Review of Economic Studies, Ltd.

Optimal Growth with Scale Economies in the Creation of Overhead Capital

Author(s): M. L. Weitzman
Source: The Review of Economic Studies, Vol. 37, No. 4 (Oct., 1970), pp. 555-570
Published by: Oxford University Press
Stable URL: http://www.jstor.org/stable/2296485
Accessed: 29-12-2015 07:34 UTC

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at http://www.jstor.org/page/
info/about/policies/terms.jsp
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.

Review of Economic Studies, Ltd. and Oxford University Press are collaborating with JSTOR to digitize, preserve and
extend access to The Review of Economic Studies.

http://www.jstor.org

This content downloaded from 18.9.62.46 on Tue, 29 Dec 2015 07:34:37 UTC
All use subject to JSTOR Terms and Conditions

Optimal
Economies

with

Growth
in

the

Scale

Creation

of

Capital1

Overhead

1. SUMMARY
Following closely the approach to optimal economic growth taken in the work of Frank
Ramsey [11], a highly simplified two-sector model is presented in which the " overhead
capital " sector exhibits increasing returns to scale. Basic properties of the optimal growth
path are discussed. From an economic standpoint, the model might be relevant in bearing
on some basic issues of development programming. Mathematically, this kind of a model
has an interesting structurebecause it is a combination of convex and concave sub-problems.

2. INTRODUCTION
In the context of development economics it is useful to distinguish two types of capital
according to how round-about a role each plays in producing output. One type, the
quantity of which is denoted K, is the ordinary directly productive quick-yielding capital
which, when it is combined with labour, creates output according to classical laws of production. A second kind of capital, Kl, is the indirectly productive infrastructurewhich lays
down the basic framework within which directly productive economic activities can function. Capital of this variety has come in for increased scrutiny by development economists.
At least in part this is due to the growing suspicion that, capital, comprising those essential
services without which ordinary production cannot operate, plays an especially important
role in the early stages of economic growth.
For the purposes of this paper the total capital stock of the economy is thought of as
being partitioned between two sectors-K. belonging to the acsector and Kl to the, sector.
This being the case, it becomes a fair question to ask for operational criteria which can be
used to distinguish acfrom, capital. Unfortunately it is difficult to be precise about this
issue. For one thing it depends upon how aggregative a view one is prepared to take.
Considering an entire economy on the most general level, ,Bmight consist of all social
overhead capital including public service facilities for education, scientific research, sanitation engineering, public health, and law enforcement, agriculturaloverheadsuch as drainage
and irrigation systems, and hardpublic utilities like transportation, communications, power
and water supply installations. A somewhat more satisfactory interpretation might limit
/ to the hard public utilities. There is even an interesting way of looking at this model
which restricts the economic scenario to manufacturing and treats, as structures, cxas
producers' durable equipment.
For the purposes of this paper probably the most useful formulation is the middle one
which treats, as overhead capital for producers' services. In any case, the basic features
are taken to be the following.
1 For their helpful comments I would like to thank D. Cass, T. C. Koopmans, and A. S. Manne.
The researchdescribedin this paper was carriedout under grantsfrom the National Science Foundation
and from the Ford Foundation.
555

This content downloaded from 18.9.62.46 on Tue, 29 Dec 2015 07:34:37 UTC
All use subject to JSTOR Terms and Conditions

556

REVIEW OF ECONOMIC STUDIES

(i) Capital of thef type is strongly complementary with cx. Investment in cxcapital will
be productive only if it has been preceded by sufficient investment in ,Bcapital.
(ii) The ,Bsector is highly capital intensive and usually consists primarily of structures
and installations. It is typically characterized by a significantly higher capital-labour ratio
than the cxsector.
(iii) There are substantial economies of scale in creating, capacity. The main reason
is that due to indivisibilities there is obvious cost lumpiness involved in creating a transportation, communications, or power and water supply system as a whole. Geometricengineering considerations are also important in the case of many structures because the
cost of an item is frequently related to its surface area while the capacity increases according
to its volume.1
(iv) Both ,Band cxcapitals are specific to the role for which they have been created and
cannot be shifted.
3. THE BASIC MODEL
The highly stylized economy under consideration is centralized and closed. A single
homogeneous output, denoted Y, is produced which is perfectly general before it has been
committed, and can be used for any purpose. The planners seek to maximize welfare by
appropriately manipulating the available instruments-in this case the destination of final
output. For simplification the following are assumed: stationary labour force and population, tastes independent of time, constant technology, no capital deterioration.
As an abstraction of proposition (ii), it is postulated that ,Bcapital has only negligible
manpower requirements. This makes the labour allocation problem trivial because all
available workers will be assigned to work with cxcapital.
If, capital were abundant at time t, Y(t) would depend only on the stock of Ka(t) and
the labour force. Since the latter is treated as constant, the production function in this
case could be written as
Y(t) = F(Ka(t)).
Decreasing returns to a single factor implies that F(K.) is concave. Purely for convenience,
we assume that a first derivative exists and that F'(K,) >0 for all K. > 0
With K,(t) plentiful, the production function would simply be
Y(t) = Kp(t).
Note the implied asymmetry in capital measurement; K. is gauged by the usual criterion of
real production cost, whereas it will prove useful to quantify K, in capacity units. Of course
strict identification of K, with " capacity " would be possible only if there were a negligible
elasticity of substitution between K, and F(K.), a condition which we readily assume
following (i).
In the general case,
Y(t) = min {F(K,(t)), Kp(t)}.
With cxcapital all investment goes into capital formation in the usual direct form
Ka = ial

