4/y^fr
■-s
NAVAL B€S£flfiCH
LOGISTICS
DECEMBER 1980
VOL. 27, NO. 4
OFFICE OF NAVAL RESEARCH
NAVSO P-1278
<tf?7-6
NAVAL RESEARCH LOGISTICS QUARTERLY
EDITORIAL BOARD
Marvin Denicoff, Office of Naval Research, Chairman Ex Officio Members
Murray A. Geisler, Logistics Management Institute
W. H. Marlow, The George Washington University
Thomas C. Varley, Office of Naval Research
Program Director
Seymour M. Selig, Office of Naval Research
Managing Editor
MANAGING EDITOR
Seymour M. Selig
Office of Naval Research
Arlington, Virginia 22217
ASSOCIATE EDITORS
Frank M. Bass, Purdue University
Jack Borsting, Naval Postgraduate School
Leon Cooper, Southern Methodist University
Eric Denardo, Yale University
Marco Fiorello, Logistics Management Institute
Saul I. Gass, University of Maryland
Neal D. Glassman, Office of Naval Research
Paul Gray, Southern Methodist University
Carl M. Harris, Center for Management and
Policy Research
Arnoldo Hax, Massachusetts Institute of Technology
Alan J. Hoffman, IBM Corporation
Uday S. Karmarkar, University of Chicago
Paul R. Kleindorfer, University of Pennsylvania
Darwin Klingman, University of Texas, Austin
Kenneth O. Kortanek, Carnegie-Mellon University
Charles Kriebel, Carnegie-Mellon University
Jack Laderman, Bronx, New York
Gerald J. Lieberman, Stanford University
Clifford Marshall, Polytechnic Institute of New York
John A. Muckstadt, Cornell University
William P. Pierskalla, University of Pennsylvania
Thomas L. Saaty, University of Pittsburgh
Henry Solomon, The George Washington University
Wlodzimierz Szwarc, University of Wisconsin, Milwaukee
James G. Taylor, Naval Postgraduate School
Harvey M. Wagner, The University of North Carolina
John W. Wingate, Naval Surface Weapons Center, White Oak
Shelemyahu Zacks, Virginia Polytechnic Institute and
State University
The Naval Research Logistics Quarterly is devoted to the dissemination of scientific information in logistics and
will publish research and expository papers, including those in certain areas of mathematics, statistics, and economics,
relevant to the over-all effort to improve the efficiency and effectiveness of logistics operations.
Information for Contributors is indicated on inside back cover.
The Naval Research Logistics Quarterly is published by the Office of Naval Research in the months of March, June,
September, and December and can be purchased from the Superintendent of Documents, U.S. Government Printing
Office, Washington, D.C. 20402. Subscription Price: $11.15 a year in the U.S. and Canada, $13.95 elsewhere. Cost of
ndividual issues may be obtained from the Superintendent of Documents.
The views and opinions expressed in this Journal are those of the authors and not necessarily those of the Office
of Naval Research.
ssuance of this periodical approved in accordance with Department of the Navy Publications and Printing Regulations,
P-35 (Revised 1-74).
STORAGE PROBLEMS WHEN DEMAND IS "ALL OR NOTHING"*
D. P. Gaver and P. A. Jacobs
Department of Operations Research
Naval Postgraduate School
Monterey, California
ABSTRACT
An inventory of physical goods or storage space (in a communications sys-
tem buffer, for instance) often experiences "all or nothing" demand: if a
demand of random size D can be immediately and entirely filled from stock it
is satisfied, but otherwise it vanishes. Probabilistic properties of the resulting
inventory level are discussed analytically, both for the single buffer and for
multiple buffer problems. Numerical results are presented.
1. INTRODUCTION
The usual storage or inventory problems involve demands imagined to occur randomly,
and to be capable of reducing any available stock to zero, or even beyond, when backordering is
permitted. Yet in many situations at least one component of total demand is "all or nothing;"
that is, it reduces inventory only if it can be entirely satisfied by the inventory present, and oth-
erwise seeks another supplier. Here are examples.
(a) A manufacturer's warehouse is filled with a certain item at the beginning of the sel-
ling season; let / denote the initial inventory. Suppose that demands occur as follows: a mes-
sage is sent requesting that 7), items be shipped from inventory, but only if the entire order can
be filled. That is, the demand is satisfied if 7^ ^ /, in which case inventory level is reduced to
7(1) = 7 — D\\ while if D[ > I the inventory remains unchanged and 7(1) = /. Allowing for
no replenishment, the second demand, of size D2, interacts with inventory 7(1), so that it is
filled if D2 < 7(1), but is not placed if D2 > /(l). The process continues along these lines
until the selling season is over and there are no more demands.
(b) A buffer storage device used to contain messages prior to their batch transmission
has capacity 7. Messages of length (7),, / = 1,2, . . .} approach the buffer successively, and are
admitted on an "all or nothing" basis, just as was true of demands for physical inventory in (a)
above. Once again rejection will occur, and more frequently to large demands (messages) than
to short ones.
(c) A system of many buffer storage devices is used to contain messages prior to their
batch transmission. Each buffer has capacity /. Messages of length [Dn i = 1,2, . . .} approach
the device and are successively admitted to the first buffer until there is a demand that exceeds
its remaining capacity. The first buffer is left forever and the demand that exceeds the first
"This research was supported by the National Science Foundation under NSF ENG 77-09020. ENG 79-01438 and MCS
77-07587, and by the Office of Naval Research under Contract Task NR042-411.
530 DP GAVER AND PA. JACOBS
buffer, plus successive demands, applies to the second buffer until one occurs that exceeds the
remaining capacity. This demand then applies to the third buffer, and so on. As a result there
will be some unused capacity in each buffer. For a similar problem see the paper of Coffman,
Hofri, and So [2]. For related, although not identical formulations, see Cohen [3], Gavish and
Schweitzer [6], and Hokstad [7].
In Section 2 we will discuss some models for the situations in Examples (a) and (b). We
compute such items as the distribution of the amount of inventory left at some time t and the
distribution of the times of successive unsatisfied demands.
In Section 3 we next consider a model for Example (c), and derive equations for the lim-
iting distribution of used capacity of a buffer and the expected used capacity of a buffer. It
seems to be difficult to obtain simple analytic solutions to these equations, but certain illustra-
tive numerical results are provided.
2. THE ONE-BUFFER INVENTORY PROBLEM
Suppose that demands for available stock occur according to a compound Poisson process:
if N, is the number of demands that occur in (0, /], then [N,\t ^ 0} is a stationary Poisson
process with rate \; the sizes of successive demands [D,} are independent with common distri-
bution F. Assume that there are no replenishments of inventory. Let {/,; t ^ 0} denote the
stochastic process describing available inventory at time /, and let {/(«); n = 0, 1, ...} be the
stochastic process of available inventory following the «th demand. It is apparent from our
assumptions that both {/,} and {/(«)} are Markov processes.
2.1 Functional Equations for the Amount of Available Inventory
Let
(2.1) 4>Ut) = E[e~sl']
be the Laplace transform of the available inventory at time t. Similarly, let
</,(*,«) = E[-*nn)].
Properties of the available inventory can be studied in terms of <f> and <//. It may be shown by
using conditional expectations that 0 satisfies the following differential equation.
(2.2) 4J..x£[.-*JoV_ „,(*)].
Further, >// satisfies the following difference equation
(2.3) if,U n + \) = ^{s,n) + E \e~snn) J"o' " (esx - l)F(rfx)].
Differentiation with respect to s at s = 0, or a direct conditional probability argument, now pro-
duce equations for £[/,] and £[/(«)]:
(2.4) -^ £[/,] - -\E [/o''x F(dx)\
and
£•[/(« + 1)] = E[I(n)] - E [fo'(n) x F(dx)\
In general, no explicit solutions for the expected values are available, but a simple lower
bound results from rewriting (2.4) as follows.
ALL OR NOTHING STORAGE PROBLEMS
(25) *
£[/,] = -\E \It f \' ±F{dx)\
0 It
> -\E[I,FU,)]
> -\FU)E[It],
from which one sees that
(2.6) £[/,] > /exp [-\F(I)t]
and similarly
£[/(«)] ^ /[l - -F(/)]",
so the expected available inventory declines by at most an exponential rate.
2.2. Explicit Solution When the Demand Distribution is Uniform
Although Equation (2.2) seems to be quite intractable for most demand distributions, it
can be solved completely when Fis uniform:
F(x) =
- 0 ^ x < c,
c
1 C > X
and c > /. In this case (2.2) can be expressed as
(2.7) M = ,£L^r/'(^_1)^l
Bt Jo c
- KE
-si ~sll I 1
1 - e ' e I,\
In other words, <f> satisfies a first-order (quasi) linear partial differential equation with ini-
tial condition 0(s, 0) = e~sI. Standard procedures (Sneddon [8]) easily yield the solution
,- Qx 1 -<f>(s,t) _ 1 - exp [-(s + (X/c)t)I]
a8) 5 " , + (X/C)t
which gives the desired transform. Passage to the limit as s — • 0 in (2.8) shows that
(2.9) £[/]=l-exp[-(Xf/c)/]
(K/c)t
This formula can also be derived by first finding an expression for the /cth moment of /,, and
then employing a Taylor series argument.
In order to invert the transform in (2.8) note that
(2.10) f'e -«P{I, >x}dx= l~f<*0 = l-exp[-(5 + (X//c))/]
^o s 5 -I- (\/c)r
which is the transform of a truncated exponential distribution. Thus, by the unicity theorem
for Laplace transforms,
532 DP. GAVER AND PA. JACOBS
Iexp[-(X//c)x] 0 ^ x < /,
0 J<*
Note that the distribution of /, is absolutely continuous in the interval (0, /) but that there is a
jump at /corresponding to the occurrence of no demand less than, or equal to, /in (0, t]:
(2.12) P{I,= l) = exp[-XtU/c)}.
2.3. The Expected Number of Satisfied Demands
Supposing that an initial inventory, or storage capacity, / prevails, it is of interest to com-
pute the probability that a demand is satisfied, and the expected number of demands satisfied in
an interval of length t. First notice that if a demand of size D{t) appears at time /, at which
moment /, is available, then
P[D(t) < I,\I,} = FU.)
is the conditional probability that the demand is satisfied. When F is uniform, as is presently
true, we may remove the condition to find that
P[D(t) ^ /,} = E[F(I,)}-
U\ = 1 -exp[-0u/c)/1
\ c\ \t
If S(t) is the number of demands satisfied during the time interval (0, f], then since demands
arrive according to a Poisson process with rate A.,
(2.13)
E[S(t)} = X J'E[FUu)]du = X r;i-exp[-(X»/c)/]
Jo Jo \u
where E\(-) is an exponential integral; Abramowitz and Stegun [1], and y — 0.5112... is
Euler's constant.
2.4 The Time of the First Unsatisfied Demand and the Amount
of Unused Inventory at that Time
As before F is the common distribution function of the successive demands. Now let t
be the time of the first unsatisfied demand. Then
P{t > t\N,= n) = P{D] < /, 2)2< ?--'£| D„ < I -D\ £>„_,}
= F{n){I)
where F(n) denotes the /rth convolution of Fwith itself. Hence,
(2.14) P{t > /} = £ e-K'-^I^F{") (/).
Explicit expressions for the distribution of t can be obtained in some cases. If F is uni-
form on [0, c] with c ^ /, then
(2.15)
ALL OR NOTHING STORAGE PROBLEMS 533
where 70(z) is a modified Bessel function of the first kind of the zeroth order. In this case,
(2.16) E[t\ = -j- exp{//c} = ^-exp{//2F[D]}.
If Fis exponential with mean l/>u. , then
(2.17) P{r > t) = £ *-*'&¥■ £ e-»l&£-
n=0 rt- k=n k-
and
/ 1
(2.18) E[t]= f [1 +^7] =
1"' £[/>]!■
Note that if /is small relative to F[/J>], then the expected time to first unsatisfied demand
when Fis exponential will be greater than the expected time when Fis uniform. However, for
/large relative to E[D] the expected time for /"exponential will be less than the expected time
when Fis uniform.
Let Yn be the amount of inventory present at the time of the «th unsatisfied demand.
Then for 0 ^ a < /
(2.19) P\YX > I- a) = fj R(dy)FU-y)
where
(2.20) R(y)= £ F(n)(v)
and
(2.21) 1(1- y) = 1- FU- y).
Again, explicit expressions for the distribution of Yx can be obtained for some distribu-
tions F. If Fis uniform on [0, c] for c > /, then
(2.22) P{YX > /- a) = 1 ■
If Fis a truncated exponential
(2.23) Fix) -
1 - e~»x . .
7 X < /,
1 - <?""'
1 x > I,
then
(2.24) P{YX > / - a] = 1 - [e^a - e^1] [1 - e^T1 exp [fia[l - e^T1}.
If Fhas an exponential distribution with mean 1//a, then
(2.25) P{YX > I- a} = e^{'-a).
In this last case, the distribution function of Yn can be computed by induction quite easily and
(2.26) P{Yn > I - a} = *-«*(/-•>.
534 DP. GAVER AND PA. JACOBS
Hence, when Fis expjnential
(2.27) E[Y„] - — [1 - frnft1.
np,
In principle, similar results can be obtained for other distributions, but we have found no sim-
ple expressions.
2.5. Inventory Costs and Policies
There are at least three monetary quantities which affect the profitability of an inventory
policy over a fixed interval of time (0, /]: the selling price, /r, the storage cost, a; and the cost
of lost demands, b. If the storage cost a is charged just on the basis of / (something like ware-
house size) then the total expected profit in (0,/] is
Z(/) = pil - £■[/,]) - al - bSU)
= (p- a)I- p\-t\ [l-exp[-(A//c)/]]
■ b\y + In \—l\ + £,|
HI
for the case of uniformly distributed demands; see ((2.9) and (2.13)). One can numerically
find the maximum expected profit for this case; nothing explicit seems to be available.
3. THE MANY-BUFFER STORAGE PROBLEM
In this section we will study a model for the situation of Example (c) in Section 1. Mes-
sages are successively admitted to the «th buffer until there is a message length that exceeds
the remaining capacity of the buffer. The total amount of this message is put in the (n + 1)5/
buffer and the «th buffer is left forever. Successive messages are then put in the (n + 1) st
buffer until there is a message whose length exceeds the remaining capacity of the (n + I) st
buffer; this message is put in the in + 2) nd buffer and so on.
Let / denote the common capacity of the buffers and D, denote the length of message /.
Assume {/),} is a sequence of independent identically distributed random variables with distri-
bution F having a density function /such that fix) > d > 0 for x 6 [0, /]. Let R ix) = £
Fin) (x) be the renewal function associated with F. If F(I) < 1, then we will assume that an
incoming message to the currently used nth buffer of length greater than / is sent to the
in + \)st buffer; when it cannot fit into the (n + \)st buffer, then it is "banished," i.e., sent to
some other set of buffers. The next message, however, will try to enter the (n + l)st buffer.
If this message has length greater than / it is banished and the following message will try to
enter the in + \)st buffer; all messages of length exceeding /will be banished until one appears
that is smaller than / and it will be the first entry in buffer in + 1).
This model has been studied for demand distributions Fwith F(I) = 1 by Coffman et al.
[2]. Their approach was to study the Markov process describing the total amount of inventory
or space consumed in successive buffers or bins. Here we study the process {/.„}, where L„ is
the size of the demand that first exceeds the remaining capacity of the nth buffer;
[L„; n = 1,2, . ..} is a Markov process. Let
K{x, [0,v]) = P{Ln + x ^ y\Ln = x).
ALL OR NOTHING STORAGE PROBLEMS
535
Note that
P[LX ^ y} = K(0, [0,y])
is the same as the sum of the forward and backward recurrence times at time / for a temporal
renewal process with interrenewal distribution F, see Feller [5]. Thus for y < /
(3.1) H,{y) = P[LX < *}-/,*<&) [Fiy) - FU - z)\.
Note that for y < I
X
(3.2)
Hence,
Kix, [0,y]) -
(3.3) Kix,dy) =
R(dz)[F(y)-F(I - x- z)]
if x < I - y;
Jo XR(dz)[F(y)-F(I - x - z)\
if / - y < x < I;
ji'_R(dz)[F(y)-F(I - z)}
if x > I.
[R(I - x) - RU - x - y)]F(dy) if x < I - y,
R(I - x) F(dy) if / > x > I -
kiy) F(dy) - JJ Ridz) f(y - z) + R(dy)F(y)
. if x = I - Y,
[RU) - RU - y)] F(dy) if x > I.
Note that for some 0 < a < b < /, there exists a 8 > 0 such that for all x
KHx.dy) ^ 8 for y € [a, b]
where K2(x, dy) = J K(x, dz) K(z, dy). Hence, hypothesis D' on page 197 of Doob [4]
satisfied. Thus, if
K"(x, A) = P[Ll+„ 6 A\LX= x)
for all Borel subsets A, then
(3.4) lim K"{x,A) = H{A)
exists and further the convergence is geometric
\K"(x,A) - H(A) | < ay"
for some positive constants a and y, y < 1 for all A.
Now let
//„(*) = P{L„ € [0,x]\LQ=0}.
Then a renewal argument can be used to show that for x ^ /
(3.5) ff„+iOc)- Jj[xHn * R(dy) [Fix)- F(f - y)}
+ [1 - //„(/)] f'_xR(dy) [Fix) - FU - y)\.
536 DP GAVER AND PA. JACOBS
Taking limits as // — ► °o it is seen that the distribution Hix) satisfies the following equation for
(3.6) Hix) = f^ H * R idy) [Fix) - F(I - y)]
+ [1 - //(/)] ff . R {dy) [F(x) - Fil - y)].
Equations (3.1) and (3.6) can be simplified for certain specific distributions F.
3.1 Exponential Demands
For the exponential distribution with mean 1 and x ^ /the equations are
(3.7) //,(*) = 1 - e~x- xe~x
and
(3.8) Hix) = xe~xHil) + //,(*) - e~x J* H(I - x + u) du.
3.2 Uniform Demands
For the uniform distribution on [0, e) with c > /they simplify to
(3.9)
and
(3.10) Hix) - -exp - (/ - x)\ fQ \\p\- -u\Hiu)du
x\ C' f 1 )
— I J exp u\ Hiu)du
+ [1 - //(/)] Hy(x),
for x ^ I. Similar expressions hold for x > /, but they are unimportant in the present con-
text.
Equations (3.6), (3.8) and (3.10) do not seem to yield explicit answers. As a result, we
have solved (3.8) and (3.10) numerically by iteration using the system of equations
(3.11) //„ + ,(*) = xe-xHniI) + H\ix) - e~x jj //„(/ - x + u)du
with H\ as in (3.7) and
(3.12) //„+,U) = -exp -(/-x) {' %xp \--u \Hniu)du
c c \ Jo c
:-H„iI) - -exp -/Ml - - f 'exp - - u\Hniu)du
c [ c )\ cjJo I c J
+ [1 - //„(/)] Hxix)
with H\ as in (3.9). For the cases carried out the convergence is rapid; after n = 5 iterations,
very little change is noted and convergence has occurred for most practical purposes.
ALL OR NOTHING STORAGE PROBLEMS 537
Next let Yn be the amount of storage space used in the «th bin; the distribution of Y„ is
denoted by G„(x), and
G(x) = lim P{Yn ^ x} = lim G„(x)
is the long-run distribution. By probabilistic arguments and (3.4)
(3.13) G{x) = fjff * R{dy) F(I - y) + [1 -//(/)] f^R (dy) F(I - y)
where F(I — y) = 1 — F(I — y) and the long-run average expected capacity of a bin that is
actually used is
A = f xG(dx).
Jo
For the case in which Fis exponential with unit mean
(3.14) A = I - [\- //(/)] [1 - <?-1 - e~l J"Q exH(x)dx.
For the case in which fis uniform on [0, c] with c ^ /
-2 (*'//(«)
Jo
)du + exp — /
I c
c' ( 1 1
1 exp u\ H{u)du
JO c 1
c exp — 1\ + c
\c \
+ — / + c exp — 1\ - c\
(3.15)
+ //(/) 2/-
Numerical solutions were obtained for Equations (3.14) and (3.15) by first computing the pro-
babilities H„(x), n= 1,2, ..., 10 iteratively from (3.7) and (3.11) for the exponential
demand case, and from (3.9) and (3.12) for the case of uniform demands. Our technique was
simply to discretize x: x, ■■ = jh, h = I/N, N being the number of x- values at which H„(x) is
evaluated (values of A/ from 200-1200 were utilized in order to obtain two-significant digit accu-
racy). The integrals were then approximated by a summation, i.e. Simpson's rule. Having the
values of H„(xj) it is possible to calculate those of Hn+\(xj), and from these the values of
G„(x) and the mean usage, E[Y„], may be calculated by numerical integration. In the case of
exponential demand very simple upper and lower bounds were obtainable; such bounds were
not tight enough to be useful for the uniform case.
The following table summarizes the numerical results. We have compared demand distri-
butions that result, as nearly as possible, in the same probability that an initial demand on an
empty bin will be rejected. We have tabulated the expected level to which the bin is filled. It
is interesting that the limiting bin occupancy is 0.75 when a uniform demand over the range of
the bin size is experienced. This result has been obtained analytically by Coffman et al. [2]; in
that paper simple and elegant analytical expressions for G and H also appear for this case. The
considerable similarity of the numbers in the rows of the table is notable; apparently the long-
run bin occupancy is only slightly larger than is that of the first bin, and the occupancy experi-
enced for uniform demand is only slightly larger than for exponential. Further investigations to
examine the reasons for this insensitivity would seem to be of interest.
ACKNOWLEDGMENTS
D. P. Gaver wishes to acknowledge the hospitality of the Statistics Department, University
of Dortmund, West Germany, where he was a guest professor during the summer of 1977, and
where part of this work was carried out.
538
DP. GAVER AND PA. JACOBS
Expected Fraction of Bin Filled
ifn-E[Y„] + I)
Rejection Probability
Exponential Demand
Uniform Demand
F(I)
A /oo
/i
/oo
0.00
- -
0.76
0.75
0.05
0.74 0.75
0.74
0.74
0.10
0.69 0.70
0.72
0.72
0.15
0.65 0.66
0.68
0.69
0.20
0.60 0.62
0.64
0.66
0.25
0.56 0.58
0.60
0.62
REFERENCES
[1] Abramowitz, M. and I. A. Stegun, Handbook of Mathematical Functions. National Bureau of
Standards, AMS 55, Washington, D.C. (1965).
[2] Coffman, E.G., Jr., M. Hofri and K. So, "A Stochastic Model of Bin-Packing," Technical
Report, TR-CSL-7811, Computer Systems Laboratory, University of California, Santa Bar-
bara, California (1978), (submitted for publication to a technical journal).
[3] Cohen, J.W., "Single-Server Queues with Restricted Accessibility." Journal of Engineering
Mathematics, 3, 253-284 (1969).
[4] Doob, J.L., Stochastic Processes, (John Wiley and Sons, New York, N. Y., 1952).
[5] Feller, W., An Introduction to Probability Theory and Its Applications, II, (John Wiley and
Sons, New York, N. Y., 1966).
[6] Gavish, B., and P. Schweitzer, "The Markovian Queue with Bounded Waiting Time."
Management Science, 23, 1349-1357 (1977).
[7] Hokstad, P., "A Single-server Queue with Constant Service Time and Restricted Accessibil-
ity." Management Science, 25, 205-208 (1979).
[8] Sneddon, I., Elements of Partial Differential Equations, (McGraw-Hill, New York, N. Y.
1957).
RELIABILITY GROWTH OF REPAIRABLE SYSTEMS
Stephen A. Smith and Shmuel S. Oren*
Analysis Research Group
Xerox Palo Alto Research Center
Palo Alto, California
ABSTRACT
This paper considers the problem of modeling the reliability of a repairable
system or device that is experiencing reliability improvement. Such a situation
arises when system failure modes are gradually being corrected by a test-fix-
test-fix procedure, which may include design changes. A dynamic reliability
model for this process is discussed and statistical techniques are derived for es-
timating the model parameters and for testing the goodness-of-fit to observed
data. The reliability model analyzed was first proposed as a graphical technique
known as Duane plots, but can also be viewed as a nonhomogeneous Poisson
process with a particular mean value function.
1. INTRODUCTION
Predicting the reliability of a system or piece of equipment during its development process
is an important practical problem. Reliability standards are often a major issue in the develop-
ment of transportation facilities, military systems, and communication networks. For commer-
cial products that are to be leased and maintained in a competitive marketplace, system reliabil-
ity estimates strongly influence predicted profitability and customer acceptance. When consider-
ing a system that is modified in response to observed failures, most classical statistical estima-
tion techniques are not applicable. This is because the system reliability is improving with time,
while most statistical techniques require repeated samples under identical conditions.
A frequently used graphical model of reliability growth of repairable systems is known as
"Duane Plots," proposed by J. T. Duane [9]. This model is based on the empirical observation
that, for many large systems undergoing a reliability improvement program, a plot of cumula-
tive failure rate versus cumulative test time closely follows a straight line on log-log paper.
Several recent papers present applications of Duane plots, e.g., [4], [9] and [10]. Estimating
the parameters of the Duane model, i.e., the slope and intercept of the straight line fit, is some-
what difficult to do directly on the graph [5]. Weighted least squares and regression techniques
are sometimes used ([9], [10]) to obtain parameter values.
An underlying probabilistic failure model that is consistent with the Duane reliability
model is the nonhomogeneous Poisson process (NHPP) whose intensity is total test time raised
to some power. (See [7] and [8]). Assuming the sample data consists of all the individual
failure times, Crow [7] derived maximum likelihood estimates for the Duane model parameters
and a goodness-of-fit test based on the Cramer-von Mises statistic (Parzen [12, p. 143]). A
more general NHPP model was proposed by Ascher and Feingold [1], which also used
•Now with Dept. of Engineering-Economic Systems, Stanford University, Stanford, CA.
540 S A. SMITH AND S.S. OREN
the Cramer-von Miies statistic for goodness-of-fit testing. Critical values of this statistic, how-
ever, must be obtained by Monte Carlo simulation for each sample size. Crow [7, p. 403] cal-
culated and tabulated values for sample sizes up to sixty. These parameter estimates and
goodness-of-fit test deal effectively with Duane model applications having small sample sizes.
The facts that all failure times must be stored and the goodness-of-fit measure must be
evaluated by simulation make this approach difficult for larger sample sizes. A recent paper by
Singpurwalla [13] proposes a time series model for reliability dynamics. This model can, of
course, be applied to any type of reliability trend data, but requires data tabulation at a larger
number of time stages and does not have the intuitive appeal of the Poisson process for model-
ing failure occurrences in certain systems.
Our paper develops statistical estimators for the Duane model parameters based on tabu-
lating the number of failures between fixed points in time. This approach has the advantage of
using "sufficient statistics" for the data collection, i.e., the dimension of the data does not
increase with sample size. Parameter estimates are obtained by maximum likelihood and a
goodness-of-fit test based on the Fisher chi-square statistic is derived. This test has the advan-
tage that chi-square tables are readily available for all sample sizes and significance levels. The
accuracy of the chi-square test decreases, however, as the sample size gets small. Sample sizes
for which the techniques of this paper apply are found in developmental systems that experi-
ence frequent, minor failures such as paper jams in photo copy machines, voltage fluctuations
in power supply systems, faults in semiconductor manufacturing processes, etc. The last sec-
tion of this paper illustrates the application of the estimation and goodness-of-fit techniques to a
representative set of simulated failure data.
Regardless of how the parameters of the Duane model are obtained, considerable caution
is required when extrapolating reliability trends beyond the observed data to future time points.
Major breakthroughs or setbacks in the reliability improvement program may cause significant
deviations from the straight line projections. Some users recommend reinitializing the model
and shifting to a new straight line fit when major changes in the program occur. Even if one is
uneasy about extrapolating the reliability growth model to estimate future reliability, it remains
a valuable tool for obtaining a "smoothed" estimate of current system reliability. While reliabil-
ity is changing, sample sizes at any point in time are not sufficient for conventional statistical
estimation techniques. With a dynamic reliability model, past and current failure data can be
combined to obtain estimates of current reliability based on fitting all observed data.
2. THE DUANE MODEL
The Duane model states that cumulative failure rate versus cumulative test time, when
plotted on log-log paper, follows approximately a straight line. More precisely, if we let N(0,t)
represent the total number of failures observed up to time /, we have that
(2.1) log[N(0,t)/t]~- b\ogt + a,
where the fitted parameters are a, b > 0. The relationship is meaningless at t = 0 but, as most
users point out ([5], [9]), a certain amount of early data is generally excluded from the fit
because it is influenced by factors such as training of personnel, changes in test procedures, etc.
Equation (2.1) therefore implies that
N(0,t)/t = at~b, where a = log a,
for t beyond a certain point. It should be emphasized that, in all applications, time t
corresponds to cumulative operating time or test time. For the results of this paper it is most
convenient to write the Duane model as:
(2.2) N(0,t) = at13, where /3 = 1 - b.
RELIABILITY GROWTH OF REPAIRABLE SYSTEMS 541
For a fairly diverse set of observed systems, Codier [5, p. 460] has found b to be generally
between 0.3 and 0.5, corresponding to /3 between 0.5 and 0.7.
3. AN UNDERLYING STATISTICAL MODEL
In this section we describe a statistical model for the failure process that is consistent with
assuming that the observed failure data fits the Duane model. Suppose the probability that the
system fails at time t (strictly speaking in a small interval [t,t + dt)), regardless of the past, is
determined by a hazard function hit). That is,
hit)dt ~ P {the system fails in the interval^,/ + dt)),
independent of its previous failure and repair history. The expected number of failures in any
time interval [tx,t2) of operating time is then given by the mean value function
(3.1) M(th t2) = j'2 h(t)dt.
Furthermore, it can be shown (See Parzen [12, Sect. 4.2]) that N(tu t2), the number of
failures observed in some future time interval [t\, /2), has probability distribution
(3.2) p{N{tx,t2) = k)= ([A/(/,, t2)]k/k\) exp[-M(tht2)} £ = 0,1,2,... .
In addition to its mathematical convenience, this model has considerable intuitive appeal. The
simple Poisson process has been used successfully to model the failure occurrences of many
devices, or collections of devices operating in series. One may think of a system having a
nonhomogeneous Poisson failure process as a large collection of simpler devices in series, with
individual device failure modes being gradually removed with time.
The mean value function
(3.3) MU\, t2) = aUf - tf ), where a, £ > 0;
corresponding to hit) = afifi~\ is of particular interest. Crow [7, p. 405] pointed out that the
number of failures from a process with this mean value function will approximate the Duane
Model by observing that
\o%[Mit)/t] = log a + (/3 - 1) log t, where Mit) --= Af (0, t).
This means that system failure data from a NHPP with mean value function Mit) will approach
the Duane model with probability one. Conversely, this process with mean value function
Mit) is the only model with independent increments that approximates the Duane model in a
probabilistic sense for sufficiently large sample sizes. We will not give a proof of these state-
ments but refer the reader to Parzen [12, ch. 4] or Donelson [8] for a complete discussion.
4. SELECTING A STARTING TIME
The Duane reliability model and the expected number of failures in Equation (3.3) are
both nonlinearly dependent on the choice of the time origin. That is, if we begin observing
failures at time t = t0 > 0 and ignore the first NiO, t0) failures and the time interval [0, /0), we
do not obtain the same parameters a and /3 by fitting the subsequent data. Since the logarithm
is a strictly concave function, there is only one choice of t0 that can give a straight line fit to the
data on log-log paper. Specifying the operating time t0 that is assumed to have elapsed before
the beginning of the modeling process is therefore an important step.
Some users of the Duane Model ([5], [10]) suggest reducing the cumulative failures and
observation time by removing early data to obtain a straight line fit. This is done graphically by
542 S.A. SMITH AND S.S. OREN
successively shifting ihe origin to the right and replotting the data until a straight line fit is
obtained. With each shift to the right, the shape of the graph of cumulative failures versus
cumulative observation time becomes more downward bending (concave), so it is not hard to
tell when the best point has been located.
Sometimes a terminal straight line trend on log-log paper is observed before the noisy
early data is dropped. If the origin is shifted further to the right, the straight line shape will
become concave. Therefore, the most that can be said in this case is that, for / greater than
some t\, the data fits the Duane model. The statistical model (3.3) can still be applied, how-
ever, by testing to see if the number of failures NU\,t) after the first N(0,t\) fit the NHPP
with mean value function M(t) for / > t\.
5. ESTIMATING THE MODEL PARAMETERS
If the Duane model is applied graphically, the user can attempt to estimate the parameters
a and /3 by drawing the best straight line through the plotted points. This is somewhat tricky
because, with cumulative failure data, the later points should be weighted more heavily in
determining the fit. This section describes a statistical estimation procedure based on the
NHPP model of the failure process. We consider two possibilities for collecting and recording
system failure times. The first is to record the occurrence time of each failure, which yields a
sequence of observed times T\, T2, . . . , TN. This case has been analyzed by Crow in [7] and
the maximum likelihood estimates are given by
(5.1) a*=N/Tft'
and
(5.2) /3* = -A7£log(r,/7V).
A goodness-of-fit test corresponding to these estimators is derived in [7] and critical values of
the error statistic are tabulated for sample sizes 2-60.
If large numbers of failures are observed, it is often convenient to record only the aggre-
gate number of failures between each pair in a sequence of fixed time points t0, t\, . . . , t„. In
this case the data is in the form N\, N2, ■■■ ,N„, where N, = number of failures observed in
the interval [^_lf ?,). Maximum likelihood estimates and a goodness-of-fit criterion for obser-
vations in this form are developed in the next few paragraphs.
Maximum Likelihood Estimates for the Aggregated Case
We first calculate the likelihood function for the data TV,, N2, ■■■ , N„, given the time
points t0, t\, . . . , t„ and the assumed form of the mean value function in Equation (3.3). The
probability of N,- system failures in the interval [t,-i, t,) is obtained from Equation (3.2). Since
the underlying model assumes that each of the time segments is independent, the likelihood
function can be written as a product of these probabilities,
(5.3) Ka.jS) = fl P[NU,-U /,) = N,} = exp{-M(?0,O} f\ UMU,-u f,)l *'/#/!).
To simplify the calculation of the estimators, we take the log of L (a, /3), noting that max-
imizing the log will yield the same maximum likelihood estimates. From (5.3) we have
(5.4) logL(a,/S) = -a(t? - t§) + £ N,[\oga + log (^ - *£,)] - 5>gfy!.
RELIABILITY GROWTH OF REPAIRABLE SYSTEMS 543
Taking the partial derivatives (dlog L)/da = 0 and (Qlog D/9/3 = 0, we obtain the equations
for the maximum likelihood estimates,
(5.5) a* = N/Ur~ t§'), where N = £ Nt
[ log ti - Pi log f,_, log t„ - p0log /0 1
(5.6) 0= £#,
1 - Pi 1 - Po
p,= (W/)"*. '= 1. 2, ... , n, and Po = (f0/O**-
Equation (5.6) is an implicit function of /3 *, but can be solved iteratively by a computer algo-
rithm or programmable calculator, because the right hand side is strictly decreasing in j8 *. To
verify this fact, consider any two times t, t' and compute the derivative
(5.7) ' (9/9/3) [log / - 7*'logf']/[l'- 7*] = - 7* (log n2/U - T^)2, where T= tit'.
This derivative is negative and decreasing in 7Tor 0 < T, /3 < 1. The derivative of the sum in
(5.6) is a sum of terms involving the difference of the derivative (5.7) evaluated at T = f,_|//,
and t0/tn. The fact that (5.6) is decreasing in ^3* follows from the fact that (5.7) is decreasing
in T, i.e., its largest or least negative value occurs at T = t0/t„. Therefore, (5.6) has a unique
solution.
6. GOODNESS-OF-FIT CRITERION
This section describes a procedure for testing the goodness-of-fit of the observed failure
data to the NHPP. We assume that the parameters a * and /3 * are obtained from the maximum
likelihood estimates (5.5) and (5.6). From the form of (5.5), it is clear that the estimate a * is
defined in such a way that the total number of observed failures N always equals the expected
number of failures for the time period [/0. '„)• That is, a * is defined so that
W= E{N\a*. (3*) = a*(tr- >D-
Therefore, there is no difference between the observed versus predicted total number of
failures. The goodness-of-fit measure must therefore be based on the differences between the
observed incremental failures N\, N2, . . . , N„, and the predicted values
(6.1) E{Ni\a*,p*} = a*(tr- t?-]), i=\,2,...,n.
Assuming the estimate (5.5) is used for a *, the likelihood function for a goodness-of-fit
statistic will be expressed only in terms of /3*. Since the NHPP has independent increments,
the probability that a given failure occurs in the interval [tt-.\, f,) is the expected number of
failures for that interval, divided by the total number of failures, This is written as
(6.2) Pt = Pi(0*)= [a*Ur-t?-\)\l[a*{f$~-4m)\, /'- 1,2. ..., n.
where the a * parameter obviously cancels out. The likelihood function for a set of observed
failures N\, N2, .... N„, given N, is therefore the multinomial
TV
N2, ..
■ H„ = N,
which depends only on /3 *. The parameter a * can be regarded as a scale parameter that guaran-
tees the model will fit the total number observed of failures N.
544 S.A.SMITH AND S.SOREN
We now show how the goodness-of-fit of the incremental failure data can be measured by
the Fisher chi-square statistic
(6.4) x2= £ (N,~ NPl)2/NPi.
The use of this statistic as a goodness-of-fit measure is based on the following theorem, which
has been restated in the context of this discussion.
THEOREM 6.1: Let the parameters p]t p2, . . . pfn with I/?, = 1, be functions of a param-
eter /3 and let a particular value /3' be determined from
(6.5) 0= Y (ty//>,)(3fl-/dj8)|
~f 1/3 - j3'
Then the statistic (6.4) with /?, = Pi(fi'), i = 1,2, . . . , «, has approximately a chi-square distri-
bution with n — 1 degrees of freedom (x2(« - 1)) for large N. The proof of this result is quite
lengthy and can be found in [6, pp. 424-434].
To apply this result to our particular problem, we must show that (3' equals the estimator
/3* defined by Equation (5.6). Using /?,()3) as defined in Equation (6.2), and differentiating
with respect to /3, one can verify that Equation (6.5) reduces to Equation (5.6). Thus, /3' = /3*
and, since (5.6) has only one solution, the value is unique.
The chi-square error statistic (6.4) has an additional intuitive interpretation for this appli-
cation. Suppose a and /3 are the "true" parameters of the underlying nohomogeneous Poisson
process, i.e., the values to which the estimators a * and /3 * must eventually converge for very
large sample sizes. Then the "true" variance of the number of observed failures in [/,_,, /,),
i.e., the limiting value for the sample variance of a large number of observations, is given by
Var{fy|a, 0} = a(f£j ~ tf) i = \, 2, ... , n.
Consider W(a*,fS*) = £ (N, - E{N,\a *, j8 *})2/Var{/V,|a, 0},
which is the sum of square errors between the observed and estimated failures, weighted by the
true variance for each of the time intervals. Suppose we minimize this with respect to a * and
0* by solving (3 W/ba*) = 0 and (d W/Sp*) = 0. If we then substitute our "best estimates",
a * for a and /3 * for /3, these two equations reduce to the maximum likelihood equations, (5.5)
and (5.6), respectively. Birnbaum [2, p. 251-2] also points that if we minimize the chi-square
statistic (6.4) with respect to /3, the estimate obtained must approach the estimate 0' that
satisfies (6.5) as the sample size approaches infinity.
This goodness-of-fit criterion measures, in effect, how well the observed data fits a NHPP
with mean value function M(t), where 0 * is the "best" growth parameter for the observed data.
If the x2(« _ 1) statistic (6.4) exceeds the critical value at a reasonable significance level, such
as 0.05 or 0.1, the model should be rejected. Since Theorem 6.1 gives only an asymptotic
result, it is important to discuss the sample size requirements for applying it. Given the popu-
larity of this test, there has been considerable experience with various types of data. A com-
mon criterion is that TV and, in this case the time points r0, t\, ... , tn, must be such that
Npi > 10 for all /. (See Birnbaum [2, p. 248]).
RELIABILITY GROWTH OF REPAIRABLE SYSTEMS 545
7. APPLICATION EXAMPLE
As an illustration, we will determine the estimators a * and /3 * and apply the goodness-of-
fit test to the sample data in Table 1. We assume that the failures of the system were only
monitored at fixed points of time so that the observed data consists of the first two columns of
the table. These data points were generated by computer simulation with failures sampled from
a NHPP with mean value function M(t), having parameters a = 10.0, /3 = 0.5. Failure times
TV T2,
(7.1)
from this distribution can be generated sequentially from a set of random samples
. from the uniform distribution by means of the transformation
TM- W-
(l/a)log £/,+,]1//3, T0
TABLE 1
■ 0, 1, 2,
Time Interval
Observed
Failures
Predicted
Failures
Standard
Deviation
Normalized
Error
1
400 - 800
63
78
8.8
2.88
2
800 - 1200
63
61
7.8
0.07
3
1200 - 1600
54
51
7.1
0.18
4
1600 - 2000
51
46
6.8
0.54
5
2000 - 2500
68
51
7.1
5.67
6
2500 - 3000
49
46
6.8
0.20
7
3000 - 3500
34
43
6.6
1.88
8
3500 - 4000
39
40
6.3
0.03
9
4000 - 4500
39
38
6.2
0.02
10
4500 - 5000
43
36
6.0
1.36
11
5000 - 5500
39
34
5.8
0.74
12
5500 - 6000
36
33
5.7
0.27
13
6000 - 6500
28
31
5.6
0.29
14
6500 - 7000
22
30
5.5
2.13
15
7000 - 7500
35
29
5.4
1.24
16
7500 - 8000
32
28
5.3
0.57
17
8000 - 8500
22
27
5.2
0.93
18
8500 - 9000
19
27
5.2
2.37
19
9000 - 9500
19
26
5.1
1.88
23.25
The data in Table 1 was used to obtain maximum likelihood estimates a* and /3 * from
Equations (5.5) and (5.6). This was done by calculating various values of the right hand side
of (5.6) as a function of /3 until the minimizing value /3 * was determined to two decimal places.
This gave /3 * = 0.52 and a * = 7.97, where a * was determined from (5.5) with {$ * = 0.52.
The accuracy of /3 * is reasonably close to the correct value /3 = 0.5, but the estimate of a *
is off by more than 20%. Other calculations with different sets of random numbers produced
errors in both directions but generally resulted in an a * error several times larger than the /3 *
error, on a percentage basis. This seems to indicate that one is more likely to estimate slopes
of the Duane Plot lines accurately than to estimate the intercepts accurately with the maximum
likelihood estimates. Naturally, as the number of observation points in Table 1 is increased,
the estimates become more accurate. Accuracy was not improved much by increasing the
number of time points from 20, as shown in the table, to 100 and the sign of the error for a
given example generally did not change as the number of observation points was increased,
546 S A. SMITH AND S.S OREN
while holding the underlying failure points fixed. Bringing the estimate a * to within 5% of the
correct value typically required 300 to 500 observation time points for the computed examples.
To illustrate the use of the goodness-of-fit test we calculate the chi-square statistic (6.4)
for this table. The "Predicted Failures" between the various time points are given by
Np, = a*(t,i\ - t,H, r = 1,2, .... 19.
The normalized error terms as in (6.4) are given by
(N,- Np,)2/{Np,).
The sum of these errors, when compared with a x2(18) error table, is less than the critical
values 25.99 and 28.87, associated with significance levels 0.1 and 0.05, respectively.
For many applications of the model it is more important to predict the number of failures
that will occur in the next time period than to obtain accurate estimates for a and /3. In such
cases the estimators obtained from 10-20 time points appear to be sufficiently accurate. This is
because there is a range of a, /3 pairs that provide almost as good a fit to the observed data as
the optimal ones and any parameters in this range provide a satisfactory predictive model.
To illustrate the prediction accuracy of the estimates /3 * = 0.52, a * = 7.97 obtained from
Table 1, we generated simulated failures out to 40,000 time units. The number of failures
predicted by extrapolating with the estimated parameters and with the true parameters are com-
pared in Table 2. The errors in predicting failures caused by inaccuracy in estimating the
parameters is much less than the random errors that occur due to stochastic variations of the
failure process. This was found to be the case in several similar experiments.
TABLE 2
Time Interval
Simulated
Estimated
True
Standard
Failures
Extrapolation
Extrapolation
Deviation
9500 - 10,000
24
25
25
5.0
9500- 15,000
235
251
250
15.8
9500 - 20,000
412
443
439
21.0
9500 - 30,000
715
766
757
27.5
9500 - 40,000
999
1041
1025
32.0
8. CONCLUSION
Choosing the fixed time points between which to tabulate failures is mainly a question of
engineering judgement. The time points might be selected, for example, to correspond to mile-
stones in the reliability development program. The parameter estimates and goodness-of-fit
tests obtained in this paper and those obtained by Crow are essentially complementary with
respect to various applications of the Duane model. It is not possible to determine the precise
sample size at which one approach becomes more advantageous than the other. Based on
experience, the chi-square goodness-of-fit test tends to reject most sample data, including data
that fits the model, when sample sizes are too small. Therefore, rejection of the model by the
chi-square test, based on data with a questionable total number of samples, might be viewed as
inconclusive and the more accurate test developed by Crow could then be applied. For large
sample sizes that have at least 10 failures between time points, the chi-square test should be
accurate and is computationally easier. Data that fails to fit the NHPP model with mean value
RELIABILITY GROWTH OF REPAIRABLE SYSTEMS 547
function M{t) based on these tests requires a more general approach. A NHPP model with a
different intensity such as discussed in [1], or a less constrained model such as [13] might then
be tested.
BIBLIOGRAPHY
[1] Ascher, H. and H. Feingold, '"Bad as Old' Analysis of System Failure Data," Proceedings of
the Eighth Reliability and Maintainability Conference (July 1969).
[2] Birnbaum, Z.W., Introduction to Probability and Mathematical Statistics (Harper and Broth-
ers, New York, N. Y., 1962).
[3] Barlow, R.E. and F. Proschan, Statistical Theory of Reliability and Life Testing (Holt,
Rinehart and Winston, Inc., 1975).
[4] Chapman, W.A. and D.E. Beachler, "Reliability Proving for Commercial Products,"
Proceedings 1977 Annual Reliability and Maintainability Symposium.
[5] Codier, E.O., "Reliability Growth in Real Life," Proceedings Annual Symposium on Reliability
(1968).
[6] Cramer, H. Mathematical Methods of Statistics, (Princeton University Press, Princeton, New
Jersey, 1946).
[7] Crow, L.H., "Reliability Analysis for Complex Repairable Systems," Reliability and
Biometry: Statistical Analysis of Lifelength, Society for Industrial and Applied Mathemat-
ics (SIAM), Philadelphia, Pennsylvania (1974).
[8] Donelson, J., "Duane's Reliability Growth Model as a Nonhomogeneous Poisson Process,"
Institute for Defense Analysis paper P-1162 (April 1975).
[9] Duane, J.T., "Learning Curve Approach to Reliability Monitoring," IEEE Transactions on
Aerospace, 2 (1964).
[10] Hovis, J.B., "Effectiveness of Reliability System Testing on Quality and Reliability,"
Proceedings 1977 Annual Reliability and Maintainability Symposium.
[11] Mead, P.H., "Duane Growth Model: Estimation of Final MTBF with Confidence Limits
Using a Hand Calculator," Proceedings 1977 Annual Reliability and Maintainability Sympo-
sium.
[12] Parzen, E., Stochastic Processes (Holden Day, 1962).
[13] Singpurwalla, N.D., "Estimating Reliability Growth (or Deterioration) Using Time Series
Analysis," Naval Research Logistics Quarterly, 25, 1-14 (1978).
ON THE DISTRIBUTION OF THE
OPTIMAL VALUE FOR A CLASS
OF STOCHASTIC GEOMETRIC PROGRAMS*
Paul M. Ellnert and Robert M. Stark
Department of Mathematical Sciences
University of Delaware
Newark, Delaware
ABSTRACT
An approach is presented for obtaining the moments and distribution of the
optimal value for a class of prototype stochastic geometric programs with log-
normally distributed cost coefficients. It is assumed for each set of values
taken on by the cost coefficients that the resulting deterministic primal program
is superconsistent and soluble. It is also required that the corresponding dual
program has a unique optimal point with all positive components. It is indicat-
ed how one can apply the results obtained under the above assumptions to sto-
chastic programs whose corresponding deterministic dual programs need not
satisfy the above-mentioned uniqueness and positivity requirements.
1. INTRODUCTION
This paper is concerned with deriving the distribution and/or moments of the optimal
value for a class of stochastic prototype geometric programs in which a subset of the cost
coefficients are lognormally distributed. The programs are assumed to be superconsistent and
soluble for all positive values that can be taken on by the cost coefficients. It is also required
that the dual of a program has a unique optimal point, 8f , with all positive components, for all
possible values that can be taken on by the components of the cost vector c. Such programs
include soluble programs with no forced constraints. Also included are superconsistent soluble
programs whose forced constraints are nonredundant (and hence active at optimality) and
whose forced constraint gradients are linearly independent at optimality, for each positive-
valued cost vector c.
The class of problems specified above, while of interest in themselves, can be used to
obtain the distribution and/or moments of the optimal value for more general classes of sto-
chastic prototype geometric programs. This will be indicated in Section 6.
The distribution and/or moments of the optimal value of a stochastic program will be
expressed in terms of the density function of a vector L A (log K0, log Ku ... , log Kd)\
where log denotes the natural logarithm function, d is the degree of difficulty of the program,
and
*This research was supported in part by the Office of Naval Research Contract N00014-75-C-0254.
tNow with U.S. Army Materiel Systems Analysis Activity, Aberdeen Proving Ground, Maryland.
550 P ELLNER AND R STARK
log Kj = £ bj{J) log c, for y € {0,1, ... ,.«/}.
In the above (c^ ... , c„)' is the vector of cost coefficients (where "'" denotes transpose) and
the bf1 ) are constants that are independent of the c,-.
One advantage to deriving the distribution and/or moments of the optimal value in terms
of the density function of L is that the vector L is normally distributed when the stochastic c,
are jointly lognormally distributed. Furthermore, under certain conditions, it is reasonable to
expect that L behaves approximately as if the vector of stochastic cost coefficients were lognor-
mally distributed even when it is not. More precisely, if the stochastic cost coefficients,
[cj\i € /}, are positive-valued and the variates {log c,\i € /} are independent with finite means,
variances, and third order absolute central moments, then one can apply a central limit theorem
for random vectors to the relation L = £ Z(/), where Z(,) A (6,(0) log ch 6,u) log c,, ... , b,(d)
log c,)' [11]. Thus, under the above conditions, one might expect that L tends to be normally
distributed provided: the stochastic c, are positive-valued, strictly unimodal, continuous vari-
ates; the number of indices in / is "large" in comparison to d + 1; and no partial sum of d + 1
of the Z(l) is "excessively" dominant in the sum for L.
The results of this paper should be of interest in instances where the operating or con-
struction costs associated with a contemplated project or engineering system can be adequately
approximated as the optimal value of a stochastic prototype geometric program with lognormally
distributed cost coefficients. In such cases a knowledge of the distribution function and/or
moments would be useful as a predictive tool in financial planning. For instance, if the distri-
bution function of the optimal value were known one would be able to predict with a given pro-
bability that a proposed system's operating or construction costs incurred over a given period
would lie within a specified set of limits.
To reflect the uncertainty as to the future costs, c,, that will be encountered during the
construction or operating period of interest a cost analyst often subjectively chooses a distribu-
tion for each cost c,. Cost analysts have frequently found families of positive-valued random
variables that are continuous and strictly unimodal useful for this purpose [9]. The lognormal
random variables form a two parameter family that meets these specifications. Recall a random
variable X is said to be lognormal iff log X is normally distributed. Properties of lognormal ran-
dom variables can be found in [2].
Cost analysts are most often concerned with the distribution of values of c, about a central
value and not with tail values. Thus, an analyst who wishes to utilize the present results might
proceed to express his uncertainty about the future value of cost coefficient c, as follows:
1. Assume c, is lognormally distributed and subjectively choose the median value of ch
denoted by £,.
2. Specify an interval of interest about £, of the form (0,-1f /* 0,£) where 0, € (1, °°).
3. Subjectively choose 8, € (0, 1) such that 1 - 8, reflects one's belief that c, € (0,_1£,,
0,£,); i.e., the more confident one is that c, € {OfHi, 0,-f ,■) the closer 1 - 8, should be chosen
to 1.
OPTIMAL VALUE OF STOCHASTIC GEOMETRIC PROGRAMS 551
4. Calculate the unique value of the standard deviation of c, that is consistent with (1)
and the equation /V(0,_1£, < c, < 0 ,•£,•) = 1-8, where Pr denotes the probability function
associated with c,.
