Sie sind auf Seite 1von 20

The Review of Economic Studies Ltd.

Continuous and Discrete Time Approaches to a Maximization Problem


Author(s): L. G. Telser and R. L. Graves
Source: The Review of Economic Studies, Vol. 35, No. 3 (Jul., 1968), pp. 307-325
Published by: The Review of Economic Studies Ltd.
Stable URL: http://www.jstor.org/stable/2296664
Accessed: 06/11/2010 05:12

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=resl.

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.

The Review of Economic Studies Ltd. is collaborating with JSTOR to digitize, preserve and extend access to
The Review of Economic Studies.

http://www.jstor.org
Continuous and Discrete Time
Approaches to a Maximization
Problemi
I. INTRODUCTION
A common research strategy in economic dynamics is to state the theory of a given
problem in continuous time and to test it in discrete time. The correspondence between
the theory and the real world relies on an implicit belief that the discrete time model
approximates the continuous time model. This suggests that if the original theory were
posed in discrete instead of continuous time then it would imply observable relations in
discrete time akin to the discrete time approximations of the continuous time model.
In general this is not so! Discrete and continuous time models may differ in many subtle
and surprising ways especially if the economic agents to which the theories apply are
assumed to optimize some function over which they have partial control.
We deal with a class of economic problems in which a firm seeks a maximum present
value of its return over an infinite horizon at a given interest rate. This problem has two
versions. In the one of our primary concern a monopolist seeks a present-value-maximizing
sales (or price) policy and faces a given demand function such that the quantity demanded
at time t depends linearly on past and expected future prices. In the second version a
firm determines its stock of capital over time so as to maximize the present value of its
net return. While both problems deal with investment, the first is less often so regarded
and so we call it the monopoly problem herein.
We show that not only are there surprising differences between the models in con-
tinuous and discrete time but also there are surprising differences between the monopoly
and investment problems. A solution exists in the monopoly problem only if there is an
upper limit on the interest rate while in the investment problem there is the usual result,
namely, a solution exists for all interest rates above some lower limit that may be zero.
1 There are few examples of applications of the calculus of variationsto economic problems. An
early one is Hotelling [8]. An extensive discussion of several monopoly problems in an insufficiently
appreciatedbook by G. C. Evans [6] treats problemsrelatedto the ones we discuss. Evans works with a
finite horizon and deals with both fixed and variable end point problems. His book also gives many
referencesto economic applications of the calculus of variations. We are grateful to Professor L. M.
Graves for calling to our attention a little known work by Henry Howes Pixley [10]. Pixley's analysis is
inspired by the work of Evans and is a very sophisticatedtreatment. He allows solutions q(t) (in our
notation) which have a finite numberof discontinuitiesin both q and q'. He assumes a finite horizon and
considers fixed and variable end-point problems. By allowing these discontinuitiesPixley shows that
there are no solutions obtainableusing the classicalcalculus of variations. Part of the explanationstems
from the use of a finite time interval which imposes strong demands on the sufficientconditions for a
maximum(see footnote 3) but the primaryexplanationis that Pixley permits discontinuities. He himself
suggests that other mathematicalapproachesare likely to yield meaningfulresults.
Recent interestin problemsrelatedto the ones treatedin the classicalcalculusof variationsis associated
with the attempt to derive rigorously the relation between investmentand the interest rate. Among the
first articleson this topic are Arrow [2] and Eisner and Strotz [5]. Arrow in particularnoted the pheno-
menon of an instantaneousjump in the stock of capital if the cost of adjustmentis linear in the rate of
investment. In a study of the optimal advertisingpolicy for a firm which treats advertisingas a stock of
good will, Nerlove and Arrow [9] also noted that a jump of the stock of advertisingis requiredwhen the
present value is linearin the rate of change of the stock of good will.
The first named author gratefully acknowledges the financial support of the National Science
Foundation.
307
308 REVIEW OF ECONOMIC STUDIES

In the continuous (C) monopoly problem we seek the maximum of the integral
0

j e-rtq(t)p(t)dt, ...
0

for a given r>0 with respect to q subject to


p(t) = f(t)-q(t)-alq'(t)-a2q"(t) ... (1-2)
and the initial conditions that q(t), f(t) = 0 for t<O. The differential equation (1.2)
represents the monopolist's demand conditions; p is the price and q is the quantity. The
functionf(t) represents the size of the market. Thus if the monopolist's customers increase
in number and affluence over time thenf(t) is an increasing function of t.
In the discrete (D) monopoly problem
00
max E ptqtpt
{qt} 0

for a given discount factor f such that 0 <f# ? 1 subject to the demand relation
Pt =ft-qt-aqt_..1.
Notice that in the C problem the demand relation is a second order differential equation
while in the D problem it is a first order difference equation. This choice is deliberate;
a first order differential equation in the C problem is a degenerate case as will be seen by
deriving the implications of a2 = 0.
As in ordinary maximum problems it is necessary to compare a given solution with
alternatives. Hence it is necessary to define an admissible quantity path.
Definition. In the C problem an admissible q is a continuous curve with a continuous
first derivative except at a finite number of points such that for a given non-negative r,
Ie-rtl2q I and Ie-rt/2q' I are bounded. In addition we define interior and boundary curves.
Definition. In the C problem if the integrals

J ertq2(t)dt and J ertq/(t)2dt ...(1.5)


are finite, then we say that the admissible curve is an interior curve; if the integrals are
not finite then the admissible q is a boundary curve.
The C problem has an interior solution for a prescribed r if the value of the integral
(1.1) is finite for all admissible interior curves. It has an edge solution for a given r if the
value of the integral is finite for all admissible interior curves and for some boundary
curve. Edge solutions exist only for a special value of r depending on a, and a2. Related
definitions for admissible q's in the discrete problems are as follows:
Definition. The set of admissible q's in the D problem for a given ,B,0 <,B 1, consists
of all sequences {qt} such that Ifftl22qt Iis bounded. An admissible sequence is an interior
sequence if
00

fltqt2
0

is finite; otherwise, an admissible sequence is a boundary sequence.