where I. denotes investment in cxcapital.2

However, with, capital there is a meaningful distinction between capital accumulation
and investment. Because of the presumed increasing returns to scale described in (iii) it
will typically be better not to invest directly in, capital. Rather it will pay to first accumulate what could be thought of as either a generalized inventory of materials or as projects
1 In addition the usual internaleconomiesof specializationand informationhandlingmay be present.
Note that the increasingreturnsrelates only to the design stage when the amount of installedcapacity is
treated as variable. Ex post the size of an installationis consideredto be fixed and unalterable.
2 A dot over a variabledenotes differentiationwith respect to time. Variablesmay not be explicitly
specifiedas functions of time if this interpretationis otherwiseclear.

This content downloaded from 18.9.62.46 on Tue, 29 Dec 2015 07:34:37 UTC
All use subject to JSTOR Terms and Conditions

OPTIMAL GROWTH WITH SCALE ECONOMIES

557

in progress, denoted X. Only after a while should some generalized inventory be transformed into new K; available for operation.'
Let AX represent a portion of generalized inventory earmarked for conversion into
operating ,Bcapital. Naturally 0 _ AX < X. Suppose that AK, units of K, are created,
where
AK, = G(AX).
Reflecting economies of scale, G is taken to be a convex, monotonically increasing, continuous function of AX defined for all AX > 0. It is assumed that lim G(AX) = so, G(O) = 0,
Ax

and
lim G(AX) =

AX-O+

AX

Xoo

0.

... (1)

Something like the latter condition is necessaryto ensure that economies of scale are taken
advantage of and that in fact generalizedinventories must be accumulated for this purpose.2
We will also find it useful to work with the function H, defined as the inverse of G.
H can be interpreted as an investment cost function relating the cost in cumulated output
units of a given, capital increase according to the schedule

AX = H(AKp) = G-1(AKp).
Displaying decreasing unit costs, the continuous, monotonically increasing, concave cost
function H is defined for all AKp > 0 and possesses the properties lim H(AK.) = co,
H(O) = 0, and

lim

H(AK

H(AK)

AK-co
K

AKp
Before turning to the main problem, we digress in the next two sections to consider a
pair of related problems whose solution will prove useful in characterizing an optimal path
for the general case.
AKP-O+

4. OPTIMAL GROWTH IN A MACROECONOMIC MODEL

The social utility of consuming amount C(t) at time t is taken to be U(C(t)). The
instantaneous utility function U is monotonic increasing, concave and differentiable. For
simplification the condition
lim U'(C)= oo
is imposed, guaranteeing non-zero consumption for all time. Finally, it is necessary to
make a boundedness qualification of the form
sup U(F(K,)) = B < oo.

KC 2 0

Ramsey called the least upper bound B the bliss level. A state of bliss would be attained
(in the limit) as consumers became sated with goods or as the effects of capital saturation
were so pronounced as to make it impossible to increase production past a certain output
no matter how much investment were undertaken.
1 Thisreasoningis easy to spell out by a simpleexample. If a certainamountof materialcan be moulded
into a pipelineto carrya givenvolume per unit time, twice as much materialwould resultin a pipelineable
to transportfour timesthe previousvolume(assuminga giventhicknessof pipe). Under thesecircumstances
it may be betterto wait a whilefor a biggerpipelineeven though largerinventorystockswould be standing
idly by in the interim.
2 Generalizingto the case where(1) does not hold is not difficult. Continuousadjustmentsare then
possible, as well as discretejumps.

This content downloaded from 18.9.62.46 on Tue, 29 Dec 2015 07:34:37 UTC
All use subject to JSTOR Terms and Conditions

558

REVIEW OF ECONOMIC STUDIES

For a given consumption path {C(t)}, Ramsey defined his social welfare criterion
V[{C(t)}] as a sum of the difference between instantaneous utility and the bliss level:
V[{C(t)}]-

[U(C(t))-B]dt.
0

There is no a priori reason why this evaluation integral ought to be finite for any given
{C(t)}. Should V be equal to - oo for each of two consumption paths, there would be
nothing to recommend one over the other; however a path yielding a finite value of V
would be preferable to both.1
Temporarily forgetting all about ,Bcapital, Ramsey's problem is to
max

[U(C)-B]dt

... (2)

subject to
C+I =F(K),

... (3)

K =I

... (4)

C, I

O,
0,.(5)

>0, given.
K(O) = Ka(O)

... (6)

Any path {K(t), C(t)} satisfying (3)-(6) is called feasible.

It is presumed that U and F are such that (2)-(6) is a meaningful problem in the sense
that an optimum exists.2 An optimal path {K(t), 2(t)} must be feasible, satisfy
-soo < V--=V[{C(t)}]
and possess the property that for any feasible path {K(t), C(t)},
V[{C(t)}] ? V.
Using the calculus of variations, Ramsey was able to characterize3 an optimal path
{K(t), C(t)} as the unique solution to the differential equations
B- U(C),

UI(C)K=
C

... (7)
.(8)

F(K=

with the given initial condition

K(O) = Ka().

..

