0% found this document useful (0 votes)
19 views48 pages

SGPE Lecture2

1. The document outlines the neoclassical growth model, which introduces household optimization into the Solow model. [2] It presents the basic discrete-time model, focusing on the household's problem of maximizing lifetime utility from consumption subject to their budget constraint. [3] The household chooses consumption and savings over an infinite time horizon, subject to a no-Ponzi condition preventing unsustainable borrowing.

Uploaded by

Zubiya Moin
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
19 views48 pages

SGPE Lecture2

1. The document outlines the neoclassical growth model, which introduces household optimization into the Solow model. [2] It presents the basic discrete-time model, focusing on the household's problem of maximizing lifetime utility from consumption subject to their budget constraint. [3] The household chooses consumption and savings over an infinite time horizon, subject to a no-Ponzi condition preventing unsustainable borrowing.

Uploaded by

Zubiya Moin
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 48

SGPE Macroeconomics 1 - Section A

Part 2: The Neoclassical Growth Model

Jan Grobovšek

University of Edinburgh

Fall 2023

1 / 96

Outline

1 Motivation
2 Basic Neoclassical Growth Model (in Discrete Time)
3 Recursive Representation and Dynamic Programming
4 Neoclassical Growth Model with Population and Technological Growth
(in Continuous Time)
5 Literature
6 Appendix: Convergence and Monotonicity
7 Appendix: Euler Condition of Value Function
8 Appendix: Example of Computational Solution
9 Appendix: From the Lagrangian to the Hamiltonian

2 / 96
Motivation
The Solow model is a good simple tool to organise thoughts and use
for empirical predictions.
However, it is very light on theory because the crucial actor (i.e. the
representative household) does not optimise anything.
The neoclassical growth model (alternatively called the
Ramsey-Cass-Koopmans model) introduces intertemporal household
optimisation.
At first sight, its predictions about growth and development will not
markedly differ from those of the Solow model.
The benefit (which needs to traded-off against the additional degree
of analytical and computational complexity) is that the model is a
launching pad for almost all modern macro models with an infinite
time horizon.
Also, any proper theory of growth and development should have at its
heart incentives, in particular those related to investment activities.
3 / 96

Roadmap

1 Basic neoclassical growth model: solving an infinite horizon problem


in discrete time.

2 Recursive representation and dynamic programming: the meaning and


usefulness of value functions.

3 Neoclassical growth model with population and technological growth:


solving an infinite horizon problem in continuous time, with additional
empirically relevant features.

4 / 96
Outline

1 Motivation
2 Basic Neoclassical Growth Model (in Discrete Time)
3 Recursive Representation and Dynamic Programming
4 Neoclassical Growth Model with Population and Technological Growth
(in Continuous Time)
5 Literature
6 Appendix: Convergence and Monotonicity
7 Appendix: Euler Condition of Value Function
8 Appendix: Example of Computational Solution
9 Appendix: From the Lagrangian to the Hamiltonian

5 / 96

Law of motion

Time is discrete and runs from 0 to ∞, so t = {0, 1, 2, ...}.

The law of motion of capital:

K (t + 1) = (1 − δ )K (t ) + I (t ) (1)

where the depreciation rate is δ ∈ [0, 1].

The economy starts off with some initial capital, i.e. K (0) = K0 > 0.

All this is identical to the discrete-time Solow model.

6 / 96
Production

Production is also identical to the discrete-time Solow model:

Y (t ) = F (K (t ), L(t ); A). (2)

The same assumptions apply:


1 F is continuous, and twice differentiable and concave in both K and L.
2 F has constant returns to scale in K and L.
3 F (0, L) = F (K , 0) = 0, and Inada conditions on marginal
productivities w.r.t K and L.

For now we do not necessarily need constant factor shares or


labour-augmenting growth.

7 / 96

Market clearing

Again, the same clearing conditions apply as in the Solow model from
the first lecture.
The market clearing conditions are:
Financial markets. Household demand assets B from some
(unspecified) financial intermediary that backs assets by the supply of
physical capital. That supply must equal demand for capital by the
representative firm, hence:

K (t ) = B (t ). (3)

Labour market. Labour of size L:

L(t ) = L. (4)

Product market:
C (t ) + I (t ) = Y (t ). (5)

8 / 96
Firm’s problem

Again, identical to the discrete-time Solow model.

Firms maximise profits in each period w.r.t. K (t ) and L(t ) so

π (t ) = max {F (K (t ), L(t ); A) − R (t )K (t ) − w (t )L(t )} .


K ( t ) ,L ( t )

which gives the first-order conditions w.r.t. K (t ) and L(t ):

FK′ (K (t ), L(t ); A) = R (t ) (6)

FL′ (K (t ), L(t ); A) = w (t ). (7)

Note: R (t ) is the gross rate that the firm pays. Since a fraction δ of capital
depreciates, the net rate received by households is defined as

r (t ) = R (t ) − δ. (8)

9 / 96

Household

This is where we depart from the discrete-time Solow model.


There is a representative household of size L with period utility
u (C (t )), i.e. the household likes to consume.
Properties of the utility function:
1 Strictly concave: u ′ (C ) > 0, u ′′ (C ) < 0;


2 Inada condition: lim
C →0 u (C ) = ∞;

The household is infinitely lived and discounts time at the discount


factor β ∈ (0, 1).
Comment: Because β < 1 the household is impatient and, all things
being equal, prefers consumption today versus tomorrow. In addition,
the concavity of the period utility function ensures that the
households prefers to smooth consumption over time. Finally, the
Inada condition ensures that 0 consumption is never an optimal
choice, which makes the problem more tractable.
10 / 96
Household’s problem
The household maximises

U= ∑ βt u (C (t )) (9)
t =0

subject to the budget constraint

B (t + 1) ≤ (1 + r (t ))B (t ) + w (t )L − C (t ) (10)

with an initial given amount of assets B (0).


Note that the budget constraint is written as an inequality (≤ instead
of =).
In addition, we impose the no-Ponzi condition:
!
t −1
1
lim B (t ) ∏ ≥ 0. (11)
t →∞ 1 + r ( s )
s =1

11 / 96

Why the no-Ponzi condition?

Imagine the problem without (11). Because this is an infinite-horizon


problem, the optimal choice is to run B (t + 1) → −∞, repay it in the
following period and borrow even more. This makes no sense because
in equilibrium B (t ) = K (t ) and negative capital is not feasible,
yielding no equilibrium.

There are several plausible constraints that we could impose to


prevent that and which would all yield the same equilibrium.

For example, we could just say that B (t + 1) ≥ −B, ∀t, where B is


some finite positive number, including 0.

The most general condition, however, is the no-Ponzi condition (11).

12 / 96
No-Ponzi condition

Condition: !
t −1
1
lim B (t ) ∏ ≥ 0.
t →∞
s =1 1 + r (s )

B (t ) is the level of assets acquired in period t − 1, and its discounted


present value is ∏st − 1 1
=1 1+r (s ) .
Imagine that time were finite, and that the world were to end in
period t = 2. Then it must be that
1
B (2) ≥ 0.
1 + r (1)

This requires that the household does not “die” with a negative
discounted present value of assets (financial markets would not let
that happen).
Here, time runs to infinity, but the same principle applies.
13 / 96

Lagrangian
Let λ(t ) be the Lagrange multiplier associated with the budget
constraint (10).
The Lagrangian is therefore

(
L= max
{C (t ),B (t +1),λ(t )}t∞=0
∑ βt u (C (t ))
t =0

)
+ ∑ λ(t ) [(1 + r (t ))B (t ) + w (t )L − C (t ) − B (t + 1)]
t =0

The choice variables are C (t ), B (t + 1), and λ(t ),


∀t ∈ {0, 1, 2, ...}.
Notice that at each t, B (t ) is a state variable. This implies that
B (0) is not a choice, but given.
In addition, we impose that the household observes the no-Ponzi
condition. In practice, we can solve the Lagrangian and then check
that indeed the outcome is consistent with that condition.
14 / 96
Household’s first order conditions

The first-order conditions with respect to C (t ), B (t + 1), and λ(t ),


∀t ∈ {0, 1, 2, ...} are:

βt u ′ (C (t )) − λ(t ) = 0; (12)

−λ(t ) + λ(t + 1)(1 + r (t + 1)) = 0; (13)


(1 + r (t ))B (t ) + w (t )L − C (t ) − B (t + 1) = 0, (14)

In addition, optimality requires the terminal condition

lim [λ(t )(1 + r (t ))B (t )] = 0. (15)


t →∞

15 / 96

Interpretation of terminal condition

Note that the Kuhn-Tucker optimality condition is actually


 
λ(t ) (1 + r (t ))B (t ) + w (t )L − C (t ) − B (t + 1) = 0,

which requires either (1 + r (t ))B (t ) + w (t )L − C (t ) − B (t + 1) = 0


or λ(t ) = 0 or both.

