You are on page 1of 9

American Economic Association

A Diagrammatic Exposition of Optimal Growth Author(s): Nissan Liviatan Reviewed work(s): Source: The American Economic Review, Vol. 60, No. 3 (Jun., 1970), pp. 302-309 Published by: American Economic Association Stable URL: http://www.jstor.org/stable/1817980 . Accessed: 01/11/2011 15:07
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.

American Economic Association is collaborating with JSTOR to digitize, preserve and extend access to The American Economic Review.

http://www.jstor.org

DiagrammaticExposition
Optimal
By

of

Growth
NISSAN LIVIATAN*

The theory of optimal growth can be analyzed, as is well known, by using optimal-control methods or by applying the technique of dynamic programming. It seems, however, that the graphical expositions for simple growth models have been developed only for the former approach. Thus it is customary to present the optimal growth path in terms of a trajectory in the phase-space.' However, as far as I know, there has been no attempt to formulate a graphical analysis which will bring out the essence of the dynamic programming approach2 to optimal growth. The purpose of this paper is to fill this methodological gap. Moreover, it will be seen that the graphical presentation of the dynamic programming approach is closely related to the traditional Fisherine diagram of intertemporal analysis and to the standard Hicksian tools of demand analysis. Indeed, we intend to show that the exposition of the optimal time path in a simple growth model does not require much more than a dynamic version of the well-known two-period Fisherine diagram, with current consurmLption one axis and on next year's capital on the other. I. A Restatement of Some Elementary Results In our analysis, which is based on a discrete time model, we shall deal with a
* The author is professor of economics at the Hebrew

one-sector model with no technological progress. The production function of the economy is assumed to be of the following form: (1) St+, + Ct+1 = F(Sty LJ)

where St denotes the stock of capital which enters as an input in the production function in period t, Lt is labor input in period t, and Ct+1 is consumption in period t+ 1. (We may assume that (1) incorporates an exponential depreciation factor.) An alternative way of writing (1), which is more useful for our purposes, is as follows. Define K, as the capital stock available (for production or consumption) in the beginning of period t. We then have (2) Kt = St + Ct

Substituting (2) into (1) we obtain (3) Kt+1 = F(Kt


-

Ct, Lt)

It should be noted that F1=aF/a(Kt-Ct) is the marginal product of capital in period t in producing the capital stock of period t+ 1. Hence (F1- 1) is the net own rate of return on capital, which we shall assume to be nonnegative. We shall assume as usual that F is subject to constant returns to scale with respect to the inputs (Kt- Ct) and Lt and that Lt grows according to (4) Lt+1=nLt n > 1, for all t

University of Jerusalem and visiting professor at University of California, Berkeley. This study was done during the tenure of a Ford Foundation Fellowship. I See, for example, the presentation in David Cass (pp. 236-38) which is based on control theory, or the

graphical exposition by Tjalling C. Koopmans based on an equivalent calculus-of-variations formulation. 2 For an analytic exposition of optimal gowth based on dynamic programming see, for example, Roy Radner.

302

LIVIATAN: OPTIMAL GROWTH where n is given exogenously. Using the property of constant returns to scale, we may divide all variablesin (3) by Lt, which yields
Lt Lt Lt
k.,

303

Dividing both sides of (5) by n and using (4), we have


(6) K Lt+l -F n

(---,1)
Lt

Lt

AA

B I t

Using lower case letters to denote per capita variables, we may write (6) as 1
(7) kt+l =-F(kt n
-ct,

1).
FIGURE 1

As is usual in growth theory, we make following assumptions about the


F(kt-ct, (8) 1): F(0, 1) = 0

(9) F, > 1,
(10) F1(0, 1) =

Fl, < O for all kt -ct > O


oo, Fi(oo,

1) =

where Fn denotes a2F/a(Kt- Ct)2. Since n is a constant throughout our analysis, it is convenient to define a new function 1
(11) f(kt
-

that our analysis will remain unaltered if we use the more genleral assumption <f(oo ) < 1. This will include as a special Of case the assumption f'> O and f'(oo =-O, implying a negative net marginal product of capital for a sufficienltlylarge capital input.) The function f(kt- Ct) is illustrated in
Figure 1. It should be noted that by as-

ct)