(9)

Equation (7) is the famous Keynes-Ramsey rule of optimal allocation. Along an

optimal path both capital and consumption grow monotonically until one or the other goes
to or asymptotically approaches its saturation level. The other variable goes to or asymptotically approaches a corresponding level which is determined from the production function. Thus the bliss level is reached at least asymptotically. An optimal solution is shown
in Fig. 1.
The overtakingcriterionmakes {C1(t)}preferableto {C2(t)}if there exists a T such that
U(C1(t))dt>

U(C2(t))dt

for all T > T. Were{C1(t)}preferableto {C2(t)}by Ramsey'sevaluationintegralit would also be preferable

by the overtakingcriterion. The converseis not necessarilytrue.
2 It is easy to see that (2)-(6) would have to be a well-definedproblemif U(F(K.))reacheda maximum
for finite K,. The various sufficiencyconditions of Gale and Sutherland[3], McFadden [8] and von
Weizsacker[14] are stated in termsof the overtakingcriterionbut many of them can be routinelymodified
to treat the Ramsey evaluationintegralcase.
3 A rigorousargumentis containedin Koopmans [7].

This content downloaded from 18.9.62.46 on Tue, 29 Dec 2015 07:34:37 UTC
All use subject to JSTOR Terms and Conditions

OPTIMAL GROWTH WITH SCALE ECONOMIES

559

Y()_F(K(t)

.
(10)

Define
... (11)

q(t) --U'(C(t)),
I()_K(t

....(12)
The dual price q(t) is interpretable as the value of an extra unit of output at time t
imputed in terms of the evaluation integral. Differentiating (7) and (8) with respect to
time, substituting from (11) and rearrangingyields
- F(K),
q =q
a standard relation of optimal growth theory. Note that q<0, Y_I' 0.
-

Y~(t)

Time t
FIGURE 1

5. OPTIMAL CAPACITY EXPANSION

Now temporarily neglecting occapital entirely, treat { Y(t)} and {q(t)} as if they were
prescribed data. { Yf(t)} is considered in the present context to be a fixed final demand
schedule which must always be fulfilled. Thinking of q(t) as the cost in current terms of
investment funds at time t, the present discounted cost of creating extra capacity AK at
time t is q(t)H(AK).
The least cost capacity expansion problem is to schedule capacity {K(t)} to meet final
demands { Y(t)} at minimum total present discounted cost. Mathematically, the problem
is to find times t1, t2, ... and capacity increments AK(t1), AK(t2), ... which'
00

min

f[{tj}, {AK(ti)}] _

i= 1

q(tj)H(AK(t1))

...(13)

1 At times other than ti, t2, ..., AK(t) is thought of as being zero. Manne [9] discussesseveralpractical
examplesof optimal capacityexpansionproblems.

This content downloaded from 18.9.62.46 on Tue, 29 Dec 2015 07:34:37 UTC
All use subject to JSTOR Terms and Conditions

560

REVIEW OF ECONOMIC STUDIES

subject to
K(t)_ Y(t),..(4
=
K(t) K(ti- 1)+ AK(ti - 1) for ti_- < t _ ti,
K(O) = Kp(O)> Y(O),given,
to - 0.
Let {?i,A1Q(?X)}
represent an optimal policy.' Define ;(15), an optimal path can alternatively be expressed by {k(t)}.

. ..(15)
.. .(16)

[{Ji,

{Ak(?i)}].

Using

Demand,
capacity

Capacityk,(t)
|
-kP(3)

IL/

Demand ?(t)

kp(72) -

K(O)= fl(7j)

Y(0

r
.I
t,

FIGURE

I,

t2

t3

Time t

1 Although the existence of an optimal policy is not proved, we indicate why it has been assumed.
Considera feasiblepolicy of installing8 units of extra capacitywheneverK Y. The cost of such a policy
would be
a(S)--

H()

q(i'1(K(O)+j8))8.

Passing to the limit,

lim

E q(5-1(K(O)+jS))8=

a6-O + j = O

F'(K)[B-

U(C)]dt < F'((O))

q(5-1(F))df=
K(O)

q(t)Ydt=
tj

ftj q(t)F'(K)Kdt

(
U(C)]dt < oo, where ti _ Yl(K(O)).
[B-

() is well defined for

all S>O. It follows that for some value S'>0 and some M< co, 0(8') = M. M provides,a finite upper
bound on I. Since zero is a lower bound, a finitegreatestlowerbound on 0 exists, denoted0. This means
that feasiblepolicies exist with costs arbitrarilyclose to ?,. In practicesuch a resultis as good as an existence theorem,given the uncertaintiesof the data.
Strict existenceof an optimal capacityschedulein the mathematicalsense could be rigorouslyproved
if we had chosen to formulatethe model in terms of period analysisinstead of continuoustime. Let Q be

This content downloaded from 18.9.62.46 on Tue, 29 Dec 2015 07:34:37 UTC
All use subject to JSTOR Terms and Conditions

OPTIMAL GROWTH WITH SCALE ECONOMIES

561

It is easy to see that along an optimal path, R(fi) = if(fi). No extra capacity will be
installed while some excess capacity already exists. With q < 0, it pays to postpone intended
construction until the day when some must be undertaken because full capacity will have
been reached.' A typical minimum cost policy is illustrated in Fig. 2.
6. FORMULATION AND SOLUTION OF THE BASIC MODEL
Having treated separately the Ramsey and Capacity Expansion problems, we are in a
position to tackle the main problem introduced in Section 3. Properties of the present
model are vaguely recognizable as some sort of a rough combination of features belonging
to the two simpler problems. As we shall see, the optimal solution will combine, in a well
defined sense, the Ramsey and capacity expansion optimal trajectories.
We use the same objective as Ramsey. However, in the context in which it is presently
employed,

[U(C)-B]dt

is denoted W[{C(t)}] to avoid confusion.