For t < ∞, the household values consumption and therefore the


shadow value of relaxing the budget constraint is strictly positive,
λ(t ) > 0, implying that the budget constraint holds with equality.

For t = ∞, however, we are beyond the planning horizon. The budget


constraint no longer holds, but we are “left” with (1 + r (t ))B (t ), so
the optimality constraint becomes limt →∞ λ(t )(1 + r (t ))B (t ) = 0.

16 / 96
Transversality condition

Equation (15) is called the transversality condition (TVC).

The TVC ensures that the present discounted value of assets beyond
the planning horizon is 0.

In the model here, we will see that B and r converge to a finite


number, so limt →∞ λ(t ) = 0.

In other models, it may be that B grows infinitely large, and in such


cases we not only require that limt →∞ λ(t ) = 0, but that λ converges
to 0 sufficiently quickly, which is precisely condition (15).

The initial condition B (0) and the transversality condition set


boundaries to this infinite sequence problem.

17 / 96

Additional note on TVC

Condition (15) is sometimes derived as

lim λ(t )B (t + 1) = 0. (16)


t →∞

Combining (13) and (16) yields

lim λ(t + 1)(1 + r (t + 1))B (t + 1) = 0,


t →∞

which is equivalent to

lim λ(t )(1 + r (t ))B (t ) = 0.


t →∞

18 / 96
Household optimality conditions
Our next goal is to characterise the optimality conditions.
Ideally, this is done with only the observed (original) variables, rather
than using multipliers λ.
Replacing λ in (13) by (12) for t and t + 1 yields
λ(t ) = λ(t + 1)(1 + r (t + 1)) ⇔
βt u ′ (C (t )) = βt +1 u ′ (C (t + 1))(1 + r (t + 1)) ⇔
u ′ (C (t ))
= β(1 + r (t + 1)). (17)
u ′ (C (t + 1))
This series of equations, ∀t ≥ 0, is the canonical Euler condition of
intertemporal substitution. It’s quite possibly the most important
equation in macroeconomics.
Another characterising equation is the budget constraint (10), that
optimally holds with equality:
B (t + 1) = (1 + r (t ))B (t ) + w (t )L − C (t )
19 / 96

Household optimality conditions (cont’d)


Are we done with the characterisation? Not quite yet. Remember, we still
have the pesky no-Ponzi (13) and transversality conditions (15).
Next we show how they are related. Replace λ in (15) by (12):

lim βt u ′ (C (t ))(1 + r (t ))B (t ) = 0.


 
(18)
t →∞

Recursively substituting in (17) yields


0 = lim βt u ′ (C (t ))(1 + r (t ))B (t ) = lim βt −1 u ′ (C (t − 1))B (t )
   
t →∞ t →∞
 
t −2 ′ 1
= lim β u (C (t − 2))B (t )
t →∞ (1 + r (t − 1))
 
t −3 ′ 1
= lim β u (C (t − 3))B (t )
t →∞ (1 + r (t − 1))(1 + r (t − 2))
..
.
" # " #
t −1 t −1
1 1
= lim u ′ (C (0))B (t ) ∏ ⇔ 0 = lim B (t ) ∏ .
t →∞ 1 + r (s ) t →∞ 1 + r (s )
s =1 s =1

20 / 96
Household optimality conditions (cont’d)

This proves that the transversality condition implies that the no-Ponzi
condition must hold with equality. Why? Well, agents do not want to
die with a positive present value of assets.

Important clarification: When we set up the problem we need the


no-Ponzi condition for the problem to make sense. However, when we
characterise the solution, we do not need it any more, since (18) and
(17) imply that no-Ponzi holds.

21 / 96

Characterisation of equilibrium

The following equations fully characterise the equilibrium:


1 Production: The definition of production (2), ∀t ≥ 0;

⇒ Obtains Y (t ), ∀t ≥ 0

2 Firm behaviour: The first-order conditions (6) and (7), ∀t ≥ 0;


⇒ Obtains R (t ) and w (t ), ∀t ≥ 0

3 Household: the Euler condition (17) and the budget constraint (10),
∀t ≥ 0, as well as the transversality condition (18);
⇒ Obtains B (t ), ∀t ≥ 0, and C (t ), ∀t ≥ 0, while the TVC puts a boundary on B (∞)

4 The law of motion of capital (1), ∀t ≥ 0, with the initial capital stock
K ( 0 ) = K0 ;
⇒ Obtains K (t ), ∀t ≥ 0

5 Definition (8), ∀t ≥ 0;
⇒ Obtains r (t ), ∀t ≥ 0

6 2 of the following market clearing conditions (3), (4), and (5), ∀t ≥ 0;


⇒ Obtains L(t ) and I (t ), ∀t ≥ 0

22 / 96
Note on market clearing

The equilibrium involves budget and feasibility constraints.

By Walras’ Law, one of the clearing conditions in each period is


redundant (i.e. we can characterise the equilibrium without it).

Consider the product market clearing (5). Replacing I (t ) from (1)


gives C (t ) + K (t + 1) = (1 − δ)K (t ) + Y (t ). The firm’s profit
maximization implies 0 profits, so
C (t ) + K (t + 1) = (1 − δ)K (t ) + R (t )K (t ) + w (t )L(t ). From (3)
and (4), and using R (t ) = r (t ) + δ, yields
B (t + 1) = (1 + r (t ))B (t ) + w (t )L − C (t ). This is the hh’s budget
constraint. As that always holds, equation (5) is not strictly necessary
in the characterisation.

23 / 96

Implications
Consider the Euler condition, the only real behavioural condition of
the system.
u ′ (C (t ))
= β(1 + r (t + 1)).
u ′ (C (t + 1))
Since u ′ (c ) is decreasing in C , this implies that:

> C (t ) if β(1 + r (t + 1)) > 1,

C (t + 1) < C (t ) if β(1 + r (t + 1)) < 1,

= C (t ) if β(1 + r (t + 1)) = 1.