Results of the paper do not require that the stochastic c, be independently distributed.
Thus, for every pair of stochastic cost coefficients ch Cj (/ ^ j) the analyst may subjectively
choose a number between —1 and 1, the correlation coefficient p,7 of log c, and log c,, to reflect
his opinion as to the interdependency of c, and Cj. The theory allows for the possibility that
pij = ±1 (i.e., with probability 1 c, = acf for some constants a € (0, °°) and /3 € (— », oo)).
In Section 2 the notation used in connection with the deterministic and stochastic
geometric programming problem and its dual and transformed dual is presented. Also the spe-
cial role of the transformed dual program in obtaining the distribution and/or moments of the
optimal value of the primal program is indicated.
Section 3 presents and discusses the assumptions placed upon the primal program
throughout Sections 3 through 5 and the appendices. Additionally, useful properties of the
density functions of L and L A (log Ku . . . , log KdV are stated.
In Section 4 we use the density functions of L and Z, together with the maximizing equa-
tions for an unconstrained transformed dual program, to obtain the density functions of r and
(r,v(Pc)). Here r denotes the random vector of the optimal point of the unconstrained sto-
chastic transformed dual program and \(PC) denotes the optimal value of the stochastic primal
program. We then obtain the density function of \(PC) as a marginal density of (r, v(Pc)).
In Section 5 we use the density function of r to derive a formula that expresses each
moment of v(Pc) as the integral of an explicitly given integrand over an explicitly specified con-
vex polyhedral subset of Rd, where d is the degree of difficulty of the stochastic primal pro-
gram.
Section 6 briefly indicates how the preceding results can be used to calculate the distribu-
tion and/or moments of v(Pc) when Pc need not satisfy all the assumptions of Section 3.
Appendix A contains the statement and proof of a lemma from which important proper-
ties of L and L immediately follow. These properties are stated in Theorem 1 of Section 3.
Finally, in Appendix B we establish that boundedness of the dual feasible set is a
sufficient condition for the existence of all the moments of v(Pc), under the assumptions of
Section 3.
2. NOTATION AND PRELIMINARIES
We shall now review the definitions and notation used in connection with prototype
geometric programming that will be utilized in the paper. In the following, for every positive
integer v, <v> A {1, . . . , v) and <v> A {0, 1, . . . , v). Also, for every matrix P, P' denotes
the transpose of P. All elements of Euclidean «-space, R", will be viewed as column vectors of
n real numbers and the zero vector will be denoted by 0.
Recall a prototype primal geometric program has the following form [4]: inf go(t) subject
to gKU) <1VkC <p> and t, > 0 V / € <w> where t = (tu ... , tm)' and gKU) A £ c,
552 P ELLNER AND R. STARK
Y\ tj" for k € </?>. In the above A = (a,7) is an n by m matrix with real entries called the
exponent matrix and c = (c1( .... c„)' is a vector of positive numbers called the vector of cost
coefficients. Also, JK A [mK,mK+x, ... , nK} where m0 = 1, mK = a?k_! + 1 for k € <p>, and
«p = «. The constraints gK(t) < 1 are called forced constraints and we allow the possibility that
a primal program has no forced constraints.
In this paper we shall be concerned with problems of the above form when some or all of
the cost coefficients are stochastic variables that are lognormally distributed. Thus, we shall
assume there exists / C < n > such that / ^ <t> and / € / iff c, is stochastic. Let
Q A (c/ , .... Cj)' where i\ < ... < /„ and I = {iv\ v € <o>>}. Thus, c, is a random vector
formed from the stochastic cost coefficients. Values taken on by c, will be denoted by ch Also
c will denote the value taken on by cost coefficient vector c when q takes on the value c,. We
shall let Pc and P- denote the corresponding stochastic and deterministic prototype primal
geometric programs. Furthermore, v(/>?) will denote the optimal value of P- and v(Pc) will
denote the stochastic variable that takes on the value \(P?) when c, takes on the value C/.
The stochastic program Pc is not convenient to work with due to possible randomness in
coefficients of the forced constraints. To find computationally tractable bounds on the solution
of a two stage geometric program with stochastic cost coefficients, Avriel and Wilde [3] con-
sidered the stochastic problem Dc in place of Pc where, for every c > 0, D- is the dual of P- as
given in [4]. The stochastic program Dc has the attractive feature that all its randomness is
confined to the objective function. To see this recall D- is the following program: sup JJ
8 p A (8) "°
(c//8,-) ' jj \K(8) " subject to the normality condition £ 8, = 1, the orthogonality conditions
£ Ay 8, • = 0 for j € <w>, and the positivity conditions 8,- > 0 for /' € <«>. In the above,
for every k € </?>, \K(8) A £ 8, for 8 6 R". Also, in evaluating the dual objective function
one uses the convention that xx = x~x =1 for x = 0. When P- has no forced constraints we
P ^ (g)
set p = 0 and define the expression JJ \K(8) K to be 1.
K=1
Under rather general conditions one has \(D-) = v(P-) for c € /?> [4, Ch. 6] (where /?>
denotes the positive orthant of R" and v(D-) denotes the optimal value of D-). This is true,
e.g., if P- is superconsistent and soluble [4, Ch. 4]. Thus, frequently the distribution function
of v(Dc) will be the same as that for v(Pc), where v(Dc) denotes the stochastic variable that
takes on the value v(D7) when c, takes on the value c;. Obtaining the distribution function
and/or moments of v(Dc) is facilitated by the fact that the constraint region for Dc is a
polyhedral convex set that depends only on the nohstochastic exponent Matrix A.
Instead of working directly with D- we shall use the transformed dual program, D-, con-
sidered in [4, Ch. 3]. Recall D- is obtained from D- by solving the normality and orthogonality
constraints of D-.
In what follows we shall assume without loss of generality that the rank of A is m and that
q € R" is not in the column space of A, where q, = 1 if i < n0 and q, = 0 if / > n0 (see [4,
Ch. 3]). As in [4] we define d to be the dimension of the solution space of the system of equa-
tions A'b = 0, q'b = 0. (Recall d is called the degree of difficulty of Pc and, under the above
OPTIMAL VALUE OF STOCHASTIC GEOMETRIC PROGRAMS 553
assumptions, equals n - m — 1.) Throughout the paper we assume d > 0. (The distribution
problem for v(Pc) when d = 0 has been studied by R. Stark in [13]J In accordance with the
terminology in [4], we define N A {b(J) | j € <d>} to be a nullity set for Pc iff Nis a basis for
the solution space of the above homogeneous system of equations. Also 6(0) 6 R" is called a
normality vector for Pc iff A'b(0) = 0 and q'bm = 1.
Let A7 A {6(;) | j € <d>) be any nullity set and Z>(0) be any normality vector for Pc.
Note 8 € 7?" satisfies the orthogonality and normality conditions for D- iff* 8 = 6<0) + £ r,6(y)
where the r, € i?1 are uniquely determined by 8. Thus, by replacing 8, in D- by 6,(0) + £
j= i
tyi/ we obtain the equivalent transformed dual problem D-:
sup K(c,bw) YlKfobU*) n8,-(r)~8'(r) flXK(r)Kir)
r j=\ f=l K=l
subject to the positivity constraint Br "> — 6(°} where r A (/-l7 ... , r^)' (the vector of basic
variables). In the above {K{c,b(J))\ j € <d>) is called a set of basic constants for P-
(corresponding to JVand b(0)) where K(c,b(l)) A f\ c, ' for y € <d>. Also, 5 is the nby d
matrix whose yth column is b(j) for j € <d>. Finally, for / € <«> and k 6 <p>,
d
8,(r) A 6,<0) + *£ r/6,0) and XK(r) A "£ 8,(r). When P- has no forced constraints we define
Note that the parameters in D^ depend on the choice of nullity set N and normality vector
6(0). However, as v(ZXr) = v(Zt) for c 6 /?> (where v(DF) denotes the optimal value of ZXJ,
the optimal value of £L is independent of the choice of N and b{0}. Thus, for any nullity set N
and normality vector bm for /><., the distribution function of v(Dc) is identical to the distribu-
tion function of v(Dc), where v(Dc) is the stochastic variable that takes on the value v(D^)
when C/ takes on the value Z7.
To obtain the distribution function and/or moments of v(Dc) we shall first obtain the
density function of the random vector L A (L0,L\, ... , Ld)' and L A (Lx, ... , Ld)' where,
for j € <d> , Lj is the random variable that takes on the value log K{~c,b(J)) when ct takes on
the value ct.
3. On the Density Functions of L and L
Unless otherwise stated, throughout the remainder of the paper we shall assume the fol-
lowing:
(1) [ev\v 6 <«>} is a set of positive-valued random variables such that, for every
/' € <n>, Cj = ati Y\ eviv for some a, € (0, °o) and fiiv € (— °°,°°), v € <u>. Further-
more, it is assumed that (log ex, ... , log eu)' is a nondegenerate normal random vector with
mean vector /j, = (/i,, . .. , fxu)' and dispersion matrix A;
(2) There exists a nullity set [b(j)\j € <rf>} for Pc such that (20)|y € <d>] is linearly
independent where s ' A fi'b(J) for y € <</> and /3 is the n x u matrix whose (/, j) entry is
554 P. ELLNER AND R. STARK
(3) For every value c7 that c, takes on the program P- is superconsistent and soluble;
(4) For every value ct that c{ takes on the program D- has a unique optimal point 8f and
8- > 0.
Many of the results obtained under the above restrictions form the basis to approaches for
calculating the distribution function and/or moments of \(PC) under less restrictive assump-
tions. This will be briefly indicated in Section 6.
Assumption (1) allows for the possibility that a cost coefficient c, is constant (J3iv = 0 for
all v € <u>). Also, (1) permits one to conveniently work with a vector of stochastic cost
coefficients ct = (c,y . . . , qj' for which (log c,|( . . . , log ct)' is a degenerate normal random
vector. Degeneracy would occur, e.g., if C\ and c2 where components of c, such that c2 = acf
for some a € (0, °°) and )8 € Rl.
To evaluate the mean /n, and variance cr 2 of log e, a cost analyst could apply steps (1)
through (4) of Section 1 to et in place of c,. After choosing £,-, the median of eh and the vari-
ance of e, by these steps the values of n, and cr,2 can easily be calculated [2].
Note Assumption (2) is satisfied if u = n and c, = e, for every i € <w>. Also, if there
exists a nullity set of Pc that satisfies (2) then every nullity set of Pc satisfies (2) (Proposition
1).
Recall, for c € /?>, P- is called superconsistent iff there exists t € Rm such that t > 0
and £ c, W t/' < 1 for every k € <p>. Also, P- is called soluble iff P- has an optimal
(€/„ 7=1
point. It can easily be shown that P- is superconsistent for all c € /?> iff there exists a linear
combination of the columns of A, say x, such that x, < 0 for all i 6 7K, k € <p> [1, p. 329].
Alternately, one can show that P- is superconsistent for all c e i?> iff the set
{8 € /?"| 8,- ^ 0 V /' € <aj>, ^'8 = 0, and q'8 = 1} is bounded [1, p. 329]. Moreover, if the
above set is bounded and contains a point 8 > 0 then P^ will be superconsistent and soluble for
every c € /?"> (by [4, p. 120, Th. 2] and [1, p. 329]).
Assumption (3) implies that \(P-) = v(D-) for every value c, taken on by c, ([4, p. 117,
Th. 1]).
Assumption (4) holds for c € i?> if P- is soluble and has no forced constraints. More
generally, one can show (4) holds at c € /?> if P^ is a superconsistent soluble program whose
forced constraints are nonredundant and whose forced constraint gradients are linearly indepen-
dent at optimality. By nonredundant we mean that the optimal value of P? is greater than the
optimal value of PK - for every k € <p>, where PK - denotes the program obtained from P7
by deleting forced constraint k.
If the components of L form a set of independent random variables then we obtain a
d
simpler formula for the density function of L since, in this case, g{lx, ... , ld) = JJ £/(//)
j-\
where g is the density function for L and gj is the density function for £,. If, in addition, L0 is
independent of the components of L then the calculation of moments of v(Pc) is simplified.
This follows from the fact that one can express v(Pc) as the product e °a)(r) where o> is a
known function of a ^-dimensional random vector r whose density function can be calculated
OPTIMAL VALUE OF STOCHASTIC GEOMETRIC PROGRAMS 555
from that of L. Thus, when L0 is independent of L we have F'CvC^)) = [Fie10)] [£"(w(r))]
where £"(0) denotes the i>th moment of random variable Q (whenever this moment exists).
If L0 is a linear function of the components of L one can obtain a function a> of r such that
v(Pc) - <w(r) from which one can calculate £"(v(Pc)). It will be shown, under the previously
listed assumptions, that one can always find a nullity set [biJ)\j €<d>} and normality vector
b{0) for Pc such that L0 is independent of L if s(0) £ span {s^l./ €<</>} and L0 is a linear
function of the components of L if s(0) € span [sU)\j € <d>) where s(0) A /3'£(0) and bi0) is
any normality vector of Pc.
Theorem 1 indicates how to obtain a nullity set for Pc, [biJ)\j € <d>}, such that the
components of the corresponding random vector L are independent normal variates whose
means and variances are known. Also, using the above nullity set, it is shown how to obtain a
normality vector for Pc, bm, such that if s(0) # span {s(j)\j £ <d>) then the components of
the corresponding random vector L are independent normal variates whose means and vari-
ances are known. The proof of Theorem 1 follows immediately from Lemma A which is stated
and derived in Appendix A. The proof of Lemma A uses the eigenvectors of the dispersion
matrix A. Fortunately, however, the calculation of the above-mentioned nullity set and nor-
mality vector and the calculation of the means and variances of the corresponding variates £,,
) € <d>, do not require any eigenvector or eigenvalue calculations.
THEOREM 1: (a) Define {b{J)\j € <d>) inductively by 6(1) A £(1) and, for
I < j ^ d, b{J) AbiJ) -^ «j87>(/), P'b(n > A)~x «P'b(J\ p'bU)>x)b{'\ where
<x,y>\ A x'Ay for (x,y) €'/?" x R". Then [bij)\j € <d>} is a well-defined nullity set of
(b) Define 6(0) A bi0) - ]T «/3'6(/), p'ba)>A)-x «j8'£0), p'bU)>x)bU) if s{0) € span
[s{j)\j €<d>], bm A b(0) otherwise. Then bi0) is a well-defined normality vector of Pc.
(c) For every j € <d> let Lj denote the random variable that takes on the value log
Kj(c,b(j)) when c7 takes on the value c,. Also, define LA (Lu .... Ld)' and
L A (Lq,L\, ... , Ld)'. Then I is a normal random vector with independent components.
Additionally, L is a normal random vector with independent components if S(0) £ span
{s{J)\j € <d>}.
(d) For every j € <d> let gj denote the density function of Lj. Then gj(0=
(T)7V27r)_1 exp (— (2t) j)~l(l — i>j)2) for every I € Rl, where vj is the expected value of L, and
rij is the variance of L}. Furthermore, v} .= <ix,p'bij)> - JT a,6,0) and i\) = </8'60),
- ,= '
P'bij)>A for every y € <d>, where < \ > denotes the usual inner product on R".
Throughout the remainder of the paper we define b^j) and L, for j € <d> , L, and Z as
in Theorem 1. We also denote Kj{c,b(j)) by Kj(c) for j £ <d>.
We shall now show that if there exists a nullity set of Pc that satisfies Assumption (2)
then every nullity set of Pc must satisfy (2).
PROPOSITION 1: If Assumption (2) holds then for any nullity set [bQ)\j € <d>] of
Pc the set [s{J)\j € <d>) is linearly independent where slJ) A fi'b(J) for every j € <rf>.
556 P. ELLNER AND R. STARK
PROOF: Let [t> J>\j € <d>) be a nullity set for Pc such that [s^lj 6 <d>) is linearly
independent where sU) A /3'£0) for every j € <d>. Let 5 A span {60)|y € <</>} and S A
span {S0)|y € <d>). Define T to be the unique linear transformation for B to S such that
r(£0)) - 50) for every j € <rf>. Thus, since B'b(j) - s0) for every j € <rf>, one has
T(b) = B'bforaW b e B.
Since {S0)|y € <d>} is a nullity set for Pc one has span {£0)|y € <</>}« A Thus,
for every y € <tf>, T(bU)) = /3'S0) - 50). Note T is an isomorphism from B onto 5 and
{50)|y € <</>} is linearly independent. Hence, {50)|y € <</>} is linearly independent.
Next we consider the assumption s(0) # span {S(7)|y € <</>}.
PROPOSITION 2: Assume 5(0) $ span {s0)|y € <</>}. Then for any nullity set
{bij)\j 6 <rf>} and normality vector 6(0) for Pc One has 5(0) # span {50')ly € <</>} where
50) A /3'60) for every j £ <d>.
PROOF: Define B A span (£0)|y €<</>} and S A span {30)|y 6 <</>}. Let Tbe
the unique linear transformation from B to S for which T(b(j)) = 5(7) for every y € <d>.
Since j8'60) = s0) for every y € <</> one has f(b) = B'b for all b € £.
Since {£0)|y^€ <</>} is a nullity set of Pc one has span (50)|y €<</>} C B. Also,
since 6(0) and b{0) are normality vectors of Pc it follows that £(0) - b{0) € span
(S^l/ € <«/>}, i.e., 6(0) € B. Thus, for every y € <d>, b{jJ € 5 and hence, 7-(£0)) =
B'b(J) = s^K Finally, observe f is an isomorphism from B onto 5 since s(0) C span
{s(7)|y € <<i>} and {s(;)|y G <</>} is linearly independent by Assumption (2). Moreover,
[b^^lj € <d>) is linearly independent. Thus {S^ly € <d>) is linearly independent.
As mentioned earlier, when u = n and c, = e, for every i € < « > then Assumption (2)
holds. In addition one has s(0) # span {S^ly € <d>) and hence by Theorem 1 the com-
ponents of L are independent. We next consider the case where Assumption (2) holds but
S(0) € span {S0)|y 6 <d>).
PROPOSITION 3: Assume s(0) € span {S0)|y 6 <d>). Then there exist v, € Rx for
j € <d> such that s(0) = Z yJs{J) where s0) A/8'60) for every y € <tf>. Furthermore,
d n [ — d ]
^o - Z -^7 + ^ where ^ is the constant £ ft/0) - Z J>*r lo8 *i-
7=1 /-i I 7-i J
PROOF: Since s(0) € span {s0)|y € <</>}, by Proposition 2 there exist v. € /J1 for
</ d J
j <e <</> such that 5(0) = z yjs • Thus 5/0) ■ Z >y** for every ' € <M>
By Lemma A, Part (Hi), L0 A log AT0(c) = log J] a,' (exp(L(e,5(0)))} = Z bi lo6
a,+ £ 5/0) l08 e-=Z */0) l08 a-+Z Z^/0) l08 e'=Z */0) l08 «/+Z
U Z*/0)l°8e' - Z */(0) lQg «/+Z ^U50))= z */(0) lQg «<+Z yj
OPTIMAL VALUE OF STOCHASTIC GEOMETRIC PROGRAMS 557
\Lj - £ b,0) log a, = x yjLJ '+ D where D k Z *<(0) loe <*• >- Z y\ Z 6«0) log «- = Z
I /-i J y-l ,-i y-i [,-i J ,=i
\btm- ZM0) log«,.
We next consider the rf-dimensional random vector r A (ru .... r^)' mentioned earlier
whose density function can be used to obtain moments of \(PC). To define r assume C/ takes
on the value c,. Then, by Assumption (4), D- has a unique optimal point 8?. Since
{b^^j € <d>) is a nullity set and bi0) is a normality vector of Pc there exists a unique point
r-c A (/^(c), ... , ^(c))' € Rd fox which 8? = ft(0) + £ (/-,(c))ft0>. We define r to be the ran-
dom vector that takes on the value r- when c{ takes on the value c{. In the next section we
shall obtain the density function of r from that of L. Also, when s(0) £ span {s0)|./ € <d>},
we shall obtain the density function of v(Pc) as a marginal density of (v(Pc),r). The density
function of (v(Pc),r) is obtained from that of L.
4. THE DENSITY FUNCTIONS OF rand v(/»c)
Assume C/ takes on the value c,. Since 8- is an optimal point for D- with all positive
components (Assumption (4)) it follows that 8F satisfies the maximizing equations for D- [4, p.
88, Th. 3]. Expressing 8- in terms of the components of r-, the maximizing equations can be
written in the form log Kj(c) = hj(r-) for every j € <d> where, for j € <d>, hj is the
function defined in Theorem 2. The above equations will be used to obtain the density func-
tion of r from that of L. From the above maximizing equations one can easily show that the
optimal value of P? satisfies the equation log K0(c) = log (\(P7)) + h0(r7) where h0 is defined
as in Theorem 2 [4, p. 88, Th. 3]. This equation, together with the maximizing equations, will
be used to obtain the density function of (v(Pc),r) from that of I when s(0> $ span
\s(j)\j € <d>).
To obtain the density functions of rand (\(Pc),r) we shall first define the functions hj for
j € <d> and establish several of their properties.
THEOREM 2: Let H A [r € Rd\Br > -_b(0)} where B is the n x d matrix whose /th
column is b{J) for j € <d> . For every j € <d> define hy. H — - Rl such that, for r € H,
fij(r) A £ bt(J) log 8,(r) - £ \K0) log \K(r)
P d
(where £ x*0) 1o8 M') A 0 if /> = 0). In the above, for every r € Rd, 8,(r) A ft/0' -I- £
r/6,C/) for ; € <n> and XK(r) A £ 8,(r) for k € <p>. Also, for every y € <d> and
k € <p>,XK0) A £ ft,0).
For every j € <rf> define hy. (0,«>) x // — Rd+i such that/for (z,r) € (0,«>) x 7/,
ft,(z,r) A hj(r) if j £ <d> and //oUr) A log z + h0(r).
Finally, define //: H ^ Rd and fe (0,°o) x H — flrf+1 such that, for every
fe,r) € (O.oo) x H, h(r) A (A^r), .... /*»)' and /J(z,r) A (A0(z,r),A1(r,r), .... Arf(z,r))'.
Then
558 r. ELLNER AND R. STARK
(a) h and h aid vontinuously differentiable in //and (O.oo) x //respectively;
(b) A is 1-1 in //and A is 1-1 in (0,«>) x //;
(c) h is onto /?rf;
(d) If S(0) tfspan [sU)\j € <</>} then Ais onto /Jrf+1.
PROOF^ (a) Clearly, for every y € <rf>, r/7 is continuously differentiable in //and, for
every j € <d>, hj is continuously differentiable in (0, oo) x H. Thus, /i is continuously
differentiable in //and h is continuously differentiable in (0,<») x H.
(b) Let r and 5 be elements of //such that Mr) = h(s). Note 8(r) > 0 and 8(s) > 0.
Also, since (log Kx(c), .... log A^(c))' is a nondegenerate rf-dimensional normal random vec-
tor, c, takes on a value, say q, for which log Kj(c) = hj(r) = hj(s) for every j € <d>.
Thus, by definition of A!,- and r/7 for j € <d>, 8(r) and 8(s) satisfy the maximizing equations
for D-. Also, 8(r) and 8(s) are feasible points of D-. Thus, by [4, p. 88, Th. 3], 8(r) and
8(s) are optimal points of D-. However, by Assumption (4), D- has only one optimal point
and hence 8(r) = 8(s). This implies r = s since the nullity set {b{j)\j € <d>) is linearly
independent.
Next, let (z1(r) and (z2,s) be elements of (0,°o) x //such that h(zur) = h(z2,s). Then,
Mr) = Ms), and hence, r = 5. Also, ^oU^r) = h0(z2,s). Hence, by the definition of h~0,
log z, = h0(z{,r) - £ 6,(0) log 8,(r) + £ AK(0) log Mr)
= ft0(z2,s) - X V0) log 8,(5) + £ AJ0' log A«(s)
and thus z, = z2.
(c) Let w € Rd. Since (log A^tc), ... , log Kd(c))' is a nondegenerate rf-dimensional
normal random vector, c, takes on a value, say C/, for which log A}(c) = u} for every
j € <d>. By Assumption (4), D- has an optimal point 8 such that 8 > 0. Let r be the
d
unique element of tf^for which 8 = 6(0) + £ r,*"' and denote 8 by 8(r).
y-i
Since 8(r) > 0 one has r £ H. Furthermore, since 8(r) is an optimal point of D-, by [4,
p. 88, Th. 3] 8(r) satisfies the maximizing equations for D-. Thus, for every j 6 <</>,
hj(r) = log /^(c) = Uj. Hence, h is onto Rd.
(d) Assume v<0) g span {50)|y € <</>}. Let u = («,, ... , ud)' <E Rd and u0 € /J1. By
Theorem 1, (A"0(c), #,(<:), .... ATrf(c))' is a nondegenerate (</ + l)-dimensional normal ran-
dom vector. Thus, c, takes on a value, say Z7, for which log Kj(c) = u} for every j € <d>.
By Assumption (4), /)- has an optimal point 8 such that 8 > 0. Let r be the unique element of
d
RdfoT which 8 = b(0) -I- £ rfb<J) and denote 8 by 8(r). Let r0 A v(D?).
OPTIMAL VALUE OF STOCHASTIC GEOMETRIC PROGRAMS 559
Since 8(r) > 0 and v(D-) > 0 one has (r0,r) € (0,<») x H. Also, since 8(r) is an
optimal point of D^, by [4, p. 88, Th. 3] 8(r) satisfies the maximizing equations for D-. Thus,
for every j € <d>,
(1) hj(r0,r) = hj(r) = log Kj(c) = Uj.
Also, by [4, p. 88, Th. 3], since 8 0) satisfies the maximizing equations for D? one has
r0 = v(Z)?) = K0(c) n 8,(r)"*'(0) f\ Kir)^ '.
Thus, log r0 = log tf0(c) - £ 6,(0) log 8,(r) + £ XK(0) log (XK(r)), i.e.
(2) h0(r0,r) = log K0(c) = «0.
By (1) and (2) his onto Rd+X.
Note by Theorem 2, for every / € Rd there exists a unique point r, € H such that
/= h(r,). Thus, we can define h~l: Rd -^ H by h~x[l) A r, for / € /?<*. Also, if s(0) tf span
{s^l j € <</>} then by Theorem 2, for every / 6 Rd+X there exists a unique point
(z7,r7) € (O.oo) x //such that 7= /t(z7,r7). Thus, when s(0) <7 span {S0)|j € <</>}, we can
define h~l: Rd+X — (0,°°) x //by A_1(7) A (z7,r7) for 7 € /^+1.
PROPOSITION 4: (a) r = h~xU);
(b) (v(Pc),r) = A_1(L) if S(0) g span {s0)|y € <</>}.
PROOF: (a) Let c7 take on the value c7. Then L takes on the value / A (log Kx(c),
log #</(?))'. By Assumption (4) and [4, p. 88, Th. 3] one has r- € //and log Kj(c) = fcy(r?)
for every j € <d>. Thus, h~x(l) = r-. Hence, h~x(L) takes on the value r? when C/ takes
on the value c7, i.e., h~x(L) = r.
(b) Assume s(0) g span {s0)|y_€ <</>}. Then h'x:_Rd+x — (0,«>) x //is well-defined.
Let C/ take on the value c7. Then L takes on the value /A (log K0(c), log AT^c), ... , log
Kd(c))'. Note v(Pc) takes on the value v(/>f) > 0. Also, by Assumption (4) and [4, p. 88,
Th. 3] one has r- € //and log Kj(c) = hj (v(P-), r-) for every / <E <</>. Thus, A-1 (7) =
(v(P?), r?). Hence, h~x(L) takes on the value (v(/>?), r?) when c7 takes on the value c7, i.e.,
A-HZ) = (v(Pc),r).
We can now obtain the density function of r.
THEOREM 3: Let t/» denote the density function of r and g denote the density function
of L. Then
JO if r g H,
x,,ir) = \{g(h(r)))(\detDh(r)\) if r € H,
where Dh(r) denotes the derivative of h at r.
560 P ELLNER AND R. STARK
PROOF: Let ct take on the value c,. Then r takes on the value r- A (r,, . . . , rd)' where
d —
h-= b{0) + £ r^'7' is the unique optimal point of D-. By Assumption (4), 8? > 0. Thus,
M
Br- > —b(0\ i.e., r- € H. Hence, rcan only take on values in H. Thus, <//(r) = 0 if r & H.
Let B be an open Borel subset of H. Note, by Proposition 4,
Pr(r 6 £) = Pr(h~xU) £ B) = Pr(L € /?(£)).
By Theorem 2, A is 1-1 and continuously differentiable in B. Also, g is integrable on /;(/?)
since g is the density function of L. Thus,
ML € h(B))'JHB)g = fB (g °h)\det Dh |
[12, Ths. 3-13 and 3-14], where g °h denotes the composition of g and h. Hence, Pr(r € B) =
Jg (# °/?)|det D/?|. This implies ^/(r) = {g(h(r))}(\dQt Dh(r)\) for r € //.
Next we obtain the density function of (v(Pc),r).
THEOREM 4: Let </) denote the density function of (v(Pc),r) and g denote the density
function of L. Assume s<0) £ span [s^'\j € <d>). Then
| 0 if (z,r) <? (0,°o) x H,
| {g(Mz,r))}(|det LVKz,r)|) if (z.iO € (0,°°) x //,
where Dh(z,r) denotes the derivative of h at (z,r).
Hz,r) = \
PROOF: By the proof of Theorem 3 it follows that (v(Pc),r) can only take on values in
(0,°o) x H. Hence, i/)(z,r) = 0 if (z,r) Q (0,~) x H.
Let B be an open Borel subset of (0,°o) x //and define z A v(Pc). Note by Proposition
4, Pr((z,r) 6 £) = Pr(h~x{L) € B) = Pr(L € £(£))._ By Theorem 2, /? is 1-1 and continu-
ously differentiable in 5. Also, £ is integrable on h(B) since g is the density function of L.
Thus, ML 6 h(B)) = f- - £ = f- (g °A)|det M| [12, Ths. 3-13 and 3-14], where £ °h
J h(B) J B„ p _
denotes the composition of g and h. Hence, Pr((z,r) € B) = J . (g °h) \del Dh\. This
implies 4>(z,r) = {g (h(z,r))} (\det Dh(z,r)\) for (z,r) € (0,«») x H.
When \(0) # span {.?(/)|y € <d>) the above theorem immediately yields the density
function of v(/*f).
COROLLARY 4.1: Let /denote the density function of v(Pc) and assume s(0) ? span
{s(n\j € <d>}. Then
/(*)-
0 if z ? (0,oo),
/rew{£(A(z,r))}(|detZM(z,r)|)dr if z € (°-00)-
Observe that to evaluate /at z € (0,«>) by Corollary 4.1 one must integrate a specified
function over the convex polyhedral set H = [r 6 flrf|5r > -b{0)}. When the degree of
difficulty d equals 1 then //will be an interval in Rl whose end points can easily be obtained.
OPTIMAL VALUE OF STOCHASTIC GEOMETRIC PROGRAMS 561
Thus, when d = 1, one can accurately approximate f(z) by applying a quadrature formula to
evaluate the integral expression for f(z). However, the quadrature rule must be modified as in
[7, Ch. 7, Sec. 6.2] to allow for the fact that the integrand is not defined at the end points of
the interval of integration.
When d > 1 the effort and expense of devising and applying a quadrature scheme to
approximate the integral expression for f(z) to a high degree of accuracy may not be justified
since frequently the distributions chosen for the stochastic c, will be subjectively determined.
In such cases a numerical Monte Carlo method could be an attractive alternative for approxi-
mating the multiple integral used to express f{z) [6, 14, 15].
Finally, under the assumption of Corollary 4.1 note the distribution function of v(Pc),
denoted by F, is given by
F(y)-
0 if^O,
j;^€(0l>)^{^^))f(Jdeti>AUr)|)^Z .fy > Q
Thus, if great precision is not required a numerical Monte Carlo technique could be attractive
for approximating F(y) as well as f(z).
5. THE MOMENTS OF v(Pc)
In the following, for each random variable Q, recall E(v){Q) denotes the moment of order
v o{ Q whenever it exists, where v 6 N (the set of positive integers). Also, let
E{Q) A E{l)(Q).
Throughout Section 5 we assume EM(\(PC)) exists for every v € N. Proposition B in
Appendix B establishes that boundedness of the dual feasible set FA (8 6 ^"|^'8 = 0,
q'8 = 1, 8, > 0 v i € <n>) is a sufficient condition for the above moments to exist. Furth-
ermore, one can show P- is superconsistent for every c € /?> iff F is bounded (see p. 554).
To calculate the moments of v(Pc) it is advantageous to use the density function of r
instead of that for v(Pf). To obtain the moments of v{Pc) in terms of the density function of r
8 ,-(/■) ' if p = Oando)(r) = f[8,(r) ' f[ XK(r) K if p > 0, for r € H.
PROPOSITION 5: \(PC) = eL°co(r).
PROOF: Let c, take on the value c,. Then e °o)(r) takes on the value K0(c)(o(r-) where
d
r-= (/-,, .... rd)' is the point in //for which 8-, = bi0) + £ rJb(J) is the unique optimal point
of D-. Since 8- > 0 one has K0(c)a>(r-) = v(P?) [4, p. 88, Th. 3]. Thus, v(Pc) = eL°a>(r).
We shall now obtain the moments of v(Pc) when v<0) € span [s j)\j 6 <d>).
THEOREM 5: Assume s<0) € span {s(J)\j € <d>). Then, for every v e TV,
F
W(v(J»)) = § ^ {£ (r) }>(/•) dr where, for r € //,
P. ELLNER AND R STARK
h(r) A
eDl\8l(r)u'll\Jry^ ifp>0.
In the above D A £ U/0) - £.vA(/) log a„ w, A £ v^/7 ' -6,(0) for / € <«>, and
vK A £ ^/AJ'1 -X.K(0) for k € </>> where (y, v^)' is the unique element of Rd for which
5'0)= £ _y/5(/>.
7=1
PROOF: We shall assume p > 0. (The modification needed in the proof for p = 0
should be clear.) By Assumption (2) and Proposition 1 the set [sU)\j 6 <d>) is linearly
independent. Hence, by Proposition 3 there exists a unique element (yu ... , yd)' of Rd for
d d
which 5(0) = £ yjS(J). Also, by Proposition 3, log K0(c) = £ >>,- log K^c) + D where
I/=i /-I
d \
b,(0) - L yM'\ logo,-. By Proposition 5 v(Pc) = K0(c)co(r). Thus,
(1) log (v(Pc)) = log KQ(c) + log (<»(/■))
- £ jey.lpg «/(<:>.-+ Z> ;*- £ XK(0) log M') - £ J»((0) log8,(/-).
/-I K=l ,= l
Let c, take on the value c,. Then /■ takes on the value r- = (r^c), ... , rd(c))'. Since
d
8- A 6(0) + £ (rji(c))6(/) is an optimal point of D- and 8- > 0 (by Assumption (4)) one has
(by [4, p. 88, Th. 3])
log Kj(c) = h,(r-) = £ 6,(/) log 8,(r-) - £ *#? log \K(r-)
for every y € <d>. Thus, by (1),
log (v(/>?)) = £ v, £ fcO> log 8,-(rr) - £ \K(/) log \K(rF)
M (,= i .= i
+ D + £ \j0) log \K(r?) - £ 6,(0) log 8,(/-r)
- ° + £ z m(/) - bil0)\ !°g 8<(rP
- £ [£MK(/)-^(0,|log\K(rr).
Hence,
v(Pc) = eD f[ 8,(r)"'n *»""" = w(r).
It follows that £M(v(/>,)) = £({v(/>(.)}") = Je// |«(r))'*(r)* [8, p. 18, Th. 1.4.3].
OPTIMAL VALUE OF STOCHASTIC GEOMETRIC PROGRAMS 563
We next obtain the moments of v(Pc) in terms of the density function of r when s(0) $
span [s(i)\j € <d>).
THEOREM 6: Assume 5(0) tf span [s(i)\j € <d>) and let v € N. Then,
(a) e ° is lognormal and independent of R A w(r);
(b) £(,,)(v(Pc)) = ^(e^^CR);
(c) EM (eL°) = expl|t M(VW " t *i*A + T" fr®' As<0))]
I l'-i '=i I 2 J
where s<0) = /37>(0);
(d) £(">(i?) = J {a>(r)}^(r)dr.
PROOF: (a) By Theorem 1 L0 is normal and hence e ° is lognormal.
Note by Proposition 4 w(r) = a>0~'(L)) where I = (Ilf .... Ld)'. By Theorem 1 L0
and L are independent. Thus, e "and i? are independent [8, p. 15, (III)].
(b) To show that E(R") exists let X A /L° and K A R". Clearly (A', K) is a continuous
random vector. Thus, ret wbe the joint density function of (X,Y). Also, let wj and w2 denote
the marginal densities of X and Y respectively. Then w(x,y) = wx(x)w2(y) for all
(x,y) € /?' x R[ since Zand Kare independent by (a).
By Proposition 5 {v(Pc)}'' = XY. Thus, by assumption, E(XY) exists and hence,
xyw (x,y) dxdy is convergent. Thus, by Fubini's Theorem [10, p. 207, Th. 2.8.7]
U.yKfl'x/?1 ^ oo oo
xvw(x,v)d!x4y = J H(y)dy where //(v) A J xyw(x,y)dx = yw2(y) J
Uy)ZRl*Rl ° rco co
x»vi(x)rfx = yw2(y)E(X). This implies E (XY) = E (X) I ^(v)^' and hence I yw2(y)dy
is convergent, i.e., E(R ) exists.
Since the expected values of e ° and /?" exist, the independence of e* ° and R" implies
E(e"L°)E(Rv) = E(e"L*Rv) [5, p. 82, Th. 3.6.2]. Thus, by Proposition 5,
EM(y{Pc)) = E({eL°RY) = E(e"L(i)E{Rv) = EM (/°)£(" >(/?).
(c) Recall by Theorem 1 E(L0) = £ /x,s,(0) - £ a,ft/0) and K(I0) = s(0)As(0) where
= /3'6(0) and K(L0) denotes the variance of L0- By [2] one has ^^'(e^0) = exp [vE(L0) +
2
(d) Using the density function </» of r we obtain £(|,)(/?) = E(R") = J {w(r)}Xr)<fr
[8, p. 18, Th. 1.4.3]. r€W
Note that to evaluate Eiv)(v(Pc)) by Theorem 5 or 6 one must integrate a specified func-
tion over the convex polyhedral set H = [r € /?^|5r > - bm). Hence, the comments made
in Section 4 concerning the evaluation of f(z) also apply to the evaluation of E{v)(v(Pc)). In
564 P. ELLNER AND R.STARK
particular, note that for a given precision the amount of work required to calculate EM(v(Pc))
by Theorem 5 or 6 should be about the same as the amount required to calculate f(z) by
Corollary 4.1. Thus, in calculating £(t,)(v(Pf)), it is advantageous to express E(v)(v\pc)) in
terms of the density function of r as in Theorems 5 and 6 rather than to express E(v)(\(Pc)) as
Jo" «"/«*•
6. EXTENSIONS
In this section we shall indicate how the preceding results can be used to obtain the distri-
bution and/or moments of \{PC) when Pc need not satisfy all the assumptions of Section 3.
However, no formal statements or proofs will be presented.
I
which are
n the following, we shall refer to strengthened versions of Assumptions (2), (3), and (4)
are stated below for a stochastic geometric program Pc that satisfies Assumption (1).
We say Pc satisfies Assumption (2') iff Pc satisfies Assumption (2) and s(0) G span
{slJ)\j € <d>\ where s(J) A p'blJ) for every j 6 <d>. Here [b(j)\j € <d>) is any nullity
set for Pc and b{0) is any normality vector for Pc. Also, the matrix /3 is defined as in Assump-
tion (1).
Pc is said to satisfy Assumption (3') iff P- is superconsistent and soluble for every
c € R'±.
Finally, Pc is said to satisfy Assumption (4') iff D- has a unique optimal point 8- and
8- > 0 for every c € R>.
Now consider a family of random cost vectors { c (e ) | e € (0,°o)}, where cU) A
(c^e), ..., c„(e))'fore € (0,°°), that satisfies the following:
(i) (log C\(e), . . . , log c„(e))' is a nondegenerate normal random vector;
(ii) £(log c,(€)) = £(log Cj) for every / € <n>\
(iii) lim Cov(log c.(e), log c.(e)) = Cov(log c., log c.) for every (i,j) € <n> x <«>,
€10
where Cov denotes covariance.
Such a family of cost vectors can easily be constructed if Pc satisfies Assumption (1).
When Pc also satisfies Assumptions (3') and (4') one can show that Pc{e) will satisfy Assump-
tions (1), (2'), (3'), and (4'), where Pc^ is obtained from Pc by replacing c with the cost vec-
tor c(e). Thus, the results of the preceding sections can be used to calculate the moments, dis-
tribution function, and density function of Pcie) for e € (0,°o). Additionally, one can establish
that the moments, distribution function, and density function of Pc(() converge to the
corresponding moments, distribution function, and density function of Pc as € tends to zero.
Next, consider the family of stochastic geometric programs [Pc(y)\y € (0,<»)} where, for
y € (0,oo), Pc{y) denotes the following stochastic program:
inf I c, ft t? n ^
OPTIMAL VALUE OF STOCHASTIC GEOMETRIC PROGRAMS 563
We next obtain the moments of v(Pc) in terms of the density function of r when sm $
span [s(J)\j e <d>).
THEOREM 6: Assume s(0) tf span [s(J)\j £ <d>) and let v € N. Then,
(a) e ° is lognormal and independent of R A w(r);
(b) £w(v(/'c)) = ^(/o^o,) (/?);
(c) E^(eL°) = expl\±nisi{0)- £ a,6/d)| + ^ ^m' Asm)\
I l=i '=i J 2 I
where s(0) = /37>(0);
(d)£(^(/?) = Xg// (<o(r)}>(r)rfr.
PROOF: (a) By Theorem 1 L0 is normal and hence e ° is lognormal.
Note by Proposition 4 a>(r) = a(h~l(L)) where L = (Lt, ... , Ld)'. By Theorem 1 L0
and L are independent. Thus, e ° and /? are independent [8, p. 15, (III)].
(b) To show that £(/?") exists let IA e"L° and Y A /r. Clearly (Z,K) is a continuous
random vector. Thus, let wbe the joint density function of (X, Y). Also, let wx and w2 denote
the marginal densities of X and Y respectively. Then w(x,y) = w\(x)w2(y) for all
(x,y) € R[ x i?1 since Zand Fare independent by (a).
By Proposition 5 {v(Pc)Y = XY. Thus, by assumption, E(XY) exists and hence,
xy»v(x,v)d!x^v is convergent. Thus, by Fubini's Theorem [10, p. 207, Th. 2.8.7]
UV)€J(IXXI ^ CO oo
J J xywOc,y)<£«/v = JQ H(y)dy where //(v) A Jq xvw(x,v)d!x = vh>2(v) JQ
(x.y)€RlxRl roo oo
xw!(x)rfx = yw2(y)E(X). This implies £(AT) = E(X) J vw2(v)d[y and hence J ytv2(y)</v
is convergent, i.e., E(R") exists.
Since the expected values of e ° and R" exist, the independence of e ° and /?" implies
E(evL°) E(RV) = E{eL»R") [5, p. 82, Th. 3.6.2]. Thus, by Proposition 5,
E(v)(y{Pc)) = E({eL°R}") = E(evL°)E(R") = £(v) (/°)£MU).
(c) Recall by Theorem 1 E(L0) = £ /a,S/(0) "'E a,V0) and K(I0) - 5(0)As(0> where
5(0) = ^^(0) and K(Lo) denotes the variance of L0. By [2] one has £M(/°) = exp [i/£(L0) +
— v2V(L0)] since L0 is normal.
(d) Using the density function <// of r we obtain EM{R) = £(i?") = J {a>(r)}>(r)</r
[8, p. 18, Th. 1.4.3]. '*"
Note that to evaluate E{v)(\(Pc)) by Theorem 5 or 6 one must integrate a specified func-
tion over the convex polyhedral set //={/•€ Rd\Br > - b(0)). Hence, the comments made
in Section 4 concerning the evaluation of /(z) also apply to the evaluation of Elv)(y(Pc)). In
564 P. ELLNER AND R.STARK
particular, note that for a given precision the amount of work required to calculate EM(v(Pc))
by Theorem 5 or 6 should be about the same as the amount required to calculate f(z) by
Corollary 4.1. Thus, in calculating ^"'(vC/^)), it is advantageous to express EM(v(Pc)) in
terms of the density function of r as in Theorems 5 and 6 rather than to express EM{y(Pc)) as
/." **/(*)<&.
6. EXTENSIONS
In this section we shall indicate how the preceding results can be used to obtain the distri-
bution and/or moments of \{PC) when Pc need not satisfy all the assumptions of Section 3.
However, no formal statements or proofs will be presented.
In the following, we shall refer to strengthened versions of Assumptions (2), (3), and (4)
which are stated below for a stochastic geometric program Pc that satisfies Assumption (1).
We say Pc satisfies Assumption (2') iff Pc satisfies Assumption (2) and s(0) & span
{sin\j € <d>\ where s</>A/3'6</, for every j € <d>. Here [b(n\j 6 <d>) is any nullity
set for Pc and b(Q) is any normality vector for Pc. Also, the matrix /3 is defined as in Assump-
tion (1).
Pc is said to satisfy Assumption (3') iff P- is superconsistent and soluble for every
c € R'i.
Finally, Pc is said to satisfy Assumption (4') iff D- has a unique optimal point 8F and
8- > 0 for every c € R">.
Now consider a family of random cost vectors (c(€)|e € (0,°°)}, where c(t) A
(c\(e) c„(€))'fore € (0,°°), that satisfies the following:
(i) (log C](e) log c„(e))' is a nondegenerate normal random vector;
(ii) £(log c,(e)) = £"(log c,) for every i € <n>,
(iii) lim CovOog c,(e), log c,(e)) = Cov(log c., log c.) for every (ij) € <n> x <«>,
610
where Cov denotes covariance.
Such a family of cost vectors can easily be constructed if Pc satisfies Assumption (1).
When Pc also satisfies Assumptions (30 and (4') one can show that Pc{e) will satisfy Assump-
tions (1), (2'), (30, and (40, where Pcif) is obtained from Pc by replacing c with the cost vec-
tor c(e). Thus, the results of the preceding sections can be used to calculate the moments, dis-
tribution function, and density function of Pc(t) for €'€ (0,°o). Additionally, one can establish
that the moments, distribution function, and density function of Pc(() converge to the
corresponding moments, distribution function, and density function of Pc as € tends to zero.
Next, consider the family of stochastic geometric programs {/V?)ly € (0,°°)} where, for
y € (0,°o), P}y) denotes the following stochastic program:
y I c, ft t? n *r
'•~ i€7„ /=1 k=1
OPTIMAL VALUE OF STOCHASTIC GEOMETRIC PROGRAMS 565
subject to £ c, Yl tj1' + zk ^ 1 for every k € <p>, t > 0, and z > 0 where
f = (flf . . . , tm)' and z = (z), .... z^,)'. One can show Pc{y) satisfies each assumption that Pc
satisfies. In addition, if Pc satisfies (3') then Pc{y) satisfies (3') and (4') (even when Pc does not
satisfy (4)). Thus, one can apply the results of the preceding sections to calculate the density
function, distribution function, and moments of Pc(y) for y € (0,°°) when Pc satisfies (1), (2'),
and (3'). Furthermore, one can establish that the moments, distribution function, and density
function of Pc{y) converge to the corresponding moments, distribution function, and density
function of Pc as y approaches zero.
Finally, for y € (0,°°) and e (0, °°), let P}?f\ denote the stochastic program obtained
from Pc{y) by replacing cost vector c by c(e) in Pciy). The family of cost vectors
{c(c)| e € (0, °o)} is assumed to satisfy the properties (i), (ii), and (iii) previously listed. One
can show, for (y,e) € (0,°o) x (0,«>), program Pc(e)iy) satisfies (1), (2'), (30, and (4') if Pc
satisfies (1) and (3'). Thus, in this instance, one can apply the results of the preceding sections
to Pc%\. This suggests that the family of programs [Pc{$ \y € (0,°o) and e € (0,°°)} may be
useful in obtaining the moments, distribution function, and density function of Pc when Pc
need only satisfy Assumptions (1) and (3').
APPENDIX A.
Theorem 1 in Section 3 is an immediate consequence of the following lemma.
LEMMA A: Define L(z,s) A £ 5, log z, for every positive-valued random vector
z A (z\, ... , zu)' and 5 € R". Also, define the inner product <-,->A on R" by
<x,v>A A x'Ay for (x,y) € R" x R". (Note <•, >A is an inner product since A is a disper-
sion matrix of a nondegenerate normal random vector and hence is positive definite.) Then
(i) (L(e,stu), ... , L(e,s(d)))' is a normal random vector with independent components
where e A (eu .... eu)\ s(1) = s(1), and
,(/) = JO) _ £ «5(/),5(/)>A)-1 «S(i),SU)>x)s{l)
for 1 < j < d,
(ii) (L (e,si0)),L(e,siu), . . . ,L(e,s{d)))' is a normal random vector with independent
d
components if si0) g span [siJ)\j € <d>) where s(0) = s(0) - £ (<s(/),s(/)>Ar'
(<s(0\s(/)>A)s(/) when s(0) <? span {sin\j € <</>}.
(iii) For every j € <d>, s0) = p'bU) and f\ c-' = ]]«,'''' {exp (L(e,sU)))} where
lw )
bw = bw and, for 1< j < rf, Z>0) -*</>_£ «fi'b{l),fi'bU)> x)~{ «p'bU),p'b(l)> A)6(/).
Also 6(0) = 6(0) if 5<0)€ span {S0)L/ € <rf>} and 6(0) = £(0) - £ «j8'6(,)^'6(/)>A)-1
(<P'b{0),p'bU)>A)bU) if 5(0) <? span {s0)|y € <d>). Furthermore, [bU)\j € <d>) is a nul-
lity set and b(0) is a normality vector of Pc\
566 P. ELLNER AND R STARK
(iv) For j e <d>, the density function <£, of L(e?,s0)) is given by <£,(/) A j= exp
- ojyv27r
r — for every / € Rx where p, is the expected value of L(e,s{J)) and cof is the vari-
l 2<u/ j
ance of L(e,s0)). Furthermore, pj ■. = *£ nisi(J) and ajj = <s0),s0)>A.
/=i
PROOF: Since A is real symmetric, A has an orthonormal set of u eigenvectors
[p\> ■ • • <Pu\- Let P be the u x u matrix whose yth column is p}. Then P is orthogonal (i.e.,
P~x = P') and A A P-1 A/Ms diagonal.
For every i e <u> let y, A log e, and y A (yi, ... , -yu)'. Let -y A P'y. Then y is a m-
variate normal vector with dispersion matrix P'AP = A ([8, Th. 2.1.1]). Since A is diagonal,
the components of y are independent.
Let (s,w) (E Ru x R" such that w = P's. We shall show L(e,s) = L(e,w) where e A
(?!, .... eu)' and e, A ey' for / € <«>. Note y = />y. Thus, for / € <w>, y, = ]£ AkTV
Hence,
(1) L(e,s)= £s, \£PiKyK - t 2>k4?«
= H wKyK = £(e,w).
For every y € <d> define w^' A P's(J). By assumption {s" L/ € <</>} is linearly
independent. Thus {w^}\j € <d>] is linearly independent since P' is nonsingular. Thus, one
can apply the Gram-Schmidt orthogonalization process to {w^ly 6 <</>} to obtain the
orthogonal set [w^^lj € <d>} with respect to <v>a where w(1) A h>(1) and, for 1 < j < 4
(2) w°') A w0)- £ (<w(/),w(/)>A)-1(<w'0')(iv(/)>A)w(/).
(Note <•, •>£ is the inner product on Ru defined by <x,y>A A x'Ay for (x,y) € R" x Ru.
<•, ->jj is an inner product on /?" since A is the dispersion matrix of the nondegenerate nor-
mal random vector y and hence is positive definite.) Also define
(3) w(0) A wi0)- j^«WU),wU)>^«w(0\WU)>^win
/=i
if si0) g span (S°H/ € <</>}; otherwise define w<0) A u>(0). Observe if .?(0) <? span
{sQ)\j 6 <</>} then w(0) £ span {w^l/ € <d>). Thus,lw0)U' 6 <</>} is an orthogonal
set in i?" with respect to <•, >K when s(0) <? span {s{n\j € <d>}.
Define s0) A Pw0) for every j € <d>. Then, for j € <^>, w(J) = P's0). Thus, by
(1),
(4) Lie.s^^ = L(e,w0)) for every J e <d>.