The problem is to select times t t, ... and to choose values for the instruments
n SK(1i) to2

Y(t) C(t)s I(t) Ia(t), IXt,X(

max

W[{C(t)}]

...(17)

[U(C(t))-B]dt
0

subject to

_
Y(t) F(Ka(t))9..(8
Y(t) _ Kf(t)
C(t) + I(t)

. . . (19)

Y(t),

...(20)

. ..(21)

Ia(t)+IX(t) =I(t)
Ka(t) = Ia(t),
X(t) = IX(t) for t #0 li,
lim X(li + e) = X(i) - AX(is),

C-0

. .. (22)
... (23)
.. .(24)

AKpQi)= G(AX(1i)),
1) for li- 1< t ? li,
Kp(t) = Kp(li- 1)+ &Kp(1i_
Y(t)

C(t)

I(t)

Ia(09IX(09

X(t)

AX(ii)>

... (25)
.. .(26)

...(27)

09

... (28)
Ka(O)q
X(O)given.
Kpy(O)q
Where not otherwise noted, t is any non-negative time and i is any positive integer. By
convention o ?0.
the Cartesianproduct set of all feasible x /AK(n). ClearlyQ is closed. With M an upper bound on ,
n=O
<
we can restrictAK(n) to values 0 AK _GC (
for n = O,1, 2,.... A standard application of the
Tychnoff theorem (cf. Kelley [6], p. 143) shows the Cartesianproduct set Q' = x
compact. It follows that the set Q" _ QnQ' is compact. Define the function 0
00

to be
G
G,
min {0, M) for all

x AK(n)E Q". Being real valued and continuousunder the producttopology on the compact set Q", 0
n=O
must attain a minimumfor some value x AK(n)E Q",concludingthe proof.
n=O
1 This simple conclusion is an exampleof a " regenerationpoint theorem". Such a result greatly
simplifiescomputationbecausethe searchfor an optimumcan be limitedto full capacityregenerationpoints
and these can typicallybe efficientlyexaminedvia the appropriatedynamicprogrammingalgorithm.
2 For economy of notation, positive values of AX and positive changes in Ka have been restricted
a priori to discrete times {f1}. In fact this is a vacuous restrictionbecause condition(1) implies that even
with the possibilityof continuous adjustmentsavailable, an optimal policy would always call for jump
adjustmentsin Ka at certain distinct times. For times t # li, it is useful to interpretAX(t) and AK0(t)as
being zero. Note that AX(fi) is definedas minusthe algebraicchange in X at time li, so that AX 2 0.
2N

This content downloaded from 18.9.62.46 on Tue, 29 Dec 2015 07:34:37 UTC
All use subject to JSTOR Terms and Conditions

562

REVIEW OF ECONOMIC STUDIES

With only insignificant loss of generality, we restrict attention to initial configurations

of the state variables Ka, K;, X obeying'

Kp(O)_ F(Ka(?))s
X(O) = 0.

The following theorem is the main result.

Define {1} and {!*} by the recursiveequations2
i-1
=

j= I

t-

*+ H(Ak(li))

i = 1,2 ....
Under the assumptions of the model, an optimal solution of (17)-(28), whose variables
are denoted with asterisks, can be describedas follows:
AKp*(t)= AX*(t) = 0, t =A7,
AK*(ts*) =

Kt)

AX*(!*) =H(Ak(ii))
K*(t) = K t-

(J*-t*)

K*(t) = 0(I),

C*(t) =

c (t

5 (tj*-tY))

X*(t) =

Y*(t) =

i.

(n

C*(t)= C )

Fortunately a relatively simple interpretation can be placed on this formidable looking

prescription.
The optimal policy is depicted in Fig. 3. Suppose that all the [t*, i*] sections of that
diagram were compressed into points and the gaps removed by pushing together and
1 Laterit shouldbecomeclearwhatto do if theseinitialconditionsshouldnot be met. In fact an
optimalpolicy will alwayscall for K0(t) ? F(Kit))-for all t ? 0 if K0(O)_ F(K~(O)),and for all sufficiently
larget if K0(O)<F(Ka(O)).
Thelatterconditionmusttherefore
havearisen,in somesense,outof mismanagement priorto time zero. The case of X(O)positive is not difficultto handle; we normalizeto zerojustfor
notational convenience.
2 The reader is remindedthat variablescapped with a circumflexare defined in Section 5, whereas
those capped with a tilde are definedin Section 4.

This content downloaded from 18.9.62.46 on Tue, 29 Dec 2015 07:34:37 UTC
All use subject to JSTOR Terms and Conditions

OPTIMAL GROWTH WITH SCALE ECONOMIES

563

connecting the {Y*(t)}, {I*(t)} and {C*(t)} curves. The resulting trajectories would be
exactly the {
{1(t)}, {C(t)} graphs of Fig. 1. In other words, the sole effect of introducing the [t*, t] regions into Fig. 3 is to stretch out into intervals what would otherwise
be isolated single points of the original Ramsey optimal trajectory.
An analogous interpretation is available for fl capital. Except for the [tn, i*] sections,
the time profile of Ki in Fig. 3 would look identical to the Fig. 2 portrayal of an optimal
capacity schedule {?i, AR (?I)}. It is easily seen that the {Ii} of Fig. 2 are exactly those
isolated points of the Ramsey optimal policy which in Fig. 3 are stretched out into time
segments of positive length.
While time seems to stand still during the period [t, t*], all investment is going into
accumulating an inventory of X(t!) starting from an initial level of zero at time t*. The
,