Interpretation is that the consumption path depends on two forces:


1 The opportunity cost of consumption today relative to tomorrow
(1 + r (t + 1) relative to 1) versus the value of utility today relative to
tomorrow (1 relative to β);
2 The degree of substitutability between consumption today and
tomorrow, which depends on the curvature of utility.
24 / 96
Implications
The interest rate r depends on a single variable, installed capital (in
equilibrium, the production function depends only on K (t ) since
L(t ) = L). From (6) this implies
r (t + 1) = r (K (t + 1)) = FK′ (K (t + 1)) − δ.
where K (t + 1) is pre-determined in t.
The Euler equation can therefore be written as
u ′ (C (t ))

= β(1 + FK′ (K (t + 1)) − δ). (19)
u (C (t + 1))
In addition, remember that by Walras’ Law the budget constraint
(10) implies the aggregate resource constraint:
K (t + 1) = F (K (t )) + (1 − δ)K (t ) − C (t ). (20)
The above two conditions (together with K (0) = K0 and the
transversality condition) fully characterise the interesting variables of
the equilibrium.
25 / 96

Steady state and dynamic properties

This economy also converges to a steady state, i.e. where all variables
become independent of time.
There are now two equations that must satisfy the steady state,
namely C (t ) = C (t + 1) = C ⋆ in (19) and K (t ) = K (t + 1) = K ⋆
in (20). The first one gives

1 = β(1 + FK′ (K ⋆ ) − δ). (21)

C ⋆ is then obtained from equation (20):

C ⋆ = F (K ⋆ ) − δK ⋆ .

If K (0) is off the steady state, capital converges monotonically to


K ⋆ , either from below (K (0) < K ⋆ ) or from above (K (0) > K ⋆ ).
For proofs, see Appendix. Monotonic Convergence

26 / 96
GDP growth

Economies starting out at K (0) < K ⋆ build up capital through time.


Because output only depends on K , output increases and we have
growth.

Conversely, economies starting out at K (0) > K ⋆ shed capital


through time, and experience negative growth.

When the economy converges to K ⋆ , GDP growth converges to 0.

27 / 96

Optimal growth

Importantly, this is optimal growth from the perspective of the


period 0 household.

It represents the sequence of feasible allocations that is


welfare-maximising from the perspective of period 0.

How can we prove that? It can be shown that a social planner


maximising the household’s utility subject to the economy’s feasibility
constraints would make exactly the same choices of consumption and
capital accumulation.

The reason for this is that markets here are competitive and complete
(completeness implies that goods, in particular all time-index goods,
can be freely traded), and preferences are locally non-satiated.

28 / 96
Steady state comparative statics

Rewriting (21) we have:

1 − β (1 − δ )
FK′ (K ⋆ ) = . (22)
β

Because FK′ is decreasing in K ⋆ we see immediately that the


steady-state value K ⋆ , and therefore Y (K ⋆ ), is increasing in β. The
more patient the household is, the more capital is accumulated in the
steady state.
As in the Solow model, K ⋆ is decreasing in δ.
Note that since FK′ (K ⋆ ) = FK′ (K ⋆ ; A), K ⋆ is increasing in A as well.
It follows that just as in the Solow model Y ⋆ = Y (K ⋆ ; A) is
increasing in A for two reasons, a direct effect and one passing
indirectly through a higher K ⋆ .
29 / 96

Consumption and the savings rate at steady state


Steady state consumption is obtained directly from the feasibility condition
C (t ) = Y (K (t )) + (1 − δ)K (t ) − K (t + 1) at steady state, C ⋆ = Y (K ⋆ ) − δK ⋆ .
The savings rate here is endogenous, and at the steady state it equals
Y (K ⋆ ) − C (K ⋆ ) K⋆
s⋆ ≡ = δ .
Y (K ⋆ ) Y (K ⋆ )

Contrast this to the Solow model where the savings rate is an exogenous
parameter.
An alternative formulation of s ⋆ is the following:
K⋆ K⋆ δβ
s⋆ = δ ⋆ = δ ′ ⋆ ⋆ = xK ( K ⋆
)
Y (K ) F (K )K 1 − β (1 − δ )
xK ( K ⋆ )

where (22) is used to replace F ′ (K ⋆ ).


If xK is constant (as e.g. when F is Cobb-Douglas) then s ⋆ is increasing in β
(more patience results in a higher saving propensity), increasing in δ (quick
depreciation requires a high saving effort), and increasing in xK (a higher factor
share of capital implies relatively more importance of capital in production, which
justifies a higher saving propensity).
30 / 96
Maximum steady state consumption
Does that imply that the resulting consumption at the steady state is the
maximum feasible one? No.
Consider again the feasibility condition C (t ) = F (K (t )) + (1 − δ)K (t ) − K (t + 1)
at steady state, C (K ) = F (K ) − δK . Maximising C with respect to K yields
FK′ (K gold ) = δ ⇔
K gold = FK′−1 (δ).
where gold refers to the so-called golden rule value of consumption-maximizing
capital at the steady state, C gold .
Our equilibrium value, meanwhile, is obtained from (22):
− −
 
1 β ( 1 δ )
K ⋆ = FK′−1 .
β
The two capital levels are not identical, it must therefore be that C ⋆ < C gold since
C gold is the maximum feasible consumption level. In particular, K gold > K ⋆ since
FK′−1 (·) is a decreasing function and because δ < [1 − β(1 − δ)]/β.
Intuition? The household maximises discounted utility and takes into account the
opportunity cost of building up capital today so as to consume tomorrow. Since
today is valued more than tomorrow, steady-state consumption is not maximized.
Note that as β → 1, K ⋆ → K gold and therefore C ⋆ → C gold .
31 / 96

What did we learn?

Compared to the Solow model, the equilibrium here yields similar


convergence towards a steady state of no GDP growth.

Notice that we have not mentioned how the growth rate of K or Y


evolves over that process. This is a bit more complicated than in the
Solow growth model because the savings rate is endogenous.

The important insight is that convergence towards 0 GDP growth


occurs even when households actively decide how much to save.

Again, the main driving force is that F has diminishing marginal


returns to capital.

32 / 96
What did we learn? (cont’d)

The concavity of the utility function ensures that the process of


convergence towards a steady state takes time.

For instance, if u (C ) were to be linear (e.g. u = C ) and the


household had sufficient assets, it would jump to the steady state
within one period (if it did not have sufficient assets, it would choose
C = 0 for some time and then jump to the steady state).

With strong curvature of u, convergence may be quite slow as the


household opts for a smooth consumption stream.

As for the dynamics outside (but close to) the steady state, the
neoclassical growth model is more useful in studying business cycles
rather than economic growth.

33 / 96

Outline

1 Motivation
2 Basic Neoclassical Growth Model (in Discrete Time)
3 Recursive Representation and Dynamic Programming
4 Neoclassical Growth Model with Population and Technological Growth
(in Continuous Time)
5 Literature
6 Appendix: Convergence and Monotonicity
7 Appendix: Euler Condition of Value Function
8 Appendix: Example of Computational Solution
9 Appendix: From the Lagrangian to the Hamiltonian

34 / 96
Solving the neoclassical growth model

Our system has an infinite number of equations:

It cannot be solved analytically.

It is also not straightforward to solve it computationally.

There is an alternative method to solve this system, which is the


bread-and-butter of modern macroeconomics:

35 / 96

Towards a recursive formulation


In the previous formulation our household starts out life with some B (0) and
then optimally chooses all current and future variables. Denote these
equilibrium (optimal) values with “hats,” i.e. Ĉ (t ) and B̂ (t ). Also,
associated to those choices are equilibrium object ŵ (t ), rˆ(t ), etc.
Let V (B (0)) denote the maximum value that the household obtains from
these choices:

V (B (0)) = ∑ βt u (Ĉ (t ))
t =0

V is a function of B (0) because that is the only variable that determines the
initial conditions (in particular B (0) itself as well as r (0) and w (0) via
B (0) = K (0)).
That function can be rewritten as follows.

