Introduction to Dynamic Programming

Lecture Notes∗
Klaus Neusser†

November 30, 2017

∗
These notes are based on the books of Sargent (1987) and Stokey and Robert E. Lucas
(1989).
†
Department of Economics, University of Bern, Schanzeneckstrasse 1, P.O. Box 8573,
CH-3001 Berne, Switzerland. email: [email protected]

1
Contents
Notation and Symbols iii

1 Introduction 1

2 Basic concepts 1

3 The Bellman equation and the principle of optimality 5

4 Examples 8
4.1 Intertemporal job search . . . . . . . . . . . . . . . . . . . . . 8
4.2 The Cake Eating Problem . . . . . . . . . . . . . . . . . . . . 11
4.3 The Neoclassical Growth Model . . . . . . . . . . . . . . . . . 13
4.4 The Linear Regulator Problem . . . . . . . . . . . . . . . . . . 19

5 Principles of Dynamic Programming: the deterministic case 22

6 Principles of Dynamic Programming: the stochastic case 27

7 The Lucas Tree Model 30

7.1 The Equity Premium Puzzle . . . . . . . . . . . . . . . . . . . 35

A Some Topological Concepts 37

B Fixed Point Theorems 40

C Derivation of Equations Characterizing the Reservation Wage 44

D Mean Preserving Spreads 45

E The Riemann-Stieltjes Integral 47

i
List of Theorems
5.1 Theorem (Solution of Bellman Equation) . . . . . . . . . . . . 25
5.2 Theorem (Properties of the Solution) . . . . . . . . . . . . . . 25
5.3 Theorem (Further Properties of the Solution) . . . . . . . . . 26
5.4 Theorem (Solution found by Iteration) . . . . . . . . . . . . . 26
5.5 Theorem (Differentiability of Solution) . . . . . . . . . . . . . 26
6.1 Theorem (Solution of the Stochastic Bellman Equation) . . . . 28
6.2 Theorem (Property of Solution: Stochastic Case) . . . . . . . 28
6.3 Theorem (Further Properties of Solution: Stochastic Case) . . 29
6.4 Theorem (Solution found by Iteration: Stochastic Case) . . . . 29
6.5 Theorem (Differentiability of Solution: Stochastic Case) . . . . 29
B.2 Theorem (Contraction Mapping Theorem) . . . . . . . . . . . 40
B.3 Theorem (Blackwell’s sufficient condition) . . . . . . . . . . . 41
B.4 Theorem (Brouwer’s Fixed Point Theorem) . . . . . . . . . . . 42

ii
Notation and Symbols

X state space
U control
µ strategy
π policy
V value function

iii
1 Introduction
The foundation of macroeconomic theory on microeconomic principles has
been one of the most important developments in economics. By now it is
standard to view the decision maker (households, firms, state) as operating
in a complex stochastic environment. In particular, agents are conceived
as players in a dynamic stochastic game. Typically, it is assumed that the
players understand the rules of the game and can foresee the consequences of
their actions on themselves and on others. The agents are thus understood
to choose strategies which maximize their objective function.
From a modeling point of view, it is therefore necessary to specify exactly
the information at the disposal of agents, the technology available to them
and the restrictions which constrain their actions. The decisions depend typ-
ically on expectations about the future. These expectations will influence
the actions of the agents already today, thereby determining the possibilities
of the agent in the future. This intertemporal interaction is visualized in
figure 1. It is also important to emphasize that this strategic intertempo-
ral interaction is typical for economics system and differentiate them from
physical systems.

2 Basic concepts
In order to understand the issues involved in Dynamic Programming, it is in-
structive to start with the simple example of inventory management. Denote
the stock of inventory at the beginning of period t by Xt , then the manager
has to decide on how much to order to replenish the stock. The order Ut is
considered to be the control variable. In each period the inventory is reduced
by satisfying a stochastic demand Zt . It is assumed that the manager does
not know the realized value of demand at the time he makes the decision.
The situation is depicted in figure 2.
In this problem the variable which characterizes the state of the inventory,
in our case Xt , is called the state variable of the system. The state variable
or shortly the state must lie in some set called the state space denoted by

1
 today tomorrow
observed observed
variables variables
(data) (data)

past economic system economic system

(households, firms, state) (households, firms, state)

expectations
about the future

unobserved unobserved
disturbances disturbances

Figure 1: Intertemporal Macroeconomics

stochastic
demand
Zt

inventory inventory
at the beginning at the beginning
of period t of period t+1
Inventory
Xt Xt+1 = Xt + Ut - Zt

Ut period cost
order c Ut + h(Xt+1)

Figure 2: Inventory management

2
X . The control variable or control, for short, takes values in some set C. As
the demand Zt is assumed to be identically and independently distributed,
the state is just the inventory carried over from last period. Because next
period’s inventory is just given by the accounting identity

Xt+1 = Xt + Ut − Zt , (2.1)

the control could as well be Xt+1 . In a more general context, demand follows
a stochastic process {Zt } governed by some transition function Q where
tomorrow’s demand depends on the demand realized today and possibly on
the control Ut . In this more general setup, the state is given by (Xt , Zt ).
In each period, the manager faces some costs. In our example the costs
are twofold. On the one hand inventory holdings are costly. This cost is
denoted by h(Xt+1 ). The inventory costs may also account for shortage cost
for unfilled order if Xt+1 < 0. On the other hand, each order produces some
cost c Ut . In each period total costs amount to:

c Ut + h(Xt+1 ) (2.2)

The transition equation (2.1) and the cost or utility function (period
valuation) are the main ingredients of the inventory problem. The objective
of the manager is to minimize expected discounted costs:
T −1
X
J(x0 ) = E0 β t (c Ut + h(Xt+1 ))
t=0
T
X −1
= E0 gt (Xt , Ut , Zt ) + gT (XT ) −→ min, 0<β<1 (2.3)
Ut
t=0

starting in period 0 with inventory x0 . The optimization is subject to a

feasibility constraint

Ut ∈ Γ(Xt ) = [0, B − Xt ] ⊆ C = [0, B] (2.4)

where B is the maximal storage capacity.

It is clear that it is not optimal to set the controls U0 , . . . , UT −1 in advance
at time 0 without knowing the realizations of demand. It is thus definitely

3
better to decide upon the order at time t after knowing the state Xt . We
are therefore confronted with a sequential decision problem. The gathering
of information, here observing the state Xt , becomes essential. This way
of viewing the decision problem implies that we are actually not interested
in setting numerical values for Ut in each period, but in a strategy, rule,
reaction function, or policy function µt which assigns to each possible state
Xt an action µt (Xt ) = Ut .1
The control Ut must lie in some subset Γ(Xt ) ⊆ C, the control constraint
set at Xt . Γ assigns to every state Xt a set Γ(Xt ) = [0, B − Xt ] and is thus a
set valued function or correspondence from X into the subsets of C. As noted
in the footnote before, in a general stochastic context Γ may depend on Xt
and Zt . It is typically assumed that Γ(x) 6= ∅ for all x ∈ X . This implies
that there is always a feasible choice to make.
Let M = {µ : X → C s.t. µ(x) ∈ Γ(x)}. We call π = (µ0 , µ1 , . . . , µT −1 ) a
feasible policy if µt ∈ M . The set of all feasible policies is called the policy
space and is denoted by Π. If π = (µ, µ, . . . , µ), π is called a stationary
policy. In our example X = [0, B], C = [0, B], and Γ(Xt ) = [0, B − Xt ].
With this notation we can rewrite our decision problem as
T −1
X
Jπ (X0 ) = E0 gt (Xt , µt (Xt ), Zt ) + gT (XT ) −→ min (2.5)
π∈Π
t=0

where Jπ (X0 ) is called the value function at X0 . The expectation is taken

with respect to the probability measure on (Z0 , Z1 , . . . ) induced by Q given
initial conditions (x0 , z0 ). An optimal policy is thus a policy π ∗ which mini-
mizes the cost functional given the initial conditions x0 . The optimal value
function or optimal cost function J ∗ (x0 ) is defined as

J ∗ (x0 ) = inf Jπ (x0 ) (2.6)

π∈Π

The optimal policy thus satisfies: J ∗ (x0 ) = Jπ∗ (x0 ). It is important to

realize the sequential nature of the problem which makes the gathering of
information about the state of the system indispensable:
1
In case {Zt } is not i.i.d. the policy function can also depend on Zt .

4
 1. The decision maker observes the state of the system (i.e. Xt ) and
applies his decision rule Ut = µt (Xt ).

2. The demand Zt is realized according to some transition function Q.

3. Cost or benefit gt (Xt , µt (Xt ), Zt ) is incurred and added to the costs or

benefits of previous periods. In the last period cost or benefit is gT (XT )
only.

4. The state in the next period is constructed according to the transition

function Xt+1 = ft (Xt , µt (Xt ), Zt ) = Xt + Ut − Zt .

5. Start again in 1.

3 The Bellman equation and the principle of

optimality
The recursive nature of the decision problem leads to the principle of opti-
mality due to Bellman. This principle is at the heart of the dynamic pro-
gramming technique and is intimately related to the idea of time consistency
(see Kydland and Prescott, 1977). Suppose we have selected an optimal pol-
icy π ∗ = (µ∗0 , µ∗1 , . . . , µ∗T −1 ). Consider then the subproblem arising in period
τ , 0 < τ < T − 1, when the system is in state Xτ :
T −1
X
Eτ gt (Xt , µt (Xt ), Zt ) + gT (XT ) −→ min .
t=τ

The principle of optimality then states that the truncated policy (µ∗τ , µ∗τ +1 , . . . , µ∗T −1 )
is also optimal for the above subproblem. The intuition is simple: if the trun-
cated policy were not optimal for the subproblem, the decision maker would
be able to reduce cost further by switching to an optimal policy for the sub-
problem once Xτ has been reached. This idea can be exploited by solving
the decision problem by backward induction. For this purpose consider the
decision problem in period T − 1 with given state XT −1 . Clearly the decision

5
maker chooses UT∗ −1 = µT −1 (XT −1 ) ∈ Γ(XT −1 ) in order to minimize

ET −1 [gT −1 (XT −1 , µT −1 (XT −1 ), ZT −1 ) + gT (XT )] .