C*, K*

C*(t)

Y*(t)

--)Time t
t* -i-1

-* -- l
ti1

i*t
i'

*
FIGURE

-L+1

till

interval [I, ti] lasts precisely long enough to build up that amount of X which will coalesce
to form the AR( i) units of P capital dictated by the solution to the capacity expansion
problem. Throughout the period [t*, !i], Y*(t), I*(t), C*(t) are maintained at the constant
levels Yi'Q),1Qi), C(Q). At time ti the entire generalized inventory X(t*) is formed into fi
capital and the economy picks up again at that point of the Ramsey trajectory where it left
off at time t* because the Kl ceiling had become binding.
In terms of economic development, [t*, t*] represents a big push period.' During
this time all investment is being funneled at a constant rate into the as yet unproductive
overhead capital project. As will presently be demonstrated, big push stages are likely to
be more predominant in the earlier stages of development. From the viewpoint of social
policy, the big push is probably a critical time because no real growth occurs and consumption is stagnant.
1 Cf. Rosenstein-Rodan[12], Hirschman[5], Scitovsky (13].

This content downloaded from 18.9.62.46 on Tue, 29 Dec 2015 07:34:37 UTC
All use subject to JSTOR Terms and Conditions

REVIEW OF ECONOMIC STUDIES

564

Let p(t) represent the value of an extra unit of output at time t imputed in terms of the
evaluation integral. Obviously p(t) = U'(C*(t)), so that

p(t)= q(t-

t)

1_ <t<t*.
t,*< t _ t*, i = 1, 2, ....

p(t) = q(ii),

For t belonging to the Ramsey growth phase (71, t, the dual price declines over
time; an extra unit of output is worth more if it is received early because it could be productively invested in a capital to yield increased future returns. However, in the big push
phase [t,, it] the social output price is stationary; whether received early or late in a big
push stage an extra unit of output cannot be used to increase returns but could only be
invested in non-directly-productivegeneralized inventory.

7. PROOF OF THE MAIN THEOREM

The strategy of proof can be easily outlined. A consumption path {Ce(t)} is efficient
in the usual sense if it is feasible and if, for any other feasible path {Cf(t)} with the property

Cf(t) ? Ce(t) for all t > 0,

00

[Cf(t)- Ce(t)]dt= 0.

We frst exhibit three obvious

efficiency conditions (a)-(c). Any candidate for an optimal path must satisfy these three
criteria. Next, considering only comparisons among paths so restricted, we show that an
optimal solution must belong to an even more exclusive family of paths with special additional features. Finally, the proposed solution is shown to be an optimal member of this
family of special paths.
The three efficiency conditions are:
(a) If F(K.(t)) < Kp(t), no investmentin X occurs at time t.
This condition specifies that all investment must go into building a capital if there is
excess ,Bcapacity. Let T be the first time later than t when F(K,-(T)) = Kp(t) (if this never
happens, T_ oo). With F(Ka(t'))<Kp(t'), there is no loss of generality in restricting
AX(t') to be zero for t ? t'<T. It certainly won't depress consumption at any future or
present time to put off until z setting AX> 0. It may even help to wait and take advantage
of increasing returns by later converting a bigger chunk of AX into more AKP. Now
X(t')dt' = y >0.

suppose contrary to the hypothesis that f

Since nothing useful is going

Jt

to get done with X until time T at the earliest, we can consider a differentpolicy. Maintain
the same consumption levels as in the original policy. Now, however, first devote all investment to building up Ka to level K{(T)= F- 1(Kf(t)) and then put all investment into creating
amount y of X. Let this alternative policy take (T' - t) time units to complete. It is easy
to see that ' <T because the second alternative takes advantage of the direct productivity
of a capital for all times when F(Ka)< K,. Thus, the identical stocks of Ka, K., and X are
achieved at time T and consumption is the same for all times belonging to [t, T'], but the
second alternative permits attainment of a strictly higher amount of consumption at level
Kp(t) = F(K(-(T))for all times of (T', T]. It follows that the original path must have been
inefficient.
(b) If F(Kx(t)) _ Kp(t), no investmentin K. occurs at time t.
This condition prohibits investment in a capital so long as full capacity already exists
in that sector. Let T be the first time after t when AK; >0. Suppose that F(K,(t')) > K,,(t')

for I

(t, T]

and that {

K,,(t')dt' = y>0. It is easy to see that a superiorpolicy is firstto

This content downloaded from 18.9.62.46 on Tue, 29 Dec 2015 07:34:37 UTC
All use subject to JSTOR Terms and Conditions

OPTIMAL GROWTH WITH SCALE ECONOMIES

565

invest only in X and then to coalesce H(AKp(T))units of X into AK,(T) as soon as X(T) - X(t)
units of X have been accumulated, say at time '< T. Only then should the y units of a
capital be built up. All the while the same consumption levels of the original policy are
maintained. This alternative ends up with the same values of Ka, K,, and X immediately
after time z, maintains an identical consumption level for all times belonging to [t, T'], and
still allows a splash of extra consumption to occur in the interval (z', T] due to taking
advantage of the direct productivity of a capital.
(c) If F(Ka(t)) > K,,(t), if X(t) = 0, and if z is thefirst time after t when AX> 0, then
AX(T) = X(T).

