Aio PDF
Aio PDF
This
should be taken into account when using and working with them.
Klaus Wälde
These lecture notes are permanently being updated. The most recent version
is available at www.waelde.com/aio
@BOOK{Waelde06,
author = {Klaus Wälde},
year = 2006,
title = {Applied Intertemporal Optimization},
publisher = {Lecture Notes, University of Würzburg},
address = {Available at www.waelde.com/aio}
}
Acknowledgments
This (soon-to-be) book is the result of many lectures given in various places,
including the Bavarian Graduate Program in Economics, the Universities of Dort-
mund, Dresden, Frankfurt, Louvain-la-Neuve, Munich and Würzburg and the Euro-
pean Commission in Brussels. I would therefore …rst like to thank students for their
questions and discussions. These conversations showed what the most important as-
pects in understanding model building and in formulating maximization problems
are. I also pro…ted from many discussions with economists and mathematicians from
many places. I especially would like to thank Ken Sennewald for many insights into
the more subtle issues of stochastic continuous time processes. I hope I succeeded in
incorporating these insights into this book.
The basic structure of this book is replicated here:
1 Introduction
This book can be used for a …rst year master class, a PhD course or for studying
for oneself. In fact, given the detailed step-by-step approach to problems, it should be
easier to understand solutions with this text than with other usually more condensed
presentations.
A …rst year master class on economic methods (or third year undergraduate stud-
ies, i.e. …rst semester “Hauptstudium”) would cover part I and part II, i.e. maxi-
mization and applications in discrete and continuous time under certainty. A lecture
series of 14 times 90 minutes (corresponding to one semester course in the German
1
system) is usually enough. It is useful to complement the lectures, however, by ex-
ercise classes which can use the same amount of time. Exercises are included at the
end of each chapter. A PhD class would review parts of part I and part II and cover
in full part III and part IV. It also requires around 14 times 90 minutes and exercise
classes help even more given the more complex material.
2
Contents
1 Introduction 1
3
3.6.4 Aggregation and reduced form . . . . . . . . . . . . . . . . . . 43
3.6.5 Steady state and transitional dynamics . . . . . . . . . . . . . 44
3.6.6 The intertemporal budget constraint . . . . . . . . . . . . . . 45
3.7 A central planner . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
3.7.1 Optimal factor allocation . . . . . . . . . . . . . . . . . . . . . 46
3.7.2 Where the Lagrangian comes from II . . . . . . . . . . . . . . 47
3.8 Exercises Chapter 3 and further reading . . . . . . . . . . . . . . . . 48
4
5.5.1 Free value at end point - Adjustment costs . . . . . . . . . . . 83
5.5.2 Fixed value at end point - Adjustment costs II . . . . . . . . . 85
5.5.3 In…nite horizon - Optimal consumption paths . . . . . . . . . 86
5.5.4 The central planner and capital accumulation . . . . . . . . . 89
5.5.5 The matching approach to unemployment . . . . . . . . . . . 92
5.6 The present value Hamiltonian . . . . . . . . . . . . . . . . . . . . . . 95
5.6.1 Problems without (or with implicit) discounting . . . . . . . . 95
5.6.2 Deriving laws of motion . . . . . . . . . . . . . . . . . . . . . 96
5.6.3 The link between CV and PV . . . . . . . . . . . . . . . . . . 97
5.7 Exercises chapter 5 and further reading . . . . . . . . . . . . . . . . . 98
5
7.5.5 Aggregation and the reduced form for the CD case . . . . . . 131
7.5.6 Some analytical results . . . . . . . . . . . . . . . . . . . . . . 132
7.6 Risk-averse and risk-neutral households . . . . . . . . . . . . . . . . . 134
7.7 Pricing of contingent claims and assets . . . . . . . . . . . . . . . . . 138
7.7.1 The value of an asset . . . . . . . . . . . . . . . . . . . . . . . 138
7.7.2 The value of a contingent claim . . . . . . . . . . . . . . . . . 139
7.7.3 Risk-neutral valuation . . . . . . . . . . . . . . . . . . . . . . 140
7.8 Natural volatility I (*) . . . . . . . . . . . . . . . . . . . . . . . . . . 140
7.8.1 The basic idea . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
7.8.2 A simple stochastic model . . . . . . . . . . . . . . . . . . . . 142
7.8.3 Equilibrium . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
7.9 Exercises chapter 7 and further reading . . . . . . . . . . . . . . . . . 144
6
9.2.1 Why all this? . . . . . . . . . . . . . . . . . . . . . . . . . . . 177
9.2.2 Computing di¤erentials for Brownian motion . . . . . . . . . . 178
9.2.3 Computing di¤erentials for Poisson processes . . . . . . . . . . 180
9.2.4 Brownian motion and a Poisson process . . . . . . . . . . . . . 182
9.2.5 Application - Option pricing . . . . . . . . . . . . . . . . . . . 183
9.2.6 Application - Deriving a budget constraint . . . . . . . . . . . 185
9.3 Solving stochastic di¤erential equations . . . . . . . . . . . . . . . . . 186
9.3.1 Some examples for Brownian motion . . . . . . . . . . . . . . 186
9.3.2 A general solution for Brownian motions . . . . . . . . . . . . 189
9.3.3 Di¤erential equations with Poisson processes . . . . . . . . . . 191
9.4 Expectation values . . . . . . . . . . . . . . . . . . . . . . . . . . . . 194
9.4.1 The idea . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 194
9.4.2 Simple results . . . . . . . . . . . . . . . . . . . . . . . . . . . 195
9.4.3 Martingales . . . . . . . . . . . . . . . . . . . . . . . . . . . . 198
9.4.4 A more general approach to computing moments . . . . . . . 199
9.5 Exercise to chapter 9 and further reading . . . . . . . . . . . . . . . . 203
7
12 Exams 247
(*) These sections are still work in progress. See the note on page 2.
8
1 Introduction
The objective of this course is to provide a toolbox for solving maximization problems
and for working with their solutions in economic models. Maximization problems can
be formulated in discrete or continuous time, under certainty or uncertainty. Various
maximization methods will be used, ranging from “solving by inserting”, via La-
grangian and Hamiltonian approaches to dynamic programming (Bellman equation).
Dynamic programming will be used for all environments, discrete, continuous, certain
and uncertain, the Lagrangian for most of them. Solving by inserting is also very use-
ful in discrete time setups. The Hamiltonian approach is used only for deterministic
continuous time setups. An overview is given on the next page.
Part I deals with discrete time models under certainty, Part II covers continuous
time models under certainty. Part III deals with discrete time models under uncer-
tainty and Part IV, logically, analyzes continuous time models under uncertainty.
The course is called “applied intertemporal optimization“, where the emphasis is
on “applied”. This means that each type of maximization problem will be illustrated
by some example from the literature. These will come from micro- and macroeco-
nomics, from …nance, environmental economics and game theory. This also means,
maybe more importantly, that there is little emphasis on formal scrutiny. This course
is about computing solutions. A more formal treatment (especially of the stochastic
part) can be found elsewhere in the literature and many references will be given.
Given that certain maximization problems are solved many times - e.g. utility
maximization of a household …rst under certainty in discrete and continuous time and
then under uncertainty in discrete and continuous time - and using many methods,
the “rules how to compute solutions” can be easily understood: First, the discrete
deterministic two-period approach provides the basic intuition or feeling for a solution.
Next, in…nite horizon problems add one dimension of complexity by “taking away”
the simple boundary condition of …nite horizon models. In a third step, uncertainty
adds expectations operators and so on. By gradually working through increasing steps
of sophistication and by linking back to simple but structurally identical examples,
intuition for the complex setups is built up as much as possible. This approach then
allows to …nally understand the beauty of e.g. Keynes-Ramsey rules in continuous
time under Poisson uncertainty, Brownian motion or with Levy processes.
1
Even though I would like this to be a nice, clean and …nished textbook,
this is unfortunately not yet the case. Some sections are incomplete
or contain only sketches of ideas. This is also due to the fact that
these notes are permanently expanding in an attempt to cover further
interesting material. The sections which are still work in progress are
marked by an asterisk (*).
In order to get quick access to various model building blocks, the following tables
can be used as guides. The tables show the numbers of the sections where the methods
are presented. Columns indicate whether the environment is of a deterministic or
stochastic nature, rows distinguish between setups in discrete and continuous time.
Solution by
deterministic stochastic
inserting
discrete 2.2.1 8.7
continuous - -
Dynamic
deterministic stochastic
programming
discrete 3.4 8
continuous 6 10
2
2 period GE
deterministic stochastic
OLG model
discrete 2.4 7.5
continuous - -
In addition to solving maximization problems, this book also shows (i) how the
structure of maximization problems is determined by the economic environment an
agent …nds itself in and (ii) how to go from solutions of maximization problems
to properties of the resulting dynamic system. The former point simply states that
budget constraints of households or …rms are endogenous and depend on technological
fundamentals of the model.
endogenous
deterministic stochastic
budget constraints
discrete 2.5.5, 3.6.3 -
continuous 4.3.4 9.2.6
After having gone through part I and part II, students will be familiar with the
two-period overlapping-generations (OLG) model, the optimal saving central planner
model in discrete and continuous time, the matching approach to unemployment,
the decentralized optimal growth model and (coming soon) an optimal growth model
with money, allowing to understand the di¤erence between nominal and real interest
rates in general equilibrium.
In part III and IV, economic insights will be extended and cover portfolio choices,
growth under uncertainty ...
Why teach a course based on this book? Is it not boring to go through all these
methods? In a way, the answer is yes. We all want to understand certain empirical
regularities or understand potential fundamental relationships and make exciting new
empirically testable predictions. In doing so, we also all need to understand existing
work and eventually present our own ideas. It is probably much more boring to
be hindered in understanding existing work and be almost certainly excluded from
3
presenting own ideas if we always spend long time trying to understand how certain
results were derived. How did this author get from equation (1) and (2) to equation
(3)? The major advantage of economic analysis over other social sciences is its strong
foundation in formal models. These models allow to discuss questions much more
precisely as expressions like “marginal utility”, “time preference rate” or “Keynes-
Ramsey rule”reveal a lot of information in very short time. It is therefore extremely
useful to …rst spend some time in getting to know these methods and then try to do
what economics is really about: understand the real world.
But, before we really start, there is also a second reason - at least for some
economists - to go through all these methods: They contain a certain type of truth.
A proof is true or false. The derivation of some optimal behaviour is true or false. A
prediction of general equilibrium behaviour of an economy is truth. Unfortunately,
it is only truth in an analytical sense, but this is at least some type of truth. Better
than none.
4
Part I
Deterministic models in discrete
time
2 OLG models and di¤erence equations
Given the idea of this book to start from simple structures and extending them to the
more complex ones, this chapter starts with the simplest intertemporal problem, a
two-period decision framework. This simple setup already allows to illustrate the basic
dynamic trade-o¤s. Aggregating over individual behaviour, assuming an overlapping-
generations (OLG) structure, and putting individuals in general equilibrium provides
an understanding of the issues involved in these steps and in identical steps in more
general settings. Some revision of properties of di¤erence equations concludes this
chapter.
where consumption when young and old are denoted by cyt and cot+1 or ct and ct+1 ;
respectively, when no ambiguity arises. The individual earns labour income wt in
both periods. It could also be assumed that she earns labour income only in the …rst
period (e.g. when retiring in the second) or only in the second period (e.g. when
going to school in the …rst). Here, with st denoting savings, her budget constraint in
the …rst period is
ct + st = wt (2.2)
and in the second it reads
Interests paid on savings in the second period are given by rt+1 : All quantities are
expressed in units of the consumption good (i.e. the price of the consumption good
is set equal to one. See ch. 2.2.2 for an extension where the price of the consumption
good is pt :).
This budget constraint says something about the assumptions made on the timing
of wage payments and savings. What an individual can spend in period two is princi-
pal and interest of her savings st of the …rst period. There are no interest payments in
period one. This means that wages are paid and consumption takes place at the end
of period 1 and savings are used for some productive purposes (e.g. …rms use it in the
form of capital for production) in period 2. Therefore, returns rt+1 are determined
5
by economic conditions in period 2 and have the index t + 1: Timing is illustrated in
the following …gure.
wt ct wt+1 ct+1
st = wt ct (1 + rt+1 )st
t t+1
Figure 4 Timing in OLG models
These two budget constraints can be merged into one intertemporal budget con-
straint by substituting out savings,
1 1
wt + (1 + rt+1 ) wt+1 = ct + (1 + rt+1 ) ct+1 : (2.3)
It should be noted that by not restricting savings to be non-negative in (2.2) or by
equating the present value of income on the left-hand side with the present value
of consumption on the right-hand side in (2.3), we assume perfect capital markets:
individuals can save and borrow any amount they desire. Equation (2.13) provides a
condition under which individuals save.
Adding the behavioural assumption that individuals maximize utility, the eco-
nomic behaviour of an individual is completely described and one can derive her
consumption and saving decisions. The problem can be solved by a Lagrange ap-
proach, by simply inserting (as e.g. in ch. 7.5.4), or in other ways, to be presented
shortly.
The maximization problem reads maxct ; ct+1 (2.1) subject to the intertemporal bud-
get constraint (2.3). The household’s control variables are ct and ct+1 : As they need
to be chosen such that they satisfy the budget constraint, they can not be chosen
independently of each other. Wages and interest rates are exogenously given to the
household. When choosing consumption levels, the reaction of these quantities to the
decision of our household is assumed to be zero - simply because the household is too
small to have an e¤ect on economy-wide variables. The corresponding Lagrangian
reads
1 1
L = U (ct ; ct+1 ) + wt + (1 + rt+1 ) wt+1 ct (1 + rt+1 ) ct+1 ,
where is a constant called the Lagrange multiplier.
The …rst-order conditions are
Lct = Uct (ct ; ct+1 ) = 0;
1
Lct+1 = Uct+1 (ct ; ct+1 ) [1 + rt+1 ] = 0;
1 1
L = wt + (1 + rt+1 ) wt+1 ct + (1 + rt+1 ) ct+1 = 0:
6
Clearly, the last …rst-order condition is identical to the budget constraint. Note that
there are three variables to be determined, consumption for both periods and the
Lagrange multiplier .1 Having at least three conditions is a necessary, though not
su¢ cient (they might, generally speaking, be linearly dependent) condition for this
to be possible.
The …rst two …rst-order conditions can be combined to give
Marginal utility from consumption today on the left-hand side must equal marginal
utility tomorrow, corrected by the interest rate, on the right-hand side. The economic
meaning of this correction can best be understood when looking at a version with
nominal budget constraints (see ch. 2.2.2).
What we learn from this maximization here is that its solution (2.4) gives us
an equation, that needs to hold when behaving optimally, which links consumption
ct today to consumption ct+1 tomorrow. This equation together with the budget
constraint (2.3) provides a two-dimensional system: two equations in two unknowns,
ct and ct+1 . These equations therefore allow in principle to compute these endogenous
variables as a function of exogenously given wages and interest rates. The next section
provides an example where this is indeed the case.
2.2 Examples
2.2.1 Optimal consumption
This …rst example allows us to indeed solve explicitly for consumption levels in both
periods. Let households have a Cobb-Douglas utility function represented by
Ut = ln ct + (1 ) ln ct+1 : (2.5)
Utility from consumption in each period, instantaneous utility, is given by the log-
arithm of consumption. As ln c has a positive …rst and negative second derivative,
higher consumption increases instantaneous utility but at a decreasing rate. The
parameter captures the importance of instantaneous utility in the …rst relative to
instantaneous utility in the second. Overall utility Ut is maximized subject to the
constraint we know from above,
1
Wt = ct + (1 + rt+1 ) ct+1 ; (2.6)
7
The …rst-order conditions are
1
Lct = (ct ) = 0;
1 1
Lct+1 = (1 ) (ct+1 ) [1 + rt+1 ] = 0;
and the budget constraint (2.6) following from L = 0: Dividing …rst-order conditions
gives 1 ct+1
ct
= 1 + rt+1 and therefore
1
ct = (1 + rt+1 ) ct+1 :
1
This equation corresponds to our optimality rule (2.4) derived above for the more
general case. Inserting into the budget constraint (2.6) yields
1 1 1
Wt = + 1 (1 + rt+1 ) ct+1 = (1 + rt+1 ) ct+1
1 1
which is equivalent to
ct+1 = (1 ) (1 + rt+1 ) Wt : (2.7)
It follows that
ct = W t : (2.8)
Apparently, optimal consumption decisions imply that consumption when young is
given by a share of the present value Wt of life-time income at time t of the individual
under consideration. Consumption when old is given by the remaining share 1
plus interest payments, ct+1 = (1 + rt+1 ) (1 ) Wt : Assuming preferences as in (2.5)
makes modelling sometimes easier than with more complex utility functions. The
drawback is that the share of lifetime income consumed and thereby the savings
decision is independent of the interest rate, which appears implausible. A way out
is given by the CES utility function which also allows for closed-form solutions of
consumption (cf. ch. 7.5.4). More generally speaking, such a simpli…cation should be
justi…ed if some aspect of an economy that are fairly independent of savings are the
focus of some analysis. For an example concerning bequests and wealth distributions,
see Kleiber, Sexauer and Wälde (2005).
8
The …rst-order conditions for ct and ct+1 are
This equation says that marginal utility of consumption today relative to marginal
utility of consumption tomorrow equals the relative price of consumption today and
tomorrow. This optimality rule is identical for a static 2-good decision problem
where optimality requires that the ratio of marginal utilities equals the relative price.
The relative price here is expressed in a present value sense as we compare marginal
utilities at two points in time. The price pt is therefore divided by the price tomorrow,
discounted by next period’s interest rate, pt+1 [1 + rt+1 ] 1 :
Let utility be given by u (c1 ; c2 ; ::::; cn ) : The MRS between good i and good j
given a consumption bundle (c1 ; c2 ; ::::; cn ) (:) is then de…ned by
@u (:) =@ci
M RSij (:) :
@u (:) =@cj
It gives the additional amount of consumption of good j that is required to keep the
utility level at u (c1 ; c2 ; ::::; cn ) when the amount of i is decreased marginally.
Why this is so can easily be shown: Consider the total di¤erential of u (c1 ; c2 ; ::::; cn ) ;
keeping all consumption levels apart from ci and cj …x. This yields
@u (:) @u (:)
du (c1 ; c2 ; ::::; cn ) = dci + dcj :
@ci @cj
Equivalent terms
9
As a reminder, the equivalent term to the MRS in production theory is the mar-
@f (:)=@xi
ginal rate of technical substitution M RT Sij (:) = @f (:)=@xj
where the utility function
was replaced by a production function and consumption ck was replaced by factor
inputs xk .
More on an economy wide level, there is the marginal rate of transformation
@G(:)=@yi
M RTij (:) = @G(:)=@y j
where the utility function was now replaced by a transformation
function G (maybe better known as production possibility curve) and the yk are
output of good k. The marginal rate of transformation gives the increase in output
of good j when output of good i is marginally decreased.
(Intertemporal) elasticity of substitution
Though our main interest is a measure of intertemporal substitutability, we …rst
de…ne the elasiticty of substition in general. The most common de…nition for two
d(ci =cj ) pi =pj
consumption goods is ij d(pi =pj ) ci =cj
: how much does relative consumption ci =cj
adjust when relative prices pi =pj change (proportionally to the current relative-price
p =p
relative-consumption ratio cii =cjj )? This de…nition can be expressed alternatively (see
ex. 5 for details) in a way which is more useful here. Here, the elasticity of substitution
is de…ned by the derivative of the log of relative consumption with respect to the log
of the marginal rate of substitution,
d ln (ci =cj )
ij : (2.10)
d ln M RSij
Inserting the MRS gives
d ln (ci =cj ) uc =uc d (ci =cj )
ij = = i j
d ln uci =ucj ci =cj d uci =ucj
The advantage of an elasticity when compared to a normal derivative as e.g. the MRS
is that an elasticity is measureless. It is expressed in percentage changes. (This can
be seen in the following example and in exercise number 5a.) It can both be applied
to static utility or production functions or to intertemporal utility functions.
The intertemporal elasticity of substitution for a utility function u (ct ; ct+1 ) is then
simply the elasticity of substitution of consumption at two points in time,
uct =uct+1 d (ct =ct+1 )
t;t+1 :
ct =ct+1 d uct =uct+1
Using as example the Cobb-Douglas utility function from (2.5), we obtain
= c1t+1 d (ct =ct+1 )
ct
t;t+1 = = 1;
ct =ct+1 d = 1
ct ct+1
10
The time preference rate
Intuitively, the time preference rate is the rate at which future instantaneous
utilities are discounted. To illustrate, imagine a discounted income stream
2
1 1
x0 + x1 + x2 + :::
1+r 1+r
where discounting takes place at the interest rate r: Replacing income xt by instanta-
neous utility and the interest rate by , would be the time preference rate. Formally,
the time preference rate is the marginal rate of substitution of instantaneous utilities
(not of consumption levels) minus one,2
TPR M RSt;t+1 1:
As an example, consider the following standard utility function which we will use
very often in later chapters,
1 t 1
U0 = t=0 u (ct ) ; ; > 0: (2.11)
1+
Let be a positive parameter and the implied discount factor, capturing the idea of
impatience: By multiplying instantaneous utility functions u (ct ) by t , future utility
is valued less than present utility. This utility function generalizes (2.5) in two ways:
First and most importantly, there is a much longer planning horizon than just two
periods. In fact, the individual’s overall utility U0 stems from the sum of discounted
instanteneous utility levels u (ct ) over periods 0; 1; 2; ... up to in…nity. The idea
behind this objective function is not that individuals live forever but that individuals
care about the well-being of subsequent generations. Second, the instantaneous utility
function u (ct ) is not logarithmic as in (2.5) but of a more general nature where one
would usually assume positive …rst and negative second derivatives, u0 > 0; u00 < 0.
The marginal rate of substitution is then
@U0 (:) =@u (cs ) (1= (1 + ))s
M RSs;s+1 (:) = = =1+ :
@U0 (:) =@u (cs+1 ) (1= (1 + ))s+1
The time preference rate is therefore given by :
Now take as example the utility function (2.5). Computing the MRS minus one,
we have
2 1
= 1= : (2.12)
1 1
The time preference rate is positive if > 0:5: This makes sense for (2.5) as one
should expect that future utility is valued less than present utility.
11
This de…nition of the time preference rate allows us to provide a precise answer
to the question whether consumption increases over time. We simply compute the
condition under which ct+1 > ct by using (2.7) and (2.8),
ct+1 > ct , (1 ) (1 + r) Wt > Wt ,
1+
1+r > ,r> ,r> :
1 1
Consumption increases if the interest rate is higher than the time preference rate.
The time preference rate of the individual (being represented by ) determines how
to split the present value Wt of total income into current and future use. If the interest
rate is su¢ ciently high to overcompensate impatience, i.e. if (1 ) (1 + r) > in
the …rst line, consumption rises.
Note that even though we computed the condition for rising consumption for our
special utility function (2.5), the result that consumption increases when the interest
rate exceeds the time preference rate holds for more general utility functions as well.
We will get to know various examples for this in subsequent chapters.