We shall next show (L(e,w(x)), ... , L(e,w(d)))' is a normal random vector with independent
components. Also, whenever s(0) $ span [S^J)\j € <d>), we shall show that
{L(e,wm),L{e,w(x)), ..., L(e,wid)))' is a normal random vector with independent com-
ponents.
OPTIMAL VALUE OF STOCHASTIC GFOMFTRIC PROGRAMS 567
For every / € <u> let 9} and t, be the variance and mean, respectively, of log eh For
i € <u> define i//, A 0,-1 (log q - t;). Note, for every 7 € <^>, L^w0*) = £ w,0) log
*/ - E 0,w,O) <//,+ £ tiwi<j)- Let r,t £ <d> such that r ^ r. Recall -y A (log
fi, ... , log eu)' is a normal vector with independent components. Thus, (<//,|/ € <w>} is a
set of independent unit normal random variables. Thus, L{e,w(r)) and L{e,w{,)) are normal
random variables. Moreover, L(e,w(r)) and L(e,wU)) are independent provided £
^VV" = 0 [8, Th. 4.1.1, p. 70].
Since, for every i € <«>, 9} is the variance of y, and A is the dispersion matrix of y
one has JT Ofw^w^ = <w(r),w(f)>A. By construction of {h>0)|./'€ <J>) one has
<M,('-)(M;(')>A = 0 for r,r 6 <d> with r ^ /. Also, <w{r) ,wU)> A = 0 for r,r € <</> with
r * t provided s(0) <2 span {s{,)\j 6 <</>}. Hence, by (4), (L(e,s(1)), ... , L{e,s(d)))' is a
normal random vector with independent components. Also by (4), (L(e,s{0)),
L(e,sw), ... , L(e,s(d)))' is a normal random vector with independent components if s(0) £
span [s{n\j € <d>).
Next, let 0t(' V) € Ru x /?" for / € {1,2} such that y(,) = P'xU). Then,
(5) < v(l),y(2)>A = <P'xa>,P'xi2)>i = 0c(1))7>A/>';c(2) = <x(1),x(2)> A.
Observe s(1) = Pwn) = Ph>(1) = s(1). Also, by (2) and (5), for 1 < j < done has
(6) s0) = /V'> = Pw(J) - £ (<*v(/\h>,,)>a)-1(<>v0),vv(/)>a)/V/)
= S(J) - £ «S(/),5(/)>A)-1«50,,5(/)>A)S(/).
Moreover, by (3) and (5) one has
(7) 5(0) = PW(0) = 5(0) - £ «5('),S(/)>A)-1 «5(0,,5(/)>A)S(/>
/=1
if 5(0) £ span [s{l)\j € <rf>}. This completes the demonstration of (i) and (ii).
Next, we shall obtain a nullity set {b{,)\j 6 <d>) and normality vector b{0) for Pc such
that s(J) = fi'b(J) and f[ c* ' = j]o*' {exp(L (e,s{J )»} for every J € <</>. Let 5A
span {S0)|j € <rf>) and B A span {£0'[/ € <rf>}. By assumption {s0)L/ € <d>) is
linearly independent and hence a basis for S. Thus, there exists a unique linear transformation
T. S — # such that r(5C/)) = £0) for every y € <rf>. Since {£(/)|./ € <</>} is a basis for
5, r is an isomorphism from 5 onto B. Also, since T~l(b^ ) = s^ = /3'biJ) for every
./€<</> one has 7^(6) = ]3'6 for every 6 6 5.
Recall s(1) = s(,) € 5. Let 1 < j ^ rf and assume sin 6 S for 1 < / < j. Then by (6)
one has s(J) € S. Hence, [s(J)\j €<d>) C S. Also {s(/ } |y € <</>} is linearly independent
since {w(/)|y 6 <d>) is orthogonal with respect to < ,>A and P is a nonsingular matrix for
568 P. ELLNER AND R.STARK
which 5(/) = Pw{j) for every j 6 <d>. Thus, [s^^lj € <d>} is a basis for 5. Since Tis an
isomorphism from S onto B, [T(s(,))\j € <d>) is a basis for 2?. Thus, [b(i)\j € <</>} is a
nullity set for P, where bU) A Hs0') for y € <rf>.
Let 5 A span [sU)\j f <rf>)andSA s_panj6(/,Jy € <d>). Suppose sj0) ? 5. Then
there exists a unique linear transformation t: S — ■ B such that f(s(/)) = b{,) for every
y €_<rf>. Also, by (7), s(0) € S. Thus, we can define b{0) A f (s(0)). Note by the definition
of Tone has T(s) = T(s) for every s € 5. Thus, by (7),
(8) 6(0) = f(5(0)) - £ «5(/))5(/)>A)-1«5(0),S(/)>A)f(5(/))
= 6<0) - £ «S(/\5(/)>/V)-1«5(0)(5(/)>,)6</)
since f(5(/)) = T(sV)) = 6(/) for / € <rf>. For y € <m> let ^ denote column j of
exponent matrix A. Recall q € R" such that <7, = 1 if / € <no> and 9, = 0 if / > aj0 where w0
is the number of elements in J0. Then by (8), for every j 6 <m>, one has <b(0\ Aj> = 0
since b{0) is a normality vector and {b(l)\l € <d>) is a nullity set of Pc, where <•,•> denotes
the usual inner product on R". Also, <bi0),q> = <£(0\<7> = 1 since <bU),q> = 0 for
every / € <d>. Thus, bi0) is a normality vector for Pe. If v(0) € S we define bm A £(0).
Thus, whether s<0) £ Sor 5(0) € Sone has that bm is a normality vector for Pc.
To show (3'bV'= 5(/> for every j € <</> first observe for j € <d> one has
p'b{l) = T-Hb^) = T[(T(s{l)) =_s0). Next suppose s(0) (? 5. Then f is an isomorphism
from S onto 5 since [b^ \j € <d>} is linearly independent and f(5(/>) = b(J) for every
y € <d>. Also, f~1(£0')) = s0) = p'8U) for every y € <</>. Hence, f"'(6) = p'b for all
6 6 B. Note 6(0) A f(5(0)) € 5. Thus, /37>(0) = f~Hb{0)) = f^f (s(0))) = s(0). Finally, sup-
pose s(0) 6 S. Then 5'M0)_= /3'£(0) = v(0) = Pwi0) = Pw(0) = s(0). Thus, from the above,
pb(j ) = s(j ) for every j € <d>.
Next let j € <d> and observe J J c, ' = Y[ \ai Yl e" \ = IT01'
0/„*i
5>„A''
IT n<V [= 11"/' 11*7' • Thus' since fi'bU)=s('\ one has fjc,-' =
n«*' n^,! = na'6' exp £ ^/,io&^ [}- na'* iexpa(e,s(y)))}.
Note 6(,) = r(5(,)) = T(Pw(l)) = r(/>H>(1>) = r(s(1)) = £(1). Also, by (6), for
KJ<d one has b(n = T(s(l)) = T(s{n) - £ «5(/),5(/)>/V)-1 «.v0),s(/)>A)
F(s(/)) - blJ) - X (<)8'*(/^/3'6<,)>A)-, (<p'b(,\p'bU)>x)bU). Recall if 5(0) € span
{s0>|/ € <rf>) then 6(0) A b(0). If 5(0) ? span {.v(/)|y € <d>) then by (8) one has
&(0) = Sm - £ «p'b(l),l3'bU)>^-H<{i'bm,(3'bu)>/i)b{l).
i=\
This completes the demonstration of (iii).
OPTIMAL VALUE OF STOCHASTIC GEOMETRIC PROGRAMS 569
Finally, let j € <d> and recall L(e,s{j)) = £ s,(/) log e,-. Thus, pj — £ S/^V/ ar>d
<s(/), 5(/)>A [8, Th. 2.1.1, p. 29] since (ji\, ... , ixu)' is the mean vector and A is the
•sion matrix of e. Also, since L(e,sij)) is normal, one has </>,(/) = j= exp
OJ , V 27T
for every /€/?'.
APPENDIX B.
2w
PROPOSITION B: (a) If //is bounded then EM(v(Pc)) exists for every v € TV.
(b) //is bounded iff the set F A (8 € /?"|8, ^ Ov / € < « >, ^'8 = 0, and ^'8 = 1} is
bounded.
PROOF: (a) Assume //is bounded. Then for every j € <d> there exist real numbers /,
and Uj such that /, ^ r < «, for every r € //. Let v £ N and define zc A v(Pc). Finally,
assume q takes on the value c7 and recall r- = (/-,(c), ... ,rd(c))' is the element of //for
which 8 A bi0) + £ (rj(c))bin is the unique optimal point of Dc.
j= i
Since 8 is the optimal point of D-, by [4, Ch. 3, Sec. 3] and Assumption 3 one has
zjf =.Ko(c)"ll Kj(£)"jm J! 8,{r,)~"h'{r;) f\ M'/**^
where TJ \K(r)k"lr) A 1 for r € /Y, the closure of //, if p = 0. Define t: /7 — Rx by
r(r) A TJ 8 ,•(/■) ' ' Q \K(r) K ' for /■ € //. In evaluating t(/0 use the convention
xx = x~x =1 for x = 0. Then t is continuous on H. Thus, since // is compact, there exists
U € (0f«>) such that 0 < t(/-) < [/for every r 6 //. Hence,
(1) 0 < z/ < ITKvCcY n ATyCc)"'^.
I. Assume s(0) 6 span {s(/,|y € <</>}. Then by Proposition 3 there exists y, ; € Rx for
J € <</> and W € (0,°°) such that A"0(c) = ^fl ^(c)'7. Thus, by (1),
(2) 0 < z/ < (£W rj Kj{-cY(yi+r'(c)) .
Let j e <d>. If 0 < tf/c) < 1 then Kj{c)"yjKj{c)vrj < KjW'KjCc) J. Also, if
1 < A-7(c) then Kj(c)vyjKj(cYr]{c) < Kj<£)vyj Kjffi"' . Thus, 0 < ^•(c)I,(v^'(c))
^ Z/c) A max (Kj(c)v(yj+lj) , Kj(c)v{yj+Uj)) . Hence, by (2),
(3) 0 < z/ < (CW TJ Z,(c).
7=1
Moreover, by the choice of {6(/)|./ € <d>), the variates AT,(c) for j € <d> are independent
and hence {Zj(c)\j € <d>} is a set of independent variates.
570 P. ELLNER AND R.STARK
By definition ox Z.y(c) one has
(4) 0 < Zj(c) < Kj(c)v(yJ+lj) + Kj(c)v(yj+Uj\
Since Kj(c) is lognormal so are KjicY J+J and Kj(cY J+"J . Thus, the expected values of
these two variates exist and hence so does E(Kj(cY yj+J + Kj{cY yj+Uj ). Thus, by (4), the
expected value of Zj(c) exists. Since the variates Zjic) for j € <d> are independent the
expected value of {UWY f\ Zj(c) must also exist [5, p. 82, Th. 3.6.2]. Hence, by (3), the
expected value of zc" exists.
II. Assume s(0) tf span {s(,)\j e <d>). Then by the choice of{b{j)\j 6 <d>) the vari-
ates Kj(c) for y€ <d> are independent. For j € <J> define Z/c) A max
(Kj(cY\Kj(xYUi) and let Z0(c) A LFK^Y. Then the variates Z,{c) for y € <^> must be
independent. Furthermore, 0 < KjicY"' c < Z;(c) for j € <c/>. Hence, by (1),
(5) 0 < zF" < n z/^)-
Note that E(Z0(c)) exists since Z0(c) is lognormal. Also, for j € <d>, EiZ^c))
exists since A^7(c) is lognormal and 0 < Z^c) < KjicY ' + Kj(cY"J. Since the variates Zf(c)
for y e <d> are independent it follows from (5) that £"(z/) exists.
By I and II we conclude £(")(v(/>c)) exists for all v € N.
_ <y
(b) Observe 8 € F iff there exists r = (r, rd)' £ H such that 8 = 6<0) + £ r,60).
_ /-i
Also, //is bounded iff //is bounded. Thus, /"is bounded iff //is bounded.
REFERENCES
[1] Abrams, R.A., "Consistency, Superconsistency, and Dual Degeneracy in Posynomial
Geometric Programming," Operations Research, 24, 325-335 (1976).
[2] Aitchison, J. and J.A.C. Brown, The Lognormal Distribution, (Cambridge University Press,
London, England, 1957).
[3] Avriel, M. and D.J. Wilde, "Stochastic Geometric Programming," Proceedings Princeton
Symposium on Mathematical Programming, Princeton, New Jersey (1970).
[4] Duffin, R.J., E.L. Peterson and C. Zener, Geometric Programming, (John Wiley & Sons,
New York, New York, 1967).
[5] Fisz, M., Probability Theory and Mathematical Statistics, (John Wiley & Sons, New York,
New York, 1963).
[6] Hammersley, J.M. and D.C. Handscomb, Monte Carlo Methods, (John Wiley & Sons, New
York, New York, 1964).
[7] Isaacson, E. and H.B. Keller, Analysis of Numerical Methods, (John Wiley & Sons, New
York, New York, 1966).
[8] Lukacs, E. and R.G. Laha, "Applications of Characteristic Functions, (Charles Griffin &
Company, Ltd., London, England, 1964).
[9] McNichols, G.R., "On the Treatment of Uncertainty in Parametric Costing," Ph.D. Thesis,
The George Washington University, Washington, D.C. (1976).
[10] Munroe, M.E., Introduction to Measure and Integration, (Addison-Wesley Publishing Com-
pany, Reading, Massachusetts, 1953).
OPTIMAL VALUE OF STOCHASTIC GEOMETRIC PROGRAMS 571
[11] Rotar, V.I., "On the Speed of Convergence in the Multidimensional Central Limit
Theorem," Theory of Probability and Its Applications, 15, 354-356 (1970).
[12] Spivak, M., Calculus on Manifolds, (W.A. Benjamin, New York, New York, 1965).
[13] Stark, R.M., "On Zero-Degree Stochastic Geometric Programs," Journal of Optimization
Theory and Applications, 23, 167-187 (1977).
[14] Tsuda, T., "Numerical Integration of Functions of Very Many Variables," Numerische
Mathematik, 20, 377-391 (1973).
[15] Turchin, V.F., "On the Computation of Multidimensional Integrals by the Monte-Carlo
Method," Theory of Probability and Its Applications, 16, 720-724 (1971).
A CLASS OF CONTINUOUS NONLINEAR PROGRAMMING PROBLEMS
WITH TIME-DELAYED CONSTRAINTS
Thomas W. Reiland
North Carolina State University
Raleigh, North Carolina
Morgan A. Hanson
Florida State University
Tallahassee, Florida
ABSTRACT
A general class of continuous time nonlinear problems is considered.
Necessary and sufficient conditions for the existence of solutions are esta-
blished and optimal solutions are characterized in terms of a duality theorem.
The theory is illustrated by means of an example.
1. INTRODUCTION
Recently Fair and Hanson [1] proved existence theorems, duality theorems, and continu-
ous time analogues of the Kuhn-Tucker Theorem for a class of continuous time programming
problems in which nonlinearity appears both in the objective function and in the constraints.
More recently this class was extended in Farr and Hanson [2] to include problems with
prescribed time lags in the constraints. In this paper we generalize these results by considering
a more general form of the constraints and by assuming a less stringent constraint qualification.
This constraint qualification is analogous to that of Kuhn and Tucker [5] and provides further
unification between the areas of finite-dimensional and continuous time programming. An
example is presented wherein these results are applied to a version of Koopmans1 [4] water
storage problem which has been modified to address the economic ramifications of an energy
crisis.
2. THE PRIMAL PROBLEM
The problem under consideration (Primal Problem A) is:
Maximize
V(z) = foTcf>(z(t),t)dt
subject to the constraints
(1) z(t) > 0, 0 < t < T,
(2) f(z(t),t) < h(y(z,t),t), 0< / < T,
574 T.W. REILAND AND M.A. HANSON
and
(3) z(t) = 0, t < 0,
where z 6L,^ [0, T], i.e., z is a bounded and measurable w-dimensional function; y is a map-
ping from L~ [0, 71 x [0, 71 into Ep defined by
(4) v(z,f) = X J0 £,-(z(s - a7), 5 - ctj)ds\
f(z(t),t), h{y(z,t),t) € £m; $(z(s - a,-), s - ay) 6 #\ j - 0, . . . , r. The set 0 = a0 <
a] < . . . < ar is a finite collection of nonnegative numbers; and </> is a scalar function, concave
and continuously differentiate in its first argument throughout [0, T].
It is further assumed that each component of — /, g,, and h is a scalar function, concave
and differentiable in its first argument throughout [0,71, that each component of the composite
function h(y(-,t),t):L™[0,T] — ► Em is concave in z, there exists 8 > 0 such that
(5) either V*//^,/) - 0 or V*/ X-q.t) >h,
and for each t and k there exists ik = ik(t) such that
(6) Vkfik (.7i,t) > 8,
where
v*/,(i»,/) - e/ydj./va^*,
/ = 1 , . . . , m, k — 1, . . . , n,
forrj e E\ f) > 0, and t € [0,7];
(7) «/Cz(r),r)- 0, t < 0,
7 = 0, ..., r;
(8) VMv.O^dhib.tydvt > 0,
for v £ Ep and f € [0, 71 ; and
(9a)
(9b) sup g.AOj) < oo, sup Vkgja(0,t) < <*>, j = 0
(9c)
(9d)
sup /;,(0,n < oo,
sup VJ/,
(0,t)
<
oo:
Ik- 1, ... , «,
SUp £;,((),/) < oo
sup
v,
:ft(0,
7)
<
<? = 1, ... , p, k =
1, ... ,
n,
inf f,(0,t) > -oo
, /= 1,
, m,
sup V^<£ (rj.r) <
oo, 7, €
E\
TJ >
0,
k
A function z 6 L~ [0, 71 is termed feasible for Primal Problem A if it satisfies the con-
straints (1), (2), and (3). The primal problem is itself said to be feasible if a feasible z exists.
It should be noted that Primal Problem A is identical to that considered in [2] if p = m
and
h{yiz,t),t) = Imyiz,t)
NONLINEAR PROGRAMMING WITH TIME-DELAYED CONSTRAINTS 575
where Im is an m-dimensional identity matrix.
3. EXISTENCE THEOREM
THEOREM 1: If Primal Problem A is feasible, then it has an optimal solution, that is,
there exists a feasible z for which
Viz) = sup Viz),
where the supremum is taken over all feasible z.
We preface the proof of this theorem with a brief discussion of weak convergence and two
lemmas.
Let J be a normed linear space and denote by X* the collection of all bounded linear
functionals on X. If we define the norm of an element / € J* by
11/11= ^ l/(x)l
and define addition and scalar multiplication of linear functionals in the obvious manner, then
X* is a Banach space and is commonly referred to as the dual space of X. A sequence {x„} in X
is said to converge weakly to x € X if fixn) — - fix) as n — - °°, for every / 6 X*.
LEMMA 1: Let the uniformly bounded sequence of scalar measurable functions [q^it)},
d = 1,2, ... , converge weakly on [0,71 to q0it). Then except on a set of measure zero
<7o(') < Hm sup qdit).
PROOF: See Levinson [6]
LEMMA 2: If q is a nonnegative integrable function for which there exists scalar con-
stants 9X ^ 0 and 62 > 0 such that
qit) < 6>, +02J*O' qis)ds, 0 ^ t < T,
then<?(r) < 0,/2', 0 < / < T.
PROOF: See Levinson [6].
PROOF OF THEOREM 1: Let z be feasible for Primal Problem A and multiply the con-
straint (2) by the m-dimensional vector (1, . . . , 1) to obtain the inequality
£ /,(z(f),/) ^ £ h,(y(z,t),t), 0 < t < T.
From the convexity of each f, in its first argument, if follows from [8, p. 242] that
£ fiizit),t) > £ /,«)./) + £ akit) zkit),
;=1 r-1 k=\
where
OfcW - £ V*/,<0ff).
3/0 1 W. KL1LAINU AINU MA. HArOUIN
Set 0O — max JO, sup I - JT /(0,/)U by (9c) and observe that by assumption (6)
A = inf min ak(t) > 0.
/ k
Since z is feasible and therefore satisfies constraint (1), it then follows that
(10) A £ zk(i) ^ 90 + £ /j,(y(z, /),/), 0 < t < T.
k= 1 /= 1
Define
[Vs/tj.s)] = (V^Gm)}^, for t, 6 £", s € [0,71. 7 = 0, ... , r,
and
[VM^,/)] = {Vjhiiy.t)}^, for*, €£',/€ [0,71,
G(z,/,s) = y /(s) [V/;(y(z,/),/)k(z(5),5)
7to [o. '-«./]
and
H(z,t,s) = Y / (s) [V/;(y(z,/),/)][Vs,(z(s),s)]
where /^O) is the indicator function of the set E.
Since h and gj are concave in their first arguments it follows from [8] and from (3), (7) and (8)
that
h(y(z,t),t) < MO./) + J" G(0,t,s)ds + J" H(0,t,s)z(s)ds.
By (9a) and (9b) we select 0, > 0 and 02 > 0, such that
sup £ /;,(0,/) + y ( Gi(0,t,s)ds\ ^ 0,
' [£1 PlJo )
and
sup max 1 52 Z/,^ (0,/,s)[ ^ 02.
' k l'-i )
From (10) we have that 0* = (90 + 9\)/A and 9$= 9?/ A are nonnegative and positive con-
stants, respectively, for which
X z,(/) ^ o? + 6$ ('£, zk{s)ds, o< / < r.
From Lemma 2 we conclude that
(11) £ zyt(/) < 0f exp (0?/) < 0f exp (0J71, 0 < / < T,
k=\
and hence the set of feasible solutions for Primal Problem A is uniformly bounded on [0,71.
Since <f> is concave and differentiable in its first argument throughout [0,71, it follows
from (9d), [8] and the uniform boundedness property that, for any feasible solutions z and z°,
NONLINEAR PROGRAMMING WITH TIME-DELAYED CONSTRAINTS 577
V(z) - V(z°) < T £ sup izkit) - zk°it)) sup Wk<t>iz*it),t) < ~
k=\ ' '
and hence V is bounded above for all feasible z.
Let V = lub Viz), where the least upper bound Hub) is taken over all feasible z. Then
there exists a sequence (z^ of feasible solutions such that
lim viz*) = V.
Since {z^ is uniformly bounded, it follows from [10] that there exists a z to which a subse-
quence of {z^ converges weakly in L„2 [0, 71. Denote this weakly convergent subsequence
itself by {z^}; the application of Lemma 1 to each component of zd then provides uniform
boundedness for z except possibly on a set of measure zero where, as will be shown later, it can
be assumed to be zero.
Since each component of the composite function hiyi-,t),t) is concave in z, it follows
from [8], (3) and the chain rule for differentiation that
hiyizd,t),t) < hiyiz,t),t) + J' Hiz,t,s)izdis) - zis))ds, 0 ^ t < T.
Since each entry of the m x n matrix Hiz,t,s) is bounded and measurable, it follows that each
row Hjiz,t,s) € L~ [0,71 Q L} [0, 71 and so, by weak convergence,
f ' Hiz,t,s) izdis) - zis))ds — 0, as d — °°.
Thus, by constraint (2)
(12) lim sup fizdit),t) ^ hiyiz,t),t), 0 ^ t < T.
Define [Vfi-qj)] = ((V^(T>,f))mX„, tj € E", tj ^ 0; by the convexity of/
fizdit),t) > fizii),t) + [Vfizit),t)] izdit) - 2(f)), 0 < f < r.
Therefore, from (12),
(13) /(z(/),r) ^ h(y(z,0,t)
except on a set of measure zero, since by [8], assumption (5) and Lemma 1 we have
lim sup [Vfizit),t)] izdit) - zit)) > 0
except on such a set.
A second application of Lemma 1 to each component of zd provides
-2(f) ^ lim sup(- zdit)) ^ 0, a.e.m [0,T],
d—oo
and consequently z is nonnegative except on a set of measure zero. From this result and
expression (13), we observe that z can violate the constraints of Primal Problem A on, at most,
a set of measure zero in [0, 71. We define z to be zero on this set of measure zero, as well as
for t < 0, and equal to 2 on the complement of this set. The feasibility of z is then established
by noting that
y(z,t) = y&t), 0 < t ^ T,
and that
lim sup fizd(t),t) > /(0,f), 0 < f < T,
578 T.W. REILAND AND MA. HANSON
by the convexity nstraint (1), and assumption (6).
By the concavity and differentiability of $
fQ 0 (zd(t),t)dt ^ fj<t> izit),t)dt + jj izdit) - zit))' V<f>(zU),t)dt.
Therefore, by weak convergence
V = lim f <t>izdit),t)dt
^ fQ <f>(z(t),t)dt = Viz).
By the definition of V and the feasibility of z, Viz) < V, thus Viz) = K and z is an optimal
solution for Primal Problem A. Q.E.D.
4. WEAK DUALITY
Before the dual to Primal Problem A is formally stated, a continuous time Lagrangian
function and its Frechet differential will be introduced.
For u € L~ [0,71 and w € L~ [0,71, define
(14) Liu,w) = fQT [<f>iuit),t) + w'it) Fiu,t)]dt
where
Fiu,t) = hiyiu,t),t) - fiuit),t), 0 ^ t < T,
and let 8iL(«,w;y) denote the Frechet differential [7] with respect to its first argument,
evaluated at u with the increment y € L„°° [0, 71. The differentiability of each of the functions
involved in L insures that the Frechet differential exists and allows h\Liu,w\y) to be deter-
mined by the simple differentiation
(15) 8,Z.(w,w;<y) = -j- Liu + ay,w)|Q=0-
da
The Frechet differential has two additional properties that will be used in the ensuing discus-
sion, namely, the linearity of 81L(«,w;y) in its increment y and the continuity of b\Liu,w\y)
in y under the norm
\\y\\n = max llyjl00.
Here ||-||°° denotes the essential supremum [9, p. 112].
If y (/) = 0 for t < 0, then from (14) we have
(16) 8,Z,(«fw;y) = J0 [[V<f>(uU),t)]'yU)
+ J*o' w'it)Hiu,t,s)yis)ds -W'it)[Vfiuit),t)]yit)}dt.
An application of Fubini's theorem [9] to interchange the limits of integration enables us to
express (16) as
(17) 8,I(w,w;y) = bViu\y) + J*Q y'it)F*iu,w,t)dt,
where
8^(w;y) = /o [V<f>(u(t),t)]'y(t)dt
NONLINEAR PROGRAMMING WITH TIME-DELAYED CONSTRAINTS 579
and
(18) F*(u,w,t) = fT H'(u,s,t)w(s)ds - [Vfiu (/),/)]' wit), 0 < t < T.
With this notation the dual of Primal Problem A will be shown to be:
Dual Problem A:
Minimize
(19) Giu.w) = L(u,w) -81I(m,w;w)
subject to the constraints
(20) uit), wit) > 0, 0 < t < T,
(21) F*iu,w,t) + [Vcj>(u(t),t)] < 0, 0 < t < T,
(22) uit) = 0, t < 0
and
(23) w(/) = 0, / > T.
THEOREM 2 (Weak Duality): If z and («,w) are feasible solutions for Primal and Dual
Problems A, respectively, then
VU) ^ G(u,w).
PROOF: By the concavity of <j> and — / in their first arguments and the concavity of the
composite function hiy{-,t),t) in z it follows that L is concave in its first argument and
L(z,w) — L(u,w) < d\L(u,w;z — u).
Thus,
Viz) - Giu,w) = Liz,w) -f w'it)Fiz,t)dt
- Liu,w) +8\Liu,w;u)
< bxLiu,w\z - u) + 8xLiu,w\u)
- f w'it)Fiz,t)dt
T
= hxLiu,w;z) - JQ w'it)Fiz,t)dt
by the linearity of the Frechet differential in its increment. By (17) we have
8xLiu,w;z)- fQ w'it)Fiz,t)dt = fQ z'it){[V<f>iuit),t)] + F*iu,w,t)}dt
~So »"</)F(2;/)ift
which is nonpositive by constraints (1), (2), (20) and (21). Q.E.D.
From Theorem 2 it is observed that if there exist feasible solutions, z and iu,w), for the
primal and dual problems and if the corresponding primal and dual objective function values,
Viz) and Giu,w), are equal, then these solutions are optimal for their respective problems.
580 T.W. REILAND AND MA. HANSON
5. THE CONSTRAINT QUALIFICATION.
The constraint qualification introduced here is motivated by the form of the Kuhn-Tucker
constraint qualification presented by Zangwill [11] and also by Property 1 given below. The
basic theory surrounding this qualification is established to provide a framework for the
theorems of Section 6.
PROPERTY 1: If
(24) 8Viz;y) = fj y'it)V7<t>izit),t)]dt > 0
where z, y € L„°° [0, T] , then there exists a scalar o- > 0 such that
Viz + ry) > V(z), for 0 < r < a.
PROOF: By (15) and (24)
lim [Viz + ry) - Viz)]/r = 8 Viz\y) > 0,
TiO
thus a positive o- can be chosen which is sufficiently small so that
Viz + ry) > Viz), for 0 < t < cr. Q.E.D.
DEFINITION 1: For each z which is feasible for Primal Problem A, define Diz) to be
the set of ^-vector functions y for which
(i) yeLn°°[0,T]
(ii) yit) = 0, for t < 0
(iii) there exists a scalar o- > 0 such that
zit) +ryit) > 0, 0< t < T,
for
Fiz + ry.t) ^ 0, 0 < t < T,
0< -
DEFINITION 2: Define Diz) to be the closure of Diz) under the norm ||-||~- that is, if
a sequence [y**] C Diz) is such that \\yd — y 1 1" — » 0, as d — - oo, then y 6 Diz).
Henceforth, the Frechet differential of the mapping Fi-.t): L™ [0,71 — - Em evaluated at z
and with increment y, will be denoted by 8Fiz\y),. It should be observed that, for any
specified value of t € [0,T], the existence of 8Fiz,y), is ensured by the differentiability of/,
gj, and h and that when y it) = 0 for r < 0, we have
(25) 8F(z;y), = $^ H iz.t.s) y is) ds - [Vfizit),t)]yit).
Similarly, the Frechet differential of a component Fji-,t) of Fi,t), evaluated at z with incre-
ment y, will be denoted by 8Fjiz;y)t.
NONLINEAR PROGRAMMING WITH TIME-DELAYED CONSTRAINTS 581
DEFINITION 3: For each z which is feasible for Primal Problem A define 2>(z) to be the
set of ^-vector functions y for which
(i) y € L,r [0,71,
(ii) y{t) = 0, for / < 0,
(iii) ykU)^0a.e. in Tlk(z), k - 1 ... , n,
(iv) bF,{z\y), > 0 a.e. in T2,(z), / = 1 m,
where
Tu(z)= {/ € [0,7l:z*(f) = 0}, *- 1, ..., /i
and
T2i(z) = {t £ [0,71: F,(zff) = 0}, / = 1 m.
In a comparison of the sets Z)(z) and 3(z) with their finite-dimensional counterparts
presented in Zangwill [11], it is observed that D(z) is analogous to the set of "feasible direc-
tions" at z and 2(z) is analogous to that set of directions for which the directional derivatives
of each of the active constraints at z are nonnegative.
PROPERTY 2: Z)(z) C 9{z).
PROOF: Part 1. Let y € D(z). Then by Definition 1, there exists a scalar o- > 0 such
thatO < t ^ a- implies z{t) + Ty(t) > 0, 0 < t ^ T. Thus, if zk(t) = 0, thenyfc0) ^ 0.
Assume that Fjizj) = 0. If 8Fj(z\y), < 0, then by the same technique used in the proof
of Property 1 , it follows that for t sufficiently small,
Fj(z +Ty,t) < F,{z,t) = 0.
This result contradicts the assumption that y € D(z) and therefore we conclude that
D(z) C Mz).
Part 2. Assume that there is a y € L„°° [0,71 and a sequence {y*} C D(z) such that max
Wyk-VkW00 — 0> as rf^oo. Then for all /such that z*(r) = 0, yj^(r) ^ 0, d = 1, 2, . . . . It
then follows from convergence in L°° [0,71 that yk(t) ^ 0 a.e. on T]k(z), k = 1, . . . , «.
Assume there exists an / and a set E of positive measure over which Fj(z,t) = 0 and
8Fj(z;y), < 0 for all / € £. By the continuity of 87](z;-), in the L°° norm [7], we can choose
a J* sufficiently large such that for d ^ d*
8/;-(z;y<0, < 0
over some subset of £ which has positive measure. This result yields _a contradiction to Part 1
since it was assumed {y^ C D(z) and we can therefore conclude that D(z) c 2)(z). q E D
DEFINITION 4 (Constraint Qualification): Primal Problem A will be said to satisfy the
Constraint Qualification if the problem is feasible and if
D (z) =9<z),
582 T.W REILAND AND MA HANSON
where z is an optimal solution to the problem.
In more general problems where convexity and concavity properties are not assumed, the
purpose of the Constraint Qualification would be to eliminate "cusps" in the feasible region.
For example, the constraints
z,(/) ^ 0, z2U) ^ 0, 0 < t < T,
and
[I - ZlU)V - z2U) > 0, 0< t < T,
do not satisfy the Constraint Qualification when z(t) = (1,0), 0 < / < T, since
(1/2,0) € #(z) but (1/2,0) <t D{z).
In problems such as Primal Problem A where convexity and concavity properties are
assumed, violations of the Constraint Qualification can be shown to arise when the constraints
take the form of equalities on some set of positive measure. For example, consider the con-
straints
ZlU) > 0, z2(t) > 0, 0 < t < T,
and
biW + z2(0 - l]2 < l - IEU), 0 < t < T,
where £ is a set of positive measure in [0,T] and IE{) is its indicator function. It is observed
that for z(t) = (1/2,1/2), we have (1,1) 6 9{z) but (1,1) $ D(z), thus the Constraint
Qualification is not satisfied.
THEOREM 3: If z is optimal for Primal Problem A, then under the Constraint
Qualification
8 K(z;y) < 0, for all y <E 2{z).
PROOF: Part 1. Suppose there exists a y 6 D(z) such that 8 V(z;y) > 0. Then by Pro-
perty 1 there exists a o- > 0 such that 0 < t ^ o- implies V(z + ry) > V(z); however, since
y € D(z) we can choose a <r0 sufficiently small so that z + a-0y is feasible for Primal Problem
A. Thus, by contradiction of the optimality of z, we can conclude that 8 V(z;y) < 0, for all
y € D(l).
Part 2. Let {y^ be a sequence of functions in D(z) and let y° be such that max
k
I \y( ~ 7 k 1 1°° ~" " 0, as d — * oo. It then follows from Part 1 and the continuity of 8 K(z;) that
8 K(z>°) = lim 8 K(z><0 ^ 0.
Thus, 8 K(z;y) < 0 for all y <E 5 (z). Q.E.D.
6. DUALITY AND RELATED THEOREMS
In proving strong duality and its related theorems two additional assumptions will be
made. These are:
(26) H(z,t,s) ^ 0, 0 < s ^ t < T
NONLINEAR PROGRAMMING WITH TIME-DELAYED CONSTRAINTS 583
and
(27) F(z,t) - 8F(z;z), > 0, 0 < t < T,
where z is an optimal solution for Primal Problem A. It will be shown in Corollary 1 that
assumption (26) is implied if z (r) = 0 is feasible.
THEOREM 4 (Strong Duality): Under the Constraint Qualification and assumptions (26)
and (27), there exists an optimal solution (u,w) for Dual Problem A such that u = z and
G(z,w)= V{z).
Before proving Theorem 4 the following linearized problem, called Primal Problem A', will
be considered:
Maximize
8 V(z;z - z)
subject to the constraints
(28) z(t) > 0, 0< / ^ T,
(29) Fat) + 8F(z;z - z)t > 0, 0 < t < T,
and
(30) z(r) = 0, for/ < 0.
LEMMA 3: Under the Constraint Qualification, z is an optimal solution for Primal Prob-
lem A'.
PROOF: If z is feasible for Primal Problem A', then
Ht) - z(t) > 0, for t € Tlk(z), k=\, .... n,
and
8Fj(z,z - z), ^ 0, for / € T2i(z), i = 1, . . . , m,
and therefore (z — z) € 2(j). It then follows from Theorem 3 that, under the Constraint
Qualification,
8 V(z;z - z) < 0
for all z satisfying (28) , (29) and (30) . The optimality of z follows since z is feasible for Primal
Problem A' and since 8 V(z;0) = 0. 0 E D
PROOF OF THEOREM 4: We rewrite Primal Problem A' in the form
maximize
J0 a'(t)z(t)dt
subject to the constraints
zit) > 0, 0< t < T,
584 T.W. REILAND AND MA. HANSON
and
B(t)zit) ^ c(t) + fi KU,s)z(s)ds, 0 < t < T,
where ait) = [V0(z(f),/)1, Bit) = [Vfizit),t)], cit) = Fiz.t) - 8F(z;z)t, and Kit,s) =
Hiz,t,s). From assumptions (5), (6), (26) and (27) it is observed that the matrices ait),
Bit), cit) and K(t,s) satisfy the requirements of Grinold's Duality Theorem [3]. Therefore,
there exists an m- vector function w satisfying
(31) wQ) ^ 0, 0 ^ t ^ T,
and
(32) B' it) wit) ^ a(f) + J £'(«*) w(s)4 0 < t < r,
such that
(33) .C^') c(t)dt = fj a'(t)z(t)dt.
Setting wit) = 0 for t > T, we observe from the identities (14), (17), and (18) that expres-
sions (32) and (33) can be expressed as
(320 F*iz,w,t) + tV0(F(f),/)] < 0, 0 < t < T,
and
(330 L(z,w) - hxLiz,w;z) = Viz),
respectively. From (31) and (320 and the fact that wit) = 0 for t > T, it then follows that
iz,w) is feasible for Dual Problem A and, from (19) and (330
(34) Giz,w) = Viz).
Finally, by the weak duality established in Theorem 2, it is concluded from (34) that iz,w) is
an optimal solution for Dual Problem A. Q.E.D
In order to apply Theorem 4 in practice, it is desirable to be able to verify conditions (26)
and (27) without prior knowledge of the optimal solution z. The following corollary provides
this capability.
COROLLARY 1: If
VkgjM,t)lfrqk = dgj,(ri.t)/B7ik > 0,
(35) j = 0 r, i = 1, ... , p, k=\, ... , n,
for tj € E", ri ^ 0, and 0 < t < T,
(36) Fi0,t) ^ 0, 0 < / < T,
then under the Constraint Qualification there exists an optimal solution iu,w) for Dual Problem
A such that u = zand Giz.w) = Viz).
PROOF: We have from (8) and (35) that
Hiz.t.s) - £ / is) [Vhiyiz,t),t)) [Vgjizis),s)] > 0, 0 < t < T,
and by (36) and the concavity of F that
Fiz,t) - 8Fiz;I), > FiO.t) > 0, 0 < t < T.
NONLINEAR PROGRAMMING WITH TIME-DELAYED CONSTRAINTS 585
From these results it follows that the conditions of Theorem 4 are satisfied. Q.E.D.
THEOREM 5 (Complementary Slackness Principle): If z and (z,w) are optimal solutions
for the Primal and Dual Problems A, then
(37) fQTw'(t)Fat)dt = 0
and
(38) fQTz'(t){F*(z,w,t) + [V<f>GU),t)])dt = 0.
PROOF: Since z(t) ^ 0 and F*{z,w,t) + [V0(z(f),f)] < 0, 0 < t < T, it follows from
identity (17) that
J0 z'(t){F*(z,w,t) + [V<f>a(t),t)]}dt = 8xL(z,w;z) ^ 0,
and therefore, by (330
L(z,w) - V(z) = fj w'(t)F(z,t)dt < 0.
Since w(t) ^ 0 and F(z,t) ^ 0, 0 ^ t < T, it also follows that
(39) J* w'U)F(z,t)dt ^ 0,
thus the equality in (37) is established.
Similarly, (33') and (39) imply that
8,£(z,w;z) > 0
and therefore, by (17)
fQTz'U){F*(z,w,t) + [V<t>(zU),t)])dt > 0.
Since z(t) ^ 0 and F*(z,w,t) + [V<f>(z(t),t)] ^ 0, 0 < t < T, we have
Jorr(f){F*(z,^) + [V0(z(/),/)]}<ft < 0
and thus the equality in (38) is established. Q.E.D.
THEOREM 6 (Kuhn-Tucker Conditions): Assume that (35) and (36) are satisfied for
Primal Problem A. Then under the Constraint Qualification z is an optimal solution if and only
if there exists an m- vector function w such that
(i) F*(z,w,t) + [V0(z(f),f)l < 0, 0 < t < r,
(ii) fJz'UHFHZw.t) + [V<f>(z(t),t)]}dt = 0
(iii) J0 w'(t)F(z,t)dt - 0
(iv) »(/) > 0, 0 < / < Tand w(f) = 0, / > T.
PROOF:
Necessity: The necessity of the conditions follows from Corollary 1 and Theorem 5, since
the m-vector function vv of the optimal solution (z,w) to Dual Problem A satisfies conditions (i)
through (iv).
586 T.W REILAND AND MA. HANSON
Sufficiency: Let z be feasible for Primal Problem A. Then since Fis concave
Viz) - Viz) < 8 V(z;z - z)
= Jj [zit) -Iit)]'[V<t>izit),t)]dt.
Since z(f) ^ 0, 0 < / < T, it follows from conditions (i) and (ii) that
Viz) - Viz) ^ - f hit) - zit)]'F*iz,w,t)dt,
♦'O
and by (18), (25) and Fubini's Theorem [9]
fo [zit)-zit)]'F*iz,w,t)dt = fo w'it)iFiz;z-z)tdt.
By (i), (iii) and the concavity of F,
- J*Q w'it)8Fiz;z - z),dt < - / w'it)[Fiz,t) - Fiz,t)]dt
= - X *'it)Fiz,t)dt
which is nonpositive since wit) ^ 0 and Fiz,t) ^ 0, 0 ^ t < T. Thus, Viz) < Viz) and zis
an optimal solution for Primal Problem A. Q.E.D.
7. EXAMPLE - WATER STORAGE PROBLEM
In the water storage problem posed in [4], the hydroelectric company incurred a penalty if
it could not meet a prescribed demand for power. This penalty was characterized in the objec-
tive function
jJ^iDit)- Pit))dt
where [0,71 represents a planning period of specified duration, Dit) is the demand rate, Pit)
is the production rate of hydroelectric power, and «/» is the penalty function which was assumed
to be strictly convex. The imposition of such a penalty favors the consumer or a middleman
utility company which retails electric power to the consumers. In short, it characterizes a
"buyers market."
If there is, in fact, a pending energy crisis, it seems appropriate to consider a "sellers
market" where the demand for power exceeds production capacity and a premium is paid to the
hydroelectric company for any power which it produces beyond some prescribed level. In the
case where the hydroelectric company is supplying power directly to the consumer, these premi-
ums may take the form of increasing prices per unit beyond some allotment level. When the
hydroelectric company is supplying a middleman, the premiums may represent an incentive pol-
icy which encourages maximum production during peak demand periods.
The premiums to the hydroelectric company will be represented by
fjir iPit)- Ait))dt
where [0,71 represents the planning period, Pit) is the power production rate, Ait) is the
prescribed aggregate allotment or incentive level, and it is the premium function which is
assumed to be differentiable and concave with a positive slope at zero.
For the dynamics of the problem, we assume a confluent system of rivers supplying water
to a hydroelectric plant on the main stream with r of its tributaries also having their own
NONLINEAR PROGRAMMING WITH TIME-DELAYED CONSTRAINTS 587
hydroelectric plants. The variables and parameters which relate to the dam, reservoir and plant
on the main stream will be subscripted by 0, and those for the r dammed tributaries by j,
j- 1. .... r.
We let Cl j denote the initial store of water in reservoir j and 0; the capacity of reservoir j.
The rate of spillage and rate of discharge through the turbines of dam ./at time /are denoted by
S/(t) and dj(t), respectively. The rates of inflow of water into the reservoirs on the dammed
tributaries are £,(/) 7=1, ... , r, and that into the main reservoir from its undammed tribu-
taries is £o(t).
It is assumed that it takes aJt j = 1, .. . , r units of time for the water released from dam
j to reach the main reservoir and that there is no spillage or discharge through the dams on the
tributaries for at least a units of time prior to the start of the planning period, where
a = max {a,}. The store of water in reservoir j at time rcan then be expressed as
Wj(t) = ft; + J0' (gjW) ~ Sj(t') - dj(t'))dt'
for j = 1 , . . . , r, and
w0(t) -
for the main reservoir.
: fto + /0'|fo('') - s0W) - d0W) + £ (sjU'-ccj) + djU'-aj))\dt'
The power production rate for a given rate of discharge d is assumed to be proportional to
d. In [4], it was necessary to assume the factor of proportionality to be unity. Here we allow
this factor to be proportional to the head of water in the reservoir, an assumption which is con-
sistent with constant turbine efficiency. The head is the difference h between the surface level
of the reservoir and the tailwaters below the dam and is therefore dependent primarily upon the
store of water Win the reservoir.
The relationship between hj(t), the head of reservoir j, and Wj(t) will be represented by
hj(t) = h* (Wj(t)), where h* is an increasing concave differentiable function. The functions
h* owe their concavity to the shapes of the reservoirs which are assumed to yield a continu-
ously disproportionate increase in reservoir surface area as the store of water increases. The
production rate for the yth hydroelectric plant is then expressible as
Pj(t) = dj(t) ° h* (WjU)) ,
in which case the production rate for the entire system becomes
PU)= £pjU).
Assuming the role of the hydroelectric company, we want to select our water storage pol-
icy (s,d) so as to maximize the premium payments over the planning period. This problem
takes the form
ris.d) = f ir (Pit) - A(t))dt
•'O
588 T.W. REILAND AND MA HANSON
subject to
0 ^ SjQ) < pj(t)
0 ^ dj(t) ^ <t>,
0 < WjU) ^ 9j
./ = 0, .... r, and
A(t) < Pit)
for 0 ^ / ^ T, where /3,-(/) is the maximum allowable spillage rate through dam 7 and </>, is the
turbine capacity of plant /
Through proper association of the terms of this model with those of Primal Problem A it
can be shown through application of Theorem 1 that feasibility ensures the existence of an
optimal water storage policy which will maximize the total premium payment.
REFERENCES
[1] Farr, W.H. and M.A. Hanson, "Continuous Time Programming with Nonlinear Con-
straints," Journal of Mathematical Analysis and Applications 45, 96-115 (1974).
[2] Farr, W.H. and M.A. Hanson, "Continuous Time Programming with Nonlinear Time-
Delayed Constraints," Journal of Mathematical Analysis and Applications 46, 41-60
(1974).
[3] Grinold, R., "Continuous Programming Part One: Linear Objectives," Journal of
Mathematical Analysis and Applications 28, 32-51 (1969).
[4] Koopmans, T.C., "Water Storage in a Simplified Hydroelectric System," Proceedings of the
First International Conference on Operational Research, M. Davies, R.T. Eddison and T.
Page, Editors (Operations Research Society of America, Baltimore, 1957).
[5] Kuhn, H.W. and A.W. Tucker, "Nonlinear Programming," Proceedings of the Second Berke-
ley Symposium on Mathematical Statistics and Probabilities, 481-492, J. Neyman, Editor
(University of California Press, Berkeley, 1951).
[6] Levinson, N., "A Class of Continuous Linear Programming Problems," Journal of
Mathematical Analysis and Applications 16, 73-83 (1966).
[7] Luenberger, D.G., Optimization by Vector Space Methods (Wiley, New York, N.Y., 1969).
[8] Rockafellar, R.T., Convex Analysis (Princeton University, Princeton, New Jersey, 1970).
[9] Royden, H.L., Real Analysis (MacMillan, New York, 1968).
[10] Taylor, A.E., Introduction to Functional Analysis (Wiley, New York, N.Y., 1958).
[11] Zangwill, W.I., Nonlinear Programming: A Unified Approach (Prentice Hall, Englewood
Cliffs, New Jersey, 1969).
EQUALITIES IN TRANSPORTATION PROBLEMS AND
CHARACTERIZATIONS OF OPTIMAL SOLUTIONS*
Kenneth O. Kortanek
Department of Mathematics,
Carnegie-Mellon University
Pittsburgh, Pennsylvania
Maretsugu Yamasaki
Department of Mathematics
Shimane University
Matsue, Shimane, Japan
ABSTRACT
This paper considers the classical finite linear transportation Problem (I) and
two relaxations, (II) and (III), of it based on papers by Kantorovich and Rubin-
stein, and Kretschmer. Pseudo-metric type conditions on the cost matrix are
given under which Problems (I) and (II) have common optimal value, and a
proper subset of these conditions is sufficient for Problems (II) and (III) to
have common optimal value. The relationships between the three problems
provide a proof of Kantorovich's original characterization of optimal solutions
to the standard transportation problem having as many origins as destinations.
The results are extended to problems having cost matrices which are nonnega-
tive row-column equivalent.
1. INTRODUCTION WITH PROBLEM SETTING
Over 25 years ago Kantorovich in his classic paper, "On the translocation of masses" [4],
formulated generalized transportation problems which are continuous analogs of the well-known
transportation problem in the theory of finite linear programming. He raised the question of
characterizing optimal solutions to those problems whose finite dimensional versions have the
same number of origins as destinations. As is well known, optimal solutions to the standard
finite dimensional transportation problem having "m" origins and "n" destinations are charac-
terized by means of a system of linear inequalities involving m row numbers and n column
numbers which together comprise a feasible list of dual variables.
Within the finite dimensional context m = n, Kantorovich's goal was to use only n
numbers in a linear inequality system characterization of an optimal solution rather than the
standard In (row plus column) numbers. In order to accomplish this, three conditions defining
a pseudo-metric were imposed on the cost coefficient matrix. Actually, the triangle inequality
condition on unit costs is what Gomory and Hu later termed "reasonable costs" in their network
*The work of the first author was supported in part by National Science Foundation Grants ENG76-05191 and
ENG78-25488.
590 K..O. KORTANEK AND M. YAMASAKI
studies [3], Section 2. Violation of this particular condition is also related to the "more for less"
paradox in the transportation model, see Ryan [7].
The original application of the pseudo-metric conditions involved subtleties which were
later clarified in Kantorovich-Rubinstein [5] but for a transformed version of the standard tran-
sportation problem, which we state as Problem III in the next section. In attempting to give a
proof of Kantorovich's characterization, Kretschmer [6] introduced yet another transformation
of the standard problem, which we shall term Problem II in the next section.
The basic purpose of this paper is to delineate the key relationships between these three
problems: the standard transportation Problem I, the Kretschmer transformed Problem II, and
the Kantorovich-Rubinstein Problem III. The results we obtain depend on how the three
pseudo-metric cost conditions, denoted (C.l) through (C.3) in Sections 3 and 4, are coupled
together.
Our main application is to obtain a proof of the originally sought for characterization of
optimal solutions of the standard transportation problem where the number of origins equals
the number of destinations. We are not prepared at this time however to state that we have
industrial or public sector applications of the type II or type III transportation models.
2. THE KANTOROVICH-RUBINSTEIN AND KRETSCHMER TRANSFORMS
OF THE STANDARD TRANSPORTATION PROBLEM
Let Cy, a, and bj(i — 1, .... n\ j =■ 1, . . . , n) be nonnegative real numbers and assume
that a, and bj satisfy
(1.1) X>/=I>,>0.
r-i j=\
The original transportation problem may be expressed as follows:
(I) Determine the minimum value Mof
(1.2) ttdjXu
i-l 7=1
subject to the condition that xtj are nonnegative and
(1.3) jtxij-a, (/=1 n),
7=1
£*/, = bj (/- 1, ... , n).
i=\
Let us consider the following transportation problems "which were studied in [5] and [6]:
(II) Determine the minimum value N of
(1.4) H^M+yy)
1=1 7=1
subject to the condition that xtj and y^ are nonnegative and
(1.5) Z (*,,->>/,) = a, 0-1, .... n),
7=1
lt(xij-yij) = bj 0-1, .... n).
EQUALITIES IN TRANSPORTATION PROBLEMS 591
(III) Determine the minimum value V of
(1.6) ttcuzu
i-l 7-1
subject to the condition that zy- are nonnegative and
(1.7) £ Zjj - £ zy, = a, - bt 0=1 n).
7=1 7=1
Program I of course is the classical transportation problem which may be solved by the
well-known row and column number method ([1],[2]) and other more modern, large scale pro-
gramming methods. The row and column number method easily extends to solving Program II.
On the other hand, the structural matrix of Program III is a network incidence matrix, and so
III is an uncapacitated network problem.