However, unlike the C problem, we show that the D problem has only interior and
no edge solutions.
Thef's must satisfy certain conditions because otherwise no solutions would exist;
thusf cannot grow too rapidly. In the D problem it is assumed that there exists a positive
number /1o such that
t
I
CONTINUOUS AND DISCRETE TIME APPROACHES 309

is bounded. In the C problem it is assumed there exists a positive ro such that


I e- rotl2f(t)l

is bounded. For technical reasons in the C problem, f must be further restricted to the
class of continuous functions of t except at a finite number of points. The requirements
imply that
{ ert If(t)2dt andE t f
O Q

are both finite for r> ro and / <fi, respectively.'


The discount factors in the two problems can be related as follows:
er-
Demand relation (1.2) can be solved for q(t) as a functional of the p's. It turns out
that if the demand equation is such as to imply that the maximum present value is bounded,
then the purchases, q(t) planned for time t as of the initial time, t = 0, must depend on
prices both before and after time t. However, in the D problem the demand equation
(1.4) has a different property if the maximum present value is bounded. In this case the
planned purchases for period t depend on past prices and past f's. The dependence on
past prices can be interpreted as due to the customer's use of past prices to forecast future
prices.
The investment problem is a slightly generalized version of one due to Eisner and
Strotz [5]. Assume the present value of the firm's net revenue depends on its capital
stock and its rate of investment. Let
s(t) capital stock at time t,
i(t) = rate of investment at time t.
Therefore,
i(t) = s'(t) plus depreciation.
Depreciation is a function of the firm's capital stock. Hence the present value of the
firm's net revenue is given by
=
0

PV ert{P[s(t)]- C[s(t), s'(t), t]j}dt. ...(16)

The term P[s(t)] represents the gross return to the firm as a function of its capital stock.
The term C[s(t), s'(t), t] represents the firm's cost as a function of its rate of investment;
this term implicitly accounts for depreciation by the presence of s(t). However, C can
depend on s for other reasons. For example, the cost of investment may be lower for
large than for small firms because the former obtain more favourable terms from their
suppliers. A quadratic form of the investment problem is as follows:
P[[s(t)] -s(= o-55) ... (1.7)
C[s(t), sA(t),t] = s'(Y0+ 71S+Y2S'). ...(1.8)
In (1.7) the average revenue at t is a linear function of the capital stock given by co - ols.
In (1.8) the average cost of changing the size of the firm (as measured by s(t)) is given by
YO+ Y,S+ y2S'. If y1<0 then these costs are lower for firmsof largersize. The discrete
approximation to this problem would replace s'(t) by s,- s,_ and we omit the details

1 Anfis givenand we are to determinewhetherthereis a a suitablediscountrate, r, or discount factor


p to give the integralor sum, respectively,a finite value. That geometricgrowth of f, at a rate less than
$-5 iS necssary and sufficientfor the convergenceof the infinitesum is a standardresult in the theory of
infiniteseries. Since we use ordinaryRiemann integralsin this note, f(t) is allowed to be discontinuous
only at a finite numberof points. Hence for all purposes of this article,geometricor exponentialgrowth
of f is the maximumallowed.
310 REVIEW OF ECONOMIC STUDIES

which are straightforward. Our main interest in this investment problem is to demonstrate
the different role of the interest rate r in (1.6) and in (1.1).
In the monopoly problem the solutions of the C and D versions differ in several
important respects. First, as we shall show, there is an exponential approach to the
equilibrium quantity path in the C problem if and only if a2 # 0. If a2 = 0, so that the
constraint (1.2) reduces to a first order linear differential equation, then the optimal path
requires instantaneous adjustment from the initial to the equilibrium path. In the D
problem there is never instantaneous adjustment to the equilibrium path if a =# 0. As
long as a =A0, qt approaches the equilibrium path at a geometric rate.
Second, the boundary conditions for the C problem raise delicate mathematical issues.
It turns out that in the C problem there is in general an initial finite jump of the first deriva-
tive at t = 0. This jump is analogous to the jump in q itself in that C problem for which
a2 = 0.
Third, if the sufficient conditions for a maximum of the C problem are satisfied then
the demand equation (1.2) is unstable. The sufficient conditions for a maximum in the
D problem are satisfied for all values of the discount factor ,Bsuch that 0 <,B < 1 if and
only if the demand relation (1.4) is stable. These results have an important economic
interpretation in terms of the implicit price forecasting of the demand relations.
Fourth, in the C problem there is a maximum positive discount rate r dependent on
a1 and a2 that yields a finite maximum present value for all admissiblef. This is strange
because intuitively one expects the higher the discount rate the lower the maximum present
value. In the D problem no upper bound is required of the discount rate. Regardless of
the stability of the demand relation, the mathematics supports our intuition because there
is always a discount factorIB1 such that for all smaller discount factors, <,</3 _ 1, there
is a bound to the maximum present value and a maximizing sequence {qt} growing at a
geometric rate less than fl5. The optimal path is uniquely determined by the boundary
conditions and the choice of a discount factor. Stability of the demand relation in the
D problem is necessary for the existence of a bounded finite maximum present value if
the problem is to have a solution for all, < 1. If, however, the discount factor is bounded
by a number smaller than one as determined by the demand relation, then unstable demand
relations also yield bounded finite maximum present values for all discount factors less
than the upper bound determined from the demand relation.
Although certain of these issues do not arise at all in the investment problem, the
investment problem and the monopoly problem illuminate each other. It is convenient,
however, to defer further discussion of the former until some analysis of the mathematical
problem has been completed.

2. THE C PROBLEM
To some extent this problem relies on the classical calculus of variations and we
begin by seeking interior solutions. It is convenient to reduce the problem to one involving
only first order derivatives by an integration by parts. Thus letting PV, for present value,
denote the integral in (1.1), find by an integration by parts that
00
PV = | &rt[f-q-a1q'-a2q"]qdt ... (2.1)
0~ ~ 0 O
-
--a2e - rtqqf + f rt[fq - q2_(al+a2r)qq'+ a2q 2]dt. ... (2.2)
o o
Since our interest lies in bounded maxima, we obtain the first condition that an admissible
path must satisfy from the finiteness of the integrals (1.5).
lim e rtqq'= O. ...(23)
t-400
CONTINUOUS AND DISCRETE TIME APPROACHES 311

Hence (2.1) becomes


PV = a2q(0)q'(0)+ e- r[fq - q2-(a1+a2r)qq'+ a2q'2]dt. ... (2.4)

Both q(0) and q'(0) are given, making the first member of the right hand side a constant.
Therefore, maximizing PV is equivalent to maximizing the integral on the right side.
Define the function
F(t, q, q') = e-rt[fq - +a2r)qq + a2q 2]. .. (2.5)
q2-(a,

In order to maximize
T F(t, q, q')dt,

it is necessary that the function q(t) satisfy the Euler equation

Fq-d Fq, = 0.