Let us consider a maximization problem maxx1 ;x2 F (x1 ; x2 ) subject to some con-
straint g(x1; x2 ) = b: (The presentation here is strongly inspired by Intriligator, 1971,
p. 28 - 30). If we rewrite the constraint as x2 = h (x1 ) (which is possible if the
function g (:) is not too complex, i.e. not e.g. of the type g (:) = ex1 x2 + x1 x2 ), the
maximisation problem can be written as
max F (x1; h (x1 )) :
x1
12
The derivatives of implicit functions
@F @g(x1 ; x2 )
= 0: (2.17)
@x1 @x1
As can be easily seen, this is the …rst-order condition of the Lagrangian
L = F (x1; x2 ) + [b g (x1; x2 )]
with respect to x1 :
Undertaking the same steps for x2 would yield the second …rst-order condition
@F @g(x1 ; x2 )
= 0:
@x2 @x2
We have thereby shown where the Lagrangian comes from.
13
2.3.2 How to understand shadow prices
The idea
We can now also give an interpretation of the meaning of the multipliers . Start-
ing from the de…nition of in (2.17) and rewriting them according to
@F=@x2 @F @F
= = ;
@g=@x2 @g @b
we see that equals the change in F as a function of b:3 By how much does F increase
(e.g. your utility) when your constraint b (your bank account) is relaxed? How much
does the social welfare function change when the economy has more capital? How
much do pro…ts of …rms change when the …rm has more workers?
A derivation
A more rigorous derivation is as follows (cf. Intriligator, 1971, ch. 3.3). Compute
the derivative of the maximized Lagrangian with respect to b,
@L(x1 (b) ; x2 (b)) @
= (F (x1 (b); x2 (b)) + (b) [b g (x1 (b); x2 (b))]
@b @b
@x @x 0 @x1 @x2
= Fx1 1 + Fx2 2 + (b) [b g(:)] + (b) 1 gx1 gx2
@b @b @b @b
= (b)
The last equality results from …rst-order conditions and the fact that the budget
constraint holds.
As L(x1 ; x2 ) = F (x1 ; x2 ) due to the budget constraint holding with equality,
@L(x1 ; x2 ) @F (x1 ; x2 )
(b) = =
@b @b
Shadow prices
x1 = f (K1; L1 ) ; x2 = g (K2; L2 ) ;
K1 + K2 = K; L1 + L2 = L:
3 @f (x1 ;:::;xn ) lim4xi !0 f (x1 ;:::;xi +4xi ;:::;xn ) f (x1 ;:::;xn )
The term @x2 cancels out as by de…nition @xi = lim4xi !0 4xi :
14
Using as multipliers p1 ; p2 ; wK and wL ; the Lagrangian reads
2.4.1 Technologies
The …rms
15
Let there be many …rms who use a technology
Yt = Y (Kt ; L) (2.18)
which is characterized by constant returns to scale. Letting …rms act under perfect
competition, the …rst-order conditions from pro…t maximization, where pro…ts are
given by t = Yt wtK K wtL L, read
@Yt @Yt
= wtK ; = wtL : (2.19)
@Kt @Lt
They equate in each period t the marginal productivity of capital to the factor price
wtK for capital and the marginal productivity of labour to labour’s factor reward wtL .
Euler’s theorem
Euler’s theorem shows that for a linear-homogenous function f (x1 ; x2 ; :::; xn ) the
sum of partial derivatives times the variables with respect to which the derivative was
computed equals the original function f (:) ;
2.4.2 Households
Individual households
Households live again for two periods. The utility function is therefore as in (2.5)
and given by
Ut = ln cyt + (1 ) ln cot+1 : (2.23)
It is maximized subject to the intertemporal budget constraint
1 o
wt = cyt + (1 + rt+1 ) ct+1 :
16
This constraint di¤ers slightly from (2.6) in that people work only in the …rst period
and retire in the second. Hence, there is labour income only in the …rst period on
the left-hand side. Savings of the …rst period are used to …nance consumption in the
second period.
Given that the present value of lifetime wage income is wt ; we can conclude from
(2.7) and (2.8) that individual consumption expenditure and savings are given by
Aggregation
We assume that in each period L individuals are born and die. Hence, the number
of young and the number of old is L as well. As all individuals within a generation
are identical, aggregate consumption within one generation is simply the number of,
say, young times individual consumption. Aggregate output in t is therefore given by
Ct = Lcyt + Lcot : Using the expressions for individual consumption from (2.24) and
noting the index t (and not t + 1) for the old yields
Ct = Lcyt + Lcot = ( wt + (1 ) (1 + rt ) wt 1 ) L:
Kt+1 = (1 ) Kt + Y t Ct : (2.26)
Yt + (1 ) Kt = Ct + Kt+1 : (2.27)
In this formulation, it re‡ects the goods market equilibrium where the left-hand side
shows supply as current production and capital held by the old. The old sell capital
as it is of no use for them, given that they will not be able to consume anything
next period. Demand for the aggregate good is given by aggregate consumption (i.e.
consumption of the young plus consumption of the old) plus the capital stock to be
held next period by the currently young.
17
simplifying the structure of the system of equations coming out of the model as much
as possible. In the end, after inserting and reinserting, a system of n equations in n
unknowns results.
Ideally, there is only one equation left and this equation gives an explicit solution
of the endogenous variable. In static models, an example would be LX = L; i.e.
employment in sector X is given by a utility parameter times total exogenous
labour supply L: This would be an explicit solution. If one is left with just one
equation but one obtains on an implicit solution, one would obtain something like
f (LX ; ; L) = 0:
We now derive, given the results we obtained by now, how large the capital stock
in the next period is. Splitting aggregate consumption into consumption of the young
and consumption of the old and using the output-factor reward identity (2.22) for the
resource constraint in the OLG case (2.27), we obtain
De…ning the interest rate rt as the di¤erence between factor rewards wtK for capital
and the depreciation rate ;
rt wtK ; (2.28)
we …nd
rt Kt + wtL L + Kt = Cty + Cto + Kt+1 :
The interest rate de…nition (2.28) shows the net income of capital owners per unit of
capital. They earn the gross factor rewards wtK but, at the same time, they experience
a loss from depreciation. Net income therefore only amounts to rt : As the old consume
the capital stock plus interest cot L = (1 + rt )Kt 4 , we obtain
which is the aggregate version of the savings equation (2.25). Hence, we have found
that savings st of young at t is the capital stock at t + 1:
Note that equation (2.29) is often present on “intuitive” grounds. The old in
period t have no reason to save as they will not be able to use their savings in t + 1:
Hence, only the young will save and the capital stock in t + 1; being made up from
savings in the previous period, must equal to the savings of the young.
In our simple dynamic model considered here, we obtain indeed the ideal case
where we are left with only one equation that gives us the solution for one variable,
4
Now we see that the interest rate in (2.24) must refer to period t + 1.
18
the capital stock. Inserting the individual savings equation (2.25) into (2.29) gives
with the …rst-order condition (2.19) of the …rm
@Y (Kt ; L)
Kt+1 = (1 ) wtL L = (1 ) L: (2.30)
@L
The …rst equality shows that a share 1 of labour income becomes capital in the
next period. Interestingly, the depreciation rate does not have an impact on the
capital stock in period t + 1. Economically speaking, the depreciation rate a¤ects
wealth of the old but - with logarithmic utility - not saving of the young.
Steady state
@Y (K ; L)
K = (1 ) L: (2.31)
@L
All other variables like aggregate consumption, interest rates, wages etc. are constant
as well. Consumption when young and when old can di¤er as in a setup with …nite
lifetimes the interest rate in the steady state does not need to equal the time preference
rate of households.
Transitional dynamics
Dynamics of the capital stock are illustrated in …gure 5. The …gure plots the
capital stock in period t on the horizontal axis. The capital stock in the next period,
Kt+1 ; is plotted on the vertical axis. The law of motion for capital from (2.30) then
shows up as the curve in this …gure. The 45 line equates Kt+1 to Kt :
We start from our initial condition K0 . Equation (2.30) or the curve in this …gure
then determine the capital stock K1 : This capital stock is then viewed as Kt such
19
that, again, the curve gives us Kt+1 ; which is, given that we now started in 1; the
capital stock K2 of period 2: We can continue doing so and see graphically that the
economy approaches the steady state K which we had computed in (2.31).
N
Kt+1
Kt = Kt+1
Kt+1 = (1 )w(Kt )L
6
-
- 6
N
0
K0 Kt
Figure 5 Convergence to the steady state
Summary
T i 1 aT T i 1 aT +1
i=1 a =a ; i=0 a =
1 a 1 a
Proof. The left hand side is given by
T i
i=1 a = a + a2 + a3 + : : : + aT 1
+ aT : (2.32)
20
Now subtract (2.33) from (2.32) and …nd
T i 1 aT
(1 a) i=1 a =a aT +1 , T
i=1 a
i
=a : (2.34)
1 a
Lemma 2
T i 1 1 aT
i=1 ia = a T aT +1
1 a 1 a
Solution by inserting
The simplest way to solve it is simply to insert and reinsert this equation su¢ -
ciently often. Doing it three times gives
x1 = ax0 ;
x2 = ax1 = a2 x0 ;
x3 = ax2 = a3 x0 :
When we look at this solution for t = 3 long enough, we see that the general solution
is
x t = at x 0 :
This could formally be proven by a proof by induction.
21
14
12
10
0
0 2 4 6 8 10
e x pw a .m
Long-run behaviour
We can now ask whether xt approaches a constant when time goes to in…nity.
This gives 8 9 8
< 0 = < a<1
t
lim xt = x0 lim a = x 0 , a=1 :
t !1 t !1 : ; :
1 a>1
Hence, xt approaches a constant only for a < 1:
A graphical analysis
For more complex di¤erence equations, it often turns out to be useful to analyze
their behaviour in a phase diagram. Even though this simple di¤erence equation
can be understood easily analytically, we will illustrate its properties in the following
…gure. This allows to understand how analyses of this type can be undertaken also
for more complex di¤erence equations.
N N
xt+1 xt+1
xt+1 = axt
-
xt = xt+1
xt+1 = axt 6
xt = xt+1
-6
N
x0 xt x0 xt
22
Figure 7 A phase diagram for a < 1 in the left and a > 1 in the right panel
The left panel shows how xt evolves over time, starting at x0 : In this case of a < 1;
we see how xt approaches zero. When we graphically illustrate the case of a > 1; the
evolution of xt is as shown in the right panel.
Solution by inserting
We solve again by inserting. In contrast to the solution for (2.35), we start from
an initial value of xt : Hence, we imagine we are in t (t as today) and compute what
the level of x will be tommorrow and the day after tommorrow etc. We …nd for xt+2
and xt+3 that
n 1 i an 1
xt+n = an xt + b i=0 a = an x t + b :
a 1
The last equality used the …rst lemma from ch. 2.5.1.
The left panel of the following …gure studies the evolution of xt for the stable case,
i.e. where 0 < a < 1 and b > 0: Starting today in t with xt ; we end up in x . As
we chose a smaller than one and a positive b; x is positive as (2.37) shows. We will
return to the right panel in a moment.
23
N N xt+1 = axt + b
xt+1 xt+1 -
xt = xt+1
? xt = xt+1
6
?
xt+1 = axt + b -
- 6
b - 6
N
x xt x xt
We can use these examples to de…ne two concepts that will be useful also at later
stages.
x = f (x ) : (2.38)
For di¤erence equations of the type xt+1 = f (xt ) ; the …xpoint x of the function
f (xt ) is the point where x stays constant. This is usually called the long-run equilib-
rium point of some economy or its steady or stationary state. Whenever an economic
model, represented by its reduced form, is analysed, it is generally useful to …rst try
and …nd out whether …xpoints exist and what its economic properties are.
For the di¤erence equation of last section, we obtain
b
xt+1 = xt x () x = ax + b () x = :
1 a
Local and global stability
Once a …xpoint has been identi…ed, one can ask whether it is stable.
The concept of global stability usually refers to initial values x0 that are econom-
ically meaningful. An initial physical capital stock that is negative would not be
considered to be economically meaningful.
unstable
De…nition 3 (Local stability and instability) A …xpoint x is if,
locally stable
diverges from
starting from an initial value x + "; where " is small, xt x.
converges to
24
For illustration purposes consider the …xpoint x in the left panel of …g. 8 - it is
globally stable. In the right panel of the same …gure, it is unstable. As can easily
be seen, the instability follows from simply letting the xt+1 line intersect the 45 -line
from below. In terms of the underlying di¤erence equation (2.36), this requires b < 0
and a > 1:
Clearly, economic systems can be much more complex and generate several …x-
points. Imagine the link between xt+1 and xt is not linear as in (2.36) but nonlinear,
xt+1 = f (xt ) : Unfortunately for economic analysis, a nonlinear relationship is the
much more realistic case. The next …gure provides an example for some function
f (xt ) that implies an unstable xu and a locally stable …xpoint xs .
N
xt+1
?
-
- 6
N
x01 xu x02 xs x03 x(t)
25
Deriving a budget constraint is in principle straightforward. One just de…nes
wealth of the household (taking into account which types of assets the household can
hold for saving purposes and what their prices are), computes the di¤erence between
wealth “today”and “tommorrow”(this is where the di¤erence equation aspect comes
in) and uses an equation which relates current savings to current changes in the
number of assets. In a …nal step, one will naturally …nd out how the interest rate
showing up in budget constraints relates to more fundamental quantities like value
marginal products and depreciation rates.
In our one-asset case, wealth at of a household is given by the number kt of stocks
the household owns (say the number of machines she owns) times the price vt of one
stock (or machine), at = vt kt : Computing the …rst di¤erence yields
Wealth changes depend on the acquisition vt+1 (kt+1 kt ) of new assets and on
changes in the value of assets that are already held, (vt+1 vt ) kt .
Savings are given by capital income (net of depreciation losses) and labour income
minus consumption expenditure, wtK kt vt kt + wt pt ct . Savings in t are used for
buying new assets in t for which the period-t price needs to be paid,
wtK kt vt kt + wt pt ct
= kt+1 kt : (2.40)
vt
We can rewrite this equation slightly, which will simplify the interpretation of subse-
quent results, as
wK kt + wt pt ct
kt+1 = (1 ) kt + t :
vt
Wealth in the next period expressed in number of stocks (and hence not in nominal
terms) is given by wealth which is left over from the current period, (1 ) kt , plus
new acquisitions of stocks which amount to gross capital plus labour income minus
consumption expenditure divided by the price of one stock. This is a di¤erence
equation in kt but not yet a di¤erence equation in wealth at :
Inserting this (2.40) into (2.39) yields
vt+1 K
at+1 at = wt kt vt kt + wt pt ct + (vt+1 vt ) kt
vt
vt+1 wtK vt+1
= at at + wt pt ct + 1 at ,
vt vt vt
vt+1 wK
at+1 = 1+ t at + wt pt ct : (2.41)
vt vt
What does this equation tell us? Each unit of wealth at (say Euro, Dollar, Yen
...) gives 1 units at the end of the period as % is lost due to depreciation
plus “dividend payments” wtK =vt . Wealth is augmented by labour income minus
consumption expenditure. This end-of-period wealth is expressed in wealth at+1 at
26
the beginning of the next period by dividing it through vt (which gives the number
kt of stocks at the end of the period) and multiplying it by the price vt+1 of stocks in
the next period. We have thereby obtained a di¤erence equation in at :
This general budget constraint is fairly complex, however, which implies that in
practice it is often expressed di¤erently. One possibility consists in choosing the
capital good as the numeriare good and setting vt 1 8t: This simpli…es (2.41) to
kt+1 = (1 + rt ) kt + wt pt ct ;
which is very similar to (3.33) from ch. 3.6.3. The simpli…cation in this expression
consists also in the de…nition of the interest rate rt as rt wtK :
An alternative consists in equalizing the price of the capital good to the price of the
consumption good, pt = vt . This is often the case in general equilibrium approaches
- think e.g. of growth models - when there is an aggregate technology which is used
both for producing the consumption and the investment good. The constraint (2.41)
then simpli…es to
wt
art+1 = (1 + rt ) art + ct ;
pt
at wK
where real wealth and the interest rate are de…ned by art pt
and rt t
pt
,
respectively.
If one is not willing to …x any prices or relative prices - which should indeed not
be done if one is interested e.g. in capital asset pricing issues, to be treated later in
ch. 8.4 - one can write the somehow cumbersome budget constraint (2.41) in a more
elegant way. To this end, divide by vt+1 and use the de…nition of at to get the number
of stocks held tommorrow as
wtK wt pt
kt+1 = 1+ kt + ct : (2.42)
vt vt vt
Rearranging such that expenditure is on the left- and disposable income on the right-
hand side yields
pt ct + vt kt+1 = vt kt + wtK vt kt + wt :
This equation has a simpler form than (2.41) and it also lends itself to a simple in-
terpretation: On the left-hand side is total expenditure in period t; consisting of con-
sumption expenditure pt ct plus expenditure for buying the number of capital goods,
kt+1 , the household wants to hold in t + 1: As this expenditure is made in t; total
expenditure for capital goods amounts to vt kt+1 . The right-hand side is total dispos-
able income which splits into income vt kt from selling all capital “inherited”from the
previous period, capital income wtK vt kt and labour income wt : This is the form
budget constraints are often expressed in capital asset pricing models. Note that this
is in principle also a di¤erence equation in kt :
27
Samuelson. For presentations in textbooks, see e.g. Blanchard and Fischer (1989),
Azariadis (1993) or de la Croix and Michel (2002). The presentation of the Lagrangian
is inspired by Intriligator (1971). Treatments of shadow prices are available in many
other textbooks (Dixit, 1989, ch. 4; Intiriligator, 1971, ch. 3.3). More extensive
treatments of di¤erence equations and the implicit function theorem can be found in
many introductory “mathematics for economists”books.
28
Exercises chapter 2
Applied Intertemporal Optimization
Optimal consumption in two-period discrete time models
Ut = ln ct + (1 ) ln ct+1
subject to
1 1
b + n + (1 + r) wt+1 = ct + (1 + r) ct+1 :
(a) What is the optimal consumption pro…le under no capital market restric-
tions?
(b) Assume loans for …nancing education are not available, hence savings need
to be positive, st 0. What is the consumption pro…le in this case?
3. Optimal investment
Consider a monopolist investing in its technology. Technology is captured by
marginal costs ct . The chief accountant of the …rm has provided the manger of
the …rm with the following information,
Assume you are the manager. What is the optimal investment sequence I1 , I2 ?
29
4. A particular utility function
Consider the utility function U = ct + ct+1 ; where 0 < < 1: Maximize U
subject to an arbitrary budget constraint of your choice. Derive consumption
in the …rst and second period. What is strange about this utility function?
5. Intertemporal elasiticity of substitution
Consider the utility function U = ct1 + ct+1
1
:
6. General equilibrium
Consider the Diamond model for a Cobb-Douglas Production function of the
form Y = K L1 and a logarithmic utility function.
7. Sums
T i 1 1 aT
i=1 ia = a T aT +1 :
1 a 1 a
The idea is identical to the …rst proof in ch. 2.5.1.
(b) Show that (exam)
k 1 k 1 s s ck4 k
s=0 c4 = :
c4
Both parameters obey 0 < c4 < 1 and 0 < v < 1: Hint: Rewrite the sum
as ck4 1 ks=01 ( =c4 )s and observe that the …rst lemma in ch. 2.5.1 holds for
a which are larger or smaller than 1.
8. Di¤erence equations
Consider the following linear di¤erence equation system
30
3 In…nite horizon models
This chapter looks at decision processes where the time horizon is longer than two
periods. In most cases, the planning horizon will be in…nity. In such a context,
Bellman’s optimality principle is very useful. Is it, however, not the only way how to
solve maximization problems with in…nite time horizon. For comparison purposes, we
therefore start with the Lagrange approach, as in the last section. Bellman’s principle
will be introduced afterwards when intuition for the problem and relationships will
have been increased.
31
where the latter is, as in the OLG case, the budget constraint. Again, we have as
many conditions as variables to be determined.
Do these …rst-order conditions tell us something? Take the …rst-order condition
for period and for period + 1. They read
t 0 ( t)
u (c ) = [1 + r] p ;
+1 t 0 ( +1 t)
u (c +1 ) = [1 + r] p +1 :
1 u0 (c ) p u0 (c ) p
0
= (1 + r) () 0
= 1 : (3.5)
u (c +1 ) p +1 u (c +1 ) (1 + r) p +1
Theorem 1 Let there be a function g (x; y) : Choose x such that g (:) is maximized
for a given y: (Assume g (:) is such that a smooth interior solution exists.) Let f (y)
be the resulting function of y;
Then the derivative of f with respect to y equals the partial derivative of g with respect
to y, if g is evaluated at that x = x(y) that maximizes g;
d f (y) @ g (x; y)
= :
dy @y x=x(y)
@
Proof. f (y) is constructed by @x g (x; y) = 0: This implies a certain x = x (y) ;
provided that second order conditions hold. Hence, f (y) = maxx g (x; y) = g (x (y) ; y) :Then,
d f (y)
dy
= @ g(x(y);y)
@x
d x(y)
dy
+ @ g(x(y);y)
@y
: The …rst term of the …rst term is zero.
32
3.3.2 Illustration
The plane depicts the function g (x; y). The maximum of this function with respect
to x is shown as maxx g (x; y), which is f (y). Given this …gure, it is obvious that the
derivative of f (y) with respect to y is the same as the partial derivative of g (:) with
respect to y at the point where g (:) has its maximum with respect to x: The partial
@g
derivative @y is the derivative in the direction of y (“Richtungsableitung”). Choosing
the highest point of g (:) with respect to x, this “Richtungsableitung” must be the
same as dfdy(y) at the back of the …gure.
f (y )
max g(x , y )
x
g(x , y )
df (y ) ∂g(x(y ), y ) dx ∂g(x(y ), y )
= +
dy ∂x dy ∂y
∂g(x(y ), y )
=
∂y y
3.3.3 An example
There is a central planner of an economy. The social welfare function is given by
U (A; B), where A and B are consumption goods. The technologies available for
producing these goods are A = A (cLA ) and B = B (LB ) and the economy’s resource
constraint is LA + LB = L:
5
This …gure is contained in the …le envelope.ds4. It was originally created in Mathematica, using
the command Plot3D[-(x-.5*y)^ 2-(y-2)^ 2,fx,0,2g,fy,0,4g, fAxes -> noaxes, PlotRange ->f-8,5g,
Boxed -> Falseg]
33
The planner is interested in maximizing the social welfare level and allocates
labour according to maxLA U (A (cLA ) ; B (L LA )) : The optimality condition is
@U 0 @U 0
Ac B = 0: (3.6)
@A @B
This makes LA a function of c,
LA = LA (c) : (3.7)
The central planner now asks what happens to the social welfare level when the
technology parameter c increases and she still maximizes the social welfare function,
i.e. (3.6) continues to hold. Without using the envelope theorem, the answer is
d
U (A (cLA (c)) ; B (L LA (c)))
dc
0 0 0
= UA (:) A [LA (c) + cL0A (c)] + UB (:) B [ L0A (c)] = UA (:) A LA (c) > 0
where the last equality follows from inserting the optimality condition (3.6). Eco-
nomically, this result means that the e¤ect of a better technology on overall welfare
is given by the direct e¤ect in sector A: The indirect e¤ect through reallocation of
labour vanishes as due to the …rst-order condition (3.6) the marginal contribution of
each worker is identical across sectors. Clearly, this only holds for a small change in
c:
If one is interested in …nding an answer by using the envelope theorem, one would
start by de…ning a function V (c) maxLA U (A (cLA ) ; B (L LA )) : Then, according
to the envelope theorem,
d @ 0 0
V (c) = U (A (cLA ) ; B (L LA )) = UA (:) A LA = UA (:) A LA (c) > 0:
dc @c LA =LA (c) LA =LA (c)
Apperently, both approaches yield the same answer. Applying the envelope theorem
gives the answer faster.