V (B (0)) = u (Ĉ (0)) + ∑ βt u (Ĉ (t ))
t =1

= u (Ĉ (0)) + β ∑ βt u (Ĉ (t + 1)).
t =0
36 / 96
Towards a recursive formulation (cont’d)
Consider the meaning of the term ∑t∞=0 βt u (Ĉ (t + 1)). It represents the maximum
present discounted utility of the household, starting off with B (1), and in fact not
any B (1), but the optimal one chosen at 0, i.e. B̂ (1).
Notice that an agent born in period 1 with B (1) = B̂ (1) would make exactly the
same choices for t ≥ 1 as our original agent.
The resulting value is the maximum possible one, and thus
∑t∞=0 βt u (Ĉ (t + 1)) = V (B̂ (1)). We can hence write
V (B (0)) = u (Ĉ (0)) + βV (B̂ (1)).
So far we have determined a recursive structure, but there are no choices to be
made because the objects in the above line are expressions evaluated at optimum
choices. To introduce choices, we can write:
V (B (0)) = max {u (C (0)) + βV (B (1))} ,
C ,B (1)
subject to some constraints.
Finally, notice that what holds between 0 and 1 equally holds between any other
time periods, so we simply have
V (B ) = max′ u (C ) + βV (B ′ ) ,

C ,B

subject to some constraints, where the symbol ′ denotes future variables.


37 / 96

Household problem in recursive form

The household’s problem can therefore be represented by the


following Bellman equation where V is called a value function:

V (B ) = max′ u (C ) + βV (B ′ )

(23)
C ,B

subject to
B ′ = (1 + r )B + wL − C .

This works because we know that B is the only state variable of the
system and thus captures all the relevant period information (in
particular, in equilibrium w and r are only functions of K = B). E.g.,
if there was an additional state variable X we would write V (B, X ).

38 / 96
Household optimality conditions

It can be shown that the value function approach yields an identical


optimality condition to that of the sequence approach:

uC ( C )

= β (1 + r ′ ) (24)
uC ′ ( C )

coupled with the budget constraint.

For a proof, please see the Appendix. Proof of Euler Condition

39 / 96

Usefulness of recursive representation

There are a number of necessary requirements to write problems


recursively (e.g. it must be that β ∈ (0, 1), V must be differentiable
and continuous).

The exact conditions are beyond this course.

What is important to note is that recursive representations are very


useful for many infinite-horizon problems, both analytically and
computationally.

The Appendix provides a recipe of how to solve this problem on a


computer. Computational Example

40 / 96
Outline

1 Motivation
2 Basic Neoclassical Growth Model (in Discrete Time)
3 Recursive Representation and Dynamic Programming
4 Neoclassical Growth Model with Population and Technological Growth
(in Continuous Time)
5 Literature
6 Appendix: Convergence and Monotonicity
7 Appendix: Euler Condition of Value Function
8 Appendix: Example of Computational Solution
9 Appendix: From the Lagrangian to the Hamiltonian

41 / 96

Law of motion

Endogenous law of motion:

K̇ (t ) = I (t ) − δK (t ), (25)

with some initial capital, i.e. K (0) = K0 > 0.

Exogenous laws of motion:

Population grows at the rate n, starting at L0 :

L(t ) = exp(nt )L0 . (26)

Technology grows at the rate, g , starting at A0 :

A(t ) = exp(gt )A0 . (27)

42 / 96
Production

As in the continuous-time Solow model we assume that technology is


Harrod-neutral (purely labour-augmenting).

Y (t ) = F (K (t ), L(t ); A(t )) = F (K (t ), A(t )L(t )).

This will make it possible to work with stationary variables, namely:


 
F (K (t ), A(t )L(t )) K (t )
y (t ) ≡ ≡f , A(t ) = f (k (t ), A(t ));
L(t ) L(t )
 
f (k (t ), A(t )) k ( t )
ỹ (t ) ≡ ≡ f˜ , 1 = f˜(k̃ (t )).
A(t ) A(t )

Also, remember that

f˜k̃′ (k̃ (t )) = fk′ (k (t ), A(t )) = FK′ (K (t ), A(t )L(t ))

43 / 96

Market clearing and firm’s problem


Market clearing:
Financial markets:
K (t ) = B (t ). (28)
Labour market:
L(t ) = exp(nt )L0 (29)
Product market:
C (t ) + I (t ) = Y (t ). (30)
Firm’s problem:
F.O.C. w.r.t. K :

FK′ (K (t ), A(t )L(t )) = R (t ) = r (t ) + δ; (31)

F.O.C. w.r.t. L:
FL′ (K (t ), A(t )L(t )) = w (t ). (32)
Definition of net rate of return

r (t ) = R (t ) − δ (33)
44 / 96
Household’s preferences

This is slightly different from the previous model, both because time
is continuous and because there is population growth.
Remember that c (t ) ≡ C (t )/L(t ).
The household’s period utility is defined in per capita terms, where
u (·) is assumed to have identical properties as before (strict concavity
and Inada condition).
The household is assumed to discount time at the subjective rate
ρ > 0, but is fully altruistic towards future generations, so that the
effective discount rate is ρ − n.
In order to ensure discounting we assume that ρ − n > 0. In fact,
when technological growth is strictly positive (g > 0), and depending
on certain properties of u we may need to assume, in addition, that
ρ − n > g × factor where factor will be discussed shortly.
45 / 96

Household’s preferences (cont’d)

The utility at period 0 is then


Z ∞
U (0) = exp(−(ρ − n )t )u (c (t ))dt (34)
0

In continuous time, the discount factor is exp(−(ρ − n )). For n = 0,


the connection to discrete time is therefore exp(−ρt ) = βt , meaning
that exp(−ρ) = β. We could have also used β (properly corrected for
n), but it is more convenient to express the discount factor in terms
of the discount rate ρ.

46 / 96
Household’s budget constraint
The budget constraint in continuous time is:

Ḃ (t ) = r (t )B (t ) + w (t )L(t ) − c (t )L(t ).

It is more useful to rewrite this in terms of assets per capita, namely


b (t ) ≡ B (t )/L(t ). For this, note that

∂B (t ) ∂(B (t )/L(t )) Ḃ (t ) B (t ) ∂L(t )


ḃ (t ) ≡ = = − 2
∂t ∂t L(t ) L (t ) ∂t
Ḃ (t ) B (t ) Ḃ (t )
= − 2 L(t )n = − b (t )n.
L(t ) L (t ) L(t )

Replacing Ḃ (t ) into the initial condition yields a more useful


representation of the budget constraint:

ḃ (t ) = [r (t ) − n ]b (t ) + w (t ) − c (t ). (35)

47 / 96

Initial and no-Ponzi conditions


Initial condition:
b (0) = b0 (36)
No-Ponzi condition: In addition, we must prevent the household from
running up infinite debt. As in discrete time, this is most easily achieved by
imposing an additional no-Ponzi constraint. In continuous time it takes the
form:   Z t 
lim B (t ) exp − r (s )ds ≥ 0.
t →∞ 0
This can be more conveniently expressed in per worker units. Replacing
B (t ) = b (t )L(t ) = b (t )L0 exp(nt ),
 Z t 
b (t )L(0) exp(nt ) exp − r (s )ds
0
 Z t   Z t 
=b (t )L(0) exp n ds exp − r (s )ds .
0 0
Since L(0) > 0 is just an uninteresting constant, we have
  Z t 
lim b (t ) exp − [r (s ) − n]ds ≥ 0. (37)
t →∞ 0 48 / 96
Optimisation

The problem is to maximise (34) subject to the budget (35), the


initial condition (36), and the no-Ponzi game constraint (37).

Because we are operating with differential rather than difference


equations, this problem is slightly different.

We can make use of the Maximum Principle from the Theory of


Optimal Control. This lets us construct the Hamiltonian, a close
relative of the Lagrangian.

49 / 96

Household’s current-value Hamiltonian

We maximise (34) subject to (35), as well as the initial and terminal


constraints (36) and (37).

The current-value Hamiltonian is written as

H(
b t, b, c, µ) = u (c (t )) + µ(t ){w (t ) + [r (t ) − n ]b (t ) − c (t )} (38)

where

1 c is the control variable;

2 b is the state variable;

3 µ is the co-state variable;

4 and ρ − n is the discount rate.