Other conditions for a maximum are more complicated in our problem because the
terminal value of q is only restricted by the condition that PV be bounded. This implies,
inter alia, that q may grow at an exponential rate not greater than r/2. Hence it is not
possible to appeal to the " simpler" sufficient conditions in the calculus of variations
because these apply to problems in which boundary conditions are given in the form of
equalities. One of the classical conditions does apply in our problem. Thus for a maximum
it is necessary that
Fqq,<O.,
We know that if the problem has a solution at all, q(t) must satisfy Euler's equation
which is the following linear second order differential equation:

-2a2q"+ 2a2rq'-(2+ a1r+ a2r2)q+f = 0. ... (2.6)

Its characteristic equation is quadratic as follows:

-2a2x2 +2a2rx-(2+a1r+a2r2) = 0. ... (2.7)


Its roots are
r/2+(-a2r2 -4a2-2a1a2r)5/2a2. ... (2.8)
Their values depend on the restriction imposed on the coefficients of (2.6) by the require-
ment that there is a finite maximum. One requirement we know, namely,
F,,= a<0 ... (2.9)
We now show that for a finite maximum it is necessary that the roots given by (2.8)
are real. To prove this we choose an admissible path

q = Meat, ? < r/2, M = q(0), ...(2.10)


and derive additional conditions on the coefficients that are necessary for this path to
yield a finite maximum. In other words if the conditions to be derived are not satisfied
then it would be possible to choose a 5<r/2 such that the present value can be made as
large a finite positive number as we please. Furthermore, this phenomenon can occur
1 See Appendix, section 1.
312 REVIEW OF ECONOMIC STUDIES

without necessarily violating an economic restriction we have not yet mentioned, namely
non-negativity of prices. For q given by (2.10), the PV in (4) becomes
00
PV = a2q(0)q'(0)+ e- rtqfdt
00

- M2 e(2 -r)t[1 + b(a + a2r)- 2a2]dt


0

= a2q(o)q'(0)+ e-rtqfdt
0

+M2(25-r) -'[1 + (a1 + a2r)5-a2 2] ...(2.11)


As 5 approaches r/2 from below (26 - r)-1 approaches minus infinity. Hence the present
value can be made an arbitrarilylarge positive number if
lim (1 +(al +a2r)b-a2 62) < 0.
2n-r-

Therefore, to rule out this possibility it is necessary that


= 0.
1 + aj(r/2) + a2(r/2)2 ... (2.12)
Since the discriminant of the roots in (2.8) can be written as follows:
- 4a2(1 + a 1(r/2) + a2(r/2)2),
this proves that the characteristicroots of the demand equation must be real if the maximum
present value of revenue is to be bounded from above.'
This argument depends on the fact that q given in (2.10) is admissible, but nothing
has been said so far about the non-negativity of prices. Thus perhaps a quantity path
that satisfies (2.10) might require steadily falling prices. Hence if non-negativity is imposed
on prices, the q given in (2.10) might be inadmissible. However, admissibility of q for
prices non-negative depends on how rapidly the market grows. Since the growth of the
market equals the growth rate off, it is sufficientfor our purposes to display anf that grows
enough to maintain non-negative prices, and that is admissible. An example of such an
f is given by
f(t) = Kert/2/(l1+t). ... (2.13)
It is readily verified that forf given by (2.13) the integral
00
e-rty2dt

is finite. Moreover, for this f, and q given by (2.10), since prices must satisfy the demand
equation (1.2), it follows that
p(t) = e't[Ke(r/2 - )t/(1 + t)-M(1 + alb + a2 2)]. . . .(2.14)
Clearly, if K> 0 and large enough so that p is never negative, the market actually grows,
and not only are prices positive, but also they become arbitrarily large. In this case the
preceding argument applies full force and inequality (2.12) is necessary for the existence
of a bounded positive maximum present value.
The inequality given in (2.12) together with the first necessary condition for a maximum,
a2 <0, implies that an upper bound on the discount rate r is necessary if the present value
is not to become unboundedly large. This result has generated so much interest that it is
helpful to see why the same phenomenon does not arise in the continuous investment
1 See Appendix, section 2.
CONTINUOUS AND DISCRETE TIME APPROACHES 313

problem described before. In the investment problem the present value of the firm's net
revenue satisfies

PV = { ert{s({0 - als)- s(0 + YlS+ y2s')}dt. ...(2.15)


0

Compare this expression with the present value given in (2.4) that applies to the monopoly
problem. The same combination of q and derivatives appears in (2.4) as in (2.15) with the
trivial exception that (2.4) lacks a term in q' alone. The significant difference is that the
coefficient of qq' is (a, + a2r), which depends on r, while the coefficient of ss' in (2.15) is Yi,
a constant independent of r. In both problems for a bounded maximum it is necessary
that the coefficient of the squared derivative term must be negative. However, in the
monopoly problem this fact together with the fact that r appears in the coefficient of
qq' " explains " why there has to be an upper bound on r. For if r gets too big, then since
a2 <0, the coefficient of qq' gets too small and this causes trouble. In the monopoly
problem the discount rate appears twice, but in the investment problem it only appears
once.
The monopoly problem can also be interpreted as a constrained maximization problem
where the constraint is a demand schedule that happens to be representableas a differential
equation. By solving for prices in terms of quantity and its derivatives, the problem is
converted into an unconstrained maximization problem and is thus more easily solved.
The investment problem assumes at the outset a form of the present value functional
which depends on the capital stock and its rate of change. If instead the firm were assumed
to face a supply schedule of capital goods such that the price of the capital goods is a
function of the quantity purchased and its rate of change then the investment problem
becomes a monopsony problem and would present exactly the same phenomena encountered
in the monopoly problem. In the version of the investment problem that we use, the cost
of changing the capital stock has no monopsonistic elements though there can be quantity
discounts. As given in (1.8) the cost of changing the capital stock does not allow for the
firm's expectations about the subsequent prices of capital goods. Were expectations to
enter, the investment problem would formally resemble the monopoly problem.
We now resume our discussion of the monopoly problem in its continuous form. We
have seen that the characteristic roots of the Euler equation must be real if there is to be
a bounded maximum present value. There is still the possibility that the roots might be
real and equal. For this to occur the discriminant must be zero. The case where the
discriminant is zero is more complicated and involves the existence of an edge solution.
Hence we first complete our analysis of the easier case where the roots are real and distinct.
One root, say A1, is less than r/2, and the other root, A2, is greater than r/2. Hence at
least one root and possibily both are positive.
Given that the roots are real and unequal, we exhibit a solution satisfying the Euler
equation and yielding a finite maximum. The general solution of (2.6) has the form