We denote wealth by xt to make clear that this constraint can represent also other
aspects than a budget constraint of a household. In fact, this di¤erence equation
6
All intertemporal utility functions in this book will use exponential discounting. This is clearly
a special case. Models with non-exponential or hyperbolic discounting imply fundamentally di¤erent
dynamic behaviour and time inconsistencies.
34
can be non-linear (e.g. in central planner problem where the constraint is a resource
constraint like (3.48)) or linear (e.g. a budget constraint like (3.14)). We treat here
this general case …rst before we go to more speci…c examples further below.
The consumption level ct and - more generally speaking - other variables whose
value is directly chosen by individuals, e.g. investment levels or shares of wealth
held in di¤erent assets, are called control variables. Variables which are not under
the direct control of individuals are called state variables. In many maximization
problems, state variables depend indirectly on the behaviour of individuals as in
(3.8). State variables can also be completely exogenous like e.g. the TFP level in an
exogenous growth model.
Optimal behaviour is de…ned by maxfc g Ut subject to (3.8), where the value of
this optimal behaviour or optimal program is denoted by
V (xt ) is called the value function. It is a function of the state variable xt and not
of the control variable ct . The latter point is easy to understand if one takes into
account that the control variable ct is, when behaving optimally, a function of the
state variable xt :
The value function V (:) could also be a function of time t (e.g. in problems with
…nite horizon) but we will not discuss this further as it is of no importance in the
problems we will encounter. Generally speaking, xt and ct could be vectors and time
could then be part of the state vector xt : The value function is always a function of
the states of the system or the maximization problem.
Given this description of the maximization problem, solving by dynamic program-
ming essentially requires to go through three steps. This three-step approach will be
followed here, in continuous time later, and also in models with uncertainty.
The …rst step establishes the Bellman equation and computes …rst-order condi-
tions. The objective function Ut in (3.1) is additively separable which means that it
can be written in the form
Ut = u(ct ) + Ut+1 : (3.10)
Bellman’s idea consists in rewriting the maximization problem in the optimal program
(3.9) as
V (xt ) max fu(ct ) + V (xt+1 )g (3.11)
ct
subject to
xt+1 = f (xt ; ct ) :
Equation (3.11) is known as Bellman equation. In this equation, the problem
with potentially in…nitely many control variables fc g was broken down in many
problems with one control variable ct . Note that there are two steps involved: First,
the additive separability of the objective function is used. Second, more importantly,
35
Ut is replaced by V (xt+1 ). This says that we assume that the optimal problem for
tomorrow is solved and we worry about the maximization problem of today only.
We can now compute the …rst-order condition which is of the form
@f (xt ; ct )
u0 (ct ) + V 0 (xt+1 ) = 0: (3.12)
@ct
In principle, this is the solution of our maximization problem. Our control variable
ct is by this expression implicitly given as a function of the state variable, ct = c (xt ) ;
as xt+1 by the constraint (3.8) is a function of xt and ct . Hence, as V is well-de…ned
above, we have obtained a solution.
As we know very little about the properties of V at this stage, however, we need
to go through two further steps in order to eliminate V from this …rst-order condition
and obtain a condition that uses only functions (like e.g. the utility function or the
technology for production in later examples) of which properties like signs of …rst and
second derivatives are known.
The second step of the dynamic programming approach starts from the “maxi-
mized Bellman equation”. The maximized Bellman equation is obtained by replacing
the control variables in the Bellman equation, i.e. the ct in (3.11) in the present case,
by the optimal control variables as given by the …rst-order condition, i.e. by c (xt )
resulting from (3.12). Logically, the max operator disappears (as we insert the c (xt )
which imply a maximimum) and the maximized Bellman equation reads
This equation is a di¤erence equation for the costate variable V 0 (xt ). This derivative
of the value function is the shadow price of the state variable xt : It says how much
an additional unit of the state variable (say e.g. of wealth) is valued: As V (xt ) gives
the value of optimal behaviour between t and the end of the planning horizon, V 0 (xt )
says by how much this value changes when xt is changed marginally. Hence, equation
(3.13) describes how the shadow price of the state variable changes over time when
the agent behaves optimally. If we had used the envelope theorem, we would have
immediately ended up with (3.13) without having to insert the …rst-order condition.
36
DP3: Inserting …rst-order conditions
Now insert the …rst-order condition (3.12) twice into (3.13) to replace the unknown
value function and to …nd an optimality condition depending on u only. This will
then be the Euler equation. We do not do this here explicitly as many examples will
go through this step in detail in what follows.
3.5 Examples
3.5.1 Individual utility maximization
The individual’s budget constraint is given in the dynamic formulation (in contrast
to an intertemporal version (3.3)) by
Note that this dynamic formulation corresponds to the intertemporal version in the
sense that (3.3) implies (3.14) and (3.14) with some limit condition implies (3.3). The
…rst part of this statement can easily be seen: Write (3.3) for the next period as
1 ( t 1) 1 ( t 1)
=t+1 (1 + r) e = at+1 + =t+1 (1 + r) w : (3.15)
Express (3.3) as
1 ( t) 1 ( t)
et + =t+1 (1 + r) e = at + wt + =t+1 (1 + r) w ,
1 1 ( t 1) 1 1 ( t 1)
et + (1 + r) =t+1 (1 + r) e = at + wt + (1 + r) =t+1 (1 + r) w ,
1 ( t 1) 1 ( t 1)
=t+1 (1 + r) e = (1 + r) (at + wt et ) + =t+1 (1 + r) w :
Insert (3.15) and …nd (3.14). The second part is part of the exercises.
The budget constraint (3.14) can be found e.g. in Blanchard and Fischer (1989, p.
280). The timing as implicit in (3.14) is illustrated in the following …gure. All events
take place at the beginning of the period. Our individual owns a certain amount
of wealth at at teh beginning of t and receives here wage income wt and spends ct
on consumption also at the beginning. Hence, savings st can be used during t for
production and interest are paid on st which in turn gives at+1 at the beginning of
period t + 1.
st = at + wt − ct at +1 = (1 + rt )st
ct ct +1
t
t +1
37
The consistency of (3.14) with technologies in general equilibrium is not self-
evident. We will encounter more conventional budget constraints of the type (??)
further below. For understanding how to apply dynamic programming methods,
however, this budget constraint is perfectly …ne.
The objective of the individual is to maximize her utility function (3.1) subject
to the budget constraint by choosing a path of consumption levels ct ; denoted by
fc g ; 2 [t; 1] : The value of the optimal program fc g is, given her initial endow-
ment with at ,
V (at ) = max Ut (3.16)
fc g
38
DP2: Evolution of the costate variable
The derivative of the maximized Bellman equation reads, using the envelope the-
orem,
@at+1
V 0 (at ) = V 0 (at+1 ) : (3.19)
@at
We compute the partial derivative of at+1 with respect to at as the functional rela-
tionship of ct = ct (at ) should not (because of the envelope theorem) be taken into
account.
From the budget constraint we know that @a@at+1 t
= 1 + rt : Hence, the evolution of
the shadow price under optimal behaviour is described by
V 0 (at ) = (1 + rt ) V 0 (at+1 ) :
We can insert the …rst-order condition (3.18) on the RHS. We can also rewrite the
…rst-order condition (3.18), by lagging it one period, as (1 + rt 1 ) V 0 (at ) = u0 (ct 1 )
and can insert this on the LHS. This gives
1 1
u0 (ct 1 ) (1 + rt 1 ) = u0 (ct ) , u0 (ct ) = (1 + rt ) u0 (ct+1 ) : (3.20)
This is the same result as the one we have obtained when we used the Lagrange
method in equation (3.5).
It is also the same result as for the 2-period saving problem which we found in
OLG models - see e.g. (2.9) or (2.43) in the exercises. This might be suprising as
the planning horizons di¤er considerably between a 2- and an in…nite-period decision
problem. Apperently, whether we plan for two periods or many more, the change
between two periods is always the same when we behave optimally. It should be kept
in mind, however, that consumption levels (and not changes) do depend on the length
of the planning horizon.
39
where f (:) is the page of quality production function: More e¤ort means more output,
f 0 (:) > 0; but any increase in e¤ort implies a lower increase in output, f 00 (:) < 0:
The objective function of our student is given by
T t T t
Ut = D =t e : (3.22)
The value of the completed dissertation is given by D and its present value is obtained
by discounting at the discount factor : Total utility Ut stems from this present value
minus the present value of research cost e . The maximization problem consists in
maximizing (3.22) subject to the constraint (3.21) by choosing an e¤ort path fe g :
The question now arises how these costs are optimally spread over time. How many
hours should be worked per day?
An answer can be found by using the Lagrange-approach with (3.22) as the objec-
tive function and (3.21) as the constraint. We will use here the dynamic programming
approach, however. Before we can apply it, we need to derive a dynamic budget con-
straint. We therefore de…ne
t 1
Mt =1 f (e )
as the amount of the pages that have already been written by today. This then implies
The increase in the number of completed pages between today and tommorrow de-
pends on e¤ort-induced output f (e ) today. We can now apply the three dynamic
programming steps.
The value function is given by V (Mt ) = maxfe g Ut subject to the constraint. The
objective function written recursively reads
T (t+1) 1 T t T (t+1) 1 T t
Ut = D =t e = D et + =t+1 e
T (t+1) T (t+1)
= D =t+1 e et = Ut+1 et :
Assuming that the individual behaves optimally as from tomorrow, this reads Ut =
et + V (Mt+1 ) and the Bellman equation reads
Again as in (3.12), implicitly and with (3.23), this equation de…nes a functional
relationship between the control variable and the state variable, et = e (Mt ) :
40
DP2: Evolution of the costate variable
To provide some variation, we will now go through the second step of dynamic
programming without using the envelope theorem. Consider the maximized Bellman
equation, where we insert et = e (Mt ) and (3.23) into Bellman equation (3.24),
Using the …rst-order condition (3.25) simpli…es this derivative to V 0 (Mt ) = V 0 (Mt+1 ) :
Expressed for t + 1 gives
V 0 (Mt+1 ) = V 0 (Mt+2 ) (3.26)
The …nal step inserts the …rst-order condition (3.25) twice to replace V 0 (Mt+1 )
and V 0 (Mt+2 ) ;
1 f 0 (et+1 )
(f 0 (et )) 1 = (f 0 (et+1 )) 1 , 0 = : (3.27)
f (et )
The interpretation is now simple. As f 00 (:) < 0 and < 1; e¤ort et is increasing under
optimal behaviour, i.e. et+1 > et : Optimal writing of a dissertation implies more work
every day.
The optimality condition in (3.27) speci…es only how e¤ort e changes over time,
it does not provide information on the level of e¤ort required every day. This is
a property of all solutions of intertemporal problems. They only give information
about changes of levels, not about levels themselves. The basic idea how to obtain
information about levels can be easily illustrated, however.
Assume f (et ) = et ; with 0 < < 1: Then (3.27) implies (with t being replaced
by ) e +11 =e 1 = , e +1 = 1=(1 ) e : Solving this di¤erence equation yields
( 1)=(1 )
e = e1 ; (3.28)
where e1 is the (at this stage still) unknown initial e¤ort level. Inserting this solution
into the constraint (3.21) yields, starting in = 1 on the …rst day,
T ( 1)=(1 ) T ( 1) =(1 )
=1 f e1 = =1 e1 = L:
41
This gives us the initial e¤ort level as (the sum can be solved by using the proofs in
ch. 2.5.1)
1=
L
e1 = T ( 1) =(1 )
:
=1
With (3.28), we have now also computed the level of e¤ort every day.
Behind these simple steps, there is a general principle. Solutions to intertemporal
problems are di¤erence or di¤erential equations. Any di¤erence (or also di¤erential)
equation when solved gives a unique solution only if an initial or boundary condition is
provided. Here, we have solved the di¤erence equation in (3.27) assuming some initial
condition e1 : The meaningful initial condition then followed from the constraint (3.21).
Hence, in addition to the optimality rule (3.27), we always need some additional
constraint which allows to compute the level of optimal behaviour. We return to this
point when looking at problems in continuous time in ch. 5.3.
3.6.1 Technologies
The technology is a simple Cobb-Douglas technology
Yt = At Kt Lt1 : (3.29)
This good can be used for consumption and investment and equilibrium on the goods
market requires
Yt = Ct + It :
Gross investment It is turned into net investment by taking depreciation into account,
Kt+1 = (1 ) Kt + It :
Taking these two equations together give the resource constraint of the economy,
Kt+1 = (1 ) Kt + Y t Ct : (3.30)
As this constraint is simply a consequence of technologies and market clearing, it is
identical to the one used in the OLG setup in (2.26).
42
3.6.2 Firms
Firms maximize pro…ts by employing optimal quantities of labour and capital, given
the technology in (3.29). First order conditions are
@Yt @Yt
= wtK ; = wtL (3.31)
@Kt @Lt
as in (2.19), where we have again chosen the consumption good as numeraire.
3.6.3 Households
Preferences of households are described as in the intertemporal utility function (3.1).
As the only way households can transfer savings from one period to the next is by
buying investment goods, an individual’s wealth is given by the “number of machines”
kt , she owns. Clearly, adding up over all individual wealth stocks gives the total
capital stock, L kt = Kt : Wealth kt increases over time if the household spends less
on consumption than what she earns through capital plus labour income, corrected
for the loss in wealth each period caused by depreciation,
rt wtK ; (3.32)
Note that the “derivation” of this budget constraint was simpli…ed as compared
to ch. 2.5.5 as the price vt of an asset is, as we measure it in units of the consumption
good, by 1.
Given the preferences and the constraint, the solution to this maximization prob-
lem is given by (see exercise 5)
Structurally, this is the same expression as in (3.20). The interest rate, however, refers
to t+1, due to the change in the budget constraint. Remembering that = 1= (1 + ),
this shows that consumption increases as long as rt+1 > .
43
To see that individual constraints add up to the aggregate resource constraint,
we simply need to take into account as in (2.21) that individual income adds up to
output Yt ; wtK Kt + wtL Lt = Yt :
The optimal behaviour of all household taken together can be gained from (3.34)
by summing over all households. This is done analytically correctly by …rst applying
the inverse function of u0 to this equation and then summing individual consumption
levels over all households. Applying the inverse function again gives
u0 (Ct ) = (1 + rt+1 ) u0 (Ct+1 ) ; (3.35)
where Ct is aggregate consumption in t:
Reduced form
We now need to understand how our economy evolves in general equilibrium. Our
…rst equation is (3.35), telling us how consumption evolves over time. This equation
contains consumption and the interest rate as endogenous variables.
Our second equation is therefore the de…nition of the interest rate in (3.32) which
we combine with the …rst-order condition of the …rm in (3.31) to yield
@Yt
rt = : (3.36)
@Kt
This equation contains the interest rate and the capital stock as endogenous variable.
Our …nal equation is the resource constraint (3.30), which provides a link between
capital and consumption. Hence, (3.35), (3.36) and (3.30) gives a system in three
equations and three unknowns. When we insert the interest rate into the optimality
condition for consumption , we obtain as our reduced form
@Yt+1
u0 (Ct ) = 1 + @K u0 (Ct+1 ) ;
t+1 (3.37)
Kt+1 = (1 ) Kt + Y t Ct :
This is a two-dimensional system of non-linear di¤erence equations which gives a
unique solution for the time path of capital and consumption, provided we have two
initial conditions K0 and C0 :
44
Transitional dynamics
at+1 = (1 + rt ) at + wt et :
where indicates a product, i.e. is=0 (1 + rt+i ) = 1+rt for i = 0 and is=0 (1 + rt+i ) =
(1 + rt+i ) (1 + rt ) for i > 0: For i = 1; is=0 (1 + rt+i ) = 1 by de…nition.
Letting the limit be zero, we obtain
1 et+i wt+i 1 e w 1 e w
at = i=0 i
= =t t =t
s=0 (1 + rt+s ) s=0 (1 + rt+s ) R
where the last but one equality substituted t + i by : We can write this as
1 e 1 w
=t = at + =t : (3.38)
R R
With a constant interest rate, this reads
1 ( t+1) 1 ( t+1)
=t (1 + r) e = at + =t (1 + r) w :
45
Equation (3.38) is the intertemporal budget constraint that results from a dynamic
budget constraint as speci…ed in (??). Comparing it with (3.3) shows that the present
values on both sides discount one time more than in (3.3). The economic di¤erence
lies in whether we look at values at the beginning or the end of a period.
The budget constraint (??) and (3.33) are the “natural”budget constraint in the
sense that it can be derived easily as above in (??) and in the sense that it easily
aggregates to economy wide resource constraints. We will therefore work with them in
what follows. The reason for not working with them right from the beginning is that
the intertemporal version (3.3) has some intuitive appeal. Showing its shortcomings
explicitly should make clearer why - despite its intuitive appeal - it is not so useful
in general equilibrium setups.
K +1 = Y (K ; L ) + (1 )K C : (3.40)
The Lagrangian
This setup di¤ers from the ones one got to know before in that there is an in…nity
of constraints in (3.40). This constraint holds for each point in time t: As a
consequence, the Lagrangian is formulated with in…nitly many Lagrangian multipliers,
1 t 1
L= =t u (C ) + =t f [K +1 Y (K ; L ) (1 ) K + C ]g : (3.41)
t
The …rst part of the Lagrangian is standard, 1=t u (C ), it just copies the ob-
jective function. The second part consists in a sum from t to in…nity, one constraint
46
for each point in time, each multiplied by its own Lagrange multiplier : In order to
make the maximization procedure clearer, we rewrite the Lagrangian as
1 t
L= =t u (C ) + s =t2 [K +1 Y (K ; L ) (1 )K + C ]
+ s 1 [Ks Y (Ks 1 ; Ls 1 ) (1 ) Ks 1 + Cs 1 ]
+ s [Ks+1 Y (Ks ; Ls ) (1 ) Ks + Cs ]
1
+ =s+1 [K +1 Y (K ; L ) (1 )K + C ];
This expression has the same interpretation as e.g. (3.5) or (3.20). When we replace s
by ; this equation with the constraint (3.40) is a two-dimensional di¤erence equation
system which allows to determine the paths of capital and consumption, given two
boundary conditions, which the economy will follow when factor allocation is opti-
mally chosen. The steady state of such an economy is found by setting Cs = Cs+1
and Ks = Ks+1 in (3.45) and (3.40).
Insert the constraint (3.40) into the objective function (3.39) and …nd
1 t
Ut = =t u (Y (K ; L ) + (1 )K K +1 ) ! max
fK g
This is now maximized by choosing a path fK g for capital. Choosing the state
variable implicitly pins down the path fC g of the control variable consumption and
47
one can therefore think of this maximization problem as one where consumption is
optimally chosen.
Now rewrite the objective function as
s 2 t
Ut = =t u (Y (K ; L ) + (1 )K K +1 )
s 1 t
+ u (Y (Ks 1 ; Ls 1 ) + (1 ) Ks 1 Ks )
s t
+ u (Y (Ks ; Ls ) + (1 ) Ks Ks+1 )
1 t
+ =s+1 t u (Y (K ; L ) + (1 )K K +1 ) :
@Y (Ks ; Ls )
u0 (Cs 1 ) = u0 (Cs ) 1 + :
@Ks
When we now go back to the maximization procedure where the Lagrangian was
used, we see that the partial derivative of the Lagrangian with respect to Kt in
(3.43) captures how t changes over time. The simple reason why the Lagrangian is
maximized with respect to Kt is therefore that an aditional …rst-order condition is
needed as t needs to be determined as well.
In static maximization problems with two consumption goods and one constraint,
the Lagrangian is maximized by choosing consumption levels for both consumption
goods and by choosing the Lagrange multiplier. In the Lagrange setup above, we
choose in (3.42) to (3.44) both endogenous variables Kt and Ct plus the multiplier t
and thereby determine optimal paths for all three variables. Hence, it is a technical
- mathematical - reason that Kt is “chosen”, three unknowns simply require three
…rst-order conditions to be determined, but economically, the control variable is Ct
which is “economically chosen” while the state variable Kt adjusts indirectly as a
consequence of the choice of Ct :
48
Exercises chapter 3
Applied Intertemporal Optimization
Dynamic programming in discrete deterministic time
U = U (C; L) ;
wL = C:
Let the individual maximize utility with respect to consumption and the amount
of labour supplied.
(a) What is the optimal labour supply function (in implicit form)? How much
does an individual consume? What is the indirect utility function?
(b) Under what conditions does an individual increase labour supply when
wages rise? (No analytical solution)
(c) Assume higher wages lead to higher labour supply. Does disutility arising
from higher labour supply overcompensate utility from higher consump-
tion? Does utility rise if there is no disutility from working? Start from the
indirect utility function derived in a) and apply the proof of the envelope
theorem and the envelope theorem itself.
(a) Compute the derivative of the Bellman equation (3.19) without using the
envelope theorem. Hint: Compute the derivative with respect to the state
variable and then insert the …rst-order condition.
(b) Do the same with (3.26)
49
(a) Show that the intertemporal budget constraint
T 1 1 T 1 1
=t k=t e = at + =t k=t i (3.46)
1 + rk 1 + rk
(b) Under which conditions does the dynamic budget constraint imply the
intertemporal budget constraint?
(c) Now consider at+1 = (1 + rt ) at + wt et : What intertemporal budget
constraint does it imply?
7. Environmental economics
Imagine you are an economist only interested in maximizing the present value of
your endowment. You own a renewable resource, for example a piece of forest.
The amount of wood in your forest at a point in time t is given by xt . Trees
grow at b(xt ) and you harvest at t the quantity ct .
50
(a) What is the law of motion for xt ?
(b) What is your objective function if prices at t per unit of wood is given by
pt , your horizon is in…nity and you have perfect information?
(c) How much should you harvest per period when the interest rate is constant?
Does this change when the interest rate is time-variable?
51
Part II
Deterministic models in continuous
time
4 Di¤erential equations
There are many excellent textbooks on di¤erential equations. This chapter will there-
fore be relatively short. Its objective is more to recall basic concepts taught in other
courses and to serve as a background for later applications.
52
4.1.2 Two theorems
Theorem 2 Existence (Brock and Malliaris, 1989)
If f (t; x) is continuous function on rectangle L = f(t; x)j jt t0 j a; jx x0 j bg
then there exists a continuous di¤erentiable solution x (t) on interval jt t0 j a that
solves initial value problem
This theorem only proves that a solution exists. It is still possible that there are
many solutions.