This condition requires that all X built up from zero for the purpose of increasing K,
must be coalesced into AK,, all at once. Suppose to the contrary that AX(T) = y < X(T).
r'

Let T' < be such that J

(t')dt' = y. If F(Kx(t)) = K1(t), it will be better to set AK(T')

and then, duplicating the reasoning of (a), to restrict further immediate investments to a
capital. The advantage of setting AK(-c')= y for the case F(K,(t)) > K,(t) is obvious. In
either event, the original path is inefficient.
The following comments apply to efficient paths.
It follows from (b) and the initial condition F(K,(O)) < Kp(O)that F(K,(t)) < Kp(t)
for all t > 0. Let {Iij represent the times (in order) at which AX>0; that is, AX(t)>O if
and only if t = ti for some positive integer i. Immediately after 1i-1, F(Ka)<Ks, and all
investment goes into a capital. Let ti be the first time after ti- I when F(K.) = K.. Define
00

U (1i-1, ti), with to

0. If t E R, F(Ka(t))<K,(t), and, from (a), I(t) = I"(t). Let

i= 1

and (b) stipulates thatI(t') = IX(t'). Thus

U [, tiJ. For t'eS, F(K,(t'))=Kp(t'),
any efficientpath can be divided into two distinct investment specialization phases. Property
(c) combined with the initial condition X(0) = 0 implies that AX(b) = X(71), or that
X(t) = 0 for t E R.
Let qi be the value of an optimal policy starting at time ti. Note that qi is a function
of KQ(ti),KQ(ti)and X(ti) but does not depend on t, explicitly because the objective function
and the constraints of the main problem are time autonomous. We know that all investment during the period (ti_1, ti) is restricted to a capital alone. Consider ti-1, Ka(t.i_1)
K,(Qi),as fixed. To be part of an optimal path, ti, treated as a variable, must be chosen to
S-

i = 1

max

I'ti
[U(C)-B]dt+

... (29)

subject to
=

C+

ii-

F(K)(30)

K=-Ia

... (31)

C,I> _ 0,

... (32)

1, KJ!i - 1), Kx(tij)given.

.. (33)

Using the calculus of variations, the solution to this free-time fixed-endpoint problem
is

U'(C)Kc
c

+ Ka

B- U(C).

... (34)

F(Kx),

... (35)

i - 1 KJ! i - ), Kx(tj) given.

... .(36)

his, of course, is nothing but a section of the Ramsey optimal trajectory (7)-(9).
Now consider an analogous problem over the time interval [ti, 7]. Let Obbe the value

This content downloaded from 18.9.62.46 on Tue, 29 Dec 2015 07:34:37 UTC
All use subject to JSTOR Terms and Conditions

566

REVIEW OF ECONOMIC STUDIES

of an optimal policy starting at time biwith capital stocks Ka(Ii), K,(ii), X(Ii). All investment during the period [ti, i] goes into building up an inventory of X(Ii) starting from
= Y(Ii) is constant for t e [ti, Ji].
zero. Also Y(t) = Y(t) = F(K=(ti))=Kp(ti)
Look upon Y(t), ti, XQ() and X(i) as constants. An optimal policy must have the
property that bi,treated here as a variable, is selected to
max

rZi

[U(C)- B]dt+0k

...

(37)

subject to
...(38)

Y(i),

C+IX=

X-=I~X,
C, >x 00,

... (39)

ti, X(tQ)= 0, X(Ii), Y(tQ)given.

... (41)

... (40)

The calculus of variations solution to this free-time fixed-endpoint problem is

U'(C)X = B- U(C),

C+X =

... (42)
(43)
...(44)

Y(i)-F(Ka()))=..

ti, X(ti) = 0, X(QI),Y(tQ)given.

Optimal values of IX(t) and C(t) are constants for t E [ti, ti]. They are equal to the
optimal values, respectively, of I,(t) and C(t) in the solution of problem (29)-(33) at t = ti,
since at that time equations (34)-(36) and (42)-(44) are identical. Using the same reasoning,
IX(t)and C(t) from (42)-(44) are also equal to the optimal values of I,(t) and C(t) at time
t = ii for the free-time fixed-endpoint problem which is identical to (29)-(33) except for
taking place over the interval (ii, ti+1) instead of (i- 1, ti).
These results justify the basic features of an optimal policy as they have been depicted
in Fig. 3. Y*(t), I*(t), and C*(t) are continuous. Each big push stage of S is just a single
Ramsey optimal point stretched out into an interval of positive length. If these isolated
points were reassembled back into the appropriate niches of R and the resulting conglomeration treated as if it were a set connected under continuous time, the complete Ramsey
trajectory (7)-(9) would emerge. A Ramsey growth stage of R ends whenever full fi capacity
is encountered and it becomes necessary to devote investment to building generalized
inventories. Ramsey growth continues as soon as all inventories have been converted into
excess, capacity.
In our search for an optimal policy, without loss of generality further attention is
restricted to the family of paths described above. The social objective W[{C(t)}] can then
be split up as follows:
W[{C(t)}]=

[U(C)-B]dt=
O

[U(C)-B]dt.

