). Thus the expected cost per cycle is

5 So VM(-R ~r~x> T)ff(r + *, T) dx + H(R, D (5-106)

The second term comes about because at t0 the inventory position is R;

here it is not convenient to try to include this term in the first term by
summing from n = 0 rather than from n = 1, as was done in the discrete
case.

The expected length of a cycle is

t Jq R - r-x-T)V(x; T) dx + V(R - r; T)J (5-107)

where 7(x; T) is the complementary cumulative of v(x; T). The average

annual cost then becomes

X(R, r, T) =
00 rR—r

T A + 1l vM(R -r-x; T)H{r + x, T) dx + H(R, T)

± _|_n = 1 JO_
T ' r * rR-r
TI Z, I nv(”-»(R -r-x; T)V(x; T) dx + V(R -r;T)\

(5-108)

We ask the reader in Problem 5-22 to work out H(r + x, T) and

}(.K — r — x; T) explicitly for the case where v(x; T) is a normal
distribution.
It is interesting to examine the behavior of the inventory position
immediately after a review. The behavior of the inventory position im¬
mediately after a review is peculiar in the sense that there is a positive
probability that the inventory position will be R, while for other values of
the inventory position, the probability of a specific value is zero (as is
true with continuous random variables), while there is a probability
p(f + y) dy that the inventory position lies between r + y and r + y + dy.
SEC. 5-10 THE (R, r, T) MODEL WHEN DEMAND IS CONTINUOUS 277

Let p{R) be the probability that the inventory position is R immediately

after a review. Then it is clear that the equations which p(R) and p(r + y)
must satisfy are

p(R) = p(fi)V{R — r;T) + f p(r + y)V(y; T) dy (5-109)

and °

p(r + y) = f p(r + x) v(x — y; T) dx (5-110)

Jv

From the average annual cost it is clear that

v{R) =

4e nv^\R -r-x; T)V(x; T) dx + V(R - r;

(5-111)

E vW(R -r-y,T)
p(r + y) =

[e
Ln = 2
r
JO
nv(n~l)(R -r-x; T)V(x; T) dx +V(R-r;'.

0 < y < R - r (5-112)

We shall leave the proof that (5-111) and (5-112) are indeed the solutions
to (5-109) and (5-110) to Problem 5-23. For any given distribution v(x; T),
it is generally very difficult by use of (5-111) and (5-112) to obtain p(R)
and p(r + y) in a simple form not involving a sum over n.
When one deals with the inventory position at a review time prior to the
placing of any order the peculiar behavior at r + x — R is not encountered.
If 6{x) is the density function for the inventory position at a review time
prior to placing any order, then it is clear, on considering the possible
transitions from one review time to the next, that 6{x) must satisfy the
equations.

6(x) = v(R — X) T)
(5-113)

6{x) = v(R — x; T)

(5-114)

In general, it is very difficult to attempt to solve (5-113), (5-114) for 6(x).

Note that if 9(x) is known, then
 278 PERIODIC REVIEW MODELS WITH STOCHASTIC DEMANDS CHAP. 5

p(x) = 6(x), r < x < R, p(R) = j e(x) dx (5-115)

There is one special form of v(x; T) for which it is possible to solve

(5-113), (5-114) explicitly for d(x), and for which it is also possible to carry
out the summation over n in (5-111), (5-112). This is the case where
v(x', T) is a gamma distribution. The difficulty of carrying out the solution
increases very quickly with the order of the gamma distribution. We
shall illustrate the procedure only for the first order case, i.e., that where
v(x; T) is an exponential distribution. We shall assume that

v(x; T) = fie*11*
Note that we do not implicitly indicate the time dependence here. We
imagine that T is fixed.
We shall first show how (5-113), (5-114) can be solve explicitly for B(x).
Note that
dv
— = -0* 6T* = -pv (5-116)

To solve (5-113), (5-114) we shall convert them to differential equations

with the aid of (5-116). Differentiating (5-113), (5-114) with respect to x,
we obtain

B'(x) = -i)’(R - x; T) 6(y) dy - J* 0(y) v'(y - x; T) dy,

x<r (5-117)

B'(x) = -v'(R - x; T) 8(y) dy - 6(x) v(0; T) - j\(y) v'(y - x; T) dy,

r <x < R (5-118)

Substitution of (5-116) into (5-117), (5-118) along with «(0; T) = p yields
B'(x) = p8(x), x < r; 6'(x) =0, r<x<R (5-119)
Integration of (5-119) yields

6(x) = h &*, x < r; B(x) = k2, r <x<R (5-120)

where fci, fa are constants yet to be determined. From (5-113), (5-114)
we see that 8(x) must be continuous at x = r. This requires that
fa = fa eBr

/
In order that 0(x) be a legitimate density function, it is necessary that
R

6(x) dx = 1 = e$r(R — r) + ki j
r

e$x dx = ki e$r ^ — r ^J
SEC. 5-10 THE <R, r, T) MODEL WHEN DEMAND IS CONTINUOUS 279

R — r +

and hence

r < x < R
•R — r +
(5-121)

\R-r + ■

Therefore from (5-115)

6(x) dx

(5-122)
p(r + y) 0 < y < R — r
R-r +

Let us now compute p(r + y) and p(R) from (5-111), (5-112) to show
that the same result is obtained. From (3-108), it follows that if v(x; T) =
pe~fixj then v{n)(x] T) is a gamma distribution of order n, i.e.,

v<*>(z; T) = = pp(n - 1; fix) (5-123)

V(x; T) = e-e*

f* T »«*-«(« - r - *; T)7(a;; T) dx = 1><">(R - r; T)

(5-124)
= p[n - 1; P(R - r)]
Hence

00 rl
Z
n-2 JO
-r-x; T)V(x; T) dx + V(R - r; T)

= Z) — 1; /3(R — r)3 = 2 (w + !)p[nJ — r)]

= Z np[n; j8(jB - r)] + Z p[n; /3(R - r)] = 0(R - r) + 1

 280 PERIODIC REVIEW MODELS WITH STOCHASTIC DEMANDS CHAP. 5

Finally

X) vM(K - r - *; T) = 0 2 p[ft - 1; KR-r- *)]

n=l n=l

00

= 0 13 - »• - z)] = /3
n=0
Hence

= ,9(ff _ r) + j! + 2/) = -"-j' 0 < y < R - r

ff-r + -

which is precisely (5-122). Thus, we have obtained the same result in two
different ways. To solve (5-113), (5-114) explicitly for 8(x) is considerably
more difficult when v(x; T) is a gamma density of order two. We leave this
case to be worked out in Problem 5-32. The procedure for solving (5-113),
(5-114) when v(x; T) is gamma distributed by reducing them to differential
equations has been presented by Karlin m Chapter 14 of [1],
To have obtained simple expressions for p(r + y) and p(R) when the
distribution of demand for a period is exponential is not especially valuable
from a practical point of view for several reasons. First, the exponential
distribution is seldom a realistic representation of the density for the de¬
mand in a period. Secondly, in order to compute H(x, T) it would be
necessary to use some other distribution such as the normal since H(x, T)
involves the demand from r to r + T, and the exponential distribution
does not provide for introducing time explicitly. Finally, the average
annual cost even in this case does not reduce to an expression which can be
easily handled manually. While the result just obtained does not seem to be
of great practical value, it is of interest from a theoretical point of view to
show how p(R), and p(r + y) may be computed in two entirely different
ways.

5-11 Comparison of the Periodic Review Operating Doctrines

In this chapter, three different operating doctrines have been considered

for use with periodic review systems. They are: (a) order up to R, (b)
nQ policy, (c) Rr doctrines. We have already noted that the order up to
R doctrine is a special case of both the nQ and Rr doctrines. The nQ doc¬
trine is also really a special type of Rr doctrine for which R — r = Q and
the order size must be a multiple of Q. Thus the doctrines are all related
to a certain extent. We have seen, however, that it is not equally easy to
make numerical computations for each type of operating doctrine. It is
relatively easy to make numerical computations manually for (R, T)
SEC 5-12 PERIODIC REVIEW VERSUS TRANSACTIONS REPORTING 281

models, but for (nQ, r, T) or (jB, r, T) models a computer will normally be

required. Because of the differences in computational effort required, it is
of interest to inquire under what circumstances will the order up to R
policy be essentially optimal, i.e., under what circumstances will the aver¬
age annual cost using an order up to R policy differ so little from the average
annual cost of an nQ or Rr policy that it is not worthwhile to make the
computations for the nQ or Rr cases.
The answer to the question just posed depends, of course, on the relative
magnitudes of the review and ordering costs. In the event that review
costs are high relative to ordering costs, then normally, it will be unde¬
sirable to have a review without placing an order since otherwise, the review
is essentially wasted. This means that an (R, r, T) model optimized with
respect to T should have the property that an order will be placed at almost
every review. However, this in turn means that the use of an (R, T) model
will yield essentially all the benefits of the (R, r, T) model, i.e., will yield
almost as low an average annual cost. Therefore, when review costs are
high relative to ordering costs, an order up to R doctrine should be essen¬
tially optimal. In the real world, it seems to be frequently true that the
review costs are considerably higher than the cost of placing an order, and
hence one would expect to find that in many practical situations one could
use an order up to R policy without great deviations from optimality.
It is only when ordering costs are high with respect to review costs that an
Rr doctrine could be considerably better than an order up to R doctrine.
In such cases the question arises as to whether it would not be better to
change over to a transactions reporting system rather than using periodic
review. If this is not the case, then it may be advantageous to use either
an nQ or Rr operating doctrine. Although the (R, r, T) model will in gen¬
eral yield a lower average annual cost than an (nQ, r, T) model, it would be
expected that the differences would be rather small, and in any real world
situation, one could probably use either one if an order up to R doctrine was
not satisfactory.

5-12 Comparison of Periodic Review and

Transactions Reporting Systems

When review costs are ignored, a transactions reporting system always has
lower average annual costs, all other things being equal, than a periodic
review system. The principal reason for this is that in the periodic review
system, sufficient safety stock must be held to offer protection for a length
of time t + T, while for a transactions reporting system r is the relevant
length of time. Thus the periodic review system will require higher safety
stocks and have higher costs. The thing that really determines which sys-
 282 PERIODIC REVIEW MODELS WITH STOCHASTIC DEMANDS CHAP. 5

tem is to be preferred is the cost of operating a transactions reporting

system, as compared with that of operating a periodic review system. These
costs can vary widely with the type of inventory system under considera¬
tion, and in any particular case, either transactions reporting or periodic
review may be favored. No general statement can be made as to whether
one is to be preferred or the other. As has been already noted, in industry
and the military, periodic review systems have in the past been used much
more extensively than transactions reporting, but with the advent of auto¬
matic data processing equipment, there has been an increasing trend both
in industry and the military towards using transactions reporting. It does
not follow that, when a system that carries many items changes from
periodic review to transactions reporting, it will do so for all items. In
many cases it will be profitable to use transactions reporting for some items
but not on others. Of course, there exist some relatively rare situations in
which one must use a periodic review system, perhaps, for example, because
orders can be placed only at certain specified times, and in these cases a
comparison with the costs of operating a transactions recording system is
irrelevant.

5-13 The Lost Sales Case

Except for the simple, approximate model of Sec. 5-2, all the models in this
chapter have concentrated on the backorders case. We noted in Chapter 4
that it was not possible to treat rigorously the lost sales case when more
than a single order was outstanding. For periodic review systems, things
are even more difficult, and trouble is encountered even in treating the case
where only a single order is outstanding.
To illustrate the difficulties, consider the simplest case, i.e., that where
the number of units demanded in any time period is Poisson distributed,
t e procurement lead time is a constant t, and an order up to R policy is
used. The additional stipulation will be imposed that there is never more
than a single order outstanding. This will be rigorously true if r < T.
In order to compute the average annual carrying costs and stockout costs,
it is necessary to determine the distribution of the on hand inventory, either
at a review time or immediately after an order arrives.
Let 6(x) be the probability that x units are on hand at a review time.
If x units are on hand, then R — x units are ordered, so that the quantity
on hand plus on order immediately after the order is placed is R. To set up
the equations which determine 6(x), it is convenient to divide the time
between two successive reviews into two sub-periods, the first extending
over a time period of length r from the time that the order is placed until
SEC 5-13 THE LOST SALES CASE 283

it arrives, and the other extending over a time period of length T — r from
the time that the order arrives until the next order is placed. It should be
noted that it is possible to incur lost sales in both of these sub-periods.
Let 4f(z) be the probability that z units are on hand immediately after an
order arrives. The time after an order arrives until the next review is
T — t. The probability that x units are on hand at the next review, if z
units are on hand immediately after an order arrives, is

p[z — X) \(T — r)], 0 < x < z

P[z;\(T - r)], x = 0 (5-125)
0, x > z
Thus
R
o(x) =
z=x
E Hz)p[z - z; HT - t)], x > o (5-126)

m = E *(*)p
2=0
X(-T - ^ (5-127)
Next, by considering the time period extending from 0 to r, we shall
express \f/(z) explicitly in terms of Q{x). This can then be substituted into
the above equations to yield a set of equations involving Q{x) only (or
involving ip{z) only). If z units are on hand after an order arrives at time r,
the quantity on hand x at the time the order was placed could not have been
less than R — z> since the size of the order is R — x. If the quantity on
hand at the review time satisfies x > R — z, then in order to have z units
on hand after the order arrives, it is necessary that the lead time demand
be R — z. If x — R — z, then for any lead time demand > R — z, the
quantity on hand after the arrival of the order must be z. Thus we have

R
}p(z) = p(R — z; Xr) ^2 z; Xr)6(R — z),
x=*R—2+1
z > 0 (5-128)

*(0) = p(p; Xr)ff(fi) (5-129)

Substitution of (5-128), (5-129) into (5-126), (5-127) yields

o(x) =E
z~x i
~ Xt)p[z ~ x> x(r - T)1 E y—R—z-\-l
e^>

+E
z=x V
_ z> Xt)pCz - x> x(r - r)]0CS ~ z) * > 0 (5-130)
 284 PERIODIC REVIEW MODELS WITH STOCHASTIC DEMANDS CHAP. 5

*(°) = 2=1
Z W ~ 2; Xr)P[2; \(T - r)]
^
J
y = R~z+l
%)}
J

+ Z0 \p(R
z= 1
~ z'> Xr)P[z; x(T — t)]8(r — z) (5-131)

These equations can be rearranged to read

e(x) = Z a(x> v)eiR -y),

y—0
X = 0,1, ... ,R (5-132)
where
it
«(*» y) = Z viR - *; Xr)p[2 - x; X(T - r)],
Z-x

y = 0, 1,. . . , x - 1; x > 0 (5-133)

“(*• y) = Z 2 =2/4-1
P(R - z; XtMz - x; \(T - r)] (5-134)

+ P(R - y; \r)p[y - a; X(P - r)], y = x, . . . , R - 1; x > 0

a(x, R) = P(0; Xr)p[P — x; \(T — r)], a; > 0 (5-135)
R

“(0, y) = Z P(S - Xr)P[2; X(P - r)]

*-v+i

+ P(22-y;XT)P[y;X(3,-T)]> y = 0, 1,. . . , P - 1 (5-136)

a(0, R) = P(0; Xr)P[P; X(T - r)] (5-137)

If the values of 0(2;) from (5-126), (5-127) are substituted into (5-128),
(5-129) we obtain the set of equations which involve ^(2) only. They are
R R
Hz) = P(R - z; Xr) J2 Hy)p\y - x; \(T - r)]
x = R — s-f-l 2/ = a:

R
+ P(P - 2; Xr) *(j/)p[y - P + z; X(T - r)],
y — R—z

z = 1,. . ., R - 1 (5-138)

iA(P) = p(0; Xr) ^2 J2 Hy)p[y - x; \(T - r)]

1
x— y=x

+ P(0; Xr) Hy)P[y, x(r - r)] (5-139)

y—0

*(0) = P(R; Xr)p[0; \(T - T)]HR) (5-140)

SEC. 5-14 STOCHASTIC LEAD TIMES 285

These equations can also be rearranged to read

R

Hz) = XI Hz, y)HR - y) (5-141)

y=0

We leave for Problem 5-25 the computation of the f3(z} y).

Unfortunately, it does not seem possible to solve the set of equations
(5-132) or (5-141) explicitly. For any given value of R, it is straightforward
to determine numerically either the 6(x) or 1p(z), but this is not what is
needed.
If one could compute the $(x) or \p(z) it would be somewhat more con¬
venient to have the ^(2), because then it would be unnecessary to divide
the period up into two sub-periods. The average number of lost sales
incurred per year would be

E(It, T) = \ X Hz) X (« - *)p(«; XT) (5-142)

and the average unit years of storage incurred per year would be

D(R, T) =
1
X Hz)
2=0
I T X (* - *)?(*; Xi) dt
XwsO
(5-143)

The average annual cost when the unit cost of the item is constant is then

3C = j, + j, P(1; XT) + ICD(R, T) + irE(R, T) (5-144)

We have now seen how to formulate the model if the distribution ^(z) was
available. Unfortunately, as we have noted, it does not seem possible to
solve for the tl/(z) explicitly, and thus the terms D(R, T) and E(R, T)
cannot be evaluated explicitly. Hence, for periodic review systems, it is
not possible to formulate exact models for the lost sales case even when
the requirement is made that never more than a single order is outstanding.
About the best one can do is to use the simple approximate treatment of
lost sales given in Sec. 5-2. Fortunately, for many practical applications,
this is adequate.

5-14 Stochastic Lead Times

It has already been noted in Sec. 5-2 that for periodic review models, unlike
(Q, r) models, it is possible to treat rigorously stochastic lead times when
more than a single order can be outstanding at any point in time provided
that the range of lead time values is less than T. In this case, there can
 286 PERIODIC REVIEW MODELS WITH STOCHASTIC DEMANDS CHAP. 5

be no overlap of the lead time densities, and one can rigorously require that
lead times be independent random variables and also that orders be re¬
ceived in the same sequence in which they were placed. The reason that
stochastic lead times can be treated rigorously in certain cases for periodic
review systems, while this is not possible with (Q, r) models, lies in the fact
that for periodic review systems, the time between the placement of orders
must be at least T, while for (Q, r) models there is a positive probability
that two orders will be placed in an arbitrarily small time interval.
Let us show how to incorporate stochastic lead times into the <nQ, r, T)
model of Sec. 5-3. The notation and terminology of that section will be
employed here. We shall assume that the lead times are independent ran¬
dom variables and that orders are received in the same sequence in which
they were placed. This can be done rigorously if the range of possible lead
time values is less than T\ otherwise, the treatment is only approximate.
If a review takes place at time t and an order is placed, and if an order is
also placed at time t + T, then let tj, t2 be the random variables repre¬
senting the lead times of the orders placed at t and t + T respectively.
The lead time density function will be written g(r). Thus the expected
number of backorders incurred in the period between the arrival of the two
orders under consideration is
1 f* /*» **+Q « f n

Q L L 23 23
Jv u=r+li = u
— w) i V[z; k(r2 + T)] — p(x; Xn) f g(.T2)g(Tj) dr2 dn
'v J
r+Q * r >
7) 23 23 (x ~ w) ]
^u = r+li=u 1
T) — h(x) y
J
(5-145)
where

Hx) = p(x; \r)g(r) dr (5446)

is the marginal distribution of lead time demand and

h(x) T) — p[x; X(r + T)]g(r) dr (5-147)

The limits on the integrals have been taken to be 0 to oo. Of course, g{r)
may be 0 over most of this range so that the limits could be replaced by
'Tmin and rmax. Clearly, the assumption that lead times are independent and
arrive in the order placed could not hold rigorously if g(r) > 0 for all r
between 0 and co. In particular, the assumption could not hold rigorously
if g(r) was represented by a gamma distribution. In this case, the treatment
would at best be approximate.
The expected value of the unit years of shortage incurred between the
arrival of the two orders is
SEC 5-14 STOCHASTIC LEAD TIMES 287

1 f00 f 00 fT2*^” ^ Q 80
q / / £ X) (* - Wg(Ti)g(r2) d£ dn dn (5-148)
''■'JO Jo Jn

It is not necessarily true that an order will be placed at each review. In

such cases it is necessary to apportion the expected number of backorders
and unit years of shortage incurred between the arrival of two successive
orders to as many periods as are involved in the time between the placement
of two successive orders. This can be done in any way we choose provided
that all backorders and unit years of shortage are counted somewhere.
A convenient way to do this is to proceed as if an order was placed at each
review, the number of units ordered being zero when an order is not in
actuality placed.* The lead time density for a zero order quantity is the
same as for a positive order quantity. With this convention, the expected
number of backorders and unit years of shortage incurred per period are in
every case given by (5-145), (5-148) respectively. Then the expected
number of backorders and unit years of shortage incurred per year are

-j r+Q 00
m, r, T) = n £ £(*-«) [h(x; T) - h(x)] (5-149)
u=r+lx=u

r-f-Q so
1 C80 f00 f T2~^* ^
B(Q, r, T)
QTJo Jo L £ £
u=rJrl x = u

(x — u)p(x; A£)0(ti)0(t2) d? dn dr2 (5-150)

These changes allow the inclusion of stochastic lead times in the model
developed previously.
It is only for very special forms for g{r) that it is at all simple to work out
explicitly E(Q, r, T) and B(Q, r, T). This can be done quite easily if r is
gamma distributed. Problem 5-27 asks the reader to do this. Of course, for
a gamma distribution, it will never be rigorously true that the range of
lead time values will be less than T. When the demand variable is treated
as continuous and normally distributed, the usual procedure used to take
account of stochastic lead times is simply to change the variance of the
normal distribution, i.e., the value of D. In this case, the value of D for the
t terms need not be the same as the value of D in the r + T terms.
Stochastic lead times can be handled in (E, r, T) models in precisely the
same way as for the (nQ, r, T) model illustrated above. The details are
left for Problem 5-28.

* The reader should be careful to note that the use of this artifice in no way changes
the fact that an ordering cost is incurred only if a positive quantity is ordered.
 288 PERIODIC REVIEW MODELS WITH STOCHASTIC DEMANDS CHAP. 5

5-15 Quantity Discounts

When quantity discounts are offered which depend on the order size an
Rr or order up to R policy need not he optimal. To operate the system
optimally, a function Q(%) would have to he computed, using the techniques
of dynamic programming discussed in Chapters 7 and X, which gives the
quantity Q to order as a function of the inventory position £ at the review
time. In the real world, one will often use an order up to It, Hr or nQ policy
even when quantity discounts which depend on the order size are available
since it is too complicated to attempt to compute and use the Q(£) function!
The quantity discounts may, however, influence the optimal values of
R }t and 1 *. I he main e fleet of quantity discounts tin a periodic review
system, if any, is to increase 7’*. We shall illustrate how to account for
quantity discounts in an (It, T) model.
Let us study the backorders case under the assumption that the number
of units demanded in any time period is Poisson distributed. Ltd- (\x)
be the cost of the units when x are procured. The probability that, x units
are procured at any given review is the probability that x units wore de¬
manded in the past period, i.e., p(x; k'l). 'I hus the expected cost, of the
units procured at any given review is

J2 C(x)p(x; XT)
XmQ

and the average annual cost is

M = i ]C <"Wp(.r; XT) (5-151)

1 X 0

Let us now evaluate this expression explicitly for the case of all
units quantity discounts. Assume that, there are m price breaks
?ij • • ■ > $>»(<?o = 0, jm+i = °o), and that the unit. cost, is (\ if the quantify
ordered x satisfies q{ < x < <•/,„. Then, the average annual cost of the
units procured is
 PROBLEMS 289

The expected unit cost V of the item is simply the average annual cost
divided by the mean rate of demand, i.e.,
7ft
P - £ - 1 ;XT)- P(qt - 1; XT)] + CJP(qm - 1; XT)
i— 1

It is this expected unit cost that must be used in the inventory carrying
cost term. Note that R* for a given T can be found without including the
M(T) term in the cost expression. However, the expected unit cost V will
now be a function of T. To determine T* it is necessary to include the
M(T) term in the cost expression.
The computational procedure in the lost sales case is precisely the same
as in the backorders case, provided that the influence of quantity discounts
on the cost of a lost sale can be ignored. This is usually a valid approxima¬
tion in practice.
We shall leave for Problem 5-29 the development of the average annual
cost of units procured when incremental quantity discounts are available.
The same general computational procedure applies in this case. However,
(5-152) must be replaced by an equation of somewhat different form.
It is straightforward but tedious to compute M and (J for nQ and Rr
operating doctrines also. We ask the reader to attempt this in Problem
5-30.

REFERENCES

1. Arrow, K. J., T. Harris, and J. Marschak, “Optimal Inventory Policy,”

Econometrica, XIX (1951), pp. 250-272.
2. Arrow, K. J., S. Karlin, and H. Scarf, Studies in the Mathematical Theory
of Inventory and Production. Stanford, Calif.: Stanford University
Press, 1958.
3. Hadley, G., and T. M. Whitin, “A Family of Inventory Models,”
Management Science, Vol. 7, No. 4, July 1961, pp. 351-371.
4. Morse, P. M., “Solutions of a Class of Discrete Time Inventory Prob¬
lems,” Operations Research, Vol. 7, No. 1, Jan.—Feb. 1959, pp. 67—78.

PROBLEMS

5-1 A warehouse has a policy of reviewing each item in its inventories

once every six months. It uses an order up to IB policy. One particular
item costs $25 per unit independently of the quantity ordered. The
 290 PERIODIC REVIEW MODELS WITH STOCHASTIC DEMANDS CHAP. 5

cost of placing an order is $2.00 and the cost of the review apportioned
to this item is $15. Demands occurring when the system is out of
stock are backordered and the cost of a backorder is taken to be
$100. The system uses an inventory carrying charge of I = 0.20.
The number of units demanded in any time period is normally dis¬
tributed with mean 240* and variance 500*, where * is the length of
the period in years. The lead time is constant at 2.5 months. What
is the optimal value of R for the presently used review period? What
is the optimal review interval and R*(T*)7 What average annual
savings could be obtained by using T* as the time between reviews?
6-2 Derive the discrete analogue of the model developed in Sec. 5-2 for
the case where demand is Poisson distributed. Consider both the
case of constant and stochastic lead times. What is the inequality
which is used to determine R*(T). Consider both the backorders and
lost sales cases.
6-3 A certain item carried in a large military supply system is reviewed
every three months. An order up to R policy is used. The number of
units demanded in any time period can be considered to be Poisson
distributed with the average annual demand being 15. The procure¬
ment lead time is essentially constant with the value five months.
The unit cost of the item is $500, the cost of a backorder is taken to
be $5000, and the cost of placing an order is $30. The system uses an
inventory carrying charge of 7 = 0.20. Determine the optimal value
of R for the specified time between reviews. What is the safety stock?
What is the cost of uncertainty?
6-4 Under what conditions will the solution to Eq. (5-9) not be unique?
Hint The solution is unique if h(x; T) >0 for all x > 0.

6-6 For the model developed in Sec. 5-2 develop the derivatives needed
for Newton’s method to solve for R* and T* simultaneously.
6-6 Apply the results of Problem 5-5 to the example given in Sec. 5-2 and
determine R* and T* simultaneously.
6-7 Discuss how the method of steepest descents could be used to deter¬
mine R* and T* for the model developed in Sec. 5-2.
6-8 Use the method of steepest descents to determine R* and T* for the
example presented in Sec. 5-2.

6-9 Prove that the cost function of Eq. (5-6) is a convex function of R
for a given T. Under what conditions is it strictly convex? Is 3C a
convex function of R and 77

6-10 Discuss the behavior of the safety stock for the model developed in
Sec. 5-2 as 7 —> 0. How do you explain this?
 PROBLEMS 291

5-11 A department store reviews the stock of a particular white shirt each
week. An order up to R policy is used. The number of units
demanded in any time period for a particular size of the shirt can
be considered to be Poisson distributed with a mean of 25 per week.
Each shirt costs the store $3.00 and the selling price is $5.95. An
inventory carrying charge of I = 0.17 is used. Demands occurring
when the store is out of stock are lost, and the cost of a lost sale aside
from lost profit is $10.00. The procurement lead time can be assumed
to be constant at 10 days. Determine the optimal value of R.
5-12 Go through the same sort of analysis employed in Sec. 4-6 for the
(Q, r) model to obtain the various time averages for the {nQ) r, T)
and (R, r, T) models, and show their relation to expected values. In
particular, show that the average annual cost of carrying inventory,
stockouts, etc. is the expected cost per period times 1/T. Also show
that the average annual cost is the expected cost per cycle times the
average number of cycles per year.
5- 13 Derive Eqs. (5-27) through (5-29).
6- 14 For the (nQ, r, T) model compute the steady state probability $(x)
that the inventory position at a review time is x (before the place¬
ment of any order). Do this by making use of the p(r + j). Also
develop the set of equations from which the \f/(x) can be computed
directly by considering the related Markov process for generating
transitions between the inventory position states at the beginning of
a review. Can you solve these equations directly?
5- 15 Compute the mean and variance of the density \j/(x) defined in
Problem 5-14.
6- 16 For the continuous version of the (I?, T) model presented in Sec. 5-6
determine the derivatives that would be needed if Newton's method
were to be used to determine R* and T* simultaneously.
6-17 Is the cost expression (5-74) a convex function of R for a given T?
Is it a convex function of R and T then. What are the implications
of this result for the determination of and consequences of R* and T*?
6-18 Carry through the limiting processes of Sec. 5-7 for the case of sto¬
chastic lead times. What approximations and/or assumptions are
needed?
6-19 Provide a step-by-step derivation of Eqs. (5-88), (5-89).
5-20 For a given value of r and R, develop a numerical procedure for
solving the set of equations (5-88), (5-89).
5-21 For a periodic review system determine an explicit form for the aver¬
age annual cost expression in the backorders case when an Rr policy
292 PERIODIC REVIEW MODELS WITH STOCHASTIC DEMANDS CHAP. 5

is used and the demand in the time between reviews has an exponen¬
tial distribution. Assume that the lead time is precisely one period
in length, x = 0, and that the inventory carrying cost for a period
depends only on the on hand inventory at the end of the period.
Determine R* and r*.
6-22 In Eq. (5-108) work out explicitly H(x, T) and v(n)(x; T) when v(x; T)
is a normal distribution.
6-23 Show that Eqs. (5-111), (5-112) are solutions to Eqs. (5-109), (5-110).
Also show that with the proper interpretation, p{R) and p(r + y)
represent a legitimate density function.
6-24 Derive in detail Eqs. (5-113), (5-114).
6-25 Evaluate explicitly the /3(z, y) in Eq. (5-141).
6-26 Develop a numerical procedure for solving the set of equations (5-132)
and (5-141) for a given value of R.
6-27 Work out the explicit form for Eqs. (5-149), (5-150) when the lead
time is gamma distributed.
6-28 Discuss the introduction of stochastic lead times into (R, r, T) models.
6-29 Develop the equations needed to handle incremental quantity dis¬
counts in order up to R models when demand is Poisson distributed.
5- 30 Study the problem of handling quantity discounts in (nQ, r, T) and
{R, r, T) models under the assumption that an nQ or Rr policy is to be
used.

6- 31 Show that the equations for quantity discounts developed in Sec. 5-15
reduce to the correct values when the unit cost of the item is a con¬
stant independent of the quantity ordered.
5-32 Obtain 6(x) and p(r + y), p(R) by solving Eqs. (5-113), (5-114), in the
case where v(x; T) is a gamma distribution of order two. Also obtain
p(r ”t~ y) and p(R) from (5-111) and (5-112). Hint: Use the proper¬
ties of v(x; T) to obtain second order differential equations for 6(x).
These equations are

6" — 2/36' + /320 = 0, x < r; 8" — 2/30' = 0, r < x < R

Determine the four constants of integration by noting that

6(R) = 0; 8(x), d’(x) are continuous at x = r; and 8(x) must be a
legitimate density function. The resulting expression for p(r + y)
and p(R) are

p(r + y) = 2/3A[l - 0 < y < R - r

p{R) = 4A, A = [3 + 2/3(R - r) + e—0]—i

 PROBLEMS 293

5-33 Suppose that instead of specifying a stockout cost, it is desired to

minimize the costs of review, ordering, and carrying inventory subject
to the constraint that the average fraction of time out of stock is not
greater than /. For each type of operating doctrine discussed in this
chapter, show how such a criterion can be used to determine the
optimal values of the parameters for the specific operating doctrine
under consideration. Show that this really imputes a value of 0 to x
and a unique value to 7r. Also discuss the case where it is desired that
an upper limit to the expected number of backorders at any point in
time be specified.
5-34 Solve Problem 5-1 under the assumption that in addition to the fixed
cost of $100 per backorder, there is also a variable backorder cost of
$2000 per unit per year.
5-35 Solve Problem 5-3 under the assumption that in addition to the fixed
cost of $5000 per backorder, there is a variable cost of $50,000 per
unit per year.
5-36 Determine T* for Problem 5-1 if all units quantity discounts are
available of the following type: the unit cost is $25 if the quantity
ordered Q lies in the range 0 < Q < 100, and is $20 if Q > 100.
5-37 Determine T* for Problem 5-3 if all units quantity discounts are
available of the following sort: the unit cost is $500 if the quantity
ordered is less than 10 and is $300 if the quantity ordered is greater
than 10. Assume that the cost of making a review is $50.
5-38 Discuss how to compute the state probability *p(x) that the inventory
position of a periodic review system is x at any randomly chosen
instant of time t. Carry out the computation to determine yp{x) when
an order up to R policy is used and demand is Poisson distributed.
Hint: Show how to compute the probability that the inventory
position is x at a time £ after a review. Then average £ over the period.
Can you use these state probabilities to obtain the same E and B
terms obtained in the text?
5-39 Consider a low demand, high cost item carried in a military supply
system which has the following characteristics. The number of units
demanded in any time period can be considered to be Poisson dis¬
tributed with a mean rate of demand of 4 per year. The unit cost of
the item is $1000 independently of the quantity ordered. All demands
occurring when the system is out of stock are backordered and the
cost of a backorder is taken to be $30,000 independently of the length
of time for which the backorder exists. The procurement lead time
shows quite a bit of variability. Its mean is 6 months and it has a
standard deviation of 3 months. It is felt that the distribution of lead
294 PERIODIC REVIEW MODELS WITH STOCHASTIC DEMANDS CHAP. 5

time demand can be represented with sufficient accuracy by a gamma

distribution. The system uses an inventory carrying charge of I =
0.20. All items in inventory are reviewed every six months and an
order up to R policy is used. Determine the optimal value of R for the
part under consideration.
5-40 Solve Problem 5-39 under the assumption that it was not possible
for the military to specify a backorder cost, but they did feel that the
average fraction of time out of stock should not be greater than 0.001.
What is the imputed backorder cost?
5-41 Solve Problem 5-1 under the assumption that instead of specifying
the cost of a backorder, it is specified that the average fraction of
time out of stock should be 0.01.
5-42 Solve Problem 5-3 under the assumption that instead of specifying
the cost of a backorder, it is specified that the average fraction of time
out of stock should be 0.001.
5-43 Solve Problem 5-11 under the assumption that instead of specifying
the cost of a lost sale, it is specified that the average fraction of time
out of stock should be 0.05.
5-44 Discuss the problems involved in handling non-linear stockout costs
in periodic review systems. Consider a system which uses an order
up to R policy. Let w(t) be the cost of a backorder, if it lasts for a
time t. If demand is Poisson distributed, can you develop the ex¬
pression for the average annual cost of backorders?
5-45 Use the notion of an ensemble average to justify the fact that the
integral of the expected on hand inventory or backorders is the
expected value of the integral.
5-46 In a military supply system, a policy is followed of reviewing a certain
critical item weekly. An nQ ordering policy is to be used. The de¬
mand for the item can be imagined to be Poisson, with the average
annual rate being 50 units. The procurement lead time is essentially
constant at 2 months. The unit cost of the item is $500 independently
of the quantity ordered, and the system uses an inventory carrying
charge of I — 0.20. The cost of placing an order is estimated to be
$200. Any demands occurring when the system is out of stock are
backordered, and the cost of a backorder is placed at $10,000. Deter¬
mine Q* and r*.

5-47 For the simple model discussed in Sec. 5-2, show how to include an
inventory carrying cost which is based on the maximum inventory
level. How does this term change the equation which determines J?*?
 PROBLEMS 295

5-48 Assume that an inventory system stocks n items. Imagine that an

order up to R policy is used for each item. Also imagine that the
time between reviews must be the same for each item. Show how to
determine T*.
5-49 Show what sorts of problems are encountered in introducing floor
space constraints, or constraints on investment in inventory when n
items are carried by a system and a periodic review policy is used for
each of them.
5-50 Derive the average annual cost for an (nQ, r3 T) model by first com¬
puting the expected cost per cycle and then multiplying by the aver¬
age number of cycles per year.
5-51 Determine the limiting form of the (R, r, T) model with Poisson
demands and constant lead times as T —» 0. Is the limiting form the
familiar (Q, r) model?
SINGLE

PERIOD MODELS

“Deliberate often; decide once.”

Publilius Syrus
6-1 Introduction

The models to be discussed in this chapter are perhaps the simplest of all
inventory models in which demand is treated as a stochastic variable. The
essential characteristic of these models is that only a single time period,
usually of finite length, is relevant and only a single procurement is made.
A rather wide variety of real world inventory problems including the
stocking of spare parts, perishable items, style goods, and special season
items offer practical examples of the sort of situations to be studied here.
Since only a single time period is to be considered, there is no steady state
associated with the models analyzed in this chapter. These models then
provide a transition between the steady state models of Chapters 2, 4, 5
and the dynamic models to be presented in the next chapter.

6-2 The General Single Period Model with

Time Independent Costs

The type of problem to be studied in this section is often referred to as the

Christmas tree problem or newsboy problem, since it can be phrased as a
problem of deciding how many trees a dealer in Christmas trees should
purchase for the season, or how many newspapers a boy should buy on a
given day for his corner newsstand.
Let us illustrate by formulating the Christmas tree problem. Assume
that a Christmas tree vendor must order his trees a month before the
season begins, and that he has no opportunity to reorder later in the season
if he needs more trees. \ Thus in a given season he has only one opportunity
to place an order. The cost of a tree is Cand the selling price is S. Any
trees not sold at the end of the season are a fotaTIoss. Let p(x) be the
probability that x trees will be demanded during the season.* Then his
expected profit glor the season if he procures h trees is

* We here ignore the fact that in reality he will stock trees of different sizes and costs,
and that the demand variables for trees of different sizes will not be independent, since
if he is out of one size a customer will often accept a different size.
297
 298 SINGLE PERIOD MODELS CHAP. 6

Q(h) = S xp(x) + Sh p(x) - Ch (6-1)

x=0 x = h+l

since the revenue received is Sx if x < h, and is Sh if x > h. The problem

is to determine the value of h which maximizes his expected profit.
The Christmas tree problem is a special case of a more general model
which can be formulated as follows: An item can be procured only at the
beginning of a period. The unit cost of the item is C, independent of the
number of units procured. The selling price per unit is S. Let p(x) be the
probability that x units are demanded during the period. If a demand
occurs when the system is out of stockJ.,a_ goodwill cost ir0 is incurred in
addition to the lost profit. Any units remaining at the encTof' tfre- period
can be sold at a price of L dollars per unit {L < C). It is desired to deter¬
mine the optimal number of units to have on hand at the start of the
period in order to maximize the expected profit for the period.
If h units are procured, the expected profit is
h "f oo

xp{x) + Sh Z p(x)
x=0 x = A-f 1

h-1

+ L E & - x) p(x) - Ch - Ti-o Yh (x ~ h) P(x)

*=0 x=A

= (S — L)fi — (C — L)h — (s + to — L) (x ~ h) p{x) (6-2)

x=h »

where p, is the expected demand for the period. The smallest h which max¬
imizes the expected profit is the largest h for which Ag(A) > 0. However

Ag(A) = (S + 7T0 — L) PQi) + L — C (6-3)

where P(x) is the complementary cumulative of p(x). Thus the largest h

such that

v>ru\ ^ ^ C — L
W>S + To-L- (S -C) + X„ + (C - L) M

is optimal, i.e., is h*. Note that C — L is the loss per unit on any units re¬
maining at the end of the period, S - C is the unit profit, and (S - C) + x0
is the cost incurred if a demand occurs when the system is out of stock.
Note also that no restrictions whatever are placed on the demand distribu¬
tion for the period p(x). For example, units need not be demanded one at
a time, and demand in non-overlapping time intervals need not be inde¬
pendent. If AS(fi* + 1) = 0, both h* and h* + 1 are optimal.
Often 4 is convenient to treat h and the demand variable as continuous.
SEC 6-3 EXAMPLES 299

Then if f(x) is the density function for demand in the period and F(x) is
its complementary cumulative, the expected profit for the period when h
units are procured is

2(h) = S j* xf(x) dx + Sh f(x) dx

+L J (h — x) f(x) dx — Ch — to J (x — h) f(x) dx

= OS - L)m L)h .— (S + 7Tq L) L (x - h) f(x) dx (6-5)

The optimal h is then a solution to dg/dh = 0, i.e., h* satisfies the equation

m s^~tL
=

In Problem 6-4 we ask the reader to show that (6-5) is a concave function
of h. Furthermore, 2(h) is usually strictly concave. These results imply
that any relative maximum of g is the absolute maximum. When g is
strictly concave, then the absolute maximum is unique. Problem 6-5 asks
for a discussion of the conditions under which the solution to (6-6) is not
unique.

