You are on page 1of 26

Model Building Techniques for the Correction of End Effects in Multistage Convex Programs

Author(s): Richard C. Grinold


Source: Operations Research, Vol. 31, No. 3 (May - Jun., 1983), pp. 407-431
Published by: INFORMS
Stable URL: http://www.jstor.org/stable/170615 .
Accessed: 18/01/2011 10:53

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

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

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

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

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

http://www.jstor.org
Model Building Techniques for the Correction of
End Effects in Multistage Convex Programs
RICHARDC. GRINOLD
University of California, Berkeley, California
(Received July 1979; accepted October 1982)

This paper considers model building procedures that can be used to reduce
end effects in multistage planning models. In particular, it emphasizes some
problems that are peculiar to energy planning applications. The paper pre-
sents and compares four methods for mitigating end effects. It concludes that
one of these four methods, the dual equilibrium procedure, is best. The major
drawback of the dual equilibrium procedure is its lack of simplicity. If it is to
be widely used, then it must be widely understood.

M ODEL BUILDERS dealingwith multistageproblemsmust decide,


either explicitly or implicitly, on the appropriate length of a
planning horizon. This paper presents, analyzes, and evaluates alternative
model building techniques that can be helpful in dealing with the planning
horizon decision. The selection of a planning horizon must balance the
increased computational burden of adding more stages against the loss in
model realism that can occur by decreasing the number of planning
stages. The errors that arise from having too short a planning horizon are
called end effects. The trade-off between end effects and computational
costs is the essence of the planning horizon decision.
One way to examine the planning horizon problem is to experiment
with planning horizons of different length and thereby obtain some
insight concerning the rate at which increasing the planning horizon
increases computational costs and mitigates end effects. One can imagine
a pair of curves depicting computational cost and end effects (below we
comment on the measurement of end effects) as a function of the length
of the planning horizon. Then one chooses the planning horizon that
strikes the proper balance between end effects and computational cost.
This is a reasonable procedure, but it is not the approach considered
here. The paper discusses four model building techniques that can be
used to alter the model. For each possible technique there will be a trade-
off between end effects and computational costs. The objective of this
paper is to present, analyze and evaluate these four techniques. This is
Subjectclassification:473 infinitetime horizonmodels,658 end effects in infinitehorizonproblems.
407
Operations Research 0030-364X/83/3103-0407$01.25
Vol. 31, No. 3, May-June 1983 ? 1983OperationsResearchSociety of America
408 Grinold

accomplished by putting the models on an equal footing. One starts with


an infinite horizon planning problem. This problem has a convex objective
and linear multistage constraints. This infinite problem is divided into
two phases: a transient phase up to the planning horizon and a secondary
phase after the planning horizon. This separation effectively decomposes
the problem in the time dimension. Each of the techniques can be viewed
as a method for dealing with this decomposition, in time, by simplifying
the secondary phase problem and its linkage with the transient problem.
In each case this leads to a convex multistage program with a finite
number of stages. Once the particular technique is selected, the length of
the planning horizon could be determined in the manner discussed above.
Notice that there is no definition of end effects in this paper. It is
possible to give many reasonable a' priori definitions of end effects. For
example, if V is the optimal value of the infinite horizon problem and V
is the objective value in the infinite horizon problem using the decisions
obtained from the approximating problem, then j V - VI is a measure of
the loss in objective value that would arise in using the approximating
problem. One objection to this measure is that the value of V is unknown
and it might be difficult to go from a solution of the approximating
problem to the objective value for the infinite problem. There is another
objection to any fixed measure of end effects. Many planning models
have a hidden agenda; the objective value of the problem may actually
be of secondary interest to the value of certain key primal and dual
variables. For example, in some energy planning models used to measure
benefits derived from a research and development program, the crucial
variables may be the extent to which certain proposed technologies are
used in the future. In another example, the main issue may be the
justification of a mineral exploration program. In that case, the items of
interest are the dual variables for the mineral resource constraints. These
variables will provide the value of adding additional minerals. Thus, there
are a wide variety of possible measures of end effect; the appropriate
gauge of end effects should depend on the intended use of the model.
An improvement in the way end effects are handled can have substan-
tial benefits. First, multistage planning models are widespread (see Ber-
genthal [1974], Connolly et al. [1977], Manne [1976], Marcuse et al. [1975]
and Nordhaus [1973] for a sample of recent planning models). For any
particular planning model the choice of a technique for mitigating end
effects and the subsequent choice of a planning horizon can greatly reduce
computational costs. The savings one realizes by reducing the time
dimension of a model can either be garnered or used to expand other
dimensions of the model.
Suppose one has a T period linear programming model with M con-
straints per period. The difficulty in solving this problem is, say, propor-
tional to (M x T)25. If we mitigate errors caused by the transition from
Correction of End Effects in Multistage Programs 409

the initial phase to the secondary phase, then we might decrease T to T'
= (3/4)T. This would reduce the computation burden by 1 - (3/4)2/5 = 51%,
or it would allow us to enrich other aspects of the model and increase M
to M' = 4/3Mand still retain the same degree of difficulty. Thus the
smaller one can make T without causing undue distortions, the more one
can save on computations or enrich other aspects of the model.
The paper is organized as follows: Section 1 describes the multistage
planning model as a multistage linear program in which the initial
transient phase is followed by a stationary phase that consists of an
indefinite sequence of equal periods. The idea is to simplify that station-
ary phase problem in order to make the entire problem easier to solve.
Section 2 presents four ways of simplifying the problem. The four
methods are truncation (ignore the stationary phase), salvage (place an
ad hoc value on resources carried over from the transient to the stationary
phase), and primal and dual equilibrium (impose an equilibrium condition
on either the primal or dual variables in the stationary phase). Section 2
describes these methods and makes some comments of how they relate
to each other and Section 3 gives an example to illustrate the use of the
procedures. Sections 4 and 5 show how to extend the framework described
in Section 1 to fit more general models and, in particular, to accommodate
some of the special features found in energy planning models. Section 4
considers the case of nonrenewable resources; constraints on resource
availability are introduced into the model and we show how the four
methods (truncation, salvage, and primal and dual equilibrium) fare with
a model that contains constraints on exhaustible resources. Section 5
considers the problem of phasing-in new technologies during the station-
ary-phase. The phasing-in would normally lead to nonstationary coeffi-
cients in the constraint matrix during a period in which we have assumed
they are stationary. In Section 5, we present an alternative way of
modeling the phase-in restrictions by moving the nonstationarity to the
right hand side. Section 6 evaluates the four techniques qualitatively,
comparing them along dimensions of generality, simplicity, flexibility,
and computational difficulty. This section is qualitative and, in large part,
reflects the author's informed opinion. This evaluation in Section 6
concludes that dual equilibrium procedure is the best way to reduce the
length of the planning horizon and to provide a smooth transition from
the initial to the secondary phase. However, the dual equilibrium proce-
dure is more difficult to understand and should be mastered before one
embarks on its use. For this reason, we recommend the use of the dual
equilibrium procedure for initiating projects. For ongoing models, the re-
engineering of the model to fit with the dual equilibrium technique may
cause more difficulty than the eventual benefit; in this case an educated
use of the salvage technique seems advisable.
The paper is closely related to four other papers. Grinold [1977, 1983]
410 Grinold