Denote the optimal cost for the last period by JT −1 (XT −1 ):

JT −1 (XT −1 ) = min ET −1 [gT −1 (XT −1 , µT −1 (XT −1 ), ZT −1 )

UT −1 ∈Γ(XT −1 )

+ gT ( fT −1 (XT −1 , UT −1 , ZT −1 ) )]
| {z }
=XT

JT −1 is clearly a function of XT −1 . In the process of computing JT −1 (XT −1 ),

we automatically obtain the optimal policy UT∗ −1 = µ∗T −1 (XT −1 ) ∈ Γ(XT −1 )
for period T − 1. Going back one period to period T − 2 with state XT −2 , the
decision maker should not just minimize expected period cost but take the
consequence of his decision for period T − 1 into account. Thus he should
minimize the expected period cost in T − 2 plus the expected period in T − 1
given that an optimal policy is followed in period T − 1. He must therefore
minimize:

ET −2 [gT −2 (XT −2 , µT −2 (XT −2 ), ZT −2 ) + JT −1 (XT −1 )]

−→ min
UT −2 =µT −2 (XT −2 )∈Γ(XT −2 )

Denoting the optimal cost by JT −2 (XT −2 ), we get:

JT −2 (XT −2 ) = min ET −2 [gT −2 (XT −2 , µT −2 (XT −2 ), ZT −2 )

UT −2 =µT −2 (XT −2 )∈Γ(XT −2 )

+ JT −1 (XT −1 )]

Going back further in time we obtain

Jt (Xt ) = min Et [gt (Xt , µt (Xt ), Zt ) + Jt+1 (Xt+1 )]

Ut =µt (Xt )∈Γ(Xt )

where Xt+1 can be substituted by Xt+1 = ft (Xt , µt (Xt ), Zt ). From this rea-
soning we can derive the following proposition.

6
Proposition 3.1. If J ∗ (X0 ) is the optimal cost, J ∗ (X0 ) = J0 (X0 ). More-
over,

Jt (Xt ) = min Et [gt (Xt , Ut , Zt ) + Jt+1 (ft (Xt , Ut , Zt ))], t = 0, 1, . . . , T − 1,

Ut ∈Γ(Xt )

and JT (XT ) = gT (XT ). Furthermore, if Ut∗ = µ∗t (Xt ) ∈ Γ(Xt ) minimizes the
right hand side above for each Xt , then the policy π ∗ = (µ∗0 , µ∗1 , . . . , µ∗T −1 ) is
optimal.

In macroeconomics, we have, typically, a maximization instead of a min-

imization problem and the following time invariant specification which leads
us to consider only time invariant policies π = (µ, µ, . . . , µ):

• gt (Xt , µ(Xt ), Zt ) = β t U (Xt , µ(Xt ));

• Xt+1 = f (Xt , µ(Xt ), Zt+1 );

• JT (XT ) = 0.

The above iteration can therefore be rewritten as follows:

Jt (Xt ) = max Et [β t U (Xt , µ(Xt )) + Jt+1 (f (Xt , µ(Xt ), Zt ))]

Ut =µ(Xt )∈Γ(Xt )

=⇒
Vt+1 (Xt ) = max U (Xt , µ(Xt )) + βEt Vt (f (Xt , µ(Xt ), Zt ))
Ut =µ(Xt )∈Γ(Xt )

where VT −t (x) = β −t Jt (x).

We might then consider the limit

lim Vt (x) = V (x).

t→∞

If this limit exists, it must satisfy the following functional equation, called
Bellman equation:

V (x) = max U (x, u) + βEV (x0 ) (3.1)

u∈Γ(x)

= max U (x, u) + βEV (f (x, u, z 0 )) (3.2)

u∈Γ(x)

7
where a prime denotes next period’s value. The expectation E is conditional
on the information available to the agent. It must be emphasized that in
the Bellman equation the unknown is not a number as in standard algebraic
equations, but a function. In this context the following mathematical issues
arise:

1. Does the limit exist? Is the limit independent from the initial functional
V0 ?

2. Does the Bellman equation have a unique solution? Can it be found

by iteration irrespective of the starting functional?

3. Does there exist a time invariant policy function µ? What are the
properties of such a function?

4. Is V and/or µ differentiable? If so we obtain an analogue to the enve-

lope theorem, the so-called Benveniste-Scheinkman formula:

∂V (x) ∂U (x, u) ∂V (x0 ) ∂f (x, u, z 0 )

= + βE × (3.3)
∂x ∂x ∂x0 ∂x

4 Examples
4.1 Intertemporal job search
Consider the following simplified intertemporal job search model. Suppose
that a worker, if unemployed, receives in each period a job offer which
promises to pay w forever. If he accepts the offer he receives w in all subse-
quent periods. Assuming that the worker lives forever, the value of the job
offer in period t is
∞
X w
β τ −t w =
τ =t
1−β
If he rejects the offer, he receives an unemployment compensation c and
the chance to receive a new wage offer next period. Wage offers are drawn
from a known probability distribution given by F (w0 ) = P[w ≤ w0 ] with
F (0) = 0 and F (B) = 1 for some B < ∞. Denoting the value of a wage offer

8
by V (w) and assuming that wage offers are independent draws, the value of
waiting one more period therefore is
Z ∞
c+β V (w0 )dF (w0 )
0

Thus the value of a wage offer must satisfy the following functional equa-
tion:  Z ∞ 
w 0 0
V (w) = max ,c + β V (w )dF (w ) (4.1)
1−β 0
From figure 3, we see that the solution must have the reservation wage prop-
erty: ( R∞
W
1−β
= c+β 0
V (w0 )dF (w0 ), w ≤ W ;
V (w) = w
(4.2)
1−β
, w ≥ W.
where W is called the reservation wage. It is determined through the follow-
ing equation:
Z ∞ Z ∞
W 0 0 β
=c+β V (w )dF (w ) ⇒ W − c = (w0 − W )dF (w0 )
1−β 0 1−β W
(4.3)
Z ∞
⇒W −c =β V (w0 − W )dF (w0 )
W
(4.4)

where the left hand side represents the cost of searching one more time having
a wage offer W at hand and where the right hand side is the expected benefit
of searching one more time in terms of the expected present value associated
with drawing an offer w0 > W .
Manipulating the above equation then leads to an alternative character-
ization of the reservation wage:
Z W
W − c = β(Ew − c) + β F (w0 )dw0 (4.5)
0

This characterization immediately shows that an increase in unemployment

compensation or a mean-preserving increase in risk causes the reservation
wage to rise.2
2
The derivation of equations (4.3), (4.4), and (4.5) can be found in Appendix C.

9
 V(w)

w/(1-b)

c + bò V(w ¢)dF(W¢)

0 reject W accept w
wage offer wage offer

Figure 3: Job search model

The search model can be used to get a simple equilibrium model of un-
employment known as the bathtub model. In each period, the worker faces a
given probability α ∈ (0, 1) of surviving into the next period. Leaving the
remaining parts of the problem unchanged, the worker’s Bellman equation
becomes:
 Z 
w 0 0
V (w) = max , c + αβ V (w )dF (w ) . (4.6)
1 − αβ
This is essentially the same equation with only the discount factor changing.
Let the implied reservation wage be w̄. Assume that in each period there is a
constant fraction 1 − α of new born workers. They replace an equal number
of newly departed workers. If all new workers start out being unemployed,
the unemployment rate Ut obeys the law of motion:

Ut = (1 − α) + αF (w̄)Ut−1 . (4.7)

The right hand side is the sum of the fraction of new born workers and
the fraction of surviving workers who remained unemployed at the end of

10
last period (i.e. those who rejected offers because they were less than the
reservation wage w̄). The steady state unemployment rate U ∗ is
1−α
U∗ = . (4.8)
1 − αF (w̄)

Let N be the length of time until a successful offer is encountered. N = 1

means that the first job offer is accepted. If λ denotes the probability that a
R w̄
job offer is rejected, i.e. λ = 0 dF (w), then P[N = j] = (1 − λ)λj−1 . Thus,
the waiting time is geometrically distributed. The mean waiting time then
becomes ∞ ∞
X X 1
jP[N = j] = j(1 − λ)λj−1 = . (4.9)
j=1 j=1
1−λ

Next we show that it is never optimal for the worker to quit. For this
purpose consider the following three options:
w
(A1): accept the wage and keep it forever: 1−αβ
.

(A2): accept the wage and quit after t periods:

w − (αβ)t w
 Z 
t 0 0
+ (αβ) c + αβ V (w )dF (w )
1 − αβ
w w − w̄
= − (αβ)t
1 − αβ 1 − αβ

V (w0 )dF (w0 ) = w̄

R
(A3): reject the wage: c + αβ 1−αβ

If w < w̄, then A1 < A2 < A3, and if w > w̄, then A1 > A2 > A3. The
three alternatives yield the same lifetime utility if w = w̄. Thus, A2 is never
optimal.