50 / 96
Recipe for first order conditions using current Hamiltonians

The derivative of H
b w.r.t. the control variable (here c) must equal 0.

The derivative of Hb w.r.t. the state variable (here b) must equal


−µ̇(t ) + ζµ(t ) where ζ is the discount rate.

The derivative of H b with respect to the costate variable (here µ)


must equal ḃ (t ).

The transversality condition equals limt →∞ [exp(−ζt )µ(t )b (t )] = 0.


From the Lagrangian to the Hamiltonian

51 / 96

First order conditions


The derivative of H
b in (38) w.r.t. c equals 0:

b c (t, b, c, µ) = 0 ⇔ u ′ (c (t )) − µ(t ) = 0.
H (39)

The derivative of H
b w.r.t. b equals −µ̇(t ) + (ρ − n )µ(t ):

H
b b (t, b, c, µ) = −µ̇(t ) + (ρ − n )µ(t )
(40)
⇔µ(t )[r (t ) − n] = −µ̇(t ) + (ρ − n)µ(t ).

The derivative of Hb with respect to µ equals ḃ (t ), and gives back the


budget constraint:

H
b µ (t, b, c, µ) = ḃ (t ) ⇔ w (t ) + [r (t ) − n ]b (t ) − c (t ) = ḃ (t ). (41)

The final necessary equation is the transversality condition:

lim [exp(−(ρ − n )t )µ(t )b (t )] = 0. (42)


t →∞

52 / 96
Optimality conditions
We would like to characterise optimality in terms of the original variables
only (i.e. without costate variables). Equation (40) features µ and µ̇ that
we would like to replace.
We have µ from (39). To obtain µ̇ as well, just derive (39) with respect to
t: u ′′ (c (t ))ċ (t ) = µ̇(t ).
Now take (40), which is just µ(t )[r (t ) − ρ] = −µ̇(t ) and replace µ̇(t ) from
above and µ(t ) from (39) to obtain

u ′ (c (t ))[r (t ) − ρ] = −u ′′ (c (t ))ċ (t )

It is more convenient to express this in terms of growth rates, ċ (t )/c (t ).


Dividing by c (t ) on both sides we have.

ċ (t ) 1
= [r (t ) − ρ ] (43)
c (t ) ε(c (t ))
u ′′ (c (t ))c (t )
where ε(c (t )) ≡ − u ′ (c (t ))
.

53 / 96

Optimality conditions (cont’d)

Equation (43) is the continuous time version of the Euler equation


with ε being the negative of the elasticity of the marginal utility
u ′ (c (t )) with respect to c (t ).

In continuous time, ε also equals the inverse of the elasticity of


intertemporal substitution (EIS), where the EIS measures the
willingness of individuals to substitute consumption over time due to
a change in the rate of return.

The Euler equation and the budget constraint (35) at any t


optimality characterise the household’s decisions.

54 / 96
Optimality conditions (cont’d)
Notice that we are still carrying around the transversality condition
(42) with the costate variable µ(t ).
We can either replace µ(t ) by u ′ (c (t )) from (39), which is fine.
Or, more conveniently, we can find another expression for µ(t ). Note
from (40) that
µ̇(t )
= −[r (t ) − ρ].
µ (t )
Integrating the above equation yields
 Z t 
µ(t ) = µ(0) exp − [r (s ) − ρ]ds
0

Replacing the above µ(t ) into the transversality condition (42) now
gives the more useful expression:
  Z t 
lim b (t ) exp − [r (s ) − n ]ds = 0. (44)
t →∞ 0

where µ(0) = u ′ (c (0)) > 0 drops out. Notice that the no-Ponzi
condition (37) is indeed satisfied, and with equality. 55 / 96

Characterisation of equilibrium

The equilibrium is fully characterised by the following relationships:


1 Production: The definition of production (43), ∀t ≥ 0;
2 Firm behaviour: The first-order conditions (31) and (32), ∀t ≥ 0;
3 Household: the Euler condition (43) and the budget constraint (35),
∀t ≥ 0, as well as the transversality condition (44);
4 The law of motion of capital (25), ∀t ≥ 0, with the initial capital stock
K ( 0 ) = K0 ;
5 Market clearing conditions (28), (29), and (30), ∀t ≥ 0;
6 The following definitions hold: b (t ) = B (t )/L(t ), c (t ) = C (t )/L(t ),
and the net rate of return (33) ∀t ≥ 0
7 Technology evolves according to (27).

56 / 96
Balanced growth path

As in the Solow model, we are interested in the balanced growth path


(BGP).

Recall the Kaldor facts that we would like the model to satisfy:

1 Constant rate of output growth: limt →∞ Ẏ (t )/Y (t ) = (Ẏ /Y )⋆ ≡ gY⋆ .

2 Constant capital-output ratio: limt →∞ K (t )/Y (t ) = (K /Y )⋆ .

3 Constant capital share in national income


limt →∞ R (t )K (t )/Y (t ) = (RK /Y )⋆ ≡ xK⋆ .

57 / 96

Implication of BGP
The following must be true on the BGP:
1 Capital grows at the same rate as output (directly from a constant
capital-output ratio);
2 The interest rate is constant, i.e. r (t ) → r ⋆ . Why? For the capital
share R (t )K (t )/Y (t ) to be constant while K and Y grow at the same
rate, R (t ) and hence r (t ) = R (t ) − δ must be constant as well;
3 Consumption grows at the same rate as output. Why? Using the law
of motion of capital (25) and product market clearing (30) gives the
aggregate resource constraint:
K̇ (t ) = I (t ) − δK (t )
⇔K̇ (t ) = Y (t ) − C (t ) − δK (t ).
From here, notice that
K̇ (t ) Y (t ) C (t )
= − −δ
K (t ) K (t ) K (t )
Since at BGP K̇ (t )/K (t ) and Y (t )/K (t ) are constant, C (t )/K (t )
must therefore be constant as well, meaning that C grows at a
constant rate, equal to that of Y and K .
58 / 96
Utility consistent with BGP

Since at BGP C (t ) grows at a constant rate, so does


ċ ⋆ (t )
c (t ) = C (t )/L(t ), namely c ⋆ (t ) = gc .

This fact, together with a constant interest rate, implies the following
for our Euler condition (43):

1
gc = (r ⋆ − ρ )
ε(c (t ))

For this to be possible with strictly positive gc , it must be that


ε(c (t )) → ε, namely (the inverse of) the elasticity of intertemporal
substitution tends to a constant.

In other words, if we want a representation that delivers a BGP, we


need to work with a utility function with that particular property.

59 / 96

Example of utility function consistent with BGP

An example of such a utility function (widely used in theoretical and


empirical work) is the constant relative risk aversion (CRRA) utility
function: ( 1− θ
c −1
u (c ) = 1 − θ , if θ ≥ 0 and θ ̸ = 1
log c, if θ = 1.

Note: It can be shown that for θ → 1, [c 1−θ − 1]/(1 − θ ) converges


to log c. The higher is θ, the more “curvature” is there in preferences
(i.e. the more the household values a smooth consumption profile).

u ′′ (c (t ))c (t )
This utility yields ε(c (t )) ≡ − u ′ (c (t ))
= θ, by construction it is
always constant.

60 / 96
Production function consistent with BGP

In addition, remember that a BGP is only possible if technology can


be represented as Harrod-neutral (purely labour-augmenting).

61 / 96

Key equations in effective units

We are interested in characterising the growth trajectory along the BGP so let us
assume that the utility is indeed such that ε(c (t )) = ε is constant.

As in the discrete time model, it is convenient to represent the system in terms of


consumption and capital only. The key equations are the Euler condition, the
aggregate resource constraint, and the transversality condition.