_-F(kt

Ct, 1)

wheref incorporatesimplicitly the growth factor n. Applying assumptions (8)-(10) to f, we may write
(8') (9') (10') f'>
-

sumption (10') the curve correspondingto f must be above the 45? line for some range of (kt-Ct) and that it must intersect the latter at some kt-ct>O, since J'(oo)
=1/n<1. It follows from the foregoing

f(O) = 0; n

f" <0;
ft(oo)=

facts that for any value of kt, O<kt < k, there corresponds a positive value of ct which will leave capital intact. To see this, consider X- ct in Figure 1. If ct= then k'- ct = k and we have kt+i
= > kt so that capital increases. Howk,+1

f'(0) = o,

where f' and f" denote the first- and second-orderderivatives of f. Note that in assumptions (8')-(10') we made use of relation f'= Fl/n. (It may be pointed out

then ever, if ct equals the segment Ao]B,x o and capital remains stationary. kt+hat Denote the relationshipbetween ktand the valueof c which leavesit intact by c= s(k)p. In view of Figure 1, it is clear that c, considered as a function of k, possesses a

304

THE AMERICAN ECONOMIC REVIEW


for ct (where ct =0), we shall assume (15) lim U'(Ct) =
?.

unique positive maximum3 at some k* O<k* <k. It is seen immediately from Figure 1 that this maximum corresponds to the point A* where the tangent to f is
parallel to the
450

line. This implies that


Fi(k-

at A* we have
(12) f'(k-c) c, 1)
=

The problem of determining the optimal program can then be stated as follows:
00

Maximize
(16) to co,cl,

U3

tGo

atu(ct) with respect


to kt+l =f(kt
-

or F1= n. In other words, when stationary


consumption per caplta is maximized (i.e., at k-=k* and c=c*), the net own rate of

. . . ,subject

Ct),

O<ct?kt,

t=O,

1,

. . . ,

given initial

return on capital (F1- 1) must equal the net rate of population growth (n- 1). This is the so-called "golden rule" of capital accumulation, with c* and k* being the golden rule4values of c and k. Another implication of Figure 1 is derived from the fact that the average product of capital is less than unity for kt> k, which means that kt+i<kt if kt> k. This implies that for any feasible program(with 0 < Ct ? kt)kt and ct are boundedfrom above
by max [k0,

capital ko. The first-order conditions, or the "Fisher conditions," for the optimal path require the equalization of the marginal rates of substitution in production to those in consumption for every t, which yields
u'(ct) -=f'(kt t
= -

Ct)

(17)

aut (Ct+1)
0 1,.

] for all t.

Let us turn now to the utility function which is supposed to represent the preferences of the economy as a whole. We assume the standard type of utility function used in growth theory, namely
00

(13)

E atu(ct) t=o

where a is a constant subjective discount factor satisfying 0 a < 1, and u is a function of per capita consumption. It is assumed that u has the following properties: (14) u' > u" < 0 for ct > 0 and in addition, to exclude cornersolutions
3This can also be verified as follows: In stationary solutions we have k =f(k-c). Differentiation with respect to k and c yields dc/dk=s'(k) = (f'- 1)/f'. (Note that by (10') we have s'(0) = 1.) Differentiating once more we obtain s"(k)=f"/(f')3<0 which shows that s(k) is concave. By assumption (10') (f'- 1) changes sign. It follows therefore that s(k) has a unique absolute maximum. At the maximum point we have s'(k) =0, i.e., in fSe as E (1P2).
4 See E. S. Phelps.