6-3 Examples

Several examples will now be given to illustrate the wide variety of prob¬
lems to which the simple model developed in the previous section is
applicable.
1. A large supermarket must decide how much bread to purchase each
day. For shopping days Monday through Thursday, past history has
shown that the daily demand can be considered to be normally distributed
with mean 300 and standard deviation 50. A loaf of bread sells for $0.25
and costs the store $0.19. Any bread not sold by the end of the day is, on
the following day, placed on a counter where it is sold for $0.15 per loaf.
All bread on this counter can be sold at this price. It is desired to deter¬
mine the optimal number of loaves to purchase to maximize the expected
daily profit.
In the notation of the previous section, S = $0.25, C = $0.19, L = $0.15,
x0 = 0. Then h* is the solution to
 300 SINGLE PERIOD MODELS CHAP. 6

Hence from the normal tables

h - 300
0.253 or h = 300 + 12.65 « 313
50

Thus 313 loaves should be purchased.

The expected number of loaves per day of bread to be placed on the
stale bread counter when h loaves are stocked is

lj\h - = h - M+ 0, - A) $(^) + ^(~^)

Then since h* = 313 and pL — 300, the expected number of loaves to be

placed on the stale bread counter each day is

13 - 13(0.40) + 50(0.3863) = 13 + 14.1 = 27.1

It is of interest to compare the expected daily profit when h* loaves

are stocked with that in the case where the number of loaves ordered is equal
to p, the expected demand. When f(x) is a normal distribution, (6-5) can
be written

s(h) = (S - L) M - (C - L) h + (S - L + to) (h - M) $ (^fi)

— <r(S — L + 7T0) 4> (^ ■"-■**)

, Then for the case at hand where tto = 0

g(h*) = 30.00 - 12.52 + 0.52 - 1.93 = $16.07

and
g(fx) = 30.00 - 12.00 + 0 - 1.99 = $16.01
Thus the increase in expected profit by using h* rather than h — p, is only
$0.06 per day, i.e., about a J percent increase in expected profit.
2. Let us now solve the above problem in the case where the super¬
market also attaches a goodwill loss of xo = $0.50 to every loaf of bread
demanded when the store is out of stock. In this case the optimal h is the
solution to
(h - 300\ 0.04 0.04
$
\ 50 ) 0.10 + 0.50 0.60
0.06667
Hence
h - 300
1.500 or h = 300 + 75 = 375
50

Now the expected number of loaves which will be placed on the stale bread
SEC 6-3 EXAMPLES 301

counter is 76.48. The expected daily profits using h* and /z are respectively

g(A*) = 30.00 - 15.00 + 3.00 - 3.88 = $14.12

800 = 30.00 - 12.00 + 0 - 11.97 = $6.03

In this case there is a very large difference between the expected profit
when A* is used and the expected profit when is used. It should be
observed that in this example, the expected daily profit is not what would
be obtained by averaging the profit obtained from the accounting records
over a long time, since it includes the influence of the goodwill loss wrhich
would not appear in the accounting records. Indeed, on the basis of what
would be taken from the accounting records, the expected profit wrould be
lower using A* = 375 than using h = 313 or h = m- The reader should
check this by computing the expected profit with 7r0 = 0 when A* = 375.
3. A fashionable candy store does not make its own chocolates, but
instead buys from a large producer which caters to a number of different
stores. The store under consideration must decide how many large choco¬
late rabbits to order for the Easter season. They must be ordered two
months in advance and there is no possibility of placing a reorder. Each
rabbit costs the store $2.50 and sells for $7.00. Any rabbits not sold at the
end of the season are a total loss. The store feels sure that it can sell at
least 100 of these rabbits and not more than 500. Any number between
100 and 500 is felt to be equally likely, i.e., the store feels that the dis¬
tribution of demand is uniform over the interval 100 to 500. If the demand
variable is treated as continuous, this means that the density function
f(x) is
fix) = 100 < x < 500; fix) = 0, otherwise

The complementary cumulative function F(x) of fix) is

fl, 0 < x < 100

Fix) = J 1 - Mx - 100), 100 < x < 500
[0, X > 500

From the data given above, S = $7.00, C = $2.50, L = 0, wo = 0.

Thus (6-6) becomes

m - 1 - J55 c* - 100) - H - 0.3571

or
h* = 357.2 « 357
The expected number of rabbits remaining at the end of the season when
h (100 < h < 500) rabbits are stocked is
302 SINGLE PERIOD MODELS CHAP. 6

J (h — x) /(x) dx = h — fi+ J xj(x) dx — hF(h)

= h - M ^ [A2 - 1000 A + 25 X 10*]

Then, since h* = 357 and jjl = 300, the expected number of rabbits re¬
maining when h* are stocked is

57 + 8^0 ['12‘75 X 104 “ 35‘7 X 104 + 25 X 1Q4J = 82.6

The expected profit on the rabbits if A (100 < A < 500) units are stocked
is

8(A) = (S — L) p. — (C — L) h

- ^ (S - L + tto) [A2 - 1000 A + 25 X 10*]

Then
S(A*) = 2100.00 - 892.50 - 178.93 = 81028.57
and
8(a0 = 1350.00 - 350.00 = 81000.00

Thus by carrying 57 rabbits above the mean demand, the expected profit
is increased by $28.57.
4. Certain spare parts for aircraft are made at the same time the aircraft
is produced. Once production on the aircraft has ceased, it is extremely
difficult to obtain these spare parts. Production is about to cease on a
certain military aircraft. It is desired to determine how many spares of a
certain low demand item should be available at the time production stops.
For the number of planes in the force, past data indicate that the demand
for the part will be Poisson distributed with the mean rate of demand
being 0.75 per year. It is uncertain precisely what the operational life of
the aircraft will be. However, it has been decided that the probability
density for the time until obsolescence can be described by a gamma dis¬
tribution with a mean of 6 years and a standard deviation of 1.5 years.
The parts cost $2000 each and their scrap value, if any remain unused at
the end of the plane's operational life, is $100 each. If a demand should
occur when no spares are available, the cost of obtaining one is estimated
to be $13,000.
In order to determine the optimal number of spares which should be on
hand, it is first necessary to determine the marginal distribution of demand
over the operational life of the aircraft. We know from (3-73) that this
distribution is the negative binomial a + 1, 0/(0 + X)], since de-
SEC 6-3 EXAMPLES 303

mands are Poisson distributed and the operational life has a gamma dis¬
tribution. To determine a, 0 we use (3-44), (3-45) which relate the mean
and standard deviation of the gamma distribution to a, p. Thus

Then, X = 0.75, /S/OS + X) = 0.781, a + 1 = 16, and

bN[x; a + 1 ;/3/G3 + X)] = (1^+J)! (0.781)16 (0.219)*

From the above data we see that C = $2000, 1/ = $100, x0 = $13,000

and $ = 0. Now maximization of profits with S = 0 is equivalent to
minimization of expected costs. Hence (6-4) applies in this case also. If
h is the quantity on hand when production ceases, the optimal value of h
is the largest h for which

00 1QOO
E bN(x; 16, 0.781) > = 0.1475

It is not too difficult to compute the optimal h, since once b&(x; 16, 0.781)
has been calculated for a given x, it is easy to compute successively values
for other x values using
Ifi 4- r
bN(x + 1; 16, 0.781) = (°*219) b*(x'>16> °*781)

Denoting the complementary cumulative by Bn(x), the cumulative func¬

tion by Bn(x), and noting that Bn(x) = 1 — Bn(x — 1), we find that
Bn{6) = 0.3061, Bn(7) = 0.1906, BN(8) = 0.1111, and BN(9) = 0.0609.
Thus it is clear that h* = 7.
If h units are on hand when production ceases, the expected number
still on hand when the aircraft becomes obsolete is

bNix; a+1, /3/03 + X)]

x=0

= h — fi + ^2 (x — h) bw[x; a. + 1, @/(]3 + X)]

x=h

Since the expected demand until obsolescence is n = 6(0.75) = 4.5, the

expected number to be scrapped if h* = 7 units are on hand when produc¬
tion ceases is

2.5 + 0.0502 + 2(0.0275) + 3(0.01421) + 4(0.00702) + ...» 2.7

304 SINGLE PERIOD MODELS CHAP. 6

6-4 Constrained Multiple Item Problems

Another class of single period inventory problems which frequently arise

in practice is concerned with the stocking of n different items when the
items are not completely independent but are related by some constraint
such as a budget or floor space constraint. The only interaction between
the items is assumed to be through the constraint. In particular, this
implies that the variables representing the demand for different items are
independent random variables. We shall discuss one typical problem
which is often referred to as a flyaway-kit problem [3]. A practical example
of such a problem is as follows: It is desired to stock a nuclear submarine
about to go on patrol with spare parts for the submarine itself and for the
missiles which it carries. The probability distributions for the demands
for the spare parts over the length of time of the patrol mission are assumed
to be known. It will be imagined that for each demand for part i which
occurs when no spares are available a cost 7rt- is incurred. The volume
available in the submarine for stocking spares is severely limited. The
problem then reduces to that of determining the quantity of each spare to
stock so as to minimize the expected stockout costs while not exceeding
the available space.
Suppose that there are n spare parts of interest. Let v% be the volume of
one unit of spare part i and V the total volume available for storage. The
probability that x units of spare part % will be demanded will be written
p%(x). Then, if hi units of part i are stocked, it is desired to determine
non-negative integers hi which minimize the expected cost

3C = TTi (x — hi) Pi(x) + . . . + 7Tn 'Yj (x “ hn) pn(x) (6-7)

x = hl X = hn

subject to the constraint that the total volume of the spares does not
exceed F, i.e.,
n

^2 Vihi = V\K\ + . . . + vnhn < V (6-8)

i= 1

Let us now consider how the h* may be determined. First, imagine that
each h* will be large enough that the hi can be treated as continuous.
Assume that p%(x) can be approximated by the continuous density fi(x),
whose complementary cumulative function is Fi(x). It is clear that when
we treat the hi as continuous, then the h* will satisfy (6-8) as a strict
equality. To solve the problem we can use the Lagrange multiplier
approach. We introduce a Lagrange multiplier 6 and form the function

W = (x — hi) fi(x) dx + 6 £ vA* - V (6-9)

.% = 1
SEC. 6-4 CONSTRAINED MULTIPLE ITEM PROBLEMS 305

Then if each h\ > 0, the M must be solutions to the set of equations

§r = 0 = -ViFiihi) + evit i=l,...,n (6-10)

ohi
or

— Fi(hi) = d, i = l,...,n (6-11)

Vi

A computational procedure is to select a d and compute the hi from (6-11).

Then compute V = £?-i Vih{. If V > V, select a larger value of 6 and re¬
peat the process. If V < V, select a smaller value of d and repeat. If V = V
the solution is optimal. One can make the computation quite readily by
hand if there are not too many items. The hi values so obtained will have
to be rounded to integral values. This should not lead to serious devia¬
tions from optimality when the hi are sufficiently large that the continuous
approximation is valid.
In many practical situations, however, the hi will tend to be very small
(i.e., 0, 1, or 2), and any attempt to use the above procedure and round the
results could lead to considerable deviations from optimality. An approxi¬
mate procedure for handling this case is to proceed as follows: Note first
of all that if we change the number of units of item i stocked from hi — 1
to hi, the expected reduction in cost is 7CiPi(hi), where P%{x) is the comple¬
mentary cumulative of Pi(x). The additional volume used in adding this
unit is Thus the expected cost reduction per unit increase in volume is
TnPi(hi) I Vi. The procedure is then to progressively assign units to the item
which yields the greatest reduction in expected cost per unit increase in
volume. The first step is to compute

If the maximum is taken on for i = j set hj = 1, and then compute

[Sp<(1)]- Sp,(2)}
The next unit is assigned to the index where the maximum is taken on, etc.
This is continued until adding an additional unit would exceed the volume
restriction. We ask the reader to explain in Problem 6-28 why this pro¬
cedure should often yield a satisfactory computational procedure and under
what conditions it might not be expected to work too well. The problem
of finding non-negative integers which maximize (6-7) subject to (6-8) can
be solved exactly by the technique of dynamic programming to be intro¬
duced in the next chapter. We shall illustrate the procedure there. It
might be pointed out that even dynamic programming cannot give the
 306 SINGLE PERIOD MODELS CHAP. 6

exact solution to the actual physical problem. The reason for this is that
specification of the unit volumes Vi and V is not sufficient. It is also
necessary to consider the actual shape of the units and of the space avail¬
able for storage. In other words it does not follow that simply because
(6-8) tells us that an additional 2 cubic feet of storage space are available
that an item having a volume of 1.75 cubic feet can be fitted into this
space. It depends on the shape of the available space and of the item.
In some cases there may be two constraints on the problem such as a
volume and weight constraint. The addition of one more constraint makes
the problem much more difficult to solve, but again the dynamic pro¬
gramming technique can be used.
Single period models involving a number of items connected by one or
more constraints can take a variety of forms. The following example
presents a situation which is slightly different from that given above, but
the method of analysis is just the same.

EXAMPLE The buyer for the housewares department of a Midwest retail

store is making his annual trip to Europe. In the real world he might be
purchasing as many as 100 or more different items. For the purpose of
this example, imagine that only three items are to be purchased. These
are: (1) copper skillets, (2) handcarved salt and pepper sets, and (3)
coffee makers. The copper skillets cost $6.50 each and retail for $15.00.
The buyer believes that the demand in the coming year for these skillets
will be normally distributed with a mean of 50 and a standard deviation
of 10. Any skillets unsold at the end of the year will be disposed of at
10 percent below cost. The salt and pepper sets cost $8.00 each and retail
for $25.00. The buyer feels that he will not sell less than 50 of these sets
or more than 200. Any number between these two values seems equally
likely. Sets not sold by the end of the year will be disposed of at 20 percent
below cost. The coffee makers cost $3.50 each and retail for $9.00. The
buyer believes the demand for this item is normally distributed with a
mean of 100 and a standard deviation of 15. Any units unsold at the end
of the year can be disposed of at cost. The buyer has an open to buy of
$1600 (i.e., he can spend up to $1600 on these three items). How many
units of each item should he procure?
If the buyer purchases hi units of item i his expected profit Qi(hi) will
have the form (6-5). He desires to maximize £ Q,(A*) subject to £ Cfc
< 1600, where C% is the unit cost of item i. Introducing a Lagrange
multiplier 0, we form the function
SEC. 6-4 CONSTRAINED MULTIPLE ITEM PROBLEMS 307

and the necessary conditions which the h* must satisfy (if they are all
positive) are
aw dQijhi)
= 0 = + 8C i
dhi dhi

or in the notation of Sec. 6-2

P \ Ci - U -dCi . - n
Fi(hi) = g' i = 1, 2, 3

Using the appropriate values for the individual items, we see that the
necessary conditions become

m , (hi - 50\ 0.65 - 6.500

Skillets: « =-gjl-

Salt and _JL_ ,90n __ h ^ 1.60 — 8.000

Pepper Sets: 150 ^ 2 18.60

Coffee As - 100\ = -3.500

Makers: \ 15 / 5.50

In this problem, it does not follow that it will be optimal to use the
entire budget. However, one sees immediately from the coffee maker
equation that if there was no budget restriction (i.e., 6 ~ 0) an infinite
number of coffee makers would be purchased (why?); Thus 0 must be
different from 0. In particular, it is clear that 0 must be negative^ Table
6-1 presents the results of the numerical computations. To the accuracy
with which rounding to integral values permits the computation to be
made, one might select either 8 = —0.99 with h± = 42, h2 = 123, hz = 95
or 0 = —0.98 with hi — 43, h2 = 124, hz = 95. There would be no practical
value in attempting to be more precise.

TABLE 6-1

Computational Results for Example

Cost of units
-8 hi h2 h (dollars)

0.1 61 181 123 2275

0.2 58 174 117 2178
1.0 42 123 95 1590
0.99 42 123 95 1590
0.98 43 124 95 1604
308 SINGLE PERIOD MODELS CHAP. 6

6-5 Single Period Models with Time Dependent Costs

The single period model introduced in Sec. 6-2 did not involve any costs
that were time dependent. We would now like to generalize that model to
include carrying costs which are proportional to the length of time that a
unit remains in inventory, and a stockout cost which is proportional to the
length of time from the point when the demand occurs until the end of the
period. As usual, the cost per unit year of keeping the item in stock will
be written IC, where C is the unit cost of the item. Furthermore, as usual,
7r will be the cost per unit year of a stockout. When including these costs,
one is forced to be somewhat less general than in Sec. 6-2, since it is neces¬
sary to introduce a distribution for the demand from the beginning of the
period to any time t in the period. The only case that can be worked out
easily when the variables are treated as discrete is that where demand is
Poisson distributed over any time period. This is the case we shall examine.
Consider first the case where the mean rate of demand X is constant over
time, and the length of the period is fixed and is not a random variable.
Then if h units are on hand at the beginning of the period, the expected on
hand inventory at time t is from A3-11
h
^2 (h — x)p(x-j \t) = h — \t + \tP(h; \t) — hP(h + l;\f) (6-12)
*=o

and the expected inventory carrying charges for the period are

ICDQi) = 1C J* [h-Xt + XtP(h; Xt) - hP(h + 1; XQ] dt

= IC |hT + P(h; XT) - 1} P(h + 2; XT)

- hTPQi + 1; XT) + h(-h + ^ P(h + 2; XT)}

= ICT^h - Y~ + yP(/i;XT) - hP(h + 1; XT")

+ ~' 2P(h + 2;XT)j (6-13)

by A3-18, A3-19.
If at time t the demand has exceeded h by y, then between t and t + dt
the variable stockout cost incurred is Tty dt. The expected value of y at
time t is by A3-10, A3-13

^2 (x — h)p(x; \t) = \tP(h; \t) — hP(h + 1; \t)

x=h
SEC. 6-5 MODELS WITH TIME DEPENDENT COSTS 309

and the expected time dependent stockout cost for the period is ttB(K)
where

B(h) = J* [XtP(h; Xt) - hP(h + 1; Xt)} dt

= P(h; XT) - hTP(h + 1; XT) + h^ ' P(A + 2; XT) (6-14)

Thus the cost terms which must be added to (6-2) are

ICT [a - y] + (IC + *W) (6-15)

Written out explicitly the expected profit becomes

8(A) = OS — L)n - (C - L - ICT)h - | ZCXT2 - n [s - L + tto

- | (* + ICO] P(A; XT) + h[S — L + vo (6-16)

- T(# + JC)]P(A + 1; XT) + IC^ h(h + 1)P(A + 2; XT)

where ju = XT. Then

Ag(h) = — (C *— L — ICT) + — L + ttq

+ (f - r) (# + zo] P(A; XT) - £ (* + IC)P(A; XT) (6-17)

The largest h for which Ag(A) is positive will be optimal. It is not so

simple to determine the optimal h using (6-17) as it is to use (6-4). How¬
ever, it only involves a straightforward tabulation and can usually be
carried out quite quickly.
The reader should be careful not to misinterpret the stockout cost
$B(h). It is not normally a backorder cost incurred while waiting for the
arrival of the next procurement, since, usually, demands occurring when the
system is out of stock are not backordered, or if they are, the backorder
cost will not be time dependent. This cost usually arises in a different way.
For example, the item may be a spare part for a machine which is used to
turn out some product. Then tt will be the rate at which revenue is lost if
a machine must halt operation for lack of the spare part.
Often in situations where the type of model under discussion would be
useful, the date of obsolescence is not known with certainty. It must be
described by a probability distribution. For example, one might assume
that the time to obsolescence has a gamma distribution. Then in place of
 310 SINGLE PERIOD MODELS CHAP. 6

the Poisson terms in (6-16) the negative binomial distribution will appear.
We leave for Problem 6-13 the derivation of the equivalent of (6-17)
when the length of the period is gamma distributed. Frequently, it is
difficult to estimate reliably what the nature of a continuous density
function for the length of the period should look like. About, the best that
can be done is to select n times T, at which the period may end, and assign
probabilities wj that the period will end at times Tj. Then if t^ft, Tj) is the
expected profit when h units are on hand at the beginning'of the. period
and the length of the period is Th the expected cost- (.averaged over the
possible period lengths) is

m = E u$(h> ?v) (6-i8)

J t

It is desired to determine the h which maximizes <;(/)). This h is the

largest h for which
»
Tj) (0-19)
y-i

is positive. Kadi 7;) has the farm (0-17) (with T replaced by Tj).
Of course, the determination of h* Ls now more complicated than in the
determination of h* using (1M7).
One final geneialization ol the model is ot ten needed. The mean rate of
demand will now be allowed to vary with time. Imagine a curve Ls avail¬
able giving the cumulative moan demand IHt) irom the beginning of the
period until time t. It will be assumed that the eumillative demand up to
time t is Poisson distributed with mean Then write hit) - t\{t).
If the rate oi demand remained constant at the value. XU) from the begin¬
ning of the period until time /, than there would be the mum* probability
of having x units demanded by time4 / as in the ease when the mean rate of
demand changes with time and the expected mean demand up to time, t Ls
D(t).
Let us first consider the ease where the mean rate of demand ean vary
with time, but the length of the period is fixed at the value T. To compute
the optimal quantity to have on hand at the beginning of the period, it is
convenient to subdivide the period into m subperiods, subperiod i extending
from time to U (4 (), tm =- 7*). Within each subperiod / the mean
rate of demand will be assumed constant at the value A, - Aitt). By
choosing U — 1 sufficiently small, the approximation made by taking the

rate of demand to be constant within each subperiod, will come arbitrarily

close to representing the actual time pattern of demands.
Ihen the expected cost oi holding inventory, and the expected time
variable stockout costs in subperiod i are, respectively
SEC. 6-5 MODELS WITH TIME DEPENDENT COSTS 311

IC jH' [h - + \itP{h; \it) - hP(h + 1; Xii)] dt = ICU [h -

- icti-t [h - + ic[a(h, u, x<) - a(k, u_l7 xo] (6-20)

and
u, Xi) — a(h, ti-1, X*)] (6-21)
where

a(h, i, X) = Y P(A; Xi) - AiP(A + 1; Xi) + P(A + 2; Xi) (6-22)

These results follow from (6-13) and (6-14) on breaking the integral from
U-i to U into the integral from 0 to U minus the integral from 0 to U~i.
The expected profit for the period is then
1 m
8(A) = (-S - i)M - (C - L - ICT)h X<(i? - <?-i)
1=1

- ix(S- L + iro)P(h-, XT) + h(S-L + TO)P(h + 1; XT) (6-23)

m
+ (x + IC) [G>(/t, h X%) — dQi} U—i, Xi)]
t=i

The optimal h is then the largest h which yields a positive value for the
following expression for Ag(ft).

Ag(A) = — ((7 — L — ICT) + (S -L + irQ)P(h; XT)

m -j

+ (x + IC)h V f [P(/i; X,-i») - P(A; Xtfi-0]

(6-24)
m
- (t + IC)J2 [<iP(A; Xifi) - ii_iP(A; Xiii-0]
1=1

m ^
- (# + IC)h V 7- [p(X; Xjii) — p(A; X,-i,-_i)]
i=i Xi

Again the optimal h can be found by tabulation, although the task could
be quite arduous to carry out by hand. Several of the expressions in (6-24)
such as P(h; X,i») — P(h; XA-i) involve the difference of two terms which
are almost equal. In Problem 6-14 we ask the reader to show how Taylor's
expansion can be used to simplify such terms and thus yield a considerably
simpler expression for Ag(h).
A variable mean rate of demand can be combined with a situation in
which the length of the period is a discrete random variable having n
 312 SINGLE PERIOD MODELS CHAP. 6

possible values Th with the probability of T} being Uj. For each Ts there
will in general be a different function describing the mean rate of demand
over time. We can then obtain expressions like (6-23), (6-24) for each possi¬
ble Tj. Let g(h, Tj) be the expected profit if the length of the period is Tj
To obtain g(A, Tj) from (6-23), T is replaced by Th X, by Xiy, and m by m,
(for each j we may wish to use a different number of subperiods). Then
n

SW = X) u&(h> Ti) (6-25)

3 = 1
and

A2(h) = ^2 wiA<o(h, Tj) (6-26)

j=l

The task of determining the optimal h from (6-26) will usually involve too
much work to be done manually. However, it can be done quite easily on
a digital computer. It is not easy to allow the mean rate of demand to be
time dependent if the length of the period is described by a continuous
density function. We ask the reader in Problem 6-15 to explain why this
Models of the type presented in this section have been discussed in
[4, 5],

6-6 Marginal Analysis

S S!t6"2 We determilled ASW by first computing g(h) and then finding

the difference g(A) - Q(h - 1). In economic terms Ag(h) is the change
m expected profits if we change the number of units on hand at the begin¬
ning of the period from h — 1 to h. It will be profitable to add the extra
unit so long as A2(h) > 0. We now wish to point out that it is easy to com-
ao/ja- k directly wlthout first computing the expected cost. Note that
A2(h) is the expected change in revenue from sales minus the expected
change m cost plus the expected change in income from liquidation sale
minus the expected change in stockout costs.
Now the expected change in revenue from sales is 8 if the additional
unit is sold and 0 otherwise. The probability that the unit will be sold is
the probability that the demand will be > h, i.e., P(h). Thus the expected
increase in revenue is SP(h). The change in cost is simply C, the cost of
the unit. The increase in revenue on liquidation will be L if the additional
unit is not sold, and 0 otherwise. The probability that the unit will not
be sold is [i -P(h)] and the expected gain of liquidation revenue is
L ( )]• Finally, note that on adding an additional unit, expected
stockout costs must be reduced or left unchanged so that the change in
expected stockout cost must be non-positive. The change will be -x0 if
SEC 6-6 MARGINAL ANALYSIS 313

the additional unit is demanded (i.e., the stockout cost is avoided on this
unit) and 0 otherwise. Hence the expected change in stockout cost is
— 7t0jP(/i) . Therefore

AS(h) = SP(h) - C + L[1 - P(h)] - (-7rQ)P(h)

= (s + iro - L)P(h) - (C - L) (6-27)

which is precisely (6-3). We have thus obtained &Q(h) without ever having
written down the expected cost.
What we have introduced above is, of course, nothing but the marginal
analysis so familiar to economists. Marginal analysis is a very useful tool
for deriving optimizing conditions, and is applicable to a very wide
variety of problems. In particular, marginal analysis can be used to obtain
the optimizing conditions for most of the models discussed in the previous
sections. The essential feature of marginal analysis is that one examines
what happens when a unit change is made in the variable under consider¬
ation. As another simple example, consider the order up to R model for
the backorders case discussed in Sec. 5-2. Observe that A3C(i?) is equal to
the expected change in carrying costs plus the expected change in backorder
costs on changing from R — 1 to R. Now when the time over which
backorders exist is negligible, then adding an additional unit implies that
it will be carried in inventory essentially all the time, and the expected
change in annual carrying costs will be IC. The change per period in the
backorder cost is — x if the additional unit is demanded in the period
and 0 otherwise. Thus, per period, the change in expected backorder cost
is —irH(R; T), and per year is —irH(R; T)/T. Therefore

A3C(iJ) = IC -j H(R; T) (6-28)

which is the same as (5-7).

Unfortunately, it is not always so easy to carry out the marginal analysis
as it was in the above two simple examples. The analysis is usually fairly
straightforward as long as there are no time dependent costs. It may be¬
come more complicated, however, when time dependent costs must be
included. The inventory carrying cost in the above example was a time
dependent cost. The analysis was made simple because of the assumption
that the unit remained in stock essentially all the time.
The simple deterministic model of Sec. 2-4 is also easy to analyze by
marginal analysis. It illustrates, however, a case where it is difficult to
proceed without having first evaluated the cost terms. Note that A3C(Q)
is the sum of the change in ordering costs and the change in carrying costs
on going from Q — 1 to Q. The average annual change in the carrying
cost is simply

TRE ROUT IIM7

tmmt INSTITUTE OF TESMOUi?
 314 SINGLE PERIOD MODELS CHAP. 6

l ~ — A ,4
XA _Q Q- i_ Q{Q - l)

and, since the average inventory is (Q - l)/2, the average annual change
in carrying costs is /C/2. Summation of these two terms yields the proper
expression for A3C(Q). In this case, however, marginal analysis did little
to simplify the analysis, since one essentially had to have the cos), terms
to difference them. In essence, one could not. analyze the effects of
adding one unit without considering its effects on the costs associated with
the other Q — 1 units.
The computation of ArIl{Q, r) for the model of Sec. -1-7 by marginal
analysis provides an illustration of a. ease when' it is easy to handle a time
dependent cost. Consider ArIi((J, ;•). If the inventory position of the sys¬
tem is j units above the reorder point at time / r, then in going from
r - 1 to r, the expected number of backorders at time t will be reduced
by one if the extra unit is demanded and will be unchanged otherwise.
The probability that the extra unit is demanded is /'(/• ( j; Xr). Thus

ArB(Q, r) ==• ^ P(r -f j; Xr) (6-29)

^'•'1

This expression can be .simplified using the Poisson properties. The direct
computation of AqB{Q, r) by marginal analysis is more* tricky and the
reader is asked to try it in Problem h-10.
Marginal analysis is olten a helplul way to look at a problem. In some
cases it (‘an simplify considerably the amount of work required to obtain
the optimization conditions. In other eases the analysis is so complicated
that, no groat .savings in time results. In principle, however, marginal
analysis could be. used to obtain the optimization equations for all the
models developed thus far in this book. Of course, often if. is desired to
have an explicit expression for the profit or cost, and in this ease, the
potential savings in labor by use of marginal analysis to avoid computing
the profit or cost expression am not realized. In such eases, however,
marginal analysis can be helpful in cheeking the result obtained by directly
dineienc.ing ^1<f cas(l or I,r<>fit expression. Infinitesimal marginal analysis
can be used when the variables are treated as continuous. Here one con¬
siders the change on going from say .r to x 1 tlx. Some examples of
infinitesimal marginal analysis are considered in the problems.

REFERENCES

1. Bowman, lb lb, and It. B. better, Anal^i.sfar / War//,,,, Manama,l,

Kev. Ed., Homewood, Illinois: Richard 1). Irwin, hie., HltH.
 PROBLEMS 315

Chapter 10 is devoted to incremental analysis. The procedure given

is not rigorous when the variables are discrete, since the authors
assume that the optimizing condition has the marginal profit equal to
zero, whereas it is not necessarily true that it can be zero when the
variables are discrete. This point is not of great importance for the
types of problems which they consider, however.
2. Fetter, R. B., and W. C. Dalleck, Decision Models for Inventory Manage¬
ment. Homewood, Illinois: Richard D. Irwin, Inc., 1961.
3. Geisler, M. A., and H. W. Karr, “A Fruitful Application of Static
Marginal Analysis/' Management Science, Yol. 2, No. 4, July, 1956,
pp. 313-326.
4. Hadley, G., and T. M. Whitin, “An Optimal Final Inventory Model,"
Management Science, Yol. 7, No. 2, January, 1961, pp. 179-183.
5. Hadley, G., “Generalizations of the Optimal Final Inventory Model,"
Management Science, Yol. 8, No. 4, July, 1962, pp. 454-457.

PROBLEMS

6-1. A dealer in Christmas trees is attempting to determine how many to

stock. Past experience indicates that demand will be normally dis¬
tributed with a mean of 200 and a variance of 300. The dealer sells
a tree for $7.50 and it costs him $4.00. Any trees not sold are a com¬
plete loss. How many trees should he procure to maximize his
expected profit? What is his expected profit if he purchased a number
of trees equal to the expected demand? What is the expected number
of trees that will remain unsold if he procures the optimal number of
trees?
6-2. Solve Problem 6-1 under the assumption that the distribution of
demand is uniform with the same mean and variance given in
Problem 6-1.
6-3. Consider a low demand item which is to be produced only once. This
item is to be used in a system which will be obsolete in a number of
years. The date of obsolescence is not known with certainty, but
there have been assigned probabilities 0.5, 0.3, 0.2 that it will become
obsolete at the end of the fifth, sixth, and seventh years respectively.
The number of units demanded in any time period is expected to be
Poisson distributed with a mean rate of 1 per year. The item costs
$50.00 and has no scrap value. All items not used when the system
becomes obsolete will be scrapped. The cost of each stockout is
316 SINGLE PERIOD MODELS CHAP. 6

$3000. How many items should be stocked to minimize the expected

costs? What is the expected number of items that will be scrapped?
6-4. Show that Eq. (6-5) is a concave function of x. A function f(x) is
said to be concave if/[aa* + (1 - a)x2] > af(xi) + (1 - a)f(x2) for
all a, 0 < a < 1, and all xh x2 in the relevant interval of interest.
Note that if /(x) is concave, -/(:r) is convex (see Problems 2-6, 2-7).
We say that f(x) is strictly concave if —f(x) is strictly convex.
Under what circumstances is g(h) strictly concave? Prove that any
relative minimum of Q(h) is also the absolute minimum. Also prove
when g(h) is strictly concave, then the absolute minimum is unique.
6-5. Under what circumstances is the solution to Eq. (6-6) not unique?
Illustrate this geometrically.

6-6. An international oil company is buying a large tanker for mcfving

crude from the Middle East to the United States. Spare parts such
as rudders must be made when the ship is built. They are very
expensive to obtain later. Consider one particular part of the steering
gear. This part costs $5000 each when procured at the time the ship
is constructed and $25,000 each if procured later. Past experience
on other ships has indicated that the number of units demanded in
any time period will be Poisson distributed with a mean of 0.2 per
year. The useful life of the ship cannot be determined precisely
ahead of time, but it is expected to be gamma distributed with a
mean of 20 years and a standard deviation of 10 years. From the
available information, how many spares of the item under consider¬
ation should be purchased when the ship is built?

6-7. Show how to incorporate “all units” quantity discounts in the model
of Sec. 6-2. Assume that the selling price remains constant. Devise
a procedure for computing the optimal order quantity. Sketch
curves equivalent to those of Fig. 2-13 for the case at hand.
6-8. Show how to incorporate incremental quantity discounts in the model
of Sec. 6-2. Devise a procedure for computing the optimal order
quantity. Assume that the selling price remains constant. Sketch
curves equivalent to those of Fig. 2-15 for the case at hand.
6-9. Solve Problem 6-1 under the assumption that “all units” quantity
discounts are available. If a quantity h is purchased, the unit cost
is $4.00 if 0 < h < 225, $3.00 if 225 < h < oo.

6-10. Solve Problem 6-1 under the assumption that incremental quantity
discounts are available. The first 225 trees cost $4.00 each and
additional trees cost $3.00 each.
 PROBLEMS 317

6-11. Consider Problem 6-1. Assume that “all units” quantity discounts
are available. If a quantity h is purchased, the unit cost is $4.00 if
0 < h < q, and $3.50 if q < h < co. Determine the value of q such
that the optimal order quantity occurs at the price break. For what
range of q values will $4.00 be the optimal unit cost. For what range
of q values will $3.50 be the optimal unit cost.
6-12. The buyer of a large West Coast department store must decide what
quantity of a high priced women’s leather handbag to procure in
Italy for the coming Christmas season. The unit cost of the handbag
to the store is $17.50 and it will retail for $50.00. The buyer is
confident that all handbags not sold by the end of the season can be
disposed of at cost. However, for every dollar invested in a handbag
not sold at the end of the season, the buyer feels that he is really
losing $0.30 since the dollar invested in something else could have
yielded this gross profit. The buyer believes that he will sell more
than 50 of the handbags but not more than 250. Sales of any number
within these limits seem equally likely. How many handbags
should the buyer procure? Suppose now that, after discussions with
other buyers, he can sharpen his sales estimate to a point where he
believes sales will be normally distributed with mean 175 and a
standard deviation of 20. Now what is the optimal quantity to
procure? How much was it worth to the buyer to gain the additional
information?
6-13. Consider the model described in Sec. 6-5 whose expected profit is
given by (6-16). Suppose now that the length of the period is gamma
distributed. Determine the expected profit g(ft) and Ag(ft).
6-14. Recall that for a first order approximation in At = ti — to, Taylor’s
theorem reads f(k) - f(t0) = f(to)At, where f(to) is the derivative of
/with respect to t evaluated at to. Use this result to simplify P(h; \iU)
— P(ft; \%ti—i), tJP(h; \%ti) ti—iP(h) p{h) \{ti) p(Ji] \iii—i)
and tip(h)\iti) - U-xpQi) \iU~i). What does Eq. (6-24) reduce to in
this case? Why is it now much easier to evaluate numerically?

6-16. Point out the problems involved in attempting to allow the length
of the period to be a continuous random variable when the mean
rate of demand is also allowed to change with time.

6-16. Attempt to compute AqB(Q, r) by marginal analysis, B(Q, r) being

the backorders term for the model of Sec. 4-7.

6-17. Derive Eq. (6-6) by computing directly the change in expected profits
when the quantity procured is changed from ft to ft + dh.
 318 SINGLE PERIOD MODELS CHAP. 6

6-18. The demand for a spare engine part is expected to behave in ac¬
cordance with a Poisson distribution with mean rate 0.01 units per
day. After 600 days elapse, it is known that the part will be obsolete
and will be scrapped for $200 less than its original cost. The costs
o carrying the unit in inventory amount to $0.10 per unit per day.
li parts are not available when engines break down, the daily costs
ot idle time are $900 per part needed and not in stock. What number
of spare parts should be acquired in advance for the 600 day period
of usage if all parts must be purchased at this time?

6-19. A farmer has to decide on which crops to take into the city on his
weekly trip. He has more than enough ripe crops to fill his truck
which can carry only 8000 lbs. of vegetables. Cabbage, cauliflower,
and tomato crops are the farmer’s leading candidates. The demands
i oooheon^ree Cr,°PS 'in P°unds) are normally distributed with means
, , and 2000 respectively. The corresponding standard
eviations are 200, 300, and 400. If the crops are not sold they are
worthless, while the profits per pound are 4 cents for cabbage, 6 cents
for cauliflower, and 7 cents for tomatoes. How should the farmer
load his truck to maximize expected profits?

6-20. A dressmaker is interested in buying a special type of imported fabric

for manufacture for the coming one week fiesta season. The fabric
must be ordered well in advance of the fiesta. The dressmaker knows
that one-third of the customers are willing to wait for the cloth to be
made into dresses while the remaining two-thirds would take their
business elsewhere. Ihe number of customers is equally likely to be
any number from 101 to 200. The cost of making the fabric into
dresses is $20 per dress made in advance of the season and $30 per
dress made during the season. The dresses sell for $100 each. Any
a nc remaining in stock at the end of the season can be returned at
f15 d^®s urut> wMe finished dresses can be sold after the season
for $50. What is the optimum number of (dress) units of fabric to
purchase. What number of dresses should be completed before the
season. What is the expected value of the dressmaker’s profits?
6-21. Suppose the dressmaker in Problem 6-20 must make all the dresses
m advance of the season (either because a negligible fraction of
customers will wait or because too much time or incremental expense
is involved m making them during the season). What is the optimal
number of dresses to make. What profits are expected?

6-22. Given that demand for a spare part is Poisson distributed with mean
0.1 umts per day and two alternative modes of delivery, namely two
ays and eight days at costs of $18.00 and $10.00 per unit respectively.
 PROBLEMS 319

The incremental costs incurred when a unit is demanded but not in

stock are $5.00 per unit per day and the inventory carrying charges
per unit per day are $1.00. Parts demanded but not in stock are
backordered. Design an inventory control system which minimizes
the sum of stockout costs, inventory carrying charges, and transpor¬
tation costs, assuming that only one of the two modes of transpor¬
tation can be used. What does such a system cost per average day
of operation? For what transportation cost for the two-day model
would the average cost per day be equal?