established theoretical underpinnings of these ideas. These papers show


that the dual equilibrium technique works in the limit, i.e. as the length
of the planning horizon increases, the optimal value and solution (of the
approximating problems) converge to the optimal value and solution of
the actual problem. In addition, Grinold [19771 shows the other tech-
niques: salvage, primal equilibrium, and truncation may not even lead to
bounded or feasible approximating problems even though the actual
problem has an optimal solution. Thus Grinold [1977, 1983] gives strong
theoretical support to the dual equilibrium procedure.
Note that there is a difference in emphasis in this paper. The previously
cited papers study the limiting properties of T period finite horizon
models as T becomes large. This paper uses that knowledge to address
the more practical question, how can one make T small and still have a
reasonable approximation of the longer range problem. For this reason,
this paper does not follow a rigorous theorem proof approach; instead, it
uses the theoretical results of Grinold's [1977, 1983] to motivate an
approach to the practical problem of dealing with end effects.
Grinold [1980] applies the ideas contained in this paper to an energy
model and compares the numerical results. These calculations indicate
that the dual equilibrium procedure is superior in reducing end effects on
key primal and dual variables that appear in the initial stage. Grinold et
al. [1976] reports on a longitudinal manpower planning model. In that
study, we used both the primal and dual equilibrium procedure. The dual
equilibrium procedure worked perfectly and the primal equilibrium pro-
cedure caused instability in primal decisions in the time periods imme-
diately prior to the end of the planning horizon.
To my knowledge this is the first paper to give a formal treatment of
end effects. Model builders have always, of course, used ad hoc procedures
for reducing end effects, and have effectively adapted special ad hoc
procedures developed for special models. The truncation and salvage
value techniques are so natural there is no formal description of them in
the literature; they are folk procedures. The truncation procedure, for
example, is the first thing one would consider: merely extend the planning
horizon to a decent interval, then ignore what comes after. The salvage
value procedure is a recognition of the fact that resources in place of the
planning horizon will have some value in the subsequent years.
Manne [1970] used the primal equilibrium technique in the study of a
dynamic economic planning model.
The dual equilibrium technique has its roots in the study of Markov
decision problems. De Ghellinck and Eppen [1967] treat the Markov
decision problem as a multistage linear program, then show how the dual
equilibrium technique can be used to obtain an exact solution in one
stage. Hopkins [1971] first applied the dual equilibrium idea to a general
linear program, and Grinold and Hopkins [1973] defined a class of
Correction of End Effects in Multistage Programs 411
problems in which the dual equilibrium technique leads to an exact
solution.

1. THE GENERAL PLANNING FRAMEWORK


The optimization problem under consideration is a multistage program
with linear constraints and a convex lower semicontinuous objective
function.
minimize
fo[xo] + af[x1] + ac2f[x2] + a3f[x3] + * + af[xt] ] +
?

subject to the constraints


Aoxo = bo
Hixo + Ax, = bi
H2xo + K1xi + Ax2 = b2

H3xo+ K2xi + KIx2+ Ax3 = b3

(1.1)

Htxo + Kt-lxl + Kt-2x2 + Kt3x3 + * + Axt=bt

xo, xt-- t'lE! .


This prototype planning model consists of two phases: a transient phase
and a subsequent stationary phase. Let (0, 1, 2, 3, ***) be an increasing
sequence of points in time. The shorthand terminology of time t is used
to mean time gt and the interval of time (tt, gt+l] after gt and up to tt+i is
called period t. The transient phase of the planning process ends at time
D1;thus, period 0 captures the entire transient phase. The transient phase
may itself consist of several subperiods, perhaps of unequal length. The
time Dlis called the length of the planning horizon. The stationary phase
of the planning problem consists of all remaining periods. For t-' 2, gt -
gt-i is constant, thus the periods are of equal length in the stationary
phase. This way of stating the problem is convenient although it does
present a hazard. Increasing or decreasing gl will change the problem's
structure during the transient stage and will cause the stationary phase
objective f to be multiplied by a scale factor in order to have attf[xt]
412 Grinold
represent present value at time 0. These dependencies on Di are not
obvious in our notation.
In period zero, the transient phase, a decision xo is selected that satisfies
the linear constraints Aoxo - bo, xo > 0. There is a cost, fo[xo], associated
with choosing xo. We assume the fo is a convex function. Choosing xo has
an impact on the set of possible future decisions. That impact is repre-
sented by the sequence of vectors Htxo for t - 1, where Htxo is the impact
of the transient decision xo on the set of possible decisions at time t.
The time 1 decision xi is selected subject to the constraints:
Ax, b - Hixo
x1 : 0.

The choice of xi produces a cost with present value af[xi] where f is a


convex function. The choice of xi has an impact Ktx1on the set of possible
xt+i decisions. In general, the set of possible decisions at time t will be
affected by the previous decisions xo, xi, ***, xt-i; and the time t decision
xt for t ' 2 is feasible if it satisfies the constraints:
Axt = bt- Htxo - EtJ` Kjxt-j, xt ? 0. (1.2)
Selection of xt yields a cost with present value atfxt[], and has an impact
Kjxt on the set of time t + j decisions for j- 1. Notice that the constraint
matrix A and objective f do not change for times t > 1. Thus for all
t ' 1, each xt and bt are of the same dimension. In general, xo and bo will
be of much larger dimension than xt and bt for t ' 2. In order to simplify
notation, we treat problem (1.1) as a mathematical program with an
indefinite planning horizon. For a detailed theoretical treatment of such
problems, see Grinold [1977, 1983].