q(t) = (Al 2)- {elt[cl +(2a2)-1 J e-A sf(s)dsj

+ eA2t [C2+ (2a2)-1 e 2sf(s)ds]} ... (2.16)

and the c's are constants to be determined by the boundary conditions. The optimal
quantity path must pass through the initial value q(0) (we shall see why in a moment),
314 REVIEW OF ECONOMIC STUDIES

and condition (3) must be satisfied for the present value to be finite. These two require-
ments imply, respectively, that

q(O) = (Cl + C2)(Al + A2)

C2 = -(2a2) 1 e 2sf(s)ds.

The latter condition ensures that the second term on the right side of (2.16) approaches
zero as required by condition (2.3). Substituting these values of the constants into (2.16)
gives the following expression for the quantity path:

q(t)= e 't {q(O)+(2a2)"1GA1_Z2)V1 [J' e Astf(s)ds-J e- 2sf(s)ds]}

+ (2a2f '-1 -1 .. . (2.17)


-
_2) eA2t e-A2sf(s)ds.
Jt

Recalling that Al <r/2, it is easy to verify that q(t) given by (2.17) actually gives a finite PV.
We now consider the boundary conditions in greater detail. Since the Euler equation
(2.6) is a linear second order differential equation with constant coefficients, it has two
degrees of freedom that determine an optimal path up to the choice of two boundary
conditions. One of these degrees of freedom, represented by the constant c2, is chosen to
guarantee that the solution satisfies condition (2.3). This leaves one degree of freedom to
fix the path at t = 0. In fact both q(O)and q'(O) are given. Hence if one is used to fix the
present value maximizing path at t = 0, the other cannot in general satisfy the path thus
determined. We now claim that q(O) and not q'(O) must be used to fix the Euler path,
and that q'(0) as given initially cannot in general equal the slope of the path at t = 0.
To see why, suppose that the Euler path were fixed by the initial value of q'(0) instead of
q(O) and that q(O) were required to jump instantaneously to the path. This jump would
make the value of the integral arbitrarily small because a2q'(t)2<0.1 Hence permitting
a jump in q would make the integral approach the value minus infinity. Therefore, jumps
in q are forbidden and the optimal path must be determined by the given initial value,
q(O), together with condition (2.3). However, once the path is thus determined, the slope
of the path at t = 0 does not necessarily equal the given initial slope q'(0). Thus there
must be a jump in q'(t) at the point t = 0. It can be shown that a finite jump in the initial
value of the derivative imposes no cost as measured by the integral.2 Hence while a jump
in q itself is prohibited, a jump in q'(t) at t = 0 is generally required.
Now consider the case in which the discriminant of the roots in (2.8) is zero. This
case opens the possibility for the existence of edge solutions. In this case both character-
istic roots of the Euler equation (2.7) equal r/2 so that the Euler equation reduces to

(- - r/2) q(t)= f(t)/a2.

For the discriminant to vanish, it is necessary for r to satisfy the following quadratic
equation: 0
a2r2 +2a,r+4 = O
Since the characteristic equation of the demand equation (1.2) is
a2x 2+ax+1 = O,

1 See Appendix, section 3.


2 See Appendix,section 4.
CONTINUOUS AND DISCRETE TIME APPROACHES 315

it follows that r/2 is also a characteristic root of the demand equation. Hence the char-
acteristic equations of the Euler equation and the demand equation have a common
root, r/2, allowing us to represent the demand equation (1.2) as follows:

p(t)P= f(t) -a2 d - r/2) d _ ar q(t).


2~~~ dt d
We now show that in this case there is a maximum growth rate of q(t) and that the
problem has a finite maximum present value so that an edge solution exists provided
f(t) ? Ce6t,where 6 is less than r/2 and r is given. The example,
f = er,12/(I + t),
which is admissible, illustrates the necessity of imposing some conditions on f for there
to be edge solutions since for this choice of f and q = ertI2,the value of the integral

T e7rtq(t)f(t)dt
0

is infinite. Hence without loss of generality let C = 1 and set


f(t) = eat.
For this f, the solution of the Euler equation becomes
q(t) = clertl2+c2tertl2 + [2a2(6- r/2)2] le5t.
It is readily verified that setting c2 = 0 and
c, = q(O)-[2a2(6- r/2)2]-

yields a finite PJVin (2.1) although


lim e tqq'o 0
t-4cGo

Moreover, in this case f < 0 since a2 < 0.


Next we observe that the rate of growth of q(t) in general is the larger of the following
two numbers: the smaller characteristic root of the Euler equation, and the growth rate
of f. The maximum value of the characteristic root of the Euler equation occurs for that
r which makes the discriminant of the roots of the Euler equation equal zero. This follows
from (2.12) and a2<0. Together these two conditions imply that the largest admissible r
is the positive root of
a2r +2a,r+4 = 0, ... (2.18)
since this equation always has two real roots of opposite sign if a2 <0. The positive root
is precisely that value of r which makes the discriminant equal zero and q(t) grow most
rapidly.
We conclude that for a given pair (a1, a2), there exists a positive r that is the largest
root of (2.18), yields a bounded maximum PV, and makes revenue grow most rapidly.
This maximum growth rate is that value of r for which the demand equation (1.2) and the
Euler equation (2.6) have a common characteristic root equal to r/2.