4.2 The Cake Eating Problem

Suppose that you want to eat a cake over the periods t = 0, 1, . . . , T . Your
intertemporal utility functional V is:
T
X
V (c0 , c1 , . . . , cT ) = β t ln ct .
t=0

11
The cake has an initial size of k0 > 0. The transition equation for the size of
the cake clearly is:
kt+1 = kt − ct .

The Euler equation for the above optimization problem is given by

U 0 (ct ) = βU 0 (ct+1 ) =⇒ ct+1 = βct .

The size of the cake in each period can therefore be computed as follows:

k1 = k0 − c0
k2 = k1 − c1 = k0 − c0 − βc0
| {z } |{z}
k1 c1
1 − β3
k3 = k2 − c2 = k0 − c0 − βc0 − β 2 c0 = k0 − (1 + β + β 2 )c0 = k0 − c0
| {z } |{z} 1−β
k2 c2

...
1 − β T +1
kT +1 = k0 − c0
1−β
From this derivation we see that the Euler equation does not uniquely deter-
mine the path of {kt }. Only if we add the transversality condition that the
cake must be completely eaten in period T , i.e. kT +1 = 0, can we solve the
problem uniquely:
1−β
c0 = k0 .
1 − β T +1
For T → ∞, we get:

c0 = (1 − β)k0

which implies

ct = (1 − β)kt .

The last equation can be interpreted as the optimal policy rule.

12
 The value of the cake when T → ∞ is thus given by
∞
X ∞
X
V (k0 ) = V (c0 , c1 , . . . ) = β t ln ct = β t ln (β t c0 )
t=0 t=0
∞ ∞
X
t
X β ln β ln c0
= ln β tβ + ln c0 βt = 2
+
t=0 t=0
(1 − β) 1−β
β ln β ln((1 − β)k0 )
= 2
+
(1 − β) 1−β
β ln β ln(1 − β) 1
= 2
+ + ln k0
(1 − β) 1−β 1−β

4.3 The Neoclassical Growth Model

The simple neoclassical growth model (Ramsey model) has become the work
horse of modern macroeconomics. We will start with a simplified determin-
istic version with inelastic labor supply. For this purpose consider a closed
economy whose production possibilities are described by a neoclassical pro-
duction function F in capital k and labor n:

F : R+ × R+ −→ R+
(k, n) −→ y = F (k, n)

It is assumed that F is continuously differentiable, strictly monotonically

increasing in both arguments, homogenous of degree one (constant returns
to scale), and strictly quasi-concave. Moreover, we assume that both capital
and labor are essential, i.e. F (0, n) = F (k, 0) = 0, and that the Inada
∂F (k,n) ∂F (k,n)
conditions hold for capital, i.e. limk→0 ∂k
= ∞ and limk→∞ ∂k
= 0.
The labor force is constant over time and normalized to one: 0 ≤ nt ≤ 1.
The output yt can either be consumed or invested:

ct + it ≤ yt = F (kt , nt ).

If output is invested, the capital stock in the next period changes according
to the transition equation:

kt+1 = (1 − δ)kt + it

13
where the depreciation rate δ ∈ [0, 1]. Thus we have the following resource
constraint:
ct + kt+1 − (1 − δ)kt ≤ F (kt , nt )

The economy is populated by a continuum of identical households. The

representative household lives forever and derives utility from consumption
only according to the additive utility functional:
∞
X
U (c0 , c1 , . . . ) = β t U (ct ), 0<β<1
t=0

where β is the subjective discount factor. The period utility function U

is assumed to be twice continuously differentiable, strictly increasing, and
strictly concave with limc→0 U 0 (c) = ∞. Thus the decision problem of the
representative household is given by the following intertemporal maximiza-
tion problem:
∞
X
β t U (ct ) −→ max
{ct ,kt+1 ,nt }
t=0

s.t. ct + kt+1 − (1 − δ)kt ≤ F (kt , nt )

ct ≥ 0, 0 ≤ nt ≤ 1, kt ≥ 0
k0 > 0 given.

This problem can be simplified through the following arguments:

• There is always a positive value for output because U 0 > 0. This implies
that the resource constraint must hold with equality. Thus choosing
kt+1 automatically implies a value for ct given kt . We can therefore
choose kt+1 instead of ct as the control variable.

• Because labor is supplied inelastically and because the marginal prod-

uct of labor is always positive, nt = 1 for all t.

• Allowing old capital to be retransformed into consumption goods, the

resource constraint can be rewritten as

ct + kt+1 = F (kt , 1) + (1 − δ)kt = f (kt )

14
With these modification the above maximization problem can be rewritten
as
∞
X
β t U (f (kt ) − kt+1 ) −→ max
{kt+1 }
t=0

s.t. 0 ≤ kt+1 ≤ f (kt )

k0 > 0 given.

Note that f inherits the properties of F . Given the properties of f , it is easy

to see that there exists a maximal sustainable capital stock kmax > 0. Setting
ct = 0 forever, this maximal capital is determined by equation k = f (k)
which has a unique positive solution, given the properties of f . The state
variable for this economy is obviously kt so that the state space is the left
open interval X = (0, max{k0 , kmax }]. This space including 0 may be taken
to be the control space, i.e. C = [0, max{k0 , kmax }]. The correspondence
describing the control constraint is Γ(kt ) = [0, f (kt )] ⊆ C. Alternatively, we
can write the maximization problem as a functional equation by applying
the dynamic programming technique:

V (k) = sup {U (f (k) − k 0 ) + βV (k 0 )}

k0 ∈Γ(k)

This functional equation can be solved by at least three methods. We illus-

trate these methods for the case of a logarithmic utility function, a Cobb-
Douglas production function, F (k, n) = Ak α n1−α , and complete depreciation
within the period (i. e. δ = 1). This is one of two specifications for which an
explicit analytic solution is available.3

Solution method 1: Value function iteration As will be justified later

on, one can approach the solution by iteration. Having found a value function
Vj in the j-th iteration, the Bellman equation delivers a new value function
Vj+1 :
Vj+1 (k) = sup {U (f (k) − k 0 ) + βVj (k 0 )} .
k0 ∈Γ(k)

3
The other specification involves a quadratic objective function coupled with linear
constraints (see Section 4.4).

15
Most importantly, we can start the iteration with an arbitrary function V0 .
Take, for example, the function V0 (k) = 0. Then the Bellman equation
simplifies to
V1 (k) = sup {U (f (k) − k 0 )} .
k0 ∈Γ(k)

The maximization problem delivers k 0 = µ0 (k) = 0. Inserting the solution

(1) (1)
into the Bellman equation leads to V1 (k) = ln A + α ln k = v0 + v1 ln k.
The equation for V2 is then given by
V2 (k) = sup {U (f (k) − k 0 ) + βV1 (k 0 )} .
k0 ∈Γ(k)

The first order condition of the maximization problem delivers:

1 1 0 αβ
= βα =⇒ k = µ 1 (k) = A kα.
Ak α − k 0 k0 1 + αβ
Inserting this optimal value into the objective function yields:
(1 + αβ)A k α − αβ A k α
V2 (k) = ln + β ln A + αβ ln k 0
1 + αβ
A Aαβ
= ln + β ln A + αβ ln + α(1 + αβ) ln k
1 + αβ 1 + αβ | {z }
(2)
=v1
(2) (2)
= v0 + v1 ln k
The equation for V3 is then given by
V3 (k) = sup {U (f (k) − k 0 ) + βV2 (k 0 )} .
k0 ∈Γ(k)

The first order condition for the maximization problem delivers:

1 1 0 αβ + α2 β 2
= αβ(1 + αβ) =⇒ k = µ 2 (k) = A kα.
Ak α − k 0 k0 1 + αβ + α2 β 2
This leads to the following value function V3 :
A Aαβ
V3 (k) = β ln + β 2 ln A + αβ 2 ln
1 + αβ 1 + αβ
A Aαβ(1 + αβ)
+ ln 2 2
+ αβ(1 + αβ) ln
1 + αβ + α β 1 + αβ + α2 β 2
+ α(1 + αβ + α2 β 2 ) ln k
| {z }
(3)
=v1
(3) (3)
= v0 + v1 ln k.

16
 The structure of the solutions V0 , V1 , V2 , V3 , and so on, leads to the
conjecture that, by taking j to infinity, the value function is (log-)linear with
V (k) = v0 + v1 ln k where v1 is given by:
(j) α
v1 = lim v1 =
j→∞ 1 − αβ

Similarly, the policy function k 0 = µj (k) converges to

Pj
0 (αβ)i α
k = lim µj (k) = lim Pji=1 Ak = αβA k α .
i
i=0 (αβ)
j→∞ j→∞

Solution method 2: Guess and verifying With this solution method

we guess the type of solution and verify that it is indeed a solution. This
is a feasible strategy because from theoretical considerations we know that
there is a unique solution. Take the guess V (k) = E + F ln k with constants
E and F yet to be determined. The maximization problem then becomes:

ln(A k α − k 0 ) + βV (k 0 ) = ln(A k α − k 0 ) + βE + βF ln k 0 −→ max

0
.
k

The first order condition is

1 1 Fβ
0
= βF 0 =⇒ k 0 = µ(k) = A kα.
Ak α −k k 1 + Fβ
Inserting this optimal value into the objective function yields:
   
α Fβ α Fβ α
E + F ln k = ln Ak − A k + Eβ + F β ln Ak
1 + Fβ 1 + Fβ

This implies
A Fβ A
E + F ln k = ln + α ln k + Eβ + F β ln + F αβ ln k
1 + Fβ 1 + Fβ
Equating the coefficients of ln k leads to following equation for F :
α
F = α + F αβ =⇒ F =
1 − αβ

Thus, k 0 = µ(k) = αβ A k α as before.