Note on variable transformations. If x (t ) ≡ X (t )/L(t ) and x̃ (t ) ≡ x (t )/A(t ),


then
∂[X (t )/L(t )] Ẋ (t ) X (t )L(t )n Ẋ (t ) − X (t )n Ẋ (t )
ẋ (t ) ≡ = − = = − x (t )n;
∂t L(t ) L2 ( t ) L(t ) L(t )

∂[x (t )/A(t )] ẋ (t ) X (t )A(t )g ẋ (t ) − x (t )g ẋ (t )


x̃˙ (t ) ≡ = − = = − x̃ (t )g .
∂t A(t ) A2 (t ) A(t ) A(t )

In particular:
ẋ (t )−x (t )g
x̃˙ (t ) x̃˙ (t ) A(t ) ẋ (t )
= = = − g.
x̃ (t ) x (t )/A(t ) x (t )/A(t ) x (t )

62 / 96
Euler equation in effective units

Replacing the interest rate from (31), the Euler equation (43) is

ċ (t ) 1
= [FK′ (K (t ), A(t )L(t )) − δ − ρ]
c (t ) ε
1
= [f˜′ (k̃ (t )) − δ − ρ].
ε

In effective units it is rewritten to


c̃˙ (t ) 1
= [f˜′ (k̃ (t )) − (δ + ρ + εg )]. (45)
c̃ (t ) ε

Notice the last term εg .

63 / 96

Aggregate resource constraint in effective units

The second equation is the aggregate resource condition, suitably


transformed. We can either obtain it from the budget constraint or
directly from the aggregate resource condition obtained from
combining (25) with (30):

K̇ (t ) =F (K (t ), A(t )L(t )) − δK (t ) − C (t )

k̃˙ (t ) = f˜(k̃ (t )) − (δ + n + g )k̃ (t ) − c̃ (t ). (46)


h i
˙
where A(t )L(t ) k̃ (t ) + (g + n )k̃ (t ) = K̇ (t ) is used, obtained from
differentiating K (t ) = k̃ (t )A(t )L(t ) w.r.t. t.

64 / 96
Transversality condition in effective units

In addition, we have the transversality condition (44). Using


b (t ) = k (t ) and replacing r (t ) by f˜′ (k̃ (t )) − δ gives
  Z t 
lim k (t ) exp − [f˜′ (k̃ (t )) − δ − n ]ds = 0.
t →∞ 0

Since k (t ) = k̃ (t )A0 exp(gt ) this condition becomes


  Z t 
lim k̃ (t ) exp − [f˜′ (k̃ (t )) − (δ + n + g )]ds = 0. (47)
t →∞ 0

where A0 > 0 drops out.

65 / 96

Implications

Equations (45) and (46) represent a system of differential equations.

They are complemented by k̃ (0) = k̃0 and the TVC (47).

There are two loci (c̃˙ = 0) and (k̃˙ = 0), and the intersection is the
point where the economy is on the BGP. From (45) and (46) these
are given by
f˜′ (k̃c̃˙ =0 ) = δ + ρ + εg ; (48)
c̃k̃˙ =0 = f˜(k̃k̃˙ =0 ) − (δ + n + g )k̃k̃˙ =0 . (49)

66 / 96
Implications

Consider a situation where k̃ (t ) < k̃c̃˙ =0 . Comparing (45) and (48), it


must be that c̃˙ (t ) > 0. Conversely, if k̃ (t ) > k̃c̃˙ =0 , then c̃˙ (t ) < 0.
This gives us an idea of how consumption moves.

Analogously, consider the situation where c̃ (t ) < c̃k̃˙ =0 . Comparing


(46) and (49), it must be that k̃˙ (t ) > 0. Conversely, if c̃ (t ) > c̃k̃˙ =0 ,
then k̃˙ (t ) < 0. This gives us an idea of how capital moves.

The BGP is then the situation where both (48) and (49) coincide.

67 / 96

Graphical analysis of the equilibrium path

68 / 96
Equilibrium path

The household is given an initial k̃ (0) = k̃0 .

The growth rates of c̃ (t ) and k̃ (t ) are given by the Euler equation


and feasibility conditions.

These equations gives us:

the whole sequence of k̃ (t );

the whole sequence of c̃ (t ) except c̃ (0).

What value does the control variable (or “jump variable”) c̃ (0) take?

Answer: such that the transversality condition holds.

69 / 96

Equilibrium path (cont’d)

Why is the equilibrium path unique?

If c̃ (0) is chosen “too high” the two characterising equations send the
equilibrium to a situation where c̃ (t ) grows at a strictly positive rate while
k̃ (t ) eventually becomes negative.

This cannot be an equilibrium, in particular because the transversality


condition is not satisfied (f˜′ (k̃ ) is not defined for k̃ < 0).

70 / 96
Equilibrium path (cont’d)
Conversely, if If c̃ (0) is chosen “too low” k̃ (t ) grows at a strictly positive
rate while c̃ (t ) eventually drops and converges to zero, which is also the
situation where k̃ (t ) reaches its highest possible value k̃max > 0 according
to feasibility from (49): 0 = f˜((k̃max ) − (δ + n + g )k̃max .
It can be shown that that level of limt →∞ k̃ (t ) = k̃max does not satisfiy the
transversality condition, and hence optimality is violated! The TVC is then
 Z t 
lim k̃max exp −[f˜′ (k̃ (t )) − (δ + n + g )]ds
t →∞ 0

In the limit as t → ∞, we have that

˜ ′ −f˜′ (k̃max )k̃max + (δ + n + g )k̃max


− f (k̃max ) + (δ + n + g ) =
k̃max
−x (k̃max )f˜(k̃max ) + f˜(k̃max )   f˜(k̃max )
= K = 1 − xK (k̃max ) >0
k̃max k̃max
so the discounted value of k̃max (and therefore the TVC) explodes to ∞.
71 / 96

BGP comparative statics

The economy is therefore characterized by a unique equilibrium path


that converges to the BGP.

From (48), f˜′ (k̃ ⋆ ) = δ + ρ + εg , the following becomes clear:

BGP capital in effective units k̃ ⋆ is decreasing in δ and g (just as in


the Solow model), as well as ρ (i.e. more impatient households
accumulate a lower k̃ ⋆ ).

What is new is that ε also tends to lower k̃ ⋆ if and only if g > 0. The
intuition is that in the context of growing technology, a household
that strongly values a smooth consumption stream settles for a
relatively lower saving rate (it knows that production is increasing
exogenously, giving rise to higher possible consumption stream in the
future, and so it cuts back on savings).

72 / 96
Final note on discounting assumption

Recall that discounting assumes that ρ − n > 0, but that this may
not be sufficient here (rather, we assumed ρ − n > g × factor ).

The reason for this is buried in the transversality condition (47).

If we are to reach the BGP, then f˜′ (k̃ ⋆ ) = ρ + δ + εg from (48) and
so it must be that ultimately at the BGP (47) becomes:
  Z t 
lim k̃ ⋆ exp − [ρ − n − (1 − ε)g ]ds = 0.
t →∞ 0

Since k̃ ⋆ > 0, it must be that ρ − n − (1 − ε)g > 0, and therefore


ρ − n > (1 − ε)g . If ε > 1 (and g > 0), the assumption ρ − n > 0 is
sufficient, but not more generally.

73 / 96

Final note on discounting assumption (cont’d)

Intuition: In the presence of technological growth g > 0 and a low ε


(i.e. the household does not care too much for smoothing
consumption through time) the BGP requires that the household be
sufficiently impatient. Otherwise, the transversality condition breaks
down.

Why is that a problem? If the TVC is not satisfied, then the


household’s utility function is not finite. In that case not only do we
not reach a BGP, but the household’s utility (from period 0) is
infinite, and all of our equilibrium characterisation is meaningless
(there are many paths to infinity).

74 / 96
What did we learn?

