You are on page 1of 8

2 WESTERN ECONOMIC JOURNAL

where again F is a concave, homogeneous-first-degreefunction.


Intertemporal efficiency requires solution, with respect to (S’t+l, . . .,
. P t + l ) , of the maximum problem:

(3) Max S’t+, = Max F[F[S’t,. . .,


P t ; !?t+l, . . ., Pt+J,

!?t+l,. . .,P t + 1 ; S1t+l, . . .,P t + J


= F&, . . . ,Pt; !?t+l, . . .,P t + l l
If all the maxima are smooth and interior, this requires
ast+,
- =o= aF[slt+l, . . . ,Pt+r; S1t+l, . . .,Pt+ll
(4) ast+, ast+,
aF[sit, . , .,P t ; . . . ,s n t + J . . ., st+,,. . . p t + d
+ aF[sit+l,
asj,+,

( j = 2,. ..,II)
the fundamental envelope conditions for intertemporal &cienq.
For T > 2 periods ahead, we have the recursive maximum problems
. . .):
solved with respect to (P~+T-~,

(5) Max S I T = Max F[FT-i[S’t, . . .,Pt; S’t+lci, ...,P t + ~ - i ] ,


Ft+T-Ij . . . , P t + T - l ; !?t+T, -.. S”t+T]

= FT[slt, . . ,P t ; !?t+T, . , .,st+,] T = 1,293, . - - e

with Fl[S1t, .. . ,P t ; Sit+l, . . .,S”t+l] = FIS1t, . . .,Pt; !?:+I, . ..,Pt+11


It is easy to verify that this goes back inductively to the envelope condi-
tions of (4) for all intermediate times and to no new conditions.
The Turnpike theorem is a statement about some of the properties of the
FT functions as T+ 00 or becomes large.
3. Alternatively, we can think of (2) and (4) as a set of difTerence
equations determining the efficient trajectories of the [ S j t ] . In particular, it
is easy to verify that a voh Neiunann balanced growth (or “turnpike”) con-
figuration satisfies these equations, with

(6) SjtJS’t = sf, a constant, ( j = 2, . ..,n )


S’t = slosi.(Y,)t

where 7. is the von Neumann maximum rate of balanced growth of the


system.
What D-S-Stried to prove in Chapters 12-2-10 and 12-2-11 was that the
linearized difference equations for the S j t / S ’ t ratios, which are associated
SAMUELSON: TURNPIKE THSOREM 3

with the system (2)-(4), have characteristic roots which necessarily appear
in reciprocal pairs. This property ensures that the diagram of the phase
space does approximate the saddlepoint form shown in Figure 12-9of D-SS.
4. Taking advantage of the homogeneity properties of the system, which
arc basic to all production Turnpike theorems, I shall reformulate the problem
in terms of ratios, namely:
(7) sft = Sft/S1t , (1 = 2, . ..,n)
r't =s t + p t

Dividing our homogeneous transformation function of (2) through by


S':, we rearrange terms to get the (per capita of S') relation
(8) = SJ:+l/S1t = F[I,ft,
"r . . .,
Zot;llt+lylt, . ,S t + I y'tl ..
Now it is easy to see, economically and mathematically, that this implicit
equation can be solved uniquely for y':, namely:
(9) = f(A,.. ., Pt;ft+1, . , Et+I)
+# .
Intatanpod diiaency can now be formulated as the following maximum
problem:

where
(11) O(.f:,St+r) - log [f(St, . . ., sat; f t + I

Now we have a standard unconstrained problem in discrete-time pro-


,.. .,J"t+l)l.
gramming, with prescribed terminal conditions (so,sr). Elsewhere,' I have
recently proved the general theorem that all motions around a stationary
point of such a variational system must possess the desired ("generalized
catenary") property of having associated characteristic roots that appear in
nciprocplpain.
'[2?]. A still more - 4rnrlogue for d i k m c e equations of the P o i d stability a m l ~ i s
far pcdodic motiOar b #iven in [171.
4 WESTERN ECONOMIC JOURNAL