This can be verified as follows. Consider an optimal program and denote the optimal values of kt by kt. If, for some t, we take kt and kt+2as given, then in the optimal program atu(ct)+at+?u(ct+i) must be maximized with respect to ct and ct+i and kt+2 subject to kt+1=f(kt-ct) . =f(kt+i-ct+1), or kt+2=f f(kt- Ct) -t+ The latter is a transformation curve between ct+i and ct with a slope of f'(kt-ct). At an optimum of the miniature system we must therefore have an equality of f'(kt- Ct) with the marginal rate of substitution between ct+1 and ct, i.e., with Since this must be true for U'(ct)/au'(ct+i). all t, we have established (17). Ordinarily these are only the necessary conditions for an interior maximum. However, under our assumptions about the discount factor and about the concavity of the production and utility functions, it follows that the Fisher conditions are sufficient as well, provided the program satisfies the condition atu'(ct)-*0 as t->oo (as, for example, in the case where ct tends

LIVIATAN: OPTIMAL GROWTH to a positive stationary limit). Moreover, the program satisfying the foregoing conditions is unique. The proof of these statements' is of no direct concern to our discussion and we shall therefore take them as given, or treat them as assumptions. An important feature of the solution of (16) is that it will remain optimal (as far as present and future consumption is concerned) when the economy reexamines it as it moves actually into the future. This "dynamic consistency" property of the optimal program6 follows directly from the form of the utility function in (13). A related aspect of (16) is that the solution is independent of the calendar time. The index t should be interpreted as indicating the "number of periods ahead" at any given calendar date. Thus the process is a stationary one. We have formulated the maximization problem subject to a given arbitrary level of initial capital. As a preliminary inquiry we may, however, disregardinitial capital and examine whether our system is at all capable of a stationary optimal solution where t= c and kt= k for all t. Substituting the constant values of c and k into (17) we obtain
(18) 1/a = f'(k
-

305

by e and k, constitute an optimal sta-

tionary solution. Using the fact that 1/a =f'(k -e) >1, we may infer that k and c are smaller than the correspondinggolden rule values, i.e., k<k* and <c*. This is illustrated in Figure 1 where the optimal stationary solution is represented by the point A1which corresponds to f'> 1. The optimal stationary consumption level is then AB <A*B*. It is also seen that as time preference decreases, so that "a" approaches unity, the optimal stationary
values approach those of the golden rule.7

II. The ReducedMaximization System We shall assume that for any kothere exists a unique solution of (16). As an indication that our assumption about the existence of a maximum is a reasonable one, we note that the utility sum of any feasible consumption sequence is bounded from above by u(M)/(1-a), where M =max [k, ko]. Consider the maximized value of U, say U as a function of k0, and
denote this function by U = v(ko).8 We

c).

We know that there must exist a positive solution to (18) if f'(0) > 1/a >f'(cc). Since 1/a > 1 this is guaranteed by our foregoing assumption (10'). It also follows from the fact that f' is monotonically decreasing that (k-c) is uniquelydeterminedby (18). In a stationary solution we must also have
(19) k =-f(k
-

then assume that v(ko) is continuous and > twice differentiablefor all ko 0. Consider a two-stage maximization of (16). In the first stage we maximize the utility function treating coand ki as given parameters,ignoringcompletely koand the relation k1=f(k0-c0). We may then write (16) as
(20) u(c0) + a max
ct+1

E atu(ct+i)
tso

c).

for a given value of k1. Note, however, that the problem of maximizing co=O atu(ct+?) for a given value of ki is formally identical with maxi7 The case where strictly a= 1 raises some mathematical problem concerning the convergence of the utility sum. It has been shown, however, by Koopmans that this problem can be overcome by a proper adjustment of the origin of the function u. 8 In the logarithmic model where u(c) =log c, and log log (kt-ct), 0<-y<l, one can verify f (kt-ct)=g+? that v (k0)= (1/1 --ya) log k0+constant (note that in this model f'(oo) =0). See, for example, Radner (pp. 96105).