It is clear that V < M and N ^ M and in this sense Problems II and III are relaxations of
Problem I. We shall study when one of the equalities V = N, V = A/, and M = N holds.
3. THE EQUALITY N = V OF PROBLEMS II AND III
First we have
LEMMA 1: The inequality V < N holds if the following condition is fulfilled:
(C.l) Cy = Cjj for all i and j.
PROOF: There exists an optimal solution xy and jvy of Problem (II), i.e., xi} and y,j are
nonnegative and satisfy (1.5) and
i-I 7=1
Then
\Lxij + i,yji\ ~ It,** + !>// - a> ~ bi-
[j-i y-i J u-i y-i j
Taking zy = xu + yJh we see that zu are nonnegative and satisfy (1.6), so that by condition
(C.l)
y<tlcuzu = N.
i-l 7-1
THEOREM 1: The equality V— N holds if condition (C.l) and the following condition
are fulfilled:
(C.2) c„ = 0 for all /'.
PROOF: There exists an optimal solution zy of Problem (III), i.e., zti are nonnegative
and satisfy (1.7) and
/-I7-1
592 K.O. KORTANEK AND M YAMASAKI
Then
£ z,j + b, , - £ z7, + a, = 4 > 0.
7=1 7=1
Let us take Xy = 0 if / ^ y and x„ = d, and put jfy = Zyt. Then xy and jty are nonnegative and
satisfy (1.5), so that
* s £ £ cMj + .v-y) - £ £ djZji = y
by conditions (C.l) and (C.2).
We show by an example that the equality N = V does not hold in general if we omit con-
dition (C.2).
EXAMPLE 1: Let n = 2 and take
C\\ = c22= 1, c12= c2i = 2,
a\ — 1, a2 = 2, 6] = 2, 61 = 1.
Then we easily see that K = 2 and Af = TV = 4.
4. THE EQUALITY M = N OF PROBLEMS I AND II
Next we show that the equality M = N does not hold in general even if both conditions
(C.l) and (C.2) are fulfilled.
EXAMPLE 2: Let n = 3 and take
cu - c22 - c33 - 0, c12=c21 = 20,
Cl3 = C31 = C23 = C32 = 1,
fl! = 3/2, a2= 1/2, a3= 1/4,
bx= b2= 1, 63 = 1/4.
optimal solution of Problem (I) is given by xn = 1, x22 = 1/2 x12 = x13 = x32 = 1/4 and
*2i = *3i - *23 = *33 = 0- we have N = I. An optimal solution of Problem (II) is given by
*n = 1, *22 = 1/2, xu = x32 = 1/2, x12 = x2, = x23 = x31 = x33 = 0, y33 = 1/4 and yu = 0 if
ay) * (3,3).
Our main result is the following one.
THEOREM 2: The equality M = N holds if the following condition is fulfilled:
(C.3) qj < ciq + cpj + cw for all /, j, p, q.
PROOF: There exists an optimal solution Xjj and yu of Problem (II). In case
Zjj = Xjj — yjj is nonnegative for each i, /, we see that z^ is a feasible solution of Problem (I) , so
that
A/<II cijzij <££ cu(x0 + y0) = N.
'=1 7=1 '-1 7=1
EQUALITIES IN TRANSPORTATION PROBLEMS 593
We consider the case where some x/y - ytj are negative. We may assume that min(x,7, yu) = 0
for all / and / There exist p and q such that 0 = xpq < ypq. Then we have by (1.5)
Z xiq kypq and £ xpj ^ yp,,.
/=1 7=1
Let us define Ah Bj and dy by
Ap = Bq - 0,
A = *fcW Z x«? U * p), Bj = xwj>w/ X Jf/y 0" * Q),
flfc - AtBjly„.
Then
(4.1) Z4/ = ^ <*,<,, td0=Bj<xPJ>
/-i ,=i
(4.2) i^-i^-jw-
,-i y-i
We define x,, and ^ by
(4.3)
X'ij = Xij + dy
if i ^ p and y ^ #,
Xpj = Xpj — Bj
if / * 9,
Xjq Xjq Aj
if / ^ a
y\j = y>j
if i ^ p or y ^ 9,
x'n-y'n- 0.
Then x\j and v,y are nonnegative and satisfy (1.5) and
N^tt Cijixlj+ylj)
/-I 7=1
= 11 CijiXij + yu) + £ £ rfylcy - c„ - cOT - c,,]
/=iy=i ,-iy-i
^ZZ tyCxfc+j^-tf
/=1 7=1
by condition (C.3). Repeating the above procedure (4.3) a finite number of times,t we obtain
xfj which are nonnegative and satisfy (1.3) and
^ = Zl¥5-ii CM + y^ = n.
/-I 7-1 '-1 /-I
Hence, M = N.
THEOREM 3: Let ky, ft and gj be nonnegative numbers and assume that condition
(C.3) holds for ky instead of Cy. If Cy = k0 + /, + #,, then M = N.
hw,'
This number is at most the number of {>»y > 0} .
594 K.O KORTANEK AND M YAMASAKI
PROOF: Denote by M(k) and Nik) the values of Problems (I) and (II) respectively if
Cjj are replaced by ktj. Then we have M = M(k) + C/g with
cfs = £ Mi + £ gjbj.
<=i j-l
Let Xjj and v,y be nonnegative and satisfy (1.5). Then
1-1 /-I /=U=1
^11 fcy(xy + j/y) + C/g > AT(/r) + CA.
1=1 y=l
Thus, N £ iV(Ar) + Cfg. Since #(*) = M(k) by Theorem 2, we have N ^ Af(A:) +
Cfg = M, and hence M = N.
As is well known, two transportation problems of type (I) with costs [cy] and
[cjj + fj + gj) respectively, are equivalent for any list of real numbers {/}}, {#,}, /' = 1, .... m;
j — 1, . . . , «. The following example shows that the nonnegativity of all the // and gj is
required in Theorem 3.
EXAMPLE 3: Let n = 2 and take fc„ = 0, kn = 1/2, fc21 = 1/2, k22 = 0, a, = 1, a2 = 1,
bx = 1/2, b2 = 3/2, /, = 1, f2 = -5/2, ?j = 2, g2 = 7/2. Then, Jl/(fc) = N(k) - 1/4 while
AT - 9/2 < 5 - Af.
5. KANTOROVICH'S THEOREM FOR PROBLEM (I)
The finite version of Kantorovich's Theorem [4] can be written as follows:
A feasible solution xtJ of Problem (I) is an optimal solution if and only if there exist
numbers w, such that
(5.1) \uj — Uj\ < cu for each /, j,
(5.2) U, - Uj= Cy ifxy > 0.
We show that this theorem is not valid as it stands. In fact, let us recall Example 2 and
let Xjj be the optimal solution obtained there. If there exist numbers w, which satisfy (5.1) and
(5.2), then we must have
"i - "2 - en = 20,
u3- u2= c32= 1,
"1 - "3= Cn= 1.
This is impossible.
In order to give another proof of Kantorovich's Theorem, Kretschmer considered Prob-
lem (II) and asserted N = M without any assumption. Notice that N < Mm Example 2.
EQUALITIES IN TRANSPORTATION PROBLEMS 595
Kantorovich's Theorem was amended by Kantorovich and Rubinstein [5; Theorem 3] in
the following form:
THEOREM 4: Assume that conditions (C.l), (C.2) and (C.3) hold. Then a feasible
solution Xy of Problem (III) is an optimal solution if and only if there exist numbers «, which
satisfy (5.1) and (5.2).
Under conditions (C.l) and (C.2), the dual problems of Problems (II) and (III) coincide
and Theorem 4 is an immediate consequence of the well-known duality theorem applied to
Problem (II). Thus, condition (C.3) can be omitted in Theorem 4.
Notice that conditions (C.l), (C.2) and (C.3) hold if and only if the cost c^ is a pseudo-
metric, i.e., Cjj satisfies conditions (C.l) and (C.2) and the following condition
(C.4) cu < cik + ckj for all /, j, k.
With the aid of Theorems 2 and 4, we have
THEOREM 5: Assume that conditions (C.l), (C.2) and (C.3) hold. Then a feasible
solution Xy of Problem (I) is an optimal solution if and only if there exist numbers w, which
satisfy (5.1) and (5.2).
ACKNOWLEDGMENT
We are indebted to a referee for helpful comments, weakening the original assumptions of
Theorem 2, in particular.
REFERENCES
[1] Charnes, A. and W.W. Cooper, Management Models and Industrial Applications of Linear Pro-
gramming, /and //, (J. Wiley and Sons, New York, N.Y., 1961).
[2] Dantzig, G.B., Linear Programming and Extensions, (Princeton University Press, Princeton,
1963).
[3] Gomory, R.E. and T.C. Hu, "An Application of Generalized Linear Programming to Net-
work Flows," Journal of the Society for Industrial and Applied Mathematics, 10, 260-283
(1962).
[4] Kantorovich, L.V., "On the Translocation of Masses," Management Science, 5, 1-4 (1958).
(English translation of Doklady Akademii Nauk USSR, 37, 199-201 (1942).
[5] Kantorovich, L.V. and G. Sh. Rubinstein, "On a Space of Completely Additive Functions,"
Vestnik Leningrad University, 13, 52-59 (1958) (Russian).
[6] Kretschmer, K.S., "Programmes in Paired Spaces," Canadian Journal of Mathematics, 13,
221-238 (1961).
[7] Ryan, M.J., "More on the More for Less Paradox in the Distribution Model," in Extremal
Methods and Systems Analysis, An International Symposium on the Occasion of Professor Abra-
ham Charnes' Sixtieth Birthday, A.V. Fiacco, K.O. Kortanek (Editors), 275-303, Volume
174 of Lecture Notes in Economics and Mathematical Systems, Managing Editors: M.
Beckmann and H.P. Kiinzi, Springer- Verlag, Berlin-Heidelberg-New York, 1980.
A NETWORK FLOW APPROACH FOR CAPACITY
EXPANSION PROBLEMS WITH TWO FACILITY TYPES
Bell Laboratories
Holmdel, New Jersey
ABSTRACT
A deterministic capacity expansion model for two facility types with a finite
number of discrete time periods is described. The model generalizes previous
work by allowing for capacity disposals, in addition to capacity expansions and
conversions from one facility type to the other. Furthermore, shortages of
capacity are allowed and upper bounds on both shortages and idle capacities can
be imposed. The demand increments for additional capacity of any type in any
time period can be negative. All cost functions are assumed to be piecewise,
concave and nondecreasing away from zero. The model is formulated as a
shortest path problem for an acyclic network, and an efficient search procedure
is developed to determine the costs associated with the links of this network.
INTRODUCTION
In a previous paper [9], we described a deterministic capacity expansion model for two
facility types. The model has a finite number of discrete time periods with known demands for
each of the two facilities in any period. At the beginning of each period, facility /'(/' =1,2)
may be expanded either by new construction or by converting idle capacity of one facility to
accommodate the demand for the other facility.
In this paper, we extend our previous work by allowing for the reduction of facility size
through capacity disposals. Furthermore, shortages of capacity are allowed and upper bounds
on idle capacities and shortages may be imposed. These generalizations allow us to deal with
more realistic situations. Capacity disposals are often initiated due to high holding cost of idle
capacity when the cumulative demand decreases over some successive periods. Capacity shor-
tages may be attractive when capacity may be temporarily rented or imported from other
sources. Also, in some applications it may be economical to permit temporary shortages and
pay a penalty for unsatisfied demand, rather than expanding the facilities at that time. Finally,
upper bounds on idle capacity and shortages are usually imposed by management.
The costs incurred include those for construction of new capacity, disposal of existing
capacity, conversion, holding of idle capacity, and for having capacity shortages. As in [9],
conversion implies physical modification so that the converted capacity becomes an integral part
of the new facility and is not reconverted automatically at the end of the period. The capacity
expansion policy consists of timing and sizing decisions for new constructions, disposals, and
conversions so that the total costs are minimized.
597
598 h. luss
The model is useful for communication network applications, such as the cable sizing
problems examined in [9]. Suppose the demands for two cable types is known for the next T
periods. Furthermore, suppose the more expensive cable can accommodate both demand
types, whereas the cheaper cable can be used only to satisfy its associated demand. Since the
construction cost functions are often concave, reflecting economies of scale, it can become
attractive to use the more expensive cable for future demand for both cables. Thus, careful
planning of the expansion policy is needed. A similar application is the planning of capacity
expansion associated with communication facilities which serve digital and analog demands.
Other areas of applications include production problems for two substitutable products, and
inventory problems of a single product produced and consumed in two separate regions; see [9]
for more details.
Many capacity expansion models and closely related inventory models have been
developed for the single facility problem with a finite number of discrete time periods. The
first such model was proposed by Wagner and Whitin [13] who examined a dynamic version of
the economic lot size model. Many authors extended this model; for example, Manne and
Veinott [11], Zangwill [16] and Love [8]. Zangwill used a network flow approach, and Love
generalized the model to piecewise concave cost functions and bounded idle capacities and
shortages.
Several models and algorithms for two facility problems have been developed. Manne
[10], Erlenkotter [1,2], Kalotay [5], and Fong and Rao [3] examined models in which it is
assumed that converted capacity is reconverted automatically, at no cost, at the end of each
period. Kalotay [6], Wilson and Kalotay [14], Merhaut [12], and Luss [9] examined models in
which converted capacity is not reconverted automatically at the end of each period.
In Section 1 we describe the generalized model. The algorithm in [9] is extended and
used to solve the new model with the additional features described before. In Section 2 a shor-
test path formulation is presented, and in Section 3 some properties of an optimal solution are
identified. These properties are used to compute the costs associated with the links of the net-
work constructed for the shortest path problem. In Section 4 the solution is illustrated by a
numerical example, and some final comments are given in Section 5.
1. THE MODEL
The model assumes a finite number of discrete time periods in which the demand incre-
ments, new constructions, capacity disposals, and capacity conversions occur instantaneously
and simultaneously immediately after the beginning of each period. We define the following
notation:
; — index for the two facilities.
/ — index for time periods (r = 1,2, . . , , T) where Tis the planning horizon.
r„ — the increment of demand for additional capacity of facility i incurred
immediately after the beginning of period t. The /-,-,' s may be negative, and
for convenience are assumed to be integers.
Ri(tut2) « £ r„, for r, < t2.
— the amount of new construction (x„ > 0), or capacity disposal (x„ < 0),
associated with facility i immediately after the beginning of period t.
CAPACITY EXPANSION WITH TWO FACILITY TYPES 599
y, — the amount of capacity converted immediately after the beginning of period
t. y, > 0 (y, < 0) implies that capacity associated with facility 1 (facility 2) is
converted to satisfy the demand of the other facility. Once converted, the
capacity becomes an integral part of the new facility.
/,, — the amount of idle capacity (/,-, > 0), or capacity shortage (/,, < 0), associ-
ated with facility /' at the beginning of period t (or equivalently, at the end of
period t — 1, / = 2,3, ... , T + 1). We assume that initially there is no idle
capacity or capacity shortage, that is, Iti = 0.
hi — lower bound on /,,, that is, the maximum capacity shortage of facility i
allowed at the beginning of period t ; the /,,'s are assumed to be integers and
- OO < /,., < 0.
wH — upper bound on the idle capacity of facility i at the beginning of period /.
The w,,'s are assumed to be integers and 0 < wtt < oo.
Ci,(xit) — the construction and disposal cost function for facility /'at time period t.
Si(y,) — the conversion cost function at time period t.
hit(Ii,t+\)— the cost function associated with idle capacity, or capacity shortage, of facil-
ity /'carried from period t to period t + 1.
All cost functions are assumed to be concave from 0 to °o and from 0 to — oo, but not
necessarily concave over the entire interval [-oo, ©o]. Such functions are called piecewise con-
cave functions, see Zangwill [15]. All cost functions are also assumed to be nondecreasing
away from zero; for example, c„(x„) is nondecreasing with xit for xit > 0, and nondecreasing
with — xit for xjt < 0. For convenience, we assume that c,,(0) = £,(0) = //,,(0) = 0.
The problem can be formulated as follows:
(1.1) Minimize £ £ cit 0c„) + hit (/, ,+1) I + g, (y,)
(1.2) /u+1 - /„ + x„ - yt - n, <
(1.3) /2,,+1 - I2l + x2l + y, - rlt
(1)
(1.4) Iu < In < "a
(1.5) 7(1 = 0
(1.6) 7,,7-+1 = 0
The objective (1.1) is to minimize the total cost incurred over all periods. Equations
(1.2) - (1.3) express the idle capacity or capacity shortage 7,,+i as a function of /,,, the actions
undertaken at period t, xit and v,, and the demand increments r„. Constraints (1.4) specify the
bounds on idle capacities and capacity shortages, and Equation (1.5) is introduced by assump-
tion. Constraint (1.6) 7, r+1 = 0 implies that idle capacity or capacity shortages are not allowed
after period T. Such a constraint is not restrictive since one can add to problem (1) a fictitious
/= 1,2, ... , T
i-l,2
600 H. LUSS
period T' = T + I with riT = max R,(X,t) - R,(l,T) (yielding R,(l,T') > U/(U) Vr),
/,r = 0,w,r = °°, and c,r(-) = hjr (•) = ?7-(0 = 0. (/,r is fixed at zero since no shortages are
allowed at the end of period T). This allows us to fix /,r+i at zero since then there always
exists an optimal solution with /,,r+i = 0. To simplify notation, we assume that period Tin
formulation (1) is the added fictitious period.
The constraints (1.2) - (1.6) form a nonempty convex set. Since each term of the objec-
tive function is nondecreasing away from zero with a finite value at zero, there exists a finite
optimal solution. Furthermore, suppose each of the variables x„, v,, and /„ is replaced in for-
mulation (1) by the difference of two nonnegative variables, for example, x„ = xl, — x ■',, where
x'u ^ 0 represents constructions and x"r ^ 0 stands for disposals. In that case, the objective
function becomes concave on the entire feasible region; hence, there exists an extreme point
optimal solution. From Pages 124-127 in Hu [4], the constraints (1.2) - (1.3) are totally uni-
modular. Thus, since rit, /„ and wu are assumed to be integers, such an extreme point solution
consists of integers. In the next sections we describe an algorithm which finds an optimal
extreme point solution.
2. A SHORTEST PATH FORMULATION
Since all cost functions are nondecreasing away from zero, it can be shown that there
exists an optimal solution in which
(2) |/„| < max[/?,(T1,r) + R2(t2,T)] = b Vi, t.
T,,T2
However, usually, better bounds than those given by (2) can be assigned. To simplify the
presentation, we assume that the lower and upper bounds on the /„ variables satisfy wit ^ b
and /„ ^ -b for all values of / and t.
Generalizing the concept of capacity point in [9], we define a capacity point as a period / in
which /„ = 0, or /„, or w„ for at least one value of i. Since an extreme point optimal solution
consists of integers, the set of capacity points is defined as follows:
(3.1) /„ = /21 = 0
(3.2) Iu = /if,0(wn and I2t = ht>0,w2t
(3)
(3.3) /„ = lu, 0,wh and I2l = l2t + 1, . . . , -1, 1, . . . , w2,,.
(3.4) I2t = l2„0,w2l and/1; = lu + 1, ... , -1,1, ... , wu.
t = 2,3, ... , T
(3.5) /, r+1 = /2 r+1 = 0.
The capacity point values can be conveniently specified by a single parameter a,. For
example, a, = 1,2, .... 9 can be used to specify the combinations given by (3.2), etc. A
complete example of a special case can be found in [9].
The set of capacity points can be limited to those satisfying
(4) (4.1) Iu + I2,< RiU,T)+ R2U,T)
(4.2) Iu + I2t> -max [/?,(t,,/- 1) + R2(t2,i- 1)].
CAPACITY EXPANSION WITH TWO FACILITY TYPES 601
Equation (4.1) states that the total idle capacity at the beginning of period /does not exceed the
cumulative demand from period t to T. Equation (4.2) restricts the maximum capacity shor-
tages to the maximum demand increase from any period prior to / - 1 up to period t — 1.
Clearly, there exists an optimal solution which satisfies (4).
We now describe a shortest path formulation which can be used to solve Problem (1).
Let
duv(au,av+l) — the minimal cost during periods u, u + \, . . . , \ associated with an
extreme point solution of (1) when u and v + 1 are two successive capacity
points with values defined by au and av+1. More specifically:
lii c,,(x,,) + hituu+A
(5) duw{au,av+\) = minimum ]£ £ cit(xu) + //„(/,, +1) -I- g,(y,)\
such that
(i) Constraints (1.2) and (1.3) are satisfied for t = u, u + 1, . . . , v,
(ii) /„ < /„ < wu and /„ ^ 0 for i = 1, 2 and t = u + 1, u + 2, ... , v,
(iii) I\u and I2u are defined by aw, and 7i,v+i and /2,v+i are defined by av+i>
(iv) xit and y, for t = u, u + 1, . . . , v satisfy the necessary conditions (to be developed
later) for an extreme point solution of (1).
Suppose that all subproblem values duv(au,av+\) are known. The optimal solution can then
be found by searching for the optimal sequence of capacity points and their associated values.
As shown in Figure 1, Problem 1 can be formulated as a shortest path problem for an acyclic
network in which the nodes represent all possible values of capacity points. Each node is
described by two values it.a,) where t is the time period and a, is the associated capacity point
value. From each node (u,au) there emanates a directed link to any node (v + l,av+i) for
v ^ u with an associated cost of duy/(au,av+i).
Let C, be the number of capacity point values at period /. Clearly, C\ = CT+\ = 1, and C,
for all other periods can be obtained from Equations (3) and (4). The total number of links N
in the shortest path problem is
T I T+l I
(6) N=± CA £ Cj\.
;=1 [j-i+l }
Since most of the computational effort is spent on computing the duv(au,av+\) values, it is
important to reduce N, if possible. One way, of course, is to reduce the values of C, through
the imposition of appropriate bounds /„ and w„ .
The shortest path problem can be solved using various algorithms. Since the network is
acyclic a simple dynamic programming formulation can be used. Let a, be described by the set
of integers 1,2, .... C,, where a, = 1 represents I\, = I2t = 0. Furthermore, let f,(a,) be the
cost of an optimal policy over periods /, t + l, ... , T, given that period t is a capacity point,
and that /t, and I2, are specified by a,. The following dynamic programming formulation is
then obtained:
/r+i(«r+i) = 0, ar+1 = 1
(7) fu(au) = min [^uv(au,av+1) + /v+1(av+1)],
K«V+)^"CV+1
u= T,T-\, ... , \
au= 1,2 Cu.
Figure 1. The shortest path foi
The first term of the minimand is the minimum cost of the optimal policy during periods «,
u + 1, .... v, given that u and v + 1 are two successive capacity points with values au and
av+1. The second term is the optimal cost for periods v + 1, v + 2, ..., T, given av+1.
3. SOLUTION OF THE SUBPROBLEMS duMu.<*r+0
Most of the computational effort is spent on computing the subproblem values. As shown
in [9], when r„ ^ 0, x„ ^ 0, /,, = 0 and wit = oo for all /and /, the subproblems are solved in a
trivial manner, however, when the r„'s are allowed to be negative the effort required to solve
the subproblems increases significantly. The additional modifications needed to solve the sub-
problems duv(au,ay+i), as defined by (5) for the generalized model, require a more careful
analysis than needed in [9], however, the resulting computational effort appears to be about the
To compute the subproblem values duv(au,av+l), it is convenient to describe Problem (1)
as a single commodity network problem. The network, shown in Figure 2, includes a single
source (node 0) with a supply of /?,(1,D + R2(l,T). There are 2T additional nodes, each
denoted by (/,/) where /specifies the facility and /specifies the time period. At each node (/,/)
there is an external demand increment /•„, possibly negative. The nodes are connected by links,
where the flows along these links represent the constructions, disposals, conversions, idle capa-
cities, and capacity shortages. The flows on each link can be in either direction, and the link
direction in Figure 2 indicates positive flows. The nodes are connected by the following links:
— A link from node 0 to each node (/,/) with flow x„. xit is positive if the flow is from
node 0 to node (/,/), and negative otherwise.
CAPACITY EXPANSION WITH TWO FACILITY TYPES
603
Rgd.T)
Figure 2. A network flow representation of the capacity expansion problem
— A link from each node (i,t) to node (i,t + 1) with flow Iu+\. Iu+\ is positive if the
flow is from (i,t) to (i,t + 1) and negative otherwise.
— A link from each node (1,/) to node (2,t) with flow y,. y, is positive if the flow is
from node (l,f) to (2,/), and negative otherwise.
As discussed before, we are interested in finding an optimal extreme point solution to a
modified version of Problem (1), in which each of the variables xit, yt, /„ is replaced by the
difference of two nonnegative variables. It can be shown that a feasible flow in the network
given in Figure 2 corresponds to an extreme point solution of Problem (1) modified as
described above, if and only if it does not contain any loop with nonzero flows in which all /„
flows satisfy /„ < Iit < wit and Iit ^ 0.
Concentrating upon a single subproblem, as shown in Figure 3, one may observe that a
feasible flow does not contain such loops if and only if the following properties are satisfied:
(8.1) xitixil2=0Ui* t2),
(8) (8.2) V,2=0(^r2),
(8.3) xlrix2,2^3=0.
i = l,2
u < fi, t2, h ^ v
For example, suppose (8.3) is violated and tx < t2 < t3, then xlri, /i,,1+i, . . . , /i,3, ytj, /2/3,
h,t3-\> ■•• • h,t2+\> xit form a loop with nonzero flows and all relevant /„ values satisfy
/„ < Iit < wu and 4 J* 0.
Equation (8) implies that in the optimal solution of duv(au,aw+i) there is at most one new
construction or disposal for each facility (8.1), and at most one conversion (8.2). Furthermore,
if two constructions or disposals (one per facility) are being considered, conversion is then not
allowed (8.3).
Figure 3. A network flow representation of a subproblem
Let Dj be the capacity change of facility /' during periods u,u + 1, . .. , v, that is:
(9) A = /,, v+1 + /?, (u, v) - /,„, i - 1,2
or, equivalents
(10) A = ± xu - y,
D2 = ±x2t+yt.
t—u
Let t\ and t2 be two time periods u < U\,t2) < v. From the optimal properties (8) shown
above, the possible policies associated with an optimal solution to any subproblem duv(au,av+{)
can be restricted to three different policies. These policies are summarized in Table 1 below.
To illustrate the table, let us concentrate on the column A ^ 0 and D2 ^ 0. Policy (a)
indicates a single disposal of A capacity units of facility 1, and a single construction of D2 units
of facility 2. Policy (b) implies a single construction of A + D2 of facility 1 if Z>! + Z)2 ^ 0, a
single disposal of A + -A of facility 1 if Dx + D2 ^ 0, and a single conversion of D2 units
from facility 1 to facility 2. Obviously, if D^ + D2= 0, no constructions or disposals take
place, and if D2 = 0, no capacity is converted. Finally, policy (c) consists of a single construc-
tion of A + Di capacity units of facility 2 if Z>i + Z)2 > 0, a single disposal of A + A units
of facility 2 if A + &i ^ 0> and a single conversion of -A from facility 1 to facility 2.
The optimal solution of a subproblem duy{auiay+x) is therefore obtained by the following
procedure:
(1) For each of the policies (a), (b), and (c) in Table 1, find the optimal values of t\ and
t2y which minimize duv(au,a^.\) as given by Equation (5), while satisfying condi-
tions (i) - (iv) given below Equation (5). If no feasible values of t^ and t2 exist, set
the value of the corresponding policy to «>.
CAPACITY EXPANSION WITH TWO FACILITY TYPES
TABLE 1. Possible Policies for Optimal Subproblem Solutions
605
— ^_01.02
D, ^ 0
Z>, < o
Z>, > 0
Z>, < 0
Policy
d2 ^ o
Z)2 > 0
Z)2 < 0
£>2 < 0
x,M = Du xu - 0 t 7* f,
construction
disposal
construction
disposal
(a)
*2,2 = D2, x2l = 0 t * t2
y, = 0 Vr
construction
construction
disposal
disposal
xUj = Z), + D2, xu - 0 f ?* fj
construction
construction
or disposal
construction
or disposal
disposal
(b)
y,; =o2,j, = 0/^/2
conversion
conversion
conversion
conversion
from 1 to 2
from 1 to 2
from 2 to 1
from 2 to 1
x2, = 0V/
*2,, - X>1 + ^2- X2, = 0 ^ r.
construction
construction
or disposal
construction
or disposal
disposal
(c)
jte-4/)!, y, = o^/2
conversion
conversion
conversion
conversion
from 2 to 1
from 1 to 2
from 2 to 1
from 1 to 2
x„ = OVf
(2) Choose as the optimal policy the best of those found in Step (1). If none of the policies
is feasible, duv(au,av+]) = «>.
The procedure above may involve spending a significant amount of computation on
finding all feasible policies and comparing the costs associated with these policies.
4. A NUMERICAL EXAMPLE
As an illustration, we solve the capacity expansion problem shown in Figure 4.
^14 = ^24 = 0 by assumption, thus, a fictitious period is not added. The cost functions are given
in Table 2 below.
The shortest path formulation is shown in Figure 5. The capacity point values are given
inside the nodes in terms of lu and I2, rather than a,. Using Equation (4.1), several capacity
point values are omitted in periods 2 and 3. Furthermore, all links from period t = 1 to
periods t = 3 and 4 are omitted since there is no feasible solution to the associated subprob-
lems with lu < hi < wu and hi ^ 0. The number associated with each link is the optimal
solution of the corresponding subproblem. The shortest path is marked by stars.
Consider the subproblem dn{a\, a2) where a\ represents the capacity point value
hi = hi = 0, and a2 represents I2\ = hi = 0- By Equation (9), Dx = 1 and D2 = 1. Using
the results of Table 1, policy (a) yields xn = x2i = 1 with a total cost of 68, policy (b) yields
xn = 2 and y\= I with a cost of 46, and policy (c) yields x2i = 2 and y\ = — 1 with a cost of
45. Hence policy (c) is the optimal one.
To illustrate further, consider d23(a2,a4), where a2 stands for 712 = -1 and 722 = 0, and
a4 stands for 714 = 724 = 0, so that D^ = 1 and D2 = 0. From Table 1, policy (a) implies that
either xu = 1 or xJ3 = 1. However, if x13 = 1 (and x12 = 0) then 713 = 0 so that d23(-) = «>.
Hence, policy (a) implies x12 = 1 with construction and holding cost of 43.2. Policy (b) yields
the same solution as policy (a), and policy (c) results in x22 = 1 and y2 = -I with a total cost
of 40.5; hence, policy (c) is optimal for that subproblem.
606
Figure 4. A network flow representation of the example
TABLE 2 - The Cost Functions
^\function
argument^\
c\,(xh)
c2l(x2t)
MW
/= 1,2
g,(yt)
positive
zero
negative
(30 + 8 xu)0.9'-1
0
6 • 0.9'"1
(20 + 10x2,)0.9'"1
0
5 • 0.9'" '
5Iu+l0.9'-1
0
-l0Iiit+l0.9'-{
0
0
-5^,0.9'-'
CAPACITY EXPANSION WITH TWO FACILITY TYPES
607
Figure 5. The shortest path problem for the example
Finally, consider tf23(a2,a4) with a2 standing for 712 = /22 = 0, and a4 standing for
A 4 = ^24 = 0- From Table 1, since Dx = D2 = 0, all decision variables are zero in all three pol-
icies and the total costs incurred are equal to 9.
After solving the subproblems for all the links of Figure 5, the shortest path can be found
using the dynamic programming formulation (7) or any other shortest path algorithm. The
shortest path in this example is 54 and consists of two links. The first link connects node
hi = hi = 0 to node hi = hi= 0> and the second link connects node l\i = /22 = 0 to node
A4 = -^24= 0- The optimal policy of the entire problem is x2i — 2, y\ =* — 1, with all other
decision variables xit and y, being equal to zero.
5. FINAL COMMENTS
This paper generalizes our previous work [9] by allowing for capacity disposals and capa-
city shortages. Furthermore, bounds on idle capacities and capacity shortages can be imposed.
The model is formulated as a shortest path problem in which most of the computational effort
is spent on computing the link costs. Using a network flow approach, properties of extreme
point solutions are identified. These properties are used to develop an efficient search for the
link costs.
Further generalizations may include bounds on new constructions and capacity disposals,
and operating costs which depend on the facility type and time period. As shown by several
authors, for example Lambrecht and Vander Eecken [7], bounded constructions or disposals
complicate considerably even the single facility problem. Introducing operating costs may
require major changes in the algorithm since the amount of each capacity type used to satisfy
the demand in each period affects the total cost.
608 H LUSS
Finally, negative costs for disposals (credit for salvage value) can be incorporated for cer-
tain cost functions c„(x„) for which the optimal solution would be finite. For example, cost
functions in which the credit per unit of disposed capacity is always smaller than the construc-
tion cost per unit of capacity. In general, however, cost functions c„(x„) that are negative for
xit < 0 may result in an unbounded solution.
REFERENCES
[1] Erlenkotter, D., "Two Producing Areas— Dynamic Programming Solutions," Investments for
Capacity Expansion: Size, Location, and Time Phasing, 210-227, A. S. Manne, Editor,
(MIT Press, Cambridge, Massachusetts, 1967).
[2] Erlenkotter, D., "A Dynamic Programming Approach to Capacity Expansion with Speciali-
zation," Management Science, 21, 360-362 (1974).
[3] Fong, CO., and M.R. Rao, "Capacity Expansion with Two Producing Regions and Con-
cave Costs," Management Science, 22, 331-339 (1975).
[4] Hu, T.C., Integer Programming and Network Flows, 124-127, (Addison Wesley, Reading,
Massachusetts, 1969).
[5] Kalotay, A.J., "Capacity Expansion and Specialization," Management Science, 20, 56-64
(1973).
[6] Kalotay, A.J., "Joint Capacity Expansion without Rearrangement," Operational Research
Quarterly, 26, 649-658 (1975).
[7] Lambrecht, M. and J. Vander Eecken, "Capacity Constrained Single Facility Lot Size Prob-
lem," European Journal of Operational Research, 2, 132-136 (1978).
[8] Love, S.F., "Bounded Production and Inventory Models with Piecewise Concave Costs,"
Management Science, 20, 313-318 (1973).
19] Luss, H., "A Capacity-Expansion Model for Two Facilities," Naval Research Logistics
Quarterly, 26, 291-303 (1979).
[10] Manne, A.S., "Two Producing Areas— Constant Cycle Time Policies," Investments for Capa-
city Expansion: Size, Location, and Time Phasing, 193-209, A.S. Manne, Editor, (MIT
Press, Cambridge, Massachusetts, 1967).
[11] Manne, A.S. and A.F. Veinott, Jr., "Optimal Plant Size with Arbitrary Increasing Time
Paths of Demand," Investments for Capacity Expansion: Size, Location, and Time Phasing,
178-190, A.S. Manne, Editor, (MIT Press, Cambridge Massachusetts, 1967).
[12] Merhaut, J.M., "A Dynamic Programming Approach to Joint Capacity Expansion without
Rearrangement," M. Sc. Thesis, Graduate School of Management, University of Cali-
fornia, Los Angeles, California (1975).
[13] Wagner, H.M. and T.M. Whitin, "Dynamic Version of the Economic Lot Size Model,"
Management Science, 5, 89-96 (1958).
[14] Wilson, L.O., and A.J. Kalotay, "Alternating Policies for Nonrearrangeable Networks,"
INFOR, 14, 193-211 (1976).
[15] Zangwill, W.I., "The Piecewise Concave Function," Management Science, 13, 900-912
(1967).
[16] Zangwill, W.I., "A Backlogging Model and a Multiechelon Model for a Dynamic Economic
Lot Size Production System— A Network Approach," Management Science, 15, 506-527
(1969).
SOLVING MULTIFACILITY LOCATION PROBLEMS
INVOLVING EUCLIDEAN DISTANCES*
Department of Systems Design
University of Waterloo
Waterloo, Ontario, Canada
Christakis Charalambous
Department of Electrical Engineering
Concordia University
Montreal, Quebec, Canada
ABSTRACT
This paper considers the problem of locating multiple new facilities in order
to minimize a total cost function consisting of the sum of weighted Euclidean
distances among the new facilities and between the new and existing facilities,
the locations of which are known. A new procedure is derived from a set of
results pertaining to necessary conditions for a minimum of the objective func-
tion. The results from a number of sample problems which have been exe-
cuted on a programmed version of this algorithm are used to illustrate the
effectiveness of the new technique.
1. BACKGROUND
It was as early as the 17th century that mathematicians, notably Fermat, were concerned
with what are now known as single facility location problems. However, it was not until the
20th century that normative approaches to solving symbolic models of these and related prob-
lems were addressed in the literature. Each of these solution techniques concerned themselves
with determining the location of a new facility, or new facilities, with respect to the location of
existing facilities so as to minimize a cost function based on a weighted interfacility distance
measure.
If one studies a list of references to the work done in the past decade involving facility
location problems it becomes readily apparent that there exists a strong interdisciplinary interest
in this area within the fields of operations research, management science, logistics, economics,
urban planning and engineering. As a result, the term "facility" has taken on a very broad con-
notation in order to suit applications in each of these areas. Francis and Goldstein [4] provide a
fairly recent bibliography of the facility location literature. One of the most complete
classifications of these problems is provided in a book by Francis and White [5].
'This work was supported by the National Research Council of Canada under Grant A4414 and by an Ontario Gradu-
ate Scholarship awarded to Paul Calamai.
609
610 P. CALAMAI ANDC. CHARALAMBOUS
This paper concerns itself with the development of an algorithm for solving one particular
problem in the area of facility location research. The problem involves multiple new facilities
whose locations, the decision variables, are points in E2 space. The quantitative objective is to
minimize the total cost function consisting of the sum of weighted Euclidean distances among
new facilities and between new and existing facilities. The weights are the constants of propor-
tionality relating the distance travelled to the costs incurred. It is assumed that the problem is
"well structured" [3].
The Euclidean distance problem for the case of single new facilities was addressed by
Weiszfeld [13], Miehle [10], Kuhn and Kuenne [8], and Cooper [1] to name a few. However,
it was not until the work of Kuhn [7] that the problem was considered completely solved. A
computational procedure for minimizing the Euclidean multifacility problem was presented by
Vergin and Rogers [12] in 1967; however, their techniques sometimes give suboptimum solu-
tions. Two years later, Love [9] gave a scheme for solving this problem which makes use of
convex programming and penalty function techniques. One advantage to this approach is that it
considers the existence of various types of spatial constraints. In 1973 Eyster, White and
Wierwille [2] presented the hyperboloid approximation procedure (HAP) for both rectilinear
and Euclidean distance measures which extended the technique employed in solving the single
facility problem to the multifacility case. This paper presents a new technique for solving con-
tinuous unconstrained multifacility location problems involving Euclidean distances.
2. PROBLEM FORMULATION
The continuous unconstrained multifacility location problem involving the lp distance
measure can be stated as follows:
Find the point X*T = (X*{, ... , X*„r)in E2n to
(PI) minimize f(X) = £ vJk \ \Xj - Xk \ \p + £ £ wM \ \Xj - At \ \p
\Hj<kHn 7-1 /-l
where
n A number of new facilities (NF's) .
m A number of existing facilities (EFs).
X] ' = (Xji Xji) A vector location of NFj in E2, j — 1, • • • , n.
Aj — (an ai2) A vector location of EFt in E2, i = 1, ,. . . , m.
\jk A nonnegative constant of proportionality relating the lp distance between NFj and NFk
to the cost incurred 1 < j < k < n.
Wjj A nonnegative constant of proportionality relating the lp distance between NFj and EFt to
the cost incurred 1 < j < n, 1 < / < m.
\\Xj - Xk\\p = {\xn - xkl\P+ \xJ2 - xk2Wlp A lp distance between NF, and NFk.
\\Xj - A,\\p - {\xji - an\" + \xJ2 - ai2\pYlpb lp distance between NF, and EF,.
Note that we make the assumption that v^ = vkJ for j,k = I, ... , n. Substituting p = 1
and p = 2 in Problem PI respectively yields the rectilinear distance problem and the Euclidean
SOLVING MULTIFACILITY LOCATION PROBLEMS 611
For the purpose of this paper Euclidean distance will be the measure used between facili-
ties located as points in E2 space. The objective function becomes
minimize f(X) = £ vJk {(xyl - xkl)2 + (xj2 - x^2)2)1/2
X \^j<k^n
(P2) + £ £ wj, {(*,, - an)2 + (xJ2 - a/2)2}1/2.
7=1 i-1
The techniques presented in this paper can also be used for problems involving facilities located
in three-dimensional space.
3. NEW FACILITY CATEGORIZATION
If we consider a current solution to Problem P2 we can think of each new facility as being
in one of the following distinct categories:
(1) Unique Point (UP)
A new facility in this category occupies a location that differs from all other facility
locations.
(2) Coinciding Point (CP)
A new facility in this category occupies a location that coincides with the location of
an existing facility but differs from the current locations of all other new facilities.
Thus, each new facility in this category has associated with it some existing facility
which has the same vector location.
(3) Unique Clusters ( UCX , ..., UCNUC)
All new facilities in the /cth unique cluster (k = 1, . . . , NUC) occupy the same vec-
tor location. This location is distinct from all existing facility locations as well as the
current locations of new facilities that are not classified in this cluster.
(4) Coinciding Clusters (CC\, ... , CCNCc)
All new facilities categorized in the /cth coinciding cluster (k = 1, ... , NCC) occupy
the same vector location. This location coincides with the location of some existing
facility and differs from the current locations of all new facilities that are not
classified in this cluster. Each of these coinciding clusters of new facilities is there-
fore associated with some existing facility with which it shares a location.
If we define the index sets J A {1, . . . , n) and / A {1, . . . , m) and let the subsets
UC0 = CC0 = 0 then the categorization can be restated as follows:
Partition the set J into the subsets UP, CP, UC\ UCNUC, CC\, ... , CCNCc where
(3.1) UP = {V7. € j\Ai * Xj * Xk; X/i € /, Vk € J - {j})
(3.2) CP = {\/j € j\Ar = Xj * Xk; ij e I, Vk £ J - {j}}
for a = 1, .... NUC
(3.3)
UCa A Vy 6 J - U UC,\A, * Xj = Xk; V; € /, k €7 - {j} - U UC,\
612 P CALAMAI ANDC. CHARALAMBOUS
for/3= 1, .... NCC
(3.4) CCP A I V, € / - V Cq\Aip = ^ = AT*; fc € /, Jfc € / - {j} - V Cci
NUC A number of unique clusters.
A^CC A number of coinciding clusters.
Note that
(a) New facility j coincides with existing facility ij for j € CP (from 3.2).
(b) The new facilities in cluster /3 coincide with existing facility ip for /3 = 1, . . . , A^CC
(from 3.4).
In order to use this notation for the derivation of the new algorithm given in the next sec-
tion define a unit vector D in E2n as follows:
DT={D{, .... AH
where
(3.5) Dj= [dji dj2], 7=1, •••- n
and
ILDll2=l.
4. THE DIRECTIONAL DERIVATIVE
Using the notation given in the last section we can write the directional derivative of the
objective function at X'm the direction D in the following useful manner:
= I [G>--I>y]
7 6 W
,/eCP
NUC
+ 11 [^•^+ I v.JlA- Dk\\2]
a=\ \j£UCa k^UCa
NCC .
(4.1) +11 \Gj-Dj+ £ VjJlDy-Alla+HUlDyl
where
(4.2a) G, = £ | fy 7 y | + I I ' J ., V, 6 W>
(42b) °'-£ife^t+£Tfe^t V^€CP
SOLVING MULTIFACILITY LOCATION PROBLEMS 613
<4-2c> GJ-Z ifr.ly.lL + X.
vjk{Xj-Xk) | ^ w7,(A}-^)
v, € f/Ca
k<lUCn
Xj-Xk\\2 +f?, WXj-AMi «-i. ■■■■ ^c
It should be noted that in each case, the expression for Gj is the gradient of that part of
the objective function f{X), which is differentiate with respect to Xj. In the case where
j € UP, the expression is the exact gradient with respect to Xj\ in all other cases, the expres-
sion for Gj can be considered a pseudo-gradient of f(X) with respect to Xj.
Since f(X) is a convex function, the point X* in E2n is a minimum for this function if
and only if the directional derivative dDf{X*) is nonnegative for all unit vectors D in Eln. This
fact will be used in the next section.
5. NECESSARY CONDITIONS FOR OPTIMALITY
THEOREM 1: If the following conditions are not satisfied at the point X in E2n, then the
directional derivative given by expression (4.1) will be negative for some unit vector D in E2n.
(5.1) (1) ||G,||2-0 V7 € UP
(5.2) (2) ||G,||2< wj,j Vj € CP
(5.3) (3) fora= 1, ... , NUC
HZ GA< £ E v;* vsc t/ca
7€5 ,/€S k£[UCa-S]
(5.4) (4) for/i
-M*6[cc„-n p
PROOF: The proofs for conditions 1 and 2 are obvious.
I Dj = R for j 6 iS
then </fl/(Ar) =£(?,-•/? +£ £ v;j|/?||2
y€5 y€5 /c€[f/C0 - S]
-H/?ll2lllZG/ll2cose+2: I V
I yes yes /ce[t/ca-s]
Therefore, rfD/U) )0 VD only if
III Gj\\2< I I ¥;» V5C UCa.
j£S j£S k€[UCa-S]
The proof for condition 4 is similar.
614 P. CALAMAI AND C CHARALAMBOUS
6. UPDATE FORMULAS
As a result of the preceeding optimality conditions the following update formulas are con-
structed:
CASE 1: If 3/ € UP such that 11(7,11 ^ 0 then the direction of steepest-ascent in the
subspace defined by X, is Gj = Gr We therefore use the following update formula for Xy
Xj * — Xj — Kj Gj
where
(6.1)
= y v* + y ZS.
CASE 2: If 3/ 6 CP such that llGylb > wjt then the direction of steepest-ascent in the
subspace defined by Xj is Gj = Gj. We therefore use the following update formula for Xy.
Xj — Xj - \j Gj
where
(6.2)
+ 1-
fe \\Xj-xk\\2 fe \\Xj-AMt
CASE 3: If 35 C UCa,a = 1,
NUC, such that
IIIG,||2>I £ v„
/€S y€5 *€[{/CQ - 5)
then the direction of steepest-ascent in the subspace defined by the subset cluster is
Gs = £ Gj. We therefore use the following update formula:
Jts
V, € 5 J; — X,-ksGs
where
(6.3)
y na + y zt
CASE 4: If 37C CQ, 0 = 1, . . . , A^CC, such .that
III^||2> l[| I vJ + mJ
jZT j€T \\k€lCCrT) P]
then the direction of steepest-ascent in the subspace defined by the subset cluster is GT =
£ Gj. We therefore use the following update formula:
Xj <— Xj - XT GT
SOLVING MULTIFACILITY LOCATION PROBLEMS 615
where
(6.4)
^tJcc, WXj-XkWi +k, \\Zj-4\\2
In each result, the expression for lambda (A) can be considered a weighted harmonic
mean [8] of the interfacility distance terms appearing in the equation for the gradient (Case 1)
or pseudo-gradients (Cases 2 through 4) .
7. A NEW ALGORITHM
Using the results derived in the preceeding section the following algorithm can be used to
solve Problem P2:
(1) Find a current solution X in E2„-
(2) Try to obtain a better solution by moving single new facilities by using Cases 1 and 2.
(3) For a = 1, .... NUC try to obtain a better solution by applying the special form of
Case 3 where \S\ = 1 (to move single new facilities) or, if this fails, applying the
special form of Case 3 where |S| = I UCa\ (to move entire clusters of new facilities).
If successful, return to Step 2.
(4) For p = 1, .... NCC try to obtain a better solution by applying the special form of
Case 4 where |r| = 1 (to move single new facilities) or, if this fails, applying the
special form of Case 4 where I T\ — |CCp| (to move entire clusters of new facilities).
If successful, return to Step 2.
(5) Try to obtain a better solution by moving subset clusters using Cases 3 and 4. If an
improvement is made, return to Step 2.
8. REMARKS ON IMPLEMENTATION
The following rules were used in implementing the algorithm described in the last section:
(a) New facility j and new facility k were considered "clustered" if:
(8.1a) ||*,.||2+|l*JI2<«i Kj<k^n
or
where e i A inputted cluster tolerance,
(b) New facility j and existing facility / were considered "coinciding" if:
(8.2a) ll*,ll2+IU/[b<€, j- 1 n; ir 1 m
or
<8-2b) iiiiif+liln! K «• J-1 n;'-1 m
616 P. CALAMAI AND C CHARALAMBOUS
where e i A inputted cluster tolerance,
(c) The update formulas were used only if:
(8.3) a||G||2>€2
where €2 A inputted step tolerance. This helped avoid the possibility of repeatedly taking small
steps. However, the step tolerance is reduced prior to the termination of the algorithm as out-
lined by the next rule.
(d) In order to ensure optimality, the following check is made prior to executing Step 5 of the
algorithm:
(8.4) \f(X(h'l)) - f(Xu'})] * 100 < e3 * f(X{"-l))
where €3 A inputted function tolerance.
If this condition is not satisfied, the step tolerance (e2) is reduced and the algorithm res-
tarted at Step 2.
9. DISCUSSION
The new algorithm has the following properties:
(a) It makes full use of the structure of the facility location problem thus avoiding the
need for any background in related nonlinear programming areas.
(b) The actual objective function, and not an approximation to it, is minimized at each
step in the algorithm.
(c) The stepsize used in this algorithm may not be "optimal" when compared with step-
sizes obtained from line-search techniques. However, the use of this stepsize has the
following advantages: a) ease of computation, b) maintenance of location problem
structure, and c) reduced computation time per update.
(d) Although Step 5 in the algorithm is combinatorial in complexity, very little computa-
tional work is necessary. This is a result of the fact that all the information needed
for this step has already been computed and stored in previous steps.
(e) The algorithm is similar to the technique devised by Kuhn for solving the single-
facility location problems with Euclidean distances [7] and the method devised by
Juel and Love [6] for the multifacility location problem with rectilinear distances.
This makes the algorithm attractive to those with experience with these methods.
(f) Currently, there is no rigorous proof that this algorithm converges. In 1973, Kuhn
[7] completed the proof of convergence for a similar scheme, introduced by
Weiszfeld [13] in 1937, for the case of single new facilities. Based on computational
experience and on the fact that the algorithm is designed to minimize the objective
function in all new facility subspaces, it is likely that the algorithm always converges.
(g) The main disadvantage of the algorithm is that the order in which each of the sub-
spaces is checked is, currently, not optimal. A method, based on projections, that
would allow us to determine "a priori" which subspace to update, is now being inves-
tigated.
SOLVING MULTIFACILITY LOCATION PROBLEMS 617
(h) Most existing methods for solving the multifacility problem lack any consideration of
the existence of constraints on the solution space [9]. This is also true of the new
method outlined in this paper; however, the addition of constraints should not
present a problem to the projection technique.
(i) It has yet to be proven that the necessary conditions for optimality for Problem P2,
given by Equations (5.1) through (5.4), are also sufficient.
10. COMPUTATIONAL EXPERIENCE
The performance of the algorithm described in this paper (MFLPV1) was tested against
the hyperboloid approximation procedure (HAP) described in Eyster, White and Wierwille [2]
and a modified hyperboloid approximation procedure (MHAP) suggested by Ostresh [11].
Two parameters were used as a basis of comparison: 1) the number of new facility loca-
tion updates needed to reach optimality, and 2) the required CPU time in minutes. In the case
of program MFLPV1, two counts were considered necessary for specifying the first parameter.
The first count represented the number of "attempted" updates (excluding those updates from
Step 5 of the algorithm). The second count represented the number of "successful" updates.
The reason for excluding the number of attempted updates from Step 5 of the algorithm is sim-
ply this: computationally, very little work is done at this step in the procedure.
Six problems were used for the comparison; the first three were taken from [5] (#5.23,
#5.7 and #5.6 respectively), the fourth appears in [2] and the last two problems summarized in
Tables 1 and 2, are the authors.
HAP and MHAP were both executed using two different initial hyperbolic constants e<0)
for these problems in order to emphasize the significance of this parameter to the performance
of these algorithms. The stopping criteria used in each case was the same as that outlined in
the paper introducing HAP [2]. Unless otherwise specified, program MFLPV1 also made use of
the following data.
(1) «i A cluster tolerance = 0.01 (from Equations (8.1) and (8.2)).
(2) e2 A step tolerance — 0.05 (from Equation (8.3)).
(3) e3 A function tolerance = 0.01 (from Equation (8.4)).