3. THE D PROBLEM'
Substituting (1.4) into (1.3) we obtain

PV= Z ftqtjft-qt- cxqt1].


0

1 Our paper [7] contains a detailed treatmentof the general discrete problem allowing the p's and
q's to be vectors and the constraintsto be given by a set of linearnth order differenceequations.
316 REVIEW OF ECONOMIC STUDIES

We seek a sequence {qt} to maximize the present value. Straightforward methods of


differential calculus apply and give the necessary conditions for a maximum. Thus for a
maximum PV it is necessary that
OPV =0
=f0-2q0-c43q1
aqo
apv = 0
= fl[f1- aqo-2q1-cf4q2]
aq,

dPV = 1] = 0.
flt[f-xq,1-2qtf-afq,
aqt
It is sufficient for a maximum that the infinite band matrix (known as a Toeplitz matrix),

2 p*5 o 0 0 0

fl,5o 2 fl.50 0
?

0 2 -50C
Q
4l5a

be positive definite. This can be shown to be equivalent to requiring that the difference
equation as follows:
ocqt_- +2qt+aflqt+j =ft ...(3.1)
must not have a characteristic root of modulus p35. This condition is analogous to the
requirement for an interior solution of the C problem that the characteristic roots of the
Euler equation must be real and not equal to r/2. The characteristic equation of (3.1) is
az-' +2+ xflz = 0 .. .(3.2)
whose roots are
(1B)'1[- 1+?(1-ot2fl)5].
Denote the roots by ,u1and 12 where I 91 I< 1 2 l. For p, not to have modulus /3., it is
necessary and sufficient that the discriminant satisfy
1 -x23> 0. ...(3.3)
Hence both roots must be real. Condition (3.3) also shows the role of the discount factor
in the D problem. If this condition must be satisfied for all ,Bsuch that 0<,B < 1, then
I a I< 1, meaning that the demand relation must be stable. However, if the discount factor
may be chosen arbitrarily in the range (0, 1), then it follows from (3.3) that for any value
of o, there is a set of f's determined by the condition that
O<13<o 2

such that for any ,Bin this range, there is a bounded maximum present value. In general
this requires geometric growth of qt at a rate less than fl-5. 1 Thus the optimal quantity
sequence satisfies the difference equation as follows:
1Equation(3.2) converts into a reciprocalequation in w by the transformation
z =3-5w.
This gives
Cfl,5w+ 2 + oc 5w-I ==0. ... .(3.2')
For a maximumto exist, it is necessaryand sufficientthat the reciprocalequation (3.2') has no roots of
modulus1. The derivationof this result is involvedand we referthe readerto our paper [7]. This implies
that the originalequation (3.1) have no roots of modulus P'5 or P-5.
CONTINUOUS AND DISCRETE TIME APPROACHES 317

00
tS
qt-l-lqt-l= =s = t E2 f4
.. (.4
s=t

where I , I is not necessarily less than one.'


In the D problem only interior solutions can yield a finite value of the maximum PV.
This follows from the fact that the characteristic equations of (1.4) and (3.1) cannot have
a common root so that if one root of (3.2) had a modulus f- 5 then so would the other,
and the maximum PV would be unbounded.

4. INTERPRETATION
The C and D problems differ in several important respects. In the C problem, if
a2 = 0, then (1.2) reduces to a first order linear differentialequation and the Euler equation
(2.6) reduces to
q = f/(2 + a1r), ... (4.1)
implying instantaneous adjustment to the optimal path. There is no cost in allowing
q(t) to be discontinuous at t = 0 in this case because the integrand does not contain the
term q'(t)2. If a2 <0, then there is in general a finite jump in q'(0). In the D problem,
constrained by a first order linear difference equation, instantaneous adjustment is never
optimal. Instead there is a geometric approach throughout. Still more striking features
distinguish the two approaches.
Consider the conditions sufficient for a maximum of the C problem given above.
These require a negative value of a2. Therefore, the differential equation (1.2) must be
unstable! (A necessary and sufficient condition for the stability of (1.2) is that both a1
and a2 are positive.2) Hence, if the differential equation (1.2) is stable and f grows
exponentially at a rate less than r/2, then the monopolist can choose an output path yielding
an infinite present value. In the D problem this is impossible if the difference equation
expressing the demand conditions is stable. What explains the surprisingdifference between
the two approaches?
A first answer to this question derives from an economic interpretation of a stable
differential equation expressing demand conditions. Equation (1.2) represents the demand
for the product by the monopolist's customers. If this is a stable differential equation,
then its characteristic equation, a2
a2x2 + alx+l =-0 ... (4.2)
has roots with negative real parts. Denote the roots by a, and a2. Solving (1.2) in order
to represent quantity as a function of prices andf's gives:

q(t) = (a1 - a2) 1 {eelt [cl + a2 f e~-ls(f- p)ds]

+ ea2t C2+ aj e-'2s(f-p)ds]} ... (4.3)

where the constants c1 and c2 are determined by the initial conditions. Observe that (4.3)
gives the quantity demanded at time t as a function of prices up to and including time t.
There is no provision for prices expected in the future beyond time t. Accordingly, given
such short-sighted consumer behaviour, it is hardly surprising that a profit maximizing
monopolist can find an output path yielding an infinite present value.
Under conditions of perfect competition a stable demand equation presents no logical
difficulty. No individual buyer would defer purchases in the hope of changing the current
1 See Appendix, section 5.
2 See Birkhoffand MacLane [3], pp. 115-116.
x
318 REVIEW OF ECONOMIC STUDIES

price because his individual purchases are too small a part of the total demand to exert
any effect on the current price. Moreover, competition among the sellers prevents any
one of them from exploiting consumer lack of foresight and maintains their return at a
competitive level. This argument does not, of course, deny the benefit to producers of
forecasting prices. However, in a competitive market, forecasting helps producers by
improving their inventory, output and employment policies and not by exploiting customers.
Now consider an unstable demand relation. This means that one of the roots of
(4.2) say a,, has a negative real part, while the other, a2, has a positive real part. Therefore,
the quantity demanded at time t depends on future as well as contemporary and past
prices. Instead of (4.3) there is

q(t) = (a! -a2)'- {eealt[1 +a- f e- is (f-p)ds] + a' f- e(t-S)af2(f-p)ds}. ...(4.4)