We may thereforedeterminefrom (18) and (19) a unique solution for c and k individually. The foregoing values, to be denoted
5 These statements can be verified by adopting the method of proof used by Cass (p. 235) to our model. Note that his equation (11) is our condition on atu(ct). 6 On the possibility of dynamic inconsistency, see the analysis by R. H. Strotz.

306

THE AMERICAN ECONOMIC REVIEW As a necessary condition for a maximum of the right-handside of (24), we have (25) u'(c0)= af'(k0
-

mizing E -o atu(ct) for a given value of ko.We may then use the foregoing v function to write
00

c)]f'(k0 - c0).

(21)

max E atu(ct+i)= v(kl).


Ct+1 t=O

Substituting in (20), we have (22) max U = u(c0) + av(k1).


C+1

Denote the maximizing value of co by c where the latter can be considered as a function of ko. Substituting in (24), we have (26) v(ko) = u(eo) + av[f(k. -c) Differentiating (26) with respect to koand using (25), we obtain v'(ko)= u'(o)deo/dko + av'Vf(k -co)]
(27) *f'(ko u' (eo)

This completes the first stage of maximization, from which we have obtained the reduced utility function (22). (It is understood, of course, that the reduced utility function depends not only on the original utility function but also on the production function.) In the second stage we treat c0 and k1as endogenous variables and maximize the reducedutility function (22) subject to the present period's production constraint k1=f(ko-c0). This determines the current period's optimal values of co and ki. The second-stageproblem9can then be written as
max [u(c0) + av(ki)]
Co t1

o)(1

deo/dko)

(23)
subject to ki = f(ko - c.), and given ko.

The recursivenature of this system is clear. Thus the value of ki determined by (23) becomes the next period's kowhich is used to determine by means of the same functions the value of the new ki and c0 and so on. Note also that by the definition of v, (23) satisfies (24) v(k0)=max {u(co) + avjf(ko -c.)I
CO

which is the fundamental functional equation of dynamic programming.10


9 It should be noted that while (23) represents a twoperiod analysis, it presupposes that an appropriate infinite horizon problem has been solved in the first stage of maximization. This, however, does not cause any difficulty as long as we are not interested in the computational aspect of the problem. 10See discussion concerning this type of equation in R. Bellman.

i.e. the marginal utility of current capital equals the marginal utility of current consumption (and consequently v'> 0). This is only natural, since one of the alternatives of using an increase in ko is consuming it in the currentperiod. It follows from (27) and (14) that v is concave if deo/dko>O. Consider the hypothesis that deo/dko<Ofor all ko. Then, if k'a increases, z, will not increase and hence ki will increase, implying that next period's "initial" capital increases. When we are actually in the next period, we face again the same kind of situation (i.e., an increase in initial capital) as in the original period. Repeating the previous argument, we shall find that no ct will increase, compared with the originalpath, in spite of the fact that an increase in some ct's is feasible.'1 Clearly this cannot happen with optimal programs. In particular it contradicts our earlier conclusion that v'>O. By using a somewhat more sophisticated argument, one could also reject the hypothesis that de0/dko<O holds for some k,. Hence
deo/dko>0 and v" <0 for all ko.

Using the foregoing results and the functional equation, we may infer that
'1 Note that we have used here the propertyof dynamicconsistency whichenablesus to relatethe planned path to the actual one.

LIVIATAN: OPTIMAL GROWTH < deo/dko 1. Thus differentiating (25) with respect to ko and rearranging terms, we obtain dco
(28) dko 1