2. APPROXIMATIONPROCEDURES
This section describes four ways of approximating (1.1) with a smaller
problem.

Primal Equilibrium
The primal equilibrium procedure assumes a constant rate of growth
(or decline) in the primal variables; that is xt+l = ytxl. This procedure is
applicable in problems that have bt+1= ytbi, and in which there is an
index I such that Hj = 0 and Kj = 0 for j > 1.
The primal equilibrium condition is enforced by adding extra con-
straints; there is a block of constraints for each time period that the
decisions xo and xi carry over into the equilibrium period. For example,
if Kj = 0 and Hj = 0 for j > I = 3, then substituting xt+i = ytxi and
bt+i = ytbi for t = 1, 2, 3, 4 yields
Correction of End Effects in Multistage Programs 413
Aoxo bo
H1xo + Ax1 = bi
H2xo + (yA + Kj)xi = yb1 (2.1)
H3xo + (y2A + yKi + K2)xl = -y2b
(y3A + y2KI + yK2 + K3)xl = y3b1, xo > 0, xl > 0.
On this example the constraints for t ? 5 are
4
yt4(y3A + y2Ki + yK2 + K3)x = yt4(y3bi)
However, this is simply yt-4 times the final constraint block in (2.1), and
is thus redundant.
The objective function can be written as

fo[x1] + a E'J= a'f[yxil]. (2.2)


The primal equilibrium approximation in the general case is
minimize fo[xo] + aEJ=1 ajf[yjxi] (2.3)
subject to
Aoxo =bo
H1xo + Ax, =b

Htxo+ (yt-lA + Et` yt 'Kj)xi = yt1b for 2 c t ' I


(y'A + >IJ-=Jly'-'Kj)xi = y'bi, xo 2 00, X1 0.
The primal equilibrium constraints (2.3) are more strict than those of
(1.2); if (xo, xi) satisfies the constraints of (2.3), then x0 and xt = yt-1x1 for
t ? 1 will be feasible for (1.1). Therefore the optimal value of (2.4) is an
upper bound on the optimal value of (1.1). Problem 2.3 will be infeasible
if no primal equilibrium solution exists. It may happen that primal
equilibrium solutions exist, but problem (2.3) is infeasible. This will occur
if it is not possible to make the transition to equilibrium in I steps.
The objective in (2.3) may be difficult to evaluate. However in several
special cases it can be evaluated exactly.
(i) If f [x] = cx - 12x'Qx and ay2 < 1, then E'=1 a/f[Tyx'] = cxi/(1
ay) - 1/2(xl'Qxl)/(1 _- 2).
(ii) If A > 0 implies f[[x] = A'f[x] for some 8 > 0, then E>=i aff[yjxi]
= f[xil]/(l - a).

If it is impossible to evaluate the objective of (2.3), then one can use


the convexity of f to produce an easily evaluated lower bound on the
objective:
414 Grinold

J=.. ajf[yjxl] = (>J=1 - a)ajf[yjxi])/(l - a)

([la)xll(l - a)]/(I - a).


The primal equilibriummethod has the advantage of not requiringa
constant value of f. As long as the sum EJ=i a-'f[y'xi] can be calculated,
the method could be used.
The primal equilibriummethod has several drawbacks.Problem 2.3
contains 1 + I blocks of constraintsin the stationary phase. The dual of
(2.3) might be solved but it will, in general, be largerthan a primal with
2 or 3 blocks. The set of solutions of (1.1) with xt+i = -ytxl might be either
empty, in which case (2.3) is infeasible, or of little interest. Also, as is
pointed out in Section 3, problemswith exhaustibleresourcesmean there
can be no resource extraction in the stationary phase. In addition, the
models reported in Grinold et al. indicate that this procedure does not
lead to realistic solutions. In particular,in those cases the calculations
producedradicalchanges in the decisionvariablesright before the end of
the planning horizon. This is an end effect correction in the decision
variables that the procedureis designed to avoid.

Truncation
The second procedureis called truncation. As the name implies, one
ignoresthe connectionbetween decisionsat times 0 and time 1 with those
in the later time periods. This leads to the problem
minimize fo[xo] + af[xi]
subject to (2.4)
Aoxo= b
Hixo + Ax, = b1, xo _ O, xi > O.
Notice that (2.4) is independent of the matrices (K1,K2, ... ) and (H2,
H3, * ). These matrices link times t = 2, 3, * with the decisions xo
and x1.
Truncating the problem disconnects the first two periods from the
remainderof the problem.This form of disconnectionassumes that any
fixed assets created up to time 1 will have no value after time 1. This, of
course, will induce a bias against building assets that are expensive and
long lived in periods 0 and 1. In addition, any liabilities incurredbefore
time 1 will not be paid back.If, for example,the model allowsfor financing
by intertemporalborrowing,then, in a naive application,there will be no
obligation to repay the loan. In extreme cases, the ability to incur
unlimited liabilities can cause the truncated problem to be unbounded.
In any case, the cost incurredin periods 0 and 1 is a lower bound on the
cost incurred by an optimal solution. This follows since whenever
Correction of End Effects in Multistage Programs 415
(xo, xi, *..) is any solution feasible for the infinite problem, (xo, x1) is
feasible for (2.4). However, we cannot, in general, obtain any bound on
the optimal value of (1.1) from the truncated problem (2.4).

Salvage
The salvage technique extends the truncation technique by placing a
value on the resources carried over from time 1 into later periods. The
time 0 decision, XO,creates a stream of goods (H2xo,H3xo, *.. ) available
at times 2, 3, ... respectively. It is possible to choose a vector doand say
that doxois the time 0 value of all the future goods (assets and liabilities)
created by decisions in time 0. For example, if xo creates physical capital,
then doxo could be the undepreciated value of that capital; i.e. we build
a new plant in period 0 and then sell the plant for its undepreciated value
at the end of the planning horizon. Similarly, the decision, x1, at time 1
creates a stream of goods (Klxl, K2x1, *.. ) available at times (2, 3, . . . ).
If we let ad1x1 be the time 0 value of these resources, then the salvage
value problem becomes:
minimize {tf[xo] - doxo} + a{tf[xi] - dixi}
subject to (2.5)
Aoxo= bo
H1xo + Ax, = b1, xo 2 O, xl > O.
We will discuss the salvage technique in more detail later in this
section.

Dual Equilibrium
The fourth technique, dual equilibrium, can be motivated from either
a primal or dual perspective. We will give the primal motivation here and
use the dual motivation later in the section to obtain a more unified view
of the truncation, salvage, and dual equilibrium approaches.
The primal approach aggregates the constraints and variables from
time t = 1 onward by placing a weight at-l(1 - a) on the time t constraint
and summing. Hence in place of constraints for t = 1, 2, * , one obtains

H(a)xo + [A(a)]x(a) = b(a)


where
(i) b(a) = (1 - a) Et'=, at-lbt,
(ii) H(a) = (1 - a) Et'=, at-1Ht,

(iii) A(a) = A+ EtXli atKt,and (2.6)


(iv) x(a) = (1 - a) Et'= at1lXt.
416 Grinold
Item (iv) in (2.6) is the rule for aggregating variables. Notice that it must
be possible to calculate b(a), H(a), and A (a) in order to use the dual
equilibrium procedure. We can always calculate H(a) and A (a) if Ht = 0
and Kt = 0 for all t sufficiently large and we can calculate b(a) if, for
example, bt = Mb for all t sufficiently large and ad < 1. Because of the
convexity and lower semicontinuity of f we can say (if >' atxt < oc) that

Et=i atf[xt] = a(t=i(l - a)at-lf[xt])/(1 - a) _ af[x(a)]/(1 - a).