6-23. Assume in a one-period model that a cost of Ai is incurred for each of

N warehouses (i = 1, 2, . . . , N) in which the quantity demanded
exceeds stock on hand. Obtain a rule for allocating x units to the N
warehouses in order to minimize the expected value of the cost, when
the demand at each warehouse is a stochastic variable. Suppose that
the cost of stockouts was Bi per unit demanded but not in stock.
How should the x units be allocated in order to minimize the stockout
costs? (Assume continuous probability density functions).
6-24. A hula hoop manufacturer can store finished hoops, or plastic which
can be made into hoops in a week. Only one-third of his customers
are willing to wait for the hoops to be manufactured. Assume that
there is one discrete time period under consideration and that the
probability density function is uniform from 100 to 200 (0 elsewhere).
Finished hoops have no liquidation value. The plastic for one hoop
costs $0.50 and can be sold as surplus for $0.20. Manufacturing costs
are $0.10 per hoop. The hoops are sold for $1.20 each. What is the
optimal number of hoops to stock? What amount of plastic should be
stored in unfinished form?
6-25. The probability density function for sales volume during a given
period is f(v) = gg300 l>2(10 — v), 0 < v < 10, and the units are in
hundreds of pieces. The cost of raw material is $3.00 per item with
ato. added cost of $2.00 per item for manufacturing cost. The raw
material is made up only as orders are received and raw material not
used up will be sold for scrap at $1.00 per piece. Find an expression
for the number of items of raw material that should be ordered at the
beginning of the period as a function of the mark-up price M per item
when expected profits are maximized. The raw material must be on
hand at the beginning of the period.
6-26. In Example 2 of Sec. 6-3, calculate the differences between §(h*)
and S(m) associated with differences in expected profits on units sold,
expected losses on units not sold, and expected losses of goodwill on
units demanded but not in stock.
320 SINGLE PERIOD MODELS CHAP. 6

6-27. Assume in Example 3 of Sec. 6-3 that a goodwill loss of $1.00 per
rabbit demanded but not in stock is incurred. What is the increase
in optimal stock? What is the change in expected profits?
6-28. Consider the computational procedure discussed in Sec. 6-4 for
handling constrained problems when the variables are treated as
discrete. Under what circumstances is the procedure exact? When
it is not exact, under what circumstances should it yield a good
approximation?
6-29. Discuss the relationship between Eq. (6-4) and Eq. (4-116) with
7T = 0 for steady state models with Q = 1. Why should they be
of the same form?
6-30. Show for the general single period model discussed in Sec. 6-2 that
h* is the solution to

where h is treated as continuous, where Ci, C2 are the cost of overage

and underage respectively. Derive this equation by incremental
analysis by balancing the expected cost of overage against the ex¬
pected cost of underage. What is the corresponding result in the
discrete case? What are cu c2 in terms of w0> C, L, SI
]
I

J
DYNAMIC

INVENTORY MODELS

“When we mean to build,

We first survey the plot, then draw the model;

And when we see the figure of the house,

Which if we find outweighs ability.

What do we then but draw anew the model

in fewer offices?“

Shakespeare, Henry IV, Part II.

i
7-1 Introduction

In a strict sense, steady state conditions are a fiction in the real world.
The essential characteristic of all economic systems is that they are con¬
tinually changing with time. For inventory systems, the processes gen¬
erating demands and lead times change with time, as do the various costs
of interest, and even the items carried by the system. In many cases,
however, the changes occur slowly enough so that for considerable lengths
of time the system can be treated as if it were in a steady state mode of
operation. In other instances, however, the changes occur with such
rapidity that they must be explicitly accounted for. Usually it is the
changes in the process generating demands which are most important.
It is the purpose of this chapter to study multi-period models in which the
mean rate of demand changes with time.
As might be expected, the difficulty of formulating and obtaining
numerical solutions to realistic dynamic inventory models is considerably
greater than for the case where it was permissible to assume that the
system was in steady state. In fact, when demand is treated as a sto¬
chastic variable whose mean is time dependent, only the most trivial
problems can be solved manually. Usually a large digital computer is
needed to obtain numerical results. The natural formalism for setting up
dynamic models in a form for numerical computations is dynamic
programming.
Before turning to dynamic inventory models themselves, we shall first
introduce dynamic programming and cover the topics in this ‘subject that
will be needed here.

7-2 Dynamic Programming

Consider the problem of determining non-negative integers xj which

minimize the function g{xll. . . , xn) defined by

g{xh • • . , Xn) = = /ifo) +/2OB2) + . * . +fn(xn) (7-1)

y-i
323
 324 DYNAMIC INVENTORY MODELS CHAP. 7

subject to the constraint

n

X
3 = 1
v>xi ^ V (7-2)

where the vj; V are specified constants, and fj{Xj) is a function of Xj only.
We shall assume that the vs, V are integers. There is no real loss in gen¬
erality in doing so, since by proper choice of the physical dimensions (i.e.,
perhaps using cubic centimeters instead of cubic feet), it is always possible
to obtain an arbitrary degree of accuracy with the v,, V being integers.
The single period, multi-item problems subject to a constraint, such as the
flyaway-kit problem discussed in Chapter 6, are special cases of the general
problem just formulated.
We shall introduce dynamic programming by studying how one might
attempt to solve numerically a problem such as that formulated above.
The important thing to notice is that we can minimize over the variables
in any way we choose, provided that the procedure we select does permit
the examination of all possible combinations of values for all the variables,
if necessary. Suppose then that we proceed as follows: We select a value
of x„ and we minimize g over Xi, ..., xn_i for this given value of xn. By
(7-2), we see that the variables xx,... , xn_t must satisfy

71 — 1

X
y=i
V’Xi ^ V - VnXn (7-3)

so that the allowable range of variation for Xi, . ... , xn-i will depend on
the value of xn selected. Note that the value selected for xn can only have
the integral values 0, 1, 2, . . ., [V/vn], where [V/vn] denotes the largest
integer less than or equal to V/vn.
Now*

n—1
min 9 = f(.Xn) + min V ffe) (7-4)
XI, , Xn-l XI, , Xn-1

and in computing
n— 1
min X fiixs) (7-5)
..

xh , xn^i must satisfy (7-3). We can observe that the minimum value
expressed by (7-5) will depend on V - vnxn because of (7-3). Denote (7-5)
by Zn-.i(V — vnxn). Then Zn-i(F — vnxn) is the minimum of 1 fj(xj)

* The notation min g will be taken to mean the absolute minimum of g(xu ...<xn)

over xi, ..., £n-i for a fixed value of xn.

SEC 7-2 DYNAMIC PROGRAMMING 325

for non-negative integers which satisfy (7-3). Equation (7-4) can there¬
fore be written
min q — fixn) ~i~ Zn—i(V VnXn) (7-6)
XI, ..., Xn-1

If g* is the optimal value of g, we see that

g* = min [fn(xn) + Zn-i(V — vnxn)} (7-7)

Xn

Hence, if we knew the function Zn-i(£) for all integral arguments £ from
0 to V, we could determine g* simply by computing /n(0) + Zn-i(F),
fn(l) + Zn-i{V - O, /»(2) + Zn-i(V - 2vn), etc., up to = [7/wn],
and picking the smallest of these. We would simultaneously determine xt,
the optimal value of xn, in doing so.
The question then arises as to how we determine Zn-x(£) for any argu¬
ment £. By definition

Zn-i(l) = min 22 (7-8)

XI, . . . , In-1 j — 1

for non-negative integers j = 1, . . ., n — 1, satisfying

n—1
22
3 = 1
VfXj < £

We can now resort to the same trick as above. Suppose that we pick a
value of £»_i, and minimize /yfe) over a*, . . . , xn_2 for non-negative
integers rcy, j = 1, . . ., n — 2, satisfying
n—2
y ^ vjXj ^
i=l

For any argument £, let us now define the function Zn-2© by

Zn-2(8 = min fj(xj) (7-9)

XI, . - . ,Xn- 2 j=1

for non-negative integers j = 1, . . ., n — 2. satisfying

2 < £ (7-10)
y=i
Then
2/n—1(£) ^ mm [/n—l(^n—l) “b 2(£ ^n—l^n—l)] (7-11)

We can continue working down in this way until we are reduced to

evaluating
Zi(£) = min /i(rci) (7-12)
Xl
326 DYNAMIC INVENTORY MODELS CHAP. 7

for non-negative integers xx satisfying xi < [f/vj. It is a straightforward

task to evaluate Zi(£) for any given £. We need the value of Zi(£) in general
for all integers £ between 0 and V (if this is not clear to the reader now it
will be later). Once we have a table giving Zi(£) as a function we can then
compute a table giving Z2(£) by use of

Z2(£) = min [/2(^2) + Zi(£ - v2x2)]

Then we can compute a table giving Z3(£), etc., until finally from the
table for Zn^i(£) we obtain #* using (7-7).
Let us now repeat the computational procedure for determining g*, and
at the same time indicate how x*, . . . , x$ are determined. We define the
sequence of functions
k
Zk(£) = min ^fj(xj), k = 1,. . . , n (7-13)
XX, ... ,Xk j—l

for non-negative integers xh j = 1,. . . , k, satisfying

X
J-l
VfXj < £ (7-14)

Then
g* = Zn(V) (7-15)

and the 2^(£) can be computed recursively using the recurrence relation

Zk(£) = min \Jk(xk) + 2*_i(£ - vkxk)], k = 2,. . ., n (7-16)

where in the minimization Xk can take on only integral values in the range
shown. In order to compute Z*(£) for a given £, we will quite possibly need
to know Zk~i(r}) for all 77 = 0, But then in order to compute
zk+i(f), we will quite possibly need Z*({) for ah £ = 0, 1, ..., f. Finally,
since g* = Zn(\), it follows that each Zk(£) may need to be known for all
£ = 0, 1, 2,. . ., F.
The computational procedure can be thought of as an n stage process.
At the first stage we compute a table of Zi(£) for £ = 0, 1, . . . , V, and at
stage 2 a table of Z2(£), and finally at stage n the single value Zn(V). Note
that once the table Zk(£) had been obtained the tables for Z*(£), i < k, are
no longer needed.
In computing Z^(£) from (7-16) or Zi(£) from (7-12) we also obtain for
each £ the value of xk which yields the minimum. Denote by £*(£) the
value of xk which yields Z&(£) in (7-16) or (7-12). Suppose now that as we
record the table of £&(£) we also record a table of £*(£). At the last step,
when Zn(V) is computed, we automatically determine x$ and no table is
SEC. 7-2 DYNAMIC PROGRAMMING 327

recorded. However, we will have ft — 1 tables giving the £&(£) for

fc = 1, . . ., ft — 1. The question now is how do we determine xi, . . . , a?S-i
from these tables. This is done by working backwards from x%. To find
xt-i, we only need to determine £*(£) for £ = F — vnx%, i.e., arj-i
= fn_i(F — vnxi). Having determined rrj-i we are now able to compute
xt-2 since xl-% = xn_2(F — vnxt — yn_i^_x). We continue this procedure
until finally we find that

Thus in general

X% = xk (v — ]T} k = 1| • • • J n — 1 (7-17)
\ j**h+ i 7

It is by no means true that the optimal solution must be unique, i.e.,

the xk(£) functions need not be single valued. In numerical computations,
if one finds at stage h that more than one value of xk yields Zk(£), one can
tabulate each xk which yields the minimum, thus making it possible at the
end to find all alternative optimal solutions, or simply tabulate one of
these xk) thus yielding finally only one of the alternative optimal solutions.
We have now presented a numerical procedure for solving exactly the
problem formulated at the beginning of the section. The procedure can
be considered to be an n stage process where at stage k we construct tables
of Z*(£) and &(£) for each integer £ = 0, 1, 2, . . . , F. To obtain each entry
in these tables, it is necessary to carry out a minimization over a single
variable. If there are a large number of stages and if F is also a reasonably
large integer, it is clear that the computational procedure would be beyond
what could be done manually. The computational procedure suggested
can usually be carried out quite easily and quickly using a digital computer.
We might note, incidentally, that depending on the v3- values, it may not
be necessary to evaluate Z&(£) for each integer £ between 0 and V. How¬
ever, the task of determining precisely what £ are needed can be so com¬
plicated (it is necessary to work back from the last stage) that it is often
simpler merely to tabulate £*(£) for each integer in the range of interest.
This is especially true if the problem is being solved on a computer.
An interesting and helpful physical interpretation can be given to the
Zjb(£). Zjt(£) is the minimum cost if only stages 1, 2, . . ., k existed, and
the quantity of the limiting resource which could be devoted to these
stages was £. The physical interpretation of (7-16) is that an optimal
policy for k stages must have the property that whatever the choice of the
decision variable for the kth. stage, the policy must be optimal for the re¬
maining k — 1 stages for this choice of xk. Bellman [1] refers to this as
the principle of optimality.
328 DYNAMIC INVENTORY MODELS CHAP. 7

The amount of numerical work required using the dynamic programming

formalism will in general be much less than if we simply attempted to
determine the minimum by examining every possible combination of
values which the variables could assume. We shall now illustrate the com¬
putational technique described above by a very simple example.

EXAMPLE Consider a grossly simplified flyaway-kit problem involving

only three items. The volume available for storage is 9 cubic feet. Item 1
has a unit volume Vi of 1 cubic foot, item 2 a unit volume v% of 2 cubic feet,
and item 3 also has a unit volume vz of 2 cubic feet. The demand for each
item is Poisson distributed with the mean for the period being 4 for item 1,
2 for item 2, and 1 for item 3. The stockout costs are $900, $700, $1400
for items 1, 2 and 3 respectively. It is desired to determine how many
units of each item to stock in order to minimize expected stockout costs
while not exceeding the available volume.
If Xi units of item i are stocked, we desire to minimize
3 r oo
x = 2 x< LnC (y ~viv, Mi)
i=l .y = xi
3
1, fti) X{P
i—1

for non-negative integers Xi which satisfy

Xx + 2x2 + 2x3 < 9

We then define the functions

jfc
Zk(g) = min V tri[fjLiP(xi — 1; m) - XiP(xi; ^)], h = 1, 2, 3

for non-negative integers Xi satisfying

k
2 v&i < £
i=i

Then, if 3C* is the optimal value of X, X* = Z3(9). For k> 2, the Zk(£)
can be computed sequentially from

Zk(g) = min {■wk\jikP{xk — 1; wt) — xicP(xk; /t*)] + — vkxk)}

os”s[»]
Zi(g) is given by

Zi(?) = min {900[4P(a;i — 1; 4) — an Pfa; 4)]}

since Vi = 1. It is clear that the minimum in the above expression will be

 DYNAMIC PROGRAMMING 329
SEC. 7-2

talari on when xx = £, since the expression in braces is the expected stock¬

out cost for item 1 when xx units are stocked. Hence

Zi(Q = 900 [4P(£ - 1;4) - £P(€;4)]

and if &(£) is the value of xi which yields Zi(£), then &(£) = 0
In the text it was suggested that the £*(£) would in general be determined
for each £ = 0, 1, . . ., V. For this simple example, however, it is easy to
work backwards to determine which arguments £ will be needed. It saves
some work here to do this at the outset. To compute Z3(9) we need
Zi(9 — 2x3) where x3 can take on every integral value from 0 to [9/2] - 4.
Thus Zs(£) is needed for £ = 9, 7, 5, 3, 1. To compute Z2(0 for any one
of these arguments we need Zi(£ — 2^2) for all integral values o x2 rom
0 to [£/2]. Thus Zi(£) will be needed only for the arguments 9, 7, 5, 3, 1.
The values of Zi(£) and 4(£) and &(£) are given in Table 7-1 for the £
values needed. For example
Zi(5) = 900 [4(0.56653) - 5(0.37116)] = 369.3

To compute Z2(£) we use

Zi(£) = min (700[2P(22 - 1; 2) - a* P(a*; 2)] + Zi(£ - 2%)}

o<^[|]
For £ = 1, it must be true that x2 — 0, and hence

Z2(l) = 1400 -f- Zi(l) = 1400 -t- 2716 4116

For £ = 3, £2 can be 0 or 1. To determine which x2 value is optimal, we

compute the quantity in braces in (7-18) for x2 = 0 and x2 = 1, and pic
the smallest value, to yield Z,(£). When *2 = 0 the quantity in braces is
2613. When z2 = 1, we need Zi(l) = 2716 which is itself greater than
the entire quantity in braces when x2 = 0. Hence Z2(3) — 2613 an
x2(3) = 0. When £ = 5, m2 can be 0, 1, 2. For x2 = 0, the quantity m
braces in (7-18) is 1769, and is 2008 if a, = 1, and is greater than 2716 if

TABLE 7-1

Data for Example

Z i(£) MO z2(0 4(a

€

2716 1 4116 0
1
3 1213 3 2613 0
369.3 5 1769.3 0
5
76.29 7 1164 1
7
11.04 9 748.2 2
9
330 DYNAMIC INVENTORY MODELS CHAP* 7

£2 = 2. Thus £2(5) = 0 and Z2(5) = 1769. The same computational

procedure is repeated for £ = 7, 9 and the results are shown in Table 7-1.
We are now ready to make the final step and compute 3C* = Zz{9). To
do this we use

Zz(9) = min {1400[P(x3 - 1; 1) - x3P(x3; 1)] + Z2(9 - 2xz)} (7-19)

0<ar3<4

Computing the quantity in braces in (7-19) for xz = 0, we obtain 2148;

and for xz — 1, 1679; and for xz = 2, the value is greater than 1769. For
xz = 3, 4 the value is still greater. Hence £3(9) = £3 = 1, and 3C* = Zz{9)
= 1679. It only remains to determine xf, xf. We know from (7-17) that
x* = f2(9 - 2x|) = £2(7) = 1. We find £2(7) in Table 7-1 for £ = 7.
Finally xf — £i(9 — 2x% — 2x|) = £i(5) = 5. Thus 5 units of item 1 and
1 unit of items 2 and 3 should be stocked. The expected stockout cost is
$1679.

7-3 Dynamic Programming Formulation of Other Problems

Let us generalize the problem studied in the previous section to the case
where two constraints are present, i.e., we now wish to minimize
n

q{xij . . ., xn) = fjixj) (7-20)

j= 1

for non-negative integers xy, j = 1,. . ., n, which satisfy the two constraints

2 < V; Yj < w (7-2i)

i=l i=1

where the vj, Wj, V, IF are integers. Precisely the same procedure can be
used as before except that now Zk(£) must be replaced by a function of
two arguments 2jk(£, 77). Let us define the set of functions
k
Zk(£, v) = min ][]/y(xy), k = 1,. . . , n (7-22)
Xp .. .,xk y = 1

for non-negative integers xy satisfying

n n,

XI< £; X) ^ ‘n (7-23)
i=1 ;=1

Then g* = Zn(F, IF). Since the minimization can be carried out in any
way that we desire provided that all possible combinations of the values
of the variables are considered if necessary, it follows that for k > 2
SEC 7-3 FORMULATION OF OTHER PROBLEMS 331

i *~j

fk(xk) + min ^2 fjixj)

Xx,...,Xk-l j=sl J
and in computing
k-1
min
Xx,...,Xk~1 j=l

the variables must satisfy

k~i k-i
Y, VjXj < £ — vkxk; WjXj < 7] — wkxk
j=l 0=1

However, by definition, this minimum is simply Zk~i(£ —- vkxk} rj — wkxk).

Thus the Zk(%, rj) can be computed using the recurrence relation

Zk(£, n) = min [fk(xk) + Zk-i(£ — vkxk, -q — wkxk)], k> 2 (7-24)

Xk

and xk can assume every integral value between 0 and

Finally
Zi(£, 7]) = mmfxixi) (7-25)

The task of numerically solving this problem involving two constraints

is much more difficult than the problem involving only a single constraint.
The reason for this is that the Zk now involve two arguments, and hence
tables of two arguments must be constructed. Thus if for the case of one
constraint £ could take on 100 values, then for the case of two constraints,
if both £ and rj can take on 100 values each, a table of 10,000 values will
be needed, since Zk must in general be computed for every combination of
£ and r\ values. It can easily require a computer one hundred times as long
to solve a problem with two constraints as it would a problem with only
a single constraint. Furthermore, a great deal more memory capacity is
needed for the larger tables. The formulation of this problem points out
one of the serious limitations of the dynamic programming technique.
Only problems which can be formulated in a way such that the Zk func¬
tions depend on a very small number of parameters are feasible computa¬
tionally. In fact, any problem with Zk involving more than three param¬
eters is outside the range of feasibility, even with the largest computer
available, and problems with three parameters require a tremendous
amount of computer time to solve.
Let us now turn our attention to a slightly different type of problem
that will be of interest to us, which can also be cast into the dynamic pro-
332 DYNAMIC INVENTORY MODELS CHAP. 7

gramming framework. Suppose that we wish to determine non-negative

integers Xj, j = 1, . . ., n, which minimize

g{xi, = 22 /AVh xj) (7-26)

;=i

where yi is specified and

yj+1 = yj + Xj - dh j = 1, . . . , n — 1 (7-27)
Intuitively, the interpretation of the above problem that will be of interest
to us is that where g is the cost of operating a deterministic inventory
system for n periods, and fj(yj, x0) is the cost associated with period j
when xj is the quantity ordered at the beginning of period j and yj is the
inventory position at the beginning of period j before an order is placed.
Equation (7-27) is the material balance equation which relates the inven¬
tory position at the beginning of period j + 1 to that at the beginning of
period j, the quantity ordered Xj at the beginning of period j, and the
demand dj in period j.
We shall now show how to cast this problem into the dynamic pro¬
gramming framework. The important thing to note is that g* will depend
on 2/1, and as the problem is stated, yi is the only parameter which can be
varied. This suggests that 2/1 should play the same role as V in the problem
of the previous section. In other words, the inventory position at the
beginning of period k will be the parameter appearing in Zk. Here again
we can imagine the problem to be an n stage process. Now, however, in
making the computations we must start with the last stage and work
backwards (physically this will usually mean working backwards in time),
since 2/1 is specified at the beginning of the first stage. In the example of
the previous section, it did not matter which stage was considered to be the
first, but here it does.
Let us define the sequence of functions

Zk(t) = min (7-28)

Xk,...,Xnj==k

for non-negative integers Xj, j = k, . . ., n, where the parameters 2/y must

satisfy
Vi+i = Vi + Xj dj, j = k,... ,n - 1 (7-29)
and
yk = $ (7-30)
Then g* = #1(2/1). Furthermore, for k < n
n

Zt(£) = min [/*(?, xk) + min x,:)]

Xk Xfc+1, . . . , Xn j — £_|_ 1
SEC. 7-3 FORMULATION OF OTHER PROBLEMS 333

where
yic+1 = £ + %k — du

and yk+2,. . . , yn must satisfy (7-29). Thus

Zk(£) = min [/*(£, as*) + Zk+i(£ + rr* — d*)] (7-31)

Xk

and
£«(£) = min/»(f, s») (7-32)

In carrying out the minimization in (7-31), (7-32), xk can take on any

non-negative integral values.
The computational procedure is precisely the same as in the previous
section. First, one computes Zn(£) for all integers £ from the minimum
possible value to the maximum possible value. Simultaneously £*(£), the
value of xn which yields Zn({), is tabulated. Then (7-31) is used to compute
recursively the other Zk(£). As these are computed, xk(g) is also tabulated.
At the last stage it is only necessary to compute Zi(yi). A table of Zi(g) is
not needed. The value of x* is found at this point also. Then

xt = x2(yi + xf - di); z$ = f3(yl + x$ — d2), etc. (7-33)

The method of solution just presented is sometimes referred to as the back¬

ward solution since it works back from the last stage (i.e., usually back¬
wards in time). The Zk(£) can be interpreted physically as the minimum
cost for periods k through n if the inventory position at the beginning of
period k is £.
Suppose now it is also required that

2/n+l — Vn “I"" dn (7-34)

where yn+i is a specified constant. This places restrictions on yn and xn.

Physically the new constraint specifies the inventory position at the end
of the last period. In this case the problem can be formulated using a
forward solution as well as a backward solution. To obtain the forward
solution formulation define the set of functions
k
£*(£> = min fiiyj, xj) (7-35)
Xlt...,Xk j=l

where
yk + Xk — dk = £

and 2/2,. . . , 2/fc satisfy (7-29). Then the recurrence relations take the form

Zk(i) = min [/*(£ + dk — xk, xk) + Zk-i(£ + dk — xk)], k > 1 (7-36)

Xk
334 DYNAMIC INVENTORY MODELS CHAP. 7

and
J?i(£) = min/i(£ + dk — xh Xi) (7-37)

so that g* = Zn(yn+i). The computational procedure is precisely the same

as above. Here the Zk(£) can be interpreted physically as the minimum
cost for periods 1 through k if the inventory position at the end of period
k is £.
Let us now turn our attention to the case where the dj are discrete,
independent random variables with distributions Pj{d3). Physically we
might think of this as a problem in which demand is a stochastic variable.
The functions f3(y3, Xj) can now be interpreted as the expected cost for
period j if the inventory position immediately after the review is yj + x3\
The function we now wish to minimize is the expected cost, i.e.,

g = X) [ n pj(dj)~\ ["Z xy)~| (7-38)

alld,->0 bi=l J Ly = 1 J

where 5Zy=i fj(y3j x3) is evaluated for a given set of y3 (the y3- being related
to the dj by (7-29)) and

n 3
p {d3) = Pi(di)p (d ) 2 2 . . . Pnidn) (7-39)
i=i

is the probability of the given set of dj, since the dj are independent
random variables.
As before, we assume that yi is specified. For a stochastic process it is not
possible in general to specify the state of the system at the end of the last
period, and hence a backward solution will always be needed in this case.
By analogy with the deterministic case we define the set of functions

Zk(£) = min J Z [~ n py(dy) Z fAVh %) L

Xk,...,Xn\ alldj>0 \J = k J =A J
'•j = k, . . . , n J
k = (7-40)
where yk = £. Then g* = Zi(yi) as before.
Let us now consider the derivation of the recurrence relations. Note
that in Zk, the function fk(yk, xk) = /&(£, xk) is independent of the dj. Also

Z [ n py(dy)l = n I" Z = 1
alldi>0 Li = * J j = Jc Lrf. = o ->
j = k, ... ,n

Hence
En ~] r -
min xk) + min n py(dy) Z
Xk l Xk+U...,Xn all flfj >0
u<ti>0 U=k JLyii+l -I
j = k, . . ., n
SEC 7-3 FORMULATION OF OTHER PROBLEMS 335

= min xk)

+ X Vk(dk) [min X {. H { X ^)}1}

djfe = 0 L.a?fc+i» all<ij>0 L? — &Hr ^ J Ly = &-{-1 J JJ
y=£-bl, . . . , n

= min -[/*(£, Xk) + ^2 Pk(dk)Zk+1(£ + Xk — <4)}-

Xk L dk=o J
Thus the recurrence relations are

^(f) = coin //*(£, £*) + pk(dk)Zk+i(i; + Xk — £4)"k k < n (7-41)

x* L <4=0 J
and
£*(£) = min/B(£, a?»)
Xk

The computational procedure is precisely the same as for the deterministic

case. It is somewhat more difficult here, however, because in (7-41) it is
necessary to sum Zk+i over the possible values of dk. In any practical
problem, some upper limit to the demand would be used rather than oo as
shown on the summation sign in (7-41). A subtle difference enters into
the determination of the x*. The formulas are precisely the same as in the
deterministic case. However, a numerical value can be obtained only for
Xi. To obtain numerical values for Xz, ... ,xn the dj must be known. This
is very plausible intuitively since, for example, one would not want to make
a decision as to how much to order in period 2 until the quantity demanded
in the first period was known. As in the deterministic case, Zk(£) can be
thought of physically as the minimum expected cost for periods k through
n if the inventory position at the beginning of period k is £.
Dynamic programming can also be used for problems in which all vari¬
ables are treated as continuous. For continuous variables, however, one
must develop some sort of search procedure to carry out the minimization
over Xk in the recurrence relations. The procedure of examining each of the
possible values suggested in the text can no longer be applied because of
the continuum of values allowed. One straightforward way to find the
minimum is to begin by using a relatively coarse grid to divide up the
interval over which xk can vary, and to determine the relative minimum
(or minima) for these points. Then progressively finer grids are used in the
neighborhood of the minimum (or minima) until sufficient accuracy is
obtained. There exist a great variety of specialized search procedures
which can be applied if the functions /y have certain special properties
(such as being concave). (Indeed, even when the xk are treated as discrete
it may be advantageous in some cases to use a specialized search procedure.)
336 DYNAMIC INVENTORY MODELS CHAP. 7

When xk is continuous, £ will usually be continuous also. In this case,

Zkii) will be tabulated for some gridwork of points, and interpolation will
be used to determine Zk(£) for { values other than those for which Zk is
tabulated. Such a procedure is legitimate, since Zk(£) will normally be a
continuous function of £. It is clear that the case of continuous variables
is more difficult to treat using dynamic programming than that where the
variables are discrete. This is just the opposite of what was true for the
computational procedures used in the previous chapters.

7-4 Dynamic-Deterministic Lot Size Model

We would here like to consider a dynamic generalization of the simple lot

size model of Sec. 2-2. In dealing with dynamic models it is difficult to use
an infinite planning horizon (i.e., consider all future times) since numerical
calculations using dynamic programming can be made for only a finite
number of periods. This is no serious restriction in practice, because the
distant future has essentially no effect on decisions to be made at the
present, and because for many situations in which a dynamic model will be
applied, the item becomes obsolete after a certain length of time, and a
planning horizon longer than the time until obsolescence would have no
meaning.
For the model to be developed it will be assumed that the demand rate
is deterministic. No backorders or lost sales are to be allowed, i.e., the
system can never be out of stock when a demand occurs. The unit cost C
of the item is assumed to be a constant which is independent of the quantity
ordered. We shall imagine that there are n times ft, . . ., ft at which orders
can be placed. The procurement lead time associated with any order placed
at time ft will be ry, here assumed to be constant. Note that the value of
the procurement lead time is allowed to vary with j. However, it is assumed
that the ry are such that orders cannot cross. The possible times at which
orders can arrive are then t, = ft + ry. The cost of placing an order at
time ft will be written Ay. Observe that an order need not be placed at
time ft. One can be placed at this time if it is desirable to do so.
The quantities Qj(Qj > 0) to order at times ft will be determined by
minimizing the sum of the ordering and carrying costs over the planning
horizon, which will be assumed to end at time f(f > f» + r»). We suppose
that the desired on hand inventory at time f is specified; it will be denoted
by yn+1- It will be noted that we have no influence whatever over the
carrying costs incurred from the time ft when the first order can be placed
until the time ft + ri = ti when the first order arrives. In addition, these
costs are independent of the Qy, and hence need not be included in the cost
SEC 7-4 DYNAMIC-DETERMINISTIC LOT SIZE MODEL 337

expression. The only holding costs that are relevant are those incurred
between h and f.
By period j we shall mean the time from t3 to ij+i (tn+1 = f). Write
T3 = tj+i — t3. Let yi be the on hand inventory at time h. The demand
rate as a function of time will be written \(t). Then the demand d3 in period
j is
rti+1
d3 = J A(t) dt

We are now ready to develop the total variable cost expression. All
variables will be treated as continuous. Let I3C be the cost of carrying one
unit in inventory for period j. Note that this is not an annual cost but the
cost for period j. If y3 is the on hand inventory at the beginning of period j
before any order placed at time fy has arrived, then the inventory carrying
charges for period j are

i c fti+i r n i t o fu+i r*
-jT Jt Vi + Qj ~ Jt Hu) du dt = I3C[y3 + Q3] - ~jT jt Jt Hu) du dt

It is convenient, for reasons which will soon become apparent, to express

this in terms of the on hand inventory at the end of the period rather than
that at the beginning of the period. This can easily be done using the
material balance equation

yj+1 = Vi + Qi - dj, j = 1, • • •, n (7-42)

Here y3+i is the on hand inventory at the end of period j (and the beginning
of period j + 1 before any order arrives). Thus the carrying costs for
period j can be written

I3Cy3+1 + I3C ^d3 ~ Y-fti f du dt~\ (7-43)

Now note that I3Cy3+i is the inventory carrying cost in period j for those
units carried into period j + 1. The other term in (7-43) is the cost in
period j of carrying the d3 units which are demanded in period j. This
latter cost is independent of the Q3 and is unavoidable, since the d3 units
demanded in period j must be on hand at the beginning of period j. Hence
the carrying cost in period j for these d3 units need not be included in the
variable cost expression.
Thus the variable costs of ordering and holding inventory which are
incurred over the planning horizon are
n

3C = y ] [A38j + I3Cy3+i] (7-44)

3=1
338 DYNAMIC INVENTORY MODELS CHAP. 7

where
0 if Q, = 0
$3 ~~
1 if Qj > 0

and the y, are related by (7-42). As formulated above, X is the variable

cost over the planning horizon and not the present worth of the variable
cost. It can equally well, though, be interpreted as the discounted cost I
if it is imagined that the discounting factors are included in the Ay and Iy.
•I
However, when discounting is used, it is necessary to include in (7-44) the
discounted cost of the units themselves. Problem 7-5 asks the reader to
work out a means for solving the problem when discounting is introduced.
The dynamic programming formalism introduced in the previous section
can be usefully employed to yield a computational procedure which can
easily be carried out by hand to determine the Q*. Since the ending inven¬
tory is specified as well as the beginning inventory, it is possible to use
either a forward or backward solution. The forward solution turns out to 4
be the most interesting, for reasons which will become clear below. The
forward solution formulation will now be developed. "i
Let us define the sequence of functions

Zk{£) = min Y] {Ayoy + Ifiyy+is, k = l,... ,n (7-45)

where yk+i = £ and (7-42) holds for the other ys. Then from (7-36), (7-37),
the recurrence relations are

Z&(£) = + min {AA + Zk-i(£ + dk — Q&)} (7-46)

Qk
Also l
if Qi > 0
Zi(Q = hQ + {q 1 if Qi = 0 (7-47)

where
'£ + di — yi if £ > 2/i — di (7-48)
.0 otherwise

We shall now restrict our attention to the case where yi = 0, i.e., the
I
time fi is chosen so that everything on hand at time fi will be used up just
as the first order arrives. There is normally no reason to have a procure¬
4
ment arrive before everything on hand is used up, and hence the case of
yx — o is the one that is usually of interest. If yi = 0, it is possible to make
several observations that will considerably reduce the computational effort
below that required if the problem was solved directly using the above
recurrence relations. Let us note first of all that if QI > 0 then yi = 0, i.e.,
the on hand inventory at the time of arrival of an order should be zero.
The reason for this is that if yk > 0, a saving in inventory carrying costs
 SEC. 7-4 DYNAMIC-DETERMINISTIC LOT SIZE MODEL 339

could be effected by increasing the quantity procured at time from

Q! to Qt + yk. There would be no increase in ordering costs on doing
this, and hence a net reduction in costs would result. The argument above
is based on the assumption that the unit cost of the item is a constant.
Given that yt = 0 if Qt > 0, it follows that Qt must be equal to an inte¬
gral number of periods demand, i.e., if Qt > 0 then

Qt — dk or dk + dk+i or ... or dk + . . . -f- dn (7-49)

This observation considerably reduces the number of Qk values which need

to be examined in determining the minimum in (7-46). It will be noted that
if the demand in period j > k is satisfied by Qt, then the demands in all
periods k + 1, . . ., j — 1 will also be satisfied by Qt If this was not so,
i.e., if Q* > 0, k < v < j, and Q* satisfies the demand in period v and per-
haps one or more of periods v -j~ 1, . . ., j — 1, then Q$yv > 0, since some
of Qt must be kept to satisfy the demand in period j. This contradicts
what we have obtained above.
Next, let us show that if for Zk(0) it is optimal to have Qt > 0 (and
hence yt = 0), it is also optimal to have Qt > 0 for Z*(£), £ > 0. This can
be proved by considering (7-46). Note first of all that

M0) = min {Akh + Zk_Adk - Qk)}

Qk

(7-50)
Ak Zk-i(fi)
= min
Zk~i(dk)

Since we have assumed that Q% > 0, i.e., Qt = dk when £ = 0 in Z*(£), it

follows that Ak + Z4_i(0) < Zk-i(dk). Furthermore, it must be true that
•£i~i(£) > Zk-i(dk) if £ > dk, for if we end period k — 1 with more units,
it was at least necessary to store these units for period k — 1, and hence the
cost will be greater for the first k — 1 periods. Now

■Z*(£) = min {4A + x(£ + dk - Qk)} (7-51)

However, we have noted above that

Ak + Zk—i(0) < Zk-i(dk) < Zk-i(£) when £ > dk.

Thus
Q% = dk > 0
and we have proved what we set out to show.
The result we have just proved has the following physical interpretation.
If it is optimal to have an order arrive at the beginning of period k when
there are only k periods and the on hand inventory is to be zero at the end
of period k} then it will be optimal to have an order arrive at the beginning
of period k regardless of how many additional periods there are.
 DYNAMIC INVENTORY MODELS CHAP. 7
340

One final computational simplification will be introduced. Note that an

optimal policy for k periods, when nothing is on hand at the end of period k
must have the form that an order arrives at the beginning of period w which
satisfies the demands in periods w through k, and an optimal policy is
followed in periods 1 through w — 1 given that nothing is on hand at the
end of period w — 1. Thus
k-1 r k -i
min 4 Aw + C w = 1, 2,. . . , k
22 ^'.2 di +Zu
i=M L i = i -4-1 —

where Zq(0) = 0, or if

Yk(w) = Aw + C 22 IJ i =j22+1 di
j = w L-
^w-i(O)

Zk{0) = min Yk{w), w = 1, 2, . . . , k (7-53)

W

Note that Yk(w) is the cost for periods 1 through k under the assumptions
that the on hand inventory at the end of period k is zero, an optimal policy
is followed for periods 1 through w — 1 given that the ending inventory
for period w — 1 is zero, and an order arrives at the beginning of period w
which satisfies the demand for periods w through k.
We can now show that if for Zk—i{0), the order which satisfies the demand
in period k — 1 arrives at the beginning of period v, then in computing
Zk(0), the order which satisfies the demand in period k will arrive at the
beginning of period v or a later period, i.e., in computing Zk(0) it is un¬
necessary to consider the possibility that the order which satisfies the
demand in period k arrives before period v. Then in (7-53), it is only neces¬
sary to allow w to range from y to k. We shall prove this by contradiction.
Assume that
fc ^ — Jc —«
Zjk(O) = Yk(u) = Au + C h % di + 0) <
j=u L i=j+1 *-*

where

Zk-1(0) = Yk-i(v) — Av + C 22 h 22 Yv-i(0)

j = v <- i=j+1 “*

and v > u. However, on factoring out in Zk{0) the carrying costs for the
items carried from period u through k to satisfy the demand in period k, we
have

Zk(0) = Cdk 22 ”f~ Yk-i{y) — Yk(u)

 SEC. 7-4 DYNAMIC-DETERMINISTIC LOT SIZE MODEL 341

But

Yk(v) = Cdk 2 h +

On subtracting Yk(v) from Zk(0), we obtain

£*(0) - Yk(v) = Cd* 2] h + Yk^{u) - F^Cr) > 0

since by assumption
k-1
F*-i(w) > F*_i(t;) = Z*-i(0) and Cd* ^ Jy > 0
j=u

Hence Zk(0) > Fa(z;), which is a contradiction.

Thus we see that the computational procedure has been reduced to a
very simple form. At stage k, we compute Zk(0) from (7-53) using (7-52)
for Yk(w), and we allow w to range over the values v to k, where in Zk„i(0)7
the order satisfying the demand for period k — 1 arrived at the beginning
of period v. The optimal policy will normally be obtained on computing
Zn(0). If it is desired to have £ units on hand at the end, then at the last
step we compute instead

Z»(£) = min
w
aw + cj_;[// J d* + f) Zw-1
- j=w «— ' —

and again w only needs to run from v to n, where in Zn-i(0) the order which
satisfied the demand of period n — 1 arrived at the beginning of period v.
The reason that the same computational procedure holds for Zn(£) as for
Zn(0) is that having £ units remaining at the end is equivalent, insofar as
the previous periods are concerned, to increasing the demand in period n
by £.
A convenient tabular format for making the computations is given in
Table 7-2. In the last line, the optimal Q3- values are indicated by enclosing
in parentheses all periods whose demands are met by the order arriving at
the beginning of the period whose number appears first in the parentheses.
Thus (3, 4, 5) would mean that it is optimal to meet the demands in periods
3, 4, and 5 with the order arriving at the beginning of period 3. Table 7-2
indicates the results which might be obtained in a typical case. The aster¬
isks on the Yk(w) indicate the minima over w, i.e., Zk(0).
In the above model the times at which orders could be placed were
specified. A more general problem would be to determine both the times
at which orders should be placed and the quantities to be ordered. The
problem of determining the precise times at which orders should be placed
when the demand rate varies with time turns out to be quite difficult even
in the simplest cases. Problem 7-6 will ask the reader to derive the equa-

i
 SEC 7-5 DYNAMIC LOT SIZE MODEL-EXAMPLE 343

tions which determine these times for the case where there are no fixed
ordering costs, and to examine the difficulties involved in solving them.
It will be noted, however, that by allowing the times at which orders can be
placed to be sufficiently close together in the model developed in this
section, it is possible to come arbitrarily close to the solution where the
optimal times at which orders are to be placed are to be determined.