17
 Inserting this results into the objective function leads to:
   
αβ αβ α
(1 − β)E = ln(1 − αβ) + ln A + ln(Aαβ) + α 1 + − ln k
1 − αβ 1 − αβ 1 − αβ
αβ
= ln(1 − αβ) + ln A + ln(Aαβ)
1 − αβ
=⇒
 
1 αβ
E= ln(1 − αβ) + ln A + ln(Aαβ)
1−β 1 − αβ
Note that first order condition delivers a first order linear difference equa-
tion in kt :
Fβ
kt+1 = A ktα = αβ A ktα
1 + Fβ
ln kt+1 = ln Aαβ + α ln kt

As 0 < α < 1 this is a stable difference equation.

Solution method 3: Difference equation The first order condition for

the maximization problem is:

U 0 (f (k) − k 0 ) = βV 0 (k 0 ).

From the Benveniste-Scheinkman formula we get:

∂k 0 ∂k 0
V 0 (k) = U 0 (f (k) − k 0 )f 0 (k) − U 0 (f (k) − k 0 ) + βV 0 (k 0 )
∂k ∂k
= U 0 (f (k) − k 0 )f 0 (k).

We therefore get the following second order difference equation:

U 0 (f (k) − k 0 ) = βU 0 (f (k 0 ) − k 00 )f 0 (k 0 )
1 1 α−1
α
=β α Aα kt+1
Akt − kt+1 Akt+1 − kt+2
α
kt+1 Aαβ kt+1
=
Aktα − kt+1 α
Akt+1 − kt+2
1 αβ
α
= α
.
(Akt /kt+1 ) − 1 1 − (kt+2 /Akt+1 )

18
 difference equation for savings rate yt with αβ = 0.3
2
45°−line

1.5
1 + αβ
SS
1

SS
0.5
αβ yt+1 = 1+αβ − (αβ/yt)
yt+1

−0.5

−1

−1.5

−2
0 0.5 1 1.5 2
y
t

Figure 4: Ramsey model

kt+2
If we set yt+1 = α ,
A kt+1
we get the following non-linear difference equation:

αβ
yt+1 = (1 + αβ) − .
yt
This equation admits two steady states: αβ and 1. The second steady state
cannot be optimal because it implies that there is no consumption. As the
second steady state is unstable, we must have yt = αβ which implies kt+1 =
Aαβ ktα . The situation is depicted in figure 4

4.4 The Linear Regulator Problem

Another prototypical case is when the objective function is quadratic and
the law of motion linear. Consider the following general setup know as the

19
Optimal Linear Regulator Problem:
∞
X
V (x0 ) = − β t (x0t Rxt + u0t Qut ) → max, 0<β<1 (4.10)
ut
t=0

xt+1 = Axt + But , x0 given,

where xt denotes the n-dimensional state vector and ut the k-vector of con-
trols. R is positive semidefinite symmetric n × n matrix and Q is a positive
definite symmetric k × k matrix. A and B are n × n, respectively n × k ma-
trices. Note that the problem has been simplified by allowing no interaction
between xt and ut .
The Bellman equation can thus be written as

V (xt ) = max −{(x0t Rxt + u0t Qut ) + βV (xt+1 )} s.t. xt+1 = Axt + But ,
ut

where V (xt ) denotes the value of x in period t. We may solve this equation
by guessing that V (x) = −x0 P x for some positive semidefinite symmetric
matrix P . Using this guess and the law of motion the Bellman equation
becomes:

−x0 P x = max −{x0 Rx + u0 Qu + β(Ax + Bu)0 P (Ax + Bu)}

u

The first order condition of the maximization problem of the right hand side
is4
(Q + βB 0 P B)u = −βB 0 P Ax.
Thus, we get the feedback rule:

u = −F x = −β(Q + βB 0 P B)−1 B 0 P Ax.

Inserting this rule into the Bellman equation and rearranging terms leads to

P = R + βA0 P A − β 2 A0 P B(Q + βB 0 P B)−1 B 0 P A.

This equation is known as the Riccati equation. It can be solved by iteration:

Pj+1 = R + βA0 Pj A − β 2 A0 Pj B(Q + βB 0 Pj B)−1 B 0 Pj A

∂x0 Ax
4
Here we used the following rules for matrix derivatives: ∂Ax
∂x = A0 , ∂Ax
∂x0 = A, ∂x =
0
(A + A )x.

20
starting from P0 = 0. A sufficient condition for the iteration to converge is
that the eigenvalues of A are absolutely strictly smaller than one.
If the optimal rule is inserted into the law of motion, we obtain the closed-
loop solution:
xt+1 = (A − BF )xt

This difference equation is stable if the eigenvalues of A − BF are strictly

smaller than one in absolute value. There is a large literature investigating
the conditions on R, Q, A, and B such a stable closed loop solution obtains.
Basically, two conditions must be met. First, A and B must be such that
the controller can drive down xt to zero starting from any initial condition
x0 . Second, the R must be such that the controller wants to drive xt to zero.
As an illustration consider the following simple numerical example with
β = 1, A = 1, B = 1, Q = 1, and R = 1. It is instructive to note that this
specification allows for the possibility that some plans yield a limit to the
infinite sum in (4.10) equal to −∞. Such plans are, however, never optimal
as they are dominated by plans with a finite limit. The general principles
of dynamic programming outlined in Chapter 5 give a theoretical account
of this possibility (see in particular Assumption (5.2)). Given the scalar
specification, the Riccati equation can be solved analytically:5

P2 √
P =1+P − =⇒ P = (1 + 5)/2 ≈ 1.618
1+P
The negative solution can be disregarded because we are looking for positive
solutions. This implies that
P
u = −F x = − x
1+P
Inserting this in the law of motion for xt gives the closed-loop solution:
 
P 1
xt+1 = 1 − xt = xt
1+P 1+P

This solution is stable despite A = 1 and β = 1. Finally, V (x) = P x2 < ∞.

5
It can also be shown that the Riccati difference equation is stable although A = 1 and
β = 1.

21
5 Principles of Dynamic Programming: the
deterministic case
As we have seen, many dynamic economic problems can be cast in either of
the two following forms: a sequence problem (SP) or a functional (Bellman)
equation (FE).
∞
X
(SP) sup β t F (xt , xt+1 )
{xt+1 }∞
t=0 t=0

s.t. xt+1 ∈ Γ(xt ), x0 given

(FE) V (x) = sup {F (x, y) + βV (y)}

y∈Γ(x)

where 0 < β < 1. The optimization problem is thus completely specified

(X , Γ, F, β) where X is the state space, Γ a correspondence from X into X
describing the feasibility constraints, a period return function F defined on
A, the graph of Γ (i.e. A = {(x, y) ∈ X × X : y ∈ Γ(x)}), and a subjective
discount factor β with 0 < β < 1.
A sequence π = {xt } is called a plan. A plan is called feasible if xt+1 ∈
Γ(xt ) for all t. The set of all feasible plan starting in x0 is denoted by Π(x0 ).
Thus any π ∈ Π(x0 ) fulfills the constraints of (SP). In order for the problem
to make sense, the following two assumptions must hold:
Assumption 5.1. ∀x ∈ X , Γ(x) 6= ∅, in particular Π(x0 ) 6= ∅.
Assumption 5.2. ∀x0 ∈ X and ∀π ∈ Π(x0 ), limT →∞ Tt=0 F (xt , xt+1 ) ex-
P

ists. Limit values of ∞ and −∞ are possible.

Note that assumption 5.2 is fulfilled if F is bounded.
For any T = 0, 1, 2, . . . define the function UT : Π(x0 ) → R by
T
X
UT (π) = β t F (xt , xt+1 ).
t=0

This is just the partial sum of discounted returns for a feasible plan π. As-
sumption 5.2 allows to define a function U : Π(x0 ) → R by
U (π) = lim UT (π)
T →∞

22
where R may now include ∞ and −∞. By assumption 5.1, Π(x0 ) is not
empty and the objective function in (SP) is well defined. This allows us to
define the supremum function V ∗ : X → R by

V ∗ (x0 ) = sup U (π)

π∈Π(x0 )

This function is well defined and is the unique function which satisfies:

(i) If |V ∗ (x0 )| < ∞, V ∗ (x0 ) ≥ U (π) for all π ∈ Π(x0 ).

(ii) If |V ∗ (x0 )| < ∞, ∀ > 0 ∃π ∈ Π(x) : V ∗ (x0 ) ≤ U (π) + .

(iii) If |V ∗ (x0 )| = ∞, there exists a sequence of plans {πk } such that U (πk )
converges to ∞.

(iv) If |V ∗ (x0 )| = −∞, there exists a sequence of plans {πk } such that
U (πk ) converges to −∞.

Thus V ∗ satisfies the (FE),

(i) if |V ∗ (x0 )| < ∞, V ∗ (x0 ) ≥ F (x0 , y) + βV ∗ (y) for all y ∈ Γ(x0 );

(ii) |V ∗ (x0 )| = ∞, ∃πk ∈ Γ(x0 ) : limk→∞ F (x0 , yk ) + βV ∗ (yk ) = ∞;

(iii) |V ∗ (x0 )| = −∞, F (x0 , y) + βV ∗ (y) = −∞ for all y ∈ Γ(x0 ).

Lemma 5.1. If the problem (X , Γ, F, β) satisfies assumptions 5.1 and 5.2,

then
U (π) = F (x0 , x1 ) + βU (π 0 ) ∀π ∈ Π(x0 ) and ∀x0 ∈ X
where π 0 = (x1 , x2 , . . . ).