Therefore the dual equilibrium problem is
minimize fo[xo]+ af[x(a)]/(1 - a)
subject to Aoxo = bo (2.7)
H(a))xo + A(a)x(a) = b(a), xo ~_O, x(a) _ O.
Thus any feasible solution (xo, x1, ...) of (1.1), for which the sum (2.6
(iv)) is finite, yields a feasible solution of (2.7), with a lower value than its
value in (1.2). Grinold [1983] demonstrates that under very general
conditions, problem (2.7) is feasible, has a finite optimal value, and
provides a lower bound on the optimal value of (1.1).
When (2.7) is solved one obtains optimal values of (xo*, x *(a)). In the
linear case, if one can find a sequence (X1, X2, X3, * * . ) such that x = (xo*,
X1, X2, X3, *. . ) is feasible for (1.2) and x*(a) = t'=, at-lxt(l - a), then the
sequence is optimal for (1.1). In general, it is not possible to find such a
sequence (XI, x2, *--); however, it is possible in some special cases. We
should not be too concerned about our inability to reconstruct an optimal,
or even feasible, primal solution for t - 1. Our objective is to reduce end
effects. We are concerned about the accuracy of our estimate of the
optimal cost and of the losses involved in using xO*from (2.7) rather than
the actual optimal xo from (1.1).

Lagrange Multipliers
This section describes another way to motivate the truncation, salvage,
and dual equilibrium procedures. The basic idea is to use Lagrange
multipliers to relax the constraints beyond the planning horizon. Then
we find that different assumptions about the values or form of the
Lagrange multipliers lead us to the three procedures (truncation, salvage,
and dual equilibrium).
Let us view bt as an external demand for goods and at time t: bt - Htxo
- Et>-51Kt-jxi is the net demand for goods at time t; i.e. the external
demand bt minus any demand satisfied or created by the earlier decision
(xo, xi, xt--,X_). The amount of goods supplied at time t is Axt. The
constraints of (1.1) say that this must equal the demand. If we relax that
constraint then we can define the unsatisfied demand as
Correction of End Effects in Multistage Programs 417

[excess [net
demand] demand] [supply] (2.8)
gt = [bt - Htxo - Eti-1 Kt-,x,] - [Axt].
If an element of gt is positive, then demand for that good has not been
met. If an element of gt is negative, then there is excess supply.
Now suppose we know, or can guess, values for the goods at each point
in time. Let ut be the time t value of goods at time t. Then utgt would be
the value of the excess demand at time t. If we viewed ut as a vector of
prices, then we would sell items in excess supply and buy the items in
order to satisfy the excess demand. These transactions would cost utgt at
time t. The time 0 value of this cost is atutgt (not really, since period 0 is
not of the same length as subsequent periods; however, the argument can
be modified slightly to compensate for that difficulty).
If we relax the constraints from period t = 2 onward, and to compensate
we make the supplier pay utgt at time t in order to satisfy the demand,
then the cost of any production plan (xO, xI, X2, ***) will be
fo[xo] + af[XI] + Xt=2 atf[xt] (2.9)
+ Xt=2 atut[bt - Htxo - tI-= Kt_-xj - Axt].

It is generally possible to reshuffle the terms in (2.9) (if the ut are


bounded, the xt are nonnegative, and Et'=,atxt is finite). After reshuffling
we get
{f[xo] - [_=, atutHt]xo} + a{tf[xil] - [ 1 aui+jj]xi} (2.10)
+ >t=2 att{f[xt] - [utA + >J=1 ajut+jKj]xt} + 't=2a utbt.
Now let us consider the problem of minimizing (2.10) subject to the
constraints
Aoxo=bo (2.11)

H1xo +Ax, = bi, xo 0, xI 0, xt>0 for t' 2.


Under certain very general conditions (see the Appendix) the optimal
value of (2.11) subject to the constraints of (2.10) will give a lower bound
on the optimal value of (1.1). This is true regardless of the selection of ut.
The intuitive reason is clear; any optimal solution of (1.1) will satisfy the
constraints of (2.11) and, since it will have gt = 0 for all t, the optimal
solution will have its optimal value in (2.9) and (2.10).
The truncation problem is a special case of this. For truncation we
assume that ut = 0 for t ' 2; i.e. we assume that the goods have no value
after time 1. In that case, the objective becomes fo[xo] + af [xi] +
Et=2 atf[xt]. The truncation technique minimizes the first two terms of
this objective subject to (2.11). If one happens to know an x* ' 0 that is
418 Grinold
the minimizer of f[x] subject to x ' 0, then the truncation technique can
provide a lower bound on the optimal value of (1.2). If (xo*, xl*) is the
optimal solution of the truncated problem, then f[xo*] + af[xl*] + a 2/
(1 - a) f[x*] is a lower bound on the optimal value of (1.2), where x*
solves minimize f[x] subject to x ' 0.
The salvage technique can be viewed in a similar way. Compare the
objective in the salvage problem (2.5) with (2.10). If
(i) do = fJ2 ajujHj,
(ii) di = J==i a', ul+jKj, and (2.12)
(iii) dt = utA + >'=i ajut+jkj for t o 2,
then (2.10) becomes
{f[xo - doxo} + at{f[xil] - d1x} (2.13)
+ Xt=2 at{f[xt] - dtxt) + Dt=2 atutbt.
The salvage value technique minimizes the first two terms of (2.13), and
ignores the third term. There are two valuable insights to be gained from
this exercise. First, the most reasonable and consistent way to estimate
salvage value is to use the formulas in (2.12). Notice that this requires
placing a value ut on the goods available at time t, and it does not
distinguish between resources produced in period 0 (Htxo) and those
produced in period 1 (Kt-lxl). This is, of course, consistent with the way
the goods are treated in the decision problem. At time t the decision xt is
constrained by

Axt = bt- Htxo- J-1 Ktjxj, Xt _ 0.