7-5 Example of the Dynamic Lot Size Model

A sub-contractor to a major missile manufacturer has a contract to deliver

at the beginning of each month for the coming year specified numbers of a
particular die cast part. The number of units to be delivered varies from
month to month, and the specific monthly values are given in Table 7-3.
This particular item is produced in lots. The setup cost for a production
run is $300. The variable production costs on the item amount to $120
per unit. The sub-contractor uses an inventory carrying charge of I = 0.20.
It is desired to determine how many production runs should be made and
when they should be made. Production decisions are made only at the
beginning of each month, and for the item under consideration, if a run is
scheduled at the beginning of a month it will be completed by the end of
the month, and any units not shipped immediately to the missile manufac¬
turer at the end of the month will be sent to the sub-contractor’s warehouse.
The sub-contractor does not want to have any units in stock at the end of
the year when the contract terminates.
In the notation of the previous section As = A = $300 independently
of when the setup is made, and Ifi = 7(7/12 = $2.00 per unit per month.
Here a period is one month, and it is convenient to imagine that the number
of units which must be delivered to the missile manufacturer at the begin¬
ning of month k + 1 represents the sub-contractor’s demand in period k.
Then

Zi(0) = A = 300 = Fi(l)

„ ,n, . (2A = 600 = Fs(2) 1
2(0) mm |2(ioo) + 300 = 500 = F2(1)J ~ 500 =

This shows that it is not optimal to have another setup at the beginning
of the second period, if the on hand inventory at the end of the second
period is to be 0. Next

U + Z2(0) = 800 = F3(3)

Z3(0) = min 2(125) -2A = 850 = F3(2)
(2(125) + 2(225) + A = 1000 = F3(l)

= 800 = F3(3)
 SEC. 7-6 STOCHASTIC DYNAMIC MODELS-FIXED HORIZON 345

In the three period case it is optimal to set up at the beginning of period 3

to satisfy the demand for period 3, and in period 1 for the demand of the
first two periods. In computing ZA{0) we know that there will be a setup in
period 3. Hence it is unnecessary to consider the possibility of producing
for period 4 in any period prior to period 3.

Zi(0) = min 'A + Z3(0) = 1100 = F4(4)

2(100) + A + Z2(0) = 1000 = F4(3).
= 1000 = F4(3)

It is optimal to set up in period 3 to satisfy period 4 demand. The re¬

mainder of the computations are presented in Table 7-3. It is seen that an
optimal solution is

Qt = 180, Qt = 0, Ql = 225, Qt = 0, Qt = 100, Qt = 0, <2? = 225,

Qt = 0, Qt = 125, <2*o = 150, Qh = 0, Qfc = 100
where Q*} is the quantity to be produced in month j. It should be noted
that the optimal solution is not unique. In periods 5, 6, 7, 9, 11, and 12
other choices are possible. The cost of setups and carrying inventory is
82950.

7-6 Dynamic Models with Stochastic Demands and a Fixed Horizon

We shall now turn our attention to dynamic models in which demand is a

stochastic variable. As in the previous section, it will be imagined that
orders can be placed only at times ft = 0, ft, . . ., ft- The planning horizon
is assumed to be of a fixed length ft As usual, we shall examine the case
where all demands occurring when the system is out of stock are back¬
ordered. The procurement lead time rs for any order placed at t,imP ft. win
be assumed to be constant (the lead time can vary withy, however, subject
to the restriction that orders cannot cross). The times at which orders can
arrive will be denoted by 4, - - -, ft, where ft = ft- + t> It is assumed that
f > ft. The time between ft and ft+i (ft+1 = ft will be referred to as period
j. The inventory position t/i at time ft when the first order can be placed is
assumed to be specified. Because of the stochastic nature of the process,
it is not possible to specify the ending inventory at time ft
It will be imagined that the cumulative expected demand from t.iW
ft = 0 to time t(t < ft can be represented by the continuous function D(t).
Furthermore, it will be supposed that the demand in any time interval, say
from t' to t" (0 < t' < ft t' < t" < ft is Poisson distributed with mean
D(f") - D(t’). Let
Ty ft+i ft, y — 1,. . ., n (7-54)
346 DYNAMIC INVENTORY MODELS CHAP. 7

Then it will be convenient to define Xy, cr3 such that

+ r/+l) = D(tj+1) - D(fy), cryTy = D(fJ+1) - £(fy) (7-55)

Thus \y is the mean rate of demand which if maintained constant during

the time from fy to t3+1 would lead to the same probability distribution for
the demand over this interval as that appropriate to the actual time varying
rate. A similar interpretation can be given to a3.
The cost of placing an order at time fy will be written Ay, and the quan¬
tity ordered will be denoted by Qy (Qy > 0). The cost of the Qy units
ordered at fy will be taken to be Cy(Qy), and we allow this cost to be non¬
linear in the quantity procured. The cost of holding a unit in stock for
period j will be written I3C3J where C3 is an average unit cost, which can be
defined in any way desired. The fixed cost of a backorder incurred in
period j will be written xy and the cost of a unit year of shortage will be
written xy. In a number of instances in which the present model is of
interest, all items remaining on hand at time f will be sold. Let L be the
unit selling price at time f.
As was the case for the deterministic model of the previous section, the
costs incurred from the time fi to ti are independent of the Qy and need not
be included in the analysis.
Let us now compute the expected carrying and backorder costs incurred
in period j if y3 is the inventory position of the system at time fy prior to
placing an order. The inventory position after any order is placed will be
Vj + Qy. We shall imagine that the fy are chosen to be sufficiently close
together that for any t value lying between t3 and t3+i we can write to a
sufficiently good approximation

W - fy) - D(t) - D(fy)

i.e., Xy can be assumed to be constant from t3 to tJ+1.

The expected costs of carrying inventory in period j are then
liC, fti+ivi+Qi

Tj + ry+1 - ry /,
JM
Z-j (&* + Qi - x)v\*)
x~0
W ~ fy)] dt

JjCj J Tj+Tj+l

Tiy -f
+ Ty+i Tj Vi + Qi ~
(7-56)

+ L (p - Vi — Q])v(p', V) r dr
x=Vi+Qi J

Ifij Vj + Qy 2 (^* + Ty+1 + Ty) + B(y3 + Qj} T3 + ry+i, ry)

where
(Ty + ry+i — Tj)B(y3 + Qy, T3 + ry+i, ry)
SEC 7-6 STOCHASTIC DYNAMIC MODELS—FIXED HORIZON 347

is the expected unit years of shortage incurred in period j. Note, however,

that this expression has been evaluated in Sec. 5-3 and is given by (5-26),
if there instead of r + T we use Tj-\- ry+i and instead of r we use ry.
The expected number of backorders incurred during period j is simply

E(yj + Qjy Tj + Ty+1, Tj)

oo

= 1] (x-yj- Qj) {p[x; Xy(Tj + Ti+1)] - p(x; X;ry)} (7-57)

x = Vi+Qi

E(Vi + Qi, Tj + r]+ , t/) can be written down explicitly by use of A3-10.
1

The expected gain from sale of units remaining at time f is

Vn+Qn
G(yn + Qn) = L y) (x — yn — Qn)p[x; X„(f — fre)] (7-58)
x=0

All the terms needed in the cost expression have now been evaluated.
The yj are related by the material balance equation

Vm = Vi + Qi - dj (7-59)
where d, is the demand from to fy+i- The probability of dj is p(dj) <jjT3).
It follows that the expected variable cost over the planning horizon of
procurement, carrying, and backorders, net of the expected gain from sales
at time f, is

3C = X) r n v(du VjTj)] [ £ hjSj + Cj(Q,)

alleZ*>0 L-J^ -J I—y = i L

+ djCj j"y, + Qj — (Tj + Tj+i + Tj) J + TTjE(yj + Qj, Tj + t,+1, t,-)

+ [(Tj + Tj+1 - Tj)TTj + IjCj]B(yj + Qj, Tj + r,+1, r,)| - G(yn + Qn) J

(7-60)
3£ can also be interpreted as the discounted cost if it is imagined that the
discount factors are included in the As, Cj(Qj), Vj, fij and L.
Dynamic programming provides a natural way to determine numerically
the functions Q*(y,). Because of the stochastic nature of the process a
backward solution is needed. Let us define the set of functions

z&) = min / [S V(dj\ *jTj) 1 ["£ Afij + Cj(Q3)

Qk,...,Qn \ ' alld,->0 1-3 = k J L/r£
L j=k, . . . , n

+ Ifii [</y + Qi ~ g (T= + TJ'+! + Ti) ] + ^iE(Vi + Qj, Tj + tj+i, Tj)

348 DYNAMIC INVENTORY MODELS CHAP. 7

+ [(Tj + Ty +1 — Tj)7Tj + IjCj)]B(yj + Qjj Tj + fy+ , ry)

1

- <?(</»+ Q»)]j, * = (7-61)

Then from (7-41), the recurrence relations are

Zk(£) = min |lA + Ck(Qk) + IkCk j^£ + Qk — ^( Tk + Tk+i + Tk) J

+ TTkE(t; + Qk; Tk + Tk+l, Tk)

+ [(Tk + Tk+l — Tk)fk + IkCk]B(% + Qk, Tk + Tk+l, Tk)

+ CkTk)Zk+i(£ + Qk — (4)1-, k = 1, . . . , n — 1 (7-62)

dk = 0 J

Zn(£) = min -^An^n + Cn(Qn) + InCn £ + Qn “ ^ (f — f» + Tn) ^

+ T-nE(% + Qn, f f«, Tn) (7-63)

+ [(r — fn “ Tn)7Tn + /nCn]#(£ + Qn, f — fn, Tn) — (?(£ + Qn)1

It should be recalled that a numerical value can be obtained only for Qi.
For the other Q/ s, we obtain a function Q*{ijj), and it is not possible to
determine Q* until y, is known, i.e., until the time is reached when Qj is to
be selected. In practice one will often only be interested in Q*, and will not
be interested in the Q%y3), j > 1. The reason for this is that the problem
will often be solved each time a procurement is to be considered, using
additional information obtained since the last time the computation was
made.
The procedure for numerically solving the problem has been outlined in
Sec. 7-3. In most cases the task is much too burdensome to carry out by
hand. It can be solved fairly readily on a large scale computer, however.
Even on a large scale computer, though, it is sometimes desirable to intro¬
duce shortcuts to save time. For example, if it is possible for Q* to be
quite large, say over 100, the task of minimizing over Qk when Qk is changed
by only one unit at a time may be excessively time consuming. To reduce
the time required to make these computations, one might, for example,
vary Qk one unit at a time between 1 and 25, five units at a time between
25 and 100, and ten units at a time over 100. In other words, Qk cannot
take on any integral value when Qk > 25, but instead can only take on the
values 30, 35, 40, etc., to 100 and then 110, 120, etc. The upper limit on £
will simply be the maximum level that it is ever expected that the inventory
position will reach. Again, to save computational time, it is advisable to
 SEC 7-7 STOCHASTIC DYNAMIC MODELS—VARIABLE HORIZON 349

keep it as small as feasible. If the expected lead time demands are suffi¬
ciently large, it will be desirable to replace the Poisson terms by their
normal approximations.
It is also possible to give a formulation of the above model in the case
where the lead times are random variables. Problem 7-9 asks the reader to
do this. The lost sales case is, as usual, very difficult to handle. All the
problems encountered in the steady state period review systems are also
present here. In the event that more than a single order can be outstanding,
it is no longer possible to have the Zt be a function of a single argument
only. This means, essentially, that at the present time, it is impossible to
solve such problems numerically. We ask the reader to examine the lost
sales case in Problem 7-10.

7-7 Dynamic Models with Stochastic Demands and a

Variable Horizon

On occasions one encounters situations in which it is desirable to use a

dynamic model of the type discussed in the previous section. However,
it is now no longer true that the planning horizon is fixed. Instead, the
planning horizon must be considered to be a random variable. A practical
example of such a situation arises in stocking spare parts for a military
aircraft which will become obsolete at some date in the future that is not
known with certainty, but which can be described probabilistically. Here
the planning horizon is the time until obsolescence. We would now like
to generalize the model presented in the previous section to the case where
7, the time at which any units still on hand are disposed of, is a random
variable.
It is very difficult to treat the case where 7 is allowed to be a continuous
random variable. Furthermore, in any real world situation it is usually
very difficult to attempt to decide on a continuous density function for 7.
About the best that can be done is to select a finite number of times
f(f), i= 1, ... ,m, and the probability p,- that the date of obsolescence will
occur at time 7(f). For example, the 7(f) may be years 5, 6, and 7 and the
pi give the probability that the plane will become obsolete at the end of
years 5, 6, and 7 respectively. We shall follow this same procedure in the
development of the model.
In general, we would expect the cumulative mean demand to depend on
the date of obsolescence. Hence, in place of the function D(t) defined in
the previous section we now have m functions Z>,(2), where Dt(t) is the
expected cumulative mean demand if 7(f) is the date of obsolescence.
Similarly, the times at which orders can be placed may depend on the date
of obsolescence. In particular, for the larger values of 7(f) it was desirable
350 DYNAMIC INVENTORY MODELS CHAP. 7

to have more times at which procurements can be made than for the smaller
f(i)- In general, we shall suppose that if f(i) is the time of obsolescence,
there are n(i) times at which procurements will be considered, and these
times will be denoted ft(i), j = 2, . . ., n(i); ft = 0, which is the time that
the first procurement is to be considered will, of course, be the same for all i,
since at and prior to ft, it is not known which value of f (i) will actually
occur. Corresponding to the Xy, <ry, Th ry defined in the previous section,
we now have a corresponding set of these for each i, i.e., Xy(i), oy(i), 2y(t),
Tj(t). Also Aj(i), Ij(i), xy(i), £y(i), L(i) replace Ah ly, Ch xy, xy, L
respectively.
As before, y\ will be taken to be the inventory position of the system at
time ft. Let W i{yx + Qx) be the expected cost if Qx units are ordered at
tune ft, f (j) is the date of obsolescence, and an optimal policy is followed
after time ft. Then

Wi(yi + Qi) — Ai(i)8± + Cx(Qx) + IxCx \tfx + Qx

— ^ [Tx(i) + r2(i) + n]j“ + TxE\_yx + Qx, Tx(i) + r2(i), n]

+ + r2(i) — ri]7Ti + IxCx}B[yx + Qx, Tx(i) + n(i), n]

00

+ L vW, <x1(i)T1(i)]Z^‘(y1 + Q1 - 4) (7-64)

di —0

and the other Zjp are defined by (7-62), (7-63), except that all relevant
parameters must now involve L
The expected cost for any choice of Qx is then
m

Y. P*Wi(yx + Qx) (7-65)

and the optimal Qx is that value which minimizes (7-65). To solve this
problem, we must essentially solve m problems of the type discussed in the
previous section, since we need the m functions 2^}(£). Then at the last
step these are joined together by (7-65) and Q* is determined. At each time
a procurement is to be considered the problem will be solved and any
additional information on the date of obsolescence obtained since the
previous solution will be incorporated.

7-8 The Prediction Problem

In Secs. 7-6, 7-7 we assumed that the distribution of demand over any time
period was Poisson and that the mean rate of demand was known as a
 REFERENCES 351

function of time. In the real world, of course, it will be necessary to use

some sort of prediction method to estimate the mean rate of demand at
future times. The mean rate of demand as a function of time will not be
known with certainty, but will be subject to errors inherent in the predic¬
tion technique used. The type of prediction technique which might be used
can vary widely. In some instances it will be based only on past demand
data for the item. In other cases there will be no historical data available
and the predictions will be based only on planned requirements. In still
other cases, the predictions will be based on a detailed economic forecast.
The accuracy with which the mean demand can be predicted will de¬
crease the farther into the future the prediction is carried. The increasing
uncertainty in the mean demand rate as one moves into the future should
be reflected in the demand distributions used in the model, i.e., the variance
should increase as one moves into the future. What we would like to use is
the distribution of forecast errors, i.e., the distribution of the difference
between the actual demand and the forecast demand. Usually, however,
it is exceedingly difficult to obtain the distribution of forecast errors.
Furthermore, this distribution may depend on calendar time, i.e., changes
as the nature of the process generating demands changes. The problem of
attempting to account for forecast errors is another difficult problem en¬
countered when dealing with dynamic models, which is usually hard to
handle easily for lack of sufficient data. It should be recognized that if one
knows very little about the future it will not be easy to make accurate
computations, regardless of how much mathematical manipulation is
performed.

REFERENCES

1. Bellman, R., Dynamic Programming. Princeton, N. J.: Princeton

University Press, 1957.
2. Bellman, R., and S. Dreyfus, “On the Computational Solution of
Dynamic-Programming Processes—X: The Flyaway-Kit Problem,”
RM-1889, The RAND Corp., April 5, 1957.
3. Bellman, R., and S. Dreyfus, “On the Computational Solution of
Dynamic-Programming Processes—IX: A Multistage Logistic-Pro¬
curement Model,” RM-1901, The RAND Corp., November 5, 1956.
4. Bellman, R., and S. Dreyfus, “On the Computational Solution of
Dynamic-Programming Processes—II: On A Cargo-Loading Prob¬
lem,” RM-1746, The RAND Corp., November 5, 1956.
6. Bellman, R., “Combinatorial Processes and Dynamic Programming,”
P-1284, The RAND Corp., February 24, 1958.
352 DYNAMIC INVENTORY MODELS CHAP. 7

6. Dreyfus, S., “Dynamic Programming Solution of Allocation Problems,”

P-1083, The RAND Corp., May 9, 1957.
7. Hadley, G., and T. M. Whitin, “A Family of Dynamic Inventory
Models,” Management Science, Vol 8, No. 4, July, 1962, pp. 458-469.
8. Wagner, H. M., and T. M. Whitin, “Dynamic Version of the Economic
Lot Size Model,” Management Science, Vol. 5, No. 1, Oct. 1958, pp.
89-96.

PROBLEMS

7-1. An isolated air base is supplied with spare parts only once every
month when a plane from the main depot arrives to replenish the
inventory. There are n parts which are stocked at the air base. Just
prior to the time the base is to be replenished, it radios to the main
depot the on hand inventory of each of the n spare parts. Let y%
be the number on hand of item i just prior to replenishment. The
plane which brings the spares can carry a volume V and the unit
volume of item i is Vi. A cost n is incurred each time a demand occurs
when the base is out of stock. Monthly demand for each of the items
is essentially Poisson distributed with mean /it. Set up as a dynamic
programming problem the problem of determining the quantity Xi of
each item i which should be loaded on the plane, so as to minimize
the expected stockout costs for the coming month.
7-2. Consider a simplified version of Problem 7-1 which involves only
three items. Assume that: V = 15 cu. ft.; vx = 5 cu. ft.; v2 = 3 cu.
ft.; ^3 = 4 cu. ft.; yx = 1; y2 = 0; yz = 2; tti = $5000, tt2 =
$2000, tz — $10,000; mi = 1, /i2 = 5, /x3 = 3. Determine the optimal
values of xh x2, and xz.
7-3. A newsboy can carry only N papers with him in the evening. There
are two papers which he sells. The first is an out-of-town paper
which costs him Ci cents each and sells for px cents. Any papers not
sold can be returned for a credit of dx cents. The other is a local
PaPer which costs him c2 cents each and sells for p2 cents. Any papers
not sold can be returned for a credit of d2 cents. Assume that if yXi
y2 are the demands for the out-of-town and local papers respectively
then yXj y2 are independent random variables with densities pi(yi)}
V*(yz)- Formulate the problem of determining the quantities xXf x2
of each of the papers to procure as a dynamic programming problem.
Solve the problem in the particular case where cx = 7, px = 10,
di = 4, c2 = 5, p2 = 7, d2 = 3. The demand for each is Poisson
 PROBLEMS 353

distributed, the mean being 7 for the out-of-town paper and 8 for
the local paper. He can carry no more than 15 papers.

7-4. Consider a flyaway kit problem in which there are only three items
to be considered. The total volume available is 11 cu. ft. The unit
volume of item 1 is 2 cu. ft., of item 2 is 3 cu. ft., and of item 3 is 4
cu. ft. The cost of stockout for item 1 is $500, for item 2 is $1000,
and for item 3 is $2000. The demand for each item is Poisson
distributed with mean 3 for item 1, 2 for item 2, and 1 for item 3.
How many of each item should be loaded in order to minimize the
expected stockout costs?

7-5. Consider the cost equation given by Eq. (7-44). Assume that a rate
of interest i is being used. Determine the discount factors which
should appear in the Ah if it is desired to have X be the discounted
cost. Include the discounted cost of the units, and work out a means
for solving the problem in this case.

7-6. Consider a situation in which the demand for an item can be taken
to be deterministic, with the cumulative demand as a function of time
being D(t). Assume that from t = 0 to t = T, n procurements will
be allowed. There is no fixed cost of making a procurement. Derive
the equations which determine the times U at which procurements
should be scheduled to arrive, so as to minimize carrying costs over
the interval 0 to T. Show that the U are solutions to

J = 2,3,.. ,n — 1
dti {j) tj-tw

Jr D(tn) = D(T) — D(tn) + Vt

where yr is the inventory at time T. Discuss the problems involved

in solving these equations. Give a geometric interpretation. Hint:
The solution to these equations need not be unique. Can you
illustrate a case where this is so?

7-7. A television manufacturer produces its own capacitors in lots. The

setup cost to produce a lot of capacitors is $200. The particular
capacitor under consideration costs $1.00 to produce. An inventory
carrying charge of I = 0.20 is used. The production schedule for
television sets over the next year calls for the following quantities of
capacitors to be used each month: 10,000, 20,000, 20,000, 15,000,
10,000, 5000, 5000, 10,000, 10,000, 20,000, 25,000, 10,000. The time
required to produce the capacitors can be ignored. Compute the
optimal times at which lots should be produced and the size of the
354 DYNAMIC INVENTORY MODELS CHAP. 7

lots, if 5000 capacitors are now on hand and it is desired to have 5000
on hand at the end of the year.
7-8. A manufacturer of actuators for rocket motors is under contract to
deliver the following quantities to a producer of rocket motors on the
first of each month for the coming year: 25, 50, 100, 175, 200, 200,
175, 150, 100, 75, 200, 150. The actuators are produced in lots and
the setup cost is $800. Each actuator costs $300 to produce. The
manufacturer uses an inventory carrying charge of I = 0.20.
Determine when setups should be made and the optimal lot sizes.
Ignore the time required to produce a lot.
7-9. Generalize the model developed in Sec. 7-6 to the case where the
procurement lead times are random variables. What assumptions
must be made?
7-10. Investigate the problems in treating the lost sales version of the
model developed in Sec. 7-6. Show that in the Zk, one must include
parameters for the size of each of the orders outstanding. Formulate
the model in the case where there is never more than a single order
outstanding.
7-11. What simplifications occur in the model developed in Sec. 7-6 if
it = 0, the terms involving r can be neglected in comparison to those
involving T + r, and the backorders term in the inventory carrying
charge can be neglected?
7-12. Demands for the three weeks of a season are deterministic and have
the values 1, 2, 2 units respectively. The total costs of the product
delivered at the beginning of each period are given by the following
table.

Weeks 12 3

No. Units
0 0 0 0
1 1.00 0.75 2.00
2 2.00 1.50 3.50
3 1.50 2.00 5.00
4 4.00 5.00 6.50
5 6.00 8.00 8.00

Inventory carrying charges are estimated at $1.00 per unit per

period, and are applied to the ending inventory each period. Calcu¬
late the optimal purchasing program or programs. Assume that
stockouts are not allowed.
 PROBLEMS 355

7-13. Parts A, B, and C may be put in a container which can hold a

maximum of 4 cubic feet of contents. The maximum cost allowable
for the inputs of the container is $4.00. Use dynamic programming
to maximize the profits from selling one container. The profits, cost,
and weight data are given below.

Cubic Feet Cost ($) Profit/unit ($)

Part A 1 3.00 5.00

Part B 1 1.00 2.00
Part C 1 o 1.50

7-14. A retailer sells an item with the following characteristics. The item
costs $5.00, has a selling price of $9.00, and a liquidation value of
$2.00. The supplier arrives with his merchandise at the start of each
week of the three week season. The initial inventory is 0 and inven¬
tory carrying charges are negligible. Demand in the three weeks is
given by the following table of probabilities, the entry in each cell
being the probability that demand is exactly equal to the number
of units indicated in the left hand column. Determine the optimal

Week 1 2 3

No. Units
1
0 i 0 "3
i 1 1
1 4 2 3
1 1
2 0 2 1

number of units to order the first week and an optimal policy for the
remaining two weeks.
7-15. A manufacturer is able to increase production in any month at a cost
of $4.00 per unit increase in the production rate (when measured in
the dimensions of units per period). Reductions in the production
rate cost $3.00 per unit decrease. It is known that demands in the
next three periods are 2, 4, and 1 units respectively. Production
costs per unit are $6.00, $10.00, and $8.00 in periods 1, 2, and 3
respectively. Inventory carrying charges are applied to ending inven¬
tory and amount to $2 per unit. Assuming that demands must be
met as they occur and that initial and final inventories are zero,
determine the least cost levels of production in the three periods.
7-16. A camera store sells its Developo Camera at $120 per unit. Its
supplier appears each Monday morning and will sell additional sets
to the firm at a specified price S or buy sets back at price B. There
356 DYNAMIC INVENTORY MODELS CHAP. 7

are no cameras in stock at the start of a two-week period and the

Developo will be obsolete at the end of the second week but have a
liquidation value of $50 per set. Given the following data, solve for
the optimal inventory policy for the two-week period. Assume that
carrying costs are negligible.

Period 1 Period 2
S(S) 80 90
B(S) 70 70

Probability Distribution of Demand

Period 1 Period 2
No. Units 2 i 0
3
3 1 1
3 4
4 X
3
1
4
5 0 1
4
6 0 1
4

7-17. A manufacturer produces a highly stylized item which will be worth¬

less at the end of the second of two periods. The purchase price for
the item in question is $9.50 the first period and $8.00 the second
period. Initial inventory is zero at the start of the first period.
The set up cost for the first period is $1.00 and increases to $4.00 the
second period. The selling price for the first period is $10.00 and the
first period’s demand is equally likely to be 0, 1, or 2 units. In the
second, period the manufacturer can charge either $10.00 or $12.00
per unit for the product. At the $10.00 price there is a probability of
0.1 that demand will be 2 units, and 0.9 that demand will be for 3
units. At the $12.00 price, the probability that one unit is demanded
is 3/8, and that 2 units are demanded is 5/8. What is the optimal
production level in the first period? What is the optimal price
and production policy for the second period? What is the level of
expected profits under optimal policy for the two periods?
7-18. Two items, A and B, have certainly known demands with the annual
demands being 1000 and 3125 units respectively. The total volume
occupied by the two items must never exceed 210 cubic feet. Item
A requires 1 cubic foot of volume, while item B requires 3 cubic
eek Ordering costs amount to $10 per order placed for either item
and inventory carrying charges are 20 percent of average inventory
value. Item A costs $10 per unit. Item B costs $20 per unit. No
stockouts are allowed to occur. Find the ordering policy which
minimizes annual costs of the two items by means of dynamic pro-
 PROBLEMS 357

gramming. Compare with the Lagrange multiplier solution which

can be obtained by the methods of Chapter 2.
7-19. Indicate how Problem 2-44 could be solved by dynamic program¬
ming.
7-20. Suppose that demand for an item is deterministic with the cumu¬
lative demand as a function of time being D(t) = t2. Three procure¬
ments are allowed between time 0 and 10. If initial inventory is 4
units and final inventory (at t = 10) is 0, calculate the minimum
cost policy by means of the equations in Problem 7-6. Does this
solution provide an absolute minimum?
7-21. Solve Problem 7-20 as a dynamic programming problem with the
length of the period being one unit of time. Compare the solution to
that obtained in Problem 7-20.
7-22. Change the example in Sec. 7-5 to allow production runs to be
scheduled at the middle of each month as well as at the beginning.
How much can management save because of the increased number
of possible production opportunities?
7-23. The value of various numbers of parts A, B, and C is indicated in the
table below, as well as the cost schedule for these parts. The total
cost cannot exceed $4.00. The total weight cannot exceed 5 pounds.
Item A weighs 3 pounds per unit. Items B and C weigh one pound
each. Maximize the value of the parts subject to the constraints on
total cost and weight. No more than three units of any item can be
acquired.

Value ($)
No. A B C

1 6 4 5
2 8 6 7
3 10 7 9

Cost ($)
No. A B C
i

1 1 1 2
2 3 2 3
3 5 4 4

7-24. A business has the average cost and revenue schedules described
below, where x1} x2, xz represent the quantities produced in periods
one, two, and three respectively, and yh y2, yz represent the quantities
demanded in periods one, two, and three.
358 DYNAMIC INVENTORY MODELS CHAP. 7

Average Cost and Revenue Schedules

Period 1 Period 2 Period 3
Average Cost 2 + 0.53! 1 -f- 0.3£2 3 + 0.6x3
Average Revenue 12 - 1.52/: 6 — O.62/2 10 - 2/3

An inventory carrying charge of $1.00 per unit will be applied to the

ending inventory of each period. Each set up costs $3.00. It is
assumed that production in each period can be carried out in time
to meet the sales, and that total production up to and including any
period must be at least as great as total sales. Calculate the optimal
price, production, and inventory policy for the three periods. Assume
that initial inventory before the first period is zero and that inventory
on hand after the three periods is without value.
7-25. The type of dynamic model presented in Sec. 7-7 can, in a simplified
form, be useful in treating other sorts of situations. Suppose, for
example, that it is not possible to predict over the planning horizon
the mean demand in each time period, and then to treat the demands
in different periods as independent random variables. Instead, the
best that can be done is to say that the mean demand will follow one
of n different curves with the probability of curve i being p*. This is
essentially the problem treated in Sec. 7-7, except for the fact that
now obsolescence may not be a factor at all. Often in practice it is
desirable to introduce a simplification of the model introduced in
Sec. 7-7 and eliminate the stochastic elements in the demand. The
only stochastic element is that which determines which demand
curve will be the one actually observed. Thus the determination of
the Wi in Eq. (7-64) involves the solution of a deterministic dynamic
programming problem, so far as the demand is concerned. Equation
(7-65) will remain unchanged. Formulate in detail this simplified
version of the problem in the case where the lead time is allowed to
be a random variable.
USES OF DYNAMIC PROGRAMMING

IN THE ANALYSIS

OF STEADY STATE MODELS

We do not what we ought;

What we ought not, we do;

And lean upon the thought

That chance will bring us through.

Matthew Arnold, Empedocles on Etna

8-1 Introduction

Dynamic programming was introduced in the previous chapter as a tech¬

nique for obtaining numerical answers to dynamic inventory problems
with a finite number of periods. The dynamic programming formalism has
also been found useful in studying certain theoretical aspects of steady
state models. In fact, it is in this way that, under suitable assumptions, the
optimality of Rr policies can be demonstrated. Furthermore, the dynamic
programming approach provides another means for obtaining the
average annual cost, both for transactions reporting and periodic review
systems. Indeed, we shall derive in this way the average annual cost
expressions for the models developed in Chapters 4 and 5. Also, expressions
for the average annual cost for transactions reporting systems when units
are not demanded one at a time can be obtained by means of dynamic
programming.
In applying the dynamic programming approach to steady state models
we imagine that there are an infinite number of periods rather than a finite
number, i.e., the system is imagined to continue to operate for all future
time. For an infinite number of stages, the recurrence relations of dynamic
programming become what is called a functional equation. It is through
the analysis of these functional equations that the various results of interest
can be derived.

8-2 The Basic Functional Equation for Periodic Review Systems

In the previous chapter where we allowed the mean rate of demand to

change with time, it was possible to include only a finite number of periods
in the formulation. Now we wish to discuss the steady state behavior of
periodic review systems, where the mean rate of demand does not change
with time. Here it will be possible and desirable to consider an infinite
number of periods, and to imagine that the system will continue operating
in steady state for all future time.
361
362 DYNAMIC PROGRAMMING USED IN SS MODEL ANAL CHAP. 8

Let us consider a periodic review system in which all demands occurring

when the system is out of stock are backordered. As usual J will denote
the review cost, A the cost of placing an order, C the unit cost of the item
(assumed to be a constant independent of the quantity ordered), I the
inventory carrying charge, t + ft the cost of a backorder, T the time
between reviews, and r the procurement lead time (here assumed to be
constant). Demands in different periods will be assumed to be independent,
and p(x\ T) will denote the probability that x units are demanded in a
period of length T.
Suppose that time J is a review time for the system. Nothing can be
done about the carrying and backorder costs incurred between time f and
the time f + r when any order placed at time f arrives. Hence, as in the
previous chapter, we shall associate with the review time f the expected
carrying costs and backorder costs incurred from time f + r to f + r + T.
Denote by /(£ + Q; T) the sum of the expected carrying and backorder
costs incurred in the period from £ + r to t+r+T, discounted to time f,
when £ is the inventory position at time f prior to the placing of any order,
and Q > 0 is the quantity ordered. Then

/« + Q) T) = ICDt(£ + Q; T) + rErti + Q; T) + *Br(£ + Q; T) (8-1)

where ICDt and tEt + tBt are the expected carrying and backorder
costs respectively incurred from f + rtof + r+ T (and discounted to
time f). Let a(T), 0 < a < 1, be the discount factor which discounts any
costs known at f + T to their present worth at time J.
We shall define £$■(£; T) to be the present worth (discounted cost) at
time f of all future costs (with the convention that no carrying or backorder
costs incurred from f to J + r are included) if an optimal quantity is
ordered at time f and at all future review times, £ is the inventory position
of the system at time f before any order is placed, and T is the time between
reviews. In order that £$•(£; T) will be finite, it is necessary to introduce
discounting and to assume that the interest rate is positive. Similarly,
2j-+r(£; T) will be defined to be the present worth at time J + T of all
future costs if an optimal quantity is ordered at time f + T and all future
times, and £ is the inventory position of the system at time f + T before
any order is placed. By convention Zf+r(£; T) will not include the carrying
or holding costs incurred between f + T and f + r + T. Then we see
from the methods introduced in the previous chapter that

T) = {J + A5 + CQ + /(£ + Q; T)
Q

+ a(T) J2 p(x; T)Z{+t(% + Q - x; T)} (8-2)

x=0
SEC 8-2 FUNCTIONAL EQUATION-PERIODIC REVIEW 363

where
if Q = 0
if Q > 0.

An important observation can now be made. It is

Z&T) = Zt+Tfc T) = T), n = 2, 3, . . . (8-3)

In other words, the present worth of all future costs at a review time, when
the inventory position is £ prior to placing an order and an optimal policy
is followed at that review time and all future reviews, is the same at any
review time. The reason is of course that we are considering the discounted
cost over an infinite planning horizon.
Therefore, if we define Z(£; T) to be the present worth at a review time
of all future costs (with the convention that no carrying or backorder costs
are included until a lead time has elapsed) when an optimal policy is
followed at the review time and at all future review times, the inventory
position of the system is £ at the review time prior to placing any order, and
the time between reviews is T, then (8-2) reduces to

Z(£; T) — min [J + A 5 + CQ + /(£ + Q) T)

Q

+ ay£ P(x; T)ZQ + Q - x; T)] (8-4)

3=0

Equation (8-4) is an equation involving the function Z(£; T). It is called

a functional equation since it involves an unknown function rather than
simply a variable. The solution to this functional equation yields Z(%; T)
and also Q*(£; T), the optimal quantity to order as a function of £. In
general, it is very difficult to solve this equation to determine #(£; T) and
q*(£. rjny La-fcer we shall show how a modified form of the functional
equation can be solved explicitly for Z(£; T). Now we shall only be inter¬
ested in saying something about the optimal ordering policy Q*(£; T).
Usually, for the analysis to follow, it will be simplest to treat Q, £, and
the demand variable as continuous. Thus, instead of p(x; T) we shall
introduce v(x; T), the demand density for a period of length T. Then to
modify (8-4) to take account of continuous variables it is only necessary
to replace p(x; T) by v(x; T) and the summation by an integral sign. This
yields

%(%', T) = min [/ + Ad + CQ + /(£ + Q) T)

Q

+ a [ v(x; T)Z(% + Q — x;T) dx] (8-5)

Jo
364 DYNAMIC PROGRAMMING USED IN SS MODEL ANAL CHAP. 8

Equation (8-5) will be the basic functional equation to be used in the

analysis of the next section.

8-3 Optimality of Rr Policies for Periodic Review Systems—

Qualitative Discussion

The functional equation approach was first employed in the study of

periodic review inventory systems in the paper by Arrow, Harris, and
Marschak [1], although the use of functional equations had appeared
earlier, for example, in the work of Masse [8]. The paper by Arrow,
Harris, and Marschak stimulated interest in the nature of optimal oper¬
ating doctrines for periodic review systems, and since its publication, a
number of papers have appeared which discuss optimal operating doctrines
for periodic review systems [2, 5, 6, 7, 9]. Perhaps the best known of these
are the two papers by Dvoretzky, Kiefer, and Wolfowitz [6, 7]. Most of
the authors made exceptionally restrictive assumptions in their analyses,
such as assuming zero lead times. Also, it was frequently assumed that the
cost of placing an order was zero. Even with these simplifying assumptions,
the mathematical analysis employed often became rather involved. The
most interesting paper from a practical point of view is the recent paper by
Scarf [9]. It allows assumptions of sufficient generality to include a
number of cases of practical importance. Interestingly enough, it is also
the simplest mathematically. In this section we would like to show how
the functional equation approach can be used to obtain information con¬
cerning the optimal operating doctrine for steady state periodic review
systems. In the next section we shall present the results derived by Scarf
[9]-
Let F(y; T) be defined by

F(y; T) =Cy+ f(y; T) + a f° v(x; T)Z{y — x; T) dx (8-6)

Jo

For a fixed value of T, F depends only on y. Then from (8-5)

A + mm F(t + Q; T)
Z(f; T) = J - Cf + min
F& T)

Note that we added and subtracted so that F could be made a function

of £ + Q only.
Let us now consider F(y; T). Clearly, when given A, the optimal operating
doctrine will depend on the shape of F(y; T) and only on the shape of
F(y; T). First of all imagine that for a fixed T, F(y; T) has the shape shown
in Fig. 8-1. Let F(E*; T) be the absolute minimum of F(y; T) with respect
SEC 8-3 Rr POLICIES FOR PERIODIC REVIEW SYSTEMS 365

to y, and R*(T) the unique point where this minimum is taken on (R* will
be a function of T). Also let r* be the value of y(y < R*) for which F(y; T)
= A+F(R*;T), i.e.,
F(r*; T) = A + F(R*; T) (8-8)

This is illustrated in Fig. 8-1.