Theorem 3 Uniqueness
If f and @f =@x are continuous functions on L; the initial value problem (4.3) has a
unique solution for
b
t 2 t0 ; t0 + min a;
max jf (t; x)j
Unique …xpoint
Let f (x) be represented by the graph in the following …gure, with x (t) being
shown on the horizontal axis. As f (x) gives the change of x (t) over time, x_ (t) is
plotted on the vertical axis.
53
N
x(t)
_
N
N
N
N
N
x x(t)
As in the analysis of di¤erence equations in ch. 2.5.4, we …rst look for the …xpoint
of the underlying di¤erential equation. A …xpoint is de…ned similarily in spirit but -
thinking now in continuous time - di¤erently in detail: A …xpoint x is a point where
x (t) does not change; this is the similarity in spirit. Formally, this means x_ (t) = 0
which, from the de…nition (4.2) of the di¤erential equation, requires f (x ) = 0 (and
not f (x ) = x as in (2.38)). Looking at the above graph of f (x) ; we …nd x at the
point where f (x) crosses the horizontal line.
We then inquire the stability of the …xpoint. When x is to the left of x ; f (x) > 0
and therefore x increases, x_ (t) > 0: This increase of x is represented in this …gure by
the arrows on the horizontal axis. Similarily, for x > x ; f (x) < 0 and x (t) decreases.
We have therefore found that the …xpoint x is globally stable and have also obtained
a “feeling”for the behaviour of x (t) ; given some initial conditions.
We can now qualitatively plot the solutions with time t on the horizontal axis.
As the discussion has just shown, the solution x (t) depends on the initial value from
which we start, i.e. on x (0) : For x (0) > x ; x (t) decreases, for x (0) < x ; x (t)
increases: any changes over time are monotonic. There is one solution for each initial
condition. The following …gure shows three solutions of x_ (t) = f (x (t)), given three
di¤erent intitial conditions.
54
N
x(t)
x03
x02
x01
N
t0 t
Figure 13 Qualitative solutions of an ODE for di¤erent initial conditions I
Of course, more sophisticated functions than f (x) can be imagined. Consider now
a di¤erential equation x_ (t) = g (x (t)) where g (x) is non-monotonic as plotted in the
next …gure.
N
x
N
N
N
N
x1 x2 x3 x4 x
As this …gure shows, there are four …xpoints. Looking at whether g (x) is positive
or negative, we know whether x (t) increases or decreases over time. This allows us
to plot arrows on the horizontal axis as in the previous example. The di¤erence to
before consists in the fact that some …xpoints are unstable and some are stable. Any
small deviation of x from x1 implies an increase or decrase of x: The …xpoint x1 is
55
therefore unstable, given the de…nition in ch. 2.5.4. Any small deviation x2 ; however,
implies that x moves back to x2 : Hence, x2 is (locally) stable. The …xpoint x3 is also
unstable, while x4 is again locally stable: While x converges to x4 for any x > x4 (in
this sense x4 could be called globally stable from the right), x converges to x4 from
the left only if x is not smaller than or equal to x3 :
If an economy can be represented by such a di¤erential equation, one would call
…xpoints long-run equilibria. There are stable equilibria and unstable equilibria and
it depends on the underlying system (the assumptions that implied the di¤erential
equation x_ = g (x)) which equilibrium would be considered to be the economically
relevant one.
As in the system with one …xpoint, we can qualitatively plot solutions of x (t)
over time, given di¤erent initial values for x (0). This is shown in the following …gure
which again highlights the stability properties of …xpoints x1 to x4 :
x4
x3
x2
x1
N
t0 t
The system
56
Consider the following di¤erential equation system,
Assume that for economic reasons we are interested in properties for xi > 0:
Fixpoint
x_ 1 = x_ 2 = 0 , (x1 ) = x2 ; x2 = b + (x1 ) 1 :
By inserting the second equation into the …rst, x1 is determined by (x1 ) = b + (x1 ) 1
and x2 follows from x2 = (x1 ) : Analysing the properties of the equation (x1 ) = b +
(x1 ) 1 would then show that x1 is unique: The left-hand side increases monotonically
from 0 to in…nity for x1 2 [0; 1[ while the right-hand side decreases monotonically
from in…nity to b: Hence, there must be an intersection point and there can be only
one as functions are monotonic. As x1 is unique, so is x2 = (x1 ) :
General evolution
Having derived the …xpoint, we now need to understand the behaviour of the
system more generally. What happens to x1 and x2 when (x1 ; x2 ) 6= (x1 ; x2 )? To
answer this question, the concept of zero-motion lines is very useful. A zero-motion
line is a line for a variable xi which marks the points for which the variable xi does
not change, i.e. x_ i = 0: For our two-dimensional di¤erential equation system, we
obtain two zero-motion lines,
x_ 1 0 , x2 x1 ;
(4.4)
x_ 2 0 , x2 b + x1 1 :
In addition to the equality sign, we also analyse here at the same time for which
values xi rises. Why this is useful will soon become clear. We can now plot the
curves where x_ i = 0 in a diagram. In contrast to the one-dimensional graphs in the
previous chapter, we now have the variables x1 and x2 on the axes (and not the change
of one variable on the vertical axis). The intersection point of the two zero-motion
lines gives the …x point x = (x1 ; x2 ) which we derived analytically above.
57
x2
x&1 = 0
IV
III
II
x& 2 = 0
b
x1
In addition to showing where variables do not change, the zero-motion lines also
delimit regions where variables do change. Looking at (4.4) again shows (why we
used the and not the = sign and) that the variable x1 increases whenever x2 < x1 .
Similarily, the variable x2 increases, whenever x2 < b + x1 1 : The directions in which
variables change can then be plotted into this diagram by using arrows. In this
diagram, there are two arrows per region as two directions (one for x1 ; one for x2 )
need to be indicated. This is in principle identical to the arrows we used in the
analysis of the one-dimensional systems. If the system …nds itself in one of these four
regions, we know qualitatively, how variable change over time: Variables move to the
southeast in region I, to the northeast in region II, to the northwest in region III and
to the southwest in region IV. We will analyse the implications of such a behaviour
further once we have generalized the derivation of a phase diagram.
The system
Consider two di¤erential equations where functions f (:) and g (:) are continuous
and di¤erentiable,
x_ 1 = f (x1 ; x2 ) ; x_ 2 = g (x1 ; x2 ) : (4.5)
For the following analysis, we will need four assumptions on partial derivatives; all of
them are positive apart from fx1 (:) ;
fx1 (:) < 0; fx2 (:) > 0; gx1 (:) > 0; gx2 (:) > 0: (4.6)
58
Note that, provided we are willing to make the assumptions required by the theorems
in ch. 4.1.2, we know that there is a unique solution to this di¤erential equation
system, i.e. x1 (t) and x2 (t) are unambiguously determined given two boundary
conditions.
Fixpoint
x2
f (x1 , x2 ) = 0
x2*
g (x1 , x2 ) = 0
x1* x1
By the implicit function theorem - see (2.14) - and the assumptions made in (4.6),
one zero motion line is increasing and one is decreasing. If we are further willing to
assume that functions are not monotonically approaching an upper and lower bound,
we know that there is a unique …xpoint (x1 ; x2 ) x :
General evolution
Now we ask again what happens if the state of the system di¤ers from x ; i.e. if
either x1 or x2 or both di¤er from their steady state values. To …nd an answer, we
have to determine the sign of f (x1 ; x2 ) and g (x1; x2 ) for some (x1 ; x2 ) : Given (4.5); x1
would increase for a positive f (:) and x2 would increase for a positive g (:) : For any
known functions f (x1 ; x2 ) and g (x1; x2 ) ; one can simply plot a 3-dimensional …gure
with x1 and x2 on the axes in the plane and with time derivatives on the vertical axis.
59
Figure 18 A three-dimensional illustration of two di¤erential equations and their
zero-motion lines
When working with two-dimensional …gures and without the aid of computers, we
start from the zero-motion line for, say, x1 :
N
x2
-
x1;rx~2 )
(~ -
f (x1 ; x2 ) = 0
-
-
N
x1
Figure 19 Step 1 in constructing a phase diagram
60
N
x2
6g(x ; x ) = 0
1 2
r
6
?
(~
x1; x~2 ) ? 6
? 6
N
x1
Figure 20 Step 2 in constructing a phase diagram
We can now represent the directions in which x1 and x2 are moving into one
phase-diagram by plotting one arrow each into each of the four regions limited by the
zero-motion lines. Given that the arrows can either indicate an increase or a decrease
for both x1 and x2 ; there are 2 times 2 di¤erent combinations of arrows. When we
add some representative trajectories, a complete phase diagram results.
x2 x&1 = 0
x&2 = 0
x1
61
4.2.4 Types of phase diagrams and …xpoints
Some de…nitions
As has become clear by now, the partial derivatives in (4.6) are crucial for the
slope of the zero-motion lines and for the direction of movements of variables x1 and
x2 : Depending on the signs of the partial derivatives, various phase diagrams can
occur. These phase diagrams can be classi…ed into four typical groups, depending on
the properties of their …xpoint.
Some graphs
62
This …rst phase diagram shows a node. A node is a …xpoint through which all
trajectories go and where the time paths implied by trajectories are monotonic. As
drawn here, it is a stable node, i.e. for any initial conditions, the system ends up in
the …xpoint. An unstable node is a …xpoint from which all trajectories start. A phase
diagram for an unstable node would look like the one above but with all directions of
motions reversed.
(iii) focus
- all integral curves move
in spirals around fixpoint.
Either towards this point
or away.
focus
(iii) a and b stable or unstable
A phase diagram with a focus looks similar to one with a node. The di¤erence
consists in the non-monotonic paths of the trajectories. As drawn here, x1 or x2 …rst
increase and then decrease on some trajectories.
(iv) centre
- no integral curve passes
through fixpoint
centre
predator-prey-model
63
A circle is a very special case for a di¤erential equation system. It is rarely found
in models with optimizing agents. The standard example is the predator-prey model,
x_ = x xy, y_ = y + xy, where , , , and are positive constants. This is
also called the Lotka-Volterra model. No closed form solution has been found so far.
Limitations
It should be noted that a phase diagram analysis allows to identify a saddle point.
If no saddle point can be identi…ed, it is generally not possible to distinguish between
a node, focus or center. In the linear case, more can be deduced from a graphical
analysis. This is generally not necessary, however, as there is a closed form solution.
where a (t) and b (t) are functions of time. This is the most general case for a one
dimensional linear di¤erential equation.
Integrals
R
De…nition 7 A function F (x) f (x) dx is the inde…nite integral (Stammfunk-
tion) of a function f (x) if
Z
d d
F (x) = f (x) dx = f (x) (4.8)
dx dx
This de…nition implies that there are in…nitely many integrals of f (x) : If F (x) is
an integral, then F (x) + c; where c is a constant, is an integral as well.
64
Leibniz rule
We present here a rule for computing the derivative of an integral function. Let
there be a function z (x) with argument x; de…ned by the integral
Z b(x)
z (x) f (x; y) dy;
a(x)
where a (x) ; b (x) and f (x; y) are di¤erentiable functions. Then, the Leibniz rule
says that the derivative of this function with respect to x is
Z b(x)
d 0 0 @
z (x) = b (x) f (x; b (x)) a (x) f (x; a (x)) + f (x; y) dy: (4.9)
dx a(x) @x
f (x, y )
z (x + ∆x )
f (x, y )
z (x )
z (x )
a (x ) b(x ) y
Let x increase by a small amount. Then the integral changes at three margins:
The upper bound, the lower bound and the function f (x; y) itself. As drawn here,
the upper bound b (x) and the function f (x; y) increase and the lower bound a (x)
decreases in x. As a consequence, the area below the function f (:) between bounds
a (:) and b (:) increases because of three changes: the increase to the left because of
a (:) ; the increase to the right because of b (:) and the increase upwards because of
f (:) itself. Clearly, this …gure changes when the derivatives of a (:), b (:) and f (:)
with respect to x have a di¤erent sign than the ones assumed here.
65
Proof. We start by observing that
0 0 0
(u (x) v (x)) = u (x) v (x) + u (x) v (x) ;
R
where we used the product rule. Integrating both sides, dx; gives
Z Z
0 0
u (x) v (x) = u (x) v (x) dx + u (x) v (x) dx:
66
To obtain one particular solution, some value x (t0 ) has to be …xed. Depending
on whether t0 lies in the future or in the past, t0 > t or t0 < t, the equation is solved
forward or backward.
We start with the backward solution, i.e. where t > t0 . Let x (t0 ) = x0 : Then the
solution of (4.7) is
Rt Z t R
a( )d a(u)du
x (t) = e t0 x (t0 ) + e t0 b( )d
t0
Rt Z t R
t
t0 a( )d
=e x (t0 ) + e a(u)du b ( ) d : (4.12)
t0
Veri…cation
We now show that (4.12) and (4.13) are indeed a solution for (4.7). The time
derivative of (4.12) is given by
Rt Rt
Z t R
t
t0 a( )d a(u)du
x_ (t) = e a (t) x (t0 ) + e t b (t) + e a(u)du b ( ) d a (t)
t0
Rt Z t R
a( )d t
= a (t) e t0 x (t0 ) + e a(u)du b ( ) d + b (t)
t0
= a (t) x (t) + b (t) :
The principle
67
The non-autonomous di¤erential equation x_ = f (t; x) can be written equivalently
as an integral equation. To this end, write this equation as dxR= f (t; x) dt or, after
t
substituting s for t, as dx = f (s; x) ds: Now apply the integral 0 on both sides. This
gives the integral version of the di¤erential equation x_ = f (t; x) which reads
Z t Z t
dx = x (t) x (0) = f (s; x) ds:
0 0
An example
Computing the derivative of this equation with respect to time t gives, using (4.9),
x_ (t) = a (t) x (t) again.
The presence of an integral in (4.14) should not lead one to confuse (4.14) with
a solution Rof x_ = a (t) x in the sense of the last section. Such a solution would read
t
x (t) = x0 e 0 a(s)ds .
Let us …rst, following ch. 2.5.5, derive the dynamic budget constraint. We start
from the de…nition of nominal wealth. It is given by a = kv; where k is the households
physical capital stock and v is the value of one unit of the capital stock. By computing
the time derivative, wealth of a household changes according to
_ + k v:
a_ = kv _
If the household wants to save, it can buy capital goods at the end of period t:
Nominal savings of the household in period t are given by s = wk k + w e; the
di¤erence between factor rewards for capital (value marginal product times capital
owned), labour income and expenditure. Dividing savings by the value of a capital
stock at the end of period t gives the number of capital stocks bought
wk k + w e
k_ = : (4.15)
v
68
Inserting into the a_ equation gives
wk + v_
a_ = wk k + w e + k v_ = (wk + v)
_ k+w e= a+w e:
v
De…ning the nominal interest rate as
wk + v_
i= ; (4.16)
v
we have the nominal budget constraint
We can now obtain the intertemporal budget constraint from solving the dynamic
one in (4.17). Using the forward solution from (4.13), we take a (T ) as the boundary
condition lying with T > t in the future. The solution is then
RT
Z T R
i( )d
a (t) = e t a (T ) e t i(u)du [w ( ) e ( )] d ,
t
Z T Z T
D ( ) w ( ) d + a (t) = D (T ) a (T ) + D( )e( )d ;
t t
R
where D ( ) e t i(u)du de…nes the discount factor. As we have used the forward
solution, we have obtained an expression which easily lends itself to an economic
interpretation. Think of an individual who - at the end of his life - does not want to
leave any bequest, i.e. a (T ) = 0 or, for a positive a (T ) ; does not want to use any
of this a (T ) for own consumption. In both cases, a (T ) = 0 in the above expression.
Then, this intertemporal budget constraint requires that current wealth on the left-
hand side, consisting of the present value of life-time labour income plus …nancial
wealth a (t) needs to equal the present value of current and future expenditure on the
right-hand side.
Real wealth
We can also start from the de…nition of real wealth of the household, measured in
units of the consumption good, whose price is p. Real wealth is then
kv
ar = :
p
The change in real wealth over time is then (apply the log on both sides and derive
with respect to time),
a_ r k_ v_ p_
= + :
ar k v p
69
Inserting the increase in the capital stock obtained from (4.15) gives
a_ r wk k + w e p_ v_(wk k + w e) p 1 v_ p_
r
= + = 1
+
a vk p v vkp v p
k w e v_ r p_ r wk r w e v_ r p_ r
, a_ r = wk + + a a = a + + a a
p p v p v p v p
wk + v_ p_ w e w e
= ar + = rar + : (4.18)
v p p p
Hence, the real interest rate r is - by de…nition -
wk + v_ p_
r= :
v p
Solving the di¤erential equation (4.18) again provides the intertemporal budget con-
straint as in the nominal case above.
Now assume that the price of the capital good equals the price of the consumption
good, v = p: This is the case in an economy where there is one homogenous output
good as in (2.26) or in (8.14). Then, the real interest rate is equal to the marginal
product of capital, r = wk =p:
Imagine you own wealth of worth v (t) : You can invest it on a bank account which
pays a certain return r (t) per unit of time or you can buy shares of a …rm which cost
v (t) and which yield dividend payments (t) and are subject to changes v_ (t) in its
worth. In a world of perfect information and assuming that in some equilibrium
agents hold both assets, the two assets must yield identical income streams,
This forward solution stresses the economic interpretation of v (t) : The value of
an asset depends on the future income stream - dividend payments ( ) - that are
70
generated from owning this asset. Note that it is usually assumed that there are no
bubbles, i.e. the limit is zero such that the fundamental value of an asset is given by
the …rst term.
For a constant interest rate and dividend payments and no bubbles, the expression
for v (t) simpli…es to v = =r:
The utility function
Consider an intertemporal utility function as it is often used in continuous time
models, Z 1
U (t) = e [ t] u (c ( )) d : (4.20)
t
This is the standard expression which corresponds to (2.11) or (3.1) in discrete time.
Again, instantaneous utility is given by u (:) : It depends here on consumption only,
where households consume continuously at each instant : Impatience is captured as
before by the time preference rate : Higher values attached to present consumption
are captured by the discount function e [ t] ; whose discrete time analog in (3.1) is
t
:
Using (4.9), di¤erentiating with respect to time gives
Z 1
d
U_ (t) = u (c (t)) + e [ t] u (c ( )) d = u (c (t)) + U (t) :
t dt
Solving this linear di¤erential equation forward gives
Z T
[T t]
U (t) = e U (T ) + e [ t] u (c ( )) d :
t
71
4.4 Linear di¤erential equation systems
A di¤erential equation system consists of 2 or more di¤erential equations which are
mutually related to each other. An example is
x_ (t) = Ax (t) + b;
72
Exercises chapter 4
Applied Intertemporal Optimization
Using phase diagrams
1. Phase diagram I
Consider the following di¤erential equation system,
x_ 1 = f (x1 ; x2 ) ; x_ 2 = g (x1 ; x2 ) :
Assume
dx2 dx2
fx1 (x1 ; x2 ) < 0; gx2 (x1 ; x2 ) < 0; < 0; > 0:
dx1 f (x1 ;x2 )=0 dx1 g(x1 ;x2 )=0
y
y& = 0
x& = 0
O
x
73
4. Local stability analysis
Study local stability properties of the …xpoint of the di¤erential equation system
(4.21).
(a) t > s;
(b) t < s:
(c) What is the forward, what is the backward solution? How do they relate
to each other?
7. Derivatives of integrals
Compute the following derivatives.
Ry
(a) ddy a f (s) ds;
Ry
(b) ddy a f (s; y) ds;
Rb
(c) ddy a f (y) dy;
R
(d) ddy f (y) dy:
Rb Rb
(e) Show that a xydt_ = [xy]ba a
xydt:
_
74
where Z
Dr ( ) = exp r (s) ds : (4.23)
t
A dynamic budget constraint reads
E (t) + A_ (t) = r (t) A (t) + I (t) : (4.24)
(a) Show that solving the dynamic budget constraint yields the intertemporal
budget constraint if and only if
Z T
lim A (T ) exp r ( ) d = 0: (4.25)
T !1 t
(b) Show that di¤erentiating the intertemporal budget constraint yields the
dynamic budget constraint.
(a) Compute time paths of the number ni (t) of …rms in country i: The laws
of motion are given by (Grossman and Helpman, 1991; Wälde, 1996)
n_ i = nA + nB Li ni (L + ) ; i = A; B; L = L A + LB ; ; > 0:
Hint: Eigenvalues are g = (1 )L > 0 and = (L + ).
A
(b) Plot the time path of n . Choose appropriate initial conditions.
75
5 Optimal control theory - Hamiltonians
One widely used approach to solve deterministic intertemporal optimization problems
in continuous time is to use the so-called Hamiltonian function. Given a certain
maximization problem, this function can be adapted - just like a recipe - to yield a
straightforward solution. The …rst section will provide an introductory example. It
shows how easy it can sometimes be to solve a maximization problem.
It is useful to understand, however, where the Hamiltonian comes from. A list
of examples can never be complete, so it helps to be able to derive the appropriate
optimality conditions in general. This will be done in the subsequent section. Section
5.3 then discusses how boundary conditions for maximization problems look like and
how they can be motivated. The in…nite planning horizon problem is then presented
and solved in section 5.4 which includes a section on transversality and bounded-
ness conditions. Various examples follow in section 5.5. Section 5.6 …nally shows
how to work with present-value Hamiltonians and how they relate to current-value
Hamiltonians (which are the ones used in all previous sections).
76
The …rst-order condition in (5.3) is a usual optimality condition: the derivative of the
Hamiltonian (5.2) with respect to the consumption level c must be zero. The second
optimality condition - at this stage - just comes “out of the blue”. Its origin will be
discussed in a second. Applying logs to the …rst …rst-order condition, ln c = ln ;
and computing derivatives with respect to time yields c=c _ = _ = : Inserting into
(5.4) gives the Keynes-Ramsey rule
c_ c_
= r, =r : (5.5)
c c
y_ ( ) = Q (y ( ) ; z ( ) ; ) : (5.7)
The utility function u (:) and other functions are presented as u ( ) : This shortens
notation compared to full expressions in (5.6) and (5.7). The intuition behind this
Lagrangian is similar to the one behind the Lagrangian in the discrete time case in
(3.41) in ch. 3.7 where we also looked at a setup with many constraints. The …rst
part is simply the objective function. The second part refers to the constraints. In the
discrete-time case, each point in time had its own constraint with its own Lagrange
multiplier. Here, the constraint (5.7) holds for a continuum of points . Hence,
instead of the sum in the discrete case we now have an integral over the product of
multipliers ( ) and constraints.
77
This Lagrangian can be rewritten as follows,
Z T Z T
[ t]
L= e u( ) + ( )Q( )d ( ) y_ ( ) d
t t
Z T Z T
= e [ t]
u( ) + ( )Q( )d + _ ( )y( )d [ ( ) y ( )]Tt (5.8)
t t
where the last step integrated by parts and [ ( ) y ( )]Tt is the integral function of
( ) y ( ) evaluated at T minus its level at t. Now assume that we could choose
not only the control variable z ( ) but also the state variable y ( ) at each point in
time. The intuition for this is the same as in discrete time in ch. 3.7.2. We maximize
therefore this Lagrangian with respect to y and z at one particular 2 [t; T ].