It is the total constraint vector that matters: not the components.
The second insight is that the salvage technique can yield a lower
bound on the optimal value if the following are true:
(i) The dt are calculated as in (2.12).
(ii) The sum X'=2 atutbtcan be calculated.
(iii) The optimal solution xt* of
(2.14)
minimize f[xt] - dtxt.
subject to xt - 0
can be found for all t.
Under these conditions an optimal solution (xo*, xi*) of the salvage
problem (2.5) will provide a lower bound {f[xo*] - doxo*} + 2t=i at{tf[xt*]
-dtxt*} + '=2 utubon the optimal value of (1.2).
Correctionof End Effects in MultistagePrograms 419
The salvage technique therefore consists of guessing the ut, calculating
the dt, and solving for the xt*, then solving (2.5). This is a formidable job;
the natural inclination for a person faced with this task is to look for
simplifications. An obvious and, it turns out, a very useful simplication is
to assume that ut = U2 for all t ' 2. This is an assumption that resource
values (present values at time 0) will decrease at the rate of discount a;
i.e. the only value in having an additional amount of resources at time t
rather than time t + 1 is the time value which is captured by the discount
factor. When ut = U2 for all t ' 2, then (2.14) simplifies to
(i) do = U2[E' 2 ajHj]

d= U2[Er?=i ajKj]

d = U2[A + E a'K1] = U2A(a) (recall (2.6)),


(ii) Calculate U2 Et=2 atbt, and (2.15)
(iii) Find the optimal solution x* of
minimize f[x] - dx
subject to x ' 0.
In this special case the lower bound is
{fo[xo*] - doxo*} + a{f[xi*] - dixi*} (2.16)
+ a -2/( - a){f[x*] - dx*1 + U2 yt=2 atbt.
This expression is much easier to handle. The terms do, di, d and the
constant term are all linear in U2. In particular, when f[x] = cx is linear,
then (2.15 (iii)) indicates one should choose U2so that
c > u2A (a) - d. (2.17)
If this is true, then x* = 0 and the third term in (2.16) can be ignored. If
(2.17) is not true, then the optimal value in (2.15 (iii)) is -oo and the
lower bound in (2.16) is an uninteresting -oo.
In general, one could choose U2 and x* simultaneously by solving
minimize f[x]
subject to A(a)x = b (2.18)
x 0
where b = (1 - a)/a 2[>2 atbt].

Suppose x* is the optimal solution of (2.18) and U2*is the Kuhn-Tucker


multiplier for the constraints in (2.18). Now use U2* to calculate the
salvage values do and di as per (2.15). The salvage value problem becomes
420 Grinold
minimize [fo[xo] - doxo] + a[f[xi] - d1x1] + a2/(1 - a)f[x*]
subject to (2.19)
Aoxo =bo
H1xo+Ax1= b1, X0>-, XI>-.
This form of the salvage value problem will give a lower bound on the
infinite problem. It is sophisticated in that it asks us to solve an additional
problem (2.18) in order to calculate reasonable salvage values.
The dual equilibrium problem proceeds in much the same spirit.
However, instead of solving (2.18) and then (2.19), the dual equilibrium
technique solves the two problems jointly.
To interpret the dual equilibrium technique using multipliers, let us go
back to (2.8) and place the multiplier au, on the time 1 constraints. Then
(2.9) becomes

{fo[xo]- [E=, ajujHj]xo} + St-2 atff[xt] - [utA (2.20)


+ J., a'ut+jKj]xt) + ET=. atutbt.

If one makes the simplifying assumption that ut = ul for all t, then


(2.20) becomes
{fo[xo- au,H(a)xo/(l - a) (2.21)
+ Et=, at{tf[xt] - uA(a)xt} + auib(a)/(l - a)
where the notation H(a), A(a) and b(a) is explained in (2.6).
Let x(a) = (T=1 at-lxt)(1 - a); x(a) is a convex combination of the
vectors {xt}, thus the convexity of f[x] ensures that (2.22), below, is a
lower bound on (2.21).

fo[xo] + a/(1 - a){t[x(a)] - ui[H(a)xo + A(a)x(a) - b(a)]}. (2.22)


One can rewrite (2.22) as

L(xo, x(a), ui) = f[xo] + a/1 - af[x(a)] (2.23)


- u[H(a)xo + A(a)x(a) - b(a)].
As demonstrated above, the problem
minimize L(xo, x(a), ul)
subject to (2.24)
Aoxo=bo, xo'O, x(a) O
will yield a lower bound on the optimal value of (1.1). Let L(ul) be the
optimal value of (2.24), then a best lower bound is obtained by maximizing
L(ul). In sum, the best lower bound is calculated by solving
Correction of End Effects in Multistage Programs 421
maximize ui minimum {L(xo, x(a), u1)}
subject to (2.25)
Aoxo = bo, xo _:O, x(a) : O.
However, standard nonlinear programming theory states that problem
2.25 is exactly the same as the dual equilibrium problem 2.7. To sum-
marize, multipliers atut are associated with the constraints of (1.1) for
times t ' 1 with ut = ui for all t. Then this objective is minimized subject
to the time 0 constraint Aoxo = bo, xo > 0, which will give a lower bound
on the optimal value of the original problem 1.1. Then u1 is selected in an
optimal way to obtain the greatest lower bound. However, this problem
of choosing the greatest lower bound is exactly the same as the dual
equilibrium problem. To see this, start with the dual equilibrium problem
2.7 and associate a multiplier ul with the constraint H(a)xo + A(a)x(a)
= b(a). This yields (2.23) as an objective. Now the route to (2.25) is the
same as was followed above. For each ul, xo and x(a) are selected (as in
(2.24)) to give a lower bound on the optimal value of (2.7). Then maxi-
mizing over all possible values of ul will give the optimal value of the
dual equilibrium problem 2.7.
This section on multipliers has demonstrated that the truncation,
salvage, and dual equilibrium techniques can all be interpreted as at-
tempts to calculate lower bounds on the optimal value of the original
problem by associating multipliers with the constraints from time t = 2
onward in the truncation and salvage cases and from time t = 1 in the
dual equilibrium case. Truncation assumes these multipliers are zero,
salvage calls for an educated guess at the multipliers, and dual equilibrium
calculates the best value of the multipliers subject to the condition that
the period t + 1 multipliers are given by atul.

