Growth Theory

Larry Blume

November 24, 2010

1 Introduction

Growth theory (aka capital theory) studies the problem of allocating resources
through time. Capital theory is as old as economics. We are covering the mod-
ern stuff, and exploring its relation to general equilibrium theory. The theory
begins with the classical Arrow-Debreu model, and specializes it by distinguish-
ing commodities according to the date at which they are available. An important
generalization, however, is that time runs to infinity. WIth infinite horizons, there
are an infinite number of goods. The urtext is Frank Ramsey (1928).

2 The One-Sector Discounted Ramsey Problem

There is a single decision maker, Robinson. There is a single good, maize.

Maize can be eaten or held over for planting the following year. The optimal
growth problem is, how much to eat, how much to plant.

1
Ramsey

2.1 Basics

Robinson’s problem is to solve

T
PT : sup ∑ δ t −1 u t ( c t )
{ct ,k t−1 }∞
t =1 t =1

s.t. ct + k t ≤ f (k t−1 ) for t ≥ 1, (1)

k0 ≤ k for given k > 0,
ct , k t−1 ≥ 0 for all t.

We are mostly interested in the case T = ∞ and ut ≡ u. A sequence {ct , k t−1 }∞ t =1

is a program or a plan. A program is feasible if it satisfies the constraints of the
optimization problems.

In a finite horizon problem, the usual conditions suffice for the existence
of a solution — continuity of f on [0, ∞) and the upper semi-continuity of u on its
range.1 When T = ∞, more is required.

Example:
Suppose u(c) = x and f ( x ) = βc for some β > 1/δ. Let (ct , k t ) be any
plan, and let τ denote the first date at which ct > 0. This plan can be improved
upon by the plan which takes c0τ = 0, k0τ +1 = k τ +1 + cτ , c0τ +1 = cτ +1 +
β(k0τ +1 − k τ +1 ) = βcτ , and otherwise equals (ct , k t ). This plan shrinks date-τ
consumption, correspondingly increases the date-τ capital stock, and then eats
the extra output at date τ + 1. The utility loss in date-τ is cτ . The utility gain
at date τ + 1 is δβcτ > cτ , and so the net is positive. It follows that no plan
having positive consumption in any period can be optimal. But the 0-consumption
plan can be beaten by many plans, including, for instance, eating k0 and then
consuming 0 from dates 1 onward. If βδ < 1, the reverse argument shows that
there is an optimal plan, which is to eat everything at the outset.

The following assumptions about primitives will guarantee the existence of an

optimum and do other things as well.

A.1: f : [0, ∞) → [0, ∞) is strictly increasing, concave and C1 .

2
Ramsey

A.2: f (0) = 0.

A.3: (a) f 0 (0) > 1, and

(b) limk→∞ f 0 (k ) < 1.

These assumptions imply:

Lemma 1. (i) There is a maximal sustainable stock b > 0 with f (b) = b. That
is, if k > b, then f (k ) < k.
(ii) 0 < k < b implies f (k ) > k.

In the finite horizon case, it is obvious that optimal plans exist. In the
infinite horizon case, it takes a bit more work. One needs to show that the supre-
mum is finite, and then that some plan achieves the supremum. The first is easy,
because the utility of any feasible plan is bounded above by u(b)/1 − β. The
rest requires some analysis. We will use one more assumption:

A.4: u is concave, C1 , and one of u and f is strictly concave.

The strict convexity assumptions imply that the optimal plan is unique. These
assumptions are way to strong, but they are at least intuitive.

Theorem 1. Assume A.1-4. Then an optimal plan exists, and is unique.

There are two ways to characterize optimal plans — the Euler equation and
transversality condition, first order conditions, and dynamic programming.

2.2 Characterization — First Order Conditions

Imagine moving a small amount h from current consumption to tomorrow, and

eating it then. The current utility loss is δt−1 u0 (ct ) h + o ( h). The output gain
tomorrow is f 0 (k t ) h. The utility gain tomorrow from this increase in consumption
is δt u0 (ct+1 ) f 0 (k t ) h + o ( h). If the plan is optimal, there is no perturbation that
can create a gain.

−δt−1 u0 (ct )h + δt u0 (ct+1 ) f 0 (k t )h + o (h) ≤ 0 for all h.2

3
Ramsey

Simplifying, dividing by h and taking a limit,

u0 (ct ) = δu0 (ct+1 ) f 0 (k t ). (2)

We have proven:

Theorem 2. Assume that utility and production are C1 . If {k∗t }∞t=0 is an optimal
∗
plan such that for all t ≥ 1, ct = f (k t ) − k t+1 > 0. Then for all t the optimal
plan satisfies equation (2).