Proposition 5.1. If the problem (X , Γ, F, β) satisfies assumptions 5.1 and

5.2, V ∗ satisfies (FE).

Proposition 5.2. If the problem (X , Γ, F, β) satisfies assumptions 5.1 and

5.2, then for any solution V to (FE) which satisfies the transversality con-
dition
lim β t V (xt ) = 0
t→∞

for all π ∈ Γ(x0 ) and for all x0 ∈ X , V = V ∗ .

23
Remark 5.1. Proposition 5.2 implies that the solution to (FE) is unique.

A feasible plan π ∗ is optimal for x0 if U (π ∗ ) = V ∗ (x0 ).

Proposition 5.3. If the problem (X , Γ, F, β) satisfies assumptions 5.1 and

5.2, any optimal plan π ∗ which attains the supremum in (SP) satisfies

V ∗ (x∗t ) = F (x∗t , x∗t+1 ) + βV ∗ (x∗t+1 ).

Proposition 5.4. If the problem (X , Γ, F, β) satisfies assumptions 5.1 and

5.2, any optimal plan π ∗ which satisfies

V ∗ (x∗t ) = F (x∗t , x∗t+1 ) + βV ∗ (x∗t+1 )

and

lim sup β t V ∗ (xt ) ≤ 0

t→∞

attains the supremum in (SP).

Every nonempty correspondence µ : X → X with µ(x) ⊆ Γ(x) is called a

policy correspondence. If µ(x) is single valued, it is called a policy function.
The optimal policy correspondence µ∗ is defined by

µ∗ (x) = {y ∈ Π(x) : V ∗ (x) = F (x, y) + βV ∗ (y)}.

The dynamic programming approach is best understood if one views T

TV (x) = sup {F (x, y) + βV (y)}

y∈Γ(x)

as an operator on some function space. We will use the space of bounded

continuous functions from X into R, denoted by B(X ), as the appropriate
function space. This space is a Banach space with the supremum norm (i.e.
kf k = supx∈X f (x)). Convergence in this space is equivalent to uniform
convergence. A solution to (FE) is therefore nothing but a fixed-point of this
operator in B(X ):
V = TV

24
Note that all function in B(X ) satisfy the transversality condition. The idea
of using this operator view is the following. Suppose V0 has some property.
Then we ask if TV0 also has this property. If this is true Tt V0 has also this
property. But then V = limt→∞ Tt V0 will also share this property because of
uniform convergence. In order to make this idea workable, we have to place
additional assumption on our decision problem.

Assumption 5.3. X ⊆ Rm is convex. Γ is nonempty, compact, and contin-

uous.

Assumption 5.4. F : A → R is bounded and continuous and 0 < β < 1.

Theorem 5.1 (Solution of Bellman Equation). Under assumptions 5.3 and

5.4 the operator T defined as

(Tf )(x) = sup {F (x, y) + βV (y)}

y∈Γ(x)

has the following properties:

(i) T : B(X ) → B(X ).

(ii) T has exactly one fixed-point V .

(iii) ∀V0 ∈ B(X ), kTt V0 − V k ≤ β t kV0 − V k.

(iv) The policy correspondence is compact and u.h.c

Assumption 5.5. For all y, F (x, y) is strictly increasing in x.

Assumption 5.6. Γ is increasing, i.e. x ≤ x0 ⇒ Γ(x) ⊆ Γ(x0 ).

Theorem 5.2 (Properties of the Solution). Under assumptions 5.3, 5.4, 5.5
and 5.6, the solution V to (FE) is strictly increasing.

Assumption 5.7. F (x, y) is concave (with strict inequality if x 6= x0 ).

Assumption 5.8. Γ is convex in the sense that for any θ, 0 ≤ θ ≤ 1, and

x, x0 ∈ X , y ∈ Γ(x) and y 0 ∈ Γ(x0 ) implies θy + (1 − θ)y 0 ∈ Γ(θx + (1 − θ)x0 ).

This assumption excludes increasing returns to scale.

25
Theorem 5.3 (Further Properties of the Solution). Under assumptions 5.3,
5.4, 5.7, and 5.8 if V satisfies (FE) and the policy correspondence G is well
defined then

(i) V is strictly concave and

(ii) G is continuous and single-valued.

Theorem 5.4 (Solution found by Iteration). Let (X , Γ, F, β) satisfy assump-

tions 5.3, 5.4, 5.7, and 5.8 and let V satisfy (FE) with a well-defined policy
correspondence G. Denote by B0 (X ) the space of bounded, continuous, con-
cave functionals and let V0 ∈ B0 (X ). Define {(Vn , gn )} by

(i) Vn+1 = TVn , n = 0, 1, 2, . . . , and

(ii) gn (x) = arg maxy∈Γ(x) [F (x, y) + βVn (y)].

Then {gn } converges point wise to g. If X is compact, the convergence is

uniform.

Theorem 5.5 (Differentiability of Solution). Let (X , Γ, F, β) satisfy assump-

tions 5.3, 5.4, 5.7, and 5.8 and let V satisfy (FE) with a well-defined policy
correspondence G. In addition let F be continuously differentiable on the
interior of the graph of Γ. If x0 ∈ intX and g(x0 ) ∈ intΓ(x0 ), then V is
continuously differentiable at x0 with derivative given by

Vi (x0 ) = Fi (x0 , g(x0 )), i = 1, . . . , m.

The subscript i denotes the derivative with respect to the i-th element of
x0 ∈ Rm .

Remark 5.2. If V is twice continuously differentiable, the monotonicity of

g could be established by differentiating Fx (x, g(x)) = bV 0 (x). Although V is
continuously differentiable under fairly general conditions, conditions ensur-
ing that V is twice continuously differentiable are quite strong.

26
6 Principles of Dynamic Programming: the
stochastic case
In the following we will consider stochastic dynamic programming problems
of the following type:
 Z 
0 0
V (x, z) = sup F (x, y, z) + β V (y, z )Q(z, dz )
y∈Γ(x,z) Z

The major difference is that the agent(s) have to take stochastic shocks into
account. This implies that the supremum has to be taken with respect to
F plus the expected value of discounted future V . The shocks are governed
by a transition function Q. Under suitable assumptions on the shocks, the
required mathematical properties of the value function V are preserved under
integration. Thus, the results for the deterministic model carry over virtually
without change.
The state space is now the product space of the measurable spaces (X , X)
and (Z, Z), i.e. (S, S) = (X × Z, X × Z), describing the possible values of
the endogenous and exogenous state variables. Q is a transition function on
(Z, Z). The correspondence Γ describing the feasibility constraints is now
a correspondence from S into X . The graph of Γ is denoted by A, i.e.
A = {(x, y, z) ∈ X × X × Z : y ∈ Γ(x, z)}. F is again a one period return
function on A and β is the subjective discount factor with 0 < β < 1.

Assumption 6.1. X ⊆ Rm is closed and convex with Borel subsets X.

Assumption 6.2. One of the following conditions hold:

(i) Z is a countable set and Z is the σ-algebra containing all subsets of Z.

(ii) Z is a compact and convex subset of Rm with Borel subsets Z and the
transition function Q on (Z, Z) has the Feller property.

Assumption 6.3. The correspondence Γ : X ×Z → X is nonempty, compact-

valued and continuous.

Assumption 6.4. The period return function F : A → R is bounded and

continuous, and β ∈ (0, 1).

27
Theorem 6.1 (Solution of the Stochastic Bellman Equation). Let (X , X),
(Z, Z), Q, Γ, F , and β satisfy assumptions 6.1, 6.2, 6.3, and 6.4. Define the
operator T on B(S) by
 Z 
0 0
(Tf )(x, z) = sup F (x, y, z) + β f (y, z )Q(z, dz ) .
y∈Γ(x,z) Z

Then the operator T has the following properties:

(i) T : B(S) → B(S);

(ii) T has a unique fixed point V ;

(iii) ∀V0 ∈ B(S), kTn V0 − V k ≤ β n kV0 − V k , n = 1, 2, . . . ;

(iv) The policy correspondence G : S → X defined by

 Z 
0 0
G(x, z) = y ∈ Γ(x, z) : V (x, z) = F (x, y, z) + β V (x, z )Q(z, dz )
Z

is nonempty, compact-valued and u.h.c.

Assumption 6.5. ∀(y, z) ∈ X × Z, F (., y, z) : Ay,z → R is strictly increas-

ing.

Assumption 6.6. ∀z ∈ Z, Γ(., z) : X → X is strictly increasing, i.e. x ≤

x0 ⇒ Γ(x, z) ⊆ Γ(x0 , z).

Theorem 6.2 (Property of Solution: Stochastic Case). Let (X , X), (Z, Z),
Q, Γ, F , and β satisfy assumptions 6.1, 6.2, 6.3, 6.4, 6.5, and 6.6. If V
denotes the fixed point of T, then, ∀z ∈ Z, V (., z) : X → R is strictly
increasing.

Assumption 6.7. ∀z ∈ Z, F (., ., z) : Az → R satisfies: F (θ(x, y) + (1 −

θ)(x0 , y 0 ), z) ≥ θF (x, y, z)+(1−θ)F (x0 , y 0 , z), ∀θ ∈ (0, 1), ∀(x, y), (x0 , y 0 ) ∈ Az
and the inequality is strict if x 6= x0 .

Assumption 6.8. ∀z ∈ Z and ∀x, x0 ∈ X , y ∈ Γ(x, z) and y 0 ∈ Γ(x0 , z)

implies

θy + (1 − θ)y 0 ∈ Γ(θx + (1 − θ)x0 , z) for any θ ∈ [0, 1].