3. AN EXAMPLE
This section presents a small example that illustrates the four proce-
dures described above. The example is simple, yet it captures the essen-
tials.
This model considers an economy in which there are two types of
resources: exhaustible and inexhaustible. Each period an amount yt of an
intermediate good y is produced from the exhaustible resources and an
amount Zt of an intermediate good z is produced from the inexhaustible
resources. Total output is then a concave function g[yt, Zt] of the inter-
mediate goods yt and Zt.
Production of yt is limited in two ways. First by the exhaustible resource
limitation and second by a capacity restriction. If rt-i is the resource
inventory at the end of period t - 1, then the equation
-rt-, + yt + rt = O (3.1)
422 Grinold
keeps track of the resource inventory. In addition, capital is needed to
extract the resource. Let ht be the capacity for resource extraction in
period t. Let vt be unused capacity. Then
Yt + Vt = ht. (3.2)
Of course, the building and maintenance of productive capacity requires
some inputs. Thus let jt-l be the amount of period t - 1 output invested
in capital for extracting the nonrenewable resource; jt-l units of output in
period t - 1 will produce ajati, units of capacity in period t. Moreover, an
amount qht-, of period t - 1 capacity will survive to period t: i.e., this
capacity deteriorates at rate 1 - q. Thus the capacity ht is given by
ht= ajt-i + qht-i. (3.3)
There is a similar restriction for producing zt with it-, being the invest-
ment, kt the capacity, eit-i the capacity produced in period t by it-,, and
pkt-, the carry over in capacity from period t - 1 to period t. Thus if wt
is unused capacity,
Zt + Wt = kt
(3.4)
kt = eit-, + pkt-1.
Total consumption, ct, in period t is output minus investment. Thus
Ct = g[yt, Zt] - it -it, (3.5)
the utility of this consumption is U [ct], where U is an increasing concave
function and future utility is discounted at rate (1 - a). Thus the objective
is to maximize
t=1 atU [ct ] = f.-l a U[g(yt,
U zt ) it j t].
To put this model in the framework of Section 1, let
(i) x = (y, z,j, i, h, k, v, w, r),
y z j i h k v w r
(ii) [1 0 0 0 0 0 0 0 11
1 0 0 0 -1 0 1 0 0
A= 0 0 0 1 0 0 0 0
0 1 0 0 0 -1 0 1 0
? O O O 0 1 0 0 0
(3.6)
(iii) -O O O O O O O O -1-

O) K= K = -a O -q O O0 0

'O O O -e O -p O O O
(iv) Kt = O for t ' 2, and
(v) bt = 0 for t_ 1.
Correction of End Effects in Multistage Programs 423
In (3.6) the five constraints are simply those in (3.1) to (3.4).
For the purpose of illustration, suppose the transient period (0, D] is
divided into five subperiods of length equal to the periods in the equilib-
rium phase. Thus we are going to wait five periods before we terminate
the problem. These five periods are aggregated into period 0. Let XTbe
the decision taken in subperiod T of the transient phase (a 9-element
vector). Then the 45-element vector
XO = (xo, Xl, X2, X3, X4)

is the decision variable in the first phase. Since the objective in (1.1) is to
minimize, the objective for the transient phase in this example becomes
fo[xo] -- =o aTU[c,] where cT is the consumption in subperiod T; c, is
computed as in (3.5).
The matrices Ao and H1 are given by

K A
Ao= K A ]
K A
K A_
H1=[O 0 0 0 K],
and Ht = 0 for t ' 2.
The vector bo has 25 elements. The first element of bois the amount of
the exhaustible resource available at time 0, the third is the amount of
capacity for process y available at time 0 and the fifth is the amount of
process z capacity available at time 0. The remaining 23 elements of bo
are equal to zero.
The objective term in later periods is given by f[xt] =--a4U [ct]. Notice
how the length of the transient phase changes Ao, fo, f, and bo.
The primal equilibrium technique, with y = 1, would have the con-
straints.
Aoxo= bo
Hlxo + Ax1 = 0
(A+K)xi=0, xo'O, x1iO.
The objective is minimize fo[xo] + af[xll/(l - a). The matrix A + K
becomes
y z j i h k v w r
1 0 0 0 0 0 0 0 0
1 0 0 0 -1 0 1 0 0
[0 0 -a 0 1-q 0 0 0 Ol
0 1 0 0 .0 -1 0 1 0
0 0 0 -e O 1-p O O O .
424 Grinold
Notice that as long as -y' 1 the exhaustible resource will not be extracted
in the stationary phase.
Application of the truncation technique is straightforward.
For salvage value, one selects a vector U2 = (, 0, A, ?, A) to associate
with the carryover Kxl from period 1 into the later period. The salvage
values are
do=O and (37)
d= (0, 0, -aa42, -aeX, -aq4, -apA, 0, 0, -a').
The variables -A and -X are the marginal value of additional capacity
for y and z at time 2, and -X is the marginal value of the exhaustible
resource at time 2. Thus to apply the salvage technique one guesses the
value of additional resources and additional capacity at time 2, calculates
d1 (3.7), and then solves
minimize fo[xo] + a[f[xi] - dixi]
subject to
Aoxo= bo
H1xo + Ax, = O, xo -0, xi-' 0.
For the dual equilibrium technique one has the constraints
Aoxo = bo
(1 -a)Hixo + (A + aK)x(a) =O, xo-' 0 x(a) 0O
with the objective
fo[xo] + af[x(a)]/(1 - a)
where
af[x(a)]/(l - a) = -a5Utg[y(a), z(a)] - i(a) -j(a)}/(l - a)
and
y z j i h k v w r
10 00 0 0 0 0 1-al
1 0 0 0 -1 0 1 0 0

0 1 0-a 0 1 -1 0 1 0
O-0 O O -ae 0 1-ap 0 0 0 .

Since a < 1, this treatment will always allow for use of the exhaustible
resource; i.e., it is possible to have y(a) > 0.
This model could be simplified a great deal. In fact, by eliminating ht,
and kt and rt one can reduce the number of equations from 30 to 13 and
Correction of End Effects in Multistage Programs 425
the number of variables from 54 to 37 for the truncation, salvage and
dual equilibrium techniques.
For the primal equilibrium, since y must be zero in the equilibrium
phase, one can reduce to 14 equations and 33 variables.

4. EXHAUSTIBLE RESOURCES
Many dynamic planning models have inputs to the production process
that are in limited supply. This of course applies to mineral resources
such as oil, coal, natural gas, and uranium. Such as resource limitation
appears to conflict with the concept of stationarity, but as is demonstrated
below, there is no conflict. The model can include exhaustible resources.
To see this, let yt be the vector of resources consumed in period t, st a
stock of vectors to be carried over from period t to period t + 1, and let
us allow the resources to be discarded. This rather artificial device of
discarding the resource is included to allow us to avoid a technical bind
caused by the equilibrium condition. The initial supply, s-1, is known,
and the accounting relationship is st + yt ' st_- for all t.
In a primal equilibrium problem (with Ht = 0 and Kt = 0 for t _ 2),
there are three sets of resource constraints
so + yo s-, for period 0

so + Si +y ' O for period 1 (4.1)


(y - 1)si + yyl ' 0 for primal equilibrium.
When y - 1, the equilibrium condition implies yi and si must be zero.
Thus the three equations reduce to
YO_ s-i. (4.2)
In case 0 < -y < 1, there will be no need to discard resources, thus the
three equations in (4.1) will hold with equality. Then, these three equa-
tions can be replaced by a single equation
yo + (1/1 - y)yi = S-i (4.3)
where yt+1 = ytyl. Additional flexibility can be obtained by using a less
than or equal to (_) inequality in (4.3).
In the truncation and salvage value techniques the resource constraint
is expressed as yo + yi ' s_i, i.e., total resources used in periods 0 and one
cannot exceed the total available.
The difference in the treatment of exhaustible resources by the salvage
and truncation techniques is in the objective function. The salvage value
technique would place some value a2v2 on the resources carried into
period 2. Thus add the expression
-a 2V2(S-1 - yO - yl) (4.4)
426 Grinold
to the objective. This procedure allows us to strike some balance between
resource use in the first two periods and resources carried over to the
future.
In the dual equilibrium technique, one multiplies the constraints - st-I
+ yt + St - 0 for t ' I by (1 - a)at-1 and sums. This yields
-(1 - &)S0+ y(a) + s(a) ' 0 (4.5)
where
(i) y(a) = (1 - a) Yt'= atlyt, and
(ii) s(a) = (1a-a) >t=l a St.