[U(C)-B]dt+
JR

JS

As we have seen, the value of the former integral along an optimal path must be RI
the optimal Ramsey objective. As for the latter integral,
[U(C)-B]dt

[U(C(i))-B](i-ti)
~~~~i=l1

= -

H
[-HUA(C(Q))I(O)1

(-t))

Ej U'(C(ti))H(AKf(li)).

This content downloaded from 18.9.62.46 on Tue, 29 Dec 2015 07:34:37 UTC
All use subject to JSTOR Terms and Conditions

...

(45)

OPTIMAL GROWTH WITH SCALE ECONOMIES

567

i-i

Think of ti = t i-

E (tj-t!) as the single Ramsey optimal point which has been

j=1
magnified into the interval [ti, 1J. In terms of the capacity expansion problem (13)-(16),
(45) is equal to
00

Z_

i=

00

E q(ti)H(AK(ti))

U'(c(t))H(AK(ti))=-

i=l
=

where ti+1
Thus

-+[{ti},

{AK(ti)}],

[U(C)-B]dt

will be maximized whenever

is minimized subject to the

constraints (14)-(16). We can translate from {1i} to {tfl and to {t*} by using the relations
=

i+

j= 1

(t -t ) and t*-t*

__/i___)

0i

This concludes the proof.' An interesting side result is that

W* =V-.

The maximum social objective is less than the optimal value of the Ramsey problem by the
social cost of implementing the cheapest feasible capacity schedule.
8. CAPACITY EXPANSION AT A CONSTANT GEOMETRIC RATE
The features of an optimal capacity schedule can be particularized by restricting the
general functions H(AK), Y(t), and q(t) to specific parameterizations. Here we consider
a constant elasticity investment cost function of the form2
H(AK) = A(AK)a,

... (46)

where A is just a positive constant of proportionality. The exponent a measures the (constant) ratio of incremental to average costs of installed capacity. Consistent with the
presumed existence of economies of scale,
O<a< 1.
We also assume that

Yt=

q(t)

..(7

for some constants 1 and s obeying

0<1,
O<s<

1 Note the criticaldependenceof our resultson the time autonomyof the systemunder consideration.

In the original Ramsey problem the introductionof discounting,population growth,and time dependent
depreciationor technicalchange does verylittle to change the basic propertiesof a solution (so long as the
time dependenceis exponential!). This is not the case for the presentmodel. If time dependentfeatures
were introduced,the qualitativepropertiesof a welfaremaximizingpath would be roughly similar, but
there would be no possibilityof cleanlydecomposingthe optimaltrajectoryinto two distinctsub-problems.
On the other hand, incorporatinginto the model such time autonomousfeaturesas Arrow's " leaming by
doing " [I] would not significantlychange the natureof an optimal trajectory. It is not difficultto extend
the model to handlethe case of two or more overheadcapitalgoodswhichmust operatein fixed proportions.
Introducinga positiveelasticityof substitutionbetweenKxand Kawouldconsiderablycomplicatethe picture,
although some featuresof an optimal policy would remainsimilar.
2 This parameterizationis popular as a pre-design approximationto investment costs of installed
capacityfor such continuousprocess industriesas chemicals,petroleum,cement,electricitygeneration,and
primarymetals. It should be emphasizedthat in the cost engineeringliteratureAXK
would referto the extra
capacitycreatedfrom the establishmentof a typical completeprocessbalancedplant and not to the creation
of semi-mythical" social overheadcapital " for the economy as a whole. Cost engineersusuallytreat the
exponenta as being equal to about two-thirds. Cf. Chilton [2], Haldi and Whitcomb[4], and Moore [10].

This content downloaded from 18.9.62.46 on Tue, 29 Dec 2015 07:34:37 UTC
All use subject to JSTOR Terms and Conditions

568

REVIEW OF ECONOMIC STUDIES

An equivalent way of writing (47) is

g(t) = sr(t),
...(48)
where g(t) _
5Yis the growth rate of final demand and r(t)
-q/q is the discount rate,
both evaluated at time t.
In the framework of the isolated capacity expansion model of section 5, expression
(48) can be looked upon purely as a convenient parameterization of the underlying data.
The parameterization (48) generalizes the case sometimes analyzed where fixed demand
grows exponentially and the discount rate is constant.
However, within the context of the two-sector model we have been analyzing, relation
(48) has an interesting additional interpretation. It could come about as a result of society
choosing to save at the constant rate s:
-

g(t) =

t = F'(K)K
y
y

s
sr(t)
~~~~~q

sF'(K)

In turn, a proportional savings function is the optimal behaviour befitting any solvable
Ramsey problem with a constant elasticity of marginal utility. Such a utility function is
of the form
U(C) = B-DC'-,
Ul,
C > 1 are given constants.
U,
With the specific parameterization (46), (47) the minimum cost capacity schedule has
a particularly simple characterization. When capacity must be increased (because no more
slack exists), it is always incrementedby the same constantpercentage of existing capacity.'
We prove this interesting result by considering a schedule {ti, AK(tj)} which is a candidate for minimizing present discounted cost f. Without loss of generality it can be presumed that K(ti) = Y(ti), so that no extra capacity is installed while excess capacity is
already in place.
where B > 0, D > 0 and 7 =-

*=

Z q(tj)H(AK(tj))

i= 1
00

i-1

Zi l[K(O)+
i=l

AK(tj)]-

s A[AK(ti)]a

j=

=1AK(O)'a-;
j

1+i3j = i AKt)~FA(~1
]
K(O) L K(O)

...(49)

Because 0 can be written in the form (49), it is apparent that the cost minimizing values
of AK(tj)/K(0) are independent of K(0). Now consider the problem of finding a least cost
capacity schedule which begins at time tn with capacity K(tn) instead of at time 0 with
capacity K(0). This is a sub-problem of the original. The cost function for the new problem
can be written in a form identical to (49) except for obvious index renumbering and the
interchanged roles of K(tj) and K(0). But the optimal incremental capacity sequence
expressed in units of initial capacity is independent of the initial capacity level. Hence,
for an optimal path {1j, AK(?j)},
AkRt) _ Ak(Q1)
K(0)
KQi)

1, 2.