It is clear that if £ lies between r* and i?*, say £i, then Q* = 0, i.e., it is
optimal not to order anything. The reason is that by placing an order
F(y; T)cannot be reduced below F(R*; T). However, F(R*; T) + A >
F(£; T) when £ lies between r* and R*, i.e., the expected cost would only

be increased if an order was placed. On the other hand if £ < r*, say £2,
then it will be optimal to place an order, and the quantity ordered should
be sufficient to make F(y;T) as small as possible, i.e., R* — £2 should be
ordered. Then the costs of placing an order and future discounted costs
come to A + F(R*; T) < F(£2; T). Furthermore, for Q different from
Q* = R* — £2j the expected discounted costs would be greater than A +
F(R*; T). It is also clear that if £ > i2* an order should not be placed.
However, if the system is operated optimally, £ will never be greater than
R*. What we have shown then is that an Rr operating doctrine is the
optimal one to use in this case. It is interesting to note that the same sort
of analysis shows that if A = 0, then an order up to R policy is optimal.
366 DYNAMIC PROGRAMMING USED IN SS MODEL ANAL. CHAP. 8

This is obvious intuitively, since if it does not cost anything to place an

order, and if the unit cost of the item is constant, then it. will he optimal to
place an order every time, and each time to bring t in* inventory position
to the same level R. We might note in passing that if T) is a convex
function of y, its shape will he such that an Hr operating doctrine will he
optimal.
If F(y, T) had the shape shown in lug. 8-2, an Hr policy would not. he
optimal. Instead, the optimal policy would have the form: If % < r* order
a quantity R* - £; if rf < | < rj, do not order; if r-* < £ < r*h order a

quantity R* — £, if £ > rf do not order. We suspect that the curve of

F(y; T) shown in Fig. 8-2 is not a type that will 1h: encountered in practice.
This is usually correct. However, it is easy to provide practical examples
where an Rr policy is not optimal. We shall do this below. First, however,
we would like to study in a little more detail the circumstances under which
we expect an Rr policy to he optimal.
Let us note that except for the yC term, F{y; T) is simply the discounted
expected cost for all future times when we liegin the period under consider¬
ation with an inventory position of y (after placing any order) and follow
an optimal policy for all future periods. We expect that if the inventory
SEC 8-4 PROOF THAT AN Rr POLICY IS OPTIMAL 367

position y is relatively low, the expected costs will be relatively high be¬
cause of high backorder costs in the present period. As t/ is increased the
expected costs should decrease until y is large enough so that the increase
in carrying charges outweighs the decrease in backorder costs. Then the
expected costs will begin to rise and continue rising as y is increased.
Addition of the linear yC term will not change the basic shape of the curve.
Hence we expect F(y; T) to yield a curve of the shape shown in Fig. 8-1
and not that shown in Fig. 8-2. Thus, we expect in general that, when the
unit cost is constant, an Rr policy should be optimal.
In the event that the unit cost of the item is not a constant, it is no
longer true that an Rr policy must be optimal. Now we can no longer
introduce a function F(£ + Q; T) which depends on { + Q alone. Instead,
the functional equation has the form

'A + min F(£, Q; T)

Z(g; T) = J + min
mo ]T)
where
F&Q;T) = C(Q)+P& + Q;T)
and

Hi + Q;T) = /(* + Q;T)+a T v(x; T)Z(£ + Q - x; T) dx

Jo
with C(Q) being the cost of Q units.
Imagine that all units quantity discounts are available so that C(Q) is
of the form shown in Fig. 2-12. In Fig. 8-3, we have shown what F(£, Q; T)
might look like for two different values of £. An Rr policy is not optimal in
this case, since the level up to which one orders depends on £. In both
cases it is optimal to order just enough to take advantage of a price break,
but it is a different price break in each case, and the level up to which one
orders is different.

8-4 Proof that an Rr Policy is Optimal When Hy\T) is Convex

and the Unit Cost is Constant

While the above intuitive argument suggests that normally we expect an

Rr doctrine to be optimal when the unit cost is constant, it is surprisingly
difficult to prove this under very general conditions. However, there is one
important practical case where the proof can be carried out. This is the
case where f(y; T) is a convex function of y (review Problem 2-6 for the
definition of a convex function). It is this case that was studied by Scarf.
To carry out the proof, we do not use (8-5) directly. Instead, we make
368 DYNAMIC PROGRAMMING USED IN SS MODEL ANAL CHAP. 8

use of the observation that the steady state case with an infinite number of
periods can be thought of as the limit as n approaches infinity of an n
period system. Let Zn(£; T) be the discounted expected cost for an n
period system when £ is the inventory position at the beginning of period
n before any order is placed (period n, however, comes first in time) and

an optimal policy is used for each decision. Note that the periods are being
numbered in an order opposite to their time sequence. Then

Z(f; T) = lim Zn(£; T) (8-9)

—»•»
n

and
Zn(i; T) = min [J + A5 + CQn + /(£ + Qn]T)
Qn

+a j v(x; T)Z„_i(£ + Q„ — x; T) dx] (8-10)

= / — C£ + min [Ad + FniZ + Qn] T)~\

Qn
SEC 8-4 PROOF THAT AN Rr POLICY IS OPTIMAL 369

where
Fn(£ + Qn; T) = (£ + Qn)C + /(£ + ; T)

+ a y* t>(®; T)Zn„i(£ + Qn — a?; T) d®, (8-11)

n = 2, 3,. . .
Also
Zi({; T) = min [J + A8 + CQ1 + /(£ + &; T)] (8-12)
Qi

The proof would be very simple if it followed that the Fn(£; T) were
convex if / (y; T) was convex. This is not in general true, however. The
sort of phenomenon that arises which prevents the Fn(£; T) from being
convex can be seen by studying just a simple two period case. Now from
(8-12)
ZlG; T) = J - Cl + min [Ad + Fl(i + Qi; T)]
Qi
where
Fi(r, T) = Cy +f(y; T)

and Fi(y; T) is a convex function of y, since f(y; T) and Cy are. Thus from
the discussion of the previous section, we know that an Rr policy will be
optimal for the one period case and hence

'J + Fitt; T) - C( = J + /(f; T), { > r*

Zifa T) (8-13)
J + A+ Fi(R*; T) - CJ, S < r*

Zi(J; T) is plotted in Fig. 8-4 for the case where f(y;T) is differentiable
everywhere with respect to y. Note that Zi(£; T) is not a convex function
of £. Zite; T) is continuous, however, although there is a discontinuity in
the derivative at £ = r*. Hence, we see from (8-11) that since F%(y] T)
involves Zi(£; T), F2(y; T) cannot be expected to be convex, in general.
By continuing the argument, it can be seen that the Fn(y; T), n > 1 will
not necessarily be convex.
It is the fixed ordering cost A which causes Zi(£; T)} or more generally
Zn(£; T) and Fn{y\ T), not to be convex even though f(y; T) is convex.
In fact, one can prove that when A = 0, the Fn(y; T) and F(y; T) are
convex functions of y if f(y] T) is. This result will follow from what is
demonstrated below. It can be obtained directly, however, and we ask the
reader to do this in Problem 8-2.
The above shows that since we do not expect F{y) T) to be convex, the
proof cannot be based on demonstrating the convexity of F(y; T). Indeed,
F may have a number of relative maxima and minima. As before, denote
by E* the point at which F takes on its absolute minimum. If this point
370 DYNAMIC PROGRAMMING USED IN SS MODEL ANAL CHAP. 8

is not unique, E* will denote the smallest value of y for which F takes on
its absolute minimum. Also, let R denote any point at which F takes on a
relative minimum (E* will be one such point). Suppose now that the fol¬
lowing conditions are satisfied. For each E, it is true that there are no
points y < R with F(y; T) > A + F(E; T) at which F takes on a relative
maximum. This means that the curve z = F(y; T) can cross the horizontal
line z = A + F(R) T) only once for y < R- Thus if r* < E is the point
where the curve z = F(y; T) crosses the horizontal line z = A + F(R*; T),
Q* = 0 if £ > r* and Q* = E* - f if £ < r* so that an Rr policy is
optimal. When the above conditions are satisfied, cases like that shown in

Fig. 8-2, and cases where there exist relative minima with E > E* such
that for certain ranges of £ > E* it is optimal to order up to R (illustrate
this graphically), are ruled out.
We shall prove below that if f(y; T) is convex and differentiable every¬
where with respect to y, then F(y; T) is continuous and differentiable
everywhere with respect to y (even though Z(£; T) need not be differ¬
entiable everywhere with respect to £), and that for any a > 0

A + F(a + y; T) - F(y; T) - aF'{y; D > 0 (8-14)

where a prime indicates the derivative with respect to y. However, the

condition (8-14) does indeed assure us that F(y; T) cannot have a relative
maximum at a point y < R if F(y; T) > A + E(E; T). To see this,
SEC 8-4 PROOF THAT AN Rr POLICY IS OPTIMAL 371

assume that F(y; T) has a relative maximum at yi < R with F(yi; T) > A
+ F(R; T). A contradiction is obtained on using (8-14), if we set R =
a + yh since at ylf F'(yi; T) = 0, and (8-14) becomes A + F(R; T) —
F(yi; T) > 0. Hence, the truth of (8-14) assures us that an Rr policy will
be optimal.
It remains, then, to prove that (8-14) holds if f(y; T) is convex and
differentiable everywhere. For carrying costs and backorder costs of the
type we use in this book f(y; T) is always differentiable everywhere with
respect to y when all variables are treated as continuous. We have previ¬
ously defined a function f(y) as convex if for any values yi, y2 and any a,
0 < a < 1, then

f[ayi + (1 - a)|/d < 0if(ya) + (1 — a)f(y2)

When f(y) is differentiable everywhere, an equivalent definition of con¬
vexity is that for any a > 0

/(<* + y) - f(y) - af(y) > 0 (8-15)

where f(y) = df/dy. We ask the reader to show this equivalence and to
give a geometric interpretation in Problem 8-3.
Scarf defines a modified form of convexity called A-convexity, which is
a useful concept in carrying out the proof. For any number A > 0, the
differentiable function/(y) is said to be A-convex if for all y and any a > 0

A +f(a + y) - f(y) - af(y) > 0 (8-16)

From (8-15) we see that a 0-convex function is convex. The following

properties of A-convex functions will be needed. These properties may be
easily proved and the details are left for Problem 8-4.
(a) If f(y) is A-convex, then f(y + %) is A-convex for any fixed x.
(b) If fi(y) is Ai-convex and f2(y) is A-convex, then drfi(y) + is
(0iAi + 02A2)'-convex when 0X, 02 > 0.
(c) If f(y) is A-convex, then

a / v(x)f(y — x) dx, 0 < a < 1

Jo

is A-convex when v(x) is a probability density function.

(d) If f(y) is Ai-convex, it is also A2-convex for A2 > Ai.
We shall now give the proof of (8-14). We shall first show that the
functions Fn(y; T) defined by (8-11) for the n stage approximation are
A-convex. This is done sequentially (i.e., by induction). We observe that
Fi(y; T) = Gy + f(y; T) is A-convex, since y and f(y; T) are 0-convex,
 DYNAMIC PROGRAMMING USED IN SS MODEL ANAL CHAP. 8
372

and by property (b) above F\(y] T) is 0-convex, and hence is A-convex by

property (d) above.
Now assume that Fi(y; T)}. . ., Fn-i(y) T) are A-convex for any positive
integers > L We shall then show that Fn(y; T) is A-convex. By induction
it thus follows that Fj(y; T) is A-convex for all positive integers j. From
(8-11), we see that Fn(y; T) will be A-convex if

a J v(x)Zn-i(y — x; T) dx

is. However, from property (c) above, this will be true when Zn-i{y) T)
is A-convex. To prove that Zn-i(y, T) is A-convex, we observe first that,
since by assumption Fn-i(y] T) is A-convex, the optimal ordering policy at
stage n — 1 will be of an Rr type, i.e., there exist numbers rS_i and Rt-i
such that if £ < rS-1, QS-i = Rt-i - £ and if £ > rS-i, QS-i = 0. Thus

J + A — C% -j- Fn—l(Rn-l', T), £ < Tn-\ /g

T)
J -CS + Fn-*faT), £>rS-i

We ask the reader to show in Problem 8-5 that if Fn-i(y; T) is everywhere

a continuous and differentiable function of y, then Zn-i(£; T) is a continuous
function of £ everywhere, and Fn (y\ T) is a continuous and differentiable
function of y everywhere. It is also true that Fi(y; T) is a continuous and
differentiable function of y everywhere if f(y; T) is. Hence, all the Fj(y; T)
are continuous and differentiable functions of y everywhere.
Let us now use (8-17) to show that Zn-i(£; T) is A-convex. This requires
that we show that (8-16) holds for all £ and any a > 0. Three cases must
be considered depending on the values of £, £ + a relative to rU_i.

Case 1: £ + «,£> rj_i.

From (8-17) we see that in this region Zn_i(£; T) is the sum of a 0-convex
and A-convex function and hence is A-convex.

Case £: £ < rl-i < £ + a.

Here we need to determine the sign of

A + Z*_x(£ + a]T) — Z^(£; T) — a Z4-i(£; T) (8-18)

where the prime indicates differentiation with respect to £. To do this

observe that from (8-17)

Z'n-i(£; T) = -C (8-19)

since £ < i. Also for £ < rS-i

SEC. 8-4 PROOF THAT AN Rr POUCY IS OPTIMAL 373

Zn-i(Z; T) = min {J + A+ CQn-i + /(£ + Qn-i) T)

Qn-l

+a J v(x; T)Z„-2(£ + Q»-i — x;T) dx} < J + A + Ca

+ /(£ + a] T) + a j v(x; T)Zn-At + a — x; T) dx (8-20)

since the optimal Qn~i will yield a value less than or equal to that for Q„_i
= a. However, when £ + a > r£_i

Z_i(£ + a; T) = J + /(£ + a; T)

+a j v(x; T)Z._2(£ + a — x; T) dx (8-21)

since it is optimal not to order in this case. Therefore, on combining (8-21)

with (8-20)
Zn.i(£; T) < A + Ca + Z»_i(£ + a; T)
or because of (8-19)
A + Z^d£ + <x]T) — Zn-i(!r, T) - aZ'n-i(£; T) > 0
as desired.
Case 8: £, £ + a < r»_i.
From (8-17), Z„_i(£; T) is linear and hence A-convex in this region.
This then completes the proof that Zn-1(£; T) is A-convex if F*-j(£; T)
is. Hence Fn(y, T) is A-convex if Fx(y, T), . . ., T) are A-eonvex.
Thus by induction the FAy; T) are A-convex for all positive integers j.
To prove that F(y;T) is A-convex, we must pass to the limit as n —» «.
It is clear that Z*(£; T) approaches Z(£; T) as n ->«., and hence Fn(y, T)
approaches a unique limit F(y; T) (proof?). Furthermore, for fixed values
of y, a, T the sequence

(7.(2/, a; 2") = A + F„(y + a; T) — Fn(j/; T) — aFh(y> T) >0

so that 0 is a lower bound on the sequence and hence (how do we know that
F' exists?)
0(y, a; T) = lim Gn(y, ID
n—>«

= A + F(y + *;T)- F(y; T) - aF'(y, T)>0

for any y and any a > 0. Thus F(y; T) is A-convex, and an Rr policy is
374 DYNAMIC PROGRAMMING USED IN SS MODEL ANAL CHAP. 8

optimal for the steady state periodic review system if f(y; T) is a convex
function of y and the unit cost is constant.
Geometrically, the fact that F(y; T) is A-convex rules out cases such as
those shown in Fig. 8-2 where an Rr policy is not optimal. It does not,
however, rule out behavior such as that shown in Fig. 8-5 where an Rr
policy is optimal.

8-5 Use of Dynamic Programming to Compute Optimal Values

of R and r

The solution to the functional equation (8-5) will yield Z(£; T) and
Q*(£; T). If an Rr policy is optimal, Q*(£; T) will yield the values of R*
and r*, since Q* = 0, £ > r*, and Q* = R* — £ for £ < r*. Hence Q* will
have the form of the curve shown in Fig. 8-6. Thus r* is the point where
Q* changes discontinuously from 0 to R* — r* and the magnitude of the
jump then determines R*.
A possible procedure for computing numerically 2?* and r* for a given T
is to approximate the infinite period system using n periods, as was done in
the previous section in proving the optimality of Rr policies. The appropri¬
ate recurrence relations for the n period case are given by (8-10) and (8-12).
Then QJ(£; T) is used as an approximation to Q*(£; T), and R* and r* can
 SEC. 8-5 USE OF DYNAMIC PROGRAMMING TO COMPUTE R AND 375

be determined from Qn(%) T) as indicated above. The n period system can

be solved numerically using a digital computer. The value of n used might
be something like 100. It is fairly easy to decide whether or not Q%(£; T)
is a close enough approximation to Q*((; T). If there has been only a
negligible change in Q£(£; T) for a number of periods, then it should be
satisfactory to use Q£(£; T) for Q*((; T). In making the numerical compu¬
tations, one would, of course, treat all variables as discrete rather than
continuous.
In order to make the computations, it is necessary to specify a value of
the interest rate, that is, it is necessary to specify a. For sufficiently

small values of the interest rate the results will be independent of the inter¬
est rate. However, when the interest rate becomes large, the computed
results may depend on the value of the interest rate selected. It will be
shown later that the results computed using the n stage approximation
will most closely approximate the R* and r* that would be obtained using
the average annual cost expression of Sec. 5-9 if the interest rate is taken
to be zero.
To determine T*, the computations would have to be made for different
T values. In order to select the optimal T it is necessary to have some costs
to compare. Now if the n used in making the computation is sufficiently
large, then the minimum expected cost Z»(£; T) will be essentially the same
for any £ lying between R and r since any one of these values are allowable
376 DYNAMIC PROGRAMMING USED IN SS MODEL ANAL CHAP. 8

at a review time when the steady state system is operated. Thus T* can be
determined by finding the T which minimizes Zn[R*(T); T]. If the reader
does not see why any Zn{£; T) for between r and R can be used, a more
rigorous justification will be presented later.
It would, of course, be an impossibly laborious task to compute manually
R*, r*, and T* by the method outlined above. However, it is quite feasible
to make the computation using a large scale digital computer. The method
outlined above will, in general, require more computer time than the
computational procedure suggested in Sec. 5-9. It does, though, provide
an alternative procedure for computing R* and r*.

8-6 The Functional Equation for Transactions Reporting Systems

The functional equation approach can be used to study transactions report¬

ing systems as well as periodic review systems. In fact, the functional
equation approach provides a way to study cases where the number of
units demanded per demand is a random variable, and cases where the
density function describing the time between demands is not exponential.
Not nearly as much has appeared in the literature discussing the functional
equations for transactions reporting systems as for periodic review systems.
In fact, the papers by Beckmann [4] and Beckmann and Muth [3] are the
only ones which discuss the subject.
Let us then consider a transactions reporting system with constant
procurement lead times, in which demands occurring when the system is out
of stock are backordered. Define Z(£) to be the discounted expected cost
over all future time when an optimal policy is followed at all future times,
the time origin is immediately after a demand occurs, and £ is the inventory
position at the time origin after taking whatever action is appropriate (i.e.,
after the placing of an order if it is desirable to do so). As above, we do not
include in Z(£) any backorder or carrying costs incurred before time r
(r being the lead time). To obtain the functional equation, we write Z(£)
as the sum of the expected discounted costs occurring until the next demand
and those occurring after the next demand.
We shall assume that the unit cost of the item is constant; A, I, C, 7r, t
will have their usual meanings. In formulating the functional equation for
^(£)j we shall allow the number of units demanded per demand to be a
random variable, and also allow the probability that a demand occurs in
the time interval t to t + dt following a demand to be a function of t. A
variety of probability distributions introduced in Chapter 3 will be used
here. Let g(t) dt be the probability that the time between two successive
demands lies between t and t + dt, d(j\t) the probability that j units are
demanded when a demand occurs if a time t has elapsed since the previous
 SEC 8-6 FUNCTIONAL EQUATION-TRANSACTIONS REPORTING 377

demand (note that we can allow the number of units to depend on the time
since the last demand), and Vn(t) the probability that precisely n units are
demanded in a time interval of length t following a demand.
Suppose that a demand occurs at time 0, and after taking whatever action
is appropriate, the inventory position of the system is £. It will be supposed
that it is never optimal to have £ < 0. Assume also that the next demand
occurs between t and t + dt. The probability of this is g(t) dt. We would
like to compute the expected cost (discounted to time 0) of carrying inven¬
tory and of backorders in the time interval r to r + t. Imagine that the
demand at time t is for j units. The probability of this is d(j\t). We shall
perform the computation in two parts, depending on whether r > t or
r < i.
In making the computations, we must, as in the periodic review case,
introduce a discount factor in order that the cost will be finite. Here there
is no natural period, and hence it is convenient to use continuous discount¬
ing. By this we mean that if we have a sum S (t) at time t, and if the interest
rate is i, then at tune t At, the sum S(t -j- At) which we have will be given
by
S(t + At) = (1 + iAt)S(t)
for small At, or taking the limit as At —> 0 and solving the resulting differ¬
ential equation, we obtain

S(t) = S(0)eu or S(0) = S(t)e~u (8-22)

Thus the present worth 3(0) at time 0 of a sum S(t) at time t is S(t)e~u so
that the discount factor is e~u.
Consider first the case where r > t. The probability that x demands
have occurred between time 0 and f, r < f < t + t, is
X

d(j 11) Vaj(f t) (8-23)

;=1

The discounted expected cost of the unit years of shortage incurred be-
r and r + t is, when averaged over t < r
- X —

*i;p\ 2] o J-1
d(j\t)Vx—j(£ 0
J
Q(f) df dt (8-24)

It is actually more difficult to compute the expected fixed backorder cost

incurred between r and r + t because of the continuous discounting. We
cannot simply compute the expected number incurred between r and r + t,
then multiply by x and a discount factor. The backorders must be costed
and discounted at the point of time when they occur. To do this we define
Wk(n + fc; t) dt to be the probability that in a time interval of length t
378 DYNAMIC PROGRAMMING USED IN SS MODEL ANAL CHAP. 8

following a demand n units are demanded, and between t and t + dt a

demand for k units occurs. It is then easy to give an expression for the
discounted expected fixed cost of backorders incurred between r and r + t.
When averaged over t < r, it is
/*r r T-\-t ca co x

n
* I
Jo Jr
/

T-\-t
E E E kW^x - j + k> r - 0 d(j\t) g(t)dt dt +
k=ix=£y=i

e_irE E Ex
1

x = lfc = £~x3=1
°° X

— y+*;r — od(j\t)g{t)d{at

The second term in (8-25) arises because x may be less than £ while x + k
is greater than £.
The Wk{n + fc; t) can be computed recursively just as the Vn(t). It is
clear that

Wk{n + fc; t) = [‘ V d(k\t - f)g(i - f)TFm(n; f) df, n > 1 (8-26)

m= 1

W*(fc; 0 = d(fc|0ff(0 (8-27)

The expected discounted cost of carrying inventory from r to r -M is,

n
when averaged over t < r
r+t £

£ (€ -
x=l
r-X -
E dO'IO^-i(r - 0
Li== 1
“I

0(f) # dt (8-28)

We must compute now the costs corresponding to those above for the
case where t>r. From r to t the on hand inventory is £ (and no backorders
exist for this time interval since by assumption £ > 0). However, from t
to r + t} the computation must be made as above. The discounted ex¬
pected cost of the unit years of shortage incurred from t to r + t averaged
over t > r is

* JtT+t E <* - ^e_i{Ed(M- f)] s(t) df dt (8-29)

Similarly, the discounted expected fixed cost of backorders incurred

from t to t ~f~ t averaged over t > r is

* [ E O' - &e~u d(j\t)g(t) dt

Jr i-(

n r-f-f w °° X

«_,T E E E kWk(x - j + k; f - t) d(j\t)g(t) df dt

kX\7=if=i
SEC 8-6 FUNCTIONAL EQUATION-TRANSACTIONS REPORTING 379

£-1 oo x
+ x
E E
a; = l &= £—x
(k + x — £)TF*0 — j + k; t — t)

d{j\t)g(t) dt dt (8-30)

The first term arises because of the possibility of incurring fixed backorder
costs when the demand occurs at time t.
The expected carrying cost from time r to t + t averaged over t > r is
ft r« rT+t £
J e-«m) dtdt + icj Jf £ (f - *)«r<r

d(j\t) - <)J g(t) dt dt (8-31)

The discounted expected cost of carrying inventory and of backorders

from r to t + t averaged over all t can then be written

m = ICD(£) + tE(Q + *B(Q (8-32)

where

D® = E (f - *)*(*) (8-33)
x=0

5(|) = ^ (x - £)ri(x) (8-34)

x-e

E® = £ £ (* + ft - £)*(M) + £ Z>(*, *)
x = 0 k—£—x =l

(8-35)
+ Z(*-0«(O.*)
a;-=£

and

= [f
Jo Jt
e *'f ^d0'|i)7*_y(f — f) g(t)dtdt
Lj_i J

nr-H r a;
+ e-if 2 d(i|i)k'I_y(r - i) g(«) df di, a; = 1, 2, 3 .

(8-36)

- e~ir]g(t) dt (8-37)
380 DYNAMIC PROGRAMMING USED IN SS MODEL ANAL CHAP. 8

and

6(k,x) = Jo J^ e_<r[s Wk(x - 3 + k'y f - 0 dm] 9® dt

+ J J w*(x - 3 + f - t) d(j\t) J g(t) dt dt,

fc = 1, 2, . . .

x = 1, 2, . . . (8-38)

0(O:,x) = J e~u d(x\t)g(t) dt, x = 1, 2,. . . (8-39)

Having computed the expected discounted costs /(£) of carrying inven¬

tory and of backorders incurred from r to r + t and averaged over t, we can
then immediately write down the functional equation for Z(£). It is

Z(& =/(£)+ (” e-*‘d(i|0g(0{nun[A5 + Cy + Z{% + y - j)]} dt

Jo jTi y>0

Thus, since Z{£ + y — j) is independent of t

Z(£) = /(£) + a X) p(i) {min[A5 + Cy + Z(£ + y - j)]} (8-40)

fTi v>o
where

e~u d(j\t)g(t) dt = ap(j) (8-41)

£
and

a= dt (8-42)

Hence

Y, p(i) = 1
i

and the pO*) can be thought of as probabilities. Note that a is simply the
expected value of the discount factor over the time until the next demand
occurs.
The functional equation (8-40) differs from other dynamic programming
problems that we have studied in that the minimization over y must be
carried out before summing over j, i.e., y* is a function of £ — j, not of £.
The intuitive reason for this is clear. The optimal quantity to procure will
depend on the number of units demanded when the next demand occurs.
 SEC. 8-8 OPTIMALITY OF Rr POLICIES 381

The solution to the functional equation will yield Z(£) and y*(£ — j). The
function y*(f - j) provides the optimal operating doctrine which we are
seeking.

8-7 Optimality of Rr Policies for Transactions Reporting Systems

Intuitively, one would expect that an Rr policy would be optimal for trans-
actions reporting systems of the type discussed in the previous section.
We here wish to show that, normally, an Rr policy will be optimal. From
the results of Sec. 8-4 we know that an Rr policy will be optimal if /(£) is
convex when the variable f is treated as continuous and the unit cost is
constant (proof?). To see that /(£) will normally be convex, note that
ICD(f) + tB(£) can be written

2
x=0
~ x)v(x) (8-43)
where
7C(f - *), £ - x > 0
KS - x) (8-44)
fi(x — £), f — x < 0

Thus Z(f - x) is a convex function of f for a given x (plot Z(f - a:) as a

function of £). Now since y{x) > 0, it follows that 7CZ)(£) + tB(£) is a
convex function of £. This is to be proved in Problem 8-7. The term rE(Q
will not necessarily be convex for all f. In Problem 8-7, we ask the reader
to explain why this is the case. However, 7ri?(£) will usually be convex for £
in the range of interest. Therefore, we normally expect/(£) to be convex,
and hence an Rr policy will be optimal, provided that the unit cost is
constant.
In the event that units are demanded one at a time and transactions
reporting is used, it is possible to show that an Rr policy is optimal under
much more general conditions than those discussed above. For this case,
an Rr policy will be optimal provided only that the expected carrying and
backorder costs incurred between r and r + t depend on £ alone and hence
can be written /(£). This will be true independently of the shape of /(£).
Here, it need not be assumed that the unit cost of the item is constant.
An Rr policy will be optimal even if quantity discounts are available.
The proof of these facts is left for Problem 8-20.
It is possible to evaluate numerically Z(£) and ?/*(£ — j), and hence
R*, r*, using the same sort of procedure suggested in Sec. 8-5 for a periodic
review system. We approximate the infinite stage process by n stages, so
that (8-40) is replaced by ’
382 DYNAMIC PROGRAMMING USED IN SS MODEL ANAL. CHAP. 8

00

Zn(Q = /(£) + a p(j) {min [A 6 + Cy + Zn-i(£ + y - j)]} (8-45)

f=i v>°
Zi®=m (8-46)
Then ?/£(£ — j) is taken to be y*(% — j). Since for the transactions reporting
system being studied, a stage refers to the time between two successive
demands, n will in general need to be quite large before yt(£) — y*(£)>
Here, the value of n required can vary widely depending on the mean rate
of demand. An n of several thousand may be required, or, for very low
mean rates of demand, an n of less than one hundred may be satisfactory.
For medium or high mean rates of demand, many more stages will be
required than for a periodic review system with the same mean rate of
demand. On the other hand, it is not necessary here to run the problem
many times to find the optimal T, as required in the periodic review case.
The task of computing the optimal operating doctrine for a transactions
reporting system in this way could be quite time consuming even for a
large digital computer; however, it can be accomplished. Later we shall
introduce a somewhat simpler procedure for determining R* and r* which
makes use of the average annual cost expression. This simpler procedure
still requires the use of a large scale computer, however.

8-8 Explicit Solution of the Functional Equation when a Poisson Process

Generates Demands and Units are Demanded One at a Time

It is an interesting fact that the functional equation (8-40) can be solved

explicitly for Z(£). We shall illustrate here how this can be done in the
case where a Poisson process generates demands and units are demanded
one at a time. In the next section, the more general case will be considered.
The explicit solution of the functional equation makes it possible to also
obtain an explicit expression for the average annual cost, thus establishing
a link between the functional equation approach and that used in Chapter
4. For the case of a Poisson process generating demands with units being
demanded one at a time, we shall obtain the average annual cost in this
way and show that it is precisely the same as that obtained in Chapter 4.
When units are demanded one at a time, the functional equation (8-40)
can be written

Z(Q = /(£) - aCf + min {aA5 + a(£ + y)C + aZ(£ + y - 1)} (8-47)
y>0
Now an Rr policy must be optimal in this case. We can write
R* = r* + Q*. The quantity procured will always be Q*. Let r* + 1 be
the largest value of £ such that in (8-47), y* > 0. Also, write
F(x) = Cx + Z(x — 1) (8-48)
 SEC 8-8 EXPLICIT SOLUTION FOR POISSON DEMANDS 383

and let minx F(x) be taken on at x = R* + 1. Then

Z(r* + 1) = fir* + 1) + aCQ* + aA + aZ(R*) (8-49)

For £ > r* + 1, no order will be placed when the next demand occurs.
Thus
Z(r* + 2) = f(r* + 2) + aZ(r* + 1) (8-50)
Z(r* + 3) = fir* + 3) + aZ(r* + 2) (8-51)

Z(R*) = fiR*) + aZiR* - 1) (8-52)

When the system is operated optimally, the inventory position will never
get above R*, since one never orders above R*. Now substitute (8-49) into
(8-50), the resulting equation into (8-51), etc. This yields

Z(r* + fc) = J ak-fir* + j) + ak[A -f CQ* -f ZiR*)],

y-i
k = 1, . . ., Q* (8-53)
In particular
Q*
ZiR*) = £ o9*~if{r* + j) + ad*[A + CQ* + ZiR*)]
y=i
or

Z(R*) = F£ + J) + aQ*iA + cq*) J1 (8-54)

1 u U=i

Here we have obtained an explicit expression for ZiR*). Substitution of

(8-54) into (8-53) yields explicit expressions for each Z(r* + k), i.e.,

Zir* + k) =

L a^r* + 3) + [x + CQ'+f) aP-ifir* + j) 1

j=i x a l y-i J

= V~Qi\A + CQ* + Sa'’/(r*+j) + £ a->fir*+j)],

L y-i i-iN-i -I

& = 1,. . ., Q* (8-55)

Recall that Z(%) can be interpreted as the discounted value of all future
costs (with the convention that no carrying or backorder costs are incurred
until a lead time has elapsed) when the inventory position is £ immediately
after a demand, and an optimal policy is followed for all future times. We
have noted previously that an Rr policy is optimal, and hence, when the
384 DYNAMIC PROGRAMMING USED IN SS MODEL ANAL CHAP. 8

system is operated optimally, £ will only have one of the values

r* + 1, . . ., r* + Q* = jB* For each of these £ values, we have in (8-55)
an explicit expression for Z(£).
We shall now define a new function 3(£; i2, r) as follows: 3(£; i2, r) is the
discounted expected value of all future costs evaluated immediately after
the occurrence of a demand (time 0) when an Rr policy is used to operate
the system with the reorder point being r, and Q = R — r the order quan¬
tity (note that R and r need not be the optimal values of these quantities),
and £ is the inventory position at time 0 after an order has been placed if it
is called for. As usual, it is assumed that no carrying or backorder costs
incurred until a lead time has elapsed are included, i.e., until time r. The
function 3(£; R, r) then satisfies the functional equation

m R, r) = /(£) + a3(£ — 1; 22, r), £ = r + 2, . . . , R

(8-56)
d(r + 1; R, r) = aA + aCQ + f(r + 1) + ad(R; R, r)

since, when this Rr doctrine is used, the only possible values of £ are
r + 1,. . ., R. Thus equations (8-49) through (8-52) are satisfied if Z is
replaced by % and the asterisks on r, R, and Q are omitted. Hence an
explicit solution for &(£; R, r) if given by (8-55), if Z is replaced by 3* and
the asterisks are omitted on r and Q. Note that

Z(r* + h) = d(r* + fc; R*} r*), k = 1, . . ., Q*

It will now be shown how to relate 3(£; R, r) to the average annual cost
3C(Q, r). To do this, consider an ensemble of identical systems in statistical
equilibrium. If over any given interval of time of length At, we determine
the ensemble average of the costs incurred over this time interval and then
divide by At, the result obtained, when expressed in the dimensions of
dollars per year, will be simply the average annual cost X (Q, r), for any
At > 0, however small. Furthermore, the rate of incurring costs X which
is obtained will be independent of time. Now consider the quantity £
defined by

£ has the following interpretation. If we begin observation of the ensemble

of systems at a given point in time, and for each system in the ensemble,
we continuously determine the total costs, discounted to time t, incurred
since we began observation, then £ would be the limit of the ensemble
average of these costs as the time over which observations are made be-
 SEC 8-8 EXPLICIT SOLUTION FOR POISSON DEMANDS 385

comes infinite. Equation (8-57) is the familiar expression relating the

present worth to average annual cost.
We must now relate £ to the 3(£; R, r). Recall that d is the discounted
expected cost for a single system, given that we begin observation imme¬
diately after a demand occurs, and the inventory position is £ at the time
observation is begun, with the further restriction that no carrying or back¬
order costs are included until a lead time has elapsed. From this definition,
the reader might feel that
1 Q
£ = 7j 2 + k’’R> r)
since as shown in Chapter 4, the probability that the inventory position
will be r + k at an arbitrary point in time is 1/Q, so that the fraction of the
systems in the ensemble which have the inventory position r + k when
we began observation should be 1/Q. This is not quite correct, however,
because of the way in which d(r + k; R, r) was defined. Recall that in the
definition of d(r + k; R, r) no carrying or backorder costs incurred between
times 0 and r were included. On the other hand, £ includes these costs.
Furthermore, S(r + k; R, r) is defined only at points in time immediately
after the occurrence of a demand and not at an arbitrary point in time.
However, as the interest rate becomes smaller and smaller, both of these
differences disappear, since then the carrying and backorder costs incurred
from 0 to r, and the correction to the cost arising from starting at an arbi¬
trary point in time rather than immediately after a demand becomes a
negligible fraction of the total discounted cost. This discussion then shows
that X should be given by

X= lim ^ 3(r + k; R, r) (8-58)

0 ^ ft- 1

When we compute
lim ib(r + Jc; R, r)
i—>0
it turns out that the same result is obtained for each k = 1,,Q. We
ask the reader to prove this in Problem 8-11. The intuitive reason for this
is that as i —> 0, the contribution of the costs which depend on the starting
inventory position become a negligible part of the total discounted cost.
Because of this we have the simpler result

3C = lim id(r + k; R, r), k = 1,. . . , Q (8-59)

i—*0
or in particular
X = lim id(R; R, r) (8-60)
%—M)
386 DYNAMIC PROGRAMMING USED IN SS MODEL ANAL CHAP. 8

Using (8-60), we shall now compute 3C when a Poisson process generates

demands, and show that precisely the same result is obtained as was ob¬
tained in Chapter 4. When a Poisson process generates demands,
g(t) = \e~u and from (8-42)

a = X = r~~—: (8-61)
Jo X+*
Thus
lim a0~> = 1, j = 0, ... ,Q (8-62)
i—>0
and
% 1 X
(8-63)
SS 1 - o« “ SS - . Q . Q(Q + 1) , _ Q
•L J- T •> 1 o\ 2 * ' * •

where in the denominator, we have expanded aQ by the binomial theorem.

Hence from (8-54)

3C = lim *(B; B, r) = £ A + XC + £ V /(r + fc) (8-64)

♦-♦o y y 3bil
where in fir + j), we set % = 0.
It remains to evaluate /(r + j). To do this we must first compute rj(x)
and 6(k,x). When units are demanded one at a time and a Poisson
process generates demands, d(l\t) = 1, d(j\t) =0, j ^ 1, V*(£) =
p(x; X£), W*(n + fc; t) = 0, fc ^ 1, and TFi(n + 1; £) = Xp(n; Xf), since
TFi(n + 1; f) dt is the probability that n demands occur in a time £ following
a demand and a demand occurs between t and t + dt.
From (8-36), we see on making the substitution of variable T = f — t,
for fixed £, that when £ = 0

rj(x) = X ff
J0 Jr—t
p(x — 1; XT)e~u dT dt + X f
Jt Jo
f p(x — 1; \T)e~u dT dt,
x = 1, 2, . . . (8-65)

We now change the order of integration of the first term by use of the
familiar geometrical analysis employed in Chapter 3. This yields for the
first term

X / j p(x - 1 ;XT)e-udtdT
JO Jr-T

which when combined with the second term yields

ri(x) = X j j p(x _ ij \T)e~u dt dT

JO Jr-T
SEC 8-8 EXPLICIT SOLUTION FOR POISSON DEMANDS 387

We can now readily evaluate this integral explicitly. It is

dT = -v{x] Xr), x = 1, 2,. . . (8-66)

From (8-37) we see that

v(0) = lim - 7 }—(i-j-X)r — p—Xr

■ p(0; Xr) (8-67)
t—>0 t- U' i A. A_

Since Wi(x + 1; t) = XF*(f) = Xp(x; Xf), we see on comparison of (8-36)

and (8-38) that
8(1, x) = p(x; Xr) (8-68)
From (8-32) it immediately follows that
Tfl co

f(r + k) = — (r + k *— £)p(#; Xr) + x ^ p(x; Xr)

*S=0 jr = r+A

+ r X) (x ~~ r — *)p(«; Xr)
A x=r-Hfe
(8-69)
rp *
= Y(r + fc-M) + T ^ p(a;; Xr)
x=r-f&

+ r (# + /C) V] (^ — r ~ k)p(x; Xr)

A x»r+fc

Hence, according to (8-64), 3C should be

3C = + XC + (r + fc - M) + §£ P(r + fc; Xr)

0 '

+ ^ X) 2 & ~ r — k)p(x; Xr)

" fc=l i=r+4

= ^A + XC' + 7C^ + | + r — + ^ P(u; Xr)

+ ^ 2 2>-*)p<*Xt)
V y=r+l

However

2/)pfo M = X) Z) (w - *0p(w + i; Xr)

ymmV X*=y j = 0 W =V

00
= ^ (w - v)P(w; Xr)
388 DYNAMIC PROGRAMMING USED IN SS MODEL ANAL CHAP. 8

on making the change of variables y = v + j, w = x — j. Thus we see

that X is indeed equivalent to the average annual cost expression obtained
in Chapter 4. Of course, here X includes the average annual cost of the
units XC.