If we would like to represent economies that have a BGP, the


requirements are becoming more demanding. Exogenous
technological growth g demands a specific form of the utility
function, and it sets new parameter restrictions.

This may be good news as it restricts the space of possible utility


functions to explore, but it may also put a question mark on our
theory.

Also, we may want to reconsider whether the BGP really is such a


natural outcome or whether it is only coincidentally observed in the
data for a couple of stable and leading economies.

Interestingly, on the BGP k̃ is not only affected by δ, ρ and g , but


also by another behavioural parameter, ε (the degree of risk-aversion).

75 / 96

Outline

1 Motivation
2 Basic Neoclassical Growth Model (in Discrete Time)
3 Recursive Representation and Dynamic Programming
4 Neoclassical Growth Model with Population and Technological Growth
(in Continuous Time)
5 Literature
6 Appendix: Convergence and Monotonicity
7 Appendix: Euler Condition of Value Function
8 Appendix: Example of Computational Solution
9 Appendix: From the Lagrangian to the Hamiltonian

76 / 96
References

Neoclassical growth model in discrete time:


Acemoglu 8.6

Dynamic programming:
Acemoglu 6; Ljungqvist and Sargent 3.

Neoclassical growth in continuous time:


Acemoglu 8.1-8.2.2, 8.3-8.5, 8.7; Romer 2.1-2.7.

Theory of optimal control:


Acemoglu 7.

77 / 96

Outline

1 Motivation
2 Basic Neoclassical Growth Model (in Discrete Time)
3 Recursive Representation and Dynamic Programming
4 Neoclassical Growth Model with Population and Technological Growth
(in Continuous Time)
5 Literature
6 Appendix: Convergence and Monotonicity
7 Appendix: Euler Condition of Value Function
8 Appendix: Example of Computational Solution
9 Appendix: From the Lagrangian to the Hamiltonian

78 / 96
Proof of convergence

Imagine we start out at T (including T = 0) and that K (T ) < K ⋆ .

We prove that K (T + 1) > K (T ). Suppose by contradiction that


K (T + 1) ≤ K (T ).

If K (T + 1) ≤ K (T ) < K ⋆ , it must be that FK′ (K (T + 1)) > FK′ (K ⋆ ), which


from (19) requires C (T + 1) > C (T ).

Replacing C in (20) this implies


C (T + 1) >C (T ),
F (K (T + 1)) + (1 − δ)K (T + 1) − K (T + 2) >F (K (T )) + (1 − δ)K (T ) − K (T + 1),
K (T + 2) − K (T + 1) < F (K (T + 1))−F (K (T )) + (1 − δ)[K (T + 1) − K (T )].

If K (T + 1) ≤ K (T ), then F (K (T + 1)) ≤ F (K (T )), and thus


K (T + 2) − K (T + 1) < 0, i.e. K (T + 2) < K (T + 1).

Repeating the above argument, we find C (T + 2) > C (T + 1) and


K (T + 3) < K (T + 2), etc.

79 / 96

Proof of convergence (cont’d)

As capital falls through time, it eventually becomes negative.

This violates the transversality condition (it is not defined because


FK′ (K ) is not defined for K < 0).

It follows that if K (T ) < K ⋆ , it must be that K (T + 1) > K (T ).

80 / 96
Proof of convergence (cont’d)

Now suppose that the starting value is such that K (T ) > K ⋆ .

We prove that K (T + 1) < K (T ). Suppose by contradiction that


K (T + 1) ≥ K (T ).

We can use the same argument as before with reversed signs to show that C now
would need to be strictly decreasing while K is strictly increasing, namely
C (T + 2) < C (T + 1), K (T + 3) > K (T + 2), etc.

Consumption now converges to 0 while the capital stock grows ever larger. This
violates the transversality condition.

It follows that if K (T ) > K ⋆ , it must be that K (T + 1) < K (T ).

81 / 96

Proof of monotonicity
Imagine we start out at T (including T = 0) and that K (T ) < K ⋆ .
From the previous we know that K (T + 1) > K (T ). Now we prove that
K (T + 1) ≤ K ⋆ . Suppose by contradiction that K (T + 1) > K ⋆ .
If K (T + 1) > K ⋆ > K (T ), it must be that FK′ (K (T + 1)) < FK′ (K ⋆ ), which
from (19) requires C (T + 1) < C (T ).
Replacing C in (20) this implies
C (T + 1) <C (T ),
F (K (T + 1)) + (1 − δ)K (T + 1) − K (T + 2) <F (K (T )) + (1 − δ)K (T ) − K (T + 1),
K (T + 2) − K (T + 1) > F (K (T + 1))−F (K (T )) + (1 − δ)[K (T + 1) − K (T )].

Since K (T + 1) > K (T ), also F (K (T + 1)) > F (K (T )), and thus


K (T + 2) > K (T + 1) > K ⋆ .
This, however, again sends us off to the situation where consumption converges to
0 while the capital stock grows ever larger, which is impossible.
It follows that if K (T ) < K ⋆ , it must be that K (T ) < K (T + 1) ≤ K ⋆ .
A similar argument can be made for when K (T ) > K ⋆ , in which case it must be
that K (T ) > K (T + 1) ≥ K ⋆ . These last two points prove that convergence is
monotonic. Back
82 / 96
Outline

1 Motivation
2 Basic Neoclassical Growth Model (in Discrete Time)
3 Recursive Representation and Dynamic Programming
4 Neoclassical Growth Model with Population and Technological Growth
(in Continuous Time)
5 Literature
6 Appendix: Convergence and Monotonicity
7 Appendix: Euler Condition of Value Function
8 Appendix: Example of Computational Solution
9 Appendix: From the Lagrangian to the Hamiltonian

83 / 96

Household optimality conditions in recursive form

To take explicitly account of the budget constraint, we can rewrite


(23) to
′ ′

V (B ) = max′
u ( C ( B, B )) + βV ( B ) (50)
B

where the budget constraint determines the function


C (B, B ′ ) = (1 + r )B + wL − B ′ .

The first thing to notice is that since V (B ) is defined as the


maximised value w.r.t. B ′ then it must be that the following
first-order condition holds: VB ′ (B ) = 0. This implies that

VB ′ (B ) = −uC (C (B, B ′ )) + βVB ′ (B ′ ) = 0. (51)

Note: I use subscripts to refer to partial derivatives to avoid confusion


(the symbol ′ is reserved for future variables).

84 / 96
Household optimality conditions in recursive form (cont’d)

What we do not know is VB ′ (B ′ ). How can we find that?


Let B ′ = h (B ) where h is the policy function that determines the
optimal B ′ given B.
Substitute this into (50) to obtain

V (B ) = u (C (B, h (B ))) + βV (h (B )).

where the max operator is no longer needed as we evaluate at the


optimal B ′ = h (B ).
Differentiate this with respect to B to obtain
VB (B ) = uC (C (B, h (B )))[1 + r − hB (B )] + βVB ′ (h (B ))hB (B )
 
= uC (C (B, h(B )))(1 + r ) + − uC (C (B, h(B ))) + βVB ′ (h(B )) hB (B ).