We might worry about boundary solutions, Kuhn-Tucker conditions and

complementary slackness conditions, but the Inada conditions guarantee that
optimal consumption is always positive and that the optimal capital stock is al-
ways positive at every non-terminal date.

Infinite horizon problems allow two kinds of perturbations. Finite period

perturbations change consumption and production over a finite number of pe-
riods, leaving the plan unchanged outside the range of the perturbation. For
example, eat less today and more tomorrow, leaving everything else unchanged.
The euler equations guarantee that no finite period perturbation can increase
utility. An example of an infinite period perturbation is to eat more today and the
same tomorrow by shrinking next period’s capital stock, then eat the same the
next day by decreasing the subsequent period’s capital stop, and so on. The
transversality condition rules out the possibility of utility increases through infinite
horizon perturbations.

Transversality conditions are tricky. For instance, if utility is bounded, the

Euler equation alone is sufficient. In many problems, transversality conditions
weaker than (5) below can be found. Intuition for a transversality condition is
as follows: Consider a finite horizon version of (1). We know that all output
will be exhausted because utility is increasing. All initial capital will be used in
production. Taking advantage of this, the problem becomes:

T
∑ δ t −1 u


PT : sup f ( k t −1 ) − k t
{k t }tT=0 t =1
(3)
s.t. k0 = k for given k > 0,
k t ≥ 0 for all t.

4
Ramsey

The KT conditions are, for each t < T :

−δt−1 u0 f (k t−1 ) − k t + δt u0 f (k t ) − k t+1 f 0 (k t ) + νt

=0 (3a)
−δ T −1 u0 f (k T −1 ) − k T + νT


=0 (3b)
νt k t =0 (3c)
νt ≥0 (3d)

The Inada condition implies that k t > 0 for t < T , else all future consumption
will be 0. Thus (3c) implies νt = 0 and we recover the Euler equations from (3a).
Equations (3b) and (3c) imply that

δ T −1 u0 (c T )k T = 0. (4)

If consumption is not satiable, this condition makes the terminal capital stock 0.

In infinite-horizon problems, the transversality condition replaces the fi-

nite last period with its limit as T → ∞: The argument is usually that the limit
condition is plausible. A real proof might be fun.
Theorem 3. Assume A.1 A.2, A.4 and A.5. If a path {k∗t }∞
t=0 is optimal, then

lim δ T −1 u0 (c T )k∗T = 0. (5)

T →∞

Now we derive the transversality condition. This proof is due to Kami-

haga (2002). A more general but more complicated trratment is Ekeland and
Scheinkman (1986). We need the following additional assumption:

A.5: For the optimal path {k∗t }∞t=0 there is a T and as 0 < λ < 1 such that for
t > T , 0 < λk∗t < f (λk∗t ).

This is a somewhat strong interiority condition for optimal paths; stronger in par-
ticular than ct > 0 for large enough t.

We need the following fact about concave functions:

Lemma 2. Let g : [0, 1] → [−∞, ∞) be concave and suppose g(1) > −∞.
Then for any 0 ≤ γ < 1 and any γ < λ < 1,
g (1) − g ( λ ) g (1) − g ( γ )
< .
1−λ 1−γ

5
Ramsey

Draw the picture to see why this is obvious. Then try to construct the
proof yourself from the picture before reading the following quick proof.

Proof. Let µ = (1 − λ)/(1 − γ), which is less than 1. Concavity implies (why)
that

µ f ( γ ) + (1 − µ ) f (1) < f ( λ ).

Manipulating as necessary,


µ f ( γ ) − f (1) < f ( λ ) − f (1).

The rest is algebra.

The transversality condition is just another first-order condition, which

means that it is a calculus way of expressing the idea that a certain perturba-
tion from the optimum is costly. Assumption A.5 guarantees that there is a T and
a 0 < λ0 < 1 such that the production plan (k∗0 , . . . , k∗T , λk∗T +1 , λk∗T +2 , . . .) is
feasible for all λ0 < λ < 1. Since the k∗ plan is optimal,

δ T −1 u f (k∗T ) − λk∗T +1 ) − δ T −1 u f (k∗T ) − k∗T +1 )+

∑ δt−1 u f (λk∗t ) − λk∗t+1 − ∑ δt−1 u f (k∗t ) − k∗t+1 ≤ 0.



t> T t> T

(Notice that we have substituted in the (binding) constraint: ct = f (k t ) − k t+1 .