In contrast with (4.3), only the constant c1 depends on the initial conditions. Nevertheless,
although the consumer demand depends on future prices, it is still possible for the mono-
polist to secure an unlimited, positive present value. This can happen if a2 <0 but condi-
tion (2.12) is not satisfied so that the discriminant is negative. Moreover, we have given
an example of a path that would yield an unbounded positive present value in this case.
Therefore, although consumer forecasting is necessary for a bounded maximum present
value, it is not sufficient.
In the D problem even if the current demand does not depend on future prices, an
unbounded maximum present value is impossible if the demand relation is stable and f
grows at a geometric rate less than fi- ", fi < 1. Moreover, the quantity demanded at time
t is implicitly a function of a weighted moving average of all past prices with geometrically
declining weights. The latter function of past prices can be interpreted as a forecast of
future prices. This result sharply contrasts with the C problem in which perfect foresight
is necessary for a finite maximum.
The basic explanation of the difference between the C and D problems depends on
the interpretation of a continuous curve possessing a continuous first derivative. For
such a curve, the slope at any point t is the same whether it is approached from the right
or left. Indeed a continuous slope at a point means that the right and left derivatives
must be equal. This equality implies that in every small neighbourhood of a given point
the future is just like the past, and hence completely predictable. It is inherent in the
geometry of a smooth curve that uncertainty cannot enter (cf. footnote 1, page 321). Once
this is understood the same phenomenon can be introduced in a suitably modified discrete
problem making the implicit treatment of uncertainty in the continuous problem emerge
with great clarity.
A first avenue of approach to a discrete version of the continuous problem approximates
the derivative by a first difference. Thus one might think that the analogue of the C
problem is a D problem in which q'(t) is replaced by qt-qt-1. Evidently, this leads to
precisely the same D problem under consideration all along. The correct discrete approxi-
mation to the C problem replaces q'(t) by a symmetric first difference as follows:

q'(t) qt+I-qt-1.
This means introducing a future q, qt+1. Nor is this all. The discount factor, f,, must also
be accounted for. Hence we require a perfect fl-symmetric first difference defined as
follows:
_
fl"qt+ I-p qt- 1.
This expression discounts the term qt+1 that refers to the future period by the factor
/"5 while the term qt- l from the past is increased by the factor p- '. Therefore, the perfect
/3-symmetricfirst difference places the past q, qt-1 and the future q, qt+ , on the same
present value footing in the time period t.
CONTINUOUS AND DISCRETE TIME APPROACHES 319
Now write a new demand relation in terms of a perfect fl-symmetricdifference equation
as follows:
2qt+ a(f-5qt+ I - 5- qt- 1) = ft - pt ..s (4.5)
We assert that (4.5) is the correct discrete analogue of the C problem for which a2 = 0.
To verify this assertion, maximize the PV given by (1.3) subject to (4.5) instead of (1.4).
The necessary maximum conditions require the optimal sequence {qt} to satisfy
2qt+ (x-5qt+1 -f* 5qt- )+2qt +a(f 5qt- I-_ 5qt+1) =ft, ...(4.6)
which simply reduces to
4qt= ft ... (4.7)
Equation (4.6) is derived in precisely the same way as (3.1), that is, by substituting the
value of pt given by (4.5) into (1.3) and then calculating the partial derivatives of PV with
respectto qo,ql, q2 .. Observethat (4.7) exhibitspreciselythe samemode of behaviour
as the C problem in which a2 = 0, namely, instantaneous adjustment to the optimal path!
Moreover, for this problem, a finite maximum present value always exists for all values
of a and without restrictions on f,; clearly the restrictions on ft must be retained.
Finally we come to the role of the discount factor. In the original D problem, or,
for that matter in the general class of D problems containing lagged q's, it is always possible
to guarantee the existence of a finite maximum present value by choosing a sufficiently
small positive value off,. Thus the D problem has only interior solutions. This is not
true in the C problem in which for special values of r there are edge solutions. We see
from (2.12) that if r is too large then there would be an unbounded positive present value.
Normally, one would think that higher discount rates would help rather than hinder
the attainment of a finite maximum present value. What explains this puzzling aspect
of the C problem?
The explanation is that, paradoxically, the C problem as stated is not symmetric
enough. This is suggested by consideration of the analogous discrete problem that is
imperfectly symmetric. We now show that in the imperfectly symmetric discrete problem
the same phenomenon occurs as in the C problem. Thus write a demand equation as
follows:
2qt + x(qt + I1-qt - 1) = ft - pt. ... (4.8)
The quantity path maximizing (1.3) subject to (4.8) necessarily satisfies the following second
order linear difference equation:
2qt +a(qt + -qt- 1) +2qt +a( l-qt - 1-fqt+ 1) = ft. ..-(4.9)

This can be verified by calculating the partial derivatives of PV with respect to the q's
seriatum. If f, = 1, corresponding to r = 0 in the C problem, then a finite maximum
PV exists although ft has to approach zero. In addition there is a finite maximum PV
if f, is close to one. However, there is no bounded maximum if f, gets too small! 1
Intuitively, the reason is that iff, is too small it makes past receipts accumulate at too
high a rate at the same time that it makes the present value of future receipts small (a low
discount factor means a high discount rate) so that eventually even small sums become
arbitrarily large. This cannot occur in a perfectly symmetric C problem. A perfectly
symmetric C problem requires the coefficients to satisfy the equation as follows:
al+a2r = 0. ...(4.10)

1 Theimperfectly:-symmetricdemandoperatorcan havean edge solutionif the characteristicequations


of (4.8) and (4.9) have a common root. This common root is p-5 and occurs for a P that satisfies
2+OC(P5-P-'5) = 0.
Thusqt grows at the maximumrate f-5 and the analysisis exactlythe same as for the edgesolutiondiscussed
at the end of section 2.
320 REVIEW OF ECONOMIC STUDIES