The results of these tests are summarized in Table 3. The numbers in this table represent
the total new facility updates required to reach optimality. The numbers in brackets ( ), under
the column headed MFLPV1, represent the number of successful updates whereas the unbrack-
eted numbers in these columns represent the number of attempted updates. The following
observations and comments can be made about the results summarized in this table:
(a) In all but Problem 5, the number of attempted updates required to reach optimality
using MFLPV1 is less than the number of updates required by HAP and MHAP.
These numbers are directly comparable.
(b) The new procedure (MFLPV1) used considerably less CPU time in solving the six
problems than did HAP and MHAP.
618
P. CALAMAI AND C. CHARALAMBOUS
TABLE 1 — Input Parameters for Problem 5
i
«/i
an
!
0.0
0.0
2
2.0
4.0
3
6.0
2.0
4
6.0
10.0
5
8.0
8.0
J
v<P)
^0)
1
0.0
0.0
2
0.0
0.0
3
6.0
10.0
4
1.0
3.0
5
6.0
10.0
6
8.0
8.0
7
2.0
4.0
8
2.0
4.0
9
6.0
10.0
(a) EF Locations
(b) Initial NF Locations
\ i
12 3 4 5
J \
1
1.0 1.0 1.0 1.0 1.0
2
1.0 1.0 1.0 1.0 1.0
3
1.0 1.0 1.0 1.0 1.0
4
1.0 1.0 1.0 1.0 1.0
5
1.0 1.0 1.0 1.0 1.0
6
1.0 1.0 1.0 1.0 1.0
7
1.0 1.0 1.0 1.0 1.0
8
1.0 1.0 1.0 1.0 1.0
9
1.0 1.0 1.0 1.0 1.0
s *
1 2 3
4 5 6 7 8 9
y\
1
Xj.o 1.0
1.0 1.0 1.0 1.0 1.0 1.0
2
^vl.O
1.0 1.0 1.0 1.0 1.0 1.0
3
1.0 1.0 1.0 1.0 1.0 1.0
4
V 1.0 1.0 1.0 1.0 1.0
5
\ 1.0 1.0 1.0 1.0
6
N. 1.0 1.0 1.0
7
Nv l.o l.o
8
>v 1.0
9
(c) Wji Weights
(d) v,* Weights
TABLE 2 — Input Parameters for Problem 6
*j\0) Xjf
i
at\
aa
1
2.0
5.0
2
10.0
20.0
3
10.0
10.0
(a) EF Locations
5.0
5.0
15.0
15.0
(b) Initial NF Locations
1
2
0.16 0.56 0.16
0.16 0.56 0.16
1 Sv 1.5
2 1 \
(c) Wjj Weights
(d) vjk Weights
SOLVING MULT1FACILITY LOCATION PROBLEMS
TABLE 3 — Comparative Test Results for Six Problems
#
MFLPV1
e(o)= 10o
6(0)= 1Q-4
X*
fix*)
HAP
MHAP
HAP
MHAP
1
564 (77)
1661
1381
2027
1407
(1.0,0.0)
(1.0,0.0)
(1.0,0.0)
(2.0,0.0)
(2.0,0.0)
38.990
2
148 (34)
647
546
4641
2281
(10.0,20.0)
(10.0,20.0)
186.798
3
63 (16)
87
70
770
197
(8.0,7.0)
(8.0,7.0)
43.351
4
31 (15)
45
45
45
45
(2.832,2.692
(5.096,6.351)
67.250
5
223 (40)
142
114
1763
975
(4.045,4.281)
(4.045,4.281)
(4.045,4.281)
(4.045,4.281)
(4.045,4.281)
(4.045,4.281)
(4.045,4.281)
(4.045,4.281)
(4.045,4.281)
201.878
6
63 (7)
242
164
3743
1869
(10.0,20.0)
(10.0,20.0)
8.540
TOTAL
1092 (189)
2824
2320
12989
6774
CPU
0.07
0.45
0.50
1.88
1.48
(c) Five of the six problems have solutions at cluster points. This appears to be the case
in many other problems. This suggests that methods using clustering information,
such as MFLPV1, will perform better than methods that disregard this information.
10. CONCLUDING REMARKS
To date, many of the methods designed for solving the multifacility location problem have
been either poorly structured, suboptimal or haphazard. In this paper, a new method is
developed for solving the multifacility location problem involving Euclidean distances. This
new method can easily be extended to accommodate problems involving item movements that
are other than Euclidean. Computational experience shows that this method outperforms tech-
niques currently in use. In addition, the proposed method takes full advantage of the structure
of the location problem.
Most current techniques used for solving location problems, including those proposed in
this paper, are designed to minimize an unconstrained objective function. This is an incom-
plete treatment since most practical problems involve some form of spatial constraints. It is
620 P CALAMAI AND C. CHARALAMBOUS
proposed that these constraints be handled and the performance of the algorithm improved
through the use of projection techniques. This approach is currently being investigated by the
authors.
BIBLIOGRAPHY
[1] Cooper, L., "Location- Allocation Problems," Operations Research, 77, 331-344 (1963).
[2] Eyster, J.W., J. A. White and W.W. Wierwille, "On Solving Multifacility Location Problems
Using a Hyperboloid Approximation Procedure," American Institute of Industrial
Engineers Transactions, 5, 1-6 (1973).
[3] Francis, R.L. and A.V. Cabot, "Properties of a Multifacility Location Problem Involving
Euclidean Distances," Naval Research Logistics Quarterly, 79, 335-353 (1972).
[4] Francis, R.L. and J.M. Goldstein, "Location Theory: A Selective Bibliography," Operations
Research, 22, 400-410 (1974).
[5] Francis, R.L. and J. A. White, "Facility Layout and Location: An Analytic Approach"
Prentice-Hall, Englewood Cliffs, New Jersey (1974).
[6] Juel, H. and R.F. Love, "An Efficient Computational Procedure for Solving the Multi-
Facility Rectilinear Facilities Location Problem," Operational Research Quarterly, 27,
697-703 (1976).
[7] Kuhn, H.W., "A Note on Fermat's Problem," Mathematical Programming, 4, 98-107
(1973).
[8] Kuhn, H.W. and R.E. Kuenne, "An Efficient Algorithm for the Numerical Solution of the
Generalized Weber Problem in Spatial Economics," Journal of Regional Science, 4, 21-
33 (1962).
[9] Love, R.F., "Locating Facilities in Three-Dimensional Space by Convex Programming,"
Naval Research Logistics Quarterly, 76, 503-516 (1969).
[10] Miehle, W., "Link-Length Minimization in Networks," Operations Research, 6, 232-243
(1958).
[11] Ostresh, L.M., "The Multifacility Location Problem: Applications and Descent Theorems,"
Journal of Regional Science, 17, 409-419 (1977).
[12] Vergin, R.C. and J.D. Rogers, "An Algorithm and Computational Procedure for Locating
Economic Activities," Management Science, 13, 240-254 (1967).
[13] Weiszfeld, E. "Sur le Point pour Lequel la Somme des Distances de n Points Donnes Est
Minimum," Tohoku Mathematical Journal, 43, 355-386 (1936).
AN EASY SOLUTION FOR A SPECIAL
CLASS OF FIXED CHARGE PROBLEMS
Patrick G. McKeown
College of Business Administration
University of Georgia
Athens, Georgia
Prabhakant Sinha
Graduate School of Management
Rutgers— The State University
Newark, N.J.
ABSTRACT
The fixed charge problem is a mixed integer mathematical programming
problem which has proved difficult to solve in the past. In this paper we look
at a special case of that problem and show that this case can be solved by for-
mulating it as a set-covering problem. We then use a branch-and-bound in-
teger programming code to solve test fixed charge problems using the set-
covering formulation. Even without a special purpose set-covering algorithm,
the results from this solution procedure are dramatically better than those ob-
tained using other solution procedures.
1. INTRODUCTION
The linear fixed charge problem may be formulated as:
(1) Min I.ejXj + 'Z.fjyj
(2) Subject to £ djjXj > bt i € /,
(F)
j 1 ifxj > 0
(3) yJ = \ 0 otherwise J '^ 7'
(4) and Xj > 0, j € J.
for /= {1, .... m) and/ = {1, ... , n).
In addition to continuous costs, the variables have fixed costs which are incurred when
the corresponding continuous variable becomes positive. All cost are assumed to be nonnega-
tive. Problem (F) is very similar to the standard linear programming problem, differing only in
the presence of the fixed costs. In spite of this similarity, it has proven to be an extremely
difficult problem to solve.
If all the continuous costs are zero, we have a special case of the fixed charge problem
which we will refer to as problem (PF). Problems of this type can occur, for example, when-
ever there is a need to find solutions with the least number of basic, nondegenerate variables.
622 P. MCKEOWN AND P. SINHA
In a network context, Kuhn and Baumol [4] discuss the need to know the least number of arcs
necessary to carry a desired flow. Also, in the survey processing field, it often becomes neces-
sary to check a record of replies to a questionnaire and to determine changes to make the
record consistent. In this case, it is necessary to know the minimum number of such changes
that are necessary for consistency. Both of these problems are examples of problem (PF) with
the former having the standard transportation constraint matrix and the latter having a general
constraint matrix which depends upon the consistency conditions.
A special case of problem (PF) occurs when all the constraint coefficients are nonnega-
tive, i.e., a,j ^ 0 for all /, j. We will refer to this problem as (PF+) since we retain the condi-
tion that all continuous costs are equal to zero. In this paper, we will demonstrate a solution
procedure for (PF+) based on a revised formulation for the problem. We then use a branch-
and-bound integer programming code to solve the revised formulation. The results from this
approach will be compared to those obtained using other procedures.
2. A REVISED FORMULATION
The problem in which we are interested may be formulated as follows:
(5) Min J^fjyj
JtJ
(PF+)
subject to (2) - (4)
(6) where av > 0 for / € /, j € J
(PF+) remains a special case of the fixed charge problem (F) so any results that are applicable
to problem (F) will also be applicable to (PF+).
Two previously derived results for (F) that are of particular interest to (PF+) are:
1) any optimal solution to (PF+) will occur at a vertex of the continuous constraint
set (2) and (4) (Hirsh and Dantzig [3]);
2) a lower bound, L0, on the sum of the fixed costs can be found by solving the
set-covering problem, Ps, below (McKeown [5]).
Min L0 - J fjyj
w
(7) Subject to £8^ > 1, i €/
(8) yj € (0, 1), j € J
i 1 if av > 0
(9) where *(/- j 0 otherwise,
EASY SOLUTION FOR FIXED CHARGE PROBLEMS
We will combine these two results to develop
summarized in Theorem 1 below.
i solution procedure, the essence of which is
THEOREM 1: Let Bg = [j\yj = 1 in an optimal solution to P8}, then there exists a feasi-
ble solution to (PF+) such that xt > 0 for j 6 Bg. Furthermore, this solution will be optimal
for (PF+).
PROOF: Given an optimal solution to P8, we must show that there exists a correspond-
ing solution to (PF+). The first thing to note is that each column of the constraint matrix (7)
of Ph in Bg has at least one nonzero element that is the only nonzero element in that row.
Otherwise, the set would be over-covered and we could reduce the objective value of /*§ by
removing that column from the optimal solution. We may use this result together with the
nonnegativity of the ai:i elements to construct a solution to (PF+) using Bg.
Assume, without loss of generality, that \Bg\ = k and that the decision variables have
been reindexed such that {1 k] € Bg, i.e., the first k variables of (PF + ) correspond to
the optimal basic variables of Ph. We can now construct a feasible solution to (PF+) using the
following two rules:
1)
X] = Max {bjaix}
an * 0
/ € /
Max \bi-
-zVJ
2)
xk = Max
n -^ a '
j— 1 J
u, aik ?= u
/ € /
aik
This proves the existence of a solution to (PF + ) corresponding to Bg. The optimality of
this solution is guaranteed by the fact that both {Ph) and (PF+) have the same objective value
and that this objective value for Ps is a lower bound on (PF+). Hence, Bg corresponds to an
optimal solution to Ph.
3. COMPUTATIONAL COMPARISONS
Since the optimal set of variables for (PF+) can be found by solving the set-covering
problem, />§, we should be able to use this result to reach quicker solutions to (PF+). We
used a mixed integer programming code based on the approach of Tomlin [7] as extended by
Armstrong and Sinha [1] to solve the set-covering problems. Special-purpose set-covering algo-
rithms can be expected to perform even better. Fixed charge test problems first generated by
Cooper and Drebes [2] were used as a basis of comparison between this set-covering approach
and two other procedures. The first such procedure is a branch-and-bound code developed by
McKeown [6] specifically for fixed charge problems while the second procedure used the same
mixed integer code as before, but solved (PF+) as a mixed integer problem.
The original test problems were of dimension 5x10, but larger problems were generated
by putting these smaller problems on the diagonal. Using these problems, the results of our
comparisons are shown in Table 1 below.
P. MCKEOWN AND P. SINHA
Problem
Set
Size
Number
of
Problems
Average Solution Time per Problem in
CPU Seconds on CDC 70/74
Armstrong
and
Sinha
McKeown
r Set.
Covering
Solutions
1
2
3
5 x 10
10 x 20
15 x 30
12
6
4
0.132
0.856
3.039
0.049
0.345
1.357
0.017
0.046
0.101
10
4
2
From the table we can see that the set covering formulation is almost three times faster
than the best alternative approach for the small (5 x 10) problems and up to 13 times faster for
the larger problems (15 x 30). We have also noted the number of problems for which the
linear programming solution was integer feasible for the set covering problems. This occurred
in over half of the cases.
4. CONCLUSIONS
In this paper, we have shown that a fixed charge problem with nonnegative constraint
matrix coefficients and all continuous costs equal to zero can be solved by solving a related set-
covering problem. Computational experience confirms that this procedure yields dramatically
better solution times than any other available solution procedure. Even quicker solution times
can be expected to result if special purpose set-covering codes are used.
REFERENCES
[1] Armstrong, R.D. and P. Sinha, "Improved Penalty Calculations for a Mixed Integer
Branch-and-Bound Algorithm," Mathematical Programming, 21, 474-482 (1974).
[2] Cooper, L. and C. Drebes, "An Approximate Solution Method for the Fixed Charge Prob-
lem," Naval Research Logistics Quarterly, 8, 101-113 (1976).
[3] Hirsch, W.M. and G.B. Dantzig, "The Fixed Charge Problem," Naval Research Logistics
Quarterly, 15, 413-424 (1968).
[4] Kuhn, H.W. and W.J. Baumol, "An Approximative Algorithm for the Fixed-Charges Tran-
sportation Problem," Naval Research Logistics Quarterly, 9, 1-15 (1962).
[5] McKeown, P.G., "A Vertex Ranking Procedure for Solving the Linear Fixed-Charge Prob-
lem," Operations Research, 23, 1183-1191 (1975).
[6] McKeown, P.G., "A Branch-and-Bound Algorithm for the Linear Fixed Charge Problem,"
Working Paper, University of Georgia (1978).
[7] Tomlin, J. A., "Branch and Bound Methods for Integer and Non-Convex Programming,"
Integer and Nonlinear Programming, 437-450, J. Abadie, Editor, (American Elsevier Pub-
lishing Company, New York, 1970).
THE BOUNDED INTERVAL GENERALIZED ASSIGNMENT MODEL
G. Terry Ross
University of Georgia
Athens, Georgia
Richard M. Soland
The George Washington University
Washington, D.C.
Andris A. Zoltners
Northwestern University
Evanston, Illinois
ABSTRACT
The bounded interval generalized assignment model is a "many-for-one" as-
signment model. Each task must be assigned to exactly one agent; however,
each agent can be assigned multiple tasks as long as the agent resource con-
sumed by performing the assigned tasks falls within a specified interval. The
bounded interval generalized assignment model is formulated, and an algo-
rithm for its solution is developed. Algorithms for the bounded interval ver-
sions of the semiassignment model and sources-to-uses transportation model
are also discussed.
1. INTRODUCTION
In general terms, assignment models represent problems in which indivisible tasks are to
be paired with agents. Given a measure of utility (or disutility) associated with each possible
pairing, the objective of the model is to optimize the collective utility associated with assigning
a set of tasks to a set of agents. In practical applications, the number of tasks typically exceeds
the number of agents, and at least one agent must be assigned two or more tasks if all tasks are
to be completed. Examples of such "many-tasks-for-one-agent" problems include the assign-
ment of engagements to a firm's personnel [20], points of distribution to facilities [15], geo-
graphic units to district centers [21], products to plants [1], inventory items to warehouses [8],
harvestable forest compartments to a labor force [12], ships to shipyards [11], scholarships to
students [18], storage compartments to commodities [19], jobs to computers [3], files to mass
storage devices [2,13], defect checkpoints to inspectors [17], and trips to ships [7]. The feasi-
bility of many-for-one assignments will depend on the agents' abilities to complete the collec-
tions of tasks assigned to them. That is, the subsets of tasks that can be assigned to each agent
are determined by the total amount of effort available to the agent and the amount of effort
that each individual task requires.
626 G.T. ROSS, R.M. SOLAND AND A. A ZOLTNERS
Several many-for-one assignment models have been developed which take into account
only upper limits on the total amount of effort that each agent may expand. Each of these
models is a special case of a model developed by Balachandran [3] and Ross and Soland [14]
called the generalized assignment model. This model has the form:
(1) (P) minimize z = £ £ CyXy
(2) subject to £ xy = 1 for all j € 7,
/€/
(3) £ rijXij < b, for all i <E /,
(4) xu = 0 or 1 for all i € /, j £ J.
where /= {1,2, ... , m) is an agent index set, J = {1,2, ... , n) is a task index set, ch
represents the disutility associated with an agent /', task j assignment, fy > 0 denotes the
resource burden incurred by agent / in completing task j, and b, is the resource available to
agent /. The decision variable x„ is interpreted as
II if agent / performs task j
0 otherwise
Constraints (2) and (4) insure that each task is uniquely assigned to a single agent, and con-
straints (3) insure that each agent expends no more than b, resource units in accomplishing
assigned tasks. Differences in the difficulty of tasks and differences in agents' abilities to per-
form the tasks are reflected in the values of the parameter r#.
The special cases of (P) place various restrictions on the form of the agent resource con-
straint (3). Francis and White [9] and Barr, Glover and Klingman [5] have addressed the prob-
lem in which constraints (3) have the form:
(3a) £ Xij < ht for all i € /.
Here b, denotes the number of jobs agent / can complete, for all jobs consume only one unit of
an agent's resource when the agent performs the task (i.e., ru = 1 for all / € /, j € /). The
model (l,2,3a,4) is a generalization of the standard assignment problem of linear programming
in that it permits an agent to undertake more than one task. It has been called the generalized
assignment problem by Francis and White and the semi-assignment problem by Barr, Glover,
and Klingman.
Caswell [6], DeMaio and Roveda [8], and Srinivasan and Thompson [16] studied the
problem in which (3) is replaced by:
(3b) £ rjXy < b, for all i € /.
The model (l,2,3b,4) explicitly considers differences in the difficulty of tasks incorporated in
the parameter /}. Srinivasan and Thompson called this model the sources-to-uses problem to
reflect the interpretation of the model as a transportation problem in which the demand at the
y'-th location, /}, is to be supplied by a single source.
Practical considerations frequently require that the agents expend a minimum total
amount of effort in completing assigned tasks. Placing both minimum and maximum restric-
tions on the resources each agent can expend, yield assignments which neither overburden nor
BOUNDED INTERVAL GENERALIZED ASSIGNMENT PROBLEM 627
underutilize the agents. Such restrictions arise in most personnel planning applications [20].
Managerial policies usually require an equitable distribution of work across agents. Analagous
restrictions crop up in other contexts as well. For example, in machine loading models, it usu-
ally is desirable to balance machine workloads rather than allowing some heavily loaded and
some lightly loaded machines. In facility location models, capacity constraints may restrict both
the minimum and maximum size of a facility to avoid diseconomies of scale associated with
plant sizes outside of a reasonable range, to permit piecewise linear approximation of concave
cost functions, or to restrict both the minimum and maximum number of facilities [15]. Simi-
larly, territory design procedures for problems of political districting, school districting, and
sales districting require an equitable distribution of some entity (such as voters, minority stu-
dents, or sales potential) among the districts. Finally, in some applications, upper limits on the
effort an agent can expend may be irrelevant, and only lower limits need be considered. Such a
situation arises in the segregated storage problem [19] which requires only that a minimal
amount of storage space be allocated to store commodities and no maximum allocation is
specified.
Thus, from the standpoint of modeling flexibility, it is desirable that assignment models
consider explicitly upper and/or lower bounds on the efforts agents must expend in completing
assigned tasks. While most "many-for-one" assignment models consider upper bounds, lower
bounds have largely been overlooked. In this paper, we introduce the bounded interval gen-
eralized assignment model and discuss how existing algorithms can be modified to accommo-
date lower bounds on agent workloads for this model and its special cases.
2. THE BOUNDED INTERVAL GENERALIZED ASSIGNMENT MODEL AND
ALGORITHMIC CONSIDERATIONS
The bounded interval generalized assignment model may be formulated as follows:
(5) IP*) minimize z = £ £ c^
/€/./€./
(6) subject to £ xu : = * f°r a^ J € ^
/€/
(7) a, < £ r0Xjj < b-, for all / € /,
(8) x,j = Oor 1 for all / € /, j£ J.
Notice that (P*) derives from (P). Fortunately, the modeling flexibility achieved through
the introduction of lower bounds a, > 0 in constraints (3), (3a), or (3b) does not complicate
significantly the computational effort required to solve any of the models described above.
Rather, as we shall show, straightforward modifications can be made to the existing algorithms
for the semi-assignment problem, sources-to-uses transportation problem, and the generalized
assignment problem. The interested reader should consult the cited references for the details
of the original algorithms.
In the case of the semi-assignment problem, the constraint matrix is totally unimodular,
and integer solutions can be obtained using the simplex method. To impose the lower limit,
(7a) J\x,j > at for all/ € /,
628 G T ROSS, R.M. SOLAND AND A. A. ZOLTNERS
one need only add an upper bounded slack variable s, < bt — a, to each of the constraints (3a)
and rewrite them as equality constraints. Optimal solutions to the resultant bounded variable
linear program will be integer valued.
Models with constraints (3) or (3b) are not totally unimodular. Hence, the solutions of
the linear programming relaxation (i.e., x,j > 0 for all i,j) need not be integer. Branch and
bound approaches have been developed for deriving integer optimal solutions which solve
linear programming relaxations for fathoming and to compute lower bounds. In the case of
(3b), a linear programming relaxation is the standard transportation problem [16]; and in the
case of (3), a linear programming relaxation is the generalized transportation problem [3]. As
in the case of the semi-assignment problem, to impose constraints (7) or
(7b) aj < £ rjXiJ < bi for all/ € /
in a linear programming relaxation, one need only add upper bounded slack variables
Sj ^ bj — d) to constraints (3) or (3b) and rewrite them as equality constraints.
The algorithm developed by Ross and Soland [14] for the generalized assignment problem
does not solve a linear programming relaxation to determine the lower bounds. Instead, a
Lagrangian relaxation is solved in the form of a series of separable binary knapsack problems.
The Lagrangian relaxation has the form:
(8) (PK) minimize ZK = £ £ cuxv + £ \j(\ - £x„)
HIJSJ j£J /€/
subject to £ rl} Xy < b, for all / € /
ju
xu = 0 or 1 for all / € /, j € J.
The value of each \j is set equal to c2y, the second smallest value of Cy for all /' € /. These \j
are optimal dual multipliers for the problem:
(PL) minimize £ £ CyXu
i£IJ€J
subject to £ xu ■ = 1 for all j € 7,
/£/
0 < xi} <■ 1 for all / € I j € J.
Thus, determining a lower bound requires two steps. First, solve (//.), then solve (/\). If the
primal solution X = (5ey) to_(PL) should also satisfy (8), then Z = ZL = Zx, and (Pk) need
not be solved. Frequently, J will not satisfy (8), and (Pk) must be solved to find Zx.
To incorporate the lower bounds a, into the algorithm, one need only replace constraints
(8) by constraints (7) giving rise to the problem (P*) with knapsack constraints bounded both
from below and from above. Seemingly, this minor modification to the form of (PK) should
have little effect on the algorithm. However, it must be noted that (Px) will involve fewer 0-1
variables and may be easier to solve than (P*). The reason is best explained by considering an
equivalent form of the objective function of (PK):
ZK = X ^j ~ maximum [£ £ (\, - cy) xJ.
BOUNDED INTERVAL GENERALIZED ASSIGNMENT PROBLEM 629
Clearly, with constraints (8), one can set any xu equal to zero which has an objective function
coefficient ikj - cu) ^ 0. Thus, using the values of Xj calculated from solving (Pi), (PK)
reduces to a problem involving at most n 0 — 1 variables. Such a reduction is not possible for
(Pt)-
In addition to providing a lower bound, the solutions to (PL) and (P*) may be used to
select a branching (or separation) _yariable for defining subsequent candidate problems. As
noted above, the solution to (PL), X, is usually not feasible to (7). In essence, the solution to
(P*), X= (Xjj), may be interpreted as recommending changes in X which must be made in
order to satisfy (7). That is, it is possible that for some j € 7, £5c,7 = 0 to avoid overloading
,€/
any agent or £ xu > 1 to insure every agent uses a minimum amount of his resource. Those
variables xu with an optimal value of one indicate agent-task pairings that should be made;
whereas, those x,7 with an optimal value of zero indicate pairings that should be avoided. Thus,
these variable values indicate changes that will reduce the aggregate infeasibility of Jin (7),
and they are helpful in choosing a branching variable.
To formalize the concept of reducing aggregate infeasibility, we define the infeasibility in
constraint / prior to taking a branch to be
A = max {0, df, d~)
where d,+ = £ r,y3cy —b„
The set /+ = {/ € l\d,+ > 0} identifies those constraints (7) for which A^exceeds the upper
bound, and /" = {/€ l\d~ > 0} identifies those constraints (7) for which A' fails to satisfy the
lower bound.
Suppose I+ ^ 0 and k€[j € J\xjj = 1 and / € /+}; if xik is set to 0 then d* and d~
become:
di+ = Z ruXij ~ bj - rik
d,r = a, • ~ £ ru Xy + rik
and the resulting infeasibility in constraint / becomes
A*= max{0, di+, d~\.
Assuming that task k is reassigned to the second least costly agent, (say agent /?, where
chk = min C\k) then the infeasibility in constraint h becomes
A*= max {0, d£, d^\
where
dh = Z n,j *hj - bh + rhk
dh = a,, - ZoyX/y - rhk.
630 G.T. ROSS. R.M. SOLAND AND A. A. ZOLTNERS
Hence, the net difference in total infeasibility is:
kD*= (Z>, + D,,)- (Df + Dlj)
If AZ)* > 0 then setting xik = 0 yields a reduction in aggregate infeasibility, and if AZ)* < 0
then such a branch will not reduce aggregate infeasibility.
Similarly, suppose that l~ ^ 0 and k € [j € y|xy = 0 and / € /""}; if xik were set to 1
then df and d~ become:
d,+ = Z rv xu - b, + rik
d,~ = a, ■ - £ ry xy - rik
J&
and the resulting infeasibility in constraint / would be
Z)/*- max {0, #\ 4-}.
If Xft is set to 1 then task k is assigned to agent / and agent ik relinquishes it, where
ik = min cik. Hence, the infeasibility for constraint ik becomes
Dtk - max {0, df , 4" }
where
df = £ V*V€ ~ *'* ~ V
/€/ '
di~ = «/A - £ r,^ *,y + r, *.
/€./
The net difference in infeasibility is
AA*= (A + D,k)- (D^+ Z)£)
where Z), and Z>, are the infeasibilities in constraints /and 4 prior to any branch. As before, if
AD* > 0 then there is reduced infeasibility following a branch on variable xlk.
Several rules for selecting the branching variable, *,.,., are formulated as follows:
I. a) Xpp is that variable for which
A £>/.*= max [ADf]
(i,j)€ H+UH~
where H+ = { (/, j) \x0 = 0 and / € /+}
//- = {(/,y)|xy= 1 and/ € /"}
b) If AZ>/.* = 0 in a) then xfy. is that variable for which
AZ>/«* = max {AZ>/}
(/, y)€ (G+~H+) U(G--H-)
where G+ = {(/, y)|3cy = 1 and / 6 I+}
Gr = {(i,j)\xu = Oand/ € /"}.
BOUNDED INTERVAL GENERALIZED ASSIGNMENT PROBLEM
II. a) x(*j* is that variable for which
min i .
piy = min
min
(/, j) e G+ '
(/, j) € G~
ADf
b) If AZ)/= 0 for all (/, j) € G+UG~ then xiT is that variable for which
max
JdF:
where £+ = {(/, j)\xu = 1 and / € /+},
F/ denotes the set of tasks assigned to agent / by prior branching.
Rules la and lb are designed to choose that variable which reduces the post branch aggregate
infeasibility by the greatest amount. Rule Ha conditions the choice of branching variable on the
additional cost incurred per unit reduction in infeasibility. Rule lib is the one used in [14]; the
variable chosen by this rule represents an agent-task pairing which should be made considering
the penalty for not doing so weighted by the fraction of the agent's remaining free resources
consumed by the assignment.
As the algorithm progresses and new candidate problems (CPs) are defined by the branch-
ing process, the additional steps given below may be taken to facilitate fathoming. These steps
are specialized adaptations of more general forcing (or variable fixing) tests suggested by Balas
[4] and Glover [10].
In solving any (CP), any x?y for which r,y > b,- — £ xj'jrj'j mav De set equal to zero.
JtFr
Here F? denotes those j € J for which x,-7 has been assigned a value of zero or one by prior
branching or variable fixing tests. Similarly, if there is an x,y for which at> — £
I
jzj-f;
then Xj'j> must be set equal to one in the solution to (CP). These
variable forcing tests may subsequently result in other variables being forced to zero or to one
when all of the resultant implications are considered. Moreover, forcing certain variables to
zero or to one in the solution to (CP) may affect the values of some of the Xj obtained from
solving (Pi). This change may, in turn, increase the value of the lower bound provided by
(P*)-
Another test may be used to check the feasibility of (P*) (or any candidate subproblem).
Summing the constraints (7) together yields the constraint (9):
(9)
B.
This new constraint, together with constraints (6), implies that for any feasible solution to (P*)
we must have:
(10)
Zo:
A and J) r
JZJ
632 O.T. ROSS. R.M. SOLAND AND A A. ZOLTNERS
where
r', = max {/•..} and r"= min {/•„}.
The values necessary for the tests (10) can be updated easily as part of the branching process in
order to apply this test to each (CP).
The algorithm terminates in the usual way when all candidate problems have been
fathomed.
3. CONCLUSION
This note has described an efficient branch and bound algorithm for the bounded interval
generalized assignment problem. The algorithm serves as a useful tool for solving a large
number of applications of this assignment model, a representative sample of which is men-
tioned in the introduction.
REFERENCES
[1] Abella, R.J. and T.E. Bova, "Optimal Plant Allocation of Stockkeeping Units," presented at
TIMS/ORSA Joint National Meeting, San Francisco, California (May 1977).
[2] Babad, J.M., V. Balachandran and E.A. Stohr, "Management of Program Storage in Com-
puters," Management Science, 23, 380-390 (1976).
[3] Balachandran, V.,."An Integer Generalized Transportation Model for Optimal Job Assign-
ment in Computer Networks," Operations Research, 24, 742-759 (1976).
[4] Balas, E., "An Additive Algorithm for Solving Linear Programs with Zero-one Variables,"
Operations Research, 13, 517-545 (1965).
[5] Barr, R.S., F. Glover and D. Klingman, "A New Alternating Basis Algorithm for Semi-
assignment Networks," Research Report CCS-264, Center For Cybernetic Studies,
University of Texas, Austin, Texas (January 1977).
[6] Caswell, W., "The Transignment Problem," Unpublished Ph.D. Thesis, Rensselaer
Polytechnic Institute (1972).
[7] Debanne, J.G. and J-N Lavier, "Management Science in the Public Sector— The Estevan
Case," Interfaces, 9, 66-77 (1979).
[8] DeMaio, A. and C. Roveda, "An All Zero-One Algorithm for a Certain Class of Transpor-
tation Problems," Operations Research, 19, 1406-1418 (1971).
[9] Francis, R.L. and J. A. White, Facility Layout and Location: An Analytical Approach,
(Prentice-Hall, Englewood Cliffs, New Jersey, 1974).
[10] Glover, F., "A Multiphase-Dual Algorithm for the Zero-One Integer Programming Prob-
lem," Operations Research, 13, 879-919 (1965).
[11] Gross, D. and C.E. Pinkus, "Optimal Allocation of Ships to Yards for Regular Overhauls,"
Technical Memorandum 63095, Institute for Management Science and Engineering,
The George Washington University, Washington, D.C. (May 1972).
[12] Littschwager, J.M. and T.H. Tcheng, "Solution of a Large-scale Forest Scheduling Problem
by Linear Programming Decomposition," Journal of Forestry, 65, 644-646 (1967).
[13] Morgan, H.L., "Optimal Space Allocation on Disk Storage Devices," Communications of
the ACM, 17, 139-142 (1974).
[14] Ross, G.T. and R.M. Soland, "A Branch and Bound Algorithm for the Generalized Assign-
ment Problem," Mathematical Programming, 8, 91-103 (1975).
[15] Ross, G.T. and R.M. Soland, "Modeling Facility Location Problems as Generalized Assign-
ment Problems," Management Science, 24, 345-357 (1977).
BOUNDED INTERVAL GENERALIZED ASSIGNMENT PROBLEM 633
[16] Srinivasan, V. and G.L. Thompson, "An Algorithm For Assigning Uses to Sources in a
Special Class of Transportation Problems," Operations Research, 21, 284-295 (1973).
[17] Trippi, R.R, "The Warehouse Location Formulation as a Special Type of Inspection Prob-
lem," Management Science, 21, 986-988 (1975).
[18] Wagner, H.M., Principles of Operations Research, (Prentice-Hall, Englewood Cliffs, N.J.,
1968).
[19] White, J. A. and R.L. Francis, "Solving A Segregated Storage Problem Using Branch and
Bound and Extreme Point Ranking," AIIE Transactions, 3, 37-44 (1971).
[20] Zoltners, A. A., "The Audit Staff Assignment Problem: An Integer Programming
Approach," Working Paper 74-34, School of Business Administration, University of
Massachusetts, Amherst, Massachusetts (September 1974).
[21] Zoltners, A. A., "A Unified Approach to Sales Territory Alignment," Sales Management:
New Developments from Behavioral and Decision Model Research R. Bagozzi, Editor,
(Cambridge, Massachusetts Marketing Science Institute, 1979), 360-376.
THE M/G/l QUEUE WITH INSTANTANEOUS
BERNOULLI FEEDBACK*
Ralph L. Disney
Virginia Polytechnic Institute and State University
Blacksburg, Virginia
Donald C. McNickle
University of Canterbury
Christchurch, New Zealand
Bell Laboratories
Holmdel, New Jersey
ABSTRACT
In this paper we are concerned with several random processes that occur in
M/G/l queues with instantaneous feedback in which the feedback decision pro-
cess is a Bernoulli process. Queue length processes embedded at various times
are studied. It is shown that these do not all have the same asymptotic distri-
bution, and that in general none of the output, input, or feedback processes is
renewal. These results have implications in the application of certain decompo-
sition results to queueing networks.
1. INTRODUCTION
In this paper we are concerned with several random processes that occur within the class
of M/G/l queues with instantaneous feedback in which the feedback decision process is a Ber-
noulli process. Such systems in the case G = M are among the simplest, nontrivial examples
of Jackson networks [8]. Indeed, they are so simple that they are usually dismissed from con-
sideration in queueing network theory as being obvious. We will show that far from being
obvious, they exhibit some important unexpected properties whose implications raise some
interesting questions about Jackson networks and their application.
In particular, Jackson [8] observed that in his networks the vector-valued queue length
process behaved as if the component processes were independent, M/M/l systems. Since those
results appeared there has developed a mythology to explain them. These arguments usually
rest on three sets of results that are well known in random point process theory: superposition,
thinning, and stretching. By examining the network flow, it will be shown that the applications
of these results are inappropriate for queueing networks with instantaneous, Bernoulli feedback.
These flows are considerably more complicated than one expects based on such arguments.
The research was supported under ONR Contracts N00014-75-C-0492 (NR042-296) andN00014-77-C-0743 (NR042-296).
635
636 RL. DISNEY. DC. MCNICKLE AND B. SIMON
It is shown that in general, both the input and output processes of the M/M/l queue with
feedback are Markov-renewal, and the kernels of those Markov-renewal processes are given.
The output of the M/G/l queue with feedback is also Markov-renewal, and that kernel is given.
It is shown that in general these processes are never renewal. The implications of these facts
are discussed in Section 4.
1.1 The Problem and Notation
We assume the usual apparatus of an M/G/l queue with unlimited waiting capacity. The
new idea is that a unit which has received service departs with probability q and returns for
more service with probability p. p + q = 1 . Without loss of generality for the processes stu-
died here, the returning customer can be put anywhere in the queue.
To establish notation it is assumed that the arrival process is a Poisson process with param-
eter X > 0. The arrival epochs are the elements of {Wn: n = 1,2, ...}. Service times are
independent, identically distributed, nonnegative, random variables, S„ with
Pr [Sn < t] - Hit), t > 0,
E[S„] < «>.
We define H*is), the Laplace-Stieltjes transform of Hit), by
H*is)= f°° e-s'dHit), Res ^ 0.
The arrival process and service times are independent processes.
Service completions occur at T0 < T] < T2 ... called the output epochs. Let
0, if the n -th output departs,
1, if the n-th output feeds back.
{ Yn} is a Bernoulli process.
Elements of the subset {tn} C [Tn] are called the departure epochs and are the times at
which an output leaves the system. The elements of the subset {t,,} c {Tn} are called the feed-
back epochs and are the times at which an output returns to the queue. {/„} U {t„} = \Tn).
The times T'n are the times at which a unit enters the queue. {7^} is called the input pro-
cess, {t;,} = [w„)\j Wn}.
There are five queue length processes to be studied. They are closely related as will be
shown. Let Qit) be the queue length (number in the system) at t. Then,
Q- („) = Q(wn - 0); Qi in) = QiVn - 0); Qt in) = QiTn + 0); Qt in) = Qitn + 0) are
respectively the embedded queue lengths at arrival epochs, input epochs, output epochs, depar-
ture epochs.
2. QUEUE LENGTH PROCESSES
The queue lengths listed in Section 1.1 are closely related. The steady state versions of
[Qi in)} and {Q^ in)} are of primary concern. They are studied in Sections 2.1 and 2.2
separately. They are related to the other processes in Section 2.3. The important special case
for G = Mis then studied in 2.4.
M/GA QUEUE WITH INSTANTANEOUS BERNOULLI FEEDBACK
feedback process
J-
arrival
input
output
process
process
Figure 1.
process
2.1 The{(?4+(/!)} Process
There are several ways to study this process. The following appears to be direct, correct,
and may help explain why these feedback problems have received such little attention in the
queueing literature. First, it is clear that
j r„_, + S;,, ifQfin - 1) > 0,
'" = | /„_, + /„ + S^, if Q+ (n-\) = 0.
Here S'n is the total service time consumed between the (n - 1) - st and «-th departure. /„ is
the idle time following tn_x when Qf (n — 1) = 0. For the M/G/l queue, the /'s are indepen-
dent, identically distributed, random variables that are exponentially distributed with parameter
Without loss of generality, since customers are indistinguishable,
S'„ = S] + s2 + ... sm,
where m is the number of services performed between the (n — 1) - st and n-th departure.
Since [Yn} is a Bernoulli process, m is geometrically distributed and it follows that {S'n} is a
sequence of independent, identically distributed, random variables. Thus, the Laplace-Stieltjes
transform of the distribution function of S'„ is easily found to be
G*(s) = qH*(s)/[\- pHHs)].
Using standard embedded Markov chain methods [3, 167-174] one finds that the probabil-
ity generating function of Jp, the limiting probability distribution of {^4" («)}, is given by
(0) (z - 1) G*(k -kz)
(1)
and
(2)
g(z)-
■ G*(\ -kz)
ir'(0)= \-kE[Sn]/q.
638
R.L. DISNEY, DC. MCNICKLE AND B. SIMON
If one is willing to assume that the M/G/l queue with instantaneous, Bernoulli feedback
has a queue length process which asymptotically has the same distribution as another M/G/l
queue without feedback, then (1) and (2) follow immediately. This assumption is valid since if
customers feedback to the front of the queue, the total service time of the n-th customer is S'„.
{S'n} is a sequence of independent, identically distributed, random variables with mean E[S„]/q.
Alternatively, one can argue that the M/G/l queue with feedback (as defined here) has the
same asymptotic distribution for its queue length process as an M/G/l queue without feedback
if one takes the arrival process parameter in the latter case to be k/q. Indeed, both of these
assumptions and several others that are used to "prove" that these queues with feedback are
trivial have now been proven by the arguments leading up to (1) and (2). That these argu-
ments can be applied more generally is easily proven. In the remainder of this paper, in Takacs
[10] and in Disney [6] it is shown that while these arguments may imply that the study of
queue lengths at departure times is trivial, the same cannot be said for other processes of
interest.
2.2 The {Q?(n)} Process
This is the queue length process embedded at output points. Since {/„} C {Tn}, {Q4 (n)}
is a process on a coarser grid than {Q? (n)}. Since one is ultimately to be concerned with both
{Qt (n)) and [Tn - r„_i}, the following study is for the joint process {Qfin), Tn - T„_|}.
The marginal results for [Q^ (n)} then will be easy to determine.
THEOREM 1: The process (C?3+ (n), Tn - Tn_x) is a Markov-renewal process with kernel
A (i,j,x) = Pr{Q? (n) * j, Tn - fw_, < x\Q$ (n - 1) = /}. If one defines
Pj(y)= (ky)Je Ky/jl,
j = 0,1,2, ..
A (i,j,x) = <
L
if; < / - 1,
i+l(y)q)dH(y),
if / ^ 0,
C (1 - e-k(x-y)) (Pj-](y)p + Pi{y)q)dH(y),\n = 0,
J > 0,
u;
(1 _ e-i(x-y)) pQ(y)qdH(y), if j = / = 0.
PROOF:
j S„, if Q$(n - 1) > 0,
Tn - rn_, - j Jn + Sn> .f Q+ {n _ 1} = 0
where /„ is the exponentially distributed idle time preceeding S„ if Q3+ (n ■
then follows directly using arguments as in [5]. □
1) = 0. The result
As x — 00, A (ij.x) — A (ij) the one step transition probability for the {Q3 (n)} process.
Then using standard embedded Markov chain results [3, 167-174] one can show that the proba-
bility generating function g(z) for the limiting probabilities n(j) of Q3+ («) are given by
tt(0) (2 - 1) (pzH*(k - \z) + qH*(k - kz))
(3)
?(z) =
z - pzH*(k - kz) - qH*(k - kz)
M/GA QUEUE WITH INSTANTANEOUS BERNOULLI FEEDBACK 639
and
(4) ir{0) = q-\E[S„l
2.3 Other Queue Length Processes
The queue length and limiting probabilities for the queueing processes, {(?f(n)},
[Qi (n)} now follow from a theorem found in Cooper [3, 155]. From this it follows that
{QU)}, {(?f (/»)}, and [Qfin)} are asymptotically, identically distributed (see Cooper [3, 65])
and Qi(n), and {Qi in)} are asymptotically, identically distributed. Clearly, [Qf in)} and
{Qf in)} are not asymptotically, identically distributed. That {Qt in)} and [Qf in)} are not
asymptotically, identically distributed can be seen as follows. First, in the set up of studying the
(C?3+ in)} process one must decide how to count the feedback customer when he appears. The
clean way to do this is to use Yn as defined in Section 1.1 and *Qf in) as the number in the
queue not including the outputting customer. Then one can study the process [Y„, *Q?in)}.
Indeed, this is precisely the direction used, for example, in d' Avignon and Disney [4]. Then
the {Qi in)} of Theorem 1 above would be the {*Qy in) + Y„} process of [4]. It then follows
that {*(?3+(«)} and (Q4+(n)} are asymptotically, identically distributed. Thus, if one does not
count the feedback customer in the queue length process, the queue length processes defined in
Section 1.1 are all asymptoticaly, identically distributed.
2.4 The Ml MIX Case
If one assumes that the service time distribution is
Hit) = 1 - e-*', t > 0,
some further clarification is possible here. From the results of Jackson [8],
ir'O')- 1- — — , 7 = 0,1,2, ....
From (3) and (4) one obtains
77 (0) =
T
- — 1
<7M j
7T(/) =
l
A_
[ x )J
\qfl)
\p
+
— 1
p 1
j= 1,2, ....
Comments in Section 2.3 explain this difference between tt(j) and ir'(j).
3. FLOW PROCESSES
To further clarify the problems here, it is useful to study the flow processes in this sys-
tem. There are five processes of interest: the arrival process, the input process, the output pro-
cess, the departure process, and the feedback process.
There have been some questions since the publication of the Jackson results concerning
the interpretation of his results [2]. In his paper Jackson showed that for his networks the joint
limiting probability for the vector of queue lengths at each server could be factored into limit-
ing probabilities for the queue length at each server. This imples that the queue lengths are
independent in the limit. Furthermore, the marginal limiting probabilities were found to be
640 R L DISNEY. DC. MCNICKLE AND B. SIMON
precisely those of an M/M/l queue. Burke (2], has argued that the Jackson results are surpris-
ing. Burke's argument is based on showing that the input to a single server queue with feed-
back is not Poisson because the interinput times (our [T'n - T'„-\l) are not exponentially distri-
buted. [2] gives the precise result
Pr\r„ - r;_, < /} = 1 - qP ~ X e~kl P±-e-»', t > 0.
fJL — k (X — A
In this section we will study some of the flows in this network and show indeed that the
Jackson results are surprising.
3.1 Departures
The departure process {t„} can be studied as in Disney, Farrell, deMorais [5] upon using
the mapping in Section 2.1. Thus we know that whenever {Sn} is a renewal process with
exponential distribution this departure process is a renewal process, and is a Poisson process.
This is the Jackson case. So we conclude that the departure process from the Jackson network
is a Poisson process.
From the results of Section 2.1 it would seem possible that the departure process is Pois-
son even if S„ is not exponentially distributed. The result that is needed for the results of [5]
to follow is that 5^ be exponentially distributed (since it is known that {S'„} is a sequence of
mutually independent, identically distributed, random variables).
LEMMA 1: The departure process from the M/G/l queue with feedback is a renewal pro-
cess if and only if S„ is exponentially distributed for every n. In that case the departure process
is Poisson.
PROOF: From Section 2.1 we have G*is), the Laplace-Stieltjes transform of the distribu-
tion functions of S{, is given by
/-«/ ^ qH*(s)
G (5) = i — u*t \ ■
1 - pH*(s)
From [5], when the queue capacity is infinite the departure process will be a renewal pro-
cess if and only if S'n is exponentially distributed with parameter a, and will be Poisson in that
case. But this implies that H*(s) must satisfy
alia + s) = qH*is)/[\- pH*is)].
The only solution here is
HHs) = -^-
aj q + s
which implies Hit) is exponential. □
3.2 Outputs and Inputs
From Section 2.2 it is clear that the output process is a Markov-renewal process whose
distributions are given by A iij.x). From these, the following results are obtained.
THEOREM 2: The output process [T„ - r„_,} is a renewal process if and only if q = 1
and Hit) = 1 - e^'.
M/GA QUEUE WITH INSTANTANEOUS BERNOULLI FEEDBACK 641
PROOF: If q = 1 and Hit) = 1 - e_M', the output process and departure process are
identical processes. Furthermore, the processes are both departure processes from a M/M/l
queue without feedback. From [5] we have that this departure process is a Poisson process and
"if follows. To prove "only if we consider the contrapositive statement and assume q ^ 1.
(The other side of the contrapositive would have Hit) ^ 1 — e-M'. But then "only if follows
trivially from [5]. Thus, we need only consider the case of q ^ 1.) Equations (3.1) and (3.2)
in [5] can be modified in such a way that one can show that if q ^ 1, there is no solution to
both of those equations simultaneously. Then using the same arguments as in [5] one has that
{T„ - Tn_x] is not a renewal process and therefore "only if is proven. □
To be more specific, Theorem 2 can be particularized as
COROLLARY 1: The output process [Tn — Tn_:} for the M/M/l queue is a Poisson pro-
cess if and only if q = 1. One can prove this result (in fact it is obvious) directly from
Theorem 2. The following is an alternate proof that exposes a bit more of the properties of
these systems. Again we use a contrapositive proof for "only if.
PROOF: Define
F(x)= Pr{T„- r„_, < x).
F(x) = 77 AU where (/is a column vector all of whose elements are 1, n is the vector of limit-
ing probabilities given in Section 2.4 for (Q3+ in)} and A is the matrix of A (i,j,x). Then from
Theorem 1 one obtains after some algebraic manipulations:
(5) Fix) = \q - -\ C [1 - e-^-y'] dHiy) + \p + ~\ Hix)
M I ° Ml
for any M/G/l queue with instantaneous, Bernoulli feedback.
For Hiy) = 1 - e_w, it follows that
(6) Fix) = 1 - qfX~k e~Kx - -££— e~»\ x > 0.
/JL — \ /J, — A.
Thus, single intervals are not exponentially distributed and the output process is not a Poisson
process if q ^ 1. On the other hand if q = 1, then we fulfill the conditions of Theorem 2.
Hence, [Tn — Tn_x) is a renewal process. But from (6) this renewal process has exponentially
distributed intervals and thus is a Poisson process. □
Formula (6) was previously found by Burke [2] for the distribution of times between
inputs. The input process can be analyzed as follows:
THEOREM 3: If Hix) = \ - e~»\ the process {Qfin), T'„ - T'n_x) is a Markov-
renewal process with kernel
YiiJ.x) = PrlQi in) = j, T'„ - T'n_x < x\Qi (n - 1) - /]
given by
YiiJ,x) -
0; ; >
i + 1,
J>-
- qe~
<*)q
dHu+x)is);
J = 0, i
> 0,
X'H
k + fX
u-
e-(\+n)(x-s
>) + p\ Q'
~jdH(
er?%\ -
- e-kx);
j =
- i + 1,
642 R L DISNEY. DC MCNICKLE AND B SIMON
where dH("+])is) = ^5)"g~M' ds.
n\
PROOF: Clearly, if j > / + 1 then YiiJ.x) = 0. If j = 0 then Y(i,j,x) is the probability
that the /' + 1 customers in line all depart before x and the first arrival occurs after the last
departure, but before x, or, the first / - 1 customers depart, but the last one feeds back before
x, and there are no arrivals while this is happening.
If 1 < j < / then Y(iJ,x) is the probability that i — j + 1 customers depart before x, no
arrivals occur during this time, but between the last departure and x, an arrival occurs before a
departure; or, / — j + 1 customers are served before x, the first / - j depart, the last one feeds
back, and there are no arrivals while this is happening.
If j = i + 1 then Y(i,j,x) is the probability that there is an arrival before x and no depar-
tures before x. Since Y(i,j,x) never depends on {Q2 ik); k < n — 1} or \Tk\ k < n}, the
process [Qi in), T'n — T'n_\\ is a Markov-renewal process. □
Now, if Y(x) is the matrix whose elements are Y(i,j,x), tt is the vector of probabilities
found in (3) and U is a vector all of whose elements are 1 then it is easy to see that
Fix) = Pr[T'n - r;,_, < x] = ttY(x)U
and
G(x,y) = Pr[r„ - r;_, < x, T'n+X - T'„ < y] = tt Y(x) Y(y)U.
where Fix) is the Fix) given by (6). Of course, if { T'n — T'n_\) is to be a renewal process then
it is necessary (but not sufficient) that
Gix,y) = Fix)Fiy).
From this we can conclude:
COROLLARY 2: The input process to the M/M/l queue with instantaneous, Bernoulli
feedback is not a renewal process unless q = \.
PROOF: If q = 1 then the input process is just the arrival process which is Poisson.
If the input process is a renewal process for q ^ 1 then it must be true that
y/x\TrYix)U = Fix)
Vxv; 7rYix)Yiy)U = Fix)Fiy) where
Fix) is given by (6) and (/is a column of l's. Thus,
Vxv; U-^ry- Yiy)U=0.
Fix)
y PH- P-y-y + np-(jx+\)y\
Some algebra yields
f
7TYix)\
Fix) J
Yiy)U =
Fix]
1- (1
- e^*) \
r
Fix
)
\ 1
if? * 1,
Fix)
-i\-e~
<
e-nx -
fiq
- k
-X
- A
fj. — X
e *y + pe
-MdS^k
M/GA QUEUE WITH INSTANTANEOUS BERNOULLI FEEDBACK 643
Thus, to show that the input process is not renewal, it suffices to show that for some y,
(X — X /JL — A
The third term of the Taylor expansion of this expression is
M >i2 _ PH M2 , P(p- +X) = p\fji , 0
fi-\ 2 fi -\ 2 2 2
so (by [1, 198] for instance), it cannot be identically zero unless p = 0 (i.e., g = 1). □
It seems obvious that the arrival process and feedback process are not independent
processes. One can show, using the above arguments:
COROLLARY 3:* Either the feedback process is not a Poisson process or the arrival pro-
cess and feedback process are not independent processes (or both) for the M/M/l queue with
instantaneous, Bernoulli feedback.