28
Theorem 6.3 (Further Properties of Solution: Stochastic Case). Let (X , X),
(Z, Z), Q, Γ, F , and β satisfy assumptions 6.1, 6.2, 6.3, 6.4, 6.7, and 6.8.
If V denotes the fixed point of T and G the policy correspondence, then,
∀z ∈ Z, V (., z) : X → R is strictly concave and G(., z) : X → X is a
continuous (single-valued) function.

Theorem 6.4 (Solution found by Iteration: Stochastic Case). Let (X , X),

(Z, Z), Q, Γ, F , and β satisfy assumptions 6.1, 6.2, 6.3, 6.4, 6.7, and 6.8.
If V denotes the fixed point of T and g = G the single-valued policy function,
then V ∈ C0 (S), the set of bounded continuous functions on S which are
weakly concave jointly in their first m arguments. Moreover, if V0 ∈ C0 (S)
and if {(Vn , gn )} is defined by

Vn = TVn−1 , n = 1, 2, . . .

and
 Z 
0 0
gn (x) = arg max F (x, y, z) + β Vn (y, z )Q(z, dz ) .
x∈Γ(x,z) Z

Then {gn } converges pointwise to g. If X and Z are both compact, then the
convergence is uniform.

Assumption 6.9. ∀z ∈ Z, F (., ., z) is continuously differentiable at (x, y) ∈

intAz .

Theorem 6.5 (Differentiability of Solution: Stochastic Case). Let (X , X),

(Z, Z), Q, Γ, F , and β satisfy assumptions 6.1, 6.2, 6.3, 6.4, 6.7, 6.8, and
6.9. If V ∈ C0 (S) denotes the unique fixed point of T and g = G the single-
valued policy function, then V (., z0 ) is continuously differentiable in x at x0
for x0 ∈ intX and g(x0 , z0 ) ∈ intG(x0 , z0 ) with derivatives given by

Vi (x0 , z0 ) = Fi (x0 , g(x0 , z0 ), z0 ), i = 1, 2, . . . , m.

The subscript i denotes the derivative with respect to the i-th element.

29
7 The Lucas Tree Model
This presentation follows closely the exposition given in Sargent (1987, chap-
ter 3). For details we refer to the original paper by Lucas (1978). Imagine an
economy composed of a large number of agents with identical preferences and
endowments. Each agent owns exactly one tree which is perfectly durable.
In each period, the tree yields a fruit or dividend dt . The fruit is the only
consumption good and is nonstorable. Denote by pt the price of the tree.
Then each agent is assumed to maximize
∞
X
E0 β t U (ct ), 0 < β < 1,
t=0

subject to

At+1 = Rt (At − ct ), A0 given.

Denoting the return Rt is given by Rt = (pt+1 + dt+1 )/pt the Euler equation
for this maximization problem is

pt U 0 (ct ) = βEt (pt+1 + dt+1 )U 0 (ct+1 ). (7.1)

Because all agents are identical and there is no satiation, every agent just
consumes this period’s dividends. Thus, we have in equilibrium

ct = dt .

Inserting the equilibrium condition into the Euler equation yields

U 0 (dt+1 )
pt = βEt (pt+1 + dt+1 ) (7.2)
U 0 (dt )

Iterating this equation forward in time and using the law of iterated expec-
tations, we get the solution
∞
"j−1 #
X Y U 0 (dt+i+1 )
pt = Et βj dt+j
j=1 i=0
U 0 (dt+i )

30
which simplifies to
∞
X U 0 (dt+j )
pt = Et βj dt+j . (7.3)
j=1
U 0 (dt )

The share price is the expected discounted stream of dividends, but with
time-varying and stochastic discount factors.
An interesting formula is obtained by taking U (c) = ln c. In this case,
the pricing formula (7.3) simplifies to
∞
X β
pt = Et β j dt = dt .
j=1
1−β

This is a simple asset-pricing function which maps the state of the economy
at time t, dt , into the price of an asset at time t.
In order for the conditional expectation in equation (7.1) to be well-
defined, it is necessary to impute to the representative agent a view about
the law of motion over time of dt and pt . The specification of an actual law
of motion for dt , which agents are supposed to know, and a perceived pricing
function that maps the history of dt into pt implies that a law of motion for
pt has been perceived. Given that Et in equation (7.1) is computed using
the perceived pricing function, equation (7.1) maps the perceived pricing
function into an actual pricing function. The notion of a rational expectation
equilibrium is that the actual pricing function equals the perceived pricing
function. In the following, we will exploit this notion in more detail.
Suppose that dividends evolve according to a Markov process with time-
invariant transition probability distribution function F (x0 , x) defined as

F (x0 , x) = P[dt+1 ≤ x0 |dt = x].

Conditional expectations are computed with respect to this law which is

supposed to be known by all agents.
Denote by st the number of shares held by the agent. The period budget
constraint is
ct + pt st+1 ≤ (pt + dt )st .

31
Lifetime utility is again given by
∞
X
E0 β t U (ct ), 0 < β < 1.
t=0

As explained in the previous paragraph, in order for the conditional expec-

tation to be well-defined we must specify a law of motion for the stock price:

pt = h(xt ).

This law of motion together with the evolution of xt defines the perceived
law of motion for the tree prices. The Bellman equation can then be written
as

V (s(h(x) + x)) = max U (s(h(x) + x) − h(x)s0 )
s0
Z 
0 0 0 0
+β V (s (h(x ) + x ))dF (x , x) . (7.4)

The first order condition for this optimization problem is

Z
h(x)U (s(h(x) + x) − h(x)s ) = β (h(x0 ) + x0 )V 0 (s0 (h(x0 ) + x0 ))dF (x0 , x).
0 0

Differentiating the Bellman equation (7.4) with respect to s, we obtain the

Benveniste-Scheinkman formula which reads as

(h(x) + x)V 0 (s(h(x) + x)) = (h(x) + x)U 0 (s(h(x) + x) − h(x)s0 )

or by simplifying

V 0 (s(h(x) + x)) = U 0 (s(h(x) + x) − h(x)s0 )

Combining the Benveniste-Scheinkman formula and the first-order condition

we get

h(x)U 0 (s(h(x) + x) − h(x)s0 )

Z
= β (h(x0 ) + x0 )U 0 (s0 (h(x0 ) + x0 ) − h(x0 )s00 )dF (x0 , x)

32
Defining w(x) by

w(x) = h(x)U 0 (s(h(x) + x) − h(x)s0 ),

we obtain
Z Z
0 0
w(x) = β w(x )dF (x , x) + β x0 U 0 (s0 (h(x0 ) + x0 ) − h(x0 )s00 )dF (x0 , x)

In equilibrium s = s0 = s00 = 1 because all agents and all trees are alike so
that consumption c(x) = s(h(x) + x) − h(x)s0 = x. Thus, we obtain the
functional equation in the unknown function w(x):
Z Z
w(x) = β w(x )dF (x , x) + β x0 U 0 (x0 )dF (x0 , x)
0 0
(7.5)

where w(x) = h(x)U 0 (x). Thus, once w(x) has been determined, the pricing
function h(x) can be recovered from w(x) = h(x)U 0 (x) as U 0 (x) is known.
Denote by g(x) the function g(x) = β x0 U 0 (x0 )dF (x0 , x), then we can
R

define the operator T as follows

Z
(Tw)(x) = β w(x0 )dF (x0 , x) + g(x)

In order to apply Blackwell’s sufficient condition for a contraction (see The-

orem B.3), we must show that T is a continuous mapping from the space of
continuous and bounded functionals, B(X ) with X ⊆ Rn , to itself. For this
purpose, we follow Lucas (1978) and assume that the utility function U (c) is
bounded by some number B and concave with U (0) = 0. The definition of
concavity of U implies that U (y) − U (x) ≤ U 0 (x)(y − x). Setting y = 0, then
leads to xU 0 (x) ≤ B. Thus,
Z
g(x) = β x0 U 0 (x0 )dF (x0 , x) ≤ βB < B.

Therefore g(x) is also bounded by B. Given some mild technical conditions

on the transition function F (x0 , x), it can be shown that g(x) is also continu-
ous. Thus, T is continuous. It is then easy to verify that T is monotone and
discounts. Let w1 , w2 ∈ B(X ) such that w1 (x) ≤ w2 (x) for all x ∈ X , then
Z
(Tw1 )(x) = β w1 (x0 )dF (x0 , x) + g(x)
Z
≤ β w2 (x0 )dF (x0 , x) + g(x) = (Tw2 )(x).

33
Thus, T is monotone. Furthermore, for any constant c ∈ R we have
Z
(T(w + c))(x) = β (w + c)(x0 )dF (x0 , x) + g(x)
Z Z
= β w(x )dF (x , x) + βc dF (x0 , x) + g(x) = (Tw)(x) + βc.
0 0

Thus, T discounts. Therefore T is a contraction so that a unique fixed point

exists. This fixed point is approached by iterations on w(x), respectively
h(x):
Z
w (j+1)
(x) = β w(j) (x0 )dF (x0 , x) + g(x),

or in terms of h(x)
Z
0
h(j+1)
(x)U (x) = β h(j) (x0 )U 0 (x0 )dF (x0 , x) + g(x) (7.6)

Thus, the operator T maps a perceived pricing function h(j) into an actual
pricing function h(j+1) . A rational expectation equilibrium is then nothing
but a fixed point of this mapping so that the actual pricing function equals
the perceived one.
As an example, take U (x) = ln x. With this utility function g(x) =
x0 (1/x0 )dF (x0 , x) = β. Guess that the solution is of the form h(x) = ax.
R
β
This implies that
Z
1 1
ax = β ax0 dF (x0 , x) + β = aβ + β
x x0

which results in
β β
a= giving the solution h(x) = x.
1−β 1−β
The Lucas tree model highlights a general principle for the construction
of asset-pricing models:

(i) Describe the preferences, technology, and endowments of a dynamic

economy. Find the equilibrium intertemporal consumption allocation.
This is typically the solution of a social planning problem.