You should ask: What does this perturbation do? The effect of the perturbation is
to increase consumption at date T and distribute the loss of future consumption
over all future periods. (How do you know it is a loss. That is, what is the
relationship between ct and f (λk t ) − λk t+1 on feasible paths?) Divide through
by 1 − λ and check that

∗ ) − λk ∗ ∗ ) − k∗



u f ( k T − u f ( k T
δ T −1 T +1 T +1
1−λ
∞ ∗ ∗ ∗ ∗



t−1 u f ( k t ) − k t+1 − u f ( λk t ) − λk t+1
≤ ∑ δ
t = T +1
1−λ
∞
∑ δt−1 u f (k∗t ) − k∗t+1 − u(0) .



≤
t = T +1

6
Ramsey

The last inequality follows from lemma 2 (Why?).

Now take limits as λ → 1, to see that

∞
0
f (k∗T ) − k∗T +1 k∗T +1 ∑ δt−1 u f (k∗t ) − k∗t+1 − u(0) .




0 ≤ −u ≤
t = T +1

Now take T → ∞. The right hand side converges to 0, which proves the condi-
tion.

Ekeland and Scheinkman (1986) prove a more general result. For any
T > 0 a sequence {k t−1 , ct }tT=+02 is a T-feasible solution iff ct + k t+1 ≤ f (k t )
for all t ≤ T + 1.

A.a u is upper semi-continuous and f is continuous.

A.b There exists a sequence f t > 0 with ∑∞ t=1 δ f t < ∞ such that, for any
t

T -feasible path (k t , ct ), we have u(ct ) < f t for 0 < t < T .

A.c For all t > 0 and real numbers h there is some constant a(t, h) such that,
for all T > t and all T -feasible paths (k t , ct ) with ∑tT=1 δt−1 ut (ct ) > h, we
have |k t l + |ct | < a(t, h).

A.d There are constants d and β such that either u(0) = −∞ or −u0 (0)c ≥
du(0) + β.

A.e u is non-decreasing.

Theorem 4. If A.a-d hold and the supremum over all paths of the objective func-
tion exceeds −∞, then for any optimal path {(c∗t , k∗t )}∞
t =0 ,

1. An optimal path exists, and the value of the infinite horizon problem is the
limsup of the value of the T -horizon problems as T grows large.

2. limt→∞ δt−1 u0 (c∗t )k∗t+1 = 0.

7
Ramsey

2.3 Dynamic Programming

Denote by V (k ) the utility achieved by the optimal plan with initial capital stock k.
In dynamic programming parlance, the capital stock is the state variable. The
function V is the value function, and under our assumptions it is convex, con-
tinuous and strictly increasing. It is not hard to see that it is C1 , but C2 is quite
difficult. Let (c∗t , k∗t−1 )∞
t=1 denote an optimal program with initial capital stock
k∗0 . Suppose at time t the planner could choose again. He would have initial
capital stock k∗t−1 , and so his feasible set for periods t and onward will remain
unchanged from what he had planned for back in period 0. But will he want to re-
vise the plan? His objective function is now to max u(ct ) + δu(ct+1 ) + · · · . This
objective function is a scalar multiple of the function he optimized for these peri-
ods in the original problem, so the solution remains unchanged. This property is
called dynamic consistency. This suggests the following method for solving the
optimization problem (1): First solve the planning problem from period t on for
every possible capital stock, and get the value function Vt (k ). Then solve the
problem of trading off current consumption for capital:
Vt−1 (k t−1 ) = sup u(ct ) + δVt (k t )
{ct ,k t }
(6)
s.t. c t + k t ≤ f ( k t −1 ),
ct , k t ≥ 0,
and do this for all t. This method is called backward induction. It seems incred-
ibly inconvenient, so why do this? “Life must be lived forward and understood
backwards.” (Kierkegaard). Moreover, it is particularly useful for infinite horizon
problems, and also for numerical schemes for solving finite horizon problems.
Equation (6) is the Bellman equation. For infinite horizon problems, shifting one
period forward in time does not change the Bellman equation: V1 = V2 · · · . The
Bellman equation becomes
V (k) = sup u(c0 ) + δV (k0 )
{c0 ,k0 }
0 0 (7)
s.t. c + k ≤ f ( k ),
c0 , k0 ≥ 0,
This defines an equation on functions. Under our conditions one can show that
there is a unique V which solves equation (7). Furthermore, it can be found by
starting with any initial guess V and iterating the equation.