For the control variable z; we get a …rst-order condition
[ t] [ t]
e uz ( ) + ( ) Qz ( ) = 0 , uz + e ( ) Qz = 0:
When we de…ne
[ t]
( ) e ( ); (5.9)
we …nd
uz + ( ) Qz = 0: (5.10)
For the state variable y; we obtain
[ t]
e uy ( ) + ( ) Qy ( ) + _ ( ) = 0 ,
[ t] [ t]
uy ( ) ( )e Qy = e _ ( ): (5.11)
Equations (5.10) and (5.12) are the two optimality conditions that solve the above
maximization problem jointly with the constraint (5.7). We have three equations
which …x three variables: The …rst condition (5.10) determines the optimal level of
the control variable z: As this optimality condition holds for each point in time ; it
…xes an entire path for z: The second optimality condition (5.12) …xes a time path for
: It makes sure, by letting the costate follow an appropriate path, that the level
of the state variable (which is not instantaneously adjustable as the maximization of
the Lagrangian would suggest) is as if it had been optimally chosen at each instant.
The constraint (5.7), …nally, …xes the time path for the state variable y:
78
5.2.2 Hamiltonians as a shortcut
Let us now see how Hamiltonians can be justi…ed. De…ne the Hamiltonian similar to
(5.2),
H = u( ) + ( )Q( ): (5.13)
In fact, this Hamiltonian shows the general structure of Hamiltonians. Take the
instantaneous utility level (or any other function behind the discount term in the
objective function) and add the costate variable times the right-hand side of the
constraint. Optimality conditions are then
Hz = 0; (5.14)
_ = Hy : (5.15)
These conditions were already used in (5.3) and (5.4) in the introductory example
in the previous chapter and in (5.10) and (5.12): When (5.13) is inserted into (5.14)
and (5.15), this yields equations (5.10) and (5.12). Hamiltonians are therefore just
a shortcut that allow to …nd solutions faster than in the case where Lagrangians are
used. Note for later purposes that both and have time as an argument.
There is an interpretation of the costate variable which we simply state at this
point (see ch. 6.2 for a formal derivation): The derivative of the objective function
with respect to the state variable at t, evaluated on the optimal path, equals the
value of the corresponding costate variable at t. Hence, just as in the static Lagrange
case, the costate variable measures the change in utility as a result of a change in
endowment (i.e. in the state variable). Expressing this formally, de…ne the value
function as V (y (t)) maxfz( )g U (t) ; identical in spirit to the value function in
dynamic programming as we got to know it in discrete time. The derivative of the
value function, the shadow price V 0 (y (t)) ; is then the change in utility when behaving
optimally resulting from a change in the state y (t) : This derivative equals the costate
variable, V 0 (y (t)) = .
79
We can use the current-value Hamiltonian (5.13) which gives optimality conditions
(5.14) and (5.15). In addition (cf. Feichtinger and Hartl, 1986, p. 20),
(T ) = 0: (5.17)
As is the shadow price of the state variable, this simply means that the value of
the state variable at T needs to be zero.
y (0) = y0 ; y (T ) = yT : (5.18)
In order for this problem to make sense, we assume that a feasible solution exists. This
“should”generally be the case, but it is not obvious: Consider again the introductory
example in ch. 5.1. Let the endpoint condition be given by “the agent is very rich in
T ”; i.e. a (T ) =“very large”. If a (T ) is too large, even zero consumption at each point
in time, c ( ) = 0 8 2 [t; T ] ; would not allow wealth a to be as large as required by
a (T ) : In this case, no feasible solution would exist.
We assume, however, that a feasible solution exists. Optimality conditions are
then identical to (5.14) and (5.15), plus initial and boundary values. The di¤erence
from this approach to the one before is that now y (T ) = yT is exogenously given, i.e.
part of the maximization problem, whereas before the corresponding y (T ) = 0 was
endogenously determined as a necessary condition for optimality.
80
5.4 The in…nite horizon
5.4.1 The maximization problem
In the in…nite horizon case, the objective function has the same structure as before in
e.g. (5.6) only that the …nite time horizon T is replaced by an in…nite time horizon
1. The constraints are unchanged and the maximization problem reads
Z 1
max e [ t] u (y ( ) ; z ( ) ; ) d ;
fz( )g t
subject to
y_ ( ) = Q (y ( ) ; z ( ) ; ) ;
y (0) = y0 : (5.19)
R1
We need to assume for this problem that the integral t e [ t] u (:) d converges
for all feasible paths of y ( ) and z ( ), otherwise the optimality criterion must be
rede…ned. This boundedness condition is important only in this in…nite horizon case.
If individuals have …nite horizon and the utility function u (:) is continuous over the
entire planning period, the objective function is …nite and the boundedness problem
disappears. Clearly, making such an assumption is not always innocuous and one
should check, at least after having solved the maximization problem, whether the
objective function converges indeed. This will be done in ch. 5.4.3.
The current-value Hamiltonian as in (5.13) is de…ned by
H = u (:) + ( ) Q (:) ;
@H _ = @H
= 0; :
@z @y
Hence, we have identical optimality conditions to the case of a …nite horizon.
where y (t) is the path of y (t) for an optimal choice of control variables.
81
A conceptionally di¤erent condition from the transversality condition is the “No-
Ponzi game”condition (4.25)
Z T
lim y (T ) exp r( )d =0
T !1 t
where r @H=@y. Note that this No-Ponzi game condition can be rewritten as
T
lim e (T ) y (T ) = lim (T ) y (T ) = 0; (5.20)
T !1 T !1
T R
where the fact that _ = = r implies e t = 0 e t r( )d was used. The for-
mulations in (5.20) of the No-Ponzi game condition is frequently encountered and is
sometimes even called TVC.
There is a considerable literature on the necessity and su¢ ciency of the TVC.
Some examples include Mangasarian (1966), Arrow (1968), Arrow and Kurz (1970),
Araujo and Scheinkman (1983), Léonard and Long (1992, p. 288-289), Chiang (1992,
p. 217, p. 252) and Kamihigashi (2001). Counterexamples that the TVC is not
necessary are provided by Michel (1982) and Shell (1969).
C ( )1 1
u (C ( )) = :
1
Assume that consumption grows with a rate of g, where this growth rate results from
utility maximization. Think of this g as representing e.g. the di¤erence between
the interest rate and the time preference rate, corrected by intertemporal elasticity
of substitution, as will be found later e.g. in the Keynes-Ramsey rule (5.33), i.e.
c=c
_ = (r )= g. With this exponential growth of consumption, the utility
function becomes
Z 1
1
U (t) = (1 ) C0 e [ t] e(1 )g[ t] 1 d :
t
82
This integral is bounded only if
(1 )g < 0:
This can formally be seen by computing the integral explicitly and checking under
which conditions it is …nite. Intuitively, this condition makes sense: Instantaneous
utility from consumption grows by a rate of (1 ) g: Impatience implies that future
utility is discounted by the rate : Only if this time preference rate is large enough,
the overall expression within the integral, e [ t] C ( )1 1 ; falls in :
This problem has been discovered a relatively long time ago and received renewed
attention in the 90s when the new growth theory was being developed. A more general
treatment of this problem was undertaken by von Weizsäcker (1965) who compares
utility levels in unbounded circumstances by using “overtaking criteria”.
5.5 Examples
This section presents various examples that show both how to compute optimality
conditions and how to understand the predictions of the optimality conditions. We
will return to a phase-diagram analysis and understand the meaning of …xed and free
values of state variables at the end points.
The model
A …rm maximizes the present value 0 of its future instantaneous pro…ts (t),
Z T
rt
0 = e (t) dt; (5.21)
0
Gross investment I (t) minus depreciation K (t) gives the net increase of the …rm’s
capital stock. Instantaneous pro…ts are given by the di¤erence between revenue and
cost,
(t) = pF (K (t)) (I (t)) :
Revenue is given by pF (K (t)) where the production technology F (:) employs capital
only. Output increases in capital input, F 0 (:) > 0; but to a decreasing extent, F 00 (:) <
83
0. Costs of …rm are given by the cost function (I (t)) : As the …rm owns capital, it
does not need to pay any rental costs for capital. The only costs that are captured
by (:) are adjustment costs. They include both the cost of buying and of installing
capital. The initial capital stock is given by K (0) = K0 . The interest rate r is
exogenous to the …rm.
Solution
0
= (r + ) pF (K (t)) ; (5.24)
0
(T ) = 0 , (I (T )) = 0:
The optimality condition for shows that the value marginal product of capital,
pF 0 (K (t)) ; still plays a role and it is still compared to the rental price r of capital,
but there is no longer an equality as in static models of the …rm. We will return to
this point in exercise 2 of ch. 6.
Optimality conditions can be presented in a simpler way (i.e. with less endogenous
variables). First, solve the …rst optimality condition for the costate variable and
compute the time derivative,
_ (t) = 00
(I (t)) I_ (t) :
Second, insert this into the second optimality condition (5.24) to …nd
00
(I (t)) I_ (t) = (r + ) 0
(I (t)) pF 0 (K (t)) : (5.25)
This equation, together with the capital accumulation constraint (5.22), is a two-
dimensional di¤erential equation system that can be solved, given initial conditions
and some boundary condition. As a solution to a di¤erential equation system is
usually a path of time variant variables, this shows that with adjustment costs, …rms
indeed have an intertemporal problem.
An example
The price to be paid per unit of capital is given by the constant v and costs of
installation are given by I 2 =2: This quadratic term captures the idea that installation
84
costs are low and do not increase fast, i.e. underproportionally to the new capital
stock, at low levels of I but increase overproptionally when I becomes large. Then,
optimality requires (5.22) and, from inserting (5.26) into (5.25),
A phase diagram using (5.22) and (5.27) is plotted in the following …gure.
I
K& = 0
I& = 0
K0 K
We now have to select one of this in…nity of paths that start at K0 : What is the
correct investment level I0 ? Given that the value of the state variable at the end
point is freely chosen, the optimal path satis…es
0
(T ) = 0 , (I (T )) = 0 , I (T ) = 0:
Hence, the trajectory where the investment level is zero at T is the optimal one.
85
5.5.3 In…nite horizon - Optimal consumption paths
Let us now look at an example with in…nite horizon. We focus on optimal behaviour
of a consumer. The classic reference is Ramsey (1928) or Arrow and Kurz (1969).
The problem can be posed in at least two ways. In either case, one part of the
problem is the intertemporal utility function
Z 1
U (t) = e [ t] u (c ( )) d : (5.28)
t
Due to the general instantaneous utility function u (c ( )), it is somewhat more general
than e.g. (5.1). The second part of the maximization problem is a constraint limiting
the total amount of consumption. Without such a constraint, maximizing (5.28)
would be trivial (or meaningless): With u0 (c ( )) > 0 maximizing the objective simply
means setting c ( ) to in…nity. The way this constraint is expressed determines in
which way the problem is most straightforwardly solved.
Solving by Lagrange
The constraint to (5.28) is given by a budget constraint. The …rst way in which
this budget constraint can be expressed is, again, the intertemporal formulation,
Z 1 Z 1
Dr ( ) E ( ) d = a (t) + Dr ( ) w ( ) d ; (5.29)
t t
R
where E ( ) = p ( ) c ( ) and Dr ( ) = e t r(u)du : The maximization problem is then
given by: maximize (5.28) by choosing a path fc( )g subject to the budget constraint
(5.29).
We build the Lagrangean with as the time-independent Lagrange multiplier
Z 1
L= e [ t] u (c ( )) d
t
Z 1 Z 1
Dr ( ) E ( ) d a (t) Dr ( ) w ( ) d :
t t
Note that in contrast to section 5.2.1 where a continuum of constraints implied a con-
tinuum of Lagrange multipliers (or, in an alternative interpretation, a time-dependent
multiplier) there is only one constraint here.
The …rst-order conditions are (5.29) and the partial derivative with respect to
consumption c ( ) at one speci…c point in time,
[ t] 0
Lc( ) = e u (c ( )) Dr ( ) p ( ) = 0
1 [ t] 1 1
, Dr ( ) e p( ) = u0 (c ( )) :
Note that this …rst order condition represents an in…nite number of …rst order con-
ditions: one for each point in time between t and in…nity. The integral is not part
of the …rst-order condition as maximization takes place seperatly for each . Note
86
the analogy to maximizing a sum as e.g. in (3.4). The integration variable here
corresponds to the summation index in (3.4). When one speci…c point in time is
chosen (say = 17), all derivatives of other consumption levels with respect to this
speci…c c are zero.
Applying logs yields
Z
r (u) du [ t] ln p ( ) = ln ln u0 (c ( )) :
t
In contrast to above, the utility function (5.28) here is maximized subject to the
dynamic (or ‡ow) budget constraint,
87
and the Keynes-Ramsey rule becomes, from inserting the CRRA u (c ( )) into (5.30),
c_ ( ) r
= : (5.33)
c( )
One e¤ect, the growth e¤ect, is straightforward from (5.33) or also from (5.30). A
higher interest rate, ceteris paribus, increases the growth rate of consumption. The
second e¤ect, the e¤ect on the level of consumption, is less obvious, however. In order
to understand it, we undertake the following steps. This is in principle similar to the
analysis of level e¤ects in ch. 3.5.2.
First, we solve the linear di¤erential equation in c ( ) given by (5.33). Following
ch. 4.3, we …nd
r
c ( ) = c (t) e ( t) : (5.34)
Consumption, starting today in t with a level of c (t) grows exponentially over time
at the rate (r ) = to reach the level c ( ) at some future > t:
In the second step, we insert this solution into the left-hand side of the budget
constraint (5.32) and …nd
Z 1 Z 1 h i1
r r 1 r
e r[ t]
c (t) e ( t)
d = c (t) e (r )[ t] d = c (t)
r e (r )[ t]
:
t t r t
The simpli…cation stems Rfrom the fact that c (t) ; the initial consumption level, can
1
be pulled out of the term t e r[ t] c ( ) d ; representing the present value of current
and future consumption expenditure. Please note that c (t) could be pulled out of
the integral also in the case of a non-constant interest rate. Note also that we do not
need to know what the level of c (t) is, it is enough to know that there is some c (t)
in the solution (5.34), whatever its level.
With a constant interest rate, the remaining integral can be solved explicitely.
First note that r r must be negative. Otherwise consumption growth would
exceed the interest rate and a boundedness condition for the objective function like in
ch. 5.4.3 would eventually be violated. Hence, we assume r > r , (1 )r < :
Clearly, this holds in a steady state where r = : Therefore, we obtain for the present
value of consumption expenditure
1 h r
i1 c (t) c (t)
c (t) r e (r )[ t]
= r =
r t r (1 )r
For the special case of a logarithmic utility function, the fraction in front of the curly
brackets simpli…es to (as = 1), a result obtained e.g. by Blanchard (1985).
After these two steps, we have two results, both visible in (5.35). One result
shows that initial consumption c (t) is a fraction out of wealth of the household.
88
Wealth needs to be understood in a more general sense than usually, however: It
is …nancial wealth a (t) plus, what could be called humanR wealth (in an economic,
1
i.e. material sense), the present value of labour income, t e r[ t] w ( ) d : Going
beyond t today and realizing that this analysis can be undertaken for any point in
time, the relationship (5.35) of course holds on any point of an optimal consumption
path. The second result is a relationship between the level of consumption and the
interest rate, our original question.
We now need to understand the derivative dc (t) =dr in order to exploit (5.35)
more. If we focus only on the term in front of the curly brackets, we …nd for the
change in the level of consumption when the interest rate changes
dc (t) 1
= f:g R 0 , R 1:
dr
The consumption level increases when the interest rate rises if is larger than one, i.e.
if the intertemporal elasticity of substitution 1 is smaller than unity. This is prob-
ably the empirically more plausible case (compared to < 1) on the aggregate level.
See however Vissing-Jørgensen (2002) for micro-evidence where the intertemporal
elasticity of substitution can be much larger than unity. This …nding is summarized
in the following …gure.
N r2 > r 1
lnc(t)
c2 (t) r1
c2 (t) >1
N
t time
Figure 27 An interest rate change and consumption growth and level for >1
This example studies the classic central planner problem: First, there is a social
welfare function like (5.1), expressed slightly more generally as
Z 1
max e [ t] u (C ( )) d :
fC( )g t
89
The generalization consists in the in…nite planning horizon and the general instanta-
neous utility function u (c ( )) : We will specify it in the most common version,
C1 1
u (C) = ; (5.36)
1
1
where the intertemporal elasticity of substition is constant and given by : The
00 0
coe¢ cient of relative risk aversion (de…ned by u (C) C=u (C)) is also constant and
given by : While the latter aspect is of importance only in models with uncertainty,
it will turn out that this close relationship between this measure of risk-aversion and
the intertemporal elasticity of substitution is not always desirable. A way out are
Epstein-Zin preferences.
Second, there is a resource constraint that requires that net capital investment is
given by the di¤erence between output Y (K; L), depreciation K and consumption
C;
K_ (t) = Y (K (t) ; L) K (t) C (t) : (5.37)
Assuming for simplicity that the labour force L is constant, this completely describes
this central planner problem.
The choice variable of the planner is the consumption level C (t) ; to be determined
for each point in time between today t and the far future 1. The fundamental trade-
o¤ consists in the utility increasing e¤ect of more consumption visible from (5.36) and
the net-investment descreasing e¤ect of more consumption visible from the resource
constraint (5.37). As less capital implies less consumption possibilities in the future,
the trade-o¤ can also be described to consist in more consumption today vs. more
consumption in the future.
The solution
90
instead of r ( ) p(
_ )
p( )
: Instead of the real interest rate on the household level, we have
here the marginal productivity of capital minus depreciation on the aggregate level.
If we assumed a logarithmic instantaneous utility function, u00 (C (t)) =u0 (C (t)) =
1=C (t) and the Keynes-Ramsey rule would look similar to (5.5) or (5.31). In our case
of the more general speci…cation in (5.36), we …nd u00 (C (t)) =u0 (C (t)) = =C (t)
such that the Keynes-Ramsey rule reads
C_ (t) YK (K (t) ; L)
= : (5.39)
C (t)
This could be called the classic result on optimal consumption in general equilib-
rium. Consumption grows if marginal productivity of capital exceeds the sum of the
depreciation rate and the time preference rate. The higher the intertemporal elas-
ticity of substition 1= ; the stronger consumption growth reacts to the di¤erences
YK (K (t) ; L) :
The resource constraint of the economy in (5.37) plus the Keynes-Ramsey rule
in (5.39) represent a two-dimensional di¤erential equation system which, given two
boundary conditions, give a unique solution for time paths C (t) and K (t) : These
two equations can be analyzed in a phase diagram. This is probably the most often
taught phase diagram in economics.
N
C C_ = 0
N
6 ?
N
N
C B K_ = 0
N
-
A ?
6
-
N
K0 K
Figure 28 Optimal central planner consumption
when the inequality signs hold as equalities. Zero motion lines are plotted in the
above …gure.
When consumption lies above the Y (K (t) ; L) K (t) line, (5.40) tells us that
capital increases, below this line, it decreases. When the marginal productivity of
91
capital is larger than + ; i.e. when the capital stock is su¢ ciently small, (5.41) tells
us that consumption increases. These laws of motion are plotted in …gure 28 as well.
This allows us to draw trajectories A; B and C which all satisfy (5.40) and (5.41).
Hence, again, we have a multitude of solutions for a di¤erential equation system.
As always, boundary conditions allow to “pick”the single solution to this system.
One boundary condition is the initial value K0 of the capital stock. The capital stock
is a state variable and therefore historically given at each point in time. It can not
jump. The second boundary condition should …x the initial consumption level C0 :
Consumption is a control or choice variable and can therefore jump or adjust to put
the economy on the optimal path.
The condition which provides the second boundary condition is, formally speaking,
the transversality condition (see ch. 5.4.2). Taking this formal route, one would have
to prove that starting with K0 at the consumption level that puts the economy on
path A would violate the TVC. Similarily, it would have to be shown that path C or
any path other than B violates the TVC as well. Even though not often admitted, this
is not often done in practice. (For an exception, see exercise 5 in ch. 5.) Whenever
a saddle path is found in a phase diagram, it is argued that the saddle path is the
equilibrium path and the initial consumption level is such that the economy …nds itself
on the path which approaches the steady state. While this is a practical approach, it
is also formally satis…ed as this path satis…es the TVC indeed.
Aggregate unemployment
The unemployment rate in an economy is governed by two factors: the speed with
which new employment is created and the speed with which existing employment is
destroyed. The number of new matches per unit of time dt is given by a matching
function. It depends on the number of unemployed U , i.e. the number of those
potentially available to …ll a vacancy, and the number of vacancies V , m = m (U; V ).
The number of …lled jobs that are destroyed per unit of time is given by the product
of the separation rate s and employment, sL. Combining both components, the
evolution of the number of unemployed over time is given by
U_ = sL m (U; V ) : (5.42)
De…ning employment as the di¤erence between the size of the labour force N and the
number of unemployed U , L = N U , we obtain
U_ = s [N U] m (U; V ) , u_ = s [1 u] m (u; V =N ) ;
where we divided by the size of the labour force N in the last step to obtain the
unemployment rate u U=N: We also assumed constant returns to scale in the
92
matching function m (:) : Note that the matching function can further be rewritten
as
V V V U
m u; = m 1; u= m ;1 u q( )u (5.43)
N U U V
and we obtain u_ = s [1 u] q ( ) u which is equation (1.3) in Pissarides (2000).
Clearly, from (5.42), one can easily obtain the evolution of employment, using
again the de…nition L = N U;
L_ = m (N L; V ) sL: (5.44)
Firm behaviour
Given this matching process, the choice variable of a …rm i is no longer employment
Li but the number of vacancies Vi it creates. Each vacency implies costs measured
in units of the output good Y: The …rm also employs capital Ki : Hence, the …rm’s
optimization problem reads
Z 1
max e r[ t] fY (Ki ; Li ) rKi wLi Vi g d
fVi ( );Ki ( )g t
subject to
Vi
L_ i = m (N sLi :
L; V )
V
The constraint now says - in contrast to (5.44) - that only a certain share of all
matches goes to the …rm under consideration and that this share is given by the
share of the …rm’s vacancies in total vacancies, Vi =V . Alternatively, this says that
the “probability” that a match in the economy as a whole …lls a vacency of …rm i
is given by the number of matches in the economy as a whole divided by the total
number of vacancies. As under constant returns to scale for m and using the de…nition
of q ( ) implicity in (5.43), m (N L; V ) =V = m (U=V; 1) = q ( ) ; we can write the
…rm’s constraint as
L_ i = q ( ) Vi sLi :
Assuming small …rms, this rate can q ( ) safely be assumed to be exogenous for the
…rm’s maximization problem.
The current-value Hamiltonian for this problem reads
The …rst condition for capital is the usual marginal productivity condition applied
to capital input. With CRTS production functions, this condition …xes the capital
to labour ratio for each …rm. This implies that the marginal product of labour is a
93
function of the interest only and therefore identical for all …rms, independent of the
labour stock, i.e. the size of the …rm. This changes the third condition to
_ = (r + s) YL (K; L) + w (5.45)
which means that the shadow prices are identical for all …rms.