PROOF: This result follows immediately from Burke's result [2] on the distribution of
the interinput arrivals. For if the feedback process is both independent of the arrival process
and is itself a Poisson process, the input process is Poisson. Thus, Burke's result contradicts
the assumption. □
3.3 Feedback
The feedback stream seems to be quite difficult to work with. From the previous section
we know that it is either not independent of the arrival stream or not a Poisson stream.
Melamed [9] has shown that this feedback process is not a Poisson process. We conjecture
further that it is not independent of the arrival process. If so, then the known superposition
theorems cannot be used to study feedbacks in terms of the arrival, feedback and input
processes.
Since the feedback stream is the result of applying a filter to the Markov-renewal output
process, it is itself Markov-renewal on the state space {1,2, . . .}. Even in the M/M/l case, the
form of the feedback stream does not appear to reduce to that of any simpler process.
4. CONCLUSIONS
There are several conjectures that one can pose concerning networks based on the results
of this paper. First with respect to queue length, busy period, and departure processes, if one
adopts the "outsiders" view [3] these processes appear to be those generated by an M/G/l
queue without feedback. However, if one adopts the "insider" view the queue length process
does not appear to behave as seen by the "outsider."
Flow processes in this network cannot be explained by appeal to superposition, stretching,
and thinning results for Poisson processes. The requisite independence assumptions both
within and between streams of events are not satisfied here. Thus, one cannot assume that
these queues which act "as if they were M/M/l queues to the "outsider" are M/M/l queues
to the "insider." In particular, this hints at the possibility that in these networks, even as simple
as Jackson networks, any attempt to decompose the network into independent M/M/l queues
is doomed to failure. This decomposition must account for the internal flows and these not
only appear to be non Poisson, they are nonrenewal and are dependent on each other.
*Melamed [91 has shown, using other arguments, that the feedback stream is not a renewal process.
644 R.L DISNEY. DC. MCNICKLE AND B. SIMON
In [9], it is shown that in the Jackson structure, the flow along any path that returns a
customer to a node that he has previously visited is not only not Poisson, it is not renewal.
Thus, if Jackson networks have loops, (direct feedback as in this paper being the simplest
example), they cannot be decomposed into sub-networks of simple M/M/l queues. In particu-
lar, these results imply that a node-by-node analysis of waiting times depending as they do on
the "insiders" view is not valid if one simply uses M/M/l results at each server. Takacs [10]
studies the waiting time problems in the system discussed in this paper. Disney [6] presents
another view of the same problem.
ACKNOWLEDGMENTS
We would like to thank Dr. Robert Foley and Dr. Robert B. Cooper for their helpful com-
ments on this paper. In particular Foley first brought to our attention that {Q3+ (n)} and
{(?4+ («)} are asymptotically, identically distributed if one does not include the feedback custo-
mers in the queue length. This point was made in his paper [7].
REFERENCES
[1] Buck, R.C., Advanced Calculus, (McGraw-Hill, New York, 1956).
[2] Burke, P.J., "Proof of a Conjecture on the Interarrival-Time Distribution in an M/M/l
Queue with Feedback," IEEE Transactions on Communications, 575-576 (May 1976).
[3] Cooper, R.B., Introduction to Queueing Theory, (MacMillan, New York, 1972).
[4] d' Avignon, G.R..and R.L. Disney, "Queues with Instantaneous Feedback," Management
Science, 24, 168-180 (1977).
[5] Disney, R.L., R.L. Farrell, and P.R. deMorais, "Characterization of M/G/l Queues with
Renewal Departure Processes," Management Science, 19, 1222-1228 (1973).
[6] Disney, R.L., "Sojourn Times in M/G/l Queues with Instantaneous, Bernoulli Feedback,"
Proceedings Conference on Point Processes and Queueing Theory, Keszthely, Hungary
(September 1978).
[7] Foley, R.D., "On the Output Process of an M/M/l Queue with Feedback," Talk given at
San Francisco Meeting, Operations Research Society of America (May 1977).
[8] Jackson, J.R., "Networks of Waiting Lines," Operations Research, 5, 518-521 (1957).
[9] Melamed, B., "Characterizations of Poisson Traffic Streams in Jackson Queueing Net-
works," Advances in Applied Probability, 11, 422-438 (1979).
[10] Takacs, L., "A Single Server Queue with Feedback," Bell System Technical Journal, 42,
509-519 (1963).
AN INVENTORY MODEL WITH SEARCH
FOR BEST ORDERING PRICE*
Woodward-Clyde Consultants
San Francisco, California
ABSTRACT
This paper presents a single-item inventory model with deterministic
demand where the buyer is allowed to search for the most favorable price be-
fore deciding on the order quantity. In the beginning of each period, a sequen-
tial random sample can be taken from a known distribution and there is a fixed
cost per search. The decision maker is faced with the task of deciding when to
initiate and when to stop the search process, as well as determining the optimal
order quantity once the search process is terminated. The objective is to
minimize total expected costs while satisfying all demands on time. We
demonstrate that a set of critical numbers determine the optimal stopping and
ordering strategies. We present recursive expressions yielding the critical
numbers, as well as the minimal expected cost from the beginning of every
period to the end of the horizon.
1. INTRODUCTION
This research is an attempt to marry some aspects of search theory and optimal stopping
with inventory theory. Following the pioneering work of Stigler [11], [12], searching for the
lowest price is considered a basic feature of economic markets. By citing examples based on
real data, Stigler [11] asserted that prices change with varying frequency in all markets, and
unless a market is completely centralized, the buyer will not know for certain the prices that the
various sellers quote at any given time. This suggests that at any time there will be a frequency
distribution of the prices quoted by sellers. If the dispersion of price quotations by sellers is
large compared to the cost of search, it will pay— on average— to obtain price quotations from
several sellers before taking an "action." The vast literature on search theory (a survey of
which can be found in Lippman and McCall [8], DeGroot [5], and Rothschild [10]) is con-
cerned with rules that the searchers should follow when the "action" is accepting or rejecting a
price. Once the price has been accepted, the decision process terminates. In many dynamic
models, the action is more complicated. In inventory models, for example, the decision not
only involves accepting or rejecting an ordering price but how much to order, an action which
will affect the search and ordering policies in future periods. In this paper we study such a
model. We seek the best search and ordering policies for a simple dynamic inventory problem
with deterministic demands where, in the beginning of each period, the purchaser can search
for the lowest price before placing an order.
*This research was partially supported by the National Science Foundation through Grant NSF ENG74-13494 and the
Air Force Office of Scientific Research (AFOSR 72-2349C).
645
646 K GOLABI
Classical optimal search considers the following problem: A purchaser can take a sequen-
tial random sample X],X2, ... from a continuous distribution with a known distribution func-
tion F. There is a fixed cost s per observation. Suppose that if the decision maker stops the
sampling (search) process after the values X\ — x\, X2 = x2, . . ., Xn = x„ have been observed,
his cost is x„ + sn. Hence, the problem is to find a stopping rule which minimizes E{XN + sN)
where N indicates the random number of observations that are taken under a specified stopping
rule. It can be shown that, whether sampling is with or without recall, the optimal stopping
rule is characterized by a unique critical number v* (usually called the reservation price) so that
an optimal sampling rule is to continue sampling whenever an observed value exceeds v* and to
stop the process as soon as some observed value does not exceed v*. Various versions of this
problem have been studied by MacQueen and Miller [9], Derman and Sacks [6] and Chow and
Robbins [2], [3] among others.
The above search model can be visualized as a one period purchasing problem in which
one unit of some commodity has to be purchased at the beginning of the period. Now consider
a dynamic multiperiod version of this problem where a demand of one unit has to be satisfied
in each period and inventory holding cost is charged for items held over for use in subsequent
periods. As in the classical search problem, in the beginning of each period a sequential ran-
dom sample XhX2, ... can be taken from a distribution with known distribution function F,
but the decision process is not terminated as soon as an acceptable value is observed. The deci-
sion maker is faced with the task of deciding how much to order so as to minimize total
expected costs while satisfying all demands on time. When the inventory level is sufficient to
satisfy the immediate demand, he has also the burden of deciding whether to initiate search at
all. This multiperiod rriodel is the subject of our study in this paper.
In Section 2, we present the model. In Section 3, we give the optimal search policy and
in Section 4, the optimal ordering policy. We show the intuitive result that an optimal strategy
prescribes that search should be initiated only when the inventory level is zero. Furthermore,
we show that the reservation price property of the classical search problem still holds. That is,
when the inventory level is zero (and therefore search has to be initiated) and n periods remain
to the end of the problem, there exists a reservation price an such that a price should be
accepted if it does not exceed a„ and rejected otherwise. In Section 4, we show that once a
price has been accepted, a finite number of critical numbers specify the optimal strategy: The
critical numbers divide the interval [O.aJ into segments so that the interval in which the
accepted price falls determines the optimal order quantity. We give recursive expressions which
yield a„ as well as the minimal expected cost for any period to the end of the horizon. We will
also obtain expressions describing the critical numbers when the holding cost function is con-
vex.
2. THE MODEL
Consider a multi-period single-item inventory model in which a demand of one unit has
to be satisfied in the beginning of each period and an inventory holding cost is charged. In
each period, a sequential random sample X\,X2, ... of ordering prices can be generated from a
continuous distribution with known cumulative distribution function F(-), E(X\) < «>, and the
Xj's are mutually independent. The cost of generating each random price is s and there is no
limit on the number of observations which can be made in each period. After receiving a price,
the decision maker has to decide whether to accept that price or generate another offer. If he
accepts the offered price, he is faced with the decision of how much to order. When the inven-
tory level is sufficient to satisfy the immediate demand, he also has to decide whether to initiate
search at all. The objective is to minimize the total expected costs.
INVENTORY MODEL WITH SEARCH 647
We assume that the length N of the planning horizon is finite, initial inventory is zero,
backlogging of demand is not allowed, the cost of holding z units for one period, h(z), is non-
decreasing in z and h (0) = 0, the purchasing cost is linear in the quantity ordered, and only
integer quantities can be ordered. We also assume prices that are not accepted immediately are
lost; in view of our results, sampling with recall (of prices in the same period) extends no addi-
tional advantage over sampling without recall, and hence would not affect the search policy.
Note that when TV = 1, this model reduces to the classical search problem.
Let n be the number of periods remaining until the end of the horizon, z the inventory on
hand with n periods remaining and x the last price received. In each period, our state space
consists of numbers (z) and pairs (z,x) corresponding respectively to the state of the system
before a search is placed and the state when a search has been placed and an offer x has been
received. A policy for period n prescribes a search decision for state (z), and a reject-accept
and ordering decision for state (z,x). We assume that for each period an optimal policy exists.
Moreover, we restrict our attention to history-independent policies; that is, once the price x has
been rejected, we are in the same position as having not placed the search at all. Schematically,
remembering that demand in each period equals one, the period-state pairs correspond to each
other as follows:
Forz ^ 1:
(nz) Search , (nzx)AccQpi x and order an amount a (n-\ z+a-\)
^ ^(A7-1(Z-1)
and
/ -i qx aearcn ^ /■ ^ \ Accept x and order an amount a /-^-i a_jv
3. OPTIMAL SEARCH POLICY
In this section, we present the optimal search policy. We show that search should only be
initiated when the inventory level is zero, and prove that in each period a single reservation
price determines the stopping rule. We also give a recursive expression which describes the
sequence of reservation prices.
To being, define
Vn(z,x) = the minimal (conditional) expected cost during the last n periods when the
inventory level with n periods remaining is z and the last price offered is x.
v„(z) = the minimal expected cost during the last n periods before the decision to
search for an offer is made, and when the inventory level with n periods
remaining is z.
un{z) = the minimal expected cost during the last n periods after the decision to
search for an offer in this period has been made, and when the inventory
level with n periods remaining is z.
w„(z) = the minimal expected cost during the last n periods after the decision not to
search for an offer in this period has been made, and when the inventory
level with n periods remaining is z, z ^ 1 .
H{z) = the total holding cost of z units to be used in z consecutive periods.
Hence, we will have the following relationships:
[ax + h(z + a - 1) + v„_,(z + a - 1)]},
1), z ^ 1,
(1)
v„(z) = min[«„(z), wn(z)]
(2)
V„(z,x) = min{v„(z), min
a€[1.2 1
(3)
u„(z) = s + Ex[V„{z,x)],
(4)
w„{z) = hiz - 1) + v„_,(z -
and
(5)
Define
//(z)=£//(z- /)= £ //(/
(6a)
In[x,a] = ax + h(a - l) + v,
and
(6b)
I Ax) = min I.,(x,a),
a€[1.2 »]
-l(fl- 1)
and let an(x) be the minimizing value of a in (6a), that is,
(6c) In(x) = lAx,an(x)}.
The quantity In(x) is the minimal expected cost attainable during the last n periods when the
inventory level with n periods remaining is zero and it has been decided to accept x, the last
price offered.
At this point it is natural to ask whether when n periods remain, there exists a single criti-
cal price an which dictates the acceptance or rejection of a price x when the inventory level is
zero. In other words, is there ana„ such that it is optimal to accept the price x (and order a
positive amount) if jc < a„ and to continue the search if x > <xn. That this is indeed the case,
is verified in Theorem 1.
Define
(7) q„e/;'[v„(0)],
and the sequence {An}"=0 by the following recursion:
(8a) ^o=0
and
(8b) A„F(a„) = 5 + C" min [ax + H(a) + An_a]dF(x) for n > 1.
J0 a6[i,2 „]
INVENTORY MODEL WITH SEARCH
649
We will show later that an exists and that A„ equals v„(0), so that an = I~HAn). These pro-
perties are exploited to verify that an optimal policy prescribes that search be initiated, and ord-
ers be placed, only when the inventory level is zero. Furthermore, we will show that if the set
of prices at which it is optimal to order one unit is nonempty, a„ = An — An-\ so that Equation
(8b) can be written as
(9)
A„FU„ - Att-0 = s -
[ax + H(a) +An_a]dF(x),
enabling us to obtain the minimal expected cost from the beginning of any period to the end of
the horizon by finding [Aj}j!Lo, the unique set of solutions to Equation (9).
THEOREM 1: If thexV inventory level with n periods remaining is zero, it is optimal to
continue the search if x, the last price offered, is greater than an and accept the price if
x < a„, where n = 1, 2, . . . , N.
PROOF: Clearly, I„{x,a) is continuous in x for each n and a, and therefore, In(x) is a
continuous function of x. Furthermore, for all positive numbers e,
I„{x + e) = In[x + e, an(x + c)] > In[x,an{x + e)] > I„[x,a„(x)] = /„(*),
and hence In{x) is strictly increasing in x. Let a„(v) be such that I„[an(y)] = v, i.e.,
a„(y) = /„-1(y). Since
v„(0) ^ v„_!(0) ^ min
fl€[l,2,...,
[h (a- l) + v„_,(a- l)] = /„(0),
it follows that an = a„[v„(0)] exists and, as In(x) is strictly increasing in x, it is unique (see
Figure 1). The first inequality of the above expression follows from the fact that for the n — 1
period problem we can always follow the optimal policy for the n period problem, so that at
each stage m, n — l^w^l, we would adopt the action prescribed by the n period optimal
policy for stage m + 1. Thus, the expected cost for the n - 1 period problem under this pol-
icy, v^_!(0), would be equal to the expected cost of the first n - 1 periods of the n period prob-
lem, and hence v^^O) < v„(0). Since v„_!(0) < v^^O), it follows that v„(0) is nondecreas-
ing in n.
W
vn(0)
ln(0)
<*n
650 K. GOLABI
From (2) and (6) we have
K„(0,x) = min [v„(0), /„(*)].
If x < a„, then In(x) < v„(0) so that V„(0,x) = In(x) and search terminates. If x > a„,
then I„(x) > v„(0) so that Vn(0,x) = v„(0), in which case it is optimal to continue the search.
Q.E.D.
Thus, when the inventory level is zero, a single critical number determines whether a
price should be accepted. We are also interested in finding the optimal strategy when the
inventory level is positive. It seems intuitive that if the immediate demand can be satisfied by
the current inventory, it would be best to postpone the search— since it is possible to incur the
same amount of expected search cost in a later period while saving on the holding cost. The
next result, the proof of which is given in the Appendix, verifies this observation. In addition,
it shows that the expected cost from any period k in which the inventory level is zero to the
end of the horizon equals Ak. Thus, the expected cost from any period can be determined by
computing the sequence [A„) from Equation (8b).
THEOREM 2. Under the assumptions of the model, for all k, k = 1, 2, ... , N,
(a) vk(0) = Ak
. (b) vfc(z) = H(z) + \k-2(0) for 1 < z ^ k.
Theorem 2 verifies that search should be initiated only when the inventory level is zero,
and Theorem 1 gives a rule for accepting or rejecting an offered price once search is initiated.
These two results however, do not completely specify the optimal strategy. Given that an
acceptable price is received, we would like to know how much should be purchased at that
price. This question is investigated in the next section.
4. OPTIMAL ORDERING POLICY
In this section we present the optimal ordering policy once an acceptable price has been
received. In Corollary 3 we show that a nonincreasing sequence of critical numbers characterize
the optimal order quantity. In other words, once a price is received that is less than the reser-
vation price for that period, the interval in which the offered price falls determines the quantity
that should be ordered at that price. In Theorem 5 we obtain expressions which describe these
critical numbers when the holding cost function is convex.
Before presenting the next result we note that when n periods remain, the inventory level
is zero, and an acceptable price has been received, the optimal order quantity is equal to a„(x).
To see this, note that
K„(0,x) = min[(v„(0), /„(*)]
by (2) and (6). This fact coupled with Theorem 1 yields Vn(0,x) = In(x) whenever x < a„.
Finally, 'since
(10) /„W = /„b„W]= min [ax + h(a-l) + w„-i(a-l)]t
it follows that ordering a„(x) minimizes the expected cost attainable during the last n periods
when the inventory level is zero and x < a„. Note also that by Theorem 2(b), Equation (10)
can be written as
INVENTORY MODEL WITH SEARCH 651
(11) /„(*)= min [ax + H(a) + A„.a].
COROLLARY 3: If n periods remain, the inventory level is zero, and an acceptable price
has been received, then the optimal order quantity is nonincreasing in the price offered, i.e.,
a„(x') < a„(x) whenever x' > x, n = 1,2, ... , N. Consequently, a nonincreasing sequence
of critical numbers [Bj{n)}-L\ characterize the optimal order quantity. Specifically, it is optimal
to order k units whenever Bk(n) < x < Bk_x{n).
PROOF: From (6c) and (11), we have
I„(x) = I„bc,a„(x)] = an(x) ■ x + H[an(x)] + An_an(x)
< In[x,an(x')] = aa(x') • x + H[a„(x')] + An^ajxV
giving
(12) x[a„(x') - an{x)] > An_Qn{x) - A^w + H[a„(x)] - H[an(x')].
If an{x') > an(x), then (12) implies
x'[an(x') - a„(x)] > An_anix) - ^.Crf + H[an(x)] - H[an(x')},
which yields
In(x') = an{x') ■ x' + An^an(x') + H[a„(x')] > an(x) • x' + An_an{x)
+ H[an(x)] = In[x',an(x)},
contradicting the fact that an(x') is optimal when x' is offered. Q.E.D.
Intuitively, we would expect that when an offered price equals the critical number an, we
would be indifferent between ordering one unit and not ordering at all. If this were indeed the
case, the expected cost when the price is rejected, v„(0), would be equal to a„ + v„_!(0) yield-
ing a„ = A„ — An_\. This result could then be used to obtain a simple expression for the
Bk(nVs when h (•) is convex. As we will show in Lemma 4, the above result holds if the set of
prices at which it is optimal to order one unit is nonempty. Unfortunately, as seen from the
following example, this is not always the case.
EXAMPLE 1. Let n = 5, 5 = 5, h (z) = 0 for all z and the price distribution be such that
P(X = 2) = 1 - e, and P(a < X < b) =^-{b- a) for 0 < a < b < 4, where 2 is
4
excluded from all intervals and e is an arbitrary small number. Suppose the offered price in the
beginning of the fifth period is 3.
The expected cost of rejecting the offered price is (approximately)
5 + 2x 5= 15,
as one would pay the search cost of 5 and almost definitely receive the price of 2, at which one
would order 5 units. However, the expected cost of ordering / units, i < 4, is (approximately)
3/ + 5 + 2(5- /)= 15 + /,
while the cost of ordering 5 units is 15. Hence, we would be indifferent between not ordering
and ordering at x = 3, which implies that a5 = 3.
Since at x = 3 we order 5 units, any price above 3 is rejected, and the optimal order quan-
tity a„(x) is nonincreasing in x, it follows that [x :a„(x) = 1] is empty.
LEMMA 4: If [x : a„(x) = 1] is nonempty, then an = An — A„_x.
PROOF: Let x be the largest x such that an{x) = 1. By Theorem 1, a „ is the highest
price at which it is optimal to order a positive quantity. Therefore, 3c < a„. Consequently, we
can conclude from Corollary 3 that an(an) ^ 1, but an(a„) is positive so that a„{a„) = 1.
From Theorem 2 and Equations (7) and (11), we have
= (a„(a„) • a„ + H[an(an)] + 4,-a>n)) = <*n + H{\) + A„.h
which yields a„ = An - A„_x. Q.E.D.
Whereas we cannot determine in advance the conditions under which Lemma 4 would
hold, we can proceed by assuming that the lemma holds, and determine the sequence [4„}!Lp
that satisfies Equation (9). We then can obtain {<*„} from an = I~x{An). If {«„} and \An) also
satisfy Equation (8b), by uniqueness of the solution, an is indeed equal to A„ - A„_x.
It is interesting to note that contrary to what one might expect, an is not monotone in n.
Before Theorem 5, we give examples where an is not monotone irrespective of whether
[x : an (x) = 1] is empty or not
EXAMPLE 2: (a) Consider again Example 1. Since we would almost definitely receive
the price of 2 after the first search, we have
v„(0) = s+ min [ax + H(a) + vfl_fl(0)].
Thus,
Vl(0) = 5 + 2=7
v2(0) = 5 + min (2 + 7,4) = 9.
From v„(0) = I„(a„), we have ax = vj(0) = 7 and
9= min (a2 + 7,2 a2)
yielding a2 = 4.5. As shown earlier, a5 = 3. Therefore, an is not monotone in n.
(b) We note that an is not necessarily monotone even if [x :an(x) = 1] is nonempty. Con-
sider the case where the price distribution is the same as Example 1 . However, there is a hold-
ing cost of 1 per unit per period and 5 = 2. Then, H{\) — 0, H(2) = 1, HO) = 3 and
7/(4) - 6 and
Vl(0) = 2 + 2=4
v2(0) = 2 + min [4+ 1,2 + 4] = 7
v3(0) = 2 + min[6 + 3,4+ 1 + 4,2 + 7]= 11
v4(0) = 2 + min [8 + 6,6 + 3 + 4,4+ 1 + 7,2+ 11]= 14.
INVENTORY MODEL WITH SEARCH 653
From v„(0) = /„(«„), we have
4-ai
7 = min [2a2 + l.«2 + 4]
11 = min [3a3 + 3,2a3 + 1 + 4,a3 + 7]
14= min [4a 4 + 6, 3a4 + 3 + 4,2a4 + 1 + 7,a4 + 11]
yielding
a\ = 4, a2 = 3, a3 =4, a4 = 3.
Note that in this example, an{an) = 1 for 1 < n < 4 and the condition for Lemma 4 holds. It
can be easily verified that a„ = v„(0) - v^CO) for all 1 < n < 4.
THEOREM 5: If the condition for Lemma 4 holds and if the holding cost function h(-)
is convex, then
(13) Bk(n) = an_k- h(k), where 1 < k < n.
PROOF: We have to show that
(a) The RHS of (13) is nonincreasing in k.
(b) It is optimal to order k units if x, the price offered, satisfies
(14) otn-k- h(k) < x < «„_<*_,)- h(k - 1).
To show (a) , we note that
An_k+X = v„_£+1(0) = /„_*+, (a„_fc+1) = min [aan-k+x
+ H(a) + 4,-fc+i-al ^ 2a„_^+1 + 7/(2) + ^„_^_!
= lUn-k+i - A„.k) + h(l) + A„-k-V,
where the first equality follows from Theorem 2, the second from (7), the third from (11) and
the last from Lemma 4. Thus, by convexity of h (•),
h(k)- h(k- \) > h(l) > An_k - An-k.x - An_k+X + An_k
= <*n-k - <*n-k+l
and, therefore, (a) is true.
To show (b), suppose x is such that (14) holds. We show that I„(x,k — j) <
I„(x,k - j - I) for each j ^ 0, and therefore ordering k units is at least as good as ordering
any amount less than k. Suppose In(x,k - j) > I„(x,k - j - 1). Then
(k-j- l)x + ^-(jfc-y-1) + H(k -j-\)<(k- j)X + An_(k_j) + H(k - j)
which yields
x > Amfr-j-Q - An_(k_j) - h(k - j - 1)
= an-(k-j-\) - h (k - j - 1) ^ a„-0fc-i) - h(k - 1),
where the last inequality follows from (a). This contradicts the right inequality of (14). There-
fore, /„ (x,k - j) < In(x,k- j - 1) .
I„(x,k + j) < I„(x,k+ J + 1) for each j > 0 by a similar proof.
Hence, it is optimal to order k units whenever (14) holds. Q.E.D.
5. REMARKS
The purpose of this study has been to investigate optimal search policies in the context of
a sequential model. The underlying inventory model has been chosen as a rather simple one.
There are no setup costs involved and the demand equals one unit in each period. It would be
interesting to investigate more general problems. We suspect that both the reservation price
property of Theorem 1 and the Wagner- Whitin [13] type result of Theorem 2 (order only when
current inventory level is zero) would still hold for models with setup costs and arbitrary deter-
ministic demands. The optimal policy would be a function of setup costs as well as the holding
cost and price distribution. The results should also hold when the price distributions are non-
stationary. Given that the initial inventory is zero, the ordering policy will be such that there is
no inventory in the beginning of periods with favorable price distributions.
Another interesting extension is the case wherein the search process is adaptive. The
searcher does not know the exact distribution of price; the price offer is used not only as an
opportunity to order at that price but also as a piece of information to update the prior distribu-
tion. When the distribution of prices is not known exactly, the form of the optimal policy is
not obvious. As Rothschild [10] points out, the reservation price property of Theorem 1 would
not necessarily hold even for a one period problem. Rothschild presents the following example.
Suppose there are three prices, $1, $2, and $3, and that the cost of search is $0.01. Prior
beliefs admit the possibility of only two distributions of prices. Either all prices are $3 or they
are distributed between $1 and $2 in the proportions 99 to 1. A man with these beliefs should
accept a price of $3 (as this is a signal that no lower prices are to be had) and reject a quote of
$2 (which indicates that the likelihood that a much better price will be observed on another
draw is high).
However, when the distribution is a member of certain families of distributions but has
one or more unknown parameters, Rothschild [10], DeGroot [5] and Albright [1] have shown
that the reservation price property holds for the one-period problem. We conjecture that when
the distribution of price is stationary but is not known exactly, search should be initiated only
when the inventory level is zero. If this is the case and the distribution belongs to one of the
families of distributions studied by Rothschild [10] and Albright [1], then the reservation price
property as well as the ordering policy presented in Section 4 should still hold.
ACKNOWLEDGMENTS
This paper is essentially Chapter 3 of the author's dissertation (1976) at the University of
California, Los Angeles. The author expresses his appreciation to Professor Steven Lippman
for his guidance and encouragement. He also appreciates several helpful comments by Profes-
sor Sheldon Ross and the referee.
BIBLIOGRAPHY
[1] Albright, C.S., "A Bayesian Approach to a Generalized House Selling Problem," Manage-
ment Science 24, 432-440 (1977).
[2] Chow, Y.S. and H. Robbins, "A Martingale System Theorem and Applications," Proceedings
of the 4th Berkeley Symposium on Mathematical Statistics and Probability, University of Cali-
fornia Press, Berkeley, California (1961).
INVENTORY MODEL WITH SEARCH 655
[3] Chow, Y.S. and H. Robbins, "On Values Associated with a Stochastic Sequence," Proceed-
ings of the 5th Berkeley Symposium on Mathematical Statistics and Probability, University
of California Press, Berkeley, California (1967).
[4] DeGroot, M.H., "Some Problems of Optimal Stopping," Journal of the Royal Statistical So-
ciety 30, 108-122 (1968).
[5] DeGroot, M.H., Optimal Statistical Decisions (McGraw-Hill, 1970).
[6] Derman, C. and J. Sacks, "Replacement of Periodically Inspected Equipment," Naval
Research Logistics Quarterly 7, 597-607 (1960).
[7] Golabi, K., "Optimal Inventory and Search Policies with Random Ordering Costs," Work-
ing Paper No. 252, Western Management Science Institute, University of California,
Los Angeles (1976).
[8] Lippman, S.A. and J.J. McCall, "The Economics of Job Search: A Survey," Economic In-
quiry 14, 155-189 (1976).
[9] MacQueen, J.B. and R.G. Miller, Jr., "Optimal Persistence Policies," Operations Research
8, 362-380 (1960).
[10] Rothschild, M., "Models of Market Organization with Imperfect Information: A Survey,"
Journal of Political Economy 81, 1283-1308 (1973).
[11] Stigler, G.J., "The Economics of Information," Journal of Political Economy 69, 213-225
(1961).
[12] Stigler, G.J., "Information in the Labor Market," Journal of Political Economy 70, 94-104
(1962).
[13] Wagner, H.M. and T.M. Whitin, "Dynamic Version of the Economic Lot Size Model,"
Management Science 5, 89-96 (1958).
APPENDIX
THEOREM 2: Under the assumptions of the model, for all k, k = 1,2, ... , N,
(a) v*(0) = Ak
(b) vA.(z) = H{z) + vfc_z(0) for 1 < z < k.
Consequently, the search process is initiated only when the inventory level is zero.
Before proving Theorem 2, we establish two elementary facts.
LEMMA A: For any two positive integers /and j, H{i + j) > //(/) + H{j).
PROOF:
HU + j) = ' £ h ik) = £ h (k) + ' £ h (Jd > J h (k) + £ h (k)
A=l k=\ k=i k=\ k=\
= HU) + Hij). Q.E.D
LEMMA B: The integral J " [y — ln(x)] dF(x) = G„(y) is strictly increasing in y,
continuous, and unbounded above.
PROOF: Since In[a„(y)] = y and /„(*) is strictly increasing in x, it follows that an{y) is
strictly increasing in y. Hence, G„(y) is strictly increasing, continuous (as Fis continuous) and
unbounded above. OFF)
PROOF OF THEOREM 2: The proof is by induction on k. From Equations (1), (3), (2)
and (6), we have
(A-l) vA(0) = uk(0) = s + Ex[Vk(0,x)}
= 5 + E min |vA(0), min [ax + h(a - 1) + vfc_, {a - \)]\
= 5 + E min [vk(0), Ik(x)].
For A: = 1, (b) is obvious. To show (a), note that by (6), /|(.v) = x. Next, from (A-l)
we have
rv,(0) r~
v,(0) = s + E min [v,(0),x] = 5 + Jq xdE(x) + Jv (Q) v,(0) dF{x),
from which we obtain
-v (0)
(A-2) v,(0)F[v,(0)] = s + Jo xdF(x).
(Note the close connection between V|(0) and the maximizing price in the house selling prob-
lem.) In order to determine whether v,(0) is the unique solution to (A-2), note that it is
equivalent to verify that s = J (y - x) dF(x) = G|(v) has a unique solution. The latter
result follows from Lemma B.
From (7) we have /i(c*i) = v,(0) and therefore a, = v,(0). Thus, (A-2) becomes
a,F(a,) = 5 + J ' xdF(x),
which coupled with (8b) for n = 1, gives A | = ot\ = v,(0) so that (a) holds for k = 1.
Assume (a) and (b) hold for k = 1,2, ... , n — 1. We show that the theorem holds for
k = n.
From (A-l), we have
v„(0) = s + E min [v„(0), I„(x)]
= s + fj" In(x) dFix) + £~ v„(0WFU)
= [F(a„)]-] \s + J"" [ min [ax + h(a - 1) + v/;_,(o - l)]|rfF0c)
= [F(a„))"'L + Jp ( min [ax + h'(a - 1) + H(a - 1) + v„_o(0)]] dF(x)\
= [F(a„)]-^ \s + Jo"" ( min [ax + H(a) + A„_a]\ dF(x)\
- [F(a„)]-*AnF(a„)
= A„f
where the second equality follows from Theorem 1, the third from a simple rearrangement of
terms, the fourth and fifth equalities from the induction hypothesis and the sixth equality from
(8b). Therefore (a) is true for k = n.
INVENTORY MODEL WITH SEARCH
657
Since we are assuming that (b) holds for k = n — 1, it follows that
AnF(an) = 5 + C" I„(x)dF(x),
which gives
(A-4) fj" [A„ - Itl(x)] dF(x) = s.
We note that by (4) and the induction hypothesis, for 1 < z < n
w„(z) = h(z - 1) + v„_,(z - 1) = h(z - 1) + H(z - 1) + vn_z(0)
= H(z) +v„_z(0),
and therefore to prove (b) for k = n, it suffices to show v„(z) = w„(z) whenever z > 1. That
is, we need to show u„{z) ^ //(z) + v„_-(0) whenever z ^ 1.
We can write
«.(z) = s + E mi
= s + E m
= 5 + E m
^ 5 + E mi
= s + E mi
where the first equality follows from (3) and (2), the second from (1), the fourth from induc-
tion hypothesis and the last equality from the induction hypothesis and (A-3). The inequality
follows from Lemma A. Hence,
(A-5) y = u„{z) - H(z) > s + E min \u„(z) - H(z), min [vn_r(0),/„--U)] .
If y were less than v„_.(0), then from (A-5) we would have
y ^ s + E min [-y,/„_-(x)]
nj«„(z),//(z) + min [ax + H{a) + A„_:_a]{
n|w„(z),//(z) + minU„_r, min (ax + H(a) + A„_:_a)][
n \u„(z),H(z) + min [v„_r(0),/„_zU)] ,
658 K CiOLABI
giving
fQa"~: \ [y - I„-:(x))dF(x) > s = fQa"~: [A„_z - I„_Ax)]dF(x),
where the equality follows from (A-4)-. Hence,
G„.z(y) > G„_:(A„-:)= G„_r[v„_r(0)],
contradicting Lemma B. Therefore, y ^ v„_2(0), which completes the induction argument.
Q.E.D.
THE UNITED STATES COAST GUARD COMPUTER-ASSISTED SEARCH
PLANNING SYSTEM (CASP)*
Henry R. Richardson
Daniel H. Wagner, Associates
Paoli, Pennsylvania
Joseph H. Discenza**
U.S. Coast Guard
ABSTRACT
This paper provides an overview of the Compuler-Assisied Search Planning
(CASP) system developed for the United Stales Coast Guard. The CASP in-
formation processing methodology is based upon Monte Carlo simulation to
obtain an initial probability distribution for target location and to update this
distribution to account for drift due to currents and winds. A multiple scenario
approach is employed to generate the initial probability distribution. Bayesian
updating is used to reflect negative information obtained from unsuccessful
search. The principal output of the CASP system is a sequence of probability
"maps" which display the current target location probability distributions
throughout the time period of interest. CASP also provides guidance for allo-
cating search effort based upon optimal search theory.
1. INTRODUCTION
This paper provides an overview of the computer-assisted search planning (CASP) system
developed for the United States Coast Guard to assist its search and rescue (SAR) operations.
The system resides on a CDC 3300 located in Washington, D.C., and can be used by all USCG
Rescue Coordination Centers (RCCs) in the continental United States and Hawaii via remote
access terminals.
The Coast Guard is engaged daily in search and rescue missions which range from simple
to complex. The amount of information available to predict the position of the search target
ranges from extremely good to almost no information at all. The process of planning, com-
manding, and evaluating these searches takes place in Rescue Coordination Centers (RCCs)
located throughout the United States in major coastal cities.
The entire planning process begins with the awareness that a distress on the water may
exist. This awareness usually results from a telephone call from a friend or relative or from a
radio communication from the boat or vessel itself.
This work was supported in part by USCG Contract DOT-CG-32489-A and ONR Contraci
**The opinions or assertions contained herein are the private ones of the author and are r
or reflecting the view of the Commandant or the Coast Guard at large
660 H.R. RICHARDSON AND J.H DISCENZA
Next all available information has to be evaluated to decide whether or not to begin a
search, and what level of effort is required given the search begins. At this point a great deal of
effort goes into deciding where the distress incident occurred. This might be considered the
first phase of planning.
The next phase involves computing where the search target will be when the first search
units arrive on scene. Among other things, this requires the prediction of ocean drift and wind
velocity and the estimation of uncertainties in these predictions.
The next questions pertain to the effort allocation process— how much effort must be
expended and in what areas? Prior to the advent of computer search programs, SAR planners
relied upon various rules of thumb as presented in the National Search and Rescue Manual
[11]. Simplicity was necessary to facilitate hand computation, but at the same time prevented
adequate treatment of the many sources of uncertainty which characterize a SAR incident.
The search phase is the actual deployment of aircraft and vessels, the conduct of preset
search patterns, and the report of results back to the RCC.
If the search is unsuccessful for that day, then the results must be reevaluated and a new
search planned for the following day.
This process continues until the target is found or until the search is terminated. In brief
(and in slightly more technical terms), the planning phases are as follows:
(1) Determine the target location probability distribution at the time of the distress
incident.
(2) Update the target location probability distribution to account for target motion prior to
the earliest possible arrival of a search unit on-scene.
(3) Determine the optimal allocation of search effort, and estimate the expected amount
of search effort required to find the target.
(4) Execute the search.
(5) If the search is unsuccessful, evaluate the results and update the target location proba-
bility distribution to account for this negative information.
(6) Repeat the planning procedures in Steps (2) through (5) until the target is found or
the search is terminated.
These planning phases are illustrated in the CASP case example given in Section 3.
The first efforts at computerization concentrated on the target location prediction process.
Oceanographic models were used to compute drift and to estimate target position. The Mon-
terey Search Planning Program and the Coast Guard's own Search and Rescue Planning Sys-
tem, SARP, represented early computer assisted search efforts. Even today, in cases where the
information available makes the planning straightforward, the SARP program does nicely.
In 1970, the Office of Research and Development in Washington funded development of
a more comprehensive approach to search planning based in part on lessons learned in the
Mediterranean H-bomb search in 1966 (Richardson [5]) and in the Scorpion search in 1968
COASTGUARD COMPUTER-ASSISTED SEARCH (CASP) 661
(Richardson and Stone [6]). In 1972, the CASP system was delivered to the Operations
Analysis Branch of Commander Atlantic Area in New York for evaluation, implementation,
and training. The system was made operational early in 1974.
CASP is now in use in 11 Coast Guard rescue centers. In addition, CASP has been used
at the Air Force Central Rescue Headquarters at Scott AFB, Illinois, to help plan and coordi-
nate search missions for lost airplanes within the continental United States. A modification of
the CASP system has also been provided to the Canadians for inland SAR planning.
At the present time, the use of CASP is limited to open ocean searches. Even though
these searches represent but a small percentage of the total U.S. Coast Guard search operations,
CASP has been credited with saving over a dozen lives.
Section 2 provides a description of the CASP methodology. Section 3 illustrates the use
of CASP in an actual SAR incident involving the 1976 sinking of the sailing vessel S/V Spirit in
the Pacific, and Section 4 describes CASP training.
2. CASP METHODOLOGY
The CASP information processing methodology is based upon Monte Carlo simulation to
obtain an initial probability distribution for target location and to update this distribution to
account for drift due to currents and winds. A multiple scenario approach is employed to gen-
erate the initial probability distribution. In the sense used here, a scenario is a hypothetical
description of the distress incident which provides quantitative inputs for the CASP programs.
Bayesian updating is used to reflect negative information obtained from unsuccessful search.
The principal output of the CASP system is a sequence of probability "maps" which
display the current target location probability distributions throughout the time period of
interest. CASP also provides guidance for allocating search effort based upon optimal search
theory.
The CASP system is composed of a number of different programs, each designed for a
different information processing function. The program components are MAP, POSITION,
AREA, TRACKLINE, COMBINATION, DRIFT, RECTANGLE, PATH, and MULTI; the
functions are as follows:
(1) display the probability maps (MAP),
(2) generate an initial distribution of target location at the time of distress (POSITION,
AREA, TRACKLINE, and COMBINATION),
(3) update the target location probability distributions for motion subsequent to the time
of distress (DRIFT),
(4) update the target location probability distributions for negative search results and com-
pute the cumulative detection probability (RECTANGLE and PATH), and
(5) calculate optimal allocations of search effort (MAP and MULTI).
These programs will be described below following presentation of an overview of the general
system design.
662 H.R. RICHARDSON AND J .11. D1SCENZA
CASP System Design
The CASP system design was motivated by a desire to provide a highly realistic probabilis-
tic description for the target's location at the time of the distress incident and for the target's
substantial motion. In view of the success achieved in the Mediterranean H-bomb search [12]
in 1966, and in the Scorpion search [5] in 1968, it seemed evident that a Bayesian approach
would provide a practical method for incorporating information gained from unsuccessful
search.
Target motion modeling posed a more difficult problem. Models which were amenable to
an "analytic" approach were not flexible enough to give a good representation of the search
facts. For example, Gaussian motion processes (or mixtures of Gaussian processes) were unsa-
tisfactory in cases where the search facts required a uniform or annular shaped target location
probability density. Markov chains based on transitions among search grid cells were unsatis-
factory in cases where one desired to change the grid in the course of an operation. In general,
these models tended to force the facts to fit the mathematics to an undesirable extent.
It was also desired to develop a modular system so that additional features and improve-
ments could be made as time went on. In order to gain the confidence of the users, the system
had to be simple to understand and require a minimum of unfamiliar inputs. The design which
seemed best suited in view of the above considerations is a hybrid approach which uses Monte
Carlo to simulate target motion and analytic methods to compute detection probabilities.
A motivation for use of Monte Carlo was the recognition that computation of the poste-
rior target location probability distribution can be viewed as the numerical evaluation of a mul-
tivariate integral of high dimensionality. In such cases (i.e., high dimensionality), classical
numerical integration techniques perform poorly (see, for example, Shreider [7]) especially
when the integrands can have jump discontinuities and are not of a simple analytic form.
These problems are typical of CASP applications. Discontinuities occur when the "target"
moves into a region where search effort is concentrated, and the joint probability density for
target position at several specified times during the search is a very complicated function.
The underlying structure of CASP is a Markov process, with a three-dimensional state
space consisting of points (X, Y, $). The variables Zand Y denote latitude and longitude and
<E> denotes search failure probability. For j= 1, ... , 7, the yth Monte Carlo replication
(Xj, Yj, $>j) represents the target's current position (time is implicit) together with the cumula-
tive probability of search failure for that particular target replication computed for its entire his-
tory. Target motion is assumed to be Markovian and successive increments of search are
assumed to be statistically independent. Thus {Xjt Yj, <!>,) completely describes the state of the
yth target replication at a given moment.
Figure 1 provides a schematic diagram for the operation of the CASP system. All of the
programs mentioned will be discussed individually in subsequent subsections. The first step is
to construct a file (called the "target-triple file") consisting of samples from the target location
probability distribution at the time of the distress incident. This file is stored on computer disc
and processed sequentially by various programs.
These initial points {Xr Yr 1) have failure probabilities <& , ■ = 1, since no search has yet
been carried out. The target positions (Xr Yj) are sampled from a probability density function
Fof the form
COAST GUARD COMPUTER-ASSISTED SEARCH (CASP)
Scenario Generation
:) Update for Negative Search Results
Figure 1. Casp syste
f= E *kfk,
where fk is the density corresponding to the kth "scenario," and wk > 0 is the scenario's subjec-
Ik |
Monte Carlo samples from a probability density F are obtained by first using one of the
"generation programs" POSITION, AREA, or TRACKLINE. Averages of densities of different
types are obtained by forming preliminary target triple files with two or more "generation" pro-
grams and then combining them with the program COMBINATION. The construction of the
prior target location probability distribution is shown schematically in Figure 1(a).
Updates for target motion (Figure Kb)) or to account for negative search results (Figure
1(c)) are carried out by reading the "old" target triple file from disc into the appropriate pro-
gram and outputting a "new" target triple file. When program DRIFT is used (Figure Kb)), the
values of Xj and Y} are modified, but the value of 4>; remains unchanged. For an update for
664
H.R. RICHARDSON AND J.H. DISCENZA
negative search results, the file is first updated for motion by use of program DRIFT. The tar-
get triple file is frozen at the mid-search time and then modified by RECTANGLE or PATH.
These programs modify <!>, by use of Bayes' theorem but the position variables Xj and Yj
remain the same since motion is frozen.
The probability distributions and optimal allocations of search effort are displayed using
program MAP or MULTI (Figure 1(d)). In both cases, this is a read-only operation, and the
target triple file is not modified.
Display
The MAP program displays the target location probability distributions in a two dimen-
sional format. Figure 2 shows an example of a probability map corresponding to an actual SAR
case. The geographical region is divided into cells oriented north-south and east-west and the
target location probabilities* for each cell are multiplied by 10,000 and displayed. Thus, the
number 1800 in a cell indicates that the target location probability is .18. Equal probability con-
tours are usually sketched to make it easier to visualize the probability distribution.
Figure 2. Target location probability distribution probabilities are multiplied at 10,000 and truncated
A "quick map" in which symbols are used to represent ranges of probabilities can also be
output. The quick map provides a compact version of the probability distribution which is suit-
able for a quick appraisal of the search situation and is convenient for inclusion in after-action
reports.
Finally, MAP can output an ordered list of the highest probability cells and the amount of
effort to be placed in each cell in order to maximize detection probability. More will be said
about search optimization in the last subsection.
The format implies higher accuracy than is warranted in view of the Monte Carlo procedures employed.
COAST GUARD COMPUTER-ASSISTED SEARCH (CASP) 665
Initial Target Location Probability Distribution
The initial target location probability distribution is constructed from "building block dis-
tributions" using a weighted scenario approach. The individual building block distributions are
generated by the use of one or more of the programs POSITION, TRACKLINE, and AREA.
Program COMBINE is used to combine the outputs of the individual "generation" programs.
In most SAR cases, there is scant information available about the target's position at the
time of distress. Sometimes, for example, a fisherman simply is reported overdue at the end of
a day. He may have been planning to fish in one of several fishing grounds but did not make
his precise intentions known.
In other cases, more information is available. For example, it might be known that a
vacationer was intending to sail from one marina to another but never arrived at the intended
destination. In some cases, it might also be known that there was bad weather along the
intended route. This would make some positions along track more likely for a distress than
others.
In order to encourage inclusion of diverse possibilities in these scenarios, it is a recom-
mended practice for two or three search planners to work out the details together. The
remainder of this subsection will describe the programs POSITION, AREA, and TRACKLINE
which are used to simulate the scenarios and generate the initial target location probability dis-
tribution.
Position. A POSITION scenario has two parts, an initial position and a subsequent dis-
placement. POSITION can be used to generate a weighted average of as many as ten scenarios.
The initial position probability distribution is modeled as a bivariate normal distribution,
and the displacement is modeled as a distribution over an annular sector. In the latter distribu-
tion, the angle and distance random variables are assumed to be independent and uniformly
distributed between minimum and maximum values input by the user. The displacement distri-
bution is useful, for example, in cases where the initial position corresponds to the last fix on
the target and where one can estimate the course and speed of subsequent movement prior to
the occurrence of the distress incident.
The displacement option can also be used in cases involving a "bail out" where it can
describe the parachute drift. The amount of displacement in this case will depend upon the
altitude of the aircraft and the prevailing winds at the time. Since these factors are rarely
known precisely, the capability to "randomize" direction and distance is an important feature.
Area. The second generation program is AREA. This program is used to generate an ini-
tial target location probability distribution in cases where a general region can be postulated for
the location of the distress incident but where a normal distribution simulated by POSITION
would be a poor representation of the uncertainty. Each scenario for program AREA deter-
mines a uniform probability density within a convex polygon. AREA might be used, for exam-
ple, when a lost fisherman's usual fishing ground is known from discussions with friends and
relatives. As with POSITION, AREA can generate a weighted average of 10 scenarios.
666
H.R. RICHARDSON AND J.H. DISCENZA
Trackline. The third and last generation program is TRACKLINE. This program is the
most complex of the generation programs and is used when target track information is available
from a float plan or some other source. TRACKLINE creates a probability distribution about a
base track. This track can be constructed from as many as 10 segments, each of which can be a
portion of a rhumb line or of a great circle.
The motion of the target about each base track segment is specified by three circular nor-
mal probability distributions corresponding to target position at the initial, mid-point, and end-
point of each segment. Each simulated target track is obtained by drawing random numbers for
target position from these distributions and then connecting the points with straight lines.
Figure 3 illustrates a typical situation. The target's point of departure and intended desti-
nation are assumed known, and a base track is constructed between these points. The base
track might be taken from the target's float plan or hypothesized from the target's past habits.
In the case illustrated by Figure 3, there are three track segments. The 50% circles of uncer-
tainty are assumed to grow in size to about midway along the track and then diminish. Since
the point of departure and intended destination are assumed to be known, the extreme end-
points of the entire track have zero uncertainty.
Figure 3. Description of irackli
COAST GUARD COMPUTER-ASSISTED SEARCH (CASP) 667
In some cases, there is information which leads one to suspect that the distress is more
likely to have occurred on one part of the track than on another. For example, as mentioned
above, the track may have passed through an area of storms and heavy seas. If desired, the tar-
get location probability distribution generated by TRACKLINE can be made to have a higher
density in such an area. This is done by specifying the highest probability point along base
track together with the odds that the distress occurred there rather than at the extreme end-
points of the track. These inputs determine a truncated triangular probability density for the
fraction of track covered before the distress incident occurred.
Updating for Target Motion
The DRIFT program is used to alter a target location probability distribution to account
for the effects of drift. Normally, the DRIFT program will cause the center of the distribution
to move to a new location and the distribution to become more diffuse.
Target motion due to drift complicates the maritime search problem. The prediction of
drift must account for the effects of both sea current due to prevailing circulation and predicted
or observed surface wind. Any object floating free on the ocean surface is transported directly
by surface current, and one component vector of drift is therefore equal to the predicted
current vector. A statistical file collected from ship reports over many years has been assem-
bled by the Coast Guard and arranged by geographical location and month of the year. The file
in use in the CASP system covers most of the North Atlantic and North Pacific Oceans.
As mentioned above, wind is also important in predicting target motion. With regard to
this factor, there are two major considerations. The first is the drift caused by the wind imping-
ing on the drifting object's surface area above water; this is called "leeway." The speed and
direction of leeway is different for different objects, and is usually difficult to predict.
The second wind consideration is the movement of the surface layer of the ocean itself;
this is called "local wind current." It is one of the most complex and least understood
phenomena in the entire drift process.
The primary data source for surface winds in the CASP system is the Navy's Fleet
Numerical Weather Central in Monterey, California. Every twelve hours their computers gen-
erate a time series for hemispheric wind circulation; three of these time series are used to pro-
duce certain geographical blocks of wind data which are transmitted to the Coast Guard for use
by CASP. All data are retained in the system for two to three months.
The process of applying the drift motion to update a CASP distribution is simple enough.
First, a set of total drift vector probability distributions is computed for various geographical
areas based upon estimates of sea current, leeway, and local wind current. Then for each target
location replication, a random vector of net drift is drawn from the appropriate probability dis-
tribution and used to move the target forward a short time. The procedure is repeated until the
entire update time is taken into account.
Updating for Negative Search Results
Once a search has actually been conducted, one of the two search update programs, REC-
TANGLE and PATH (depending upon the type of search), is run to revise the target location
probabilities to account for unsuccessful search. The effect is to reduce the probabilities within
the area searched, and to increase them outside.