If the coefficients satisfy (4.10) then a finite maximum PV exists for the C problem for all
positive values of the discount rate r. Thus (4.10) is the sought for condition on the
coefficients of the demand relation that makes the C problem perfectly symmetric. In
terms of the roots of the characteristic equation of the demand relation (1.2),
alla2'-(- a1+a2) and a2 = (a1a2) '.
Therefore, (4.10) implies that
l+Uf2 = r. ...(4.11)
In other words, the arithmetic mean of the roots is r/2. In the perfect fl-symmetricdemand
relation given by (4.5), denote the characteristic roots by s1 and s2. For (4.5) we have
S1 S2 =-2(o4f5f)-' and S1S2 = -- '.
Therefore, the geometric mean of the roots is ,B5. Recalling that ,Be-r, this is precisely
the same relation as in the perfectly symmetric C problem.
The symmetry implied by (4.11) is directly visible if the demand relation is converted
into an expression relating the quantity demanded at time t to past and future prices.
Since
2= r-u1>O, a,<O,
equation (4.4) implies that

q(t) = (2a, -ry'


{[cleritt+ a- j' ea1(t-S)(f-p)ds]

+a2f1 er(s-t)ea1(st)(f_ p)ds} ...(4.12)

Equation (4.12) shows that the quantity demanded at time t depends symmetrically on
past and discounted future prices. Thus prices for an interval (t - s) into the past receive
a weighte' (t - s) while pricesfor an interval(s - t) into the futurereceivea weighte" (s - t)
discounted by the factor er(st).
This symmetry condition extends the well known symmetry of price effects on quantity
of classical static demand theory to a dynamic situation. In the static theory a consumer
is assumed to choose n goods so as to maximize utility subject to a budget constraint.
It is shown in this theory that if real income is held constant in either a Slutsky or Hicksian
sense then the cross price slopes are equal. More precisely, let the prices of the n goods
be denoted by pj, j = 1, ..., n and the quantities be denoted by qj. The well known result
of classical demand theory is that

qj _= Oqi, real income constant, i #1.' ...(4.13)


api @pj
In the present situation a commodity available at a time t can be considered as a different
good than the same physical commodity available at another time t'. The symmetry
condition of the dynamic case given by (4.10) is akin to the classical demand condition
(4.13) since it makes quantity demanded at time t depend symmetrically on past and
discounted future prices!
The role of symmetry is interesting in that C problem for which a2 = 0. Consideration
of a path q(t) = Meat as in (2.10) implies that
1 +a1r/2>0 ...(4.14)
is necessary for the existence of a finite maximum PV. Hence the factor (2 + a,r) appearing
in (4.1) must be positive. It is curious that in this C problem it is necessary to impose a
condition on a, while in the perfect fl-symmetricD problem constrainedby (4.5) no condition
1 See Allen [1], Chapter 19.
CONTINUOUS AND DISCRETE TIME APPROACHES 321

on a is necessary. The explanation is that if a2 = 0, the corresponding C problem is not


symmetric enough! Condition (4.10) is pertinent and suggests that for perfect symmetry
in a C problem for which a2 = 0, it is also necessary that a, equal zero. Therefore, if
the differential equation (1.2) lacks the term q" then it also cannot include the term q'.
In other words there is no perfectly symmetric first order linear differential equation.
This interpretation explains the paradoxical role of the interest rate in the continuous
problem by showing that in its original form the C problem was insufficiently symmetric.
By introducing condition (4.10) the dynamic demand relation becomes symmetric in the
same sense as the classical static demand schedule. The most important moral to be
drawn from this is why symmetry of a demand relation over time is unreasonable in the
real world. This is because such symmetry would destroy the distinction between the
past and future. Symmetry in the static case arises essentially because the order in which
the goods are chosen does not matter. But in choices over time the order in which goods
are chosen does matter for two reasons. First, it is impossible to change the decisions
of the past. Second, the impossibility of changing past decisions would not matter only
if there were perfect foresight because then there would never be need of changing past
decisions. Perfect foresight implies the absence of mistakes. In the real world there is
uncertainty and perfect foresight is impossible. Hence current choices depend on the
fact that one must adapt to new unforeseen circumstances as well as possible. In a discrete
model the mathematics allows us to take into account the irreversibility of past decisions
and the effects of uncertainty.'

Universityof Chicago L. G. TELSER


R. L. GRAVES

APPENDIX
Section 1
In the classical calculus of variations problem with fixed initial and terminal conditions,
two conditions would be sufficient for a weak relative maximum:
(i) Fq,q,< 0
known as Legendre's condition, and the negative definiteness of the second variation,
which is equivalent to requiring negative definiteness of the matrix:
1
(ii) [Fqq Fqq,
Fqq, Fq,q,
The latter reduces to a matrix of constants in our case as follows:
[-2 -(al+a2r)]
L-(al +a2r) 2a2
and (ii) implies that
4a2+(al+a2r)2<O.
Since the Euler equation has a unique solution satisfying given boundary conditions, a
weak relative maximum is actually a strong relative maximum and these conditions
together furnish a complete characterization of the solution for which initial and terminal
points are given. However, these conditions are too restrictive in our problem because
we require only the existence of a finite maximum and do not actually fix the terminal value
of q. Calculus of variations problems with variable end-points are quite complicated
and the only condition we can readily use is that of Legendre (i). See Bliss [4], sections.
23, 24, and Chapter VI.
1 For example,a random walk traces a continuous curve which nowherehas a derivative. Thus one
always can tell whereone is but not whereone is going.
322 REVIEW OF ECONOMIC STUDIES

Pixley [10] shows that if the admissible paths permit a finite number of discontinuities
of both q and q', then extremals must satisfy the following analogue of the Euler equation:
Fq = Fq,= 0.
In the present application this requires a2 = 0, a result which agrees with ours as we shall
see. See also Arrow [2] and Nerlove and Arrow [9].

Section 2
The following brief proof of a necessary condition for a maximum PV easily generalizes
to nth order linear differential equations. Let
A(x) = a2x2 + alx +.