The …rst-order condition for vacencies, written as = q ( ) says that the mar-
ginal costs of a vacancy (which in this special case equal average and unit costs)
must be equal to revenue from a vacancy. This expected revenue is given by the
share q ( ) of vacencies that yield a match times the value of a match. The value of
a match to the …rm is given by ; the shadow price of labour. The link between
and the value of an additional unit of labour (or of the state variable, more generally
speaking) is analyzed in ch. 6.2.
General equilibrium
q_ ( ) q( )
=r+s [YL (K; L) w] (5.46)
q( )
q_ ( ) U_ V_ V_ U_ V_
= + = (1 ) :
q( ) U V V U V
94
Hence, the vacency equation (5.46) becomes
U_ V_ q( )
+ (1 ) =r+s [YL (K; L) w] ,
U V
V_ (N L) L V
(1 ) =r+s [YL (K; L) w] + s
V V 1 N L (N L)1
subject to
y_ ( ) = Q (y ( ) ; z ( ) ; ) (5.48)
y (t) = yt ; y (T ) free (5.49)
where y ( ) 2 Rn ; z ( ) 2 Rm and Q = (Q1 (y ( ) ; z ( ) ; ) ; Q2 (:) ; ::: ; Qn (:))T : A
feasible (zulässig) path is a pair (y ( ) ; z ( )) which satis…es (5.48) and (5.49). z( )
is the vector of control variables, y( ) is the vector of state variables.
Then de…ne the (present-value) Hamiltonian H as
HP = F ( ) + ( ) Q ( ) ; (5.50)
HzP = 0; (5.52)
_( )= HyP ; (5.53)
and
(T ) = 0;
(Kamien and Schwartz, p. 126) in addition to (5.48) and (5.49). In order to have a
maximum, we need second order conditions Hzz < 0 to hold (Kamien and Schwartz).
95
Understanding its structure
z( ) = z(y( ); ( )):
Let the objective function be (5.47) that is to be maximized subject to the con-
straint (5.48) and, in addition, a static constraint
G (y ( ) ; z ( ) ; ) = 0: (5.54)
96
This is now maximized with respect to y ( ) and z ( ), both the control and the
state variable. We then obtain conditions that are necessary for an optimum.
These necessary conditions will be the one used regularly in maximization problems.
The shortcut
Simpli…cation
97
Note that the argument of the costate variables is always time (and not time t).
When we use this de…nition in (5:64), this …rst-order condition reads
[ t]
e _( )= Gy (:) Qy (:) :
e [ t]
_( )= _( ) e [ t]
( )= _( ) ( ):
which follows from computing the time derivative of the de…nition (5.66), we get
_ = Gy ( )Qy : (5.67)
which de…nes the link between the present value and the current value Hamiltonian
as
H c (t) = e [ t] H P (t)
Summary
Hence, instead of …rst-order condition (5.63) and (5.64), we get (5.65) and (5.67).
As just shown these …rst-order conditions are equivalent.
98
Exercises chapter 5
Applied Intertemporal Optimization
Hamiltonians
subject to
p (t) c (t) + A_ (t) = r (t) A (t) + w (t) :
by
2. Adjustment costs
Solve the adjustment cost example for
(I) = I:
What do optimality conditions mean? What is the optimal end-point value for
K and I?
3. Consumption over the life cycle
Utility of an individual, born at s and living for T periods is given at time t by
Z s+T
u (s; t) = e [ t] ln (c (s; )) d :
t
where
Z Z s+T
DR ( ) = exp r (u) du ; h (s; t) = Dr ( ) w (s; ) d :
t t
This deplorable individual would like to know how it can lead a happy life but,
unfortunately, has not studied optimal control theory!
99
(a) What would you recommend him? Use a Hamiltonian approach and dis-
tinguish between changes of consumption and the initial level. Which in-
formation do you need to determine the initial consumption level? What
information would you expect this individual to provide you with? In other
words, which of the above maximization problems makes sense? Why not
the other one?
(b) Assume all prices are constant. Draw the path of consumption in a (t,c(t))
diagram. Draw the path of asset holdings a(t) in the same diagram, by
guessing how you would expect it to look like. (You could compute it if
you want)
4. Optimal consumption
(a) Derive the optimal allocation of expenditure and consumption over time
for
c ( )1 1
u (c ( )) = ; >0
1
by employing the Hamiltonian.
(b) Show that this function includes the logarithmic utility function for =1
(apply L’Hospital’s rule).
(c) Does the utility function u(c(t)) make sense for > 1? Why (not)?
5. A central planner
(Rebelo 1990 AK-Model)
You are responsible for the future well-being of 360 Million Europeans and
centrally plan the EU by assigning a consumption path to each inhabitant.
Your problem consists in maximizing a social welfare function of the form
Z 1
i
U (t) = e [ t] C 1 1 (1 ) 1d
t
K_ = BK C (5.68)
(a) What are optimality conditions? What is the consumption growth rate?
(b) Under which conditions is the problem well de…ned (boundedness condi-
tion)? Insert the consumption growth rate and show under which condi-
tions the utility function is bounded.
(c) What is the growth rate of the capital stock? Compute the initial con-
sumption level, by using the No-Ponzi-Game condition (5.20).
(d) Under which conditions could you resign from your job without making
anyone less happy than before?
100
6. Optimal consumption levels
(a) Derive a rule for the optimal consumption level for a time-varying interest
rate r (t) : Show that (5.35) can be generalized to
1
c (t) = R 1 R (1 )r(s)
fW (t) + a (t)g ;
ds
t
e t d
(a) Use the result of 5 c) and check under which conditions investment is a
decreasing function of the interest rate.
(b) Perform the same analysis for a budget constraint a_ = ra + w c instead
of (5.68).
K_ = Y (K; L) C K:
a_ ( ) = r ( ) a ( ) + wL ( ) c( ):
Formulieren Sie das Maximierungsproblem der Haushalte und lösen Sie es.
101
(d) Aggregieren Sie über die Haushalte und beschreiben Sie dadurch die opti-
male Konsumentwicklung aggregierten Konsums.
(e) Formulieren Sie das Gleichgewicht auf dem Gütermarkt.
(f) Zeigen Sie die Konsistenz der Budgetrestriktion der Haushalte mit der
Kapitalakkumulationsgleichung der Ökonomie.
(g) Leiten Sie die reduzierte Form
K_ ( ) = Y ( ) C( ) K( )
@Y ( )
C_ ( ) @K( )
=
C( )
durch Befolgen dieser Schritte her und erläutern Sie diese verbal.
102
6 Dynamic programming in continuous time
This chapter reanalyzes maximization problems in continuous time that are known
from the chapter on Hamiltonians. It shows how to solve them with dynamic pro-
gramming methods. The sole objective of this chapter is to present the dynamic
programming method in a well-known deterministic setup such that its use in a sto-
chastic world in subsequent chapters becomes more accessible.
a_ = ra + w pc: (6.2)
As in models of discrete time, the value V (a (t)) of the optimal program is de…ned
by V (a (t)) maxfc( )g U (t) subject to (6.2). When households behave optimally
between today and in…nity by choosing the optimal consumption path fc ( )g ; their
overall utility U (t) is given by V (a (t)) :
The derivation of the Bellman equation is not as obvious as under discrete time
as under continuous time, even though we do have a today t, we do not have a
clear tommorrow (like a t + 1 in discrete time). We therefore need to construct a
tommorrow by adding a “small” time intervall t to t: Tommorrow would then be
t + t: Note that this derivation is heuristic and more rigorous approaches exist. See
e.g. Sennewald (2006) for further references to the literature.
Following Bellman’s idea, we rewrite the objective function as the sum of two
subperiods,
Z t+ t Z 1
[ t]
U (t) = e u (c ( )) d + e [ t] u (c ( )) d ;
t t+ t
where t is a “small” time intervall. When we approximate the …rst integral (think
of the area below the function u (c ( )) plotted over time ) by u (c (t)) t and the
103
discounting between t and t + t by 1+1 t and we assume that as of t + t we behave
optimally, we can rewrite the value function V (a (t)) = maxfc( )g U (t) as
1
V (a (t)) = max u (c (t)) t + V (a (t + t)) :
c(t) 1+ t
By doing so, we are left with only one choice variable c (t) instead of the entire path
fc ( )g :When we …rst multiply this expression by 1 + t, then divide by t and
V (a(t))
…nally move
n t
to the right hand side, weoget V (a (t)) =
V (a(t+ t)) V (a(t))
maxc(t) u (c (t)) [1 + t] + t
: Taking the limit lim t!0 gives the
Bellman equation,
dV (a (t))
V (a (t)) = max u (c (t)) + : (6.3)
c(t) dt
This equation again shows Bellman’s trick: A maximization problem, consisting
of the choice of a path of a choice variables, was broken down to a maximization
problem where only the level of the choice variable in t has to be chosen. The
structure of this equation can also be understood from a more intuitive perspective:
The term V (a (t)) can best be understood when comparing it to rK; capital income
at each instant of an individual who owns a capital stock K and the interest rate
is r: A household that behaves optimally “owns” the value V (a (t)) from optimal
behaviour and receives a utility stream of V (a (t)) : This “utility income” at each
instant is given by instantaneous utility from consumption plus the change in the
value of optimal behaviour. Note that this structure is identical to the capital-market
no-arbitrage condition (4.19), r (t) v (t) = (t) + v_ (t) - the capital income stream
from holding wealth v (t) on a bank account is identical to dividend payments (t)
plus the change v_ (t) in the market price when holding the same level of wealth in
stocks.
While the derivation just shown is the standard one, we will now present an
alternative approach which illustrates the economic content of the Bellman equation
and which is more straightforward. Given the objective function in (6.1), we can ask
how overall utility U (t) changes over time. To this end, we compute the derivative
dU (t) =dt and …nd (using the Leibniz rule 4.9 from ch. 4.3.1)
Z 1
d
U_ (t) = e [t t]
u (c (t)) + e [ t] u (c ( )) d = u (c (t)) + U (t) :
t dt
Overall utility U (t) reduces as time goes by by the amount u (c (t)) at each instant
(as the integral becomes “smaller” when current consumption in t is lost and we
start an instant after t) and increases by U (t) (as we gain because future utilities
come closer to today when today moves into the future). Rearranging this equation
gives U (t) = u (c (t)) + U_ (t) : When overall utility is replaced by the value function,
we obtain V (a (t)) = u (c (t)) + V_ (a (t)) which corresponds in its structure to the
Bellman equation (6.3).
104
We will now follow the three-step procedure to maximization when using the
dynamic programming approach as we got to know it in discrete time setups is section
3.4. When we compute the Bellman equation for our case, we obtain for the derivative
in (6.3) dV (a (t)) =dt = V 0 (a (t)) a_ which gives with the budget constraint (6.2)
The …rst-order condition (6.5) together with the maximized Bellman equation de-
termine the evolution of the control variable c (t) and V (a (t)). Again, however, the
maximized Bellman equation does not provide very much insight. Computing the
derivative with respect to a (t) however (and using the envelope theorem) gives an
expression for the shadow price of wealth that will be more useful,
(ii) In the second step, we compute the derivative of V 0 (a) with respect to time,
giving
dV 0 (a)
= V 00 (a) a_ = ( r) V 0 (a) ;
dt
where the last equality used (6.6). Dividing by V 0 (a) and using the usual notation
V_ 0 (a) dV 0 (a) =dt, this can be written as
V_ 0 (a)
= r: (6.7)
V (a)
This equation describes the evolution of the costate variable V 0 (a), the shadow price
of wealth.
105
The derivative of the …rst-order condition with respect to time is given by (apply
…rst logs)
u00 (c) p_ V_ 0 (a)
c
_ = + :
u0 (c) p V 0 (a)
Inserting (6.7) gives
V_ 0 (a)
= r;
V 0 (a)
as just shown in (6.7), we obtain the same equation as we had in the Hamiltonian
approach for the evolution of the costate variable, see e.g. (5.4) or (5.38). Comparing
…rst-order conditions (5.3) or (5.14) with (6.5), we see that they would be identical if
we had chosen exactly the same maximization problem. This is not surprising, given
our applied view of optimization: If there is one optimal path that maximizes some
objective function, this one path should result, independently of which maximization
procedure is chosen.
A comparison of optimality conditions is also useful for an alternative purpose,
however. As (6.7) and e.g. (5.4) or (5.38) are identical, we can conclude that the
derivative V 0 (a (t)) of the value function with respect to the state variable, in our case
a, is identical to the costate variable in the current-value Hamiltonian approach,
V 0 (a) = : This is where the interpretation for the costate variable in the Hamiltonian
approach in ch. 5.2.2 came from. There, we said that the costate variable stands
for the increase in the value of the optimal program when an additional unit of the
state variable becomes available; this is exactly what V 0 (a) stands for. Hence, the
interpretation of a costate variable is similar to the interpretation of the Lagrange
multiplier in static maximization problems.
106
Our agent has a standard objective function
Z 1
max U (t) = e [ t] u (c ( )) d :
t
It is maximized subject to two constraints. They describe the evolution of the state
variables wealth a and human capital h;
We do not give explicit expressions for the functions f (:) or g (:) but one can think of
a standard resource constraint for f (:) as in (6.2) and a functional form for g (:) that
captures a trade-o¤ between consumption and human capital accumulation: Human
capital accumulation is faster when a and h are large but decreases in c: To be precise,
we assume that both f (:) and g (:) increase in a and h but decrease in c:
dV (a; h)
V (a; h) = max u (c) + = max fu (c) + Va f (:) + Vh g (:)g :
dt
There are simply two partial derivatives of the value function after the u (c) term
times the da and dh term, respectively, instead of one as in (6.4), where there is only
one state variable.
Given that there is still only one control variable - consumption, there is only
one …rst-order condition. This is clearly speci…c to this example. One could think
of a time constraint for human capital accumulation (a trade-o¤ between leisure and
learning - think of the Lucas (1988) model) where agents choose the share of their
time used for accumulating human capital. In this case, there would be two …rst-order
conditions. Here, however, we have just the one for consumption, given by
@f (:) @g (:)
u0 (c) + Va + Vh =0 (6.9)
@c @c
When we compare this condition with the one-state-variable case in (6.5), we see
that the …rst two terms u0 (c) + Va @f@c(:) correspond exactly to (6.5): If we had speci…ed
f (:) as in the budget constraint (6.2), the …rst two terms would be identical to (6.5).
The third term Vh @g(:)
@c
is new and stems from the second state variable: Consumption
now not only a¤ects the accumulation of wealth but also the accumulation of human
capital. More consumption gives higher instantaneous utility but, at the same time,
decreases future wealth and - now new - also the future human capital stock.
107
As always, we need to understand the evolution of the costate variable(s). In
a setup with two state variables, there are two costate variables, or, economically
speaking, a shadow price of wealth a and a shadow price of human capital h: This
is obtained by partially di¤erentiating the maximized Bellman equation, …rst with
respect to a; then with respect to h: Doing this, we get (employing the envelope
theorem right away)
@f (:) @g (:)
Va = Vaa f (:) + Va + Vha g (:) + Vh ; (6.10)
@a @a
@f (:) @g (:)
Vh = Vah f (:) + Va + Vhh g (:) + Vh : (6.11)
@h @h
As in ch. 6.1, the second step of DP2 consists in computing the time derivatives
of the costate variables and in reinserting (6.10) and (6.11). We …rst compute time
derivatives, inserting (6.8) in the last step,
dVa (a; h)
= Vaa a_ + Vah h_ = Vaa f (:) + Vah g (:) ;
dt
dVh (a; h)
= Vha a_ + Vhh h_ = Vha f (:) + Vhh g (:) :
dt
Inserting then (6.10) and (6.11), we …nd
V_ a @f (:) Vh @g (:)
= ;
Va @a Va @a
V_ h Va @f (:) @g (:)
= ;
Vh Vh @h @h
allows to give an interpretation that links them to the standard one-state case in
(6.7). The costate variable Va evolves as above only that instead of the interest rate,
we …nd here @f@a(:) + VVha @g(:)
@a
: The …rst derivative @f (:) =@a captures the e¤ect a change
in a has on the …rst constraint; in fact, if f (:) represented a budget constraint as in
(6.2), this would be identical to the interest rate. The second term @g (:) =@a captures
the e¤ect of a change in a on the second constraint; by how much would h increase if
there was more a? This e¤ect is multiplied by Vh =Va ; the relative shadow price of h:
An analagous interpretation is possible for V_ h =Vh :
108
The …nal step consists in computing the derivative of the …rst-order condition (6.9)
with respect to time and replacing the time derivatives V_ a and V_ h by the expressions
from the previous step DP2. The principle of how to obtain a solution therefore
remains unchanged when having two instead of one state variable. Unfortunately,
however, it is generally not possible to eliminate the shadow prices from the resulting
equation which describes the evolution of the control variable. Some economically
suitable assumptions concerning f (:) or g (:) could, however, help.
Studying the behaviour of central banks allows to …ll many books. We present the
behaviour of the central bank in a very simple way - so simple that what the central
bank does here (buy bonds directly from the government) is actually illegal in most
OECD countries. Despite this simple presentation, the general result we obtain later
would hold in more realistic setups as well.
The central bank issues money M (t) in exchange for government bonds B (t) : It
receives interest payments iB from the government on the bonds. The balance of the
central bank is therefore iB + M_ = B.
_ This equation says that bond holdings by the
central bank increase by B_ either when the central bank issues money M_ or receive
interest payments iB on bonds it holds.
The government’s budget constraint reads G + iB = T + B: _ General government
expenditure G plus interest payments iB on government debt B is …nanced by tax
income T and de…cit B: _ We assume that only the central bank holds government
bonds and not private households. Combining the government with the central bank
budget therefore yields
M_ = G T:
109
This equation says that an increase in monetary supply either is used to …nance
government expenditure minus tax income or, if G = 0 for simplicity, any monetary
increase is given to housholds in the form of negative taxes, T = M_ :
6.4.2 Households
Preferences of households are described by a “money-in-utility”function,
Z 1
m( )
e [ t] ln c ( ) + ln d :
t p( )
In addition to utility from consumption c (t), utility is derived from holding a certain
stock m (t) of money, given a price level p (t) : Given that wealth of households consists
of capital goods plus money, a (t) = k (t) + m (t) ; and that holding money pays no
interest, the budget constraint of the household can be shown to read (see exercises)
a_ = i [a m] + w T =L pc:
Tax payments T =L per representative household are lump-sum. If taxes are negative,
T =L represents transfers of the government to households. The interest rate i is
de…ned according to
wK + v_
i :
v
When households choose consumption and the amount of money held optimally,
consumption growth follows (see exercises)
c_ p_
=i :
c p
6.4.3 Equilibrium
The reduced form
Equilibrium requires equality of supply and demand on the goods market. This is
obtained if total supply Y equals demand C + I: Letting capital accumulation follow
K_ = I K; we get
K_ = Y (K; L) C K: (6.13)
This equation determines K: As capital and consumption goods are traded on the
same market, this equation implies v = p and the nominal interest rate becomes with
(6.12)
wK p_ @Y p_
i= + = + : (6.14)
p p @K p
110
The nominal interest rate is given by marginal productivity of capital wK =p = @Y =@K
(the “real interest rate”) plus in‡ation p=p.
_
Aggregating over households yields aggregate consumption growth of C=C _ =i
p=p
_ : Inserting (6.14) yields
C_ @Y
= : (6.15)
C @K
Aggregate money demand is given by
pC
M= : (6.16)
i
Given an exogenous money supply rule and appropriate boundary conditions, these
four equations determine the paths of K; C; i and the price level p:
One standard property of models with ‡exibel prices is the dichotomy between
real variables and nominal variables. The evolution of consumption and capital - the
real side of the economy - is completely independent of monetary in‡uences: Equation
(6.13) and (6.15) determine the paths of K and C just as in the standard optimal
growth model without money - see (5.37) and (5.39) in ch. 5.5.4. Hence, when
thinking about equilibrium in this economy, we can think about the real side on the
one hand - independently of monetary issues - and about the nominal side on the
other hand. Monetary variables have no real e¤ect but real variables have an e¤ect
on monetary variables like e.g. in‡ation.
Needless to say that the real world does not have perfectly ‡exible prices such
that one should expect monetary variables to have an impact on the real economy.
This model is therefore a starting point to well understand structures and not a fully
developed model to analyse monetary questions in a very realistic way. Price rigidity
would have to be included before doing this.
A stationary equilibrium
111
N
p_
p
pC @Y
M @K
! !
N
0
p p
@Y
@K
Now assume there is money growth, M_ > 0: The …gure then shows that the
price level p increases as the slope of the line becomes ‡atter. Money growth in an
economy with constant GDP implies in‡ation. By looking at (6.14) and (6.16) again
and by focusing on equilibria with constant in‡ation rates, we know from (6.14)
that a constant in‡ation rate implies a constant nominal interest rate. Hence, by
di¤erentiating the equilibrium (6.16) on the money market, we get
p_ M_
= : (6.17)
p M
In an economy with constant GDP and increasing money supply, the in‡ation rate is
identical to the growth rate of money supply.
A growth equilibrium
p_ M_ C_
= :
p M C
In‡ation is given by the di¤erence between the growth rate of money supply and
consumption growth.
112
The current thinking about central bank behaviour di¤ers from the view that the
central bank chooses money supply M as assumed so far. The central bank rather
sets nominal interest rates and money supply adjusts (where one should keep in mind
that money supply is more than just cash used for exchange as modeled here). Can
nominal interest rate setting be analysed in this setup?
Equilibrium is described by equations (6.13) to (6.16). They were understood to
determine the paths of K; C; i and the price level p; given a money supply choice M
of the central bank. If one believes that nominal interest rate setting is more realistic,
these four equations would simply determine the paths of K; C; M and the price level
p; given a nominal interest rate choice i of the central bank. Hence, by simply making
an exogenous variable, M; endogenous and making a previously endogenous variable,
i; exogenous, the same model can be used to understand the e¤ects of higher and
lower nominal interest rates on the economy.
Due to perfect price ‡exibility, real quantities remain una¤ected by the nominal
interest rate. Consumption, investment, GDP, the real interest rate, real wages are all
determined, as before, by (6.13) and (6.15) - the dichotomy between real and nominal
quantities continues given price ‡exibility. By (6.14), a change in the nominal interest
rate a¤ects in‡ation: high (nominal) interest rates imply high in‡ation, low nominal
interest rates imply low in‡ation. From the money market equilibrium in (6.16) one
can then conclude what this implies for money supply, again both for a growing or
a stationary economy. Much more needs to be said about these issues before policy
implications can be discussed. Any analysis in a general equilibrium framework would
however partly be driven by relationships presented here.
113
Exercises chapter 6
Applied Intertemporal Optimization
Dynamic Programming in Continuous
Time
(a) Solve this maximization problem by using the dynamic programming ap-
proach. You may choose appropriate (numerical or other) values for para-
meters where this simpli…es the solution (and does not destroy the spirit
of this exercise).
(b) Show that in the long-run with adjustment cost and at each point in time
under the absence of adjustment cost, capital is paid its value marginal
product. Why is labour being paid its value marginal product at each
point in time?