8-9 Explicit Solution of the Functional Equation in the General Case

It is possible to solve (8-40) explicitly for Z(£) in the general case, but the
task is somewhat more difficult than when units are demanded one at a
time. Assume that an Rr policy is optimal. Then let r* be the value of the
inventory position such that if £ < r* after a demand occurs an order will
be placed. Assume that when an order is placed the inventory position is
brought up to a level 22*. If the inventory position is r* + 1 immediately
after a demand occurs, it is certain that an order will be placed after the
next demand occurs. Similarly, if the inventory position is r* + 2 imme¬
diately after a demand, it will be r* + 1 after the next demand if only a
single unit is demanded and will be 22* if two or more units are demanded,
since in this case an order will be placed. Thus
Z(r* + 1) = /(r* + 1) + a[A + Z(22*)]

+ aC£ (R* -r*+j- l)pU) (8-71)

3 = 1

Z(r* + 2) = /(r* + 2) + aP{2)[A + Z(R*)] + ap(l)Z(r* + 1)

00

+ (R* -r*+j- 2 Mj) (8-72)

i-2

Z(r* + 3) = fir* + 3) + «P(3)[A + Z(R*)] + ap(l)Z(r* + 2)

+ ap(2)Z(r* + l)+aCj2 (R* -r*+j- 3 )p(J)

J=3

(8-73)

Z(R* - 1) = f(R* - 1) + aP(R* + Z(R*)]

+ ap(l)Z(R* - 2) + ap(2)Z(R* - 3)
+ . . . + ap(R* - r* - 2)Z(r* + 1)
00

+ aC 2 (j + l)p(j)
(8-74)
SEC. 8-9 EXPLICIT SOLUTION IN GENERAL CASE 389

Z(R*) = f(R*) + aP(R* - r*) [A + Z(R*)]

+ ap(l)Z(R* - 1) + ap(2)Z(R* - 2)

+ ■ • • + ap(R* - r* - 1 )Z(r* + 1) + aC ^ 3V(j)

j = R*-T*

(8-75)
where P(j) is the complementary cumulative of p(j).
Let us now solve explicitlyfor Z(R*) by substituting (8-74) into (8-75),
then Z(R — 2) into the resulting expression, etc. On substitution of (8-74)
into (8-75) we obtain

Z(R*) = f(R*) + ap(l)f(R* - 1) + a[P(R* - r*)

+ ap(l)P(R* - r* - 1)][A + Z(R*)] + [o*p*(l)
+ ap(2)]Z(R* - 2) + [a2p(l)p(2) + ap(3)]Z(R* - 3)
+ . . . + [a2p(l)p(R* — r* — 2)
+ ap(R* - r* - 1 )]Z(r* + 1)

+ ac\ J2 jp(j) + ap( 1) (j + 1) pO-)! (8-76)

Li = i2*-r* j = R*-r*-l J
On continuing the substitution, one quickly determines that the coefficient
of /(i?* — h) will be the probability of having h units demanded in 1 or 2
or, . . . , or h demands, and with a discount factor included for each of the
demands. Thus, if XQi) is the coefficient of /(2?* — h)

X(l) = ap(l), X(2) = a2p2( 1) + a?>(2),

X(3) = a*p*(l) + 2a2p(l)p(2) + ap(B)
or in general

X(h) = n [ap(J)] Ui (8-77)

U\\ U2I . . .Uhlj = 1

where the Uj are non-negative integers satisfying

h
=h
3 = 1

It is also possible to generate the X(h) in a different way. Define the

quantities XnQo) recursively as follows

X„(fc) = ap(j)Xn-i(k - j), n = 2,3,; k = 1,2,... (8-78)

390 DYNAMIC PROGRAMMING USED IN SS MODEL ANAL CHAP. 8

and
Xi Qc) = apik), k = 1, 2,. . .
and
Xn(0) = 1, n = 1, 2,. ..
The quantity X«(fc) is almost the n-fold convolution of ap(k) as defined in
Sec. 3-9. It differs, however, in that we sum from j = 1 in (8-78) rather
than from j = 0, and that we use the artificial convention Xn(0) = 1,
whereas in reality, ap(0) = 0, because when a demand occurs at least one
unit must be demanded. Intuitively, Xn(k) is the probability, suitably
discounted, that k units are demanded in 1 or 2 or ... orn demands. Thus

X(h) = Xh(h) (8-79)

The final coefficient of [A + Z(R*)] will be

a X(h)P(R* - r* -h)
h=0

where X(0) = 1. This is the probability, suitably discounted, that 1 or

2 or, . . ., or R* — r* demands are required before the inventory position
moves from R* and crosses (or reaches) the value r*. The coefficient of C
will be

R*— r* — 1 r- 00 —|

5(a) = a £ X(h)\ £ O' + h)p(J) (8-:80)

A=0 Lj r*—A. J

and is simply the expected size of the order suitably discounted.

With the above information, we can then write

Z(R*) = £ X(h)f(R* -h) + CQ(a)

h=0

~R*—r*—l
+ 0 £ X(h)P(R* — r* — h) [A + Z(R*)] (8-81)
- h=0 J
or

Z(R*)
itT—r- —i

aA £ X(h)P(R* -r*-h) + CQ(a) + V

2 X(h)f(R* - h)
h=0 ^=0
‘ -
R*-r*-l

1 -a £ X(h)P(R* -r* -h)

SEC 8-10 EXPLICIT SOLUTION-PERIODIC REVIEW 391

We leave for Problem 8-12 the details of obtaining explicit expressions

for the other Z(r* + k). Problem 8-12 also asks the reader to show that
limi_*o [iZ(r* + &)] is the same for each k, k = 1, . . ., Q, as would be
expected.
To relate what we have obtained here to the average annual cost X(R, r)
we introduce the function 8(£; R, r) defined in the previous section. We
note that equations (8-71) through (8-75) still hold if we remove the aster¬
isks on R and r and replace Z(r* + k) by d(r + k; R, r). Thus an explicit
solution for 3(i?; R, r) is given by (8-82) when the asterisks on R and r are
deleted. Furthermore, lim,-*o [i3(r + fc; 1?, r)] is the same for each
k, k = 1, . . ., Q. Therefore, from the discussion of the previous section,
it should be true that
X(R, r) = lim %b{R\ R, r) (8-83)
i—>0
Here, we have a means for obtaining an explicit expression for the average
annual cost which could be used to determine the optimal values of R and r.
For any particular process generating demands, such as a stuttering Poisson
process, it is rather cumbersome to evaluate X(R,t) directly, although it
can be done. In Problem 8-21 we ask the reader to derive X(R, r) from
(8-83) in the case where the number of units demanded in any time interval
has a stuttering Poisson distribution. Problem 8-22 then requires that
X(R, r) in this case also be derived by computing the expected cost per
cycle and multiplying by the average number of cycles per year, and shown
to be the same as the expression obtained from (8-83). It would not be
easy to compute manually numerical values of R* and r* using X (R} r).
However, the computation can be carried out quite easily on a digital
computer. The determination of R* and r* on a digital computer using
X (RjT) would normally require considerably less time than the solution of
the n stage dynamic programming problem suggested in Sec. 8-7.

8-10 Explicit Solution of the Functional Equation for the Steady State
Periodic Review Model

The material developed in the last few sections suggests that it should also
be possible to solve explicitly the functional equation for a periodic review
system. We shall show that this can indeed be done. This in turn will
provide another means for computing the average annual cost of an
(R} r, T) model. It will, of course, be the same expression that was obtained
in Chapter 5. In order to solve explicitly the functional equation, it must
be formulated so that the £ in Z(£; T) is the inventory position at a review
time after any order has been placed, not before the placing of an order.
392 DYNAMIC PROGRAMMING USED IN SS MODEL ANAL CHAP. 8

Let us then define Z(£; T) to be the discounted expected costs over all
future time when an optimal policy is followed for all future times, and
we begin at a review time with the inventory position being £, after any
order has been placed. As usual, we do not include any carrying or back¬
order costs incurred until a lead time has elapsed. Then it is clear that the
functional equation becomes

Z((; T) =M)T)

+ a ^ p(x; T){min [Ad + Cy + Z(£ + y — T)]} (8-84)

x=0 y>0

which has precisely the same form as (8-40). We have omitted the J which
appears in (8-4), since it has no influence on the computations for a given T.
Now, of course, p(x; T) is the probability that x units are demanded in a
period, and /(£; T) is the expected cost of carrying inventory and of back¬
orders in the period from t to r + T, discounted to time 0, if the inventory
position is £ at time 0 (the review time) after the placing of any order.
Assume that an Rr policy is optimal. Then if the inventory system is
operated optimally, the inventory position at a review time after any order
has been placed will only have the values r* + 1, r* + 2,. . ., R*. Thus,
for these values, (8-84) becomes
Z(r* + 1; T) = f(r* + 1;T) + op(0; T)Z(r* + 1; T)
+ aP{ 1; T)[A + Z(R*-, T)]

+ aCY/(R*-r* + x- 1 )p(x; T) (8-85)

X=1

Z(r* + 2; T) = f(r* + 2; T) + op(0; T)Z(r* + 2; T)

+ ap( 1; T)Z(r* + 1; T) + aP(2; T)\A + Z(R*; T)]

+ aC Y, (R* — r* + x — 2)p(x; T) (8-86)

1=2

Z(R* - 1; T) = f(R* - 1; T) + ap{0; T)Z(R* - 1 ;T)

+ ap(l; T)Z(R* - 2; T)
+ . . . + ap(R* - r* - 2; T)Z(r* + 1; T)
+ aP(R* - r* - 1; T)[A + Z(R*; T)]
ao

+ aC ^ (x+l)p(x-,T) (8-87)
x = R*-f*-l
SEC. 8-10 EXPLICIT SOLUTION-PERIODIC REVIEW 393

Z(R*-T) = f(R*; T) + ap(0; T)Z(R*; T) + ap( 1; T)Z(R* - 1; T)

+ ap(2; T)Z(R* - 2; T)

+ ... + ap{R* - r* - 1; T)Z(r* + 1; T)

+ aP(R*-r*;T)[A + Z(R*;T)] + aC £ xp(x; T) (8-88)

where P(x; T) is the complementary cumulative of p(x; T). These equa¬

tions differ from (8-71) through (8-75), because in the periodic review case,
it is possible to have no demands occur in a period, and hence p(0; T)
appears. Equations (8-85) through (8-88) can be used to solve explicitly
for Z(R;T). The solution, which the reader is asked to obtain in Problem
8-13, is
<*> R*—r*—1

aA ^ ^2 o"p<n)(A; T)P(R* — r* — h; T)
71 = 0 h — 0_
Z(B*; T) = oo i2* —7-*—1

1 - a J2 J2 a”pM(h-, T)P(R* - r* - h; T)
n=0 h=0

oo 1

CQ+22 X) o,npM(h} T)f(R* — h;T)

, _n—0 h = 0_ _
“* « R*-,*-! (8-89)
1 — E £ anp(n)(h; T)P(E* - r* - h; T)

where p{n)Qi; T) is the n-fold convolution of p(x\ T), and by definition

p(0)(0,* T) == 1; p^)(/i;r)==0j A = ij2,...

We leave for Problem 8-14 the evaluation of the other Z(r* + k; T),
k = 1,. . ., R* — r* — 1. In Problem 8-15 we ask the reader to show that
lim iZ(r* + k; T), k — 1,. . . , R* — r*
x—>0

is the same for each k} and hence the average annual cost is this common
limit.
The above results can be used to obtain an explicit expression for the
expected annual cost. To do this we as usual define a function R, r, T)
to be the discounted expected cost at a review time if an Rr policy is used
with the critical numbers R and r (note that R and r can be different than
B* and r*), and if the inventory position is £ after taking whatever action is
appropriate. It then follows that equations (8-85) through (8-88) still
hold if Z(£; T) is replaced by 2(£; R, r, T) and the asterisks are removed
from R and r. Thus an explicit solution for d(R; R, r, T) is given by
394 DYNAMIC PROGRAMMING USED IN SS MODEL ANAL CHAP. 8

(8-89) on removing the asterisks from R and r. It also follows that

limi-40 R, r, T) is the same for each £, £ = r + 1,. . ., R. Conse¬
quently, the average annual cost, exclusive of review costs, is

3Zt(R, r) = lim i&(R; R} r, T) (8-90)

%—>o

The average annual cost including the review costs is

X{R, r,T) = ~+ XT(R, r) (8-91)

In this way one can obtain an explicit expression for the expected annual
cost of a periodic review system using an Rr policy. Problem 8-16 requires
the proof that the result obtained here is the same as that obtained in
Chapter 5.

8-11 The Lost Sales Case

The functional equation approach of dynamic programming is of little

assistance in studying steady state models for the lost sales case. To illus¬
trate the problems encountered, we shall study a periodic review system
in which the procurement lead time is constant, and precisely m > 1 orders
are outstanding at any given review time before any order is placed. In
this case the Z function cannot be made to involve just a single argument.
Suppose that £ is a review time, and that quantities wh . . ., wm were
ordered at times £ — T} £ — 2T,. . . , £ — mT respectively. Also assume
that £ is the on hand inventory at time £. Z will now depend on
fc Wl, . . . , Wm•

Let Z(£, Wij, wm) be the discounted expected cost over all future
times when an optimal decision is made at time £ and at all future times,
and the time between reviews is T. Here we shall begin computing carry¬
ing costs and the cost of lost sales at time £, and the delay of a lead time
will not be introduced. The task of setting up the functional equation for
Z is not quite so straightforward as in the backorders case. Note first that
when lead times are constant, only the order placed at time £ — mT will
arrive between £ and f + T. Suppose that this order for a quantity
v>m > 0 arrives at time f + i. Let pi(xi) be the probability that xi units
are demanded from £ to f + t and p2(x2) be the probability that x2 units
are demanded in the interval £ + i to f + T. If xx < £ and x2 < wm
+ £ — %i, the on hand inventory at time £ + T is £ + wm — X\ — x2. If
xi > £ and x2 < wm, the on hand inventory at time £ + T is wm — x2.
Finally, the on hand inventory at time £ + T will be zero if x\ < £ and
x% > wm + £ — xi or if xi > £ and x2 > wm.
 REFERENCES 395

Thus, if Xi, x2 are independent random variables

Z(£, wh . . . , wm) = J + /(£, wm) + min {Ad + CQ

Q
Wm-j-£ — Xl g —1

+a 2 E Pi(xi)P2(x2)2(| + wm — Xi — X2, Q, Wi,.. ., wv_i) (8-92)

X2 = 0 Xl = 0

Wm 00

+ a ^2 ^2 pi(xi)p^(x2)Z(wm — Xi, Q,wi,, u>m_i)

X2=0xi=£

+aL E
X2 = U>m+£~2l-f 1 Xl=0
E pi(x1)p2(a:2)
oo oo —j

+ E E Pi(xi)pi(x2) J Z(0, Q,w\,, w^i)

X2 = Wm+l XI = £

where /(£, tum) is the expected cost of carrying inventory and of lost sales
incurred from f to f + T (not f + rtof + r+ T). We ask the reader in
Problem 8-17 to write out /(£, wm) explicitly.
Z is now a function of m + 1 variables rather than just a single variable
as in the backorders case. None of the methods used previously can be
applied to the functional equation (8-92). Indeed, it does not seem easy
to use it either in discussing the optimality of an Rr policy or as an aid in
obtaining the average annual cost. We shall not attempt to carry any
farther the analysis of the lost sales case. In Problem 8-18 we ask the reader
to obtain the functional equation in the case where no orders are outstand¬
ing at a review time.

REFERENCES

1. Arrow, K. J., T. Harris, and J. Marschak, “Optimal Inventory Policy,”

Econometrica, XIX, 1951, pp. 250-272.
2. Arrow, K. J., S. Karlin, and H. Scarf, Studies in the Mathematical
Theory of Inventory and Production. Stanford, California: Stanford
University Press, 1958.
3. Beckmann, M., and R. Muth, “An Inventory Policy for a Case of Lagged
Delivery,” Management Science, Vol. 2, No. 2, Jan. 1956, pp. 145-155.
4. Beckmann, M., “An Inventory Model for Arbitrary Interval and
Quantity Distributions of Demand,” Management Science, Vol. 8,
No. 1, Oct. 1961, pp. 35-37.
5. Bellman, R., I. Glicksberg, and 0. Gross, “On the Optimal Inventory
Equation,” Management Science, Vol. 2, No. 1, Oct. 1955, pp. 83-104.
 DYNAMIC PROGRAMMING USED IN SS MODEL ANAL CHAP. 8
396

6. Dvoretzky, A., J. Kiefer, and J. Wolfowitz, “The Inventory Problem.

I, Case of Known Distributions of Demand; II, Case of Unknown
Distributions of Demand,” Econometrica, XX, Nos. 2—3, 1952,
pp. 187-222 and 450-466.
7. Dvoretzky, A., J. Kiefer, and J. Wolfowitz, “On the Optimal Character
of the (s, S) Policy in Inventory Theory,” Econometrica, XXI, 1953,
pp. 586-596.
.
8 Masse, P. B. D., Les Reserves et la regulation de Vavenir dans la vie
economigue. Paris: Hermann and Cie, 1946.
9. Scarf, H., “The Optimality of (S, s) Policies in the Dynamic Inventory
Problem,” in Mathematical Methods in the Social Sciences, (K. J.
Arrow, S. Karlin, and P. Suppes, editors). Stanford, California:
Stanford University Press, 1960, pp. 196-202.

PROBLEMS

8-1. Write out explicitly Et(£ + Q; T), Brit + Q) T) defined in Eq. (8-1)
for the case where the number of units demanded in any time interval
is Poisson distributed.
8-2. For the model discussed in Sec. 8-2, prove that if A = 0, then Z„(£; T)
and Z(£; T) are convex, thus implying that an order up to R policy
is optimal. Carry out the proof directly, and do not simply specialize
the A-convex arguments to this case.
8-3. When f(y) is differentiable everywhere, show the equivalence be¬
tween the definition of a convex function given in Problem 2-6 and
that represented by Eq. (8-15).
8-4. Prove the four properties of A-eonvex functions listed in Sec. 8-4.
8-5. Prove that when f(y;T) is a convex function which is everywhere
differentiable, then Fn(y, T) defined by Eq. (8-11) is continuous and
differentiable everywhere. To do this, first show that Fi(y; T) is
continuous and differentiable everywhere. Then show that Zi(£; T)
is continuous everywhere. Thus show that Fi(y, T) is continuous and
differentiable everywhere, etc.
8-6. Consider the definition of ri(x) for x > 0 given by Eq. (8-36). Show
that if units are demanded one at a time

y(x) = J* Vx_i(t - t)e-» g(r)er« df dt

Can you make a similar transformation for 0(1, x)?

 PROBLEMS 397

8-7. In Sec. 8-7, prove that ICD(£) + ttB(%) is a convex function of f.

Explain why E(£) need not be convex. Hint: Consider what can
happen for small £.
8-8. Modify Eqs. (8-29), (8-30) and (8-31) for the case where £ is allowed
to be negative. What influence does this have on the functional
equation for Z(£)?
8-9. In Sec. 8-6, show how to simplify the computation of the expected
number of backorders incurred between r and r + t when the interest
rate i = 0.
8-10. Derive in detail Eq. (8-55).
8-11. Compute
lim iZ(r* + k), k = 1,. . . , Q*
i-> o

where Z(r* + k) is given by Eq. (8-55), and demonstrate that the

same answer is obtained for each value of k.
8-12. By use of Eq. (8-82) obtain an explicit expression for the other
Z(r* + k), k = 1,. . ., Q* — 1. Show that lim [iZ(r* + k)] is the
same for each k.
8-13. Obtain Eq. (8-89) from Eqs. (8-85) through (8-88). Hint: There
are many ways in which one can manipulate the equations in at¬
tempting to solve explicitly for Z(R; T). Difficulty can be encoun¬
tered unless the substitutions are carried out in a particular way.
Begin with Eq. (8-88). Replace the Z(i?*; T) whose coefficient is
ap{0; T) by the right hand side of Eq. (8-88); replace Z(R* — 1; T)
by the right hand side of Eq. (8-87), etc. This yields

Z{R*; T) = f(R*; T) + ap(0; T)

+ 1; T)f(R* - 1; D + ap(2; T)f(R* - 2; T)

+ . . . + ap(R* — r* — 1; T)f(r* + 1; T)

+ a2p(2)(0; T)Z(R*; T) + o*p«(l; T)Z(R* - 1; T)

+ . . . + aY2)(R* - r* - 1; T)Z(r* + 1; T)

+ a{[l + ap{0; T)]P(R* - r*; T) + ap{ 1; T)P(R* - r*

- 1; T) + . . . + ap(R* — r* — 1; T)P( 1; T)} [A

00

+ Z(R*-, D] + aC{ 1 + ap(0; T)] E xp(x; T)

x = R*—r*
 dynamic programming used in ss model anal CHAP. 8
398

+ a*Cp(l;D £ U* b T)
gm —I

+ «S(>(2; T) f] IX b 2)pu-; Tl f . . .

+ a'JCp(tf* - r* - 1; Tl £) itf* r* » / - ll/Hx; T)

J# *» I

Now repent the substitution process with the above.

8-14. Evaluate explicitly, using Kq. (H-Htiq the other 'Mr* t k; T\ fora
periodic review system,
8-15. By use of the results obtained in Problem K-U show that
lim iZ(r* 1 k; Tl, k 1.H* r*
% *0

Ls the same for each value of k,

8-16. Compute explicitly X.rUt,n defined by Kq. IS '.«»! making use of
Kq. (H-811), and show that Kq, » is precisely what was obtained
in Chapter 5.
8-17. Write out explicitly fik, «‘J defined in Sec. Hit.
8-18. Derive the equivalent of Kq. (B-tf'J) for the case where no orders are
outstanding at a review time.
8-19. Prove that for the simple deterministic system discussed in Chapter
2 when! no stockouts are allowed that an Ur operating doctrine in
optimal. Show that an Hr doctrine remains optimal when backorder**
are allowed.
8-20. Prove that for transact ions reporting systems in which units are
demanded one at a time, an Hr ojterating doctrine is optimal provided
only that the expected carrying ami backorder eosts incurred in the
time period r to r 1 t following a demand I r being the constant lead
time and t the time until the next unit is demandedt depends only
on the inventory position of the system immediately following the
demand after whatever action is appropriate has been taken. Show
that, this is true even if the unit cost of the item is not constant.
Hint: Consider first the case where the unit cost is constant,. Ixit
F(y) he defined as in Sec. H-H. Supfmse that the ,-deiolute minimum
of F(y) is taken on at y H. Then show both for the ease when* the
demand variable is treated as discrete and where it is treated as
continuous, that, when the system is operated optimally, an Hr policy
is optimal regardless of what the shaj«* of Fuji may be. Show that
 PROBLEMS 399

this is true even if R is not unique. How is r determined? Is it neces¬

sarily true that r is the largest value if y < R such that F(r) > A
+ F(R)? Next consider the case where the unit cost is not constant.
8-21. Compute X(R, r) using Eq. (8-83) for the case where the number of
units demanded in any time interval has a stuttering Poisson distri¬
bution.
8-22. Compute X(R, r) for the case where the number of units demanded
in any time interval has a stuttering Poisson distribution by com¬
puting the expected cost per cycle and then multiplying by the
average number of cycles per year. Show that result is the same
as that obtained in Problem 8-21. Hint: A cycle may require 1,
2, . . ., R — r demands. What then is the expected length of a cycle?
Let / (r + k) be the expected cost incurred from time f-j-rtof + T + tf
when the inventory position is r + k immediately after a demand at
time and t is the time until the next demand. At the beginning of
a cycle the inventory position is R. What is the probability that it
will be r + k at some time during the cycle? Hence, what is the
expected cost per cycle?
8-23. Read Problem 2-69. Show that when demand is treated as a sto¬
chastic variable, it still remains true that when minimizing the
present worth of all future costs, rather than the average annual
cost, one should not include the rate of return in the inventory carrying
charge. Show, however, that if this is done, then after performing
the limiting processes discussed in this chapter to obtain the average
annual cost, it is necessary in the resulting expression to add back
into the carrying charge the rate of return.
8-24. In Sec. 8-4 we proved that if f(y;
T) is convex then F(y; T) cannot
have a relative maximum at a point y for which F(y; T) > A +
F(R*; T). Show that it is allowable for F(y; T) to have a relative
maximum at a point y such that F(y; T) = A + F(R*; T), without
destroying the optimality of an Rr policy. Illustrate this geo¬
metrically.
8-25. Use Eq. (8-2) to provide a rigorous proof of Eq. (8-3).
8-26. Under what conditions is an Rr policy optimal for the dynamic model
of Sec. 7-6 (possibly with different values of R and r for each period)?
1

PROBLEMS

OF PRACTICAL APPLICATION

11 How chances mock,

And changes fill the cup of alteration

with divers liquors!'1

Shakespeare, Henry IV, Part II

 9-1 Introduction

The old adage that there can be many a slip from the cup to the lip seems
to be especially descriptive of the problems encountered in any attempt
to implement an operating doctrine obtained by analysis of a mathematical
model in a real world inventory system. In this chapter we wish to turn our
attention from the main theme of this work, which has been the con¬
struction and analysis of mathematical models of inventory systems, to
problems of implementation, i.e., of practical application. We do not in¬
tend this to be a set of instructions on how a model should be implemented,
since it is not clear that such a set of rules could be devised to cover all
situations, and even if it could, more than a single volume would be re¬
quired to discuss these rules. Rather, we would like to point up the sorts
of problems that can arise, and insofar as possible, suggest ways to avoid
them or solve them. Unfortunately, however, there are no easy solutions
to most of these problems, and in a fairly large number of cases, no satis¬
factory solutions at all have been devised. We shall find it convenient to
discuss the various problems under the headings of: the relevance of the
model, data problems, multi-item problems, personnel problems and pro¬
cedural problems, and problems of evaluation. If nothing else, this chapter
should make the reader aware that a successful application will require
more than the formulation of an appropriate mathematical model.

9-2 Relevance of the Model

It has been noted previously that some simplifications and approximations

will always be needed when constructing a mathematical model of any
real world system. However, one must be extremely careful in developing
a model to be sure that it represents with sufficient accuracy the essential
characteristics of the system which are important determinants of what
the operating doctrine should be. If this is not done, the results obtained
by use of the model can easily lead to operating rules which are worse than
those currently in use, or worse than those which could be derived from
simple heuristic intuitive considerations. It would hardly seem necessary
401
 CHAP. 9
PROBLEMS OF PRACTICAL APPLICATION
402

to emphasize this point, and yet, the authors have encountered many
instances in practice where attempts have been made to use models m
situations where they were completely inapplicable, simply beca^m
model was available for use. One rather obvious misuse of a model which
the authors have encountered was in a military supply system where an
^t was made to apply a steady state model of the ^cussed m
Chapter 4 to spare parts for a military aircraft which would be pdased n
and out in a relatively few years. The model was completely inapplicable
since the mean rate of demand was continually changing with time, and
since explicit account should have been taken of the fact that the parts
would be obsolete when the aircraft was phased out. A more appropriate
model would have been some form of dynamic model such as those dis¬
cussed in Chapter 7. . , ,
The inadequacy of a model for some particular application is not always
so obvious as that referred to above, but nonetheless can lead to equally
disastrous results if applied. A good example of this concerns an apphcaUon
made in that part of a military supply system concerned with stockage of
electronic components. A large number of different electronic components
were carried in stock. The basic idea was to use a steady state model such
as the one described in Chapter 4, to set order levels and reorder points for
each of the items. Previously, the system had been operating using periodic
review. This in itself was reasonable enough, since the mean rate of demand
for many of the items was essentially constant over a fa r period of time,
and since the system had installed a digital computer which made it possible
to use what was, for practical purposes, a transactions reporting system.
However, it was recognized that the order quantities could not neces¬
sarily be used directly, since each year only a certain sum of money de¬
termined by the budget could be spent on procurements Hence it was
necessary to take into account, in one way or another, the budget con¬
straint on procurement expenditures. At this point, the following pro¬
cedures were attempted. It was decided to choose the order quantities Q,
and reorder points r* for each of the n items, by minimizing the average
annual cost X for all n items, in such a way that the expected procuremen
costs would not exceed the budgetary limitation. To dothis, the cost of a
backorder for unit j was assumed to have the form BtVCh where C,»the
unit cost of the item, t is the length of time for which the backorder exists,
and 0 is a parameter independent of j, to be determined in such a way that
the budget constraint was to be satisfied. Note that in this model there is
no fixed cost of a backorder. (The reasoning which led to the choice ot
= eVcTj is somewhat obscure, but is probably irrelevant, since the same
problems would have in all likelihood been encountered if was taken
equal to 0/3- for any arbitrarily selected number J>) Then the function to
hft mrniinized was
 SEC. 9-2 RELEVANCE OF THE MODEL 403

§ +1+ T‘ - « + (^Cy + ICy)£J(Qy,ry)}

where all quantities not defined above have their usual meaning. The
minimization was carried out for various values of 8 yielding X*(8), Qf(8),
r*(0)> j = 1,... ,n. Then 8 was chosen so that the expected procurement
expenditures were equal to the sum available in the budget.
It happened that, in the particular year in which the model was intro¬
duced, the budget was extremely tight. Because of this, 8 turned out to be
a very small number, so that the cost of a unit year of shortage was small.
For example, on some tubes, the cost of a unit year of shortage came to the
absurdly small value of $0.75. The reorder points determined in this way
were quite low. In many cases, they were so low that the safety stock was
negative (i.e., r,- was less than yty). The r, were in general considerably lower
than the on hand inventory at the beginning of the year. On the other
hand, the Qj was not especially small, since in the model, the contribution
of the Tj + Qj terms to the backorder cost were ignored, and thus Qj could
not be less than the Wilson Q.
The Qj and r,- values obtained from the above model were used in the
real world system. What happened was as follows: The system began
using up the on hand inventory for all items, and when (the very low)
reorder points were reached, orders for the rather large quantities Qy were
placed as needed. Before the end of the year, the entire budgeted sum had
been spent, and the system was about to run out of stock on many items—on
some because an order had been placed but the safety stock was negative,
and on others because the reorder points had to be passed because no funds
were available. At this point a crisis was reached which required special
legislation to obtain emergency funds for restocking through the sending of
expedited procurement orders to suppliers.
The application just discussed provides an excellent example of a case
where the results obtained using a mathematical model were much worse
than would have been obtained using the standard procedures that were
already in existence. It was entirely inappropriate, of course, to attempt
to apply a steady state model to a situation where there was a fixed a.nmia.1
procurement budget, and even worse to do it on the basis of introducing a
constraint on expected expenditures where the backorder cost was varied
to bring expected expenditures in line with the budget (since varying the
backorder cost did not have a great influence on the Qj). The proper
procedure for operating a system in the face of a tight budget is very com¬
plex. One must in general use hand to mouth buying (i.e., buy in very
small quantities), and insofar as possible attempt to postpone as many
procurements as possible until right after the beginning of the next budget¬
ary year when more funds will be available. None of the models developed
 PROBLEMS OF PRACTICAL APPLICATION CHAP. 9
404

in this text are capable of handling such a problem and, indeed, no such
models have been developed. It is clear that on the average, the budget
must be sufficient to procure the average annual demand, or the system
cannot continue to operate. However, by clever manipulation, it can
survive particular years with tight budgets.
There exist many other ways in which a model can be improperly applied.
We shall now briefly mention several of these. For example, consider a
situation in which the mean rate of demand is fairly low while the number
of units demanded per demand varies widely. Here it would be very
dangerous to use the steady state model developed in Chapter 4 for Poisson
demands, even if the mean rate of demand remained constant over time,
since the variance in the order size could cause the safety stock determined
from the model to be much too small. Other misapplications have involved
attempts to apply a steady state model to a department store item, 50 per¬
cent of which was sold during the Christmas season, or the attempt to use
results computed from a transactions reporting model in a periodic review
system.
The above have illustrated actual cases where models have been mis¬
applied. The reader should not infer from the above that a model will never
be satisfactory unless it replicates very closely the behavior of the real
world. For example, it may be perfectly satisfactory to apply a steady
state model to a situation in which there is a moderate amount of season¬
ality in demand, i.e., the additional savings which could be obtained by
using an operating doctrine based on a dynamic model over the costs in¬
curred by using a {Q, r) model with a fixed Q and r (computed using the
average demand rate) may not justify its use, perhaps because it is too diffi¬
cult to get the clerical people to use a changing Q and r. It is quite easy to
obtain a rough estimate of the differences in ordering and holding costs ob¬
tained by using a fixed Q and one that varies with the season. All one needs
to do is compare the results of the simple deterministic lot size model of
Chapter 2 with the deterministic dynamic lot size model of Chapter 7.
Sample calculations quickly reveal that even with sizable seasonal patterns,
the cost differences can be very small. Thus in many instances the use of
a constant Q could be easily justified. It is somewhat more difficult to
determine whether or not it is reasonable to keep r constant at the value
determined from the steady state model also (i.e., have a constant safety
stock). This depends in part on the relative magnitude of the stockout
costs and carrying costs. A natural way to investigate whether or not the
use of a constant r value determined from the (Q, r} model would be
satisfactory would be to use simulation.
Simulation is a very useful tool for studying how any operating doctrine
can be expected to behave in the real world system. In particular, simu¬
lation is helpful in studying what sort of behavior can be expected when
SEC 9-3 DATA PROBLEMS 405

using an operating doctrine obtained from a mathematical model under

conditions different from those assumed in deriving the operating doctrine.
The procedure is to simulate in time the behavior of the system when using
a specified operating doctrine. This is done by generating random numbers
giving the times between the occurrence of demands and the number of
units demanded. These random numbers are generated in such a way that
they have the statistical properties which characterize the stochastic
processes generating demands and the order sizes for the actual system.
Given the time pattern of demands, one can then determine the inventory
level as a function of time, the points at which orders should be placed (if
lead times must be represented as stochastic variables, random numbers
will be generated giving the length of the lead time), the backorders as a
function of time, etc. By use of this information, one can compute annual
costs, the fraction of time out of stock, etc. Hence one can compare an
operating doctrine obtained from a mathematical model with the operating
doctrine actually in use, to see how much improvement can be expected; or
an operating doctrine obtained from a simplified model ’with that obtained
from a more exact model, to see if the cost differences are significant enough
to warrant using the more complicated doctrine; or one can study some
heuristic modifications of an operating doctrine obtained from a simplified
model to see if an even better operating doctrine can be obtained.
The simplest simulations can be carried out by hand. Usually, however,
a digital computer will be needed, since the simulation must be carried out
for a sufficient length of time, or repeated a sufficient number of times, so
that the various measures of effectiveness which one wants to compute
will be representative of their expected values. This may require the
generation of 100,000 or more random numbers. The arithmetic associated
with the bookkeeping for all this becomes impossibly laborious, and hence
a computer is needed. In practice, simulation has not been used extensively
to analyze operating doctrines for a system. The reason is that it is too
expensive and too time consuming. In the future, however, as computers
become faster, and cheaper to use, one might expect to see simulation used
much more than it is at present.

9-3 Data Problems

Once a satisfactory mathematical model of some real world situation has

been developed, there remains the problem of determining empirically the
values of the various parameters, and perhaps also the nature of certain
functions which appear in the model, before it is possible to use it to obtain
an explicit operating doctrine. In order to obtain these parameters and
functions, it is necessary to use data obtained from the actual system.
 CHAP. 9
406 PROBLEMS OF PRACTICAL APPLICATION

Unfortunately, for a variety of reasons which will be examined in more

detail below, it is often quite difficult to obtain the data needed. Even when
some data are available, it is frequently exceptionally difficult to compute
the parameters and functions needed in the model with great accuracy.
Fortunately, the optimal profit or cost often does not change radically with
small changes in the parameters, and hence good results can be obtained
without knowing many of the parameters with great accuracy. It is very
helpful, however, to determine which parameters, when changed, have the
greatest influence in changing the optimal profit or cost, since the greatest
effort should then be concentrated on obtaining accurate values of these
parameters. In order to determine the critical parameters, it is necessary
to perform a sensitivity analysis, simply by determining what happens when
a parameter is changed. Some examples of this have been given in previous
chapters.
In the next several sections we wish to discuss in more detail the problems
involved in determining the parameters and functions needed in the
mathematical model. It will be convenient to subdivide the discussion of
these data problems into the following categories: (a) the demand distri¬
bution, (b) the lead time distribution, and (c) the costs.

9-4 The Demand Distribution

All inventory models require some information concerning the demand for
the item under consideration. In the simplest deterministic model only a
single parameter, the demand rate is needed. For other models it may be
necessary to determine the nature of the process generating demands, and
the nature of the distribution of the quantity demanded per demand. It may
even be necessary to predict how these will change over some future time
period. In order to make even the simplest estimate of the demand rate,
historical data are needed. It is surprising how often one finds that no
demand data whatever are available in usable form, and hence, before any
attempt can be made to apply the operating doctrine obtained from the
mathematical model, it is necessary to collect such data. In essentially all
cases, the data available will be sales data, and these will be equivalent to
demand data only if no sales are lost. When lost sales are possible, then
sales data will not yield precisely demand data. Even in the case where all
demands are backordered, sales data can distort the picture if they are
based on the time of filling an order, because, when some demands must
be backordered, the time pattern of filling orders will be different from the
time pattern of the original demands. One normally has no alternative but
to use the sales data as demand data. In the lost sales case, one could, if
the average fraction of the time out of stock was known, correct the mean
 SEC 9-4 THE DEMAND DISTRIBUTION 407

rate of demand obtained from sales data by dividing by the average fraction
of time for which the system has stock on hand. Generally, however, the
fraction of the time out of stock is not available and is difficult to measure.
Fortunately, it is often small and hence not too much of a problem.
Let us now assume that some sort of demand (i.e., sales) data are avail¬
able, and let us then see what problems arise in attempting to use these to
obtain the parameters of the demand distribution needed in the model.
First, we shall study the case where the model being used assumes that the
stochastic process generating demands does not change with time. Before
any attempt at parameter estimation is made, one should examine the
situation to determine whether or not the assumption of a constant mean
rate of demand is reasonableT"Usually, in the real world, the mean rate of
demand will not be strictly constant, and what one really wants to decide is
whether over a suitable time interval (perhaps a year or so) the mean rate of
demand is essentially constant. It is very difficult to make this decision on a
quantitative basis since this would require comparing the cost using an
operating doctrine obtained from a model which assumes a constant de¬
mand rate and those using an operating doctrine obtained from a dynamic
model. To make such comparisons could require a great deal of work, and
might even require the use of a large computer to make a detailed simula¬
tion. To do all this is usually much too time consuming and expensive (and
even if it was done the results might be misleading because of the difficulties
in predicting how the mean rate of demand would change with time for use
in the dynamic model). Instead one makes the decision on a very qualitative
basis, often without a detailed study of the data. Frequently, it is sufficient
to know what sort of item is being dealt with. If it is a staple item, which
does not have an extreme seasonal pattern, and for which obsolescence is
not an immediate problem, one will usually use a model which assumes that
the mean rate of demand is constant, simply because the probable savings
that could accrue from using a more complicated dynamic model do not
justify the relatively high costs of its use. If the mean rate of demand is
changing somewhat with time, one can simply recompute the operating
doctrine at some regular interval, such as once a year. On the other hand, if
the item is a high fashion item, which is only sold during one period of the
year, or is an item which is to be phased in and out over a relatively short
period, it will frequently be ill advised to use a model that assumes the
mean rate of demand to be constant.
If one wants to determine the behavior of the mean rate of demand over
time in more detail, it is convenient to use a control chart type of analysis
in doing so. As a measure of the demand rate one will use the demand per
time period, i.e., the daily, weekly, or monthly demand. This will be
plotted on a graph as a function of time. Using the demand per time period
data, one will then compute from the data available the average demand
408 PROBLEMS OF PRACTICAL APPLICATION CHAP. 9

per period 1 and the standard deviation of the average demand jper period
Sd- On the graph where the demand data are plotted, lines for d, d + ZSd
and d — 3Sd are also plotted. This may yield results such as those shown
in Figs. 9-1 and 9-2. The chances that a point will lie outside the control
limits d + ZSd or d — 3Sd are very small if the mean rate of demand is
indeed constant.
If the data behave like those shown in Fig. 9-1, then statistically there
is no reason to believe that the data were not generated by a stochastic
process in which the mean rate of demand is constant. On the other hand,
if the data behaved like those in Fig. 9-2, it seems clear that the mean rate
of demand is increasing with time. In both cases, however, one might use

an operating doctrine obtained from a model which assumed that the mean
rate of demand was constant. Indeed one could use graphs such as Figs.
9-1, and 9-2 to indicate when it was desirable to change the parameters in
the model and recompute the operating doctrine. The recomputation (or
at least an investigation of whether or not a recomputation should be made)
would be triggered by having a point fall outside the control limits. For
such purposes, however, it would not necessarily be desirable to have the
control limits lie at d + 3Sd and d — ZSd- It might be noted that the sort
of results one obtains on graphs like Fig. 9-1 can be strongly dependent on
the time period chosen, i.e., whether one uses the daily, weekly, or monthly
demand. For example, the daily demand on Mondays may always be very
low while it is always very high on Fridays. Hence for this unit of measure,
a control chart would indicate that the mean rate of demand was not
SEC. 9-4 THE DEMAND DISTRIBUTION 409

constant. On the other hand, the mean rate of demand on a weekly basis
might be quite constant. The interval should be chosen to be as large as
possible consistent with the data available, and any other requirements
such as those that will be discussed later. It should not, however, be larger
than either the lead time or the average time between the placement of
orders.
We shall now turn to the details of estimating the demand parameters
or the demand distribution needed in the model. If the model is a determin¬
istic one which requires only the mean rate of demand, we would divide
the total demand, over the longest possible time period for which it is be¬
lieved that the demand rate is representative of the current demand rate,

Weekly i
demand
•
•
J, oo
CL + uOjjr
•

• # *
d -.-.•. '■ 1 ■

•
d-SSd •
•

0 Time

FIGURE 9-2.

by the length of the period in years to obtain X. When demand is treated

as a stochastic variable, the model may use some theoretical distribution,
for which it is only necessary to estimate the parameters in the distribution,
or it may be possible to use an empirically determined distribution, so that
in such a case, it is necessary to determine the entire distribution.
Let us first examine the case where the model allows the use of an em¬
pirically determined distribution, and it is desired to use such a distribution.
To be specific, let us imagine that we are using one of the simple (Q, r)
models developed in Secs. 4-2 and 4-3. Assume also that the lead time can
be imagined to be a constant r. We shall defer a discussion of the additional
complications introduced by stochastic lead times to the next section. To
determine the density p(x; r) that x units are demanded in the lead time,
we might proceed by dividing the total time period over which we have
410 PROBLEMS OF PRACTICAL APPLICATION CHAP. 9

useful data into intervals of length r. For each such time interval we de¬
termine how many demands occurred. Then we construct a histogram by
finding the fraction of the time intervals which have demands lying in each
of some suitably chosen demand intervals of length Ax. When plotted
graphically one might get a result such as that shown in Fig. 9-3.
Unfortunately, the procedure just suggested often cannot be used. The
reason for this is that if r is fairly long, only a very small number of time in¬
tervals of length r will be available in the time period over which useful data
are available. In order to use the above technique, one should have some¬
thing like at least 50 intervals of length r in the demand history, if one is to
obtain a histogram which contains any detail. Often this cannot be done.