85 / 96

Household optimality conditions in recursive form (cont’d)


Notice from (51) that −uC (C (B, h (B ))) + βVB ′ (h (B )) = 0. Using
this so-called envelope theorem, the previous equation reduces to
VB (B ) = uC (C (B, h (B ))(1 + r ) or simply
VB (B ) = uC (C )(1 + r )

Forwarding the previous equation by one period yields


VB ′ (B ′ ) = uC ′ (C ′ )(1 + r ′ )

Plugging this back into (51) gives


−uC (C ) + βuC ′ (C ′ )(1 + r ′ ) = 0
or simply our familiar Euler equation
uC ( C )

= β (1 + r ′ )
uC ′ ( C )
The other characterising equation is the budget constraint. Back

86 / 96
Outline

1 Motivation
2 Basic Neoclassical Growth Model (in Discrete Time)
3 Recursive Representation and Dynamic Programming
4 Neoclassical Growth Model with Population and Technological Growth
(in Continuous Time)
5 Literature
6 Appendix: Convergence and Monotonicity
7 Appendix: Euler Condition of Value Function
8 Appendix: Example of Computational Solution
9 Appendix: From the Lagrangian to the Hamiltonian

87 / 96

Reformulation of the value function

To make things easier, let us rewrite equation (50) by noting that


B = K , that wL + rK = F (K ) − δK , and using the budget
constraint:

u (F (K ) + (1 − δ)K − K ′ ) + βV (K ′ )]

V (K ) = max

(52)
K

Numerically this can be solved super quickly. In particular, because V


is a so-called contraction mapping, the following computational
algorithm can be applied.

88 / 96
Recipe for value function iteration
1 Create a dense grid of N points K1 < K2 <, ..., < KN (these do not denote time
periods, just a sequence of values ranging from the smallest to the largest possible
K , e.g. 0.01, 0.02, etc.) It makes sense to centre the values around K ⋆ which can
be easily computed.
2 Guess a sequence of values {V0 (Kn′ )}N n=1 associated to each point Kn . For

example V0 (Kn ) = 0, ∀n.
3 Now maximize, for each Kn , the value in (52). Obviously, since we guessed
V0 (Kn′ ) = 0, ∀n (i.e. there is no benefit from accumulating K ′ ), the maximized
value for each Kn is K ′ (Kn ) = K1 , i.e. the smallest possible capital level. The
result is a new sequence {V0 (Kn )}N n=1 , but this time around it is increasing in Kn
because utility is increasing in K via F .
4 Set {V1 (Kn′ )}N N
n=1 = {V0 (Kn )}n=1 and repeat the same maximization of (52) with
the new sequence. Now there exists a trade-off between current consumption and
the future value, and we obtain a more reasonable {V1 (Kn )}N n =1 .
5 This is repeated J times until the distance, at each point Kn , between
{VJ (Kn )}N ′ N
n=1 and {VJ (Kn )}n=1 is arbitrarily small. At that point we have found
a solution. The contraction mapping guarantees convergence to that solution.
6 Most importantly, we found policy functions: for each Kn we know the optimal
K ′ (Kn ), and by extension also all other variables such as C (Kn ), Y (Kn ), etc.
Back
89 / 96

Outline

1 Motivation
2 Basic Neoclassical Growth Model (in Discrete Time)
3 Recursive Representation and Dynamic Programming
4 Neoclassical Growth Model with Population and Technological Growth
(in Continuous Time)
5 Literature
6 Appendix: Convergence and Monotonicity
7 Appendix: Euler Condition of Value Function
8 Appendix: Example of Computational Solution
9 Appendix: From the Lagrangian to the Hamiltonian

90 / 96
Lagrangian in continuous time
Combining the utility function (34) and the budget constraint (35),
the Lagrangian is
Z ∞ 

L= exp(−(ρ − n)t )u (c (t ))dt + λ(t ) [r (t ) − n]b (t ) + w (t ) − c (t ) − ḃ (t ) dt
0

+ lim λ(t )b (t )
t →∞
Z ∞ 
= exp(−(ρ − n)t )u (c (t ))dt + λ(t ) ([r (t ) − n]b (t ) + w (t ) − c (t )) dt
0
Z ∞
− λ(t )ḃ (t )dt + lim λ(t )b (t )
0 t →∞

where the terminal optimality (transversality) condition is


lim λ(t )b (t ) = 0. (53)
t →∞

Next, we derive w.r.t. c (t ) and b (t ). For the latter (the state


variable), this is complicated by the presence of ḃ (t ).
We can rewrite the Lagrangian using integration by parts, namely:
Z ∞ Z ∞
λ(t )ḃ (t )dt = lim λ(t )b (t ) − λ(0)b (0) − λ̇(t )b (t )dt. (54)
0 t →∞ 0
91 / 96

Lagrangian in continuous time, FOCs


Replacing (54) in the Lagrangian obtains
Z ∞ 
L= exp(−(ρ − n)t )u (c (t ))dt + λ(t ) ([r (t ) − n]b (t ) + w (t ) − c (t )) dt
0
Z ∞ (55)
+ λ (0)b (0) + λ̇(t )b (t )dt.
0

Now, the first-order conditions w.r.t. c (t ) and b (t ) are


exp(−(ρ − n )t )u ′ (c (t )) − λ(t ) = 0 (56)
and
λ(t )[r (t ) − n ] = −λ̇(t ). (57)
The budget constraint and the TVC (53) are
[r (t ) − n]b (t ) + w (t ) − c (t ) − ḃ (t ) = 0 (58)
and
lim λ(t )b (t ) = 0. (59)
t →∞
These equations fully characterize the equilibrium.
92 / 96
Hamiltonian
Next, we show that the exact same equations as above can be
obtained more succinctly using the Hamiltonian.
Define the Hamiltonian as
H(t, b, c, λ) = exp(−(ρ − n)t )u (c (t )) + λ(t ) ([r (t ) − n]b (t ) + w (t ) − c (t )) .
The Lagrangian (55) can therefore be rewritten as
Z ∞ Z ∞
L= H(t, b, c, λ)dt + λ(0)b (0) + λ̇(t )b (t )dt.
0 0
The first-order conditions are therefore
Hc (t, b, c, λ) = 0
and
Hb (t, b, c, λ) = −λ̇(t ).
Notice that these conditions are identical to (56) and (57). Besides,
we still have the same TVC:
lim λ(t )b (t ) = 0. (60)
t →∞
93 / 96

Current-value Hamiltonian
Analogously, we can construct the current-value Hamiltonian and
show that the optimality conditions remain the same.
Define the current-value Hamiltonian as
H(
b t, b, c, µ) = u (c (t )) + µ(t ) ([r (t ) − n ]b (t ) + w (t ) − c (t )) .

where
λ (t )
µ (t ) ≡ = λ(t ) exp((ρ − n)t ). (61)
exp(−(ρ − n )t )
The Lagrangian (55) can be rewritten as
Z ∞ Z ∞
L= exp(−(ρ − n )t )H(
b t, b, c, µ)dt + λ(0)b (0) + λ̇(t )b (t )dt.
0 0
The first-order conditions are therefore
exp(−(ρ − n )t )H
b c (t, b, c, µ) = 0 (62)
and
exp(−(ρ − n )t )H
b b (t, b, c, λ) = −λ̇(t ). (63)
94 / 96
Current-value Hamiltonian (cont’d)

Notice that (62) is simply

H
b c (t, b, c, µ) = 0. (64)

Next, derive (61) w.r.t. time:

µ̇(t ) = λ̇(t ) exp((ρ − n )t ) + (ρ − n )λ(t ) exp((ρ − n )t )


⇔λ̇(t ) = exp(−(ρ − n)t )µ̇(t ) − (ρ − n)λ(t )
⇔λ̇(t ) = exp(−(ρ − n)t )µ̇(t ) − (ρ − n)µ(t ) exp(−(ρ − n)t )

where in the last line we use µ(t ) = λ(t ) exp((ρ − n )t ).


Now, replace λ̇(t ) in (63) to find

H
b b (t, b, c, λ) = −µ̇(t ) + (ρ − n )µ(t ). (65)

95 / 96

Current-value Hamiltonian (cont’d)

The budget constraint is still

[r (t ) − n]b (t ) + w (t ) − c (t ) − ḃ (t ) = 0. (66)

Finally, using (61), the TVC (60) is now

lim exp(−(ρ − n )t )µ(t )b (t ) = 0. (67)


t →∞

Equations (64)-(67) fully characterize the equilibrium.


Back

96 / 96

You might also like