When (4.6) is combined with the time 0 constraint one obtains a single
equivalent constraint:
yo + y(a)/(l -a) S<s-. (4.6)
To summarize, the truncation and salvage value add the constraint
(4.4) and in addition (4.5) is added to the objective in the salvage
technique. For primal equilibrium,one adds (4e2)if y > 1, and (4.3) if y
< 1. Finally, the dual equilibriumtechnique adds (4.7).

5. INTRODUCTION RATES AND NEW TECHNOLOGIES


It is reasonable,indeed essential, to place some constraintson the rate
of introduction of new technologies. For example, Manne [1976] con-
strains the rate of introductionof the fast breeder,the light water reactor,
and an advanced generation technology (first available in 2020). These
constraintsare included to representour naturalreluctanceto jump to a
new and untried technology.
Consider one possible way of including the constraints on the rate of
introductionof the untriedtechnology.Supposeyitrepresentsthe amount
of new capacity in period t for plant type i where i = 1, 2, ** , n. The
total new capacity is Zi=, yit. The rate of introduction constraint states
that at most a fractionkjtof the new capacity can be of type j; i.e.
yit -5 ky Y'i1=1
yit- -)
or (1 - kjt)yjt- ioj kjtyit- 0.
This constraint is not compatible with our structure since it contains
nonstationarycoefficients in the constraintmatrix for t ! 1.
One can step around this problem by using a surrogate for the total
new plant to be built in period t. For example,suppose demand in period
t can be expressed as ytd, that in equilibriumnew additions to plant are
yty, and that plants have a useful life of m + 1 periods. In that case, the
total capacity is the sum of capacity added in the past m + 1 years:
Correction of End Effects in Multistage Programs 427

E=, yt'-jy.In equilibrium, this must equal demand, i.e. ytd. From this one
can solve for y/d, the fraction of total capacity that is new: y/d =
ym(y - 1)/(ym+l - 1). When y = (1.025)5 = 1.1314, and m = 5, then y/d
= 0.2219; i.e., 22%of all plants will be of the most recent vintage. Now let
us use information to model introductory rates.
Let d, be the demand in period t, then substitute 0.2219dt for the sum
Ei=1yit in Equation 5.1. This yields Yit ' kjt(0.2219)dt. This form of the
rate of expansion constraint is slightly more ad hoc than (5.1), however,
it appears (see Grinold [1980]) to serve the same purpose.
When a constraint of this sort is used in a problem approximated by
the dual equilibrium procedure, one obtains a constraint yj(a) '
at-lkjtdt](1 - a) where yj(a) = [E=-1 at`yjt](1
LZWt=i - a).

6. QUALITATIVECONCLUSIONS
This section contrasts and ranks the four methods for treating end
effects according to four criteria: simplicity, ease of computation, gener-
TABLE I
METHOD-QUALITY SCORECARD

Method-Quality Simplicity ComPuta-


tion Generality Flexibility

Truncation 3 3 1 3
Primal equilibrium 1 1 2 2
Salvage 2 3 2 1
Dual equilibrium 1 3 3 3

ality, and flexibility. Table I summarizes the results. The rankings are
relative; three reflect the best, one the worst, and two a median. Some of
the judgments expressed in Table I can be defended formally. Others,
however, are based on intuition and are more subjective. Table I should
not be considered the last word; it is merely an organization of the
author's judgments.

Simplicity
It is obvious that the truncation procedure is the simplest both in
interpretation and ease of use. The data need not be altered and the
truncation procedure can be interpreted as ignoring the periods after the
planning horizon. The other three procedures require some modification
to the problem structure; the primal equilibrium approach adds additional
constraints; the salvage value approach modifies the objective function;
and the dual equilibrium approach modifies the last set of constraints.
The salvage value technique is easier to interpret than either of the
equilibrium techniques; it takes time and experience to understand the
equilibrium methods.
428 Grinold
Computation
The primal equilibrium approach requires additional constraints and
therefore entails more difficult computations. In addition, computational
experience with manpower planning problems (Grinold et al.) revealed
some numerical difficulties with the primal equilibrium approach. In
those calculations the primal solutions did not move toward the true
optimal solution in a smooth way; there were abrupt changes in the
primal solutions toward the end of the planning horizon. Since an end
effect correction is intended to make a smooth transition between the
planning period and the period after the planning horizon, the primal
equilibrium technique appeared to be working in the wrong direction.

Generality
The truncation procedure is a special case of the salvage value tech-
nique in which one sets U2= 0, i.e., there is no value to resources available
in period 2.
In one way, the dual equilibrium approach is more general than the
salvage technique. In the dual equilibrium procedure, the variable U2is
calculated rather than specified. In other words, the dual equilibrium
procedure is more general than the salvage value technique since it
requires a specific relationship between ut and ut+i, i.e., ut+1- ut = 0. The
generality of calculating ul rather than specifying it in an ad hoc manner
is considered more important, so the dual equilibrium procedure is ranked
as more general.
A theoretical study (Grinold et al.) shows that in general, when the
desired planning horizon is infinite, a T period approximation using dual
equilibrium will give improving approximations of the true optimal solu-
tion as T increases, and perfect approximations in the limit. None of the
other techniques will do this in the general case. This is significant, since
the actual objective of the model may be (as was pointed out in the
introduction) to determine the values of some critical primal and dual
variables. In fact, in some cases (see de Ghellinck and Eppen, Grinold et
al., Grinold and Hopkins [1973] and Hopkins [1971]), the dual equilibrium
technique can be shown to give the optimal solution.