34
 (ii) Open up a specific market for an asset which represents a specific claim
on future consumption. Assume no trade restrictions and derive the
corresponding Euler equation.

(iii) Equate the consumption in the Euler equation to the general equilib-
rium values found in (i). Then derive the associated asset price.

7.1 The Equity Premium Puzzle

A far reaching application of the Lucas tree model has been presented by
Mehra and Prescott (1985). They investigate the empirical implications for
the risk premium, i.e. the difference between the average return on equities
and the risk-free rate. They work with a discrete state-space version of the
model. Suppose that there are n ≥ 2 state of the world which determine the
growth rates of dividends:
dt+1 = xt+1 dt
where the gross growth rate xt+1 takes one of possible n values. i.e. xt+1 ∈
{σ1 , . . . , σn }. The transition from one state to the next is governed by a
Markov chain given by

P[xt+1 = σj |xt = σi ] = Pij > 0.

Pn
Clearly, j=1 Pij = 1. Given that all entries Pij are strictly positive the
chain is regular and therefore ergodic because for a given number of periods
each state of the chain can move to each state of the chain with a positive
probability.
The pricing of securities is ex-dividends and ex-interest. Assuming that
the utility function is of the constant relative risk aversion type, i.e. U (c) =
c1−α −1
1−α
, 0 < α < ∞, the application of the Euler equation (7.3) delivers the
price of equities:
∞ 0 ∞ α
jU
(dt+j ) j dt
X X
pt = Et β d t+j = Et β dt+j .
j=1
U 0 (dt ) j=1
dαt+j

The state is given by (dt , xt ) and dt+j = dt xt+1 . . . xt+j . This implies that the
price of the equity in state (d, i), p(d, i), is homogeneous in dt . The Euler

35
equation (7.2) becomes
n
X
p(d, i) = β Pij (σj d)−α [p(σj d, j) + σj d]dα .
j=1

As the price is homogeneous in the dividends, we can write p(d, i) as p(d, i) =

wi d and the above equation can be written as
n
X
wi d = β Pij σj−α [wj σj d + σj d]
j=1

which simplifies to
n
X
wi = β Pij σj1−α [wj + 1].
j=1

This is a system of n equations in n unknowns wi . This equation system can

be written in matrix form as

w = βP Σw + βP Σi

which results in the solution

w = β(In − βP Σ)−1 P Σi

where i is a vector of ones and where Σ is a diagonal matrix with diagonal

entries σj1−α , j = 1, . . . , n.
In order to bring the model to the data, we compute the mean return
for equities, Re , and for the risk-free rate, Rf . Given that the economy is in
state i the expected return on the equity is
n
X p(σj d, i) + σj d − p(d, i) σj (wj + 1)
Rie = e
Pij rij , e
where rij = = − 1.
j=1
p(d, i) wi

Thus, the average return on equity is

n
X
Re = πi Rie
j=i

36
where π = (π1 , . . . , πn ) is the stationary distribution for P = (Pij ), i.e. π
solves π = πP .
The risk-free asset pays one unit of consumption irrespectively of the
state. Thus, the price of this asset, pf (d, i) in state (d, i) is
n
X Xn
pf (d, i) = β Pij (σj d)−α dα = β Pij (σj )−α
j=1 j=1

Thus, the return is state i is

1
Rif = − 1.
pfi
This gives the average return on the risk-free asset as
Xn
f
R = πi Rif .
i=1

In their paper Mehra and Prescott (1985) calibrate this model to the U.S.
data over the period 1889 to 1978 setting n = 2. They find
! ! !
0.43 0.57 σ1 1.054
P = and = .
0.57 0.43 σ2 0.982
They then compute the combinations of the average risk-free rate and the
average risk-premium Re − Rf for values of the risk aversion parameter α
ranging from small positive numbers to a maximum of 10 and for the discount
factor β ranging from 0.925 to 0.9999. This delineates an admissible region
reproduced in Figure 5. The empirical average over this period for the risk
premium is 6.18 percent and 0.80 percent for the risk-free rate. These values
are way of the admissible region, the model is clearly incompatible with
the data. This result turned out to be very robust and is since then called
the equity premium puzzle. This puzzle is still not completely resolved and
remains an active area of research (see the survey by Mehra and Prescott,
2003).

A Some Topological Concepts

Definition A.1. The set X ⊆ Rn is convex iff
θx + (1 − θ)y ∈ X for all x, y ∈ X and 0 < θ < 1.

37
 1.4

1.2
average risk premium Re−Rf

1 US economy 1889 to 1978:

real return on riskless security: 0.80 percent
risk premium: 6.18 percent
0.8

0.6
admissible region
0.4

0.2

0
0 0.5 1 1.5 2 2.5 3 3.5 4
average risk−free rate

Figure 5: The Equity Premium Puzzle

Definition A.2. Let f be a function from X ⊆ Rn to R where X is a convex

set. Then f is convex on X iff

f (θx + (1 − θ)y) ≤ θf (x) + (1 − θ)f (y) for all x, y ∈ X and 0 < θ < 1.

The function f is said to be strictly convex if the inequality above is strict.

The function is said to concave, respectively strictly concave, if the inequality
is reversed.

Definition A.3. A compact-valued correspondence Γ : X → Y is upper

hemi-continuous (u.h.c.) at x if Γ(x) is nonempty and if, for every sequence
xn → x and every sequence {yn } with yn ∈ Γ(xn ) for all n, there exists a
convergent subsequence of {yn } whose limit point y is in Γ(x).

Definition A.4. The function f : X ⊆ Rn → R where X is convex is

quasi-concave iff

f (y) ≥ f (x) ⇒ f (θy + (1 − θ)x) ≥ f (x) for all x, y ∈ X and 0 ≤ θ ≤ 1.

38
The function f is said to be strictly quasi-concave iff

f (y) ≥ f (x) ⇒ f (θy + (1 − θ)x) > f (x)

for all x, y ∈ X with x 6= y and 0 ≤ θ ≤ 1.

Remark A.1. f is quasi-concave iff {x ∈ X : f (x) ≥ α} is convex for all

α ∈ R.

Remark A.2. f is concave implies that f is quasi-concave. The converse is

not true.

Definition A.5. A metric (distance) on some set X is a mapping d : X ×

X → [0, ∞) such that

(i) d(x, y) = 0 if and only if x = y (definiteness)

(ii) d(x, y) = d(y, x) (symmetry)

(iii) d(x, y) ≤ d(x, z) + d(z, y) (triangle inequality)

for all x, y, z ∈ X . A metric space is a pair (X , d) with X being a set and d

being a metric on it.

Definition A.6. A sequence {xn }n∈N in a metric space (X , d) is called a

cauchy sequence if d(xn , xm ) → 0 as n, m → ∞, i.e. if

∀ > 0 ∃N ∈ N such that ∀n, m ≥ N : d(xn , xm ) < .

Lemma A.1. Let (X , d) be a metric space. Then the following assertions

hold:

(i) Each convergent sequence in X is a Cauchy sequence.

(ii) Each Cauchy sequence is bounded.

(iii) If a Cauchy sequence has a convergent subsequence, then it converges.

Definition A.7. A metric d on a set X is called complete if every Cauchy

sequence converges. A metric space (X , d) is called complete if d is a com-
plete metric on X .

Theorem A.1. The Euclidean metric on Rn is complete.

39
B Fixed Point Theorems
Definition B.1. Let (X , d) be a metric space, then an operator T : X → X
is a contraction mapping (with modulus β) if there exists a β ∈ (0, 1) such
that
d(Tx, Ty) ≤ βd(x, y) for all x, y ∈ X .

Theorem B.1. If T is a contraction mapping then T is uniformly continuous

on X , and thus continuous

Proof. For any  > 0 chose δ = , then for all x, y with d(x, y) < δ
d(Tx, Ty) ≤ βd(x, y) < d(x, y) < δ = . Thus, T is uniformly continuous
and therefore also continuous.

Theorem B.2 (Contraction Mapping Theorem). If (X , d) is a complete

metric space and T : X → X is a contraction mapping with modulus β then

(i) T has exactly one fixed point x∗ ∈ X ;

(ii) For all x0 ∈ X , d(Tn x0 , x∗ ) ≤ β n d(x0 , x∗ ), n = 0, 1, 2, . . .

Proof. Uniqueness: Suppose x and y are two fixed points of T, then

0 ≤ d(x, y) = d(Tx, Ty) ≤ βd(x, y)

which implies by β < 1 that d(x, y) = 0, therefore x = y.