114
m ( ) =p ( ) in the utility function where m ( ) is the amount of cash and p ( )
is the price level of the economy. Let the budget constraint of the individual be
a_ = i [a m] + w T =L pc:
115
6.6 Looking back
This is the end of part I and II. This is often also the end of a course. This is a good
moment to look back at what has been acomplished. After 14 or 15 lectures and the
same number of exercise classes, the amount of material covered is fairly impressive.
In terms of maximization tools, this …rst part has covered
Solving by inserting
Hamiltonian
the optimal saving central planner model in discrete and continuous time
Most importantly, however, the tools having been presented here allow students to
“become independent”. A very large part of the economics literature (acknowledging
that game theoretic approaches have not at all been covered here) is now open and
accessible and the basis for understanding a paper in detail (and not just the overall
argument) and for presenting own arguments in a scienti…c language are laid.
Clearly, models with uncertainty present additional challenges. They will be pre-
sented and overcome in part III and part IV.
116
Part III
Stochastic models in discrete time
The following chapter starts part III of this book. We got to know many optimization
methods for deterministic setups, both in discrete and continuous time. All economic
questions that were analysed were viewed as “su¢ ciently deterministic”. If there was
any uncertainty in the setup of the problem, we simply ignored it or argued that it
is of no importance for understanding the basic properties and relationships of the
economic question. This is a good approach for many economic questions.
Real life has few certain components - death is certain, but when? Taxes are
certain, but how high are they? We know that we all exist - but don’t ask philosophers.
Part III and part IV will take uncertainty in life seriously and incorporate it explicitly
in the analysis of economic problems. We follow the same distinction as in part I and
II; we …rst analyse the e¤ects of uncertainty on economic behaviour in discrete time
setups and then move to continuous time setups in part IV.
F (x) = Prob (X x) :
117
probability function (Wahrscheinlichkeitsfunktion) or probability mass function. The
probability that X has the realization of x is given by f (x) :
When the RV X is continuous, the …rst derivative of the distribution function F
dF (x)
f (x) =
dx
is called the probability density function f (Dichtefunktion). The probability that the
Rb
realization of X lies between, say, a and b > a is given by F (b) F (a) = a f (x) dx:
Hence the probability that X equals a is zero.
7.1.2 An illustration
Discrete random variable
Consider the probabilistic experiment ’tossing a coin two times’. The possibility
space is given by fHH; HT; T H; T T g. De…ne the RV ’Number of hats’. The range
of this variable is given by f0; 1; 2g : Assuming that the coin falls on either side with
the same probability, the probability function of this RV is given by
8 9 8
< :25 = < 0
f (x) = :5 for x = 1 :
: ; :
:25 2
Think of next weekend. You might consider going into some pub to meet friends.
Before you go there, you do not know how much time you will spend there. If you
meet many friends, you stay longer, if you just drink one beer, you leave soon. Hence,
going to a pub on a weekend is a probabilistic experiment with a chance outcome.
The set of possible outcomes with respect to the amount of time spent in a pub
is the possibility space. Our random variable T maps from this possibility space
into a set of numbers with a range from 0 to, let’s say, 4 hours (as the pub closes
at 1 am and you never go there before 9 pm). As time is continous, T 2 [0; 4] is a
continuous random variable. The distribution function F (t) gives you the probability
that you spend a period of length t or shorter in the pub.R 2 The probability that you
spend between 1.5 and two hours in the pub is given by 1:5 f (t) dt where f (t) is the
density function f (t) = dF (t) =dt.
118
7.2.1 Discrete random variables
Discrete uniform distribution
range x 2 fa; a + 1; :::; b 1; bg
probability function f (x) = 1= (b a + 1)
An example for this RV is the die. Its range is 1 to 6, the probability for any
number (at least for fair dice) is 1=6.
The Poisson distribution
range x 2 f0; 1; 2; :::g
x
e
probability function f (x) = x!
119
The mean and standard deviation of the RVs are denoted by i and i : The parameter
is called the correlation coe¢ cient between X1 and X2 is de…ned as the covariance
12 devided by the standard deviations (see ch. 7.3.1).
The nice aspects of this two-dimensional normally distributed RV (the same holds
for n-dimensional RVs) is that the density function of each individual RV Xi ; i.e.
the marginal density (Randverteilung), is given by the standard expression which is
independent of the correlation coe¢ cient,
1 1 2
x
~
f (xi ) = p 2
e 2 i : (7.3)
2 i
7.3.2 Lemmas
Basic
120
On expectations
E [a + bX] = a + bEX
RR
z = g (x; y) ) EZ = g (x; y) f (x; y) dx dy
On variances
On covariances
Advanced
We present here a theorem which is very intuitive and highly useful to analytically
study pro- and countercyclical behaviour of endogenous variables in models of business
cycles. The theorem says that if two variables depend “in the same sense” on some
RV (i.e. they both increase or decrease in the RV), then these two variables have a
positive covariance. If e.g. both GDP and R&D expenditure increase in TFP and
TFP is random, then GDP and R&D expenditure are procyclical.
Theorem 6 Let X a random variable and f (X) and g (X) two functions such that
f 0 (X) g 0 (X) R 0 8 x 2 X: Then
121
7.3.3 Functions on random variables
We will occasionally encounter the situation where we compute density functions of
functions of RVs. Here are some examples.
An exponential transformation
Consider now a general transformation of the type y = y (x) where the RV X has
some density f (x) : What is the density of Y ? The answer comes from the following
Theorem 7 Let X be a random variable with density f (x) and range [a; b] which can
be ] 1; +1[: Let Y be de…ned by the monotonically increasing function y = y (x) :
Then the density g (y) is given by g (y) = f (x (y)) dx
dy
on the range [y(a); y(b)]:
122
This theorem can be easily proven as follows. The proof is illustrated in …g. 31.
The …gure plots the RV X on the horizontal and the RV Y on the vertical axis. A
monotonically increasing function y (x) represents the transformation of realizations
x into y:
Y N
Y (b) Y (x)
Y (e
x)
Y (a)
N
a e
x b X
Figure 31 Transforming a random variable
Proof. The transformation of the range is immediately clear from the …gure:
When X is bounded between a and b; Y must be bounded between y (a) and y (b) :
The proof for the density of Y requires some more steps: The probability that y is
smaller than some y (~ x) is identical to the probability that X is smaller than this
x~: This follows from the monotonicity of the function Y (X) : As a consequence, the
distribution function (cumulative density function) of Y is given by G (y) = F (x)
where y = y (x) or, equivalently, x = x (y) : The derivative then gives the density
d d
function, g (y) dy G (y) = dy F (x (y)) = f (x (y)) dx
dy
:
where the stochastic components are distributed according to some distribution func-
tion on some range which implies a mean and a variance 2 ; "t ~ ( ; 2 ) : We do not
make any speci…c assumption about "t at this point. Note that the stochastic com-
ponents "t are iid (identically and independently distributed) which implies that the
covariance between any two distinct "t is zero, cov("t ; "s ) = 0 8t 6= s. An alternative
representation of xt would be xt = axt 1 + + vt with vt ~ (0; 2 ).
123
Solving by inserting
and eventually
The mean
A stochastic di¤erence equation does not predict how the stochastic variable xt
for t > 0 actually evolves, it only predicts the distribution of xt for all future points
in time. The realization of xt is random. Hence, we can only hope to understand
something about the distribution of xt . To do so, we start by analysing the mean of
xt for future points in time.
Denote the expected value of xt by x~t ;
x~t E0 xt ; (7.9)
i.e. x~t is the value expected in t = 0 for xt in t > 0: The expectations operator E0
captures our “knowledge” when we compute the expectation for xt . The “0” says
that we have all information for t = 0 and know therefore x0 ; but we do not know "1 ;
"2 ; etc.
By applying the expectations operator E0 to (7.7), we obtain
x~t = a~
xt 1 + :
This is a deterministic di¤erence equation which describes the evolution of the ex-
pected value of xt over time. There is again a standard solution to this equation
which reads
1 at
x~t = at x0 + t
s=1 a t s
= a t
x 0 + ; (7.10)
1 a
where we used ts=1 at s = a0 + a1 + ::: + at 1 = ti=01 ai and, from ch. 2.5.1, ni=0 ai =
(1 an+1 ) = (1 a), for a < 1: This equation shows that the expected value of xt
changes over time as t changes. Note that x~t might increase or decrease (see the
exercises).
The variance
124
Let us now look at the variance of xt : We obtain an expression for the variance
by starting from (7.8) and observing that the terms in (7.8) are all independent from
each other: at x0 is a constant and the disturbances "s are iid by assumption. The
variance of xt is therefore given by the sum of variances (compare (7.5) for the general
case),
t s 2 a2t
t s t 1 i 21
V ar (xt ) = 0 + s=1 at V ar ("s ) = 2 t
s=1 a2 : = 2
i=0 a2 =
1 a2
(7.11)
We see that it is also a function of t: The fact that the variance becomes larger
the higher t is intuitively clear. The further we look into the future, the “more
randomness”there is: equation (7.8) shows that a higher t means that more random
variables are added up.
Long-run behaviour
When we considered deterministic di¤erence equations like (2.36), we found the
long-run behaviour of xt by computing the solution of this di¤erence equation and
by letting t approach in…nity. This would give us the …xpoint of this di¤erence
equation as e.g. in (2.37). The concept that corresponds to a …xpoint in an uncertain
environment, e.g. when looking at stochastic di¤erence equations like (7.7), is the
limiting distribution. As stated earlier, stochastic di¤erence equations do not tell
us anything about the evolution of xt itself, they only tell us something about the
distribution of xt : It would therefore make no sense to ask what xt is in the long run
- it will always remain random. It makes a lot of sense, however, to ask what the
distribution of xt is for the long run, i.e. for t ! 1:
All results so far were obtained without a speci…c distributional assumption for "t
apart from specifying a mean and a variance. Understanding the long-run distribution
of xt in (7.7) is easy if we assume that the stochastic component "t is normally
distributed for each t; "t ~N ( ; 2 ). In this case, starting in t = 0; the variable x1 is
from (7.7) normally distributed as well. As a weighted sum of two random variables
that are (jointly) normally distributed gives again a random variable with a normal
distribution (where the mean and variance can be computed as in ch. 7.3.2), we
know from (7.7) also that x1 ; x2 ... are all normally distributed. Hence, in order to
understand the evolution of the distribution of xt , we only have to …nd out how the
expected value and the variance of the variable xt evolves.
Neglecting the cases of a 1, the mean x~ of our long-run normal distribution is
given from (7.10) by the …xpoint
1
lim x~t x~1 = ; a < 1:
t!1 1 a
The variance of the long-run distribution is from (7.11)
2 1
lim var (xt ) = :
t!1 1 a2
Hence, the long-run distribution of xt is a normal distribution with mean = (1 a)
and variance 2 = (1 a2 ).
125
The evolution of the distribution of xt
Given our results on the evolution of the mean and the variance of xt and the fact
that we assumed "t to be normally distributed, we know that xt is normally distributed
at each point in time. We can therefore draw the evolution of the distribution of x
as in the following …gure.
0.3
0.25
0.2
limiting distribution
0.15
0.1
0 1 0.05
2 3
4
5
6
7 0
8
9
10
19
11
17
12
15
13
13
14
11
15
16 9 time
7
17
5
xt 18
3
19
1
Remember that we were able to plot a distribution for each t only because of
properties of the normal distribution. If we had assumed that "t is lognormally or
equally distributed, we would not have been able to easily say something about the
distribution of xt : The means and variances in (7.10) and (7.11) would still have been
valid but the distribution of xt for future t is generally unknown for distributions of
"t other than the normal distribution.
Given information in period t = 0, it predicts the density function f (xit ) for wealth
in all future periods, given some initial value xi0 . Shocks are identically and indepen-
dently distributed (iid) over time and across individuals.
Assuming a large number of agents i, more precisely a continuum of agents with
mass n, a law of large numbers can be constructed along the lines of Judd (1985):
126
Let all agents start with the same wealth xi0 = x0 in t = 0. Then the density f (xit )
for wealth of any individual i in t equals the realized wealth distribution in t for
the continuum of agents with mass n: Put di¤erently: when the probability for an
individual to hold wealth between some lower and upper bound in t is given by p, the
share of individuals that hold wealth between these two bounds in t is also given by
p. Total wealth in t is then deterministic (as are all other moments) and is given by
Z 1
xt = n f (xit )xit dxit = n t ; (7.12)
1
127
Assuming we are interested in the expected value, we would obtain
Ex1 = 1k=0 1k E k "1 k
Exk = 1
0 E 0 1
" Ex0 + 1
1 E 1 0
" Ex1
= E" + E Ex:
E"
Solving for Ex yields Ex = 1 E
= 1 a
:
7.5.1 Technology
Let there be an aggregate technology
Yt = At Kt L1t (7.13)
The total factor productivity level At is uncertain. We assume that At is identically
and independently distributed (iid) in each period. The random variable At is pos-
itive, has a mean A and a variance 2 : No further information is needed about its
distribution function at this point,
2
At ~ A; ; At > 0: (7.14)
Assuming that total factor productivity is iid means, inter alia, that there is no
technological progress. One can imagine many di¤erent distributions for At . In
principle, all distributions presented in the last section are viable candidates. Hence,
we can work with discrete distributions or continuous distributions.
7.5.2 Timing
The sequence of events is as follows. At the beginning of period t, the capital stock
Kt is inherited from the last period, given decisions of the last period. Then, total
factor productivity is revealed. With this knowledge, …rms choose factor employment
and households choose consumption (and thereby savings).
ct
At wt
Kt rt
128
Figure 33 Timing of events
Hence, only “at the end of the day”one really knows how much one has produced.
This implies that wages and interest payments are known with certainty at the end
of the period only as well.
The state of the economy in t is completely described by Kt and the realization
of At : All variables of the model are contingent on the state.
7.5.3 Firms
As a consequence of this timing of events, …rms do not bear any risk and they pay
the marginal product of labour and capital to workers and capital owners at the end
of the period,
@Yt
wt = pt ; (7.15)
@Lt
1
@Yt Lt
rt = p t = p t At : (7.16)
@Kt Kt
All risk is therefore born by households through labour and capital income.
We consider an agent that lives for two periods and works and consumes in a
world as just described. Agents consume in both periods and choose consumption
such that they maximize expected utility. In all generality concerning the uncertainty,
she maximizes
max Et fu (ct ) + u (ct+1 )g ; (7.17)
where is the subjective discount factor that measures the agent’s impatience to
consume. The expectations operator has an index t to indicate, similar to E0 in
(7.9), that expectations are based on the knowledge about random variables which
is available in period t. We will see in an instant whether the expectations operator
Et is put at a meaningful place in this objective function. Placing it in front of both
instantaneous consumption from ct and from ct+1 is the most general way of handling
it. We will also have to specify later what the control variable of the household is.
The agent has to choose consumption for the …rst and the second period. When
consumption is chosen and given wage income wt ; savings adjust such that the budget
constraint
wt = ct + st (7.18)
for the …rst period holds. Note that this constraint always holds, despite the uncer-
tainty about the wage. It holds in realizations, not in expected terms. In the second
period, the household receives interest payments on savings made in the …rst period
and uses savings plus interests for consumption,
129
(1 + rt+1 ) st = ct+1 : (7.19)
One way of solving this problem (for an alternative, see the exercise) is to insert
consumption levels from these two constraints into the objective function (7.17). This
gives
max Et fu (wt st ) + u ((1 + rt+1 ) st )g :
st
This nicely shows that the household in‡uences consumption in both periods by
choosing savings st in the …rst period. In fact, the only control variable the household
can choose is st :
Let us now take into consideration, as drawn in the above …gure, that consumption
takes place at the end of the period after revelation of productivity At in that period.
Hence, the consumption level in the …rst period is determined by savings only and
thereby certain.7 The consumption level in the second period is uncertain, however, as
the next period interest rate rt+1 depends on the realization of At+1 which is unknown
in t when decisions about savings st are made. The objective function can therefore
be written as
max u (wt st ) + Et u ((1 + rt+1 ) st ) :
st
Let us now, for illustration purposes, assume a discrete random variable At with
a …nite number n of possible realizations. Then, this maximization problem can be
written as
n
max u (wt st ) + i=1 i u ((1 + ri;t+1 ) st )
st
where i is the probability that the interest rate ri;t+1 is in state i in t + 1: This is
the same probability as the probability that the underlying source of uncertainty At
is in state i: The …rst-order condition then reads
u0 (wt st ) = n
i=1 i u
0
((1 + ri;t+1 ) st ) (1 + ri;t+1 ) = Et u0 ((1 + rt+1 ) st ) (1 + rt+1 )
(7.20)
Marginal utility of consumption today must equal expected marginal utility of con-
sumption tomorrow corrected by interest and time preference rate. Optimal behav-
iour in an uncertain world therefore means ex ante optimal behaviour, i.e. before
random events are revealed. Ex post, i.e. after resolution of uncertainty, behaviour
is suboptimal as compared to the case where the realization is known in advance:
Marginal utility in t will (with high probability) not equal marginal utility (corrected
by interest and time preference rate) in t + 1. This re‡ects a simple fact of life: “If
I had known before what would happen, I would have behaved di¤erently.”Ex ante,
behaviour is optimal, ex post, probably not. Clearly, if there was only one realization
for At , i.e. 1 = 1 and i = 0 8 i > 1; we would have the deterministic …rst-order
condition we had in exercise 1 in ch. 2.
The …rst-order condition also shows that closed form solutions are possible if
marginal utility is of a multiplicative type. As savings st are known, the only quantity
7
If consumption ct (or savings) was chosen before revelation of total factor productivity, house-
holds would want to consume a di¤erent level of consumption ct after At is known.
130
which is uncertain is the interest rate rt+1 : If the instantaneous utility function u (:)
allows to separate the interest rate from savings, i.e. the argument (1 + ri;t+1 ) st in
u0 ((1 + ri;t+1 ) st ) in (7.20), an explicit expression for st and thereby consumption can
be computed. This will be shown in the example following right now and in exercise
5.
Et f ln ct + (1 ) ln ct+1 g : (7.21)
131
identity between number of workers and number of young, Lt = N . Note that this
expression would hold identically in a deterministic model. With (7.16), consumption
of old individuals amounts to
The capital stock in period t + 1 is given by savings of the young. One could show
this as we did in the deterministic case in ch. 2.4.2. In fact, adding uncertainty would
change nothing to the fundamental relationships. We can therefore directly write
Kt+1 = N st = N (1 )wt = (1 ) (1 ) Yt ;
where we used the expression for savings for the Cobb-Douglas utility case from (7.24).
Again, we succeeded in reducing the presentation of the model to one equation in one
unknown. This allows us to illustrate the economy by using the simplest possible
phase diagram.
N
Kt+1 K2 = (1 )(1 )A1 K1 L1
K3 = (1 )(1 )A2 K2 L1
K1 = (1 )(1 )A0 K0 L1
K0 K1 K 2 K3 K(t)
This phase diagram looks in principle identical to the deterministic case. There
is of course a fundamental di¤erence. The Kt+1 loci are at a di¤erent point in each
period. In t = 0, K0 and A0 are known. Hence, by looking at the K1 loci, we can
compute K1 as in a deterministic setup. Then, however, TFP changes and, in the
example plotted above, A1 is larger than A0 : Once A1 is revealed in t = 1; the new
capital stock for period 2 can be graphically derived. With A2 revealed in t = 2; K3
can be derived an so on. It is clear that this economy will never end up in one steady
state as in the deterministic case as At is di¤erent for each t: As illustrated before,
however, the economy can converge to a stationary stable distribution for Kt :
132
In order to better understand the evolution of this economy, let us now look at
some analytical results. The logarithm of the capital stock evolves according to
ln Kt+1 = ln ((1 ) (1 )) + ln At + ln Kt + (1 ) ln Lt :
Assuming a constant population size Lt = L, we can rewrite this equation as
2
t+1 = m0 + t + t; t N( ; ); (7.25)
where we used
t ln Kt (7.26)
m0 ln [(1 ) (1 )] + (1 ) ln L (7.27)
and t ln At captures the uncertainty stemming from the TFP level At : As TFP is
iid, so is its logarithm t . Since we assumed the TFP to be lognormally distributed,
its logarithm t is normally distributed. When we remove the mean from the random
variable by replacing according to t = "t + ; where "t N (0; 2 ); we obtain
t+1 = m0 + + t + "t :
The expected value and the variance
We can now compute the expected value of t by following the same steps as in
ch. 7.4.1, noting that the only additional parameter is m0 . Starting in t = 0 with
some initial value 0 ; and solving this equation recursively gives
t
t 1 t 1 t 1 s
t = 0 + (m0 + ) + s=0 "s : (7.28)
1
Computing the expected value from the perspective of t = 0 gives
t
t 1
E0 t = 0 + (m0 + ) ; (7.29)
1
where we used E0 "s = 0 for all s > 0 (and we set "0 = 0 or assumed that expectations
in 0 are formed before the realization of "0 is known). For t going to in…nity, we
obtain
m0 +
lim E0 t = : (7.30)
t!1 1
Note that the solution in (7.28) is neither di¤erence- nor trend stationary. Only in
the limit, we have a (pure) random walk. For a very large t; t in (7.28) is zero and
(7.28) implies
t t 1 = "t 1 :
The variance of t can be computed with (7.28), where again independence of all
terms is used as in (7.11),
2t
t 1 t 1 s t 1 2(t 1 s) 2 1 2
V ar ( t ) = 0 + 0 + V ar s=0 "s = s=0 = 2
(7.31)
1
In the limit, we obtain
2
lim V ar ( t ) = 2
: (7.32)
t!1 1
133
Relation to fundamental uncertainty
For a more convincing economic interpretation of the mean and the variance, it
is useful to express some of the equations not as a function of properties of the log
of At ; i.e. properties of t ; but directly of the level of At . As (7.6) implies, the mean
and variances of these two variables are related in the following fashion
!
2
2 A
= ln 1 + (7.33)
A
!
2
1 2 (7:33) 1 A
= ln A = ln A ln 1 + (7.34)
2 2 A
We can insert these expressions into (7.29) and (7.31) and study the evolution over
time or directly focus on the long-run distribution of t ln Kt . Doing so by inserting
in (7.30) and (7.32) gives
2 2
1
m0 + ln A 2
ln 1 + A
ln 1 + A
A A
lim E0 t = ; lim V ar ( t ) = 2
:
t!1 1 t!1 1
These equations tell us that uncertainty in TFP as captured by A not only a¤ects
the spread of the long-run distribution of the capital stock but also its mean. More
uncertainty leads to a lower long-run mean of the capital stock. This is interesting
given the standard result of precautionary saving and the portfolio e¤ect (in setups
with more than one asset). (t.b.c.)
Households
8
If individuals were identical, i.e. if they had identical preferences and experienced the same
income streams, no loans would be given in equilibrium.
134
The budget constraints (7.18) and (7.19) of all individuals i now read
U i = u wt sit + Eu 1 + i rt+1 + 1 i
r sit ! maxi (7.37)
sit ;
by now choosing two control variables: The amount of resources not used in the …rst
period for consumption, i.e. savings sit ; and the share i of savings held in the risky
asset.