307

family
u(c0) + av(ki) = constant.

~~~~1
_ __ _

a(vf" +

f'2v")

which is between zero and one. It also follows from these results that dkl/dko =f *(1-deo/dko) > 0. III. The DiagrammaticAnalysis of OptimalGrowth The short-runequilibriumat some given date is illustrated in Figure 2 by means of a Fisherine diagram, where on the horizontal axis we have co and on the vertical axis k1.The production possibility frontier for initial capital kois given by AB which representsthe function k1=f(k,,- co) in the (co, ki) plane. The slope of this curve at any point is given by df/dco0= <0. -J' Similarly,a2f/3c=-f" <0. The transformation curve relating ki and c,, is then of the usual concave shape. The indifference curves II' in Figure 2 correspond to the

A,

A,, 0

From our earlier results we know that v'>O and v" <0, just as is the case with the derivatives of u(c0). It follows therefore that the indifferencecurves of our reduced utility function have the usual convex shape. If the initial capital is ko Figure 2, in then the short-run equilibrium is determined at Eo so that the equilibriumvalues are given by (c?, k'). Suppose now that initial capital increases to kl. Then the production frontier moves uniformly"2to the right, and is represented by (say) A ,B, in Figure 2. What can be said about the new short-run equilibriumpoint El? We know from our foregoing analysis that both dco/dkoand dkl/dko are positive, i.e., the "Income Consumption Curve" EE' in Figure 2 must be upward sloping.'3Hence El must be above Eo and to the right of it. So far we dealt with short-run comparative statics. We have to determine now how the system actually moves from one time period to the next. Suppose that the economy is initially at Eo in Figure 3. Then in the next period the production frontier moves to (say) A lBl. Consider now the intersection point QO the horiof zontal line correspondingto ki with A ,Bl. The value of consumption corresponding to this point, say c', or ON in Figure 3, is determined by ki=f(1k-c') where k= k, is next period's initial capital. We may then write
(29)
=

?ko

?o-ko--.

R E V

Cf(k,-c')

~~~~I
co 0

80,B

k0 00

k'

co 0

FiGux.i2

12 Moreover, it can be seen that the production frontier moves to the right parallel (horizontally) to itself. Thus, for example, the slope of the production frontier at R is the same as at Eo. To see this, note that both E0 and R correspond to the same value of k1, i.e. to k'. Since ki is a monotonic function of ko-c0, this means that both Eo and R correspond to the same value of ko-co, which implies that f'(k0 -c0) is also the same. 13In the logarithmic model (described in fn. 8) the formula for the income consumption curve is given by ki=Acy where A is a positive constant.

308

THE AMERICAN ECONOMIC REVIEW initially at R0. The production frontier of the next period will then pass through QO. Some additional features of Figure 3 should be noted. We know from our earlier discussion that the point G correspondsto the golden rule levels of c and k. At this point the slope of the production frontier ABis unitary, i.e.,f'(k*-c*) = 1. Similarly, for any point (c, k) on SS' below G the slope of the production frontier is f'(k-c) >1, and above G we havef'(k-c) <1. In fact, the slope (in numerical value) of the production frontiers decreases steadily as we move upwardalong the SS' curve. This is so since as we move up the SS' curve we increase steadily the value of ki and hence
of ko-c0 (via ki =f(ko-c0)) so thatf'(k-c)

A0~~~~~0

80

81~~~R

ko

k, K)

kCo

FiGuE 3

from which it follows that c' is the stationary value of consumption which corresponds to ki. Using our earlier notation, we have c'=s(ki) 14 Let us now introduce in Figure 3 the (stationary state) function c=s(ki). This is represented by the SS' curve whose properties are derived from Figure 1. We can then see that if we start at Eo we may find a point (QO) the next period's proon duction frontier by the intersection of the horizontal line passing through Eo and the SS' curve. If in the next period the economy chooses E1 as its equilibrium point, then in the following period the production frontier must pass through Ql. A similar analysis applies to the possibility where the equilibriumpoint is to the right of SS', as in the case when the economy is
14 If k>k (see Figure 1), then c'=s(ki) <0, implying that we have to use the extended part of SS' in Figure 3 which continues to the left of the ki axis. Clearly, in this case, c' does not represent actually maintainable consumption. It does, however, serve the purpose of determining graphically the next period's production frontier, as in the case with c'>O.