8
Ramsey

We formalize this by describing an operator that answers the question, if

h(k) is the value function today, what is the value function for yesterday? Let Fbc
denote the set of bounded-above, continuous functions from R+ to R. Define the
set Cbc of concave elements of F. Backward induction defines a function T that
maps Cbc into itself. (This has to be proved.) The function T is called the Bellman
operator. For h in the domain of T , the function Th takes the value ( Th)(k ) given
by
( Th)(k) = sup u(c0 ) + δh(k0 )
{c0 ,k0 }
0 0 (8)
s.t. c + k ≤ f ( k ),
c0 , k0 ≥ 0,
In solving the finite horizon problem, we proceeded by starting with h = 0. The
maximizers are optimal in the last period, and T0 is the value function for period
T − 1. More generally, if the value function in period t is h, then the problem in
equation (8) determines the value function Th and period t − 1, and the maxi-
mizers are the optimal actions. In general, the value function k periods back is
T · T · · · Th ≡ T k h.

It is clear that for finite horizon dynamic programming we start with 0, or a

function expressing the value of the terminal capital stock, and work backwards.
For infinite horizon dynamic programming, there is no end at which to start. But
no matter where we start, we end up at the same place: It is easy to show the
following.

Theorem 5. There is a unique V ∈ Cbc such that TV = V . Furthermore, for

any h in Fbc , limt→∞ supk≥0 || T t h − V || = 0.

This V , the fixed point for T , is the value function for the infinite horizon
discounted dynamic program. The optimal capital stock t + 1 to carry forward
from any period t with initial capital stock k t is that which solves equation (7).

How does this relate to the Euler equation and the transversality condi-
tion? Suppose for a moment that V was C1 . Then we can rewrite the optimization
problem (7) as
V (k) = sup u f (k0 ) − k0 + δV (k0 )


k0
s.t. k 0 ≤ f ( k ), (9)

k0 ≥ 0,

9
Ramsey

Suppose the optimal k∗ is interior. Then k∗ solves

−u0 f (k∗ ) − l ∗ + δV 0 (k∗ ) = 0.

Since the Euler equation also applies,

−u0 f (k∗ ) − k∗ + δu0 (k∗ ) = 0.

Consequently, V 0 (k∗ ) = u0 (k∗ ). This result is occasionally called the envelope

theorem of dynamic programming. The only thing to prove here is? that V
is in fact C1 . This is quite easy to do, and a proof can be found in “On the
Differentiability of the Value Function in Dynamic Models of Economics”, L. M.
Benveniste and J. A. Scheinkman, Econometrica 47(3) (1979), 727-732.

The transversality condition is embedded in the fact that V is bounded.?

The fundamental paper on discounted dynamic programming is: “Con-

traction Mappings in the Theory Underlying Dynamic Programming”, Eric V. Denardo,
SIAM Review 9(2) (1967), 165-177. Economists are fascinated with discounted
dynamic programming largely because they are too inept to do much else. How
different would all this look if we averaged over all the generations?

3 What Do Optima Look Like?

3.1 The Golden Rule

Any capital stock k∗ between 0 and b can be infinitely sustained. Define c∗ =

f (k∗ ) − k∗ . Since f (k ) lies above the diagonal for k in this range, c∗ is positive,
and so the plan with k t = k∗ and ct = c∗ for all t, with k0 = k∗ . is feasible. Not
all stationary plans are optimal, as one can see by checking the Euler equations.
Suppose, however, a stationary plan satisfied the conditions

δ f 0 (k∗ ) = 1.

Then it satisfies the Euler equation as well, and so is the optimal plan for initial
capital stock k∗ . When δ = 1, (a problematic case for many reasons), the optimal

10
Ramsey

stationary plan maximizes steady state consumption, and is known as the golden
rule plan. When δ < 1, we refer to modified golden rule plans.

Notice that the stationary optimal plan solves the maximization problem

sup δ f (k ) − k
k (10)
s.t. k≥0

The stationary capital stock is independent of preferences, and depends only on

technology and the discount factor. This result is known in the older literature as
the dynamic non-substitution theorem.

3.2 The Turnpike Theorem

Modified golden rule plans are important because they characterize the long-run
behavior of optimal plans.

Theorem 6 (Turnpike Theorem). The optimal plan for the infinite horizon problem
(1) converges monotonically to the modified golden plan with discount rate δ.

Proof. Take as given the properties of the value function: Increasing and con-
cave. Look at equation (7). In the objective function, both “commodities” are
normal goods. The usual analysis from undergraduate consumer theory ap-
plies. Thus ct and k t are both increasing in k t−1 . The Euler equation says that
δ f 0 (k t ) = u0 (ct )/u0 (ct+1 ). If k t < k∗ , then δ f 0 (k t ) > 1 because of the con-
vexity of f . Consequently ct < ct+1 because of the convexity of u. This implies
that k t < k t+1 because of monotonicity. Similarly for k t on the other side of the
golden rule.