Thus we must use a different procedure to obtain p(x; r) empirically. An

alternative procedure is to proceed as follows. The demand data are used
to obtain a histogram for the demand over a shorter time period of length
£, say one day or one week; t can be chosen so that r = n£ where n is an
integer. Then, if we assume that demands in different periods of length t
are independent, p(x; r) = p(7l)(x; £), i.e., by taking the n-fold convolution
of p(x; t)} one obtains p(x; r). However, it is very tedious to attempt to
compute by hand the n-fold convolution of an empirical density function,
especially if n is fairly large and x can take on a considerable range of values.
The computation can be made either directly from the definition or through
use of the generating function. The computation could be carried out quite
readily on a computer, however. The results of the computation, when
plotted, would yield a histogram such as Fig. 9-3.
SEC 9-4 THE DEMAND DISTRIBUTION 411

Once p(x; r) is obtained then it is a simple matter to obtain P(x; r),

the complementary cumulative of p(x; r). When plotted graphically, it
would look something like that shown in Fig. 9-4. In principle, then one
can proceed to use the model. In addition to P(r; r), one needs ij(r) =
(x — r)p(x; r), which can be easily computed numerically (at least
on a computer it would be easy). Frequently, another trouble arises
in using P(x; r). When stockout costs are high, P(r*; r) will be very small,
i.e., the only part of the distribution which is important is that far out on
the right hand tail of p(x; r). However, events lying far out on the right
hand tail of p(x; r) are quite rare, and for this reason an extremely long

FIGURE 9-4.

history of demand would be needed to obtain an accurate representation

of the tail of the distribution. Only under special circumstances can one
get an accurate representation of the tail of the distribution, which is
really the only important part for inventory problems. When the tail of
the distribution is poorly represented, then serious errors can be made in
computing the reorder level and safety stock.
In periodic review models, any attempt to use an empirical distribution
involves ail the difficulties mentioned above plus an additional difficulty.
For example, for the simplest order up to R model, recall that R depends
on the demand in the time t + T. Hence, if one desires to optimize with
respect to T, it is necessary to generate a new empirical distribution for
412 PROBLEMS OF PRACTICAL APPLICATION CHAP. 9

each different T value used. This is impossibly laborious to carry out by

hand, although it could be done quite readily on a computer.
The above discussion has pointed out that, in many instances, severe
difficulties may be encountered in any attempt to use an empirical demand
distribution—even in the simplest of inventory models. It should be clear
that the difficulty would be compounded if a model with t 0 was being
used, since then the distribution would need to be known as a function of
time. Perhaps the most serious deficiency in attempting to use an empirical
distribution is the fact that often only the tail of the distribution is impor¬
tant, and not enough data are available to represent it accurately. Thus,
although there do arise situations in which one can and will find it desirable
to use an empirical distribution, there are equally well many situations
where it is exceptionally difficult to use an empirical distribution, and, in
such cases, one is essentially forced to use a theoretical distribution.
There is one type of inventory model where one can frequently use an
empirical distribution with relative ease. This is the single period model.
There are two reasons why an empirical distribution can often be used here.
First, one can obtain the histogram for the period’s demand directly, and
without performing a convolution. Secondly, there may be no stockout cost
or a relatively low stockout cost, so that one does not need to move so far
out on the tail of the distribution to determine the order quantity. Thus
it is not so important that the extreme tail of the distribution be so ac¬
curately known.
Let us next consider the case where the model makes use of a theoretical
distribution, and it is only necessary to estimate the parameters of this
distribution. When the lead time demand is relatively low, the theoretical
distribution usually used will be the Poisson. The only parameter needed
is the mean rate of demand X, and it can be determined simply by dividing
the total number of units demanded over the relevant historical period by
the length of the interval in years. In addition to doing this, it is helpful,
if possible, to check to see how well the Poisson assumption seems to apply.
In the real world, it will normally be true that the number of units de¬
manded in any time period will not be strictly Poisson distributed, if for
no other reason, simply because the order size is not always unity. Often
it is very difficult to obtain information on the distribution of the order
sizes. It is desirable to learn as much as possible about this, however. One
crude but simple way of checking on the goodness of the Poisson as¬
sumption is to obtain a histogram for the demand in some period such as a
day or week, and compute the mean and variance of the histogram. If the
variance is quite different from the mean, this suggests that the assumption
that the number of units demanded in any time period is Poisson is not
too good. In cases where the lead time demand is so low that the normal
approximation cannot be applied and the data indicate that the Poisson
SEC 9-5 DEMAND PREDICTION 413

assumption is far removed from reality, then one may be forced to go to

some other theoretical distribution of demand such as the stuttering
Poisson. One would do this only if the item was very important since a
considerable amount of extra effort would be needed to program a computer
to make the computations which determine the optimal operating doctrine.
When the normal distribution is the theoretical distribution used in the
model, X is determined as above. Now, however, one also needs Z>, where by
definition a2 = Dt for any time interval t. To compute D one would, as
usual, select a time period t such as a day or a week and compute the vari¬
ance S2 of the demand over this period; D would then be obtained from
D = S2/L
It remains to discuss cases where the mean rate of demand changes with
such rapidity that a dynamic model must be used. The dynamic models
developed in this text assume that for each future time period, the system
knows precisely what the distribution of demand will be. In the real world,
of course, considerable difficulties are encountered merely in attempting
to predict what the mean demand will be for each future period, not to
mention the precise nature of the distribution. Indeed, it is impossible to
predict even the mean demand for any future period with complete ac¬
curacy. We shall now consider briefly some of the problems involved in
making demand predictions, and we shall also consider some of the tech¬
niques which can be used for this purpose.

9-5 Demand Prediction

Central to any attempt to use a dynamic model is the procedure used for
making predictions or forecasts. The nature of the predicting technique
used can vary widely. It may, for example, involve only the use of histor¬
ical data on the item itself, it may involve predictions on the general state
of the economy, or it may be based on future planned requirements (in the
case of spare parts, etc.).
We shall not discuss prediction techniques in detail, since different
circumstances require widely varying approaches. We shall, however,
briefly examine several of the more simple ones. Consider first the problem
of predicting the demand for things like spare parts, which depend on the
usage of some other piece of equipment. This is an important problem in
any military supply system. To be specific, imagine that we are attempting
to predict demand for a spare part for a certain aircraft over a given future
period of time. The typical procedure which has been used in the military
to do this is to establish, on the basis of historical data, a usage rate, given
as the average number of spares needed per flying hour. Then the total
number of expected flying hours for all the aircraft of the type under con-
414 PROBLEMS OF PRACTICAL APPLICATION CHAP. 9

sideration during the interval of interest is estimated from current plans.

Multiplication of the usage rate by the expected number of flying hours
gives the expected demand for the spare part over the interval under
consideration. The accuracy of the forecast depends on both the accuracy
of the usage rate and the accuracy of the prediction of the number of flying
hours. In practice predictions are often considerably in error, both because
it is difficult to obtain an accurate estimate of the usage rate, and because
it can be exceptionally difficult to predict what the flying hours will be.
It is especially difficult to obtain accurate values for the usage rate for very
low demand spare parts (this includes a majority of the spare parts), since
the aircraft is obsolete before a sufficient demand history is obtained to
yield an accurate value. A great deal of effort has gone into trying to find
ways to accurately predict usage rates, but, as yet, none of these has been
very successful.
It should be noted that in the above technique for predicting spare parts
demand, no attempt was made to introduce probabilistic considerations.
Clearly, in a situation like this it is very difficult to attempt to describe
what the probability density for the demand for the spare part will be in
some future period. There would be randomness in the demand for spares
even if the flying hours could be predicted with accuracy. In cases like this,
when working with a dynamic model which requires the specification of a
demand distribution, one is essentially forced to use a theoretical distri¬
bution such as the Poisson. If the distribution also requires the specifi¬
cation of a variance, about the best one can do is to make a crude guess as
to what its value may be.
It might be of interest to point out the sort of operating doctrine that has
actually been used by the military for some years for spare parts. A
periodic review system is used, with the review period being normally three
months, six months, or one year. Let T be the time between reviews in
months. The operating procedure is to attempt to always keep a safety
stock of k months demand on hand (k may vary from 1.5 to 12 months).
At a review time supply officers send in their estimates of what demand will
be for the period of length r + T + k, where r is the lead time in months.
These estimates are aggregated over all air bases and repair facilities to
yield the expected demand d over this period. Suppose that at the review
time the inventory position of the system is R. Then the quantity ordered
iscZ — iJifrf — i2>0, and nothing otherwise. This operating doctrine
does not take into account costs or probabilistic considerations. The value
of k is usually set arbitrarily, and is the same for broad categories of items.
At the present time the military is attempting to improve its procedures
for controlling inventories of spare parts. It is not a simple matter to obtain
really useful mathematical models, however, because of the prediction
problems referred to above.
SEC. 9-5 DEMAND PREDICTION 415

Let us now turn our attention to situations in which predictions are

made using historical data and nothing else. We shall assume that a
dynamic periodic review model of the type discussed in Chapter 7 is being
used, and that at each review time a new prediction is made. We shall
first examine techniques that are useful when there is no strong seasonal
pattern. The two procedures which have gained widest acceptance are (1)
to fit a least squares line to the historical data, and (2) to use exponential
smoothing.
The least squares technique assumes that the demand dj in period j can
be predicted using dj = aj + f>, where a, b are determined from historical
data by minimizing

F = ^ (dj - d’3y = L (dj - aj - 6)2 (9-1)

3 3

In the above, dj is the predicted demand for period j and dj is the actual
demand for period j. Assume that a, b are to be determined by using the
demands in the previous N periods. We shall imagine that the current
time £ is a review time, and that the time period from ttot — T,T being
the time between reviews, will be referred to as period 0, the period from
t - T to t - 2T as period -1, etc. Thus the numbers of the N periods
to be used in determining a and b will be 0, — 1, . . . , —-(A — 1). To
determine a and 6, we set dF/da = dF/db = 0 and solve the resulting
equations for a and b. Then from (9-1)
„ -(V-1)
7- = -2 L j(dj - aj - £>) = o,

= -2 L — aj - b) = 0

— (JV—1> 1
. -N(N - 1) ; V—' J-
L ji = ^ N(N — l)(2iV — 1)
L 3= 2 ’ j=0 D
i=0

Also, let
— UV— l)

U= L v= £ 3di

Tlien the solution to the equations (9-2) is

r JV - 1 TT1 , U , (N - 1\
N(N -1)(N + 1)
416 PROBLEMS OF PRACTICAL APPLICATION CHAP. 9

SO that

$=f+
u + 12 V + ^-^U
[j + ^y-1) i = 1) 2, 3,. . . (9-4)
N
N(N - 1 )(N + 1)J
Equation (9-4) tells how to compute the predicted demand for period j,
j = 1,2,, using data for the demands in periods 0, — 1,. . ., — (N — 1).
Note that period 1 occurs between t and t + T, period 2 between t + T
and t + 2T, etc. Observe also, that U/N is simply the average demand in
periods 0, — 1,. . ., — (N — 1), so that (9-4) indicates that the predicted
demand is the average over the past N periods plus a trend correction.
In (9-4) only U and V depend on historical data. It is unnecessary to
recompute U and V from their definitions each time a new prediction is
made. Instead, it is easier to compute the new values of U and V from the
previous values. Let & and V’ refer to the values of these variables com¬
puted at time t + T, i.e., using di, do,..., d-(N-2), and U and V the values
computed at time t, i.e., using do, d-i, . . ., d-w-1>. Then it is easy to see
that
ty = + U - d-(w_ i); V = v-U + Nd-ur-1) (9-5)
It is much easier to compute t) and V from (9-5) rather than directly from
their definitions.
By use of (9-4) and (9-5), it is quite simple, either manually or on a com¬
puter, to predict the demand in each of the future periods needed and also
to predict (at least approximately) the lead time demands needed.
The uncertainty in the predicted demand will increase with j, i.e., the
farther one moves into the future. It is desirable when possible to get some
estimate of this uncertainty as a function of j (the number of periods in
the future for which prediction is made). If sufficient historical data are
available, this can be done by computing a histogram for the distribution of
forecast errors. This is accomplished by using (9-4) to predict the demand
j periods in the future and then comparing it with the actual demand.
The difference between the actual demand and the forecast demand is
called the forecast error. By doing this for as many historical points as
possible, one can obtain a histogram of forecast errors for each value of j.
Let Sf be the variance of the distribution of forecast errors when predicting
j periods into the future. If the dynamic model uses a theoretical distri¬
bution such as the normal, then for period j one can use Sf for the variance
of the period's demand. Similarly, be adding the appropriate variances,
one can at least roughly estimate the variance of the lead time demands.
We gave no indication above as to how to choose N} the number of past
SEC. 9-5 DEMAND PREDICTION All

periods demand that are used in making the predictions. One way to select
N is to pick the N which minimizes the variance of the distribution of
forecast errors for a given /, or minimizes some combination of the variances
over all /. This could be a tedious computation to make manually, but it
could easily be carried out on a computer.
One difficulty with using a least squares line for prediction is that it is
always necessary to have available the demand N periods back. This can
take up an unnecessarily large amout of storage in a computer when many
items are being handled. The exponential smoothing method of predicting
eliminates this problem. Given any time sequence of data with/,- being the
value of the variable for period /, then the smoothed value of the variable
for period /, denoted by /,-, is
U = afj + (1 - «)//_i, 0 < a < 1 (9-6)
i.e., the smoothed value for period / is a times the value of the variable for
period / plus (1 — a) times the smoothed value for period / — 1, where a
is a positive number less than unity. Since /,- involves /,*_i, it is clear that all
previous data are included in obtaining /,-, i.e., from repeated substitution

h = <xfj + <*(1 —• «)/y-i + (1 — °0 2/i—2

= afj + a{ 1 — a)fj-i + a(l — a)2/y„2 + . . . (9-7)

Let us now see how exponential smoothing can be used in demand pre¬
diction. The same terminology will be used as in the least squares case.
The present time is t and we wish to predict the demand in period/, j = 1,
2,. . . , i.e., in the time interval t + (j — 1)T to t + jT. The demand in
the interval t — T to T will be denoted by do, in the interval t — 2T to
t — T by d-1, etc. We might, using the exponential smoothing formula,
compute dj, the predicted value of dj, from
dj = do = ado + (1 — ot)d—i (9-8)
This would be equivalent to using dj = U/N in the least squares case.
It will always introduce a lag when there is a trend in the data.
Suppose that demand has been increasing linearly, i.e., at a constant
rate. We shall now determine the trend correction needed in (9-8). Assume
that the demand can be written 8 + pj where p is the increase in demand
per period. Then
do = ad + a( 1 — a) [5 — p] + a( 1 — a)2[5 — 2p]
+ a( 1 - a)*l8 - 3p] + . . .

= 8 — a(l — ol)p[1 + 2(1 — a) + 3(1 *— a)2 + . . .]

(9-9)
a
418 PROBLEMS OF PRACTICAL APPLICATION CHAP. 9

In reality, dj = 5 + pj. Thus c?o must be increased by

1 — a
P
a

to obtain the correct result.

In the real world we must estimate p also. We can do this using ex¬
ponential smoothing. Let
Pj = dj - 3U (9-10)

Then the smoothed value of p will be computed using

po = apo + (1 — a)p~ i (9-11)

The formula for dj will then be

dj = do + ~ J po (9-12)

Equation (9-12) is the equivalent for exponential smoothing of (9-4) for

least squares prediction. Again one can determine a histogram for the
distribution of forecast errors for each jt and also determine the variance
Sf. The value of a can be chosen so as to minimize the variance of the distri¬
bution of forecast errors for a given j.
To see why (9-6) is referred to as exponential smoothing write a = yLt
and take the limit as At —* 0 while holding y constant. This yields the
following differential equation for continuous exponential smoothing

if+/ = /(« (M3)

The solution to this equation is

/ = 7 j[ e-y<f-nf(r)dt (9-14)

if /(—o°) = 0, and/ is an exponentially weighted average of all past values

of/.
In situations where a seasonal pattern is the most important aspect-of
the demand pattern, it is common to predict that sales this season will be
equal to sales last season at the same point in the season plus a percentage
correction appropriate to the general state of business, or any other relevant
figures. Being at the same point in the season as last year does not neces¬
sarily mean the same date, since the season may be determined by the date
of occurrence of some special event such as Easter, and hence the seasonal
dates can fluctuate from year to year. It is an interesting fact that for many
seasonal items, the percentages of the total season’s sales which are sold
in given periods of time measured from the start of the season remain
SEC. 9-6 THE LEAD TIME DISTRIBUTION 419

remarkably constant from one season to the next. In this way, the total
sales for the season can be predicted with some accuracy after only the first
few weeks. At least one very large retail chain makes use of these seasonal
percentages to make early estimates of the total season’s sales for purposes
of placing additional orders, or for markdowns.

9-6 The Lead Time Distribution

It was pointed out in the previous section that it is often true that demand
data needed in the application of an inventory model are not available.
This is even more true with lead time data, especially if lead times are
fairly long and orders are not placed too frequently. It is rare indeed when
sufficient data are available to yield a detailed histogram for the lead time
distribution. Sometimes, about the best one can do is to obtain crude
estimates of what the maximum and minimum lead times are. In such a
case, about all that can be done is to average them to yield the mean lead
time, and to estimate the standard deviation to be the range divided by
six (since often six standard deviations, three on either side of the mean,
will include essentially all of a density function).
In practice many other problems may arise with lead times besides those
of trying to estimate the lead time distribution. There may not, in fact,
be any stationary lead time distribution, since lead times may be continu¬
ally changing with time. Furthermore, it may turn out on occasions that
orders are split and the entire order is not shipped at one time. In addition,
there may exist the possibility of expediting orders if an out of stock con¬
dition appears imminent. Although some of these additional complications
can be included in the mathematical model, it usually turns out that the
model becomes so much more difficult to work with and requires so much
additional data which are hard to obtain, that it is not worthwhile to
attempt to include them rigorously.
We have noted in the previous section the difficulties involved in at¬
tempting to use an empirically determined demand distribution in a mathe¬
matical model. The difficulties are compounded, of course, if one attempts
to use both empirically determined lead time and demand distributions,
since then one must generate numerically the marginal distribution of lead
time demand. This in turn requires that the demand distribution for each
possible lead time must be found. It is only in very special cases that this
marginal distribution can be determined with sufficient accuracy to
warrant the effort required to obtain it. One can, however, use an empiri¬
cally determined lead time distribution with a theoretical, discrete, demand
distribution such as the Poisson, for if it is assumed that the lead time r
can only assume one of a finite number of values U with probability Z(k),
420 PROBLEMS OF PRACTICAL APPLICATION CHAP. 9

then the marginal distribution of lead time demand is given by (3-66).

The use of such a marginal distribution can complicate hand computations,
but would not be hard for a computer to work with. Of course, one can also
use some theoretical distribution, such as a gamma distribution, for a lead
time density, determining its mean and variance from the empirical data.
Then the marginal distribution of lead time demand can be determined
analytically, when a theoretical distribution is also used to represent the
number of units demanded over any interval. If the demand variable is
treated as continuous, or is represented by the normal density, one will usu¬
ally assume that the marginal distribution of lead time demand is also
normal with its mean and variance given by the results of Problem 3-12,
where the mean and variance of the lead time are estimated from the
empirical data, as are X and D for the demand variable.
When the lead time variance is significant, it can be dangerous to use a
model which assumes a constant lead time, if one uses the mean lead time
in the model, since this can lead to seriously underestimating the average
fraction of the time that the system is out of stock. Consequently, the
safety stock determined from the model may be too low. However, one
can often get by quite well using a model which assumes that the lead time
is constant, if instead of using the mean lead time in the model one uses
something like the maximum lead time or the mean lead time plus one
standard deviation. In situations where the lead time data are so meager
that it is essentially impossible to say much about the lead time distri¬
bution, one has little choice but to do something like the above.

9-7 Determination of Costs

The types of costs which are relevant in working with inventory models
have been discussed in Chapter 1. These costs include the costs of the units,
the fixed ordering costs, carrying costs, stockout costs, and the cost of
operating the information processing system. All these costs can be difficult
to determine in certain cases, and some are almost always difficult to
determine. Regarding this determination, there is not a great deal that
can be said to be really useful in a wide variety of situations; so we shall
only make a few brief remarks.
Stockout costs are often the most difficult to determine. They cannot
normally be measured directly, since they usually include such intangibles
as good will losses. Consequently, the usual procedure, if stockout costs
are specified at all, is for someone to make a guess as to what they are.
Fortunately, as we have illustrated previously by example, the optimal
policy tends not to be very sensitive to these costs, and hence an estimate
that is of the right order of magnitude will often suffice. In the text it was
SEC 9-7 DETERMINATION OF COSTS 421

shown how to obtain operating doctrines by specifying the maximum

allowable average fraction of time out of stock (and/or the expected
number of backorders at any point in time). If such a technique is used,
one should, after computing the operating doctrine, examine the imputed
stockout cost to see whether or not it is reasonable. If it is not, the basis for
specifying the average fraction of time out of stock should be re-examined.
Recall from Chapter 1 that except for costs such as warehouse rental,
which depend on the maximum inventory level, the rate of incurring carry¬
ing costs is assumed to be proportional to the investment in inventory, and
that the factor of proportionality is the carrying charge I, which has the
dimensions of dollars per year per dollar invested in inventory. Further¬
more, I can be written /i + 12 + • • • , since I is the sum of several contri¬
butions to the carrying cost. As has been noted in Chapter 1, the most
important component of the carrying cost is often an opportunity cost—the
rate of return. It is exceptionally difficult in any specific case to determine
precisely what maximum rate of return an organization could obtain by
investing funds elsewhere. Instead, a firm will frequently specify the rate
of return it desires on invested capital, and this can be used for the rate of
return in the carrying charge. Usually, this value will be at least 10 per¬
cent so that Ii, the contribution to I from the rate of return factor, will be
0.10 or greater. For nonprofit organizations, such as military supply
systems, the rate of return in the usual sense is not so relevant because it
would not be possible to invest funds for profit (except perhaps by earning
interest in a bank). However, there is very definitely an opportunity cost
in this case too, since if funds were not invested in inventory, they could be
used to buy more new planes or missiles which would have provided
additional security, etc. It is harder to obtain a measure of the opportunity
cost in this case, but it seems that here the rate should also be reasonably
high, probably 10 percent or even more.
The other contributions to the carrying charge may include such things
as insurance, breakage and pilferage, and taxes. Recall that taxes should be
included in the carrying charge only when no special efforts are made to
reduce inventories at certain times to avoid taxes. When no such efforts
are made, then, on the average, the annual tax costs will be proportional
to the average investment in inventory, so that J2, the contribution to I
from taxes, will be the tax rate in dollars per year per dollar invested in
inventory. This tax rate will usually not be linear over wide ranges, and
that range should be used which one expects will include the optimal av¬
erage inventory. Often, management will attempt to make insurance
cover the expected average inventory for the period of coverage, and in
this case J3, the contribution to I from insurance costs, will simply be the
insurance cost in dollars per year per dollar investment in inventory,
at the rate appropriate to the expected average inventory level. Other
422 PROBLEMS OF PRACTICAL APPLICATION CHAP. 9

contributions to I can be evaluated in the same way. For example, to

compute I4, the contribution to I of the costs of breakage and pilferage,
we can add up the total costs from this source over the length of time for
which relevant historical data are available, and divide by the integral of
the investment in inventory over that time, i.e., divide by the dollar-years
of storage incurred. The final value of I is determined by adding up all its
components h, h, etc. Typically, the final value of I will be at least 0.20.
Reasonable real world values for I range from something like 0.15 to 0.35.
Firms often make a mistake of using a value of I which is much too low in
computing inventory carrying charges. For example, the military long
used a value of I = 0.03, and has changed this only recently.
In principle, the task of determining A, the fixed cost of placing an order
is straightforward. One traces through in detail the process of placing and
receiving an order. Then the costs of forms, telephone calls, computer time,
etc., are determined. In addition, the amount of time that each individual
spends in processing the order is determined. These times are multiplied
by the appropriate wage rate (including benefits) to yield the labor costs.
All these are totaled to yield A. If a factory is part of the system, and the
item is produced in lots, then the setup cost for a production run will also
be included in A.
Various sorts of difficulties can arise in the determination of A. For ex¬
ample, if the system is quite small and there are only a few people on the
payroll, the steplike behavior shown in Fig. 1-4. This sort of behavior can
be accounted for with the models developed as follows: Assume that n
people are to be employed, and let hn be the maximum number of orders
which they can process on the average per year. Then, using the Lagrange
multiplier technique, determine the minimum average annual cost Xt when
hn is the upper limit on the average number of orders placed per year. Now
add the annual wage costs of the n employees to Xt Repeat the process for
n + 1 employees. In this way, the optimal number of employees can be
determined. In military systems that use military personnel to process
orders, it may not be the wage rates of the personnel which are relevant in
determining A; they might be in the service regardless of whether or not
any orders are placed. Instead, it is an opportunity cost that is of interest
in this case, i.e., what could they be doing if they were not processing orders?
Clearly, this opportunity cost may be somewhat difficult to determine.
The cost of the units themselves is usually the easiest of the costs to
determine. Recall, however, that because of the way the models were set
up in this text, the cost of the units includes transportation costs and any
other costs of placing an order which vary with the order size. In the text
we investigated incremental and “all units” quantity discounts, and
noted that the computational effort required to determine Q* and r* may
be increased considerably in this case. There are occasions in the real
SEC. 9-8 MULTI-ITEM PROBLEMS 423

world where discounts of this sort are important and must be taken into
account. Fortunately, however, there is another type of quantity discount
which seems to be more popular, and which is also easier to handle. This
discount is based on the total quantity purchased per year—not on the size
of any individual order. For situations where the mean rate of demand is
fairly constant, the average amount procured per year will be fixed inde¬
pendently of the operating doctrine (at least for the backorders case, and
also for the lost sales case if the system is seldom out of stock) so that the
average discount is predetermined. In this case one can treat the unit cost
as a constant, independent of the operating doctrine. The unit cost should,
however, be modified to reflect the discount.
Finally, let us consider the costs involved in operating the information
processing system. In general, these costs will include all the costs associ¬
ated with maintaining inventory records, including wage costs, computer
costs, material costs, etc. Also included will be the costs of making physical
inventory counts and the cost of making demand forecasts. In the text, it
was assumed that the cost of a review for periodic review systems and the
cost of operating a transactions reporting system were independent of the
operating doctrine. This need not- be true when the parameters of the
operating doctrine are varied over a wide range. If one could determine
precisely how these costs varied with the parameters of the operating doc¬
trine, these appropriate functions could be included in the cost expression.
Generally, however, it is adequate to assume that they are independent of
the operating doctrine.

9-8 Multi-Item Problems

Most inventory systems stock a rather large number of items. It is not at

all uncommon to find that 10,000 or more items are stocked, and even
100,000 or more will be stocked by a large department store. The control
of such a large number of items presents many problems that do not arise
in considering just a single item. It is one thing to try to develop an optimal
operating doctrine for just a single item, but it is something quite different
to attempt to develop optimal operating doctrines for 10,000 or 100,000
items.
It has often been typical of both industry, the retail trade, and the
military to treat broad categories of items in the same way. For example,
every six months a military installation might review many items,^ and then
order an amount that would bring the inventory position up to a six months
supply, plus average lead time demand, plus k months demand for safety
stock. The same value of k would be used for each item. This is in general
a very poor policy to follow. If nothing else, the models developed in this
424 PROBLEMS OF PRACTICAL APPLICATION CHAP. 9

text have clearly indicated that different items should be treated differ¬
ently, depending on the nature of the costs and stochastic processes in¬
volved. However, it can become impossibly expensive if one attempts to
develop and use sophisticated operating doctrines on each of 100,000 items.
For example, if 10,000 items were being controlled, one might need a very
large computer. If 100,000 items were being managed, several of the largest
computers available might be needed. Just the operation of several com¬
puters could easily amount to an expense of two or three million dollars
per year. The answer to this problem lies in dividing the items up into a
number of groups, with items in the different groups being treated differ¬
ently.
Recent studies in both the military, retail stores, and industry have all
reached the same interesting conclusion that, in general, a very small
fraction of the total number of items stocked account for a very large
fraction of the dollar volume of business involved. Frequently this will be
something like ten percent of the items, accounting for eighty to ninety
percent of the dollar volume. These studies have led many large inventory
systems to change the way in which they control the items they stock.
Now, the items are broken down into several categories, usually three, and
items in different categories are treated differently. The items in the three
different categories might be referred to as high, medium, and low value
items. In the military, for example, the high value items would be con¬
trolled very closely, using the best means available. For items with a
relatively constant demand rate, transactions reporting might be used along
with a (Q, r) model. For items with strongly time dependent demands, the
best available dynamic model might be used. In the case of medium value
items, somewhat less costly control procedures would be used. For example,
periodic review might be used for all items, with only two or three possible
review intervals allowed. Here also, an effort would be made to account
for the nature of the demand distribution and the costs in determining the
safety stocks to be kept. In the case of low value items, no attempt would
be made to use a sophisticated operating doctrine. All items might be
reviewed once per year. The safety stock for all these items would be more
or less arbitrarily set at k months of supply. By use of a control system of
this sort, one keeps down the cost of control while still doing a good job of
controlling the most important items. A similar sort of procedure might be
used in a department store. In this case, however, periodic review would
probably be used for all items, with shorter review periods for the high value
items. The review periods might range from one day to two weeks for high
value items, one week to one month for medium value items, and two weeks
to six months for low value items.
We have previously suggested that there will often be interactions be¬
tween items carried by the inventory system. Almost any interactions
SEC. 9-9 PERSONNEL AND PROCEDURAL PROBLEMS 425

involving a large number of items are difficult to account for in practice.

We studied in previous chapters how to handle constraints on floor space,
the number of orders which can be placed, or the allowable investment in
inventory. With 10,000 or 100,000 items it would be very difficult to
attempt to compute optimal operating doctrines in the presence of such
constraints. We also noted that it is difficult to give a satisfactory treat¬
ment of such constraints in the presence of uncertainty. Fortunately, in
the real world, constraints of this sort are usually not sufficiently stringent
that they need to be introduced explicitly. Perhaps the most important
real world constraints are budgetary restrictions on the amount that can
be spent on procurements. These are more important in the military than
in private industry, where there can be a greater degree of flexibility.
However, as we noted previously, there is no simple way of including such
budgetary constraints in a model.

9-9 Personnel and Procedural Problems

The development of an appropriate mathematical model and the gathering

of the data necessary for the formulation of an operating doctrine that
controls an item or items in an inventory system are only part of the job
required to achieve a successful application. One must, in addition, take
steps to make sure that the operating doctrine will be used properly in
practice. This means that one must make every effort to ensure that the
people who operate the system perform their tasks correctly. In order to
do so, they must receive the proper training. This is not sufficient, however.
It is also necessary that the system of incentives which is operative in the
system actually encourages the proper performance of the duties. Often,
for one reason or another, there will be antagonism on the part of certain
individuals toward the introduction of a new system. Every effort must
be made to overcome this antagonism, or if this is impossible, to introduce
checking procedures to make certain that these individuals do not com¬
pletely destroy the advantages of the new system by intentionally doing
the wrong thing.
In order to be sure that all personnel perform their tasks properly, it will
normally be necessary to prepare detailed, step by step instructions to
cover every possible contingency which may arise. Generally, it is a difficult
t,a,RV to make these instructions clear and understandable. Often, because
there are so many different cases to consider, the instructions become so
complicated that not even a lawyer can interpret them. By devising suit¬
able forms for keeping records and making reports, it is possible to increase
considerably the efficiency with which the system can be operated. As with
instructions, these forms must be such that those using them can under-
426 PROBLEMS OF PRACTICAL APPLICATION CHAP. 9

stand them. Frequently, one will have to break down the details for using
them to the point where even the color of pencil to be used in making
certain notations is indicated.
Some examples, taken from actual real world situations, may help to
emphasize some of the above points. One large department store instituted
a lot size-reorder point operating doctrine for a great many items in the
housewares department where the mean rate of demand is reasonably
constant over time, with the exception of the Christmas season. Initially,
the buyers were strongly opposed to the introduction of such a system.
Usually, the buyer in a department store has a great deal of freedom in
determining what to buy, when to buy, how much inventory to carry, when
to take markdowns, etc. The buyer himself makes all these decisions for
the items under his jurisdiction. With the new system, the buyer had no
control over how much to stock or when to place orders. In fact, all the
ordering decisions which were previously made by the buyer were, under
the new system, made by clerks in the accounting office which previously
had nothing to do with these decisions. The buyers no doubt opposed the
introduction of the new system partly because they were afraid that their
jobs were being downgraded. Probably another reason that they opposed
it was because they were afraid of what would be discovered. On making
checks prior to introducing the system, all sorts of shocking things came to
light, such as the appearance of a fifteen years7 supply of some items on
hand, the stocking of a huge number of different brands of a particular item
such as coffee pots (in different parts of the store, without one buyer know¬
ing what the other was doing), items that could no longer be sold because
they were obsolete and the buyer had not taken a markdown at the proper
time, etc.
Initially, things did not go at all smoothly on introducing the new system.
The greatest difficulty was in getting people to do the right things. The
clerks would mix up stock numbers and count the wrong items, or they
would not make the counts correctly. In the accounting department,
orders would not be placed when the reorder points were reached, the
wrong item would be ordered, or the wrong quantity would be ordered.
In time, these difficulties were corrected to a considerable extent. Further¬
more, most of the buyers did become more enthusiastic about the system
since it relieved them of a considerable amount of work and, in addition,
made it possible for them to blame someone else if things went wrong.
The operation of some military supply systems provides an excellent
example of situations where the system of rewards and punishments for
the operating personnel tends to make them behave in a way completely
contrary to the behavior desired. Just like the buyer in a department store,
the base supply officer at a military installation has considerable freedom
in deciding how much of various items to stock despite rules and regulations
SEC 9-10 THE EVALUATION PROBLEM 427

which, theoretically, determine what the inventory levels should be. The
supply officer’s commanding officer is the line officer in command of the
base. The main concern of the base commander is that all planes should
be in flying condition, or that all ships be seaworthy and with all systems
operational. Consequently, if a plane is grounded because there is a stock¬
out of some spare part, the base supply officer will be heavily penalized
and may receive a black mark on his record. On the other hand, there are
no equivalent demerits for keeping too much inventory. Thus, the logical
thing for the supply officer to do is to keep as much of everything on hand
as possible. This is precisely what he does do, and this leads to much higher
stocks at bases than are really needed.
Another example of this same sort of phenomenon occurs when some
part is in short supply throughout the entire military supply system. Then
the item is placed on what is called a critical list, and all base supply
officers are supposed to report how much they have of this item, so that
if someone needs it desperately, the item can be shipped from one base to
another. What actually happens when an item is placed on the critical list
is that all base supply officers attempt to hoard the item, and will not part
with any of it under any circumstances. Thus the situation is really made
worse rather than better.
To get some feeling as to the problems that can be involved in attempting
to write down a precise set of instructions as to how a system should be
operated, one only need look at a set of military supply manuals. These
consist of a set of thirty or more volumes that presumably state in detail
precisely how the system should be operated. However, on attempting to
determine what should be done by reading these manuals, one soon becomes
hopelessly confused. It is essentially impossible to discover from them how
the system should be operated. This is in part due to the complexity of the
whole system, and in part due to the way in which the manuals are written.
No one person knows enough about the whole system to say how it should
be operated. Different people write different volumes in the set, and,
consequently, this makes it difficult even to have them consistent, let alone
coherent.

9-10 The Evaluation Problem

Frequently, the most neglected part of any application of an inventory-
model to a real world situation is an objective evaluation of how well it
worked. Even under ideal conditions, it can often be difficult to quantify
in monetary terms precisely how much improvement there has been. In
some cases, the improvement is so startling that even though one cannot
say precisely what the average annual reductions in cost are, it is clear that
 428 PROBLEMS OF PRACTICAL APPLICATION CHAP. 9

they are substantial; a more detailed evaluation is unnecessary. However,

in such situations, it is usually also true that almost any rational procedure
for operating the system would have yielded a remarkable improvement,
and hence the improvement cannot be attributed to the use of any parti¬
cular mathematical model. In cases where the use of a mathematical
model leads to a reduction in average on hand inventory, a reduction in
the average fraction of time out of stock, and a reduction in the average
annual rate of placing orders, then, provided that it was not necessary to
hire extra people to help operate the new system, it is clear that there has
been an improvement. However, there still remains the question as to
whether the new system will effect sufficient cost reductions to pay for the
cost of installing it. The cost of installing a new system based on the use
of mathematical models is often fairly high. In very large inventory
systems where a very small percentage improvement could mean several
hundred thousand to a million dollars per year reduction in carrying costs
alone, it is very easy to justify spending rather large sums in an attempt to
make improvements. As the system becomes smaller and smaller, however,
it becomes increasingly difficult to justify large expenditures for the intro¬
duction of a sophisticated control system.
The department store referred to in the previous section provides a good
illustration of the sort of problem that one encounters in attempting an
evaluation of any application. The management had no idea at all as to
whether costs were reduced or profits increased as a result of the change.
Furthermore they knew of no way to determine this. They did feel that a
great deal was gained by having better control over inventories and by
having a clearer picture of what was going on. Furthermore, they knew
that something was gained by relieving the buyers of the task of controlling
the inventories of all the items, but they did not know how to balance this
against the additional costs incurred in accounting. The authors have
investigated applications in several department stores throughout the
country and in every case the situation was the same. The management
really had no idea as to how much better the new system was than the old
one, or whether it was really any better at all. The same sorts of difficulties
encountered by the management in the department stores is also en¬
countered in industrial applications. The situation is even more difficult
in military installations, where it is almost unpossible to make an evalu-
ation on a cost basis alone, since so many of the costs are not directly
measureable (viz., the cost of having an ICBM inoperable because of the
lack of some spare part).
There is another reason why evaluations are often difficult to make. This
reason is especially applicable when an outside organization is called upon
to make the application. The management personnel who were responsible
for an application will do everything possible to make it seem a success,
 REFERENCES 429

and they will make every effort possible to cover up any defects in the
system. On the other hand, if it really is a spectacular success, then in some
industries the management will also attempt to keep the entire operation
a secret from the outside world and their competitors.