668 H.R. RICHARDSON AND J.H. DISCENZA
Updating the target location probabilities for negative search results is carried out by an
application of Bayes' theorem. Recall that the target triple file contains J records of the form
(Xj, Yj, <t>j) for 1 < j < J, where the pair (XJt Yj) represents target position, and $,
represents the probability that the target replication would not have been detected by the cumu-
lative search effort under consideration. The overall cumulative probability of detection taking
all simulated targets into account is called search effectiveness probability (SEP) and is com-
puted by the formula
SEP = 1 - X 4),-//
/=i
Let C be a region in the search area, and let fl, denote the event, "target corresponds to
the yth replication and is in region C" The posterior probability A(C) that the target is located
in C given search failure is computed using Bayes' theorem by
j
A(C) = Pr {Target in C\ Search failure} = £ Pr[B,\ Search failure}
j
= £ /V {Search failure | B,\ Pr[B, }/Pr{ Search failure}
,/€r y-i
where r = [j : (Xh Yj) € C) denotes the set of indicies corresponding to target replications in C.
Now suppose that q} denotes the probability of failing to detect the yth target replication
during a particular update period. Using the independence assumption, the new individual
cumulative failure probability 4>, is computed by
4>; = q^'j,
where <$>', denotes the cumulative failure probability prior to the last increment of search.
The computation of the conditional failure probability qt is carried out in CASP by use of
a {M, B, o-)-detection model as described below. Recall (e.g., see Koopman [2]) that the
"lateral range" between searcher and target (both with constant course and speed) is defined as
the distance at closest point of approach. The "lateral range function" gives single sweep cumu-
lative detection probability for a specified lateral range for a specified period of time. The
integral of the lateral range function is called the "sweep width" of the sensor.
The CASP programs* are based upon the assumption that the lateral range function for
the search unit is rectangular and is described by two parameters, M and B. Here M denotes
the total width of the swept path, and /3 denotes ihe probability that the target would be
detected for lateral ranges less than or equal to Mil. The sweep width W for the rectangular
lateral range function described above is given by
W = BM.
Navigational uncertainties ("pattern error") are introduced into the detection model by
assuming each sweep is a random parallel displacement from the intended sweep. The random
*An option is also provided to use an "inverse cube" lateral range function as defined in [2] together with search pattern
COAST GUARD COMPUTER-ASSISTED SEARCH (CASP) 669
displacements are assumed to be independent identically distributed normal random variables
with zero mean and standard deviation. This model was introduced by R. K. Reber (e.g., see
Reber [4]) and used extensively in certain Navy search analyses.
Rectangular lateral range functions are a useful way of approximating more complex
lateral range functions. If the actual lateral range function has sweep width M and is nonzero
over an interval of width M, then one may define /3 to be the average detection probability over
the effective range of the sensor, i.e., /3 = WjM. Appendix A of [4] shows that replacement of
the actual lateral range function by a rectangular lateral range function with average probability
/3 usually does not lead to significant errors in the computed value of probability of detection
for parallel path search. Cases where there is significant disagreement occur when the lateral
range function is close to zero over a large part of its support.
Let G„ denote the cumulative normal probability distribution function. Let (u, v) denote
the target's position in a coordinate system where the origin is at the midpoint of a given
sweep, and where the w-axis is parallel to the sweep and the v-axis is perpendicular to the
sweep. Then for fixed M, /3, and o-, the single sweep probability p(u, v) of detecting the target
is given by
m I \ Jr I _l L\ r \ L)l \r \ j. M\ r ( M
(1) p\u,\) = j8 GJ« + — — Ga \u — — \G(T\y + — - G„ v — —
where L denotes the length of the sweep.
If there are K search legs to be considered, and if {uf, wf) denotes the coordinates of the
yth simulated target position relative to the kx\\ search leg, then the failure probability <?, is
given by
(2) qj= n [1 -/>(«/*. v/,)].
The application of these formulas in programs PATH and RECTANGLE can now be discussed.
Path. Program PATH is used to represent general search patterns constructed from
straight track segments. For example, PATH can be used to compute detection probabilities for
a circle diameter search where the search tracks are intended to cover a given circle by making
repeated passes through its center. PATH makes direct use of (1) and (2).
Rectangle. Program RECTANGLE has been designed for the special case where a rectan-
gle is searched using parallel sweeps. RECTANGLE reduces the computing time and amount
of input that otherwise would be required using program PATH. For a point outside the desig-
nated rectangle, the probability of detection qi is assumed to be 0. For a point inside the desig-
nated rectangle, "edge" effects are ignored and an average probability of detection is computed
as if there were an infinite number of sweeps, each infinitely long.
The following line of reasoning originated with R. K. Reber. Reber [4] presents results in
the form of curves and tables, and these have been adapted to program RECTANGLE by use
of polynomial approximations. Let S denote the spacing between sweeps. Since the sweeps are
assumed to be parallel and of infinite extent, the coordinate v/ expresses the lateral range for
the kth sweep and the yth simulated target location and is given by
v/ = fij + kS
for — oo < /c < oo and a number /x , such that |/x / • | ^ 5.
670 H.R. RICHARDSON AND J.H. DISCENZA
Now for arbitrary /jl, refer to (1) and (2) and define g by
(3)g(fx,S)= n ll-p(u,H + kS)]= f[ /3 \GAfjL + kS + ~ - G,
Note that since the sweeps are assumed to be of infinite length, one has u = °° and g defined
by (3) does not depend upon u. The function g is periodic in its first argument with period fi.
Let g(S) denote the average value of g(jx, s) with respect to the first argument. Then
1
Z{S)= i Jo ^•S)d»-
S
The function g has been tabulated in [4] and is used in program RECTANGLE to
represent the failure probability q} = g(S) for a point lying within the designated search rectan-
gle. RECTANGLE and PATH agree (as they should) when PATH is used to represent a paral-
lel path search.
Search Optimization
Two programs, MAP and MULTI, are used for optimizing the allocation of search effort.
MAP provides a quick way of determining the search cells which should receive effort based
upon a constraint on total track line miles available. MULTI determines search areas for multi-
ple search units under the constraint that each unit must be assigned a uniform coverage of a
rectangle and that the rectangles for the various search units do not overlap.
The method used in program MAP is based upon use of an exponential detection function
(see Stone [8]) introduced by Koopman [3] and does not impose constraints on the type of
search pattern employed. The primary usefulness of this program is to provide the search
planner with a quick method for defining the area of search concentration. The following para-
graphs give a brief sketch of the methods used in these optimization programs.
Map. Let there be TV search cells, and for 1 < n ^ N let p„ and a„ denote, respectively,
the target location probability and the area associated with the nth cell. The probability density
for target location in the nth cell is given by d„ = p„/an. Suppose that total search effort is
measured by the product of track line miles and sweep width.
Let y denote an allocation of search effort where y(n) denotes the amount of search
effort (measured in area swept) allocated to the nth cell. Probability of detection PD[y] is com-
puted using an exponential effectiveness function, i.e.,
PdW= Z PnU - exp(-y(n)/a„)].
The objective is to maximize PD subject to a constraint on total effort available. This is easily
done using the techniques introduced by Koopman [3]; easier proofs are provided in Stone [8]
and Wagner [12].
It can be shown that under the above assumptions, the initial increments of effort should
be concentrated in the highest probability density cells, and that there should be a succession of
expansions to cells having lower target location probability density.
In order to derive the formulas used in program MAP, a new collection of equi-density
search regions is formed made up of the unions of all cells having equal probability density.
Let
COAST GUARD COMPUTER-ASSISTED SEARCH (CASP)
671
K = the number of equi-density regions
dk = the probability density for region k
Ik = the set of indices corresponding to the cells comprising region k
Ak = the area of region k.
Using the above notation
/6/fc
Let Ek denote the total effort which must be expended before the optimal search expands into
the kth region. Assume that the equi-density regions have been ordered beginning with the
region having the highest density. Since search begins in the highest density region, we have
E\ = 0. It can be shown that in general for k ^ 2
(4)
Ek = Ek_{ + (ln^_, - In dk) £ Am.
m=\
Figure 4 shows output from program MAP illustrating the use of (4). The list shows the
25 highest probability cells specified by the latitude and longitude of the southeast corner. Each
cell is 15 minutes wide, and the numbers in the last column correspond to the values Ek given
by (4). The planning advice given in [10] is to apply search effort to any cell for which the
value in the effort column is less than the total effort available.
TOP 25
PROBABILITY
LOCATION
(S.E. CORNER)
EFFORT
1
0.05133
43-ON
69-45W
2
0.04167
42-45N
69-30W
35.0
3
0.04133
43-0N
70-OW
36.3
4
0.04100
43-0N
69-30W
40.3
5
0.03567
43-15N
69-30VV
129.4
6
0.03467
43-15N
69-45W
152.8
7
0.03333
42-45N
69-15W
199.6
8
0.03267
42-30N
69-15W
227.5
9
0.03267
42-45N
69-45W
222.2
10
0.03267
43-15N
70-0W
210.1
11
0.03200
42-30N
69-30VV
264.1
12
0.02800
43-ON
69-15W
491.5
13
0.02733
42-45N
70-0W
547.2
14
0.02533
43-ON
70-15W
701.3
15
0.02267
42-30N
69-0W
976.5
16
0.02233
43-15N
69-15W
983.1
17
0.02167
42-30N
69-45W
1095.1
18
0.02133
43-15N
70-15W
1104.4
19
0.02100
42-45N
69-OW
1175.3
20
0.01867
43-30N
69-30W
1505.8
21
0.01867
43-30N
69-45W
1505.8
22
0.01800
43-ON
69-0W
1659.9
23
0.01667
42-30N
68-45W
1968.0
24
0.01600
42-15N
69-0W
2137.7
25
0.01600
42-15N
69-15W
2137.7
Figure 4. Optimal allocation of effort produced by Map
672
H.R. RICHARDSON AND J.H. DISCENZA
Notice that the numbers in the effort column are not necessarily increasing. This is
because the list is ordered according to containment probability rather than probability density.
Multi. As mentioned above, program MAP does not take into account "simplicity" con-
straints which are considered important in operational planning. Program MULTI was designed
to overcome this drawback in cases where multiple search units are deployed in the same search
area.
The first simplicity constraint introduced is that each unit will be assigned to uniformly
search a rectangle. Figure 5 shows the dimensions of the optimal rectangle and the resulting
probability of detection under the assumption that the target location probability distribution is
normal. In order to use this figure, one first computes the normalized effort E* by the formula
E- -—*?—.
crmaxo-rnJn
where R is the sweep rate of the unit, T is the total search time, and o-max and crmln are the
standard deviations of the normal distribution when referred to principal axes. The optimal
search rectangle will have half side given by U*a-max and £/*crmin where the size factor U* is
given by the designated curve with values read along the outer vertical scale.
£
Tonxh
b°f
5
— 1 0
ootir
«l searc
h plan
Opti
n.
Plan
RT
Probability of Detectio
b
— .9
E* =
(Inn
er Scale)
Half
sides of
rectangle
I 4
Q
"""" 8
given
by U*
?maxand U* omin
s
— .7
2
Q
o 3
— 6
l
"*" Optimal Rectangle
Plan
|
1
5
— 5
2
2 2
— .3
— 2
1
1
"*" Optimal Rectangle
(Outer Scale)
Normalized Effort E'
Figure 5. Optimal search reciangle
Figure 5 provides curves to determine the probability of detection for the optimal rectan-
gle plan and for the unconstrained optimal plan. It is interesting to note that in all cases the
probability of detection provided by the optimal rectangle plan is at least 95% of that provided
COAST GUARD COMPUTER-ASSISTED SEARCH (CASP) 673
by the unconstrained optimal plan. Thus, under the assumption stated, uniform search of the
optimal rectangle can be recommended without hesitation since, in most cases, the simplicity of
the rectangle plan is more important than the small improvement in effectiveness obtained by
the more complicated optimal plan.
MULTI is capable of allocating the effort of up to 5 search units to nonoverlapping rectan-
gles in a way which is intended to maximize overall probability of detection. The first step in
this procedure is to approximate the target location probability distribution by the weighted
average of k bivariate normal distribution where / < k < 3. This is done by locating the three
highest local maxima in the smoothed cell distribution and then associating each simulated tar-
get position with the nearest cluster point. If three local maxima cannot be found, then the
procedure is carried out with one or two local maxima. The mean and covariance matrix of
each cluster are calculated to determine the parameters of the approximating normal distribu-
tion.
The program next considers all possible assignments of search units to one of the three
approximating probability distributions. Since there are a maximum of five units and three dis-
tributions, there are at most 35 — 243 different ways of assigning units to distributions. For
each assignment, the program sums up the total effort available to search each distribution and
then computes the resulting optimal rectangle and associated probability of detection. If Pk
denotes the conditional probability of detecting the target with optimal rectangle search given
that the target has the Ath distribution (1 < k ^ k), then probability of detection A for the
allocation is given by
A = i PkDk.
The program prints the allocation which gives the maximum probability of detection and
notes whether any of the rectangles overlap. If overlap occurs, then the next ranking allocation
is printed, and so on. This continues until an allocation without overlap is found or until the
top five allocations have been listed together with their associated detection probabilities.
Finally, when several units are assigned to the same rectangle, it is subdivided in a way which
preserves the uniform coverage.
Recently an alternative method for multiple unit allocation has been developed (see Dis-
cenza [1]) based upon integer programming considerations.
3. CASP CASE EXAMPLE
On 12 September 1976 the sailing vessel S/V Spirit departed Honolulu enroute San Fran-
cisco Bay. The owner, who was awaiting its arrival in San Francisco, reported concern for the
vessel to the Coast Guard on 14 October 1976 after it had failed to arrive. An Urgent Marine
Information Broadcast (UMIB) was initiated on 17 October. The following day, a merchant
vessel the M/V Oriental Financier reported recovering a life raft with two survivors from the
S/V Spirit which had sunk in heavy seas in mid-Pacific on the morning of 27 September. Sur-
vivors indicated three more crewmembers in a separate raft were still adrift. This information
opened an extensive six day air and surface search for the missing raft that eventually located
the raft with one of the missing persons on board.
674 H.R. RICHARDSON AND J H DISCENZA
Each day's search was planned utilizing computer SAR programs. Initial distress position
information was gained by radio-telephone debriefing of the survivors aboard the M/B Oriental
Financier on several occasions. The search began 19 October based on a SARP* datum for a
raft without a drogue from an initial reported position of 36N 136W. The second day's search
was based on a SARP datum for a position 160 nautical miles to the northeast from the previ-
ous position (this position being determined from further debriefing of the survivors over
radio-telephone). The third through the six days' searches were planned utilizing CASP output
from a POSITION scenario consisting of an ellipse with a 160 mile major axis and a 60 mile
minor axis. The CASP program was updated by RECTANGLE and DRIFT daily, and search
areas assigned to cover the highest cells which could be reached taking into account search unit
speed and endurance.
The following chronology is based upon the official USCG report and describes the utiliza-
tion of CASP in the search planning. This case is a good illustration of the many uncertainties
which must be analyzed during a search and the way both negative and positive information
contribute to eventual success.
21 October 1976
Search planning for the day's operations utilized the CASP program for the first time.
New probable distress position information given by the survivors was evaluated and the CASP
program was initiated using a POSITION scenario with center length 160 miles and width 60
miles oriented on 046°T, with the southwest end at position 36N 136W. This scenario was to
be used for the rest of the search. A search plan was generated for the 21 October search cov-
ering approximately 8 of the 10 highest CASP cells as given in MAP. Ten units were desig-
nated for the day's efforts and consisted of 3 Coast Guard, 2 Navy, and 4 Air Force aircraft and
the USS Cook.
The first aircraft which arrived on scene for the day's search reported the weather in the
search area as ceiling varying 200-1500 feet (scattered), wind from 330° at 8 knots, seas 4 feet,
and visibility unlimited except in occasional rain showers.
At 3:06 PM an aircraft located what appeared to be the life raft of recovered survivors in
position 35-38N 138-12W. M/V Oriental Financier had been unable to recover this raft when
the survivors were rescued. The USS Cook investigated and reported negative results.
Figure 6 shows the search plan for 21 October. Note that the target was eventually found
on 24 October in the first designated area C-l. There is, of course, no way of knowing where
the target was on the 21st.
22 October 1976
Planning for day's search was done using updates from the CASP program. Search units,
consisting of 17 aircraft (3 Coast Guard, 6 Navy, and 8 Air Force) and the USS Cook, were
designated areas totaling 67,920 square miles for the day's effort. Areas assigned were deter-
mined from the MAP's twelve highest cells. High altitude photographic reconnaissance flight
utilizing U-2 aircraft was also scheduled, cloud coverage permitting, to cover an area of 57,600
square miles.
*A computer program implementing methods described in the National SAR Manual and a precursor to CASP.
COAST GUARD COMPUTER-ASSISTED SEARCH (CASP)
CHARLIE SEARCH AREA
NORTH
PACIFIC
OCEAN
1st Raft Recovered
36-15N 139-23W
19 0130Z
2nd Raft Recovered
35-53N 138- 10W
24 2137Z
my
^—Distress P<
^ Two
IN 133-
"SPIRIT" Sank
I" JApprox. Position
36N 136W 27 1900Z
Hawaiian Islands
**
Tropic of Cancer
2nd Raft Recovered
"SPIRIT" Sank
Note:
POD is the estimated conditional
probability of detection given
the target is in
the designated
area.
CHARLIE SEARCH PLAN
AREA
UNIT
POD
C-l
NAVY P-3
70%
C-2
NAVY P-3
70%
C-3{S) AF HC-130
78%
(N) AF HC-130
72%
C-4
AF HC-130
70%
C-5
CG HC-130
78%
C-6
CG HC-130
55%
C-7
AF HC-130
78%
C-8
CG HC-130
77%
C-9
AF HC-130
70%
C-ll
CG HC-130
78%
Search plan based on CASP high
probability areas /distress position ellipse
Figure 6. Search plan for 21 October
676 H.R. RICHARDSON AND J.H. DISCENZA
The first aircraft on scene for the day's search reported the weather in the general area as
ceiling 1800 feet (broken), winds from 150° at 6 knots, seas 2 feet, and visibility 15 miles.
Search conducted during daylight hours utilized 15 aircraft, the USS Cook, and a U2 high
altitude reconnaissance flight. The USS Cook was unable to relocate debris sighted during pre-
vious day's search. Two Air Force aircraft failed to arrive on scene prior to darkness and were
released. Aircraft on scene searched 88 percent of 67,920 square miles assigned and obtained
POD's ranging from 50 to 82 percent. The high altitude photographic reconnaissance flight was
conducted from an altitude of approximately 50,000 feet.
The CGC Campbell arrived on scene and relieved the USS Cook.
23 October 1976
The Rescue Coordination Center (RCC) was advised by the Air Force that development
of high altitude film had shown an "orange dot" in position 35-16N 139-05W. The photo-
graphed object was described as a round orange object, approximately 7 feet in diameter, float-
ing on the surface of the water.
Search planning was done using updates from the CASP program. Search units, consist-
ing of the CGC Campbell and 8 aircraft (2 Coast Guard, 3 Navy, and 3 Air Force), were
assigned areas of highest CASP cells. The object photographed by reconnaissance aircraft was
drifted by SARP and the CGC Campbell and 1 aircraft dedicated to locate it.
The first aircraft on scene for the day's search reported weather in the search area as ceil-
ing 2000 feet, wind from 200° at 12 knots, seas 2 feet, and visibility 15 miles.
Search conducted during daylight hours utilized 8 aircraft and CGC Campbell. Search
units covered 97 percent of the assigned 34,300 square miles with POD's ranging from 50 to 92
percent. Several sightings of assorted flotsam were reported but none linked to Spirit or rafts.
The object photographed by the high altitude reconnaissance flight on 22 October was not relo-
cated by search units.
Figure 7 shows the search plan for 23 October. Although not indicated in the chart, the
position where the target was found on the 24th is in the second highest probability density cell
from the CASP map.
24 October 1976
Search planning for the day's operations was done using updates from the CASP program.
Search units consisting of the CGC Campbell and 5 aircraft (2 Coast Guard and 3 Navy) were
assigned areas of highest CASP probability totaling 18,082 square miles, with CGC Campbell
and one Coast Guard aircraft designated for location of the object reported by the reconnais-
sance flight.
The position of the reconnaissance flight sighting of 22 October was drifted utilizing
SARP and the new position passed to CGC Campbell for search purposes. The 11:00 AM
SARP datum was computed to be 35-29. 4N 138-39. 2W with standard first search radius of 16.9
miles. The search plan is shown in Figure 8.
COAST GUARD COMPUTER-ASSISTED SEARCH (CASP)
677
NORTH
PACIFIC
OCEAN
1st Raft Recovered
36-15N 139-23W
19 0130Z
\
-*f
Distress Position Two
37-54N | 133- 36W
2nd Raft Recovered
35-53N 138- 10W U.
24 2137Z ' 1
-B
"SPIRIT" Sank
Approx. position
36N 136W
27 1900Z
UNITED
STATES
ECHO SEARCH AREAS
Hawaiian Islands
_ Xrpgic o{_ Cancer _>
: Raft Recovered
"SPIRIT" Sank
ECHO SEARCH PLAN
AREA UNIT
E-l NAVY P-3
E-2 NAVY P-3
E-3 NAVY P-3
E-4 AF HC-130
E-5N/S CG HC-130
E-6 AF HC-130
E-7 CG HC-130
E-8N/S AF HC-130
E-9 CG HC-130
POD
62%
52%
50%
78%
58/74%
92%
64%
50/60%
74%
Search plan based on CASP high probability
areas, distress position ellipse, and
reconnaissance sighting (E-l).
Note: POD is
the estimated
conditional probability
of detection given
the target is in the
designated area.
138W
134W
Figure 7. Search plan for 23 October
678
H.R. RICHARDSON AND J.H. DISCENZA
NORTH
PACIFIC
OCEAN
Distress Position T
37-54N 133-36W
1st Raft Recovered L
36-15N 139-2 3W "t
19 0130Z \ |\
UNITED
STATES
x-eu
2nd Raft Recovered
35-53N 138-10W^-
24 2137Z ^
Hawaiian Islands
D
b'SPIRIT" Sank
Approx. position
36N 136W
27 1900Z
FOXTROT SEARCH AREAS
Tropic of Cance"
120W
FOXTROT SEARCH PLAN
AREA UNIT
^Raft Recovered
35-53N 138- 10W
F-3 NAVY P-3
F-4 NAVY P-3
F-5 CG HC-130
34N
F-2
32N
F-4
F-5
Search plan based on CASP high
probability area and reconnaissance
sighting (F-'5).
135W
Figure 8. Search plan for 24 October
COAST GUARD COMPUTER-ASSISTED SEARCH (CASP) 679
The first aircraft on scene for the day's search reported weather in the search area as ceil-
ing 1500 feet, wind from 000° at 7 knots, seas 3 feet, and visibility 10 miles.
The CGC Campbell reported locating a rusty, barnacle encrusted 55 gallon drum in posi-
tion 35-27.2N 138-39.0W.
At 12:05 PM the search met with success! A Coast Guard HC-130H reported sighting a
raft in position 36-03N 138-00W with at least one person on board. The CGC Campbell pro-
ceded enroute to investigate, and at 2:37 PM CGC Campbell reported on scene with the raft in
position 35-53N 138-10W. A small boat was lowered to recover the survivor, and at 3:01 PM
all search units were released from the scene.
4. TRAINING
CASP training began with an operational testing phase in cooperation with the New York
RCC. This operational testing was useful in orienting the personnel to the benefits derived
from more detailed search planning, and provided an idea of what the full training problem was
going to be like.
Coincident with this, a training manual [9] and a completely new combined operating
handbook [10] were developed encompassing all of the operational computer services available.
At the time of official implementation in February 1974, a special four-day class was con-
ducted in the operation of the CASP system; this class was attended by one representative from
each Rescue Coordination Center. It was intended that these persons would learn the system
thoroughly and return to their respective commands and teach others. This plan was marginally
successful, and worked only in those cases where an extremely capable individual was selected
for attendance.
During the next six months, personnel from the Operations Analysis Branch visited each
East Coast RCC for one week apiece in order to provide additional training. Subsequently, the
same visit schedule was repeated on the West Coast.
Another valuable tool for training has been telephone consultation. Fortunately, all mes-
sages into and out of the computer are monitored at New York, and personnel can be helped
with the details of input and output with a quick telephone call on the spot.
Finally, the National Search and Rescue School has made CASP training a regular part of
its curriculum. The school, located on Governors Island, is responsible for initial training of all
RCC personnel (among many others) in the techniques of search and rescue. The present SAR
school training session is four weeks in duration with the fourth week devoted to computer
search planning systems training. Over half of this time is devoted directly to CASP.
The Coast Guard is currently in the process of separating its administrative and opera-
tional systems by establishing an Operational Computer Center. This new Center will give res-
cue coordinators direct access to CASP through on-line terminals and will improve CASP's
availability and reliability. Interactive program control will make the modules easier to use.
The application of CASP in operational situations has been quite successful, in spite of
significant encumberances associated with computer and communications services.
680 H.R. RICHARDSON AND J.H. DISCENZA
Continued oceanographic research programs will expand CASP's applicability to important
in-shore regions. Implementation of the new multi-unit allocation algorithm [1] is expected to
simplify the search area assignment problem. These additional capabilities coupled with
improved computer access and reliability should make CASP an even more valuable planning
tool in the future.
ACKNOWLEDGMENTS
The development, implementation, training, and utilization of CASP represents the con-
tributions of individuals far too numerous to mention by name in this paper. Foremost among
these are the officers and men who use CASP in the RCCs and without whom the system
would be useless. The contributions of the following individuals to the support and develop-
ment of CASP are specifically acknowledged: C. J. Glass, R. C. Powell, G. Seaman, V.
Banowitz, F. Mittricker, R. M. Larrabee, J. White, J. H. Hanna, L. D. Stone, D. C. Bossard, B.
D. Wenocur, E. P. Loane, and C. A. Persinger.
REFERENCES
[1] Discenza, J.H., "Optimal Search with Multiple Rectangular Search Areas," Doctoral Thesis,
Graduate School of Business Administration, New York University (1979).
[2] Koopman, B.O., "The Theory of Search, Part II, Target Detection," Operations Research,
4, 503-531 (1956).
[3] Koopman, B.O., "The Theory of Search, Part III, The Optimum Distribution of Searching
Effort," Operations Research, 5, 613-626 (1957).
[4] Reber, R.K., "A Theoretical Evaluation of Various Search/ Salvage Procedures for Use with
Narrow-Path Locators, Part I, Area and Channel Searching," Bureau of Ships,
Minesweeping Branch Technical Report, No. 117 (AD 881408) (1956).
[5] Richardson, H.R., Operations Analysis, February (1967). Chapter V, Part 2 of Aircraft
Salvage Operation, Mediterranean, Report to the Chief of Naval Operations prepared by
Ocean Systems, Inc. for the Supervisor of Salvage and the Deep Submergence Systems
Project.
[6] Richardson, H.R. and L.D. Stone, "Operations Analysis During the Underwater Search for
Scorpion," Naval Research Logistics Quarterly, 75, 141-157 (1971).
[7] Shreider, Yu. A., The Monte Carlo Method (Pergamon Press, 1966).
[8] Stone, L.D., Theory of Optimal Search (Academic Press, 1975).
[9] U. S. Coast Guard, Commander, Atlantic Area, CASP Training Course, 19-22 February
(1974).
[\0] U. S. Coast Guard, Computerized Search and Rescue Systems Handbook (1974).
[11] U. S. Coast Guard, National Search and Rescue Manual (1970).
[12] Wagner, D.H. "Nonlinear Functional Versions of the Neyman-Pearson Lemma," SI AM
Review, 11, 52-65 (1969).
CONCENTRATED FIRING IN MANY-VERSUS-MANY DUELS
A. Zinger
University of Quebec at Montreal
Montreal, Canada
ABSTRACT
A simple stochastic-duel model, based on alternate firing, is proposed. This
model is shown to be asymptotically equivalent, for small hit probabilities, to
other known models, such as simple and square duels. Alternate firing intro-
duces an interaction between opponents and allows one to consider multiple
duels. Conditions under which concentrated firing is better or worse than
parallel firing are found by calculation and sometimes by simulation. The only
parameters considered are the combat group sizes (all units within a group are
assumed identical), the hit probabilities and the number of hits necessary to
destroy an opposing unit.
1. INTRODUCTION
Two extremes for the modeling combat attrition are given by the so-called Lanchester
theory of combat, which treats combat attrition at a macroscopic level, and by the theory of sto-
chastic duels, which treats combat attrition at a microscopic level and considers individual firers,
target acquisition, the firing of each and every round, etc. (see Ancker [1, pp. 388-389] for
further details). Actual combat operations are, of course, much more complex than their
representation by such relatively simple attrition models and may also be investigated by means
of much more detailed Monte Carlo combat simulations. Unfortunately, such detailed Monte
Carlo simulations usually fail to provide any direct insights into the dynamics of combat without
a prohibitive amount of computational effort. In the paper at hand, we will consider a relatively
simple stochastic-duel model to develop some important insights into a persisting issue of mili-
tary tactics (namely, what are the conditions under which concentration of fire is "beneficial").
In his now classic 1914 paper, F.W. Lanchester [10] (see also [11]) used a simple deter-
ministic differential-equation model to quantitatively justify the principle of concentration, i.e.,
a commander should always concentrate as many men and means of battle at the decisive point.
From his simple macroscopic model, Lanchester concluded that the "advantage shown to accrue
from fire concentration as exemplified by the n square law is overwhelming." However, this
conclusion depends in an essential way on the macroscopic differential-equation attrition model
used by Lanchester [10], [11] (see Taylor [14] for further discussion) and need not hold for
microscopic stochastic-duel models of combat attrition. In fact, this paper shows that for such
microscopic duel models it is not always "best" to concentrate fire.
Subsequently, many investigators have commented on the benefits to be gained from con-
centrating fire. For example, in his determination of the probability of winning for a stochastic
analogue of Lanchester's original model, Brown [6] stressed the fact that the model applied to
681
682 A /IN(,I K
cases of concentrated firing by both sides. Other investigators of deterministic Lanchester-type
models from the macroscopic combat-analysis point of view have also stressed this point (e.g.
see Dolansky [7], Taylor [13], and Taylor and Parry [15]). Recently, Taylor [14] has examined
the decision to initially commit forces in combat between two homogeneous forces modeled by
very general deterministic Lanchester-type equations. He showed that it is not always "best" to
commit as much as possible to battle initially but that the optimal decision for the initial com-
mitment of forces depends on a number of factors, the key of which is how the trading of
casualties depends on the victor's force level and time.
The first reference to problems of strategy in multiple duels is found in Ancker and Willi-
ams [2], who study the case of a square duel (2 vs 2) and arrive at the right conclusion that
parallel firing is better than concentrated firing. This is a natural conclusion since only one hit
is necessary to achieve destruction, and in concentrated firing there is a certain amount of
over-killing. In 1967, Ancker [1] makes suggestions for future research concerning mutliple
duels and states explicitly that the difficulties lie in the strong interaction between the contes-
tants. The possibility of needing more than one hit to achieve destruction in the simple duel
situation was introduced by Bhashyam [4] in 1970.
The purpose of this paper is to combine some of the above mentioned concepts, in order
to gain insight concerning a problem of strategy in multiple duels— should one concentrate
one's fire or not?
2. ASSUMPTIONS AND NOTATION
Let us consider two forces A and B that meet each other in combat. A consists of M units
and B of N units.
The following assumptions are made:
1. Firing is alternating, volley after volley, i.e., A fires all weapons simultaneously, then B
and so on until all units of a force are destroyed. This is contrary to the usual assumption of
either simultaneous firing or random firing within some time intervals as found in Robertson
[12], Williams [17], Helmbold [8], [9], Thompson [16], Ancker [3]. It is felt, and will be
shown in a few cases, that for relatively small probabilities of hitting, this approach gives results
comparable to Ancker and Williams [2]. We will denote by Vt\j the probability of / winning if j
shoots first /', j = A,B. The unconditional probability of winning will be denoted by VA or VB.
2. Hit probabilities are constant and are respectively pA and pB, with q, = 1 — p,, /' = A,B.
3. Each unit of force A requires KA hits to be destroyed. Same for B and KB.
4. The supply of ammunition is unlimited.
5. There is no time limit to score a hit.
6. In a multiple duel (more than 1 vs 1) the units of A concentrate their fire on a single
unit of B while the units of B each fire at a different unit of A, or spread their fire over all
available units of A, this last case occurs when M < N. B has to allow an amount of concen-
tration in order not to lose some shots. Concentration will be kept at a minimum to preserve
as much parallelism as possible. For example if M = 3 and N = 7 the pattern of fire for B has
to be
CONCENTRATED FIRING IN MANY-VERSUS-M ANY DUELS
7. The most general notation, for example, VA\B(M,N,KA,KB,pA pB) will be avoided if
possible and replaced by an appropriate simpler form.
Before proceeding, a general remark ought to be made: most of the difficulties come from
the asymmetry in the situation and from the interaction between the opponents. The same
model has to express concentration, dispersion and partial concentration of fire. Moreover, the
probability of winning depends upon the whole past history of the duel.
3. MULTIPLE DUEL. ONE HIT SUFFICIENT TO DESTROY
Let KA -
1 and let E(i,j) be the state of group A with i units, and of group B with
If A fires first, the next state is
E(i,j) with probability q'A and
E(i,j — 1) with probability 1 — q'A.
When B fires, let us first consider the case when j < i. Then,
E(i,j) becomes E(i - k,j), k = 0, . . . , j with probability , \pB q
E(i,j — 1) becomes E(i — k, j — 1), k = 0,
1 with probability
If on the other hand j > / some regrouping has to be done.
Let j = ai + b with b < r, a, b € I+. The regrouping which spreads the fire the most is
given by
a shots are fired with a probability of success
1 — (1 — pB)a = 1 — A0 at each of / — b targets
a + 1 shots are fired with a probability of success
1 - (1 - pB)a+l — 1 — Ax at each of b targets.
Define r = min(ij). Then both cases j < /and j ^ /are identical if one defines the probabil-
ity of transition from state E(i,j) to state E(i — k, j) when B fires as
( r - b) kx b - k\ k0 r - b - k0
* (l-^l) A (1 — ^0> A
*o Ax AQ
(3.1) 0(i,j,k,pB)= X
*o=0,l....r-
*.=0. 1
In the case j < i, a = 0, k0 = 0 and k\ = k.
684 A ZINGKR
It follows that if A starts and B returns fire once, the intial state Ed, j) can become
Ed, j) with probability qA 0 (/, j, 0, pB) = qA qfc
Ed - k, j) with probability qA 0 (/', j, k, pB), k = 1, .... r
E(i - kj - 1) with probability (1 - q'A) 0 (/, j - 1, k, pB), k = 0, 1, . . . , r'
where r' = min (/, j — 1).
If B starts, the initial state E(i, j) can become
E(i, j) with probability qA 0 (/', j, 0, pB) = qA qB
E(i — k, j) with probability qA~k 0(i,j,k,pB), k = 1, . . . , r
E(i - k,j - 1) with probability (1 - q'Ak) 0 d, j, k, pB), k = 0, 1, .... r"
where r"= min(/ — 1, j).
Let VB\A(M, AO denote the probability that group B wins with initial state E(M, N) and
A starts firing. Then
(3.2) VBU (M, N) = q? q$ VB{A (M, N)
+ q% £ 0 (M, N, k, pB) VBU{M-k, N)
k=\
+ (1 - Qa) £ 9 (M, N - 1, k, pB) VBlA (M- k, N- 1).
A:=0
This corresponds to a decomposition into all the mutually exclusive and exhaustive ways for B
to win if A fires once and then B returns fire.
In a similar way
(3.3) VB]B(M, N) = qjfqg VBlB (M, N)
+ Z QA~k 9 (M, N, k, pB) Vb\b(M - k, N)
k=\
+ Z (1 - qH?~k) 9(M, N, k, pB) VB\B(M - k, N - 1).
k=Q
Since we have
VBU (M, 0) = VB\B(M, 0) = 0 all M
and
Vb\a®, N)= VBlB(Q,N)= 1 all AT
we can calculate in succession all required probabilities. For example, since 0 (1,1,1, pB) = pB,
one finds VB\A{\, 1) = qApBl (1 - qAqB). Using ^^(1,1) and 0(1, 1,0, pB) = qB,
0(\,2,\,pB) = 1 - qh one finds VB[A(l,2).
Explicitly, one gets, by assuming that A starts half the time,
1
= \pb4<m-x)I1 (i + mi n (i - ftci).
CONCENTRATED FIRING IN M ANY-VERSUS-M ANY DUELS
One can also obtain for qA = qB = q
v (2 2) = l + 4q + Aql + lq3 + Aq* + 3qS + q6
B ' 2(\ + q)2 (\ + q2) (\ + q + q2)
A comparison with the triangular duel and the first square duel [2] for p — » 0, q — • 1
VB{2,\) =
plq(\ + q2)
q(\ + q2)
► 1/6
and VB(2,2)-
in [2].
2 (l_^2)(1_93) 2(1 + q)(l + q + q2) f-1
»l/2 which are the same limits as the one obtained from Equation 29 and 37
Table 1 gives some results for VB(M, N, pA, pB).
TABLE 1 - (x 104)
M
N
Pa 0.3
0.3
0.5
0.5
0.7
0.5
0.7
Pb 0.3
0.5
0.3
0.5
0.5
0.7
0.7
1
1
5000
6538
3462
5000
3824
6176
5000
2
2
5166
7307
3100
5317
3850
6873
5447
3
3
5678
8227
3405
6418
5081
8343
7386
3
5
9634
9982
8869
9913
9805
9998
9994
5
3
1292
3806
0368
1780
0997
3907
2832
5
5
7258
9614
5118
8940
8359
9920
9848
5
7
9831
9999
9422
9994
9986
10000
10000
7
5
3418
7843
1626
6060
5075
9090
8629
7
7
8850
9978
7538
9919
9853
10000
10000
10
10
9900
10000
9708
10000
10000
10000
10000
with M. We conclude: Parallel firing is better.
No simple relationship exists in the case pA ^ pB. Neither MpA vs NpB, nor M2pA vs
N2pB are sufficient to decide if VB > —.
4. SIMPLE DUEL. K HITS NECESSARY TO DESTROY
Let M = N = \ and let VB\A (KA,KB) denote the probability that B wins the simple duel
if A starts firing and KA hits are necessary to destroy A and KB for B.
It is evident that
Vb\a(Ka,Kb) = pA VB]B(KA, KB-l) + qA VBlB(KA, KB)
VBlB(KA, KB) = pB VBlA(KA - 1, KB) + qB VBlA(KA, KB).
686 A ZINGER
This gives
(4.1) (1 - qA qB) VBU(KA, KB) - pApB VBlA(KA -\,KB- 1)
- Pa Qb VbU{Ka, KB-\)- qApB VB\A{KA - 1, KB) = 0.
In order to solve this difference equation, following Boole [5] , let us define
x - KA, y - KB
ux,y= VBUKx - 1, y - 1)
Dxu = ux+Xy and Dyu = uxy+\.
Substituting these into Equation (4.1) we get
Kl - Qa Qb)Dx Dy - pApB- pA qB Dx - qA pB Dy]u - 0.
Let Dy = a.
((1 ~ Qa Qb)<* ~ Pa Qb)Dxu - pB(aqA + pA)u
which gives
« = PbQ>a+ Qa Dy)x[(l - qA qB)Dy - pA qB]~x 0 (y)
where 0(y) is arbitrary. Then,
« = dt Pa QV dA (1 ~ Qa Qb)~x D~x [l - j gg^ dA 0(y).
Since D~x0(y) = 0(y - x) and
,)* -L I •*+./' — l| ( P^ Qtt Y
l- QA<i
we get
. p* 1 ziur ■ lL+'«r'«* (i - ^ to)-y % - / - ;).
1 - «* Qb J £o j-o I 'J I ^ i
Taking into account that
PbQa
VB\A(hD=-
[~ QaQb
a good choice for 0(t) is
0{t) = 1 if t > 0
= 0 if r < 0.
Defining r = min (A^, KB — 1) the solution becomes
,^-1(^)1^+7-1) ^ Jfc.-I _^
(4.2) VB{A{KA,KB) = \? £ / / J** A Ci QbU-QaQb) aJ
with
W^ 0) = 0 and KflU(0, KB) = 1.
CONCENTRATED FIRING IN M ANYVERSUS-MANY DUELS
One can verify by substitution that this is a solution.
One can evaluate the other probabilities of winning by
va\b^a, kb> Pa. Pb) = vb\a^b, &a> Pb. Pa).
VB\B(KA, KB, pA, pB) = 1 - VA\B(KA, KB, pA, pB),
Va\a(Ka, Kb, pA,pB) = 1 - VBU(KA, KB, pA,pB).
Table 2 gives some results for VB\A (KA, KB, pA, pB) and VB\B(KA, KB, pA pB).
TABLE 2 - (x 104)
687
Pa = -3
Pb= -5
Pa =
Pb='-5
Pa = -5,
Pb = -1
Ka
KB
VB\A
Vb\b
Vb\a
Vb\b
Vb\a
VB\B
1
1
5385
7692
3333
6667
4118
8235
5
3
4257
5010
1139
1728
2576
3579
5
5
8201
8630
4512
5488
7414
8381
7
5
5955
6541
1674
2266
4159
5278
7
7
8695
8981
4599
5401
7981
8669
10
10
9160
9330
4671
5329
8545
9002
This table indicates that VB = 1/2 if KA = KB and pA = pB= 1/2, VB increases towards 1
if KA = KB and pB > pA and | VB\A - VB\B\ decreases if KA and KB increase.
An interesting comparison is to be made with the results given by Bhashyam [4]. Under
an assumption of an exponential distribution for interfiring times he finds that the probability of
B winning is, using our notation,
P(B)= 1 - / Pa (Kb, Ka)
Pa+Pb
where Ix is the incomplete Beta function. The correspondance in the notations being \p for pA ,
A *p* for pB, R for KB and R * for KA .
Table 3 shows at what rate a model with alternate firing converges towards Bhashyam's
model.
Alternate firing gives a good approximation if p is small. In fact, consider KA and KB
fixed and pA = c pB with pB ~* 0.
One can show that
lim VB\A -
1
(1 + c)^ ~o
and this limit from a well known theorem is
1 - I c (KB,KA).
KA+j-\
1 + c\
= lim VB\B
TABLE 3 - Rate of C
onvergence of VB
to P(B)
Pa
Pb
Ka
KB
vB
P(B)
0.4
0.2
5
5
0.1054
0.2
0.1
0.1265
0.02
0.01
0.1431
0.002
0.001
0.1447
0.1449
0.1
0.2
5
2
0.3391
0.01
0.02
0.3501
0.001
0.002
0.3511
0.3512
0.1
0.2
10
10
0.9491
0.01
0.02
0.9366
0.9352
5. SQUARE DUEL. 2 HITS NECESSARY TO DESTROY
Let M — N = 2 and KA = KB = 2. One can represent the state of the two forces by (;b
h> J\> Jt) with i\, h> J\» h = 0. 1.2, representing the number of hits necessary to destroy. For
example, (1, 1; 0, 2) means that A has 2 units that can be destroyed by one hit each and B has
one unit that has been destroyed by 2 hits and one unit untouched.
All attempts to arrive at one or two difference equations have been in vain. Two
equivalent approaches have been used. In the first, taking pA = pB = 1/2, and defining A, as
the matrix of the transitional probabilities corresponding to the case when A fires first, and B
the corresponding matrix when B fires first one obtains:
VA by summing all the probabilities for the events (/, j\ 0, 0) in lim (AB)n and VB by
„->oo
summing all the probabilities for the events (0, 0; i, j) in lim (BA)n.
The matrices are 29 x 29. The possible states of A are such that i\ ^ i2. The possible
states of B are such that jy < j2 and exclude jx = j2 = 1 since A concentrates its fire until des-
truction is achieved.
Assuming the ordering i\ > /'2, two variations are possible. In Case 1, when the state is
(2,1; 0, j) with j = 1 or 2 and B fires, B chooses at random among the two units of A. In Case
2, B fires on the second unit of A, which can be destroyed by one shot. We find
In Case 1 VA = 0.5586
and in Case 2 VA = 0.5396
In both cases concentrated firing is better.
The other approach consists in writing down all the equations that define the battle. For
example,
VAlA(2,2-\,2) = (1 - flD ^|fl(2,2;0,2) + q} VAlB(2,2-\,2).
The difference between Case 1 and 2 is seen by considering
^|S(2,l;0,l) = 0.5/7flF^M(l,l;0,l)+0.5pfl ^ (2, 0;0, 1)
+ qBVA]A(2,\-0,\)
CONCENTRATED FIRING IN MANY-VERSUS-MANY DUELS 689
VAlB(2, 1;0, 1) - pB VAU(2, 0;0. 1) + qB VaU{2, I;0, 1).
A third variation is possible in which no ordering is assumed for the iKs. Only the states with
ix = 0 are eliminated. In this case, B fires always upon the last unit of A but 2 states are con-
sidered
VA]B(2, 1;0, 1) = pB K^(2.0;0. 1) + qB VAlA(2, 1,0, 1)
and
VMB{\. 2;0, 1) = pB VA]A (1, 1;0, 1) + qB VaU (2, 1;0, 1).
In this case VA = 0.5553 for pA = pB = 0.5.
The total system consists of 35 pairs of equations and is solved by iterations.
Table 4 gives some results for the square duel in this last case. As in the two preceeding
cases, concentrated firing is better.
An extension of this last case is considered in the next section.
6. MULTIPLE FAIR DUELS. K HITS NECESSARY TO DESTROY
Let us restrict ourselves to the case of a fair duel, i.e., one such that M = N = n,
Pa= Pb = P and KA = KB = K.
All nondestroyed units of A concentrate their fire on a single unit of B, volley after volley
until destruction is achieved. For the next volley they concentrate their fire on the next undes-
troyed unit of B.
There are nK + '.
1 possible states for B
K, K,
.., K
K - 1 K,
.., K
1, K,
.., K
0, K,
..., K
On the other hand B spreads its fire over all units of A and all states are possible, eliminating
only the destroyed units.
Since there are K"~j different states with j zeros the number of possible states for A is
This means that in order to find VA (K, ... , K;K, ... , K) we will have to solve a linear sys-
tem consisting of {nK + 1) (Kn+1 - l)/(K - 1) pairs of equations of the form
Va \a (state) = linear combination of VA \B (outcome of A firing)
va\b (state) = linear combination of VA\A (outcome of B firing).
TABLE 4 - (x 104) Square Duel
Number
of
Pa = Pb = P
Va\a
VA\B
Va
Iterations
0.999
9980
40
5010
3
0.99
9809
382
5096
3
0.95
9193
1599
5396
4
0.9
8666
2573
5620
5
0.7
7573
3850
5712
8
0.5
6781
4324
5553
13
0.3
6117
4786
5452
25
0.1
5598
5198
5398
80
0.05
5487
5292
5389
157
0.025
5434
5337
5385
303
0.02
5423
5346
5385
373
0.01
5402
5364
5383
844
Unfortunately, the number of possible states increases very rapidly. A few values are
Number of States
n = 2
35
3
105
4
279
5
693
6
1651
n = 2
91
3
400
4
1573.
This, however, is much better than (K + l)2", which is the number of possible states without
any restrictions.
Since writing down the necessary equations is an impossible task, a computer program was
written to build the equations and solve them by iteration. The main steps are:
(1) define the necessary states,
(2) define VA \A = 0 for all states
Va \b = 0 f°r aN states if B is not destroyed
va \b = 1 if 5 is destroyed.
These will be the initial conditions.
(3) For each state determine the number of effective units MA and NB. If A fires, the
number of targets is T = 1 and the degree of concentration is c = MA . If B fires, the number
of targets is T= min(MA, NB). If MA ^ NB, the degree of concentration is c = 1 and if
NB > MA, then NB = a MA + b and cx = a for Tx = MA- b units and c2 = a + 1 for T2 = b
units.
CONCENTRATED FIRING IN MANY-VERSUS-M ANY DUELS 6V1
(4) Let Qc(i, j) denote the probability for a unit to go from state K = / to state K = j if
submitted to fire of concentration c. Then the matrix Q2, for example, has the form
0
1 2 tf
1
2
K
l-q2
<72
2pq q*
\ x \
V X2W V
In general, for / = 1, K and j = 0, 1, . . . , AT
(/ ^ J Z''"7 <7c-'+y for y ^ 0
1- £&('. 7') fory = 0.
aa y) =
All required matrices are constructed.
5) For each state the equation giving VA \A is constructed.
Let ;' denote the state of the target unit.
Let j denote the states of this unit after A has fired, the rest of B being unaffected.
Then,
VaiaUm)- J:QMaU,j) Va\bUJ)
the corresponding equation for VA\B is of the general form
VAlB(ihi2, .... iT-B) = £ n Qce (4 Je)\ Va\A<JIs, ...,JTS\B).
For example,
K4|5(l,2,0,0,0;l,2,2,2,2)= £ 02(U,) Q3(2j2) VAU OiJ2.0f0f0;lf 2, 2, 2, 2).
y 1-0.1
6) When all possible states are gone through, the last calculated value is
VAlB(K,K, .... /s:;/r, ..., K).
It is compared, usually within 10-6, to the previously calculated value and the process is
iterated until convergence is achieved.
Table 5 gives results for several values of M and K. The dimension of the linear system
is twice the number of states. The probability of a hit is taken as p = 0.5. Time is given for
some cases. The computer used was a CDC6400.
The value p = 0.5 was chosen because time increases very fast if p decreases, as is seen
from Table 4.
TABLE 5
- (K^ x 104), Multiple Fair duel. Exact Results
Number
Number
Ti
M= N
K
of
Equations
of
Iterations
Va
(in seconds)
2
2
70
13
5553
3
182
15
5988
4
378
17
6364
5
682
19
6661
3
2
210
13
5537
3
800
15
6211
4
2210
17
6822
5
4992
19
7289
1700
4
2
558
13
5152
3
3146
15
6132
4
11594
17
6872
6141
5
2
1386
12
4429
3
11648
15
5931
8960
Since exact calculations of VA become too time consuming, some results were obtained by
simulation. Table 6 gives some results. The number of trials was 2000 for p ^ 0.5 and 6000
for p = 0.5, A started the duel in half the cases.
TABLE 6 - (VA x 103),
Multiple Fair Duel. Simulation Results
p
0.1
0.3
0.5
0.7
0.9
M
K
2
2
550
542
561
569
564
4
574
562
522
485
403
6
583
490
351
148
4
8
580
382
120
2
0
10
562
249
11
0
0
2
3
600
600
611
622
575
4
652
632
599
628
642
6
672
606
564
494
270
8
705
554
444
181
2
10
725
482
233
6
0
2
4
586
616
642
691
778
4
715
694
685
666
722
6
774
700
653
672
651
8
796
684
608
548
205
10
797
643
524
204
1
2
5
630
646
674
708
705
4
754
753
760
748
556
6
812
786
725
665
852
8
838
777
668
698
601
10
878
740
639
594
134
CONCENTRATED FIRING IN MANYVERSUS-MANY DUELS 693
We note that for large values of p the behaviour of VA is erratic. This is due to the deter-
ministic issue of a battle for p = 1 as a consequence of alternative firing. For example, if
M = 6, k = 2 and A starts firing, the sequence of states is B. 022222, A: 211111, B: 002222, A:
210000, B: 000222, B wins.
Two independent estimates of the error can be made; one by comparing the results of the
simulation with the calculated values in Table 5 for p = 0.5, M = N = 2 or 4 and K = 2,3,4,
giving 5 = 0.0093, the other estimate is given by assuming a binomial distribution with 6000
trials giving s = 0.0065. To be on the safe side one can conclude that concentrated firing is
better if the simulation gives VA ^ 0.519 and parallel firing is better if the simulation gives
VA ^ 0.481. This does not take into account the bias introduced by alternate firing for "large"
values of p. Since the sign of the bias is evident, one can adjust one's conclusions, for example
for M = 10, K = 4 and p = 0.5 the observed value 0.524 is pulled down and almost certainly A
wins more often than B. On the other hand for M = 8, K = 3 and /? = 0.5 the value 0.444 is
certainly pulled down and one can hardly conclude that B wins more often.
Table 7 summarizes all the results obtained.
TABLE 7 — Better Strategy of Firing
Concentrated
Parallel
Border cases
/> = 0.1
K > 2
K= 1
/> = 0.3
K= 3
2 ^ M < 4
tf= 2 M^l
tf=2 M=5or6
K = 3
2 < M ^ 8
# = 3 m = 9 or 10
tf = 4,5
2 < M ^ at least 10
/> = 0.5
K = 2
2 < M ^ 3
K=2 M > 5
# = 2 M= 4
K=3
2 < M ^ 6
/<: = 3 M ^ 7
K = 4
2 < M ^ 10
K=5
2 < M < at least 10
One can conclude that concentrated firing is better if the combination of group size and
hit probability does not produce a high degree of overkilling. For K > 2 a rough rule could be
concentrate firing if pM ^ K (the exception is p = 0.5, # = 4 and M = 9 or 10).