Flexibility
In judging the methods, we considered two aspects of flexibility: first,
the ability to adapt to changes in the problem data, i.e., in sensitivity
analysis, and second the ability to handle variations of the basic model.
In the salvage technique, part of an optimal solution must be specified,
i.e., u2. If the data of the problem are changed, then the guess at these
solutions should change also. In practice, this will either make sensitivity
Correction of End Effects in Multistage Programs 429

analysis more cumbersome by adding the extra burden of revising the


estimate of U2 or it could lead to inaccuracies if the initial estimates are
not revised. Notice that the truncation and both equilibrium approaches
adapt automatically to changes in the problem data.
If limitations on the supply of natural resources are added, then the
dual equilibrium, salvage, and truncation procedures can be easily
adapted to handle this extra complication. The primal equilibrium tech-
nique is not consistent with limitation in natural resources, i.e., limited
natural resources cannot be used at a fixed rate through the indefinite
future.

APPENDIX
This technical appendix provides the theoretical support for the La-
grange multiplier arguments made in Section 2. It is based on two Grinold
[1977, 1983] papers. Let us assume that the matrix A has m rows and n
columns. The norm notation refers to the 11norm and the three-
line equality means "is defined to be." Now define
(i) b (b1, b2, ** ), a sequence of vectors in IRm
(ii) Hxo (H1xo,H2xo, *.. ), a sequence of vectors in IRm
(iii) X x3
(X1, X2, *X3, a sequence of vectors in IRn

(iv) F [x] -lmSpo t=i atf [xt]


(v) X(b - Hxo) _{xjE,-'1 Kt_jxj+ Axt = bt- Htxo, xt 2 O} (A.1)
(vi) V(b - Hxo) Inf{F[x]IxEX(b - Hxo))
(vii) Xo(bo) {xolAoxo = bo,xo - 0)
(viii) fO+:JRn _> R is the recession function of f (see Grinold
[1983] for details)
(ix) A(A = A + Zj=IXjKj.
The following assumptions are made:
(i) Xo(bo) is bounded and nonempty.
(ii) There exists a p > a such that for any xo E Xo(bo)
*t2=i Ptll bt - Htxo11< oo.

(iii) For every xo E Xo(bo). There is an x E X(b - Hxo)


a tIIxtII< ?.
such that Et'?=,

(iv) For j > I the matrices Kj = 0.


(v) fo and f are convex and lower semicontinuous. (A.2)
430 Grinold
(vi) For every A, 0 c A c a the optimal value of the program
minimize fol[y]
subject to
A(A)y = 0=, Yi= 1, y-0
has a positive optimal value.
Assumptions i-iv above ensure enough regularity to connect the tran-
sient problem with the equilibrium problem. Assumptions v and vi are
the same as those made by Grinold [1983]. They are discussed in detail
there. They are minimal assumptions in the following (rough) sense; if
they do not hold, then the optimal value function V(b) is improper, i.e.,
equal to +oo or -oo almost everywhere.
Under these assumptions one can restrict attention to solutions (xo, x)
with Et'= al xt 11< o. The result is:
THEOREM 1. Under the assumptions A.2, any solution (xo, x) with
Et=, atIIxtII= +oo, has fo[xo]+ aF[x] = +X0.
Now assume that 11Ut B for all t. Then the following proposition is
valid.
PROPOSITION 1. If
(i) 11ut 11C B, and
(ii) Zto=i at 11xt11< Xo
then

t=2 atut[Axt + Htxo + Et`5LKt-jxj - bt] = >Ji=2 atujHj)xo


+ {t=i aUlj+KjK}xl + Zt=2 at{utA + Zy I ajutjKj)xt -Zt2 atutbt
and the sums are finite.
Note that Proposition 1 gives conditions under which the functions in
Equations 2.8 and 2.9 are equal. The final assumption is that the optimal
value of (1.1) is finite. In that case, Theorem 2 of Grinold's [1983]
indicates that there is an optimal solution of (1.1). Let (xo*, x *) be that
optimal solution of (1.1) with value V* = fo[xo*] + aF[x*]. Then the
feasibility of (xo*, x *) implies that the value of (xo*, x *) in (2.8) is V* and
by virtue of Proposition 1 the value in (2.9) is V*.
Thus with xo = xo*, xI = xI* one can always do at least as well as V*
in minimizing (2.9) subject to (2.10). The optimal value will obviously be
a lower bound on V*.
Moreover x *(a) = (1 - a) Yit=l atXt* and xo* are feasible for the dual
equilibrium problem 2.7. Thus the optimal value of (2.7) will indeed be a
lower bound on the optimal value of (1.1).
Correction of End Effects in Multistage Programs 431
ACKNOWLEDGMENTS
I would like to thank Rich Richels and Jerry Karaganis for their
interest and support. The referee and the editors provided many useful
suggestions; I hope I have done them justice.

REFERENCES
BERGENTHAL, G. 1974.A LinearProgrammingStudy of the Marketfor Electric-
ity, WorkingPaper 74-7,EuropeanInstitute of AppliedManagement,Brussels.
CONNOLLY, T. J., G. B. DANTZIG AND S. C. PARIKH. 1977. The Stanford PILOT
Energy-EconomicModel. In Advances in the Economics of Energy and Re-
sources, R. S. Pindyck (ed.). JAI Press, Greenwich,Conn.
DEGHELLINCK, G. R., AND G. EPPEN. 1967. Linear Programming Solutions for
SeparableMarkovianDecision Programs.Mgmt. Sci. 16, 371-393.
GRINOLD, R. 1977. Finite Horizon Approximationsof Infmite Horizon Linear
Programs.Math. Program. 72, 1-17.
GRINOLD, R. 1980.Time Horizonsin Energy PlanningModels. In Energy Policy
Modeling, S. Schwartzand W. T. Ziemba (eds.). Nijhoff, Boston.
GRINOLD, R. 1983.ConvexInfiniteHorizonPrograms.Math. Program.25, 64-82.
GRINOLD, R., AND D. S. P. HOPKINS. 1973. Computing Optimal Solutions for
Infinite-HorizonMathematical Programswith a Transient Stat e. Opns. Res.
21, 179-187.
GRINOLD, R., R. M. OLIVER AND K. T. MARSHALL. 1976. Longitudinal Manpower
Planning Models. Naval Res. Logist. Quart. 23, 245-260.
HOPKINS, D. S. P. 1971.InfiniteHorizonOptimallyin an EquipmentReplacement
and CapacityExpansionModel.Mgmt. Sci. 18, 145-156.
MANNE, A. S. 1970. Sufficient Conditionsfor Optimalityin an Infinite Horizon
Development Plan. Econometrica 38, 18-38.
MANNE,A. S. 1976. ETA: A Model for Energy Technology Assessment. Bell J.
Econ. 7, 379-406.
MARCUSE, W., L. BODIN, E. CHERNIAVSKYAND Y. SANDBORN. 1975. A Dynamic-
Time Dependent Model for the Analysis of Alternative Energy Policy, Brook-
haven National Laboratory, BNL-19406 (July).
NORDHAUS, W. 1973. Allocation of Energy Resources, Brookings Papers on
Economic Activity, No. 3.

You might also like