9-11 Summary

In this chapter we have attempted to point out some of the difficulties

which can be encountered in making practical applications. We have not
discussed how one can treat complex multi-echelon systems which is the
form taken on by many military supply systems, where a single organization
operates the entire system. Here one has all the problems dealt with in this
chapter, plus the difficult task of attempting to make all parts of the system
mesh properly. As has been noted in Chapter 1, no general models have
been developed that show how the various parts should mesh together.
Hence, here, one has to face unsolved theoretical problems as well as
problems of practical application. Even within the range of systems
considered in this chapter, however, it was seen that many formidable
practical problems must be overcome before one can achieve a successful
practical application.
It is not true, of course, that in every application all the problems dis¬
cussed in this chapter will be major obstacles to be overcome. Some ap¬
plications are relatively easy and some are difficult. The problems which
can arise were considered, not to discourage the reader from attempting
practical applications, but, rather, to prepare him for the sorts of difficulties
which may be encountered, in order that he may be better prepared to
cope with them.

REFERENCES

1. Brown, R. G., Statistical Forecasting for Inventory Control. Hew York:

McGraw-Hill Book Co., 1959.
A number of different techniques for forecasting demand using only
historical data are examined. Included are discussions of exponential
smoothing and the use of a least squares line. Usually, no derivations
are given, however.
2. Fetter, R. B., and W. C. Dalleck, Decision Models for Inventory Manage¬
ment. Homewood, Illinois: Richard D. Irwin, Inc., 1961.
3. Magee, J. F., Production Planning and Inventory Control. New York:
McGraw-Hill Book Co., 1958.
430 PROBLEMS OF PRACTICAL APPLICATION CHAP. 9

4. Whitin, T. M., The Theory of Inventory Management^ Rev. ed., Princeton,

N.J.: Princeton University Press, 1957.

PROBLEMS

9-1. The daily sales of a certain item in a department store have been
observed for 60 days. It has been found that never more than three
units are demanded and the frequencies with which 0, 1, 2, 3 units
were demanded are 15, 10, 20, 10, 5 respectively. The procurement
lead time is always four days. Determine from the empirical data
the marginal distribution of lead time demand.
9-2. Suppose that in Problem 9-1, one additional day is observed, and the
demand is for 6 units. How does this change the marginal distribution
of lead time demand?
9-3. Consider the demand data given in Problem 9-1. Assume that the
lead time is always either 2 days or 3 days, with the probability of
two days being 0.3. Determine the marginal distribution of lead time
demand.
9-4. The weekly demand for an item has had in the recent past the follow¬
ing values on successive weeks: 25, 10, 15, 16, 9, 30, 17, 18, 8, 25, 26,
14, 12, 9, 15, 18, 7, 11, 19, 15, 28, 17. By use of a moving average
having N = 3 and without a trend correction, estimate the demand
two periods in advance for each week, beginning with the third in the
demand history, and compare with the actual demand. Repeat the
procedure using Eq. (9-4), which includes the trend correction, to
make the prediction. Plot the results obtained, showing the predicted
values and actual demand as a function of time. Compute a histo¬
gram of forecast errors for the case where the trend correction was
included in making the prediction.
9-5. Repeat Problem 9-4 using exponential smoothing with a = 0.15.
9-6. Assume that for all past time the demand per period has been 10
units. Then beginning at time 0, the demand per period begins to
increase at the rate of 2 units per period, so that dy = 10 + 2j. Plot
dj and dj = U/N when N = 4. Also plot dy and dj = dj = adj +
(1 — a)d3-1, a = 0.20. Does the lag dj — dj behave in the way one
would expect it to?
9-7. Assume that for all past time, the demand per period has been 10
units. Suddenly at t = 0, the demand per period jumps to 20 units
and remains constant at that value. Plot dy and dj, where dj is com-
 PROBLEMS 431

puted from Eq. (9-4) for N = 4, and the prediction being made for
the next period.
9-8. Repeat Problem 9-7 using Eq. (9-12) with a = 0.2, and again making
the prediction for the next period.
9-9. Solve Eq. (9-13) when
0, f<0
/(f) = 6f, t > 0

Plot /(f) and /(f) on the same graph.

9-10. Solve Eq. (9-13) when
f(t) _ / °> *<°
•'W"\20, f>0

Plot /(f) and /(f) on the same graph.

9-11. Discuss the way one could introduce an expedited procurement into
a {Q, r) model of the type discussed in Chapter 4. Assume that an
order is expedited if the on hand inventory falls to a given value, and
that an expedited order takes a time r < r to arrive. If the tune until
the arrival of the next order is less than f, no expedite order is placed.
The cost of expediting is *■*. What sorts of complications develop?
Would it be easy to work out Q* and r* for this model?
9-12. Discuss in detail the way in which a generating function could be
used to obtain the n-fold convolution of an empirical demand
distribution.
9-13. Discuss how one might compute the n-fold convolution of an em¬
pirical demand distribution on a computer, directly from its defi¬
nition, and without the use of generating functions.
APPENDIX 1

A1-1 Constraints and Lagrange Multipliers

Consider the problem of minimizing the continuous and differentiable

function z = /(asi,. . . ,xn) subject to the constraint g{xh... •, z») = «>
where g(xu . . ., xn) is also continuous and differentiable. Since the var¬
iables are not independent, it is no longer true that the optimal set of re¬
values must satisfy df/dx,- = 0, j = 1, . • •, n. However a procedure for
solving this problem would be to use the constraint to solve for one of the
variables, say to yield rn = h(x1> xn-f) This expression for r„is
then substituted into / to yield a function / of n - 1 variables. In tins
form the methods for finding unconstrained minima can then be applied,
i.e., we solve df/dx, = 0, j = 1,1 where, since

}(xi,. . ., rn-i) — flXl> • • •» x*~1> h(Xl> • • • ’

(Al-1)
j= + j = i
dXj dXj dxn dXj

However, from g(xi,. . ., x») = «, we see that

ia+ia.^ = 0 or — = - ^0
dXj dxj dg/dxn

JL
3^ _ j9f _ dXn d£_ = 0, j = 1,.. (Al-2)
~ dXj dg dXj
dXn
Now if rt . . ., x* are the minimizing values, write y = -df/dxn/dg/dxn
where df/dxn and dg/dxn are evaluated at x\, ...,xt Therefore the opti¬
mal Xj values must satisfy the n + 1 equations

M. _l „ M = o, j = 1,. - -, «; g(xi, ...,*») = <* (Al-3)

3r3 dry
Here we have » + 1 equations to be solved for x% . .., xt and y. The form
fAl-31 is very convenient to use, since all the variables are treated sym-
aid it is equivalent to (Al-2) if */»*. * 0 at *!,.... * In
433
434 APPENDIX 1

fact, we can use the equations (Al-3) to determine x*,. . ., xZ provided

that not all derivatives dg/dxj vanish at x*}. . . , x$. It seldom happens
that these derivatives do all vanish at x% . . ., xt Hence the above pro¬
cedure will work normally.
The necessary conditions (Al-3) which the x* must satisfy can be
obtained simply as follows. Form the function

F(Xi} . . . , Xn, 7}) f{x i, . . . , Xn) -f- • • * , Xn) — cc] (Al-4)

Then
eF , is. • 1 BF g — CL (Al-5)

Hence, on setting dF/dxj = 0, and dF/drj = 0, we obtain (Al-3). The

parameter rj which we introduced above is referred to as a Lagrange
multiplier. The procedure of determining the conditions satisfied by
the Xj, j = 1, . . ., n, that minimize z subject to the constraint g = a, by
forming the function F and setting the partial derivatives of F with respect
to the Xj and rj equal to zero, is referred to as the method of Lagrange
multipliers.
The Lagrange multiplier technique can also be used when there are two
or more constraints. Suppose that we wish to minimize z = fixi, . . . , xn)
subject to gi(xi,. .., xn) — ah . . ., xn) = a2- We introduce two
Lagrange multipliers rj, d and form the function

F(xh . . .,Xn,V,B) = fix1, . . . , Xn) + . . . , Xn) — «i]

+ 0[giixi, ..., Xn) — Oii] (Al-6)

The necessary conditions which the Xj that minimize z subject to the two
constraints must satisfy are

Sx3- ~ 0 ~ te, + v dxs + 6 dXj

(Al-7)
dF dF
gi — a = 0; a/i 92 0
dti dd

Here we have n + 2 equations to solve for the n + 2 variables

xh j — 1) • • •, n, ij, and 8. This procedure will work provided that the rank
of the (2 X n) matrix \\dgi/dXj\\, i = 1, 2; j = 1, is 2 at the mini¬
mizing point. The proof that this procedure works is the same as for the
one variable case, except that now one must use the two constraints to
solve explicitly for two of the variables.
Often when solving constrained optimization problems of the type dis¬
cussed above, there are the additional restrictions that x, > 0 for some or
all of the variables. In this case, the set of optimal xj need not satisfy
 INTERPRETATION OF LAGRANGE MULTIPLIERS 435
SEC. A1-2

(Al-3) or (Al-7), since one or more x* may be 0, i.e., be on the boundaries

of the region defined by x, > 0, j = 1,. . -, n (this is called the non¬
negative orthant). If there exists the possibility that the optimum has one
or more Xj = 0, then one must also find all the relative minima lying on
those boundaries of the non-negative orthant which correspond to the
variables which are required to be non-negative. To do this, one goes
through the same procedure outlined above first holding one of the variables
required to be non-negative at a zero level. This is done for each variable
required to be non-negative. Then the problem is solved for two of the
variables set equal to zero. This is done for all possible combinations of two
of the variables required to be non-negative set equal to zero. Then three
variables are set equal to zero, etc. The absolute minimum will be the
smallest of all the relative minima obtained with 0, 1, 2,.. . variables set
to zero. It can be a very arduous task to check all the boundary minima.
Fortunately, for the problems of interest in this text, these boundary
minima, do not usually occur.

A1-2 Interpretation of Lagrange Multipliers

Let us return to the study of the problem which seeks to find non-negative
values of xu . . ., xn which minimize z = /(mi, ...,Xn) subject to the
constraint g(x1,If 2* is the minimum value of z subject to
the constraint, then z* will in general depend on the value of a, i.e.,z is a
function of a. Similarly, the Lagrange multiplier v will also depend on a,
i.e. v is a function of «. We would now like to show that dz*Jda = -fl
where both sides of the equation are evaluated at the same value of a.
To prove that dz*/da = -v, let x\,. . ., xt be a set of */ values which
yield z*. These x*} values will also be functions of a. Thus

da fzi $x* da
However, because of (Al-3)

dxj V dxf

so
dz*_y' dg dx* (Al-8)
da~ n 34i dx* da

Next we note that the constraint g(xi,. .., a;») - a must always be satis¬
fied (i.e., it must be an identity in a). Hence, differentiating the constraint
with respect to a we obtain
436 APPENDIX 1

y-* $9 dxj __ l
dx*J da (Al-9)
J =1

or on substitution of (Al-9) into (Al-8), dz*/da = —ij which is what we

wanted to prove. The same procedure can be used to show that if there
are m < n constraints, then dz*/da = —i}3, i = 1,. . . , m. The proof is
more difficult for this case, however, and will not be presented here.
The above result can be given an interesting economic interpretation.
In this book z usually represents the average annual cost, and the con¬
straints will represent limitations on the physical resources such as capital
or floor space. Then, by its dimensions, m must be a value or cost per unit
of resource i. Intuitively, we see from dz*/daj = —Vi that Vi is the amount
by which the minimum cost can be reduced by adding one additional unit
of resource i. The Lagrange multipliers can thus be considered to be the
imputed values or shadow prices of the resources.

Al-3 Inequality Constraints

The constraints on a problem which are of interest to us in this work usually

take the form gj(xi,. . ., x„) < rather than gt(xi,.. ., xn) = on. Intui¬
tively, the constraint requires that no more than a given amount of a
resource can be used; it is permissible to use less than the maximum
amount, however. We now wish to show that the Lagrange multiplier
technique can be generalized to handle inequality constraints.
To begin the discussion imagine that we wish to determine non-negative
variables xh . . ., xn which minimize z = f(xh . . . ,x„) subject to
g(xh... ,xn) < a. Let x\,...,xt be the set of non-negative Xj
values which minimize z subject to the constraint. Then either:
(1) d(%i, •••>£*) = “ or (2) g(x*,. . ., xt) < a. If (1) holds the constraints
it is said to be active; a constraint is active if it holds as a strict equality at
the minimizing point. A constraint is called inactive if it holds as a strict
inequality at the minimizing point. A constraint must be either active or
inactive. The important thing to observe is that if the constraint is inactive
then the minimizing point is the same whether or not we consider the
constraint when solving the problem. If the constraint is active, then the
minimizing point is the same as would be obtained if we solved the problem
assuming that the constraint always held as a strict equality.
After one final observation, we shall be able to present a method for
solving the problem of interest. Let zZ be the minimum value of z for non¬
negative Xj in the absence of the constraint, and let z* be the minimum
value of z for non-negative Xj subject to the constraint g(xh . . ., xn) = a.
Then it is true that zt < z*, since the point which yields z* is also an allow-
SEC Al-3 INEQUALITY CONSTRAINTS 437

able solution to the unconstrained problem. More generally, if Zm is the

minimum value of 2 for non-negative xj subject to the m constraints
gi(x 1, ...,£»)= on, and Zm+i is the minimum of 2 for non-negative Xi
subject to m + 1 constraints (m of the constraints being the same as those
referred to above) of the form gi(xi, . . ., xn) = then Zm+i ^ Zm, since
the point which yields z^+i is an allowable solution to the problem with m
constraints.
It is now clear how to solve the problem of finding non-negative
Xj which minimize z = fix1,. . . , xn) subject to the inequality con¬
straint g(x 1, ... ,xn) < a. We first find the point which minimizes z for
non-negative Xj ignoring the constraint, i.e., we first solve the unconstrained
problem. If the solution so obtained satisfies the constraint, it follows from
the previous paragraph that it is the optimal solution. If the point so
obtained does not satisfy the constraint, then the constraint will be active,
and we solve the problem of finding non-negative xj which minimizes 2
subject to the equality constraint g(x 1,. . . , xn) = <*• This problem can be
solved by introducing a Lagrange multiplier as discussed in Sec. Al-L
The minimizing point so obtained will also be the optimal solution to the
given problem. When the constraint is an inequality, more effort may
be required to solve the problem than for the case where the constraint is
a strict equality, since it is also necessary to solve the unconstrained
problem.
Consider now the case where we wish to find non-negative variables
which minimize z = f(x 1, . . ., xn) subject to the two inequality constraints
gi{xh . . . , xn) < ah g2(xh ...,&»)< <*2. The procedure is to first solve
the problem ignoring both constraints. If the minimizing point so obtained
satisfies both constraints it is the optimal solution to the given problem.
If it does not satisfy both constraints, then at least one of the constraints
will be active. In such a situation, the next step is to solve the problem of
finding non-negative Xj which minimize z subject to the constraint
gi(xh . . ., xn) = oil (ignoring the other constraint). If the solution so
obtained satisfies the other constraint, it is optimal. If it does not, we next
find the set of non-negative Xj which minimize z subject to g2{xi,. . ., xn) =
a2, (ignoring the constraint gi(xi, . . ., xn) < «i)* If the solution so ob¬
tained satisfies gi(xi,. . ., xn) < ai it is optimal. When it does not, we
are sure that both constraints will be active, and we determine the set of
non-negative x, which minimize z subject to the two constraints
gi(xh ...,Xn) = Oil, 02(31,. . ., Xn) = a*. The resulting set of xs will be
the optimal solution to the given problem.
The same procedures can be used to solve problems involving three or
more inequality constraints. However, the effort required to solve the
problem increases rapidly with the number of constraints.
APPENDIX 2

A2-1 Introduction

To determine optimal values of the parameters associated with an operating

doctrine, it is often necessary to solve numerically an equation such as
fix) = 0 or a set of equations such as fi(x, y) = 0, fi{x, y) = 0. We shall
discuss a useful technique for solving numerically equations such as the
above. It is known as Newton’s method. We shall also discuss a technique
called the method of “steepest descents” for minimizing the cost expression
directly, rather than solving a set of equations obtained by setting the
derivatives of the cost equal to zero.

A2-2 Newton’s Method

Consider the problem of solving the equation fix) = 0. Suppose that x0 is

an approximate solution, i.e., /(x0) is close to 0. Then using the first two
terms of the Taylor expansion we have

/(*) « fix,) + ® (* - *o) (A2-1)

when x is close to xq. In (A2-1) df(xo)/dx is the derivative of / with respect

to x evaluated at x0. We wish to determine x such that fix) = 0. Setting
fix) = 0 in (A2-1), we obtain

x — Xq = Ax = -~f(xo)/(df(xo)/dx) (A2-2)

We can then try xi = x0 + Ax as the new estimate of the solution to

fix) = 0. This process is repeated and at the (n + l)st stage

Axn = —fixn)/(dfixn)/dx) (A2-3)

and xn+i = xn + Axn. The procedure just outlined is an iterative procedure

for finding the solution to f{x) = 0, and is referred to as Newton’s method.
It is not necessarily true, however, that the procedure will converge to the
solution to fix) = 0. The method will converge if df/dx and d2f/dx2 do not
change signs in the interval between xQ and the root of f(x) = 0.
438
 APPENDIX 2 439

The same procedure can be used to solve fi(x, y) = 0, fiix, y) = 0.

Assume that at the (n + l)st stage the approximate roots are xn} y%* Then
write

dfl(,Xn) y-n)
0 = fi(xnj yn) + {Xn-ri Xn)
dx

+ dy
yn) fl(Xn+l} yn+l)

0 = U{xnj yn) + ^ (a5»+1 x*)

+ (Vn+l ~~ yn) ttf2(xn+i} yn+1)

dy
so that
A _ _ _ I ~fi dfi/dy
AXn - *n+l *n~ j _/f
1 - / \ dfl(Xni yn) f ( dfljXm yn)
. \ dfljXni
(A2A)
= j U(xn, 2/») Jl{Xn} yn) ^

A _ i a/i/dz -A
Aj/„ - yn+1 yn- j d^bx —y2

1 r/ / ,. % dfz(xn, 2/") f (V w d_fi(xn, yn)' (A2-5)

= j \fl(xn, 2/n) 3a. M^n, 2A dx
where
dfl(xn, Vn) dMxn, yn) _ dfl(xn, yn) dpljXn, yn) (A2-6)
J = dy dx
dx dy

Then xn+i = xn + Axn, yn+1 = Vn + Aya. Again it is not necessarily true

that this iterative scheme will converge. The conditions for convergence
are not so simple as in the one variable case, and we shall not consider them.
Of course, if the roots of f(x) = 0 or h(x, y) = 0, /*(*, y) = 0 are not
unique, Newton’s method does not provide a way for finding all roots.
There does not seem to be any simple way to do this in general.

A2-3 The Method of Steepest Descents

For most of the models formulated in this text, the optimal operating
doctrine was determined by minimizing some cost expression such as t e
average annual cost X. Newton’s method provided a way of solving equa¬
tions such as dX/dQ = dX/dr = 0, which are necessary conditions that Q
and r must satisfy if they minimize X. The method of steepest descents
works with X directly, and proceeds to minimize X rather than to solve
440 THE METHOD OF STEEPEST DESCENTS SEC A2-3

equations such as dX/dQ = dx/dr = 0. Assume that X is a function of

two variables Q and r, and that it is desired to determine the absolute
minimum of X.
Suppose that we begin with any values Qq, r0, and compute X(Qo, r0).
We wish to determine values Qh n such that X(Qi, n) < X(Qo, n). Now
dX/dQ, dX/dr tell us how X is changing with respect to Q and r at any
given point, i.e., they indicate how Q and r should be changed so as to
decrease X. Consider

Qi = Qo - r1 = r0- 6 > 0 (A2-7)

Then if 8 is sufficiently small, X(Qi, rO < X(Qo, r0) unless dX/dQ =

dx/dr = 0 at Q0j r0. When 8 is the same in both expressions in (A2-7), we
say that we are using the method of steepest descents, since it can be shown
that we move in the direction of the greatest rate of decrease in X.
The procedure then is to compute a Qi and n using (A2-7). The value
of 8 is selected arbitrarily. The proper values to use will automatically
suggest themselves as the problem is solved, since if 8 is too large, the
minimum will be overshot, and if 8 is too small, it will take a long time to
reach the minimum. Then X(Qh n) is computed. If X(Qi, n) < X(Qo, n>),
then dX/dQ, dX/dr are recomputed at Qi, r± and new values of
Q, r are found from (A2-7), where now Qh n become Qo, When
X(Qi, t*i) > X(Qo, r0) select a smaller 8 and repeat the process. Just as
with Newton’s method troubles with convergence may be encountered with
the method of steepest descents. The value of 8 may be changed at each
iteration, if desired.
To use Newton’s method, the first derivatives of /i, /2 are needed. These
are the second derivatives of X. The method of steepest descents uses
only the first derivatives of X. If the derivatives are difficult to evaluate
numerically, the method of steepest descents requires less computation
per iteration. On the other hand, it may require more iterations unless 8
is wisely chosen each time.
APPENDIX 3

In the following properties, p(r; p), P(r; p), y(t) a, 0), bw(j; n, p) are
defined by
00

p(r; n) = e-*1; P(r; n) = 2 PC'j r = ' * *

T. j=T

mYe-e*
7it', «, P) r(a + 1)
O' > ~ 1, jS > 0

birti', n, p) = ^rtnf p"(‘1 ~ p^’J ~ °’ *’ 2’'

The properties hold for all m > 0 and all non-negative integers r*. The
differencing operator A{/(r)} is defined by A{/(r)} =/(r) —/(r 1).
T(x) is the gamma function of argument x.
1. rp(r; p) = pp(r — 1; p)

2. Let Mm(r) = ']jjmp(jm, m = 0, 1, 2,. . .

j=r
Then,

Pm(r) = P L (m a; X) m = X’ 2’ • ■ ■

3- p) = pP(r - p)
y=r

4. |><i; ft) = pP(r — l; p) + p2P(r — 2; p)

j=r

5. 2 j*pO'; p) = P3p(r - 3; p) + 3p2P(r - 2; M) + Pp(r “ 1>^

j=r

* Although r is to be non-negative, it is possible in some of the properties to have a

negative argument for p or P when r < 3. These properties will be correc or sue
the convention is used that p(j; /x) = 0, P(j; m) = 1 for i a ne^ative mteger*
441
442 APPENDIX 3

6. -PO"; m) = ftP(r — 1; n) + (1 — r)P(r; /*)

00

7. Let 0*(r) = m), m = 0,1, 2,. . .

Then,

^m(r) = {E (-+ *) a.(r) + Mm+iW

- (r - l)m+1P(Y; m = 1,2,

8. fj iP(i! /*) = *5 P(r - 2; M) + t*P(r - 1; m) - r-^f^ P(r; /*)

j—T

9. !>P(j; m) = [-£ + ^ - j] P(r; m) + pP(r - 1; m)

+ ^P(r-2;M) + ^P(r-3;M)
eo eo 00

10. X) O' - to(r,v) = ]CipO + r, m) = X) p0‘; m)

y»r i“0 i»r+l

= mPO - 1;m) - ^PO;m)

= np(r-, n) + (ji — r)P(r + 1; m)
r

11. 2 (r “ toO'; I*) = r - m + mPO; aO - rP(r + 1; m)

j-o
12. rP(r + 1; m) = rP(r; M) - np(r - 1; /t)

13. mPO; m) — rP(r + 1; m) = mPO — 1; m) — rP(r; fi)

14. = P(r ~ 1; ^ ~ P(‘r’ ^

16‘ ^d^ = P(r “ 1; ^

16.
Jo
f p(r; Xt) dt = 7 P(r + 1; XT)
X

17. jTfpCr; Xt) dt = JL^^±^!P(n + r + 1; XT), n = 0,1, 2,. .

18. (TP(r; Xt) dt = i £ PO; XD = PP(r; XT) - f P(r + 1; XT)

Jo Aj~r+1 X
 APPENDIX 3 443

rpn+l
U\T)=^P(r-,\T)
IS. f*l-P(r;U)dt-±£^P<.* +

_ _J_ / 1 \ (n + r)\
Xn+1 \n + 1 y (r — 1)!
n = 0,1, 2,. .

20. f (T - t)np{r-, X# dt« ]C ^ ^ p(* + r + 1>'Xr)

21. fr (T - 0»P(r; Xt) <tt = £ (“ ^(s) x^ (J+l) Kxr^1F(r; XP)

JO 2=0 x /

(x + r)!
P(x + r + 1; XT)]
(r - 1)!

22. Jq y(f; a, 0)p(r; Xi) dt =bN(r;a+ 1, x £

23. f“ y(t; <x, 0)p[r; X(t + T)] * = £ P(r “ P> xT>6" (* ° + lj M^s)
Jo 3=0

n
24. JJ J*q r(*;«>0)p(r; Xt) dT * = x [x “ %blT ij'’a + 1’ x + *)]
t+T
y(t- a, P)p(r; Xt) dr dt

- iA [i
L
- j»o»=o
Z £ rO - <; («■«
v
+ !■ r+l)]
26. A{j*i(r)} = -(r - l)p(r ~1;m) = ~MP(r ~ 2> **)

27. A{1 - m(r + 1)} = A-£|jjpO»} = rp(j] ^ = ^ ~

28. a O' - r)pO; m)| = **)

APPENDIX 4

In the following properties, $(r) are defined as follows:

1
e~rl/2; $(r) = £ 4>{x) idx

where r can be positive, negative, or zero.

1. J x4>(x) dx = <t>(r)

2. j xn4>(x) dx = 7*n~V(r) + (n — 1) j xn~2<l>(x) dx, n = 2,3,.

3. y* x2<£(#) etc = <i>(r) + r<Kr)

4. y* x?<t>(x) dx = (r2 + 2)<£(r)

/ * i
xn<3?(x) dx = ———rn+1$(r) + n
1 z*00
± J xn+14>(x) dx,
n = 0,1,
6. y* <£(x) etc = <^>(r) — r$(r)

7. / x#(x) dx = i[(l — r2)<i>(r) + r0(r)]

8. / x2<£(x) dx = |[(r2 + 2)0(r) — r3<i>(r)]

9- !rl Ti> ~ W"(l’ T'h T" T’ > °

^ ^ ^ [|^] - I ^ [^]
 APPENDIX 4 445

io' /r (aT* [^i * ■ ^ t,) - ^ ^ ^ t■> °

IF.fe D - 2 (l +1) ♦ * [t^s]

+ X* (* " f) *““<t [<OT)‘°]

11. Let

J.(z,T,,T,)- f*l4tp*\jp^\i‘’ " * 2’ ■' ■ ; T‘J T‘ > °

Then
2(DTi)wTt Fx - XTxl 2(DTi)inT% Fx - XT21
Jn+i(a:, Ti, Ti) = ^LC-DTi)1,2J X2 ^ L(^^2)1/2 J

+ (2n ^ l)D J^x> T'> T*) + fj^n-x(x, Tl, ro,

n = 0, 1, 2,.. .

“■ J"'-w'(x’T‘>’ T*T‘>0
, m if ,, sdx . 3Dn. rx-xri
w2(x, T)=^|* + X + X2
X ^ X2 J^Lc^)1/2]
_
_2CP?y/2rT - aoi .nx — xr~i
x2 |_ x2 J L (DT)1,2J
+

13. r $ * [f^] * = "M* TO - Ti), T„ T, > 0

1r ox~u
6z2Z> ,, 15xD2
±ojsJL/ , 15Z)3
W.(*, T) = + — + “5?“ + 14^1

14. Let
Rn{x, Ti, T2) = JJ* [|r^] dt, n = 0, 1, 2,... ; Ti, T2 > 0

Then _ XT 1
«.(*, ^1, Ts) = r§+1 $ [4rp] “ ^+1n+1 $ [w4J

- 2(n+l) Jn+l(a:’ ^ Tl) ~ 2(» + 1) ^ Ts)

446 APPENDIX 4

16. fx Rn(x, Ti, T2) = —Jn(x, Tu T2)

and
RJx, Th T2) = - £ Ml Th T2) d,Z

16- JTt * dt = v°(x> - v0(x, M, Th t2> o

it / - \t — x -■ D 1 a \x - xr1 i {DT)m f* ~ xrl

7o(x, T) L? X 2X2J ^ 1_(DT)1/2J X ^ L(-Z>r)1/2J

_|_ JL e2X*/D $ [" 5_+J^"|

‘ 2X2 6 l(DTyi*]

17. fT't$ [^=y^] dt = Fi(a;, T2) - Di), r„ T2 > 0

y (x f\ — ir f2 _ — _ $ r ~ ^~i
x.
J> - 2L X2 X3 2X4J L (DT) «* J
, (DT)ll2V,rr , 3D , 1 . fx - xn
+ 2X2 LXT+ X +Xrl(DT)u*_
D_ _ MTI ,n ^ fx + XT1
2X3 [X
i 2XJ'e L (DT)1’2]
Tx + xn [x - xn
18. eMD
*L(Dtyi*] “ *L(Dt)"*]
dMf)
19.
dr
= ~4>(r)
Vx - xn i rx - xn
20. j-$
dx L(D<)1/2J_ (doi/20L(doi/2J

21- [^DO^] = 2 [(DO1,2 + D1'2*3'2] * [(DO1'2]

__ 3 . fx — xn r x — xn r x — xn
22- 3x * L (Dt)1/2 J “ L Dt J * L (Dt) J
23- 1 * \jpty*\ Wt = “ x2<] * [w^\
26. 1 jf ’ (z - ,)* * - /■ * *
 APPENDIX 4 447

26. j x$ dz

= | j(<r2 + m2 - r2)$ + <Kr + M)4> (rV£)}

27. jT“ (5^) dr = | (m3 + 3/icr2 - r*)$

+ | (r2 + Mr + M2 + 2<r2)4> ^^

28. jT ' « (ttJ!) * - | {* (^r6) - «’*"** (^j

Ackoff, R. L., 68 Costs (cont.):
Arnoff, E. L., 68 unit cost, 11, 422
Arrow, K. J., 3, 4, 25, 26, 2X2, 219, 289, constant, 162, 237, 248, 336, 367
364, 395 function of quantity (see Quantity
discounts)

Backorders, 9, 42 Dalleck, W. C., 69, 219, 315, 429

approximate treatment, 162, 237 Demands:
exact treatment, 175, 245 deterministic, 29, 336
Beckman, M., 395 dynamic, 323
Bellman, R., 327, 351, 364, 395 forecasting, 350, 4D6, 413, 417
Bharucha-Reid, A. T., 149 independent, 362
Binomial distribution (see Probability) negative binomial, 203, 210, 310
Bowman, E. H., 68, 314 example, 302
Brown, R. G., 429 integrality, 40
Bryan, J., 150 normally distributed, 167,169, 191, 256
Barrington, R. S., 149 examples, 172, 242, 299, 306
one-at-a-time, 111, 161, 176, 382
Poisson, 109,181, 204,214, 245, 274,308
Christmas tree type problems, 297
examples, 190, 195, 253
examples, 299 stochastic, 106, 159, 237, 265, 268, 334,
Churchman, C. W., 68
345, 349
Constraints: time dependent, 310, 323, 332, 336, 345,
dollar investment, 57, 213, 306, 402
349
dynamic programming, 305
Discounting:
examples, 59, 306 continuous time, 80, 377
floor space, 54
discrete periods, 23
fraction of time out of stock, 217
Discounts (see Quantity discounts)
inequalities, 436
Dreyfus, S., 351, 352
number of backorders, 219
Dvoretzky, A., 3, 26, 364, 396
number of orders, 56, 213 Dynamic programming (see also Func¬
several, 57, 330 tional equations):
volume, 304, 324 continuous variables, 335
Convexity, 171, 220-222, 259, 367 deterministic demand lot size model, 336
A-convexity, 371
example, 343
optimality of (Rr) policy for convex
exposition of technique, 323
costs and constant unit cost, 367
flyaway kit problem, 324
(Qr) model, 194
example, 328
Convolutions (see Probability):
use in (R,r) models, 390 Exponential distribution (see Probability)
use in (R, r, T) models, 268 Exponential smoothing, 417
Costs:
backorder, 18 Feller, W., 149
goodwill, 18, 298 Fetter, R. B., 68, 69, 219, 314, 315, 429
inspection, 11 Flyaway kit problem, 304, 324
inventory carrying charges, 13, 421 Forecasting (see Demands)
lost sales, (see Lost sales) Fredholm integral equation, 136
procurement, 11, 422 Friedman, L., 69
receiving, 12 Functional equations:
review, 235, 237, 263, 281 explicit solution:
setup, 12, 50 general case, 388
stockout, 18, 25, 217, 420 periodic review model, 391
transportation, 11 Poisson demand case, 382
449
450 INDEX

Functional equations (cont.): Lead time (cont.):

periodic review systems, 361 gamma, 203
transactions reporting systems, 376 stochastic, 160, 200, 204, 285, 419
Lost Sales, 9
Galliher, H. P., 184, 203, 207, 219 approximate treatment, 168, 241
Gamma distribution (see Probability) certain demand lot size models, 47
Generating functions (see also Moment constant lead time with one order out¬
generating functions): standing, 197
geometric, 119 costs, 18
negative binomial, 119 dynamic programming, 394
number of orders, 216 exponential lead time when Q = 1, 211
Poisson, 119 periodic review models, 241, 282
stuttering Poisson, 123 (Qr 7) models, 168, 197
sum of independent random variables, Lot size models:
121 approximate models, 162, 168
Geisler, M. A., 304, 315 example, 172
Geometric distribution (see Probability) constrained models (see Constraints)
Glicksberg, I., 364, 395 det erministic demand:
Gross, O., 364, 395 backorders not permitted, 29
example, 39
Hadley, G., 254, 289, 315, 352 backorders permitted, 42
Harris, F. W., 3, 26, 34 example, 47
Harris, T., 3, 25, 289, 364, 395 discrete review formulation, 62
Horizon (see Planning horizon) dynamic, 336
example, 343
Inventory: finite production rate, 50
carrying charges (see Costs) example, 53
net, 46, 181 historical, 2
on order, 37, 187, 255 lost sales, 47, 197
position, 46, 181 no stockouts, 29
systems (see also Ordering policies): Poisson demands and constant lead
multiechelon, 4 time, 181
multiple sources, 5 stochastic lead time, 160, 165, 200
multiple stocking points, 5 Lundh, B., vi, 188, 253
structure of, 4
Items, 7-8 Magee, J. F., 69, 429
Marginal analysis, 312
Jewell, W. S-, 150 Marginal probability distribution (see
Joint probability distributions (see Prob¬ Probability)
ability) Markov processes, 176, 203, 206, 211, 246,
266
Karlin, S., 4, 26, 212, 219, 280, 395 continuous in space and time, 135
Karr, H. W., 304, 315 discrete in space and continuous in
Kemeny, J. G., 150 time, 129
Kiefer, J., 3, 26, 364, 396 example, 133
discrete in space and time, 127
Lagrange Multipliers, 55, 214, 217, 304, discrete in time and continuous in
433 space, 135
Lead time, 10 Marschak, J., 3, 25, 289, 364, 395
constant, 29, 181, 197, 245 Mass6, P. B. D., 364, 396
crossing of orders, 201, 207, 345 May, D. C., 149
exponential, 203, 205, 211, 286 Mirkil, H., 150
INDEX 451

Molina, E. G, 150 Order costs (see Costs, Procurement)

Moment generating function (see also Gen¬ Ordering policies, 21 (see also Operating
erating functions): doctrines)
exponential, 126 (nQ, r, T) (see (nQ, r, T))
gamma, 126 (Q, r) (see (Q, r»
normal, 125 (R, r) (see (R, r»
sum of independent random variables, (R, r, T) (see (R, r, T»
123 (R, T) (see (R, T))
Morgenstern, 0., 24, 26
Morris, J. E., vi
Morris, W. T., 69 Palm, C., 219
Morse, P. M., 150, 184, 203, 207, 219, 289 Periodic review systems, 235
Multiechelon systems, 4 approximate models, 237, 253, 255
Multiple sources, 5 certain demands, 62, 336
Multiple stocking points, 5 comparison of doctrines, 280
dynamic programming (see Dynamic
Negative binomial distribution (see Prob¬ programming)
ability, Demands) (nQ, r, T) model (see (nQ, r, T))
Net inventory (see Inventory) Poisson demands and constant lead
Newsboy type problem, 297 times, 245
examples, 299 quantity discounts, 288
Normal distribution (see Probability) (R, r, T) model (see (R, r, T))
(nQ, r, T}: (R, T) model (see (R, T))
approximate form for large Q, 255 stochastic lead times, 285
exact formulation of backorders case for Planning horizon:
Poisson demands, constant lead fixed, 308, 345
times, 245 stochastic, 309, 349
example, 253 Poisson distribution (see Probability)
normally distributed demands, 256 Prediction (see Demands)
(Q, r) model as limit when T approaches Probability:
0, 263 basic laws, 83
binomial distribution, 87, 99
Obsolescence, 8 examples, 89
date known, 308, 345 conditional probability, 84, 117
stochastic, 309, 349 continuous random variables, 91
On order inventory (see Inventory) convolutions, 118 (see also Convolu¬
Operating doctrines (see Ordering policies) tions)
comparison of periodic review doctrines, examples, 119
235, 280 discrete random variables, 86
comparison of periodic review and trans¬ Erlang process, 112
actions reporting doctrines, 235, expected values, 98
245, 281 exponential distribution, 96
determination of without specifying gamma distribution, 95, 102, 112, 126
stockout costs, 217 examples, 96
Optimal policy: generating function (see Generating
calculation of optimal (R, r) by dynamic functions)
programming, 374 geometric distribution, 101, 113, 119
conditions for optimality of (R, r) policy, joint distributions, 114
364, 367 marginal distribution, 115,166, 203, 210
conditions for optimality of (R, r, T) example, 115
policy, 364, 367 Markov processes (see Markhov proc¬
effect of variable unit cost, 288, 367 esses)
452 INDEX

Probability (cont.): {R, r, T) model (cont.):

moment generating function (see Mo¬ gamma distribution, 278
ment generating function) normal distribution, 276
negative binomial distribution, 90, 100, optimality conditions, 364, 367
114, 117, 119, 122, 203, 210 Poisson distribution, 274
normal distribution, 94, 101, 125, 140, (R, T) model:
191 approximate model, 237
properties, 143, 447 comparison with (nQ, r, T) model, 253
Poisson distribution, 90, 100, 119 example, 242
properties, 136, 443 exact model in backorders case, 260
statistical independence, 116 lost sales case, 241
stuttering Poisson distribution, 113,123, normal demand distribution, 261
246, 248 Poisson demand, 260
Procurement lead time (see Lead time)
Production rate, 50
Sasieni, M., 69
Scarf, H., 4, 26, 212, 219, 364, 367, 371,
Q — 1 models, 204 395, 396
(Q, r) model: Simond, M., 184, 203, 207, 219
approximate treatment of backorders Single period models, 297
case, 162 constrained multiple item problems, 304
approximate treatment of lost sales example, 306
case, 168 time dependent costs, 308
example, 172 Snell, J. L., 150
comparison of lost sales and backorders Standard deviation, 99
cases, 169 Stockout costs (see Costs)
exact formulation for backorders case Stuttering Poisson distribution (see Prob¬
with Poisson demands and con¬
ability)
stant lead time, 181
example, 195
limit of (nQ, r, T) as T approaches 0, Takdcs, L., 9, 207, 219
263 Teicholz, P., vi, 188, 253
normal approximation, 191 Thompson, G. L., 150
Quantity discounts: Transactions reporting, 159, 376
all units, 62, 212, 288
example, 65
Utility, 24
effect on optimal policy, 288, 367
incremental, 66, 212, 289
example, 68 Variance, 99
Queuing, 129 (see also Markhov processes) Von Neumann, J., 24, 26

Raymond, F. E., 3, 26, 69 Wadsworth, G. P., 150

Reporting systems: Wagner, H. M., 352
comparison of periodic review and trans¬ Welch, W. E., 69
actions reporting, 281 Whitin, T. M., 3, 26, 69, 220, 254, 289,
periodic review, 160 315, 352, 430
transactions reporting, 159 Wilson, R. H., 3, 34
(R, r) model, 364, 367, 381 Wolfowitz, J., 3, 26, 364, 396
(R, r, T) model, 265
continuous demands, 275
dynamic programming (see Dynamic Yaspan, A., 69
programming) Youngs, J. W. T., 220