The efficiency conditions of optimality are the Euler-like variational rela-


tions for (10):
aqSt,St+I) aqJt-l,St)
aiiog (si,/sql = o =
(12) asj, asj, + asj,
(1 = 2, . , .,n), ( t = I,. . . , T-1)
It is straightforward to show the equivalence of these to (4)'s envelope
relations of efficiency, and that these equations have a stationary solution at
the von Neumann Turnpike, namely that st = s. does satisfy (12).
Because CP is not necessarily concave in all its variables, first-order necessary
conditions need to be supplemented by secondary conditions sufficient to
ensure a maximum and not a minimum or mere stationary value. Actually,
the concavity of the original F rules out Jacobi conjugate points and other
pathologies. Similar remarks are in order concerning f in (9') later and its
relation to the F of (1').
Associated with the n-1 second-order non-linear difference equations (12)
are the linear equations:
(13) 0 = +"yt+t + (+I1 + +")yt+1 + 4"yrt
where
Y t E [sit - ]:S .
( i = 2,. . ,n)

In the case n = 2, (13) becomes a single difference equation:


(14) ayt+, + byt+l + ayt = 0
Its solution is of the form:
Yt =z +
C~hi* C A t

where (Xi) are roots of the characteristic quadratic:


(15) ah' + bh + a = a(h - X 1 ) ( A - A1-') =0

where AI = -
2a
-b
+ (6' -2a46)"'
Except for singular initial conditions,

(16) limit yt = limit [CiAi' + G A I - ~ ]= * OO


t j - 1 - W t-,+oo
SAMUELSON: TURNPIKE THEOREM 5

For A, I 1 > 0 and singular initial conditions that make C, = 0, we have


(17) limit y t = 0
t+w

the razor‘s-edge case that represents approach to the turnpike.


In D-SSFigure 12-9,the phase space was incorrectly framed in terms of
the variables here written as ( y t J J t ) . Actually, the equations for these vari-
ables are shown there to be of the form:
-
~ t + i St =A St = M(Jt,Yt)

Yt - 7:-J = v y t = N(St,yt)

These are not explicitly in causal form since both receding and advancing
differences are involved. As I have shown in a paper [19] on Hamiltonian
canonical equations for a discrete-time system, this is no accident: generally
receding anb advancing differences will be involved. For small time intervals,
the discrepancy will not matter much.
By solving some implicit equations, or by redefining new phase variables
- s u c h as (stJst+l)-we can plot the vectors of motion as we go from
+
( s t , S t + J ) to ( s t +
A Stjt+l A ~ t + , ) . These approximate the smooth curves
of D-SSFigure 12-9 and the saddlepoint form shown there when n = 2.
For n > 2, the characteristic roots are given by the determinantal poly-
nomial of (2n-2)”‘ degree:

(18) A (A) = ded+”A* + (+I1 + +*’)A + +#I]

= A**’dd+J* + (+I, + +t*)A-J + +*JA-’]


= A*m-td&+tJ + (+I1 + +U)A-J + +J*A-’l’
-
- AS*’ A(A-I)

Hence, except for the singular case where A = 0, Xi-’ is a root where Aj is,
and all roots do come in reciprocal pairs.
For n > 2, there may be complex roots, and so our catenaries and saddle-
points are really generalizations of these simple concepts.
5 . This completes our corrected proof. Because the calculus of variations
is better known than discrete-time maximization, I shall quickly sketch the
parallel treatment of the continuous-time case, amplifying slightly my 1959
Stanford-symposium treatment. By using a parallel notation, the exposition
should be self-explanatory.
Corresponding to (1) and (2), we now have :
6 WESTERN ECONOMIC JOURNAL