1 Note that if g(t) (and hencer(t) also) is constant, an optimal policy would call for schedulingextra
capacity increments at equally spaced time intervals of length l In I +
g
w pK(Oe

I. This constant cycle time

result was provedby Srinivasinin Manne [9] for the special case mentionedabove.

This content downloaded from 18.9.62.46 on Tue, 29 Dec 2015 07:34:37 UTC
All use subject to JSTOR Terms and Conditions

OPTIMAL GROWTH WITH SCALE ECONOMIES

569

Let
_AQ)

i_=

1,2, ...,

be the constant fraction by which capacity is always incremented. It is not difficult to

demonstrate that 8y48a<0 and 04u/as>0.
9. CONCLUDING REMARKS
The time spent in a typical big push period will be H(AKfl)/i. A Ramsey growth phase
will last approximately AK/Ytime units.' The fraction of time spent in big push stages
will be approximately
+K
H(AKp)

AKFK

H(AK O)
+1

F(
AKP_
H(AKP)+1
I
F'Kzk
AK#
The earlier the stage of development, the higher the anticipated values of F'(KD)and
H(AKp)/AK.2 Thus, it is to be expected that the percentage of time spent in big push
stages should decline over time.
This quantifies the generally accepted notion that infrastructure is somehow a much
more important ingredient in the growth of an underdeveloped than of a mature economy.
The increased significance of big push stages during the early years of development means
more time spent in no-growth stagnant consumption phases awaiting the completion of
overhead facilities. Of course the present model over-emphasizes certain structural rigidities, but the conclusions accord well with the customary feeling that the creation of social
overhead capital is a more formidable barrier to growth in a less developed economy.
H(AIK) +

Cowles Foundation
Yale University

~Kpi H(AKO)
Y
1~I

M. L.

WEITZMAN

First version received May 1969; final version received December 1969
REFERENCES
[1]

Arrow, K. J. "The Economic Implications of Learning by Doing ", Review of

Economic Studies (June 1962).

[2]

Chilton, C. H. (ed.). Cost Engineeringin the Process Industries(New York, McGrawHill, 1960).

[3]

Gale, D. and Sutherland, W. R. "Analysis of a One Good Model of Economic

Development ", in Dantzig and Veinott (eds.) Mathematics of the Decision Sciences,
Part 2 (Providence, American Mathematical Society, 1968).

1 This is just a first approximationwhich will be increasinglyaccurateas Y is close to being constant

over the Ramsey growthphase. We are comparinga big push period (correspondingto a single Ramsey
time point) with the growth phase which directly precedesor follows it. Both I and Y are evaluated at
this single Ramsey time point.

It has alreadybeen noted that d- F'(K) = q < 0. Under a varietyof conditionsit can be provedthat
will declinewith time (due to takinggreateradvantageof economies
the averagecost of capacityH(AKft)/AKft
of scale in later periods). For example, the condition Y >0 would be sufficientto demonstratethat AKa
increasesover time. Note that in Section 8 H(AK0)/1AK,ldeclines over time irrespectiveof the sign of ?.
2

This content downloaded from 18.9.62.46 on Tue, 29 Dec 2015 07:34:37 UTC
All use subject to JSTOR Terms and Conditions

570

REVIEW OF ECONOMIC STUDIES

[4]

Haldi, J. and Whitcomb, D. " Economies of Scale in Industrial Plants ", The
Journal of Political Economy (August 1967), 373-385.

[5]

Hirschman, A. 0. The Strategy of Economic Development (New Haven, Yale

University Press, 1958).

[6]

Kelley, J. M. GeneralTopology (New York, Van Nostrand, 1955).

[7]

Koopmans, T. C. " On the Concept of Optimal Economic Growth ", Pontificae

Academiae ScientiarumScripta Varia (1965), 225-300.

[8]

McFadden, D. " The Evaluation of Development Programmes ", The Review of

Economic Studies, January 1967, 25-50.
Manne, A. S. Investmentsfor Capacity Expansion (Cambridge, M.I.T. Press, 1967).

[9]
[10]

Moore, F. T. " Economies of Scale: Some Statistical Evidence ", QuarterlyJournal

of Economics (May 1959), 232-245.

[11]

Ramsey, F. P. " A Mathematical Theory of Savings ", EconomicJournal(December

1928), 219-226.

[12]

Rosenstein-Rodan, P. N. "Notes on the Theory of the 'Big Push' ", Ch. 3 in

H. S. Ellis (ed.) Economic Development for Latin America (I.E.A. Conference)
(London, Macmillan, 1961).

[13]

Scitovsky, T. " Growth-Balanced or Unbalanced? " in Abramovitz (ed.) The

Allocation of Economic Resources (Stanford, Stanford University Press, 1959), 207217.

[14]

von Weizsacker, C. C. "Existence of Optimal Programs of Accumulation for an

Infinite Time Horizon ", Review of Economic Studies (April 1965), 85-104.

This content downloaded from 18.9.62.46 on Tue, 29 Dec 2015 07:34:37 UTC
All use subject to JSTOR Terms and Conditions