This permits us to express the demand equation in powers of the differential operator d
dt
as follows:
=fA (d q)

PV = Xe-rt qf-qA (d dt

For a finite PV, the choice of q = Meat, r - 26 >0 requires that

-M2 e(26-r)tA(b)dt =-M2(r-26)-1A(6) < 0

for all 3 such that r -2 > 0. In particular,


A(r/2) > 0,
which is (2.12) in the text.
A similar argument proves directly the necessity of the condition a2<0 for a finite
maximum PV. To prove this, suppose a2>0, choose 6<0 and large in absolute value
and observe that PV can then be made arbitrarily large.
In the general case where A(x) is a polynomial of degree n, the corresponding necessary
conditions for a maximum PV are
(i) A(x) > 0 for 0 < x ? r/2
and
(ii) a,( -1) <O0.

Section 3
To prove that a jump in q(O)makes PV arbitrarilysmall, it is sufficient to consider the
term in the integrand given by
a2e tq (t)2,

taking note that a2 <0. Denote the given initial q by q(O-) and pass a straight line segment
through the points q(O-) and q(h) where q(h) is on the Euler path and h>0. Let q(t) be
a point on this line segment and define m so that
q(t)-q(h) = q(O0)-q(h) = m
t-h -h
Since the point q(h) satisfies the Euler equation,
lim q(h) = q(O+) A q(O-),
h-O+
CONTINUOUS AND DISCRETE TIME APPROACHES 323

meaning that q jumps from q(O-) to q(O+) which is on the Euler path. On the given line
segment, q'(t) = m. Therefore,
rh
e-rtm2dt = -(1Ir)(e- rh - e-rk)m2, O<k<h.
k

lim -(1Ir)(e-rh-e-rk)m2 =-(1/r)m2(e-rh- 1)


k-O+
e -rh -1 -rh+O(h2)
lim (-1/r)m2(e-rh-1) = lim hm2
h-O + h-O +

= lim [q(0-)-q(h)]2/h
h-O+

= lim [q(O-)-q(O+)]2/h
h-O+

= 00.

Hence the maximizing curve has to be continuous at t = 0. The same argument shows
that it must be continuous everywhere.

Section 4
To show that a jump in q'(0) is costless if the curve is continuous, we approximate
the path by two line segments of different slopes joined at t = 0. Let q(O-) and q'(O-)
refer to the given initial conditions and let
q(O-) = q(O+) = q(O)
q'(O-) :0 q'(O+)
Consider a path such that
q(t)-q(k) = q'(O-)(t-h), if t, k<O, t>k.
q(t)-q(h) = q'(O+)(t-h), if t, h>O, t<h.
The two terms in (2.2) pertinent to the argument are as follows:
00 00
-a2 -rtqqt +a2 e-rtqf2dt.

We perform the integration on a line split into two segments, [k, h], (h, oo]. For the first
term above, we have (omitting a2),
+(O)q'(O+)_ q(-)q'(O-)+ lim e -rtqq'- q(? )q (? )
t- oo
which reduces to
-q(O-)q'(O-).
The terms directly involving the integrals on the segment [k, h] give the following
rh ro rh
e-rtqf2dt e-rtq(o-)2dt+ e-rtq'(0+)2dt
Jk Jk sO

11r)[q'(0 )2( -e- rk)- q'(0+)2(1 - e-r)]


= (-

As kO-, the first term approaches zero, and as h-*0+, so does the second. The integral
on the segment (h, oo) is straightforward.
The Euler path is continuously differentiable for all t>0 and the only discontinuity
required is the jump of q'(0) to get on the Euler path. Such broken extrema are not
uncommon in the calculus of variations.
324 REVIEW OF ECONOMIC STUDIES

Section 5
Thereis an equationanalogousto (3.4) in the C problem. The differentialequation
(2.7) can be writtenas follows:
-2a2 (d 21 d 2 q-f
dit dA.ct q
whichreducesto
(- -21) q = (2a2)' (d -A2)fi

The latterexpressiondenotesan integral. To provethis let

g = (d _22)f,

so that g is the solutionof the followingfirstorderlineardifferentialequation:

I'd
_-22 0

Its solution,as is readilyverifiedby differentiation,is

g = eA2t
(C+ Xe-A2fsr) ds
The solutionis boundedif
00
c= - { e2S f(s)ds
o
since 2 >0. Therefore
g _ _eA2t e- 2Sf(s)ds.
Hence finally
q'-1 q = (2a2)-'1 e-2(S-t)f(s)ds.
t
Similarly, we can write the solution of (3.4) in a form like (2.17).

REFERENCES
[1] Allen, R. G. D. Mathematical Economics (MacMillan, London, 1956).
[2] Arrow, K. J. " Optimal Capital Adjustment ", Studies in Probability and Manage-
ment Science, edited by K. J. Arrow, S. Karlin and H. Scarf (Stanford University
Press, Stanford, 1962).
[3] Birkhoff, Garrett and Saunders MacLane. A Survey of Modern Algebra, rev. ed.
(MacMillan Co., New York, 1953).
[4] Bliss, Gilbert A. Lectures on the Calculus of Variations (University of Chicago
Press, Chicago, 1946).
[5] Eisner, Robert and Strotz, Robert H. "Determinants of Business Investment,
Impacts of Monetary Policy, Commission on Money and Credit (Prentice-Hall,
Incorporated, Englewood Cliffs, New Jersey, 1963, esp. pp. 71-72, 79-81).
[6] Evans, Griffith C. Mathematical Introduction to Economics (McGraw-Hill, New
York, 1930, Chapter XIV).
CONTINUOUS AND DISCRETE TIME APPROACHES 325

[7] Graves, Robert L., and Telser, Lester G. "An Infinite-Horizon Discrete-Time
Quadratic Program as Applied to a Monopoly Problem", Econometrica, 35
(April 1967).
[8] Hotelling, H. "The Economics of Exhaustible Resources ", Journal of Political
Economy, 39 (1931), 137-175.
[9] Nerlove, Marc, and Arrow, K. J. "Optimal Advertising Policy Under Dynamic
Conditions ", Economica N.S., 29 (1962), 129-142.
[10] Pixley, Henry Howes. "A Problem in the Calculus of Variations Suggested by a
Problem in Economics ", Contributions to the Calculus of Variations 1931-1932
(University of Chicago Press, Chicago, pp. 131-189).

Das könnte Ihnen auch gefallen