(2') S' = F(S',P,. . . ,P;Y, . . .,!?) = P(S,P, . . . ,P;F,. . .,s";


o,o, . . .,0 )
where F is a concave, homogeneous-first-degreefunction?
Corresponding to (4)'s envelope relation, we have the Euler external
equations:
d aF aF aF aF
(4) z-7
as
=
as
7 + - 1 * -
as a t
(i = 2,. . .,n)
which are derived from the constrained Lagrangian expression:

As in (7), we define new variables #(t) = S j ( t ) / S ( t ) and convert


(2') into:

(8') d/s' = F(I,P, . . .,5";; 4- P(k/s'), . , .,> 4- s" (k/s'))


(9 = f(f, . . ,,I";.?, . . .,P), by Implicit Function Theorem.
Now efficiencyis equivalent to the simple maximum problem:

(10') Mu SoT $-[log S1]dt = Max f(.?, . .. ,E;>, . . . ,>) d:


{Si ( t )1
This implies the Euler diflierential equations

These have stationary turnpike solutions (which are actually maximum


steady-state values)

'In my 1959 paper 1181, quation (1) is written in the more specid form:

SI=f[sl, .. .,P;++ c,, ...,S.+ CJ - c1, r= n


where the WI here arc written there as ISt].
SAMUELSON: TURNPIKE THEOREM 7

Associated with the non-linear equations (12') are the linear equations:

(13')
where

y ( t ) = [ss'(t) - si*] (i = 2, . . .,n)


The general solution of (13') consists of sums of terms of the form c k
exp ( A k t ) , where Ak are characteristic roots of the characteristic polynomial
and G is, at worst, a polynomial in t:
A(A) = d&pA' + ( p- f " ) A -f']
= deUf'A' + (f" - f ' ) A -PI' by above symmetry relations
= deup(-A)' + (p' - f'")(-A) f"] -

= A(+)
This proves a classical PoincarC theorem which asserts that the characteristic
roots associated with the stationary solutions of a Hamiltonian system occur
in opposite-signed pairs. This confirms the generalized catenary or saddle-
point property of the turnpike. Now the phase diagram, Figure 12-9 of D-S-S,
becomes exact.
r'
When n = 2, - p' = 0,and the roots are real and opposite in sign;
a simple saddlepoint is then assured. (If f were convex in s and concave in
J, as in the pendulum problems of classical mechanics, the roots would be
opposite-signed pure imaginaries, = *
i(Jf'/l/lf'l)%, giving rise to
sinusoidal oscillations. If f is quadratic in the velocities, as in mechanical
Prl .P.q
x. r u (s*)s s - V(5)
problems involving no gyroscopic terms, we have f = ;;;
and f'" = 0 = y'. If V(s*) is a minimum of potential energy, and])s(".[
is positive definite, then all the roots will be pure imaginaries of opposite
signs, and the fluctuations will be conservatively stable. In economics, when
n > 2, we may have, as Morishima [ I l l has observed, opposite-signed com-
+
plex roots, u iv and - u - iv, resulting in oscillations anti-damped for large
8 WESTERN ECONOMIC JOURNAL

A t . In 4 or more dimensional phase space, we have generalizations of


saddlepoints and catenaries, but the Turnpike theorem remains
OIn 118,p. 2981, I first omitted the p‘ -
f” term and had to add an erratum slip to the p u b
lished Stanford book. Unfortunately, this correction was not made in the Collected Scientijc Papers.
In both cases, equations (34) and (35) should respectively have added in them middle terms,
reading as:

”I- -
o=z
,=2

a’F* A‘+
a’F*
ax,#ax,l
(XI

(- a*F*
- XI*)“ + [-a’F*
ay,/ax,
a*p a7*
-pF* l(x,
ax, ax,‘
- XI*)# - -
arF*
ax, ax,
(XI - XI*)
I
ca,,.a,ll ax,lax, - ax,ax, J =0
For n >
2, it will not be the case that A’ is necessarily real. Actually, opposite-signed complex
roots can definitely occur, as Professor Morishima had pointed out to me.
REFERENCES
1. D. Cass, ”Optimum Growth in an Aggregation Model of Capital Accumulation: A Turnpike
Theorem,” Econometrira, Oct. 1966,34,833-SO.
2. - and M. Yaari, “Individual Saving, Aggregate Capital Accumulation, and Efficient
Growth,” in Essays on the Theory of Optimal Economic Growth, ed., K . Shell, Cambridge
1967, 269-80.
3. H. Furuya and K. Inada, “Balanced Gmwth and Intertemporal Efficiency in Capital Accumu-
lation,” Internat. Econ. Rev., Jan. 1962,3, 94107.
4. F. Hahn, “Equilibrium Dynamics with Heterogeneous Capital Goods,” Qnart. Jonr. Econ.,
NOV. 1966,80,633-46.
5. J. R Hicks, C a p i d und Growth. Oxford 1965.
6. - -, “Prices and the Turnpike (I) The Story of a Mare’sNest,” Rev. Econ. Stud.,
Feb. 1961,28, 77-88.
7. T. C.Koopmans, “Economic Growth at a Maximal Rate,” Qnart. Jour. Econ., Aug. 1964,78,
35 s-94.
8. L. W. McKenzie, “The Dorfman-Samuelson-Solow Turnpike Theorem,” Internat. Econ Rev.,
Jan. 1963,4, 29-43.
9. - - , “Turnpike Theorems for a Generalized Leontief Model,” Econometrica, Jan. 1963,
31, 165-80.
10. M.Morishima, Eqnilibrinm Stability and Growth. Oxford 1964.
11. - -, “Proof of a Turnpike Theorem: The ‘NoJoint Production Case,”’ Rev. Econ.
Stud., Feb. 1961,28, 89-97.
12. H . Nikaido, “Persistence of Continual Growth near the von NeumaM Ray: A Strong Version
of the Radner Turnpike Theorem, Econometrira, Jan.-Apr. 1964,32,151-62.
13. E. 5.Phelps, Golden Ruler of Economic Growth. New York 1966.
14. R. Radner, “Paths of Economic Growth that are Optimal with Regard Only to Final States:
A Turnpike Theorem,” Rev. Econ. Stud., Feb. 1961,28, 98-104.
15. P. A. Samuelson, “A Catenary Turnpike Theorem Involving Consumption and the Golden
Rule,’’ Am. Econ. Rev., June 1965,55, 486-96.
16. ___, “A Turnpike Refutation of the Golden Rule in a Welfare-Maximizing Many-Year
Plan,” in Essays on the Theory of Optimal Economic Growth, ed., K. Shell, Cambridge 1967,
269-80.
17. -,. Classical Orbital Stability Deduced for Discrete-Time Maximum Systems, Jan.
1968, unpublished.
18. ____, “Efficient Paths of Capital Accumulation in Terms of the Calculus of Variations.”
in Malbematical Methods in Social Sciences, 19S9,eds., K.J. A r m , S. Karlin, and P. Suppes,
Stanford 1960, 77-88;also in The Collected Scientijc Papers of Paul A. Samuelson, Vol. 1,
cd., J. Stiglitt, Cambridge 1966,287-98.
19. , Hamilton‘s Canonical Equations for Discrete-Time Variational Systems, Jan. 1968,
unpublished.
20. ---, “Indeterminacy of Development in a Heterogeneous-Capital Model with Constant
Saving Propensity,” in Essays on the Theory of Optimal Economic Growth, ed., K. Shell,
Cambridge 1967,219-32.
21. -, “Market Mechanisms and Maximization (111) Dynamics and Linear Programming,”
a 19-49 RAND memo reproduced in Essays on tbe Theory of Optimal Economic Growtb, d.,
K. Shell, Cambridge 1967, 489-92.
22. . “Reciprocal Characteristic Root Property of Discrete-Time Maxima,” weftem
Econ. JouT., 6, 90193,
23. K. Shell and J. Stiglltz, “The Allocation of Investment in a Dynamic Economy,” enart. JOUT.
Econ., Nov. 1967,81, 592-609.

You might also like