Existence: Fix x0 ∈ X and define xn as Tn x0 . The the triangular inequality
implies for any k ∈ N

d(x0 , xk ) ≤ d(x0 , x1 ) + d(x1 , x2 ) + · · · + d(xk−1 , xk )

= d(x0 , x1 ) + d(Tx0 , Tx1 ) + · · · + d(Tk−1 x0 , Tk−1 x1 )
1 − βk d(x0 , x1 )
≤ (1 + β + . . . β k−1 )d(x0 , x1 ) = d(x0 , x1 ) ≤ .
1−β 1−β
Hence for any m, n ∈ N with m = n + k, we obtain
βn
d(xn , xm ) = d(Tn x0 , Tn xk ) ≤ β n d(x0 , xk ) ≤ d(x0 , x1 ).
1−β

40
 45o-degree line
b

Tx
x*

a
a x* b

Figure 6: Contraction Mapping

Thus, {xn } is Cauchy sequence because β n → 0 as n → ∞. As X is a com-

plete metric space, the sequence converges to some x∗ ∈ X , i.e. limn→∞ xn =
x∗ . Finally, the continuity of T implies that x∗ is fixed point:

Tx∗ = T( lim xn ) = lim Txn = lim xn+1 = x∗ .

n→∞ n→∞ n→∞

Remark B.1. Let X be a closed interval [a, b] and d(x, y) = |x − y|. Then
T : X → X is a contraction mapping if for some β ∈ (0, 1),

|Tx − Ty|
≤ β < 1, for all x, y ∈ X with x 6= y
|x − y|

Thus, T is a contraction mapping if its slope is uniformly smaller than one

in absolute value (see Figure 6).

Theorem B.3 (Blackwell’s sufficient condition). Let X ⊆ Rn and B(X ) the

space of bounded functionals with the sup norm. If an operator T : B(X ) →
B(X ) satisfies

41
 (i) (monotonicity): f, g ∈ B(X ) with f (x) ≤ g(x) for all x ∈ X implies
Tf (x) ≤ Tg(x);

(ii) (discounting): There exists β ∈ (0, 1) such that for any function which
is constant to some value c ≥ 0, [T(f + c)](x) ≤ (Tf )(x) + βc, for all
f ∈ B(X ), x ∈ X ,

then T is a contraction with modulus β.

Proof. For any f, g ∈ B(X ), f ≤ g + kf − gk. Applying properties (i) and

(ii) leads to
Tf ≤ T(g + kf − gk) ≤ Tg + βkf − gk.

Reversing the role of f and g

Tg ≤ T(f + kf − gk) ≤ Tf + βkf − gk.

Combining the two inequalities, we obtain kTf −Tgk ≤ βkf −gk as required.

Theorem B.4 (Brouwer’s Fixed Point Theorem). Any continuous map of a

nonempty compact convex subset of a finite-dimensional normed space into
itself has at least one fixed point.

An illustration of Brouwer’s fixed point theorem is given in Figure 7 where

f : [a, b] → [a, b]. In this simple example it is easy to convince oneself that
Tx must cross the 45◦ –line if we go from the left edge of the square to the
right edge. Mathematically speaking, define g(x) = f (x) − x) then g(a) ≥ 0
and g(b) ≤ 0. Hence by the mean value theorem, there exists ξ such that
g(ξ) = 0 which is equivalent to f (ξ) = ξ.

Definition B.2. Let (Z, Z) be a measurable space. A transition function is

a function Q : Z × (Z) → [0, 1] such that

(i) Q(z, .) is a probability measure on (Z, Z) for all z ∈ Z

(ii) Q(., A) is a Z-measurable function for all A ∈ Z.

42
 45o-degree line
b

x*
Tx

a
a x* b

Figure 7: Illustration of Brouwer’s Fixed Point Theorem

Associated to any transition function Q is an operator T, called Markov

operator : Z
(Tf )(z) = f (z 0 )Q(z, dz 0 )

This operator maps the space of nonnegative, Z-measurable, extended real-

valued functions into itself. This operator is said to have the Feller property
if T is an operator on C(Z), the space of bounded continuous functions on
Z.

43
C Derivation of Equations Characterizing the
Reservation Wage
In this appendix we derive equations (4.3), (4.4), and (4.5). From equa-
tion (4.2) we get
Z w Z ∞
w w w0
= dF (w ) + dF (w0 )
1−β 1−β 0 1−β w
Z w Z ∞
w 0 w
=c+β dF (w ) + β dF (w0 )
0 1−β w 1 − β
Z w Z ∞
0 1
=⇒ w dF (w ) − c = (βw0 − w)dF (w0 )
0 1−β w

R∞
adding w w
dF (w0 ) on both sides leads to
Z ∞ Z ∞
β 0 0
w−c= (w − w)dF (w ) = β V (w0 − w)dF (w0 )
1−β w w

These are equations (4.3) and (4.4). Finally,

Z ∞
β
w−c= (w0 − w)dF (w0 )
1−β w
Z w Z w
β 0 0 β
+ (w − w)dF (w ) − (w0 − w)dF (w0 )
1−β 0 1−β 0
Z ∞ Z w
β 0 0 β
= (w − w)dF (w ) − (w0 − w)dF (w0 )
1−β 0 1−β 0
Z w
β β β
= Ew − w− (w0 − w)dF (w0 )
1−β 1−β 1−β 0
Z w
w − (1 − β)c = βEw − β (w0 − w)dF (w0 )
0

integrating by parts
Z w
w − c = βE(w − c) + β F (w0 )dw0
0

This is equation (4.5).

44
D Mean Preserving Spreads
Consider a class of distribution functions Fr (x) indexed by r with support
included in [0, B], 0 < B < ∞, and having the same mean. Note that if a
random variable X is distributed according to Fr then its mean EX is given
RB
by EX = 0 [1 − Fr (x)] dx.6 Therefore, if Fr and Fs have the same mean,
then Z B
[Fr (x) − Fs (x)] dx = 0.
0
Two distributions Fr and Fs are said to satisfy the single-crossing property
if there exists x̄ such that

Fs (x) − Fr (x) ≤ 0 (≥ 0) when x ≥ x̄ (x ≤ x̄).

If the two distributions satisfy the above two properties, we can regard Fs as
being obtained from Fr by shifting probability mass towards the tail of the
distribution keeping the mean constant. The two properties imply
Z y
[Fs (x) − Fr (x)] dx ≥ 0 0 ≤ y ≤ B.
0

Rothschild and Stiglitz (1970) regard the move from Fr to Fs as defining a

“mean-preserving increase in spread”. Figure 8 illustrates the single–crossing
property where Fs (x) is viewed as being more riskier than Fr (x).
If Fr is differentiable with respect to r, we can give an alternative char-
acterization of a mean-preserving increase in risk:
Z y
∂Fr (x)
dx ≥ 0, 0≤y≤B
0 ∂r
R B r (x)
with 0 ∂F∂r dx = 0.
6
Rb
This can be seen applying the integration–by–parts formula: a
(1 − F )(x) dx = (1 −
Rb Rb Rb
F (x))|ba − a x d(1 − F ) = − a x d(1 − F ) = a x dF = EX.

45
 1
Fr(x)

Fs(x)

0 x B x

Figure 8: Two distributions satisfying the single–crossing property

E The Riemann-Stieltjes Integral

The usual Riemann integral is defined as
Z b Xn
f (x)dx = lim f (ξi )(xi − xi−1 )
a n→∞
i=1

where f is a real and bounded function defined on some interval [a, b]. The
x0 , x1 , . . . , xn partition the interval [a, b] such that a = x0 < x1 < . . . <
xn−1 < xn = b. The ξi are arbitrary numbers from the interval [xi−1 , xi ]. Note
that the limit does not exist for every function. Thus, there are functions
which are not Riemann integrable. Monoton or continuous functions are
Riemann integrable.
This definition can be generalized to the so-called Riemann-Stieltjes inte-
gral by replacing (xi −xi−1 ) by (g(xi )−g(xi−1 )) where g is some real bounded
function on [a, b]:
Z b n
X
f (x)dg(x) = lim f (ξi ) (g(xi ) − g(xi−1 )) .
a n→∞
i=1

46
The Riemann-Stieltjes integral exists if f is continuous and g is a function
of bounded variation.7 For g(x) = x, we get the Riemann integral above.
Moreover, if g is continuously differentiable with derivative g 0 ,
Z b Z b
f (x)dg(x) = f (x)g 0 (x)dx.
a a

The Riemann-Stieltjes integral is particularly relevant when g is a cu-

mulative distribution function. Such functions are right-continuous, nonde-
creasing with existing left-hand limits. Moreover, they can have countably
many discontinuity points (jumps) as will be the case, for example, with a
discrete probability distribution. Thus, the expectation of a random variable
X with distribution function F can be written as
Z ∞ ( R∞
−∞
xf (x)dx, if F has density f ;
EX = xdF (x) = P
−∞ i xi p i , pi = P[X = xi ].

In the latter case F is a step function.

7
A function g on [a, b] is called a function of bounded variation if there exists
a constant M > 0 such that for all partitions x0 , x1 , . . . , xn of the interval [a, b]
Pn
i=1 |g(xi ) − g(xi−1 )| ≤ M . Continuous and monoton functions are of bounded vari-
ation.
47
References
F. Kydland and E. C. Prescott. Rules rather than discretion: The inconsis-
tency of optimal plans. Journal of Political Economy, 85:473–492, 1977.

R. E. Lucas, Jr. Asset prices in an exchange economx. Econometrica, 46:

1426–1445, 1978.

R. Mehra and E. C. Prescott. The equity premium puzzle. Journal of Mon-

etary Economics, 15:145–161, 1985.

R. Mehra and E. C. Prescott. The equity premium in retrospect. In M. H.

George M. Constantinides and R. M. Stulz, editors, Handbook of the Eco-
nomics of Finance, chapter 14, pages 887–936. North-Holland, Amster-
dam, Netherlands, 2003.

M. Rothschild and J. Stiglitz. Increasing risk I: A definition. Journal of

Economic Theory, 2:225–243, 1970.

T. J. Sargent. Dynamic Macroeoconomic Theory. Harvard University Press,

Cambridge, Massachusetts, 1987.

N. L. Stokey and J. Robert E. Lucas. Recursive Methods in Economic Dy-

namics. Harvard University Press, Cambridge, Massachusetts, 1989.

48