Up to this time we have compared two strategies: parallel firing and concentrated firing.
In the next section we will attempt to define the concept of partial concentration.
7. MULTIPLE FAIR DUELS. PARTIAL CONCENTRATION OF FIRE
Let M = N = n, pA = pB = p and KA = KB = K. Let cx be the maximal number of non-
destroyed units of A that are allowed to concentrate their fire on a single unit of B, volley after
volley until destruction is achieved.
firing.
If cx = I, A uses parallel firing in the same manner as B. If cx = n, A uses concentrated
Under partial concentration the number of targets for A is given by the integer function
f n + cx -
and the number of possible states for B is
(»- ta + D* " + -
which reduces to «AT + 1 for TA = 1 and (Ar"+1 — 1)/ (K — 1) for 7^ = n. The number of
linear equations to be solved becomes
(n- TA + 1) K '
The value cx = \ (TA = n) was used to determine the precision of the obtained results,
since VA = 0.5. For p = 0.5 the maximum error found was 3 x 10~5 and for p = 0.1 it was
1 x 10~4.
Table 8 gives some calculated results for p = 0.5.
TABLE 8
— (K^ x 104), Partial Concentration
Number
Number
M= ./V
K
cX
of
Equations
of
Iterations
VA
3
2
2
330
13
5404
3
2
1760
15
5950
4
2
6290
16
6412
4
2
2
930
12
5639
3
930
13
5192
3
2
7502
14
6304
3
7502
15
6127
5
2
2
3906
12
5541
3
2394
12
5311
4
2394
12
4523
Comparing the results of Table 5 and Table 8, one sees that partial concentration with
cx = 2 is better than total concentration for the cases M = 4 and k = 2 or 3 and any partial
concentration is better for the case M = 5 and k = 2. Further investigations are needed.
8. SUMMARY
The proposed model is an idealization of combat between small groups of individual
identical firers and is very far from the very complicated process of real combat. However, it
has provided, through the use of alternate firing as an expression for the interaction between
opponents, some important insights into combat dynamics that could be further investigated
with, for example, a high-resolution Monte Carlo simulation. It has been shown that alternate
firing gives the same results for small hit probabilities as some previously developed models. It
has also been shown that the relationship between the size, the hitting capacity and the resis-
tance of the opponents is a complex one and that concentrated firing is better than alternate
firing if the amount of over-killing is not too high. Moreover, some evidence suggests that par-
tial concentration can be even more effective.
CONCENTRATED FIRING IN MANY-VERSUS-M ANY DUELS 695
ACKNOWLEDGMENTS
This research was supported by an FIR grant from Universite du Quebec a Montreal. The
author wishes to thank the Service de lTnformatique for its help in providing computing facili-
ties and also the referee and an Associate Editor for their many helpful comments.
REFERENCES
[1] Ancker, C.J., Jr., "The Status of Developments in the Theory of Stochastic Duels-II,"
Operations Research 75, 388-406 (1967).
[2] Ancker, C.J., Jr. and T. Williams, "Some Discrete Processes in the Theory of Stochastic
Duels," Operations Research 13, 202-216 (1965).
[3] Ancker, C.J., Jr., "Stochastic Duels with Bursts," Naval Research Logistics Quarterly 23,
703-711 (1976).
[4] Bhashyam, N., "Stochastic Duels with Lethal Dose," Naval Research Logistics Quarterly
17, 397-405 (1970).
[5] Boole, G., A Treatise on the Calculus of Finite Differences (MacMillan and Co., London,
1860).
[6] Brown, R.H., "Theory of Combat: The Probability of Winning," Operations Research 11,
418-425 (1963).
[7] Dolansky, L., "Present State of the Lanchester Theory of Combat," Operations Research
12, 344-358 (1964).
[8] Helmbold, R.L., "A Universal Attribution Model," Operations Research 14, 624-635
(1966).
[9] Helmbold, R.L., "Solution of a General Non- Adaptive Many versus Many Duel Model,"
Operations Research 16, 518-524 (1968).
[10] Lanchester, F.W., "Aircraft in Warfare: The Dawn of the Fourth Arm-No.V, The Principle
of Concentration," Engineering 98, 422-423 (1914) (reprinted on pp. 2138-2148 of the
World of Mathematics, J. Newman, Editor (Simon and Schuster, New York, 1956).
[11] Lanchester, F.W., Aircraft in Warfare; the Dawn of the Fourth Arm, (Constable and Co.,
London, 1916).
[12] Robertson, J.I., "A Method of Computing Survival Probabilities of Several Targets versus
Several Weapons," Operations Research 4, 546-557 (1956).
[13] Taylor, J., "Solving Lanchester-Type Equations for 'Modern Warfare' with Variable
Coefficients," Operations Research 22, 756-770 (1974).
[14] Taylor, J., "Optimal Commitment of Forces in Some Lanchester-Type Combat Models,"
Operations Research 27, 96-114 (1979).
[15] Taylor, J. and S. Parry, "Force-Ratio Considerations for some Lanchester-Type Models of
Warfare," Operations Research 23, 522-533 (1975).
[16] Thompson, D.E., "Stochastic Duels Involving Reliability, Naval Research Logistics Quar-
terly 19, 145-148 (1972).
[17] Williams, T., "Stochastic Duels-II," System Development Corporation Document, SP
1017/003/00, 31-61 (1963).
SPIKE SWAPPING IN BASIS REINVERSION*
R. V. Helgason and J. L. Kennington
Department of Operations Research
and
Engineering Management
Southern Methodist University
Dallas, Texas
ABSTRACT
During basis reinversion of either a product form or elimination form linear
programming system, it may become necessary to swap spike columns to effect
the reinversion and maintain the desired sparsity characteristics. This note
shows that the only spikes which need be examined when an interchange is re-
quired are those not yet processed in the current external bump.
I. INTRODUCTION
An important component of a large scale linear programming system is the reinversion
routine. This paper addresses an important ancillary technique for implementing a reinversion
routine utilizing the pivot agenda algorithms of Hellerman and Rarick [5,6]. Production of fac-
tors during reinversion typically involves a left-to-right pivoting process. Unfortunately, during
the left-to-right process, a proposed pivot element of a spike column may be zero, in which
case columns are interchanged in an attempt to obtain a pivotable column while maintaining
desired sparsity characteristics. In this paper we show that the only columns which need be
considered for the interchange with a nonpivotable spike are other spikes lying to the right
within the same external bump.
II. PRODUCT FORM OF THE INVERSE
Let B be any m x m nonsingular matrix. One of the most common factorizations for B~{
is the product form which corresponds to the method for solving a system of linear equations
known as Gauss- Jordan reduction (see [3, 4]). This procedure is used to represent B~x (or a
row and column permutation of B~l) as the product of matrices each of the form
/
Z = z , «— y'th row
'This research was supported in part by the Air Force Office of Scientific Research under Contract Number AFOSR
77-3151.
698 R.V. HELGASON AND J.L. KENNINGTON
where z is an m-component column vector, and j is called the pivot row. A few observations
concerning Z are obvious.
PROPOSITION 1: Z is nonsingular if and only if z, ^ 0.
PROPOSITION 2: Let 3 be any m-component vector having Bj = 0. Then ZB = B.
PROPOSITION 3: Let B be any m-component vector having Bj ^ 0, and let ej denote
the vector having yth component 1 and all other components zero.
\-Bk/Bj, ifk*A
Letz*=( 1//3,, if*-; )• ThenZ^-e/.
Let B(i) denote the rth column of the matrix B. Consider the following algorithm.
ALG 1: Product Form Factorization
0. Initialization
Interchange columns of B, if necessary, so that the first component of 5(1) is nonzero.
Set / «- 1,0 «- 5(1), and go to 3.
1. Update Column
Setfi^ Ej~x ...EXBU).
2. Swap Columns If Pivot Element Equals Zero
If Bj ^ 0, go to 3; otherwise, there is some column B(j) with j > i such that the fth
component of y — E'~l . . . ElB{j) is nonzero. Interchange B(J) and BU) and set/8 «— y.
3. Obtain New Elementary Matrix
Set
1//3,, for A: = i
-Bk/Bj, otherwise,
E'+-
4. Test for Termination
If / = m, terminate; otherwise, i «— / + 1 and go to 1. At the termination of ALG 1,
Em . . . E1 is a row permutation of B~l.
In the following two propositions we show that if in Step 2, fi, = 0, then the proposed
interchange is always possible. Consider the following:
PROPOSITION 4: For / ^j,EJ... ElB(i) = e1.
SPIKE SWAPPING IN BASIS REINVERSION
699
PROOF: By the construction of E' and Proposition 3, E' ... E]B(i) = e'. By Proposition
2, EJ ... Ei+le'— e'. So EJ . . . ExB(i) = e'. Using Proposition 4 we may now show the fol-
lowing:
PROPOSITION 5: For 2 < / < m, let B = E'~
j > i such that [EHl . . . ExB(j)]i * 0.
. ElB(i). If Bi = 0, there is some
PROOF: Suppose [E'~l ... ElB(j)]i = 0 for all j > i. By the construction of
Ex, ... E'~l, in ALG 1, and Proposition 1, each factor is nonsingular. Since B is nonsingular,
E'~x ... EXB is nonsingular. By Proposition 4, E'~x ... EXB(J) = ej for 1 ^ j ^ / — 1.
Hence, the /th row of E/_1 ... EXB is all zero, a contradiction.
III. BUMP AND SPIKE STRUCTURE
In order to minimize the core storage required to represent the ETA file, i.e.,
El, ... , Em, the rows and columns of B are interchanged in an attempt to place B in lower tri-
angular form. If this can be accomplished, then the m nonidentity columns of E1, ... , Em,
have the same sparsity structure as B. Consider the following proposition:
Et
PROPOSITION 6: If the first j -
...EXB(J) = B(J).
1 components of B(j) are zero for j > 2, then
PROOF: This follows directly from successive application of Proposition 2. Therefore, if
B is lower triangular, the factored representation of B~x may be stored in approximately the
same amount of core storage as B itself. In practice it is unneccessary to calculate the elements
\/Bk and -8j/Bk in Step 3 of ALG 1. It suffices to store k and the elements of Bt. It may
prove advantageous to store \/Bk, in addition. If Proposition 6 applies for B(k), then
B = B(k) and the only additional storage required is for the index k (and possibly \/Bk).
Clearly, this results in substantial core storage savings compared to storing B~x explicitly.
If B cannot be placed in lower triangular form, then it is placed in the form:
Bx
Hi
B2
111
te
2?3
where Bl and B3 are lower triangular matrices with nonzeroes on their diagonals. We assume
that if B2 is nonvacuous, every row and column has at least two nonzero entries, so that no
rearrangement of B2 can expand the size of Bx or B2. B2 is called the bump section, the merit
section or the heart section. We further require the heart section to assume the following form:
B2-
Fx
Ijjg
Gl
^m
F2
^nn
G2
^^^^^^^^^^
mm
HHt§
b§^
*■ i
700 R.V. HELGASON AND J.L KENNINGTON
where Gk's are either vacuous or lower triangular with nonzeroes on the diagonal. The only
partitions in B having columns with nonzeroes above the diagonal are the /*'s which are called
external bumps. The columns extending above the diagonal are called spikes or spike columns.
An external bump is characterized as follows:
(i) the last column of an external bump will be a spike with a nonzero lying in the top-
most row of the external bump, and
(ii) the nonspike columns have nonzero diagonal elements.
The algorithms of Hellerman and Rarick [5,6] produce such a structure for any nonsingular
matrix, and we shall call a matrix having this structure an HR matrix. It should be noted that if
one applies ALG 1 to an HR matrix, then the only columns which may require an interchange
are spike columns. We now prove that the only columns which need be considered for this inter-
change are other spikes in the same external bump.
Consider the following result:
PROPOSITION 7: Let B(i) with / > 2 correspond to the first column of some external
bump, /* and let B(j) be a spike in Fk. Then EHl . . . ElB(j) = B(J).
PROOF: Note that the first i - 1 components of B(j) are zero. Therefore, by successive
application of Proposition 2, the result is proved.
Note that Proposition 6 allows one to eliminate all of the calculation required in Step 1 of
ALG 1 for nonspike columns and Proposition 7 allows one to eliminate some of this calculation
for spikes. We now address the issue of spike swapping. Consider the following propositions:
PROPOSITION 8: Any spike B(J) which is not pivotable cannot be interchanged with a
spike B(k), k > j, from another external bump, to yield a pivotable column.
PROOF: Since B(k) is from an external bump lying to the right of the external bump
containing B(j), Bj(k) = 0. By repeated application of Proposition 2, Ej~x...Ex
B(k) = B(k). Thus B(j) cannot be interchanged with B(k) to yield a pivotable column.
PROPOSITION 9: Any spike B(J) which is not pivotable cannot be interchanged with a
nonspike column B(k), k > y, to yield a pivotable column.
PROOF: Let B(k), with k > j correspond to any nonspike column. From Proposition 6,
Ej~x ... ExB(k) = B(k). Since the yth component of B(k) is zero, B(j) cannot be inter-
changed with B(k), to yield a pivotable column. We now present the main result of this note.
PROPOSITION 10: Any spike column B(j), which is not pivotable can be interchanged
with a spike, B(k), with k > j within the same external bump, to yield a pivotable column.
PROOF: If B(J) is not pivotable, then by Proposition 5 there exists a column Bik) with
k > j which is pivotable. By Proposition 8, Bik) cannot be a spike from a different external
bump. By Proposition 9, B (k) cannot be a nonspike. Hence B(k) must be a spike from the
same external bump.
SPIKE SWAPPING IN BASIS REINVERSION 701
In practice, the zero check in step 2 is replaced by a tolerance check. Discussions of prac-
tical tolerance checks may be found in Benichou [1], Clasen [2], Orchard-Hays [7], Saunders
[8], Tomlin [9], and Wolfe [10].
REFERENCES
[1] Benichou, M., J. Gauther, G. Hentges, and G. Ribiere, "The Efficient Solution of Large-
Scale Linear Programming Problems— Some Algorithmic Techniques and Computa-
tional Results," Mathematical Programming, 13, 280-322 (1977).
[2] Clasen, R.J., "Techniques for Automatic Tolerance Control in Linear Programming," Com-
munications of the Association for Computing Machinery, 9, 802-803 (1966).
[3] Forsythe, G.E. and C.B. Moler, Computer Solution of Linear Algebraic Systems, (Prentice-
Hall, Englewood Cliffs, New Jersey, 1967).
[4] Hadley, G. Linear Algebra (Addison Wesley Publishing Co., Inc., Reading, Massachusetts,
1964).
[5] Hellerman, E. and D. Rarick, "The Partitioned Preassigned Pivot Procedure CP4)," Sparse
Matrices and Their Applications, D. Rose and R. Willoughby, Editors, (Plenum Press,
New York, New York, 1972).
[6] Hellerman, E. and D. Rarick, "Reinversion with the Preassigned Pivot Procedure,"
Mathematical Programming, 1, 195-216 (1971).
[7] Orchard-Hays, W., Advanced Linear Programming Computing Techniques, (McGraw-Hill,
New York, New York, 1968).
[8] Saunders, M.A., "A Fast, Stable Implementation of the Simplex Method Using Bartels-
Golub Updating," Sparse Matrix Computations, 213-226, J.R. Bunch and D.J. Rose, Edi-
tors (Academic Press, New York, New York, 1976).
[9] Tomlin, J. A., "An Accuracy Test for Updating Triangular Factors," Mathematical Program-
ming Study 4, M.L. Balinski and E. Hellerman, Editors, (North-Holland, Amsterdam,
1975).
[10] Wolfe, P., "Error in the Solution of Linear Programming Problems," Error in Digital Com-
putation, 2, L.B. Rail, Editor (John Wiley and Sons, Inc., New York, New York 1965).
AN ALTERNATIVE PROOF OF
THE IFRA PROPERTY OF SOME SHOCK MODELS*
C. Derman and D. R. Smith
Columbia University
New York, New York
Let Hit) = £ , {t) P(k), 0 < t < oo, where A it)/t is nonde-
_*=0 k- _
creasing in t, {Pik)xlk\ is nonincreasing. It is known that Hit) = 1 - H(t) is
an increasing failure rate on the average (IFRA) distribution. A proof based
on the IFRA closure theorem is given. Hit) is the distribution of life for sys-
tems undergoing shocks occurring according to a Poisson process where P(k) is
the probability that the system survives k shocks. The proof given herein
shows there is an underlying connection between such models and monotone
systems of independent components that explains the IFRA life distribution oc-
curring in both models.
1. INTRODUCTION
In Barlow and Proschan [1, p. 93] a fairly general damage model is considered. A device
is subject to shocks occurring in time according to a Poisson process with rate k. The damage
caused by shocks is characterized by a sequence of numbers {Pik)}, where P(k) is the proba-
bility that the device will survive k shocks. The Pik)'s as shown in [1] can arise in different
models. For example, the damage caused by the rth shock can be assumed to be a nonnegative
random variable Xh where X\, X2 ... are independent and identically distributed; failure of the
k
device occurs at the /cth shock if £ Xh the cumulative damage, exceeds a certain thres-
hold. In this case Pik) = /VJ£ Xj ^ vL where y is the threshold. Ross [2] has failure occur-
ring when some nondecreasing symmetric function D(X\, .... X„) first exceeds a given thres-
hold; i.e., DiXx, ... , A"*) is a generalization of £ Xr Here, P(k) = Pr [D(XX, . . . Xk) < y\.
t=\
Let Hit) denote the probability that the device survives in the interval [0, t]. Then
Hit) - £ J
*=o
In Barlow and Proschan [1] (Theorem 3.6 p. 93) it is proven that if {Pik)]/k} is a nonincreasing
sequence then Hit) = 1 - Hit) is always an increasing failure rate on the average (IFRA)
*Work supported in part by the Office of Naval Research under Contract N0014-75-0620 and the National Science
Foundation under Grant No. MCS-7725-146 with Columbia University.
703
704 C. DERMAN AND DR. SMITH
distribution function; i.e., is nondecreasing in t.
Ross [2], generalizes by allowing the Poisson process of successive shocks to be nonho-
mogeneous with rate function \ (/) such that
A it) _ Slx{s)ds
t t
is nondecreasing in /. That is, the same assertion can be made when Hit) is given by
™Ai
_ OO -A(t) A (f\k _
(1) Hit) = £- rj^-Pik), 0 ^ t < -
*=0
The proof given in [1] is based on total positivity. arguments. Ross's technique for prov-
ing the IFRA result is obtained by making use of recent results [3] pertaining to what he calls
increasing failure rate average stochastic processes.
Our proof below shows that all such results are a consequence of one of the central
theorems of reliability theory, the IFRA Closure Theorem ([1] p. 83). This theorem asserts
that a monotone system composed of a finite number of independent components, each of
which has an IFRA life distribution, has itself an IFRA distribution.
It is remarked in [1, p. 91] that the coherent (or monotone) system model and the shock
models under consideration are widely diverse models for which the IFRA class of distribution
furnishes an appropriate description of life length, thus reenforcing the importance of the IFRA
class to reliability theory. The implication of our proof is that the models are not as widely
diverse as supposed.
The idea of the proof is the construction of a monotone system (of independent com-
ponents, each of which has the same IFRA life distribution) whose life distribution approxi-
mates Hit). The proof is completed by allowing the number of components in the system to
increase in an appropriate way so that the approximating life distributions converge to Hit)',
the IFRA property being preserved in the limit.
2. APPROXIMATING SYSTEMS APPROACH
For each m, m = 1,2 ... let Sm„, n = 1,2, ... be a monotone system of n independent
components. Let
(1) Pmnik) = Pr {no cut set is formed | exactly k components of Sm „ are failed}
where all of the n components are equally likely to fail. (A cut set is a set of components such
that if all components of the set fail, the system does not function) . Assume
(2) Pm,„ik) = 0, if k > m for every n,
(3) lim Pm_n ik) = Pmik), for every k
(4) lim Pmik) = Pik), for every k.
We can state
IFRA PROPERTY OF SOME SHOCK MODELS
Hit) = 1 - Hit) given by (1) is IFRA.
PROOF: Assume every component in Smn is independent with life distribution
Lit) = 1 - e~AU>/". Then every component has an IFRA distribution. Let Qmnik,t) denote
the probability that exactly k units fail within [0, t]. That is
M( =*mki =m.\*-*
(5) Qm,„ (k,0 = L U - e " ) U " J .
Let Hm nit) denote the probability that Smi„ works for at least / units of time, then
(6) Hmjlit)= £ Qmj,ik,t)PmJk).
k=0
By the IFRA Closure Theorem, Hm,„it) is IFRA.
However,
(7) Hmit) = lim Hmj,it)
= £ Hm Qm,nik,t) Pjk)
= 1
k=0
i
k\
k = 0
by (2), the Poisson limit of binomial probabilities, and (3). Since the IFRA_property is
preserved in the limit, Hmit) is IFRA. That is, since Hmit) = lim Hmnit) and
-(log Hm nit))/t\s nondecreasing in t, then so is -(log Hmit))/t. However,
-Ait)k lim Pmik)
k=o K-
= Hit).
Since again the IFRA property is preserved in the limit, it follows that Hit) is IFRA, proving
the theorem.
We emphasize that the IFRA Closure Theorem is invoked only to show that that Hmnit)
is IFRA. The condition that A it)/ 1 is nondecreasing is needed so that all components of Sm „
have an IFRA distribution.
3. APPLICATION OF THEOREM
The condition that [Pik)xlk) is a nonincreasing sequence is not used in the proof nor does
it appear in the statement of Theorem 1. That the condition is implicit is due to a recent
remarkable result of Ross, Shashahani and Weiss [4] that [Pik)xlk) is necessarily nonincreasing.
706 C DERMAN AND DR. SMITH
To apply Theorem 1 for our purpose we must show
THEOREM 2: Let {P(k)} be any sequence such that 0 < P(k) ^ 1 and {P(k)Uk} is
nonincreasing. Then there exist the monotone systems [Smn) such that (2), (3), and (4) hold.
PROOF: Let {P(k)} be any sequence with the hypothesized properties. Let F be any
increasing continuous distribution function over [0, °°) and {yk} the nonincreasing sequence of
nonnegative numbers such that
F(yk) = P(k)Uk, k = 1,2, ...
For each m(m < n) let Smn be a set of n components, / = 1, ... , n. The cut sets are con-
structed in the following way. The rth component has an associated value xh / = 1, ... , n
where the values are assigned so that
#{/U, ^ x) = [n Fix)], 0 < x ^ v,,
= n, x > y\,
where # means "number of" and [ ] is the greatest integer designator. Every set of k com-
ponents is a cut set if k > m\ if k ^ m a set (i\ ik) of components is a cut set if and
only if
max (x/, . . . , Xj ) > yk.
Since [yk] is nonincreasing, Sm „ is, indeed, a monotone set. But here,
fe-l [nF(yk)] - i
Pm.n(k) = n ^3- , k ^ m
■■ 0 , k > m.
Thus,
Pm(k) = lim Pmj,(k)
Fk(yk), if k < m
0 , if k > m
lim Pm(k) = Fk{yk)
= P{k) , k = 1,2, ... .
This proves Theorem 2.
Theorems one and two yield the slightly more general version of Theorem 3.6 [1, p. 93].
The Ross [2] generalization follows by defining the cut sets to be determined by a nonde-
creasing symmetric function D(x]t .... xk)\ i.e., a set ih .... ik of components is a cut set of
Smj, if k > m or, if k ^ m, when D(xit ... , xk) > v, a given threshold value. From the
construction of Theorem 2, Theorem 1 and the result referred to in [4] it follows that the
sequence {P(k)} of this model satisfies the monotonicity condition. For the special case of
k _
D{X\, ... , Xk) = £ Xh it is known that the sequence {P(k)Uk} is nonincreasing (see [1] p.
96).
IFRA PROPERTY OF SOME SHOCK MODELS 707
REFERENCES
[1] Barlow, R. and F. Proschan, Statistical Theory of Reliability and Life Testing, Probability
Models, (Holt, Rinehart and Winston, New York, 1975).
[2] Ross, S.M., "Generalized Poisson Shock Model," Technical Report, Department of Indus-
trial Engineering and Operations Research, University of California, Berkeley, California
(1978).
[3] Ross, S.M., "Multivalued State Component Systems," Annals of Probability (to appear).
[4] Ross, S.M., M. Shashahani and G. Weiss, "On the Number of Component Failures in Sys-
tems whose Component Lives are Exchangeable," Technical Report, Department of Indus-
trial Engineering and Operations Research, University of California, Berkeley, California.
NEWS AND MEMORANDA
Defense Systems Management College
Military Reservist Utilization Program
Military reservists from all U.S. Services now have a unique opportunity for a short tour
at the Defense Systems Management College, Ft. Belvoir Virginia. By volunteering for the
Reservist Utilization Program, an individual can increase proficiency training, maintain currency
in DOD Research, Development & Acquisition Policy, contribute to the development and for-
mulation of concepts that may become the bases of future DOD policy and help solve critical
problems facing the acquisition community.
Once accepted for the program, a reservist may be assigned to one of three areas:
research, education or operations. As a research associate, the individual researches and
analyzes an area compatible with his training and experience. Many reservists in this category
currently assist in the preparation of material for a comprehensive textbook on systems acqusi-
tion. The text will be used at DSMC by the faculty and students as well as by the systems
acquisition community. As an academic consultant, a reservist provides special assistance to
the College faculty by reviewing course material in his area of expertise and researching and
developing training materials. In the operations/administration category, reservists administer
the program by recruiting other reservists for the program, processing these reservists, and
maintaining files and records.
Because of the complexity and broad scope of the systems acquisition business, the Reser-
vist Utilization Program requires a large number of reservists from many diverse career fields.
Some examples of career fields used include: engineering, procurement, manufacturing, legal,
financial, personnel, administration and logistics. Reservists whose reserve duty assignments
are not in these types of career fields, but who have civilian experience in these areas, are also
urged to apply.
Many reservists perform their annual tours with the Reservist Utilization Program office.
Others perform special tours of active duty or "mandays." When tour dates are determined and
coordinated with your organization and the RUP office, submit the proper forms through your
reserve organization at least 45 days prior to the tour date for an annual tour or 60 days for a
special tour.
To apply for active duty or to get additional information, telephone Professor Fred E.
Rosell, Jr. at commerical (703) 664-5783 or AUTOVON 354-5783. Reservists outside of Vir-
ginia may call on toll-free number (800) 336-3095 ext. 5783.
NEWS AND MEMORANDA
List of Referees
The Editors of the Naval Research Logistics Quarterly are grateful to the following indivi-
duals for assisting in the review of articles prior to publication.
A. Hax
P. Heidelberger
D. P. Heyman
A. J. Hoffman
P. Q. Hwang
E. Ignall
P. Jacobs
A. J. Kaplan
U. Karmarkar
A. R. Kaylan
J. L. Kennington
P. R. Kleindorfer
D. Klingman
J. E. Knepley
K. O. Kortanek
D. Kreps
W. K. Kruse
G. J. Lieberman
S. A. Lippman
D. Luenberger
R. L. McGill
W. H. Marlow
C. Marshall
K. T. Marshall
M. Mazumder
P. McKeown
K. Mehrotra
C. B. Millham
D. Montgomery
R. C. Morey
J. G. Morris
J. A. Muckstadt
S. Nahmias
M. F. Neuts
I. Olkin
J. Orlin
S. S. Panwalkar
J. H. Patterson
M. Posner
D. Reedy
H. R. Richardson
E. E. Rosinger
S. M. Ross
H. M. Salkin
R. L. Scheaffer
B. Schmeiser
P. K. Sen
J. Sethuraman
M. L. Shooman
M. Shubik
D. O. Siegmund
E. Silver
N. D. Singpurwalla
R. Soland
Henry Solomon
Herbert Solomon
R. M. Stark
L. D. Stone
W. Szwarc
H. A. Taha
J. G. Taylor
G. Thompson
W. E. Vesley
H. M. Wagner
A. R. Washburn
C. C. White
T. M. Whitin
J. D. Wiest
J. W. Wingate
R. T. Wong
M. H. Wright
S. Zacks
CORRIGENDUM:
STOCHASTIC CONTROL OF QUEUEING SYSTEMS
Dr. A. Laurinavicius of the Institute of Physical and Technical Problems of Energetics,
Academy of Sciences, Lithuania, USSR, has pointed out an error in the statement of Theorem
1 of this paper [1]. The expression for the generator given there is valid only for x > 0, and a
different expression holds for x = 0, the proof for this case being similar. Moreover, the
domain of the generator can be extended. The correct statement is as follows.
THEOREM 1: Let the function f(t,x) be continuous and such that the directional deriva-
tives
(1) Dp/Ux) = lim /(' + *.*-*)-/(/,*)• (x > 0)
r h-~ o+ h
(2) Bg /a o) - to /<' + *-o)-/ao) _ i± /(t0)
y h~ o+ h at
where P = (1, — 1) and Q = (1,0), exist, be continuous from one side and bounded. Then
the infinitesimal generator of the semigroup {T,} is given by
(3) Af(t.x) = Dp/(t,x) -kfU,x) + X f~f(t.x + v)B(dv) for x > 0
= £>£ /(f,0) - X/U0) + X JJ f{t,\)B{d\) for x = 0.
As a consequence of this error the example of Section 3 does not lead to the stated result.
A correct example is provided by the following. Let r(r), the revenue per unit time, and c(t),
the operating cost per unit time, be given by
r(/) = r for 0 < t <: t0, and = 0 for t > t0
c(t) = C] for 0 ^ t < t0, and = c2 for t > t0.
The profit from operating the system up to a time Tis given by f(T,WT), where
(4) fU.x) = r min(xf0) _ C\h ~ ?! max (0,f + x - f0).
This leads to the following correct version of Theorem 3.
THEOREM 3: Let W0 = w < t0 and assume that
(5) Xc2 r°° [1 - B(\)]dv < r < Xc2/3
where /3 is the mean service time. Then the optimal time is given by
(6) Ta = inf{/ > 0: t + W, > a)
where a is the unique solution of the equation
(7) Xc2 f~ [l-fi(v)]rfv= r.
•"o-a
711
712 CORRIGENDUM
PROOF: It is found that for x > 0
C IfUx + v) - f(t,x)]B(dv) = -c2 C . [1 - £(v)]rfv
where (r0 - f - x)+ = max (0,f0 - t - x). Also,
Dp /(f,x) = r for / < r0, and - 0 for / ^ t0 (x > 0)
Dq fUO) = r for t < t0, and = -c2 for r ^ f0.
Therefore, the generator in this case is given by
Af(t,x) = r -Kc2 f °° x [1 - B(\)]dv for / < t0, x > 0
J(t0-t-x) +
= -x2/3 for r ^ f0. x > 0
(8) = - c2 - XC2/3 for f ^ r0, x = 0.
In applying Theorem 2 we note that Af(t.x) < 0 for / ^ f0> * ^ 0, so it suffices to consider
Af(t,x) for / < /0> x ^ 0. We can write
Af(t.x) = (f>(t + x) for / < /0, x > 0,
where
(9) 0(0 = r - \c2 J°° [1 - B(v)]dv.
(to t)
We have
0(0) = r - Ac2 J*°° [1 - B(v)]dv > r - Xc2 J"~w [1 - fl(v)]</v > 0
<t>(t0) = r -\c2B < 0
on account of (5). Also, <f>(t) is a decreasing function of t. Therefore, there exists a unique
value a such that <f>(t) > 0 for 0 < t < a and <f>(t) < 0 for a < t ^ t0. Since <f>(t) ^ 0 for
f ^ /0, we have <f>(t) < 0 for r ^ a. This means that Af(t,x) < 0 for / + x ^ o, so the set
/? of Theorem 2 is given by R = [(t,x): t + x ^ a}, and the time of the first visit to R is
given by (6). Since the process t + W, is monotone nondecreasing with probability one, the
set R is closed. Moreover, Ta ^ a with probability one and also E(Ta) < <». Thus, the condi-
tions of Theorem 2 are satisfied, and Ta is optimal at fVQ = w, as was required to be proved.
A particular case. Let B(x) = 1 - e'** (x ^ 0, 0 < fi < °°). The conditions (5) reduce
to
(10) w < t0 log l^yl < /„
and the Equation (7) gives
(11) a=t0-— log — .
H [ fir J
On account of (11) we have a > w.
REFERENCE
[1] Prabhu, N.U., "Stochastic Control of Queueing Systems," Naval Research Logistics Quar-
terly 21, 411-418 (1974).
N.U. Prabhu
Cornell University
INDEX TO VOLUME 27
ALBRIGHT, S.C., "Optimal Maintenance-Repair Policies for the Machine Repair Problem," Vol. 27, No. 1, March
1980, pp. 17-27.
ANDERSON, M.Q., "Optimal Admission Pricing Policies for M/Ek/\ Queues," Vol. 27, No. 1, March 1980, pp. 57-64.
BALCER, Y., "Partially Controlled Demand and Inventory Control: An Additive Model," Vol. 27, No. 2, June 1980,
pp. 273-280.
BARD, J.F. and J.E. Falk, "Computing Equilibria Via Nonconvex Programming," Vol. 27, No. 2, June 1980, pp. 233-
255.
BAZARAA, M.S. and H.D. Sherali, "Benders' Partitioning Scheme Applied to a New Formulation of the Quadratic As-
signment Problem," Vol. 27, No. 1, March 1980, pp. 29-41.
BEN-TAL, A., L. Kerzner and S. Zlobec, "Optimality Conditions for Convex Semi-Infinite Programming Problems,"
Vol. 27, No. 3, September 1980, pp. 413-435.
BERREBI, M. and J. Intrator, "Auxiliary Procedures for Solving Long Transportation Problems," Vol. 27, No. 3, Sep-
tember 1980, pp. 447-452.
BOOKBINDER, J.H. and S.P. Sethi, "The Dynamic Transportation Problem: A Survey," Vol. 27, No. 1, March 1980,
pp. 65-87.
CALAMAI, P. and C. Charalambous, "Solving Multifacility Location Problems Involving Euclidean Distances," Vol.
27, No. 4, December 1980, pp. 609.
CHANDRA, S. and M. Chandramohan, "A Note of Integer Linear Fractional Programming," Vol. 27, No. 1, March
1980, pp. 171-174.
CHANDRAMOHAN, M. and S. Chandra, "A Note on Integer Linear Fractional Programming," Vol. 27, No. 1, March
1980, pp. 171-174.
CHARALAMBOUS, C. and P. Calamai, "Solving Multifacility Location Problems Involving Euclidean Distances," Vol.
27, No. 4, December 1980, pp. 609.
CHAUDHRY, M.L., D.F. Holman and W.K. Grassman, "Some Results of the Queueing System E£/M/cV Vol. 27,
No. 2, June 1980, pp. 217-222.
COHEN, E.A., Jr., "Statistical Analysis of a Conventional Fuze Timer," Vol. 27, No. 3, September 1980, pp. 375-395.
COHEN, L. and D.E. Reedy, "A Note on the Sensitivity of Navy First Term Reenlistment to Bonuses, Unemployment
and Relative Wages," Vol. 27, No. 3, September 1980, pp. 525-528.
COHEN, M.A. and W.P. Pierskalla, "A Dynamic Inventory System with Recycling," Vol. 27, No. 2, June 1980, pp.
289-296.
COOPER, M.W., "The Use of Dynamic Programming Methodology for the Solution of a Class of Nonlinear Program-
ming Problems," Vol. 27, No. 1, March 1980, pp. 89-95.
DERMAN, C. and D.R. Smith, "An Alternative Proof of the IFRA Property of Some Shock Models," Vol. 27, No. 4.
December 1980, pp. 703.
DEUERMEYER, B.L., "A Single Period Model for a Multiproduct Perishable Inventory System with Economic Substi-
tution," Vol. 27, No. 2, June 1980, pp. 177-185.
DISCENZA, J.H. and H.R. Richardson, "The United States Coast Guard Computer-Assisted Search Planning System
(CASP)," Vol. 27, No. 4, December 1980, pp. 659.
DISNEY, R.L., DC. McNickle and B. Simon, "The M/G/l Queue with Instantaneous Bernoulli Feedback." Vol. 27,
No. 4, December 1980, pp. 635.
ELLNER, P.M. and R.M. Stark, "On the Distribution of the Optimal Value for a Class of Stochastic Geometric Pro-
grams," Vol. 27, No. 4, December 1980, pp. 549.
ENGELBERG, A. and J. Intrator, "Sensitivity Analysis as a Means of Reducing the Dimensionality of a Certain Class
of Transportation Problems," Vol. 27, No. 2, June 1980, pp. 297-313.
FALK, J.E. and J.F. Bard, "Computing Equilibria Via Nonconvex Programming," Vol. 27, No. 2, June 1980, pp. 233-
255.
GAVER, D.P. and P.A. Jacobs, "Storage Problems when Demand Is 'All or Nothing"' Vol. 27, No. 4, December 1980,
pp. 529.
GLAZEBROOK, K.D., "On Single-Machine Sequencing with Order Constraints," Vol. 27, No. 1, March 1980, pp. 123-
130.
GOLABI, K., "An Inventory Model with Search for Best Ordering Price," Vol. 27, No. 4, December 1980, pp. 645.
GOLDEN, B.L. and J. R. Yee, "A Note on Determining Operating Strategies for Probabilistic Vehicle Routing," Vol.
27, No. 1, March 1980, pp. 159-163.
713
714 index to Volume 27
GRASSMAN, W.K., D.F. Holman and M.L. Chaudhry, "Some Results of the Queueing System E£/M/c*: Vol. 27,
No. 2, June 1980, pp. 217-222.
GREENBERG, I., "An Approximation for the Waiting Time Distribution in Single Server Queues," Vol. 27, No. 2,
June 1980, pp. 223-230.
HANSON, M.A. and T.W. Reiland, "A Class of Continuous Nonlinear Programming Problems with Time-Delayed
Constraints," Vol. 27, No. 4, December 1980, pp. 573.
HANSOTIA, B.J., "Stochastic Linear Programs with Simple Recourse: The Equivalent Deterministic Convex Program
for the Normal Exponential, Erland Cases," Vol. 27, No. 2, June 1980, pp. 257-272.
HAYNES, R.D. and W.E. Thompson, "On the Reliability, Availability and Bayes Confidence Intervals for Multicom-
ponent Systems," Vol. 27, No. 3, September 1980, pp. 345-358.
HELGASON, R.V. and J.L. Kennington, "Spike Swapping in Basis Reinversion," Vol. 27, No. 4, December 1980, pp.
697.
HILDEBRANDT, G.G., "The U.S. Versus the Soviet Incentive Models," Vol. 27, No. 1, March 1980, pp. 97-108.
HOLMAN, D.F., W.K. Grassman and M.L. Chaudhry, "Some Results of the Queueing System E£/M/e*" Vol. 27, No.
2, June 1980, pp. 217-222.
HSU, C.L., L. Shaw and S.G., Tyan, "Optimal Replacement of Parts Having Observable Correlated Stages of Deteriora-
tion," Vol. 27, No. 3, September 1980, pp. 359-373.
INTRATOR, J. and M. Berrebi, "Auxiliary Procedures for Solving Long Transportation Problems," Vol. 27, No. 3, Sep-
tember 1980, pp. 447-452.
INTRATOR, J. and A. Engelberg, "Sensitivity Analysis as a Means of Reducing the Dimensionality of a Certain Class
of Transportation Problems," Vol. 27, No. 2, June 1980, pp. 297-313.
ISAACSON, K. and C.B. Millham, "On a Class of Nash-Solvable Bimatrix Games and Some Related Nash Subsets,"
Vol. 27, No. 3, September 1980, pp. 407-412.
JACOBS, PA. and D.P. Gaver, "Storage Problems when Demand Is 'All or Nothing'," Vol. 27, No. 4, December 1980,
pp. 529.
JOHNSON, C.R. and E.P. Loane, "Evaluation of Force Structures under Uncertainty," Vol. 27, No. 3, September 1980,
pp. 511-519.
KENNINGTON, J.L. and R.V. Helgason, "Spike Swapping in Basis Reinversion," Vol. 27, No. 4, December 1980, pp.
697.
KERZNER, L., A. Ben-Tal and S. Zlobec, "Optimality Conditions for Convex Semi-Infinite Programming Problems,"
Vol. 27, No. 3, September 1980, pp. 413-435.
KORTANEK, K.O. and M. Yamasaki, "Equalities in Transportation Problems and Characterizations of Optimal Solu-
tions," Vol. 27, No. 4, December 1980, pp. 589.
LAW, A.M., "Statistical Analysis of the Output Data from Terminating Simulations," Vol. 27, No. 1, March 1980, pp.
131-143.
LAWLESS, J.F. and K. Singhal, "Analysis of Data from Life-Test Experiments under an Exponential Model," Vol. 27,
No. 2, June 1980, pp. 323-334.
LEV, B. and D.I. Toof, "The Role of Internal Storage Capacity in Fixed Cycle Production Systems," Vol. 27, No. 3,
September 1980, pp. 477-487.
LOANE, E.P. and C.R. Johnson, "Evaluation of Force Structures under Uncertainty," Vol. 27, No. 3, September 1980,
pp. 499-510.
LUSS, H., "A Network Flow Approach for Capacity Expansion Problems with Facility Types," Vol. 27, No. 4, De-
cember 1980, pp. 597.
McKEOWN , P.G., "Solving Incremental Quantity Discounted Transportation Problems by Vertex Ranking," Vol. 27,
No. 3, September 1980, pp. 437-445.
McKEOWN, P.G. and P. Sinha, "An Easy Solution for a Special Class of Fixed Charge Problems," Vol. 27, No. 4, De-
cember 1980, pp. 621.
McNICKLE DC, R.L. Disney and B. Simon, "The M/G/l Queue with Instantaneous Bernoulli Feedback," Vol. 27,
No. 4, December 1980, pp. 635.
MILLHAM, C.B. and K. Isaacson, "On a Class of Nash-Solvable Bimatrix Games and Some Related Nash Subsets,"
Vol. 27, No. 3, September 1980, pp. 407-412.
MORRIS, J.G. and H.E. Thompson, "A Note on the 'Value' of Bounds on EVPI in Stochastic Programming," Vol. 27,
No. 1, March 1980, pp. 165-169.
OREN, S.S. and S.A. Smith, "Reliability Growth of Repairable Systems," Vol. 27, No. 4, December 1980, pp. 539.
PIERSKALLA, W.P. and J. A. Voelker, "Test Selection for a Mass Screening Program," Vol. 27, No. 1, March 1980, pp.
43-55.
RAO, R.C. and T.L. Shaftel, "Computational Experience on an Algorithm for the Transportation Problem with Non-
linear Objective Functions," Vol. 27, No. 1, March 1980, pp. 145-157.
REEDY, D.E. and L. Cohen, "A Note on the Sensitivity of Navy First Term Reenlistment to Bonuses, Unemployment
and Relative Wages," Vol. 27, No. 3, September 1980, pp. 525-528.
REILAND, T.W. and M.A. Hanson, "A Class of Continuous Nonlinear Programming Problems with Time-Delayed
Constraints," Vol. 27, No. 4, December 1980, pp. 573.
RICHARDSON, H.R. and J.H. Discenza, "The United States Coast Guard Computer-Assisted Search Planning System
(CASP)," Vol. 27, No. 4, December 1980, pp. 659.
ROSENLUND, S.I., "The Random Order Service G/M/m Queue," Vol. 27, No. 2, June 1980. pp. 207-215.
INDEX TO VOLUME 27 715
ROSENTHAL, R.W , "Congestion Tolls: Equilibrium and Optimality," Vol. 27, No. 2, June 1980, pp. 231-232
ROSS, G.T., R.M. Soland and A. A. Zoltners, "The Bounded Interval Generalized Assignment Problem," Vol. 27, No.
4, December 1980, pp. 625.
SETHI, S.P. and J.H. Bookbinder, "The Dynamic Transportation Problem: A Survey," Vol. 27, No. 1, March 1980, pp.
65-87.
SHAFTEL, T.L. and R.C. RAO, "Computational Experience on an Algorithm for the Transportation Problem with
Nonlinear Objective Functions," Vol. 27, No. 1, March 1980, pp. 145-157.
SHAPIRO, R.D., "Scheduling Coupled Tasks," Vol. 27, No. 3, September 1980, pp. 489-498.
SHAW, L., C-L. Hsu and S.G. Tyan, "Optimal Replacement of Parts Having Observable Correlated Stages of
Deterioration," Vol. 27, No. 3, September 1980, pp. 359-373.
SHEN, R.F.C., "Estimating the Economic Impact of the 1973 Navy Base Closing: Models, Tests, and an Ex Post
Evaluation of the Forecasting Performance," Vol. 27, No. 2, June 1980, pp. 335-344.
SHERALI, H.D. and M.S. Bazaraa, "Benders' Partitioning Scheme Applied to a New Formulation of the Quadratic As-
signment Problem," Vol. 27, No. 1, March 1980, pp. 29-41.
SHERALI, H.D. and CM. Shetty, "On the Generation of Deep Disjunctive Cutting Planes," Vol. 27, No. 3, September
1980, pp. 453-475.
SHETTY, CM. and H.D. Sherali, "On the Generation of Deep Disjunctive Cutting Planes," Vol. 27, No. 3, September
1980, pp. 453-475.
SIMON, B., R.L. Disney and D.C. McNickle, "The M/G/l Queue with Instantaneous Bernoulli Feedback," Vol. 27,
No. 4, December 1980, pp. 635.
SINGHAL, K. and J.F. Lawless, "Analysis of Data from Life-Test Experiments under an Exponential Model," Vol. 27,
No. 2, June 1980, pp. 323-334.
SINGPURWALLA, N.D., "Analyzing Availability Using Transfer Function Models and Cross Spectral Analysis," Vol.
27, No. 1, March 1980, pp. 1-16.
SINHA, P. and P.G. McKEOWN, "An Easy Solution for a Special Class of Fixed Charge Problems," Vol. 27, No. 4,
December 1980, pp. 621.
SMITH, D.R. and C. Derman, "An Alternative Proof of the IFRA Property of Some Shock Model," Vol. 27, No. 4,
December 1980, pp. 703.
SMITH, S.A. and S.S. Oren, "Reliability Growth of Repairable Systems," Vol. 27, No. 4, December 1980, pp. 539.
SOLAND, R.M., G.T. Ross and A. A. Zoltners, "The Bounded Interval Generalized Assignment Problem," Vol. 27,
No. 4, December 1980, pp. 625.
STARK, R.M. and P.M. Ellner, "On the Distribution of the Optimal Value for a Class of Stochastic Geometric Pro-
grams," Vol. 27, No. 4, December 1980, pp. 549.
TAYLOR, J.G., "Theoretical Analysis of Lanchester-Type Combat between Two Homogeneous Forces with Supporting
Fires," Vol. 27, No. 1, March 1980, pp. 109-121.
THOMPSON, HE. and J.G. Morris, "A Note on the 'Value' of Bounds on EVPI in Stochastic Programming," Vol. 27,
No. 1, March 1980, pp. 165-169.
THOMPSON, W.E. and R.D. Haynes, "On the Reliability, Availability and Bayes Confidence Intervals for Multicom-
ponent Systems," Vol. 27, No. 3, September 1980, pp. 345-358.
TOOF, D.I. and B. Lev, "The Role of Internal Storage Capacity in Fixed Cycle Production Systems," Vol. 27, No. 3,
September 1980, pp. 477-487.
TYAN, S.G., L. Shaw and C-L Hsu, "Optimal Replacement of Parts Having Observable Correlated Stages of Deteriora-
tion," Vol. 27, No. 3, September 1980, pp. 359-373.
VOELKER, J. A. and W.P. Pierskalla, "Test Selection for a Mass Screening Program," Vol. 27, No. 1, March 1980, pp.
43-55.
WASHBURN, A.R., "On a Search for a Moving Target," Vol. 27, No. 2, June 1980, pp. 315-322.
WEISS, L., "The Asymptotic Sufficiency of Sparse Order Statistics in Tests of Fit with Nuisance Parameters," Vol. 27,
No. 3, September 1980, pp. 397-406.
WUSTEFELD, A. and U. Zimmermann, "A Single Period Model for a Multiproduct Perishable Inventory System with
Economic Substitution," Vol. 27, No. 2, June 1980, pp. 187-197.
YAMASAK1, M. and K.O. Kortanek, "Equalities in Transportation Problems and Characterizations of Optimal Solu-
tions," Vol. 27, No. 4, December 1980, pp. 589.
YEE, J.R. and B.L. Golden, "A Note on Determining Operating Strategies for Probabilistic Vehicle Routing," Vol. 27,
No. 1, March 1980, pp. 159-163.
ZIMMERMANN, U. and A. Wustefeld, "A Single Period Model for a Multiproduct Perishable Inventory System with
Economic Substitution," Vol. 27, No. 2, June 1980, pp. 187-197.
ZINGER, A., "Concentrated Firing in Many-Versus-Many Duels," Vol. 27, No. 4, December 1980, pp. 681.
ZLOBEC, S., L. Kerzner and A. Ben-tal, "Optimality Conditions for Convex Semi-Infinite Programming Problems,"
Vol. 27, No. 3, September 1980, pp. 413-435.
ZOLTNERS, A. A., R.M. Soland and G.T, Ross, "The Bounded Interval Generalized Assignment Model," Vol. 27, No.
4, December 1980, pp. 625.
ZUCKERMAN, D., "A Note on the Optimal Replacement Time of Damaged Devices," Vol. 27, No. 3, September
1980, pp. 521-524.
INFORMATION FOR CONTRIBUTORS
The NAVAL RESEARCH LOGISTICS QUARTERLY is devoted to the dissemination of
scientific information in logistics and will publish research and expository papers, including those
in certain areas of mathematics, statistics, and economics, relevant to the over-all effort to improve
the efficiency and effectiveness of logistics operations.
Manuscripts and other items for publication should be sent to The Managing Editor, NAVAL
RESEARCH LOGISTICS QUARTERLY, Office of Naval Research, Arlington, Va. 22217.
Each manuscript which is considered to be suitable material tor the QUARTERLY is sent to one
or more referees.
Manuscripts submitted for publication should be typewritten, double-spaced, and the author
should retain a copy. Refereeing may be expedited if an extra copy of the manuscript is submitted
with the original.
A short abstract (not over 400 words) should accompany each manuscript. This will appear
at the head of the published paper in. the QUARTERLY.
There is no authorization for compensation to authors for papers which have been accepted
for publication. Authors will receive 250 reprints of their published papers.
Readers are invited to submit to the Managing Editor items of general interest in the field
of logistics, for possible publication in the NEWS AND MEMORANDA or NOTES sections
of the QUARTERLY.
NAVAL RESEARCH
LOGISTICS
QUARTERLY
DECEMBER 1980
VOL. 27, NO. 4
NAVSO P-1278
CONTENTS
ARTICLES
Storage Problems when Demand is 'All or Nothing'
Reliability Growth of Repairable Systems
On the Distribution of the Optimal Value for a
Class of Stochastic Geometric Programs
A Class of Continuous Nonlinear Programming
Problems with Time-Delayed Constraints
Equalities in Transportation Problems and
Characterizations of Optimal Solutions
A Network Flow Approach for Capacity Expansion
Problems with Two Facility Types
Solving Multifacility Location Problems
Involving Euclidean Distances
An Easy Solution for a Special Class
of Fixed Charge Problems
The Bounded Interval Generalized
Assignment Model
The M/G/l Queue with Instantaneous
Bernoulli Feedback
An Inventory Model with Search for Best Ordering
The United States Coast Guard Computer-Assisted
Search Planning System (CASP)
Concentrated Firing in Many-Versus-Many Duels
Spike Swapping in Basis Reinversion
An Alternative Proof of the IFRA Property
of Some Shock Models
News and Memoranda
Corrigendum
Index to Volume 27
Page
D. P. GAVER 529
P. A. JACOBS
S. A SMITH 539
S. S. OREN
P. M. ELLNER 549
R. M. STARK
T. W. REILAND 573
M. A. HANSON
K. O. KORTANEK 589
M. YAMASAKI
H. LUSS 597
P. CALAMAI 609
C. CHARALAMBOUS
P. G. MCKEOWN 621
P. SINHA
G. T. ROSS 625
R. M. SOLAND
A. A. ZOLTNERS
R. L. DISNEY 635
D. C. MCNICKLE
B. SIMON
Price K. GOLABI 645
H. R. RICHARDSON 659
J. H. DISCENZA
A. ZINGER 681
R. V. HELGASON 697
J. L. KENNINGTON
C. DERMAN 703
D. R. SMITH
OFFICE OF NAVAL RESEARCH
Arlington, Va. 22217
Live Music Archive
Librivox Free Audio
Metropolitan Museum
Cleveland Museum of Art
Internet Arcade
Console Living Room
Books to Borrow
Open Library
TV News
Understanding 9/11