4 Comparative Statics

The comparative statics question is, how do endogenous variables change with
respect to parameters? In the one sector growth model, the only parameter of

11
Ramsey

interest is the discount factor. “Statics” is a funny word to use in describing a

dynamic economy, but this is the conventional terminology.

The comparative statics of the golden rule are interesting. It is easy to

check the first-order conditions for the golden rule optimization problem (10) to
see that ∗ ∗
dc dk
>0 > 0.
dδ dδ
The first result is known as non-paradoxical consumption behavior, and the sec-
ond is known as capital deepening.

Capital deepening and non-paradoxical consumption behavior have their

analogues in the solution to the optimal growth problem:

Theorem 7. Consider two optimal growth problems with identical preferences,

technologies and initial capital stocks, and 0 < δ1 < δ2 < 1. Then k1t < k2t for
all t > 0, and for some T and all t > T , c1t < c2t .

5 Competitive Equilibrium

There are several ways one might think of writing down an equilibrium model
corresponding to the optimization problem (1). The most straightforward is to
assume that the planner is in fact a single consumer, and he owns a sequence
of firms which operates in every period, turning today’s maize into tomorrow’s
maize. In addition, we will have a sequence of prices { pt }∞ t=0 . Profits for the
firms are πt = pt f (k t−1 ) − pt−1 k t−1 . The firms’ optimize by choosing a pro-
duction plan to solve the program

P: πt ( p) = sup pt f (k t−1 ) − pt−1 k t−1

kt
s.t. k t ≥ 0 for all t.

12
Ramsey

The consumer’s problem is to choose a consumption plan to solve the program

∞
C: sup ∑ δ t −1 u ( c t )
{ct }∞
t =1 t =1
∞
s.t. ∑ p t c t ≤ p0 k 0 + ∑ π t
t =1 t
ct ≥ 0 for all t.

Definition 1. A competitive equilibrium is a program (ct , k t−1 )∞

t=1 such that
1. The production plan k solves P;

2. The consumption plan c solves C;

3. The program is feasible:

k0 = k and for all t ≥ 1, ct + k t = f (k t−1 ).

Prices are present value prices. Thus pt is the value of date t consump-
tion in units of date 0 consumption. Prices can be described by the rate of interest
rt . Take p0 = 1, and define pt−1 = (1 + rt ) pt .

Equilibrium will require that profits be finite. Let P denote the set of all
prices p such that π ( p) < ∞. This set is a convex cone. Equilibrium can be
characterized by first order conditions. For firms,

p t +1 f 0 ( k t ) − p t = 0 (11)

and

δt−1 u0 (ct ) − λpt = 0 (12)

CONSUMER TRANSVERSALITY CONDITION.

(13)

Not surprisingly, these conditions together imply the optimality first order condi-
tion (1) and the corresponding transversality condition. Conversely, an optimal

13
Ramsey

program is an equilibrium. Take p0 = 1. Define pt+1 = pt / f 0 (k t ), so that equa-

tion (11) is satisfied. Then the optimality condition becomes δu0 (ct+1 )/u0 (ct ) =
pt+1 /pt which is equivalent to the Euler equation (12).

The turnpike theorem has implications for equilibrium prices. In the mod-
ified golden rule, k t and ct are stationary. From the first order condition (12) it
follows that pt+1 /pt = δ, or in terms of the interest rate, 1/(1 + r ) = δ. As a
consequence of the turnpike theorem, equilibrium prices for optimal paths con-
verges to the modified golden rule prices.

6 Bibliography

Kamihigashi, T. A simple proof of the necessity of the transversality condition.

Economic Theory 20 (2002), 427–33.

Ekeland, I., and J. Scheinkman, Transversality conditions for some infinite hori-
zon discrete time optimization problems. Mathematics of Operations Research
11 (1986), 216–229.

Be sure to read Romer, D., It’s fourth down and what does the Bellman equation
say, 2003, http://elsa.berkeley.edu/users/dromer/papers/nber9024.pdf, to see what
dynamic programming can do for you.

14
Ramsey

Notes

1
A function f from Rn to R is upper semi-continuous iff { x : f ( x ) < y}
is open, or equivalently, iff for all x0 ∈ R and sequences xn converging to x0 ,
lim supn f ( xn ) ≤ f ( x0 ).
2
Recall that o ( h) is a term with the property that limh→0 o ( h)/h = 0.

15