decreases. Another point which should be noted is that at any point on SS' above G (such as H), the slope of SS' itself is always steeper, i.e., greater in numerical value, than the slope of the A B curve which passes through that point. This can be seen by differentiating the equation k=f(k- c) which determines the SS' curve. This yields -f' [ dk 11 (30) -== 1 dc 1-f' l s'(k)j Since above G we have 0 <1' <1 ,15 it follows that Idk/dc| > J-f'| as stated. Finally, let the equation of OG,along SS', be denoted by k= y(c), so that y(c) is the inverse of s(k) for k < k*. Then as c tends to zero, y(c) tends to zero, and consequently, f'(k-c) tends to infinity by (10'). It follows therefore from (30) that y'(0) = 1. The optimal path can now be easily determined if we introduce the incomeconsumption curve EE' into the picture, as we do in Figure 4. Denote the equation corresponding to EE' by k= e(c). Then from our foregoing analysis we know that e'>O. Suppose that at the origin we have e'(0) >y'(0) (= 1).16 Then because of e' >0
15 In fact, by (9') we always havef'> 1/n>O. 16In the logarithmic model, where e(c) =A cy. we . have e'(O)= + oo

LIVIATAN: OPTIMAL GROWTH we must have an intersection of EE' and SS' for some positive k and c. Let M be an intersection point of the two curves. Since this point is both a stationary solution and a short-run equilibrium,it must represent a stationary optimum. We have seen earlier that in a stationary optimum i'(k- C) = 1/a where k and j are determined by (18) and (19). It follows therefore that
the intersection of EE' and SS' is unique
k, St

309

A4-

and that the intersection point must be below G17 as drawn in Figure 4. Suppose alternatively that at the origin e'(O)<1. Then, since a (unique) positive stationary optimum is known to exist, the EE' curve must intersect18SS' from below at some positive k. However, this implies an additional intersection, which contradicts the uniqueness of the stationary optimum solution. Suppose now that the economy starts with some ko. Then the short-run equilibrium is determined in Figure 4 by the intersection of Ao0Bo EE' at the point and The next period's production frontier E,. is then A AB1 which passes through S0 and the new short-run equilibriumis at E1. It can then be seen that the system converges monotonically and asymptotically to the (stationary optimum) point M from below. Similarly, had we started with ko> k, the system would converge to M from above. Consider the golden rule point G. Since this point is necessarily above EE', it follows that at G the marginal rate of substitution of k1 for c0 on the consumption side is greater than on the production side, i.e., co is more valuable (at the margin) in consumption than in production. Thus if the economy is given an initial capital
17 Since l/a=fJ'(k-e)>1.

E~~

E~~~~E
E0

VS

B,,
koC' kO

B,

|B2

IB3

~~k
FIGuE 4

|B4 k'

CO

equal to k*, it will not stay at G but will rather increase consumption immediately to the point R, and then, in subsequent periods, will reduce consumption and capital graduallyto the point M.
REFERENCES R. Bellman, Dynamic Programming, Princeton 1957, pp. 11-16. D. Cass, "Optimum Growth in an Aggregative Model of Capital Accumulation,"Rev. Econ. Stud., July 1965, 32, 233-40. T. C. Koopmans, "On the Concept of Optimal Economic Growth," Pontificia Academia Scientiarus, Rome 1965, pp. 276-79. E. S. Phelps, "The Golden Rule of Accumulation: A Fable for Growthmen," Amer. Econ. Rev., Sept. 1961, 51, 638-42. R. Radner, Notes on the Theory of Economic Planning, Athens 1963. R. H. Strotz, "Myopia and Inconsistency in Dynamic Utility Maximization,"Rev. Econ. Stud., 1956, No. 3, 23, 165-80.

18The possibility of tangency can be ruled out by simple considerations of continuity. Thus suppose that EE' is below SS' except at M, where the two curves are tangent. Then for initial ko>k the sequence ct will converge to e, while for ko<k Ct will converge to zero. This implies that v(ko) is discontinuous at k, contrary to out assumptions.

You might also like