First-order conditions for sit and i , respectively, are
135
by u0 cit+1 rt+1 : The share of wealth held in the risky asset then depends on the
increase in utility from realizations of rt+1 accross various states and the expecta-
tions operator computes a weighted sum of theses utility increases: Et u0 cit+1 rt+1
n 0 i
j=1 u cjt+1 rjt+1 pj ; where j is the state, rjt+1 the interest rate in this state and pj
the probability for state j to occur. An agent is then indi¤erent between holding two
assets when expected utility-weighted returns are identical.
Ert+1 = r: (7.42)
The …rst-order condition on how to invest implies together with (7.41) that the en-
dogenous interest rate for loans is pinned down by the time preference rate,
1= [1 + r] , r = (7.43)
where we used (7.36) for the second equality and the fact that st as control variable
is deterministic for the third. Hence, as in the last section, we can derive explicit
expressions. Use (7.44) and (7.35) and …nd
1
sat = cat , wt cat = cat , cat = wt :
1+
a a
sat = wt ; cat+1 = [1 + rt+1 + (1 ) r] wt :
1+ 1+
This is our closed-form solution for risk-averse individuals in our heterogeneous-agent
economy.
Let us now look at the investment problem of risk-averse households. The deriva-
tive of their objective function is given by the LHS of the …rst-order condition (7.39).
136
Expressed for logarithmic utility function, and inserting the optimal consumption
result yields
1 1+ rt+1 r
E [rt+1 r] = E a a
cat+1 1 + rt+1 + (1 )r
1+ rt+1 r 1+ X
= E a E : (7.45)
1 + (rt+1 r) + r a+ X
The last step de…ned X rt+1 r as a RV and a as a constant.
It can now easily be shown that (7.45) implies that risk-averse individuals will
not allocate any of their savings to the risky asset, i.e. a 0: The derivative of the
expression EX= (a + a X) with respect to a is negative
d X X2
aE =E < 0 8 a:
d a + aX (a + a X)2
The sign of this derivative can also easily be seen from (7.45) as an increase in a
implies a larger denominator. Hence, when plotted, the …rst-order condition is down-
ward sloping in a : By guessing we …nd that, with (7.42), a = 0 satis…es the …rst-order
condition for investment,
a X
=0)E = EX = 0:
a + aX
Hence, the …rst-order condition is zero for a = 0: As the …rst-order condition is
monotonously decreasing, a = 0 is the only value for which it is zero,
X a
E =0, = 0:
a + aX
This is illustrated in the following …gure.
N
E a+x a x
N
a
0
137
a
Figure 35 The …rst-order condition for the share of savings held in the risky asset
The …gure shows that expected utility of a risk-averse individual is the higher,
the lower the share of savings held in the risky asset. If we exclude short-selling,
risk-averse individuals will allocate all of their savings to loans, i.e. a = 0. They
give loans to risk-neutral individuals who in turn pay a certain interest rate equal to
the expected interest rate. All risk is born by risk-neutral individuals.
The individual
E fu (c0 ) + u (c1 )g :
This individual can invest in the risky and in a riskless investment form. It earns
labour income w in the …rst period and does not have any other source of income.
This income w is used for consumption and savings. Savings are allocated to several
risky assets j and a riskless bond. The …rst period budget constraint therefore reads
n
w = c0 + j=1 mj pj0 +b
where the number of shares j are denoted by mj ; their price by pj0 and the amount
of wealth held in the riskless asset is b. The individual consumes all income in the
second period which implies the budget constraint
n
c1 = j=1 mj pj1 + (1 + r) b;
138
Solution and asset pricing
This is the standard …rst-order condition for consumption. The …rst-order condition
for the number of assets to buy reads for any asset j
E [u0 (c1 ) (1 + r) ( pj0 ) + pj1 ] = 0 , E [u0 (c1 ) pj1 ] = (1 + r) pj0 Eu0 (c1 ) :
The last step used the fact that the interest rate and pj0 are known in 0 and can there-
fore be pulled out of the expectations operatior on the right-hand side. Reformulating
this condition gives
1 u0 (c1 )
pj0 = E 0 pj1 (7.46)
1 + r Eu (c1 )
which is an expression that can be given an interpretation as a pricing equation. The
price of an asset j in period zero is given by the discounted expected price in period
one. The price in period one is weighted by marginal utilities.
139
7.7.3 Risk-neutral valuation
There is a strand in the …nance literature, see e.g. Brennan (1979), that asks under
what conditions a risk neutral valuation of contingent claims holds (risk neutral val-
uation relationship, RNVR) even when households are risk averse. This is identical
to asking under which conditions marginal utilities in (7.46) do not show up. We will
now brie‡y illustrate this approach and drop the index j.
Assume the distribution of the price p1 can be characterized by a density f (p1 ; ) ;
where Ep1 . Then, if a risk neutral valuation relationship exists, the price of the
contingent claim in zero is given by
Z
1
v (p0 ) = g (p1 ) f (p1 ; ) dp1 ; with = (1 + r) p0 :
1+r
This is (7.48) with the expectations operator being replaced by the integral over
realizations g (p1 ) times the density f (p1 ). With risk averse households, the pricing
relationship would read under this distribution
Z
1 u0 (c1 )
v (p0 ) = g (p1 ) f (p1 ; ) dp1 :
1+r Eu0 (c1 )
This is (7.47) expressed without the expectations operator. The expected value
is left unspeci…ed here as it is not a priori clear whether this expected value equals
(1 + r) p0 also under risk aversion. It is then easy to see that a RNVR holds if
u0 (c1 ) =Eu0 (c1 ) = f (p1 ) =f (p1 ; ) : Similar conditions are derived in that paper for
other distributions and for longer time horizons.
140
factor productivity) does not grow smoothly over time as in most models of exogenous
or endogenous long-run growth but that productivity follows a step function.
N
log of
productivity
x
smooth productivity growth
x h
x h
x h
x h productivity step function
x h
h
N
time
Figure 36 Smooth productivity growth in balanced growth models (dashed line) and
step-wise productivity growth in models of natural volatility
With time on the horizontal and the log of productivity on the vertical axis, this
…gure shows as the dashed line a smooth productivity growth path. This is the smooth
growth path that induces balanced growth. In models of natural volatility, the growth
path of productivity has periods of no change at all and points in time of discrete
jumps. After a discrete jump, returns to investment go up and an upward jump in
growth rates results. Growth rates gradually fall over time as long as productivity
remains constant. With the next jump, growth rates jump up again. While this step
function implies long-run growth as productivity on average grows over time, it also
implies short-run ‡uctuations.
The precise economic reasons given for this step function - which is not simply
imposed but always follows from some deeper mechanisms - di¤er from one approach
to the other. A crucial implication of this step-function is the implicit belief that
economically relevant technological jumps take place once every 4-5 years. Each
cycle of an economy and thereby also long-run growth go back to relatively rare
events. Fluctuations in time series that are of higher frequency than these 4-5 years
either go back to exogenous shocks, to measurement error or other disturbances of
the economy.
The step function sometimes captures jumps in total factor productivity, some-
times only in labour productivity of a most recent vintage of a technology. This di¤er-
ence is important for the economic plausibility of the models. Clearly, one should not
build a theory on large aggregate shocks to TFP as those are not easily observable.
Some papers in the literature show indeed how small changes in technology can have
large e¤ects. See Francois and Lloyd-Ellis (2003) or Wälde (2002, 2005). See also
Matsuyama (1999).
This section presents the simples stochastic natural volatility model which allows
to show most easily the di¤erence to exogenous shock models of the business cycle.
141
7.8.2 A simple stochastic model
This section presents the simplest possible model that allows to understand the dif-
ference between the stochastic natural volatility and the RBC approach.
Technologies
Let the technology be described by a Cobb-Douglas speci…cation,
Yt = At Kt Lt1 ;
where At represents total factor productivity, Kt is the capital stock and Lt are hours
worked. Capital can be accumulated according to
Kt+1 = (1 ) Kt + It
The technological level follows
At+1 = (1 + qt ) At
where
0 pt
qt = with probability : (7.49)
1 1 pt
The probability depends on resources Rt invested into R&D,
pt = p (Rt ) :
Clearly, the function p (Rt ) in this discrete time setup must be such that 0 p (Rt )
1:
The speci…cation of technological progress in (7.49) is probably best suited to point
out the di¤erences to RBC type approaches: The probability that a new technology
occurs is endogenous. This shows both the “new growth literature” tradition and
the di¤erences to endogenous growth type RBC models. In the latter approach, the
growth rate is endogenous but shocks are still exogenously imposed. Here, the source
of growth and of ‡uctuations all stem from one and the same source, the jumps in qt
in (7.49).
Optimal behaviour of households
The resource constraint the economy needs to obey in each period is given by
Ct + It + Rt = Y;
where Ct ; It and Rt are aggregate consumption, investment and R&D expenditure,
respectively. Assume that optimal behaviour of households implies consumption and
investment into R&D amounting to
Ct = sK Yt ; Rt = s R Y t ;
where sK and sR are two constant saving rates. This would be the outcome of a two-
period maximization problem or an in…nite horizon maximization problem with some
parameter restriction similar to e.g. Benhabib and Rustichini (1994) or, in continuous
time, Xie (1991,) or Wälde (2005). As the natural volatility literature has various
papers where the saving rate is not constant (Matsuyama, 1999, assumes a constant
saving rate, but see Matsuyama, 2001), it seems reasonable not to fully develop an
optimal saving approach here as it is not central to the natural volatility view.
142
7.8.3 Equilibrium
Equilibrium is determined by
Kt+1 = (1 ) Kt + Yt Rt Ct
= (1 ) Kt + (1 sK s R ) Yt :
plus the random realization of technology jumps, where the probability of a jump
depends on investment in R&D,
pt = p (sR Yt ) :
(1 ) kt + (1 sK sR ) k t
kt+1 = : (7.50)
1 + qt
The convergence behaviour is illustrated in the following …gure.
The …gure plots the current capital stock per e¤ective labour kt on the horizontal
and kt+1 on the vertical axis. As long as there is no technology jump, i.e. as long
as qt = 0; the capital stock kt converges to its temporary steady state kq just as in
deterministic models. It follows by setting kt+1 = kt = kq in (7.50) and is given by
(1 ) kq + (1 sK sR ) kq
kq = , (qt + ) kq = (1 sK sR ) kq
1 + qt
1=(1 )
1 sK sR
, kq = : (7.51)
qt +
143
With a new technology, kt = Kt = (At L) decreases and the kt+1 line increases, as
shown by the dashed line in …g. 37. As a consequence, as can also be seen in (7.51),
the steady state increases. Subsequently, the economy approaches the steady state
again. As growth is higher immediately after a new technology is introduced (growth
is high when the economy is far away from its steady state), growth rates after the
introduction of a new technology is high and then gradually falls. Cyclical growth
is therefore characterized by a Sisyphus-type behaviour: kt permanently approaches
the steady state but eventually is thrown back due to the arrival of a new technology.
t.b.c.
144
Exercises Chapter 7
Applied Intertemporal Optimization
Stochastic di¤erence equations and
applications
(a)
A with p
At = ;
A with 1-p
¯
(b) the density function
f (At ); At 2 A,Ā ;
¯
(c) the probability function g(At ); At 2 fA1 ; A2 ; :::; An g ;
(d) At = Ai with probability pi and
(e) At+1 = At + "t+1 : Make assumptions for and "t+1 and discuss them.
What is the expected total factor productivity and what is its variance? What
is the expected output level and what is its variance, given a technology as in
(7.13)?
145
4. Saving under uncertainty
Consider an individual’s maximization problem
(a) Solve this problem by replacing her second period consumption by an ex-
pression that depends on …rst period consumption.
(b) Consider now the individual decision problem given the utility function
u (ct ) = ct : Should you assume a parameter restriction on ?
(c) What is your implicit assumption about ? Can it be negative or larger
than one? Can the time preference rate ; where = (1 + ) 1 ; be nega-
tive?
Ut = Et c1t + (1 1
) ct+1 = c1t + (1 ) Et c1t+1
(b) Discuss the link between the savings expression st and the one for the loga-
rithmic case in (7.24). Point out why ct+1 is uncertain from the perspective
of t: Is ct uncertain from the perspective of t - and of t 1?
146
7. Asset pricing
Under what conditions is there a risk neutral valuation formula for assets? In
other words, under which conditions does the following equation hold?
1
pjo = Epj1
1+r
147
8 Multi-period models
Uncertainty in dynamic models is probably most often used in discrete time models.
Looking at time with a discrete perspective has the advantage that timing issues
are very intuitive: Something happens today, something tomorrow, something the
day before. Time in the real world, however, is continuous. Take any two points
in time, you will always …nd a point in time in between. What is more, working
with continuous time models under uncertainty has quite some analytical advantages
which make certain insights - after an initial more heavy investment into techniques
- much simpler.
We will nevertheless follow the previous structure of this book and …rst present
models in discrete time. They are also, as mentioned, probably the more widely used
models in the literature.
This constraint shows why we need to form expectations about future utility: The
value of the state variable xt+1 depends on some random source, denoted by "t : Think
of this "t as uncertain TFP or uncertain returns to investment.
The optimal program is de…ned in analogy to (3.9) by
The additional term in the value function is "t : It is useful to treat "t explicitly as a
state variable for reasons we will soon see.
In order to solve the maximization problem, we again follow the three step scheme.
148
It again exploits the fact that the objective function (8.1) is additively separable,
despite the expectations operator, assumes optimal behaviour as of tommorrow and
shifts the expectations operator behind instantaneous utility of today as u (ct ) is
certain given that ct is a control variable. The …rst-order condition is
@xt+1 @f (:)
u0 (ct ) + Et Vxt+1 (xt+1 ; "t+1 ) = u0 (ct ) + Et Vxt+1 (xt+1 ; "t+1 ) = 0: (8.3)
@ct @ct
It corresponds to the …rst-order condition (3.12) in a deterministic setup. The “only”
di¤erence consists in the expectations operator Et : Equation (8.3) provides again
an (implicit) functional relationship between consumption and the state variable,
ct = ct (xt ; "t ) :
The derivative of the maximized Bellman equation with respect to the state vari-
able is (using the envelope theorem)
@xt+1
Vxt (xt ; "t ) = Et Vxt+1 (xt+1 ; "t+1 ) :
@xt
Observe that xt+1 by the constraint (8.2) is given by f (xt ; ct ; "t ) ; i.e. by quantities
that are known in t: Hence, the derivative @xt+1 =@xt = fxt is non-stochastic and we
can write this expression as (note the similarity to (3.13))
This step makes clear why it is useful to include "t as a state variable into the ar-
guments of the value function. If we had not done so, one could get the impression,
for the same argument that xt+1 is known in t due to (8.2), that the shadow price
Vxt (xt+1 ; "t+1 ) is non-stochastic as well and the expectations operator would not be
needed. Given that the value of the optimal program depends on "t ; however, it is
clear that V 0 (xt+1 ; "t+1 ) is random in t indeed. (Some of the subsequent applications
will treat "t as an implicit state variable and not always be as explicit as here.)
Using that @f (:) =@ct = fct is non-random in t; we can rewrite the …rst-order con-
dition as u0 (ct ) + fct Et Vxt+1 ((xt+1 ; "t+1 )) = 0: Inserting it in (8.4) gives Vxt (xt ; "t ) =
fxt 0 fx
f ct
u (ct ) : Shifting this expression by one period yields Vxt+1 (xt+1 ; "t+1 ) = fc t+1 u0 (ct+1 ) :
t+1
Inserting this into the costate equation (8.4) again, we obtain
fct
u0 (ct ) = Et fx u0 (ct+1 ) :
fct+1 t+1
149
8.2 Household utility maximization
Let us now look at a …rst example - a household that maximizes utility. The objective
function is given by (8.1),
t
Ut = Et 1=t u (c ) :
It is maximized subject to a budget constraint of which we know from (??) or (3.33)
that it …ts well into a general equilibrium setup,
at+1 = (1 + rt ) at + wt pt ct :
Note that the budget constraint must hold after realization of random variables, not
in expected terms. From the perspective of t; all prices (rt , wt , pt ) in t are known,
prices in t + 1 are uncertain.
Having understood in the previous general chapter that uncertainty can explicitly
be treated in the form of a state variable, we limit our attention to the endogenous
state variable at here. It will turn out that this keeps notation simpler. The value of
optimal behaviour is therefore expressed by V (at ) and the Bellman equation can be
written as
V (at ) = max fu (ct ) + Et V (at+1 )g :
ct
Di¤erentiating the maximized Bellman equation gives (using the envelope theo-
rem)
dat+1
V 0 (at ) = Et V 0 (at+1 ) , V 0 (at ) = [1 + rt ] Et V 0 (at+1 ) : (8.6)
dat
Again, the term [1 + rt ] was put in front of the expectations operator as rt is
known in t: This di¤erence equation describes the evolution of the shadow price of
wealth in the case of optimal consumption choices.
150
0
Inserting the …rst-order condition (8.5) gives V 0 (at ) = [1 + rt ] u p(ct t ) : Inserting this
expression twice into the di¤erentiated maximized Bellman equation (8.6) gives a nice
Euler equation,
u0 (ct ) u0 (ct+1 ) u0 (ct ) u0 (ct+1 )
[1 + rt ] = [1 + rt ] Et [1 + rt+1 ] , = Et :
pt pt+1 pt (1 + rt+1 ) 1 pt+1
(8.7)
Rewriting it as we did before with (7.40), we get
( )
u0 cit+1 pt
Et =1
u0 (cit ) (1 + rt+1 ) 1 pt+1
which allows to given the same interpretation as the …rst-order condition in a two-
period model, both deterministic (as in eq. (2.9) in ch. 2.2.2) and stochastic (as in
eq. (7.38) in ch. 7.6) and as in deterministic in…nite horizon models (as in eq. (3.5) in
ch. 3.2): Relative marginal utility must be equal to relative marginal prices - taking
into account that marginal utility in t + 1 is discounted at and the price in t + 1 is
discounted by using the interest rate.
Using two further assumption, the expression in (8.7) can be rewritten such that
we come even closer to the deterministic Euler equations: First, let us choose the
output good as numeraire and thereby set prices pt = pt+1 = p constant. This allows
us to remove pt and pt+1 from (8.7). Second, we assume that the interest rate is
known (say we are in a small open economy with international capital ‡ows). Hence,
expectations are formed only with respect to consumption in t + 1: Taking these two
aspects into account, we can write (8.7) as
u0 (ct ) 1
0
= : (8.8)
Et u (ct+1 ) 1= (1 + rt+1 )
This is now as close as possible to (2.9) and (3.5). The ratio of (expected discounted)
marginal utilities is identical to the ratio of relative prices.
151
DP1: Bellman equation and …rst-order conditions
The Bellman equation reads V (Kt ) = maxCt ;Ot fu (Ct ) + Et V (Kt+1 )g : The …rst-
order condition for consumption is
dV (Kt+1 )
u0 (Ct ) + Et [ 1] = 0 , u0 (Ct ) = Et V 0 (Kt+1 ) (8.9)
dKt+1
For oil it reads
@ @Yt
Et V 0 (Kt+1 ) [Yt qt O t ] = 0 , = qt :
@Ot @Ot
The derivative of the Bellman equation with respect to the capital stock Kt gives
(using the envelope theorem)
@ @Yt
V 0 (Kt ) = Et V (Kt+1 ) = 1 + Et V 0 (Kt+1 ) (8.10)
@Kt @Kt
h i
@Yt
just as in the economy without oil. The term 1 + @Kt
can be pulled out of the
@Yt
expectation operator as At and thereby @Kt
is known at the moment of the savings
decision.
Following the same steps as above without oil, we again end up with
@Yt+1
u0 (Ct ) = Et u0 (Ct+1 ) 1 +
@Kt+1
The crucial di¤erence is that now expectations are formed with respect to technolog-
ical uncertainty and with respect to uncertainty about the price of oil.
152
8.4.1 The model
Technologies
The technology used to produce a homogenous output good is of the simple Cobb-
Douglas form
Y t = A t Kt L 1 ; (8.11)
where TFP At is stochastic. Labour supply L is exogenous and …x, the capital stock
is denoted by Kt :
Households
Preferences of households are standard and given by
1 t
Ut = Et =t u (c ) :
We start from a budget constraint that can be derived as the budget constraint
(2.42) in a deterministic world. Wealth is held in units of capital Kt where the price
of one unit is vt : When we de…ne the interest rate as
wtK
rt (8.12)
vt
the budget constraint (2.42) reads
wt pt
kt+1 = (1 + rt ) kt + ct : (8.13)
vt vt
Goods market
Investment and consumption goods are traded on the same goods market. Total
supply is given by Yt ; demand is given by gross investment Kt+1 Kt + Kt and
consumption Ct . Expressed in a well-known way, goods market equilibrium yields
the resource constraint of the economy,
Kt+1 = Kt + Yt Ct Kt : (8.14)
153
8.4.3 The pricing relationship
Let us now turn to the main objective of this section and derive an expression for the
real price of one unit of capital, i.e. the price of capital in units of the consumption
good. Starting from the Euler equation (8.16), we insert the interest rate (8.12) in
its general formulation, i.e. including all prices, and rearrange to …nd
vt vt+1
= Et t+1 dt+1 + : (8.18)
pt pt+1
Note that all variables uncertain from the perspective of today in t appear behind
the expectations operator.
Now assume for a second we were in a deterministic world and the economy is in
a steady state. Equation (8.18) hcould then bei written with t+1 = 1 and without the
vt vt+1
expectations operator as pt = dt+1 + pt+1 : Solving this linear di¤erential equation
forward, starting in vt and inserting repeatedly gives
vt vt+4
= dt+1 + dt+2 + dt+3 + dt+4 +
pt pt+4
2 3 4 4 vt+4
= dt+1 + dt+2 + dt+3 + dt+4 + :
pt+4
Continuing to insert, one eventually and obviously ends up with
vt T s T vt+T
= s=1 dt+s + :
pt pt+T
The price vt of a unit of capital is equal to the discounted sum of future dividend pay-
ments plus its discounted price (once sold) in t+T . In an in…nite horizon perspective,
this becomes
vt s vt+T
= 1 s=1 dt+s + lim T :
pt T !1 pt+T
In our stochastic setup, we can proceed according to the same principles as in
the deterministic world but need to take the expectations operator
h and thei discount
vt+1 vt+2
factor t into account. We replace pt+1 in (8.18) by Et+1 t+2 dt+2 + pt+2 and then
154
vt+2 =pt+2 and so on to …nd
vt vt+3
= Et t+1 dt+1 + Et+1 t+2 dt+2 + Et+2 t+3 dt+3 +
pt pt+3
2 3 3 vt+3
= Et t+1 dt+1 + t+1 t+2 dt+2 + t+1 t+2 t+3 dt+3 + t+1 t+2 t+3
pt+3
3 vt+3
= Et s=1 s dt+s + Et 3 ; (8.19)
pt+3
where we de…ned the discount factor to be
s s s s u0 (Ct+n ) su
0
(Ct+s )
s n=1 t+n = n=1 = :
u0 (Ct+n 1 ) u0 (Ct )
The real price vt =pt amounts to the discounted sum of future dividend payments dt+s :
The discount factor is s which contains marginal utilities, relative prices and the
individual’s discount factor : The term limT !1 Et T vt+T is a “bubble term”for the
price of capital and can usually be set equal to zero. As the derivation has shown, the
expression for the price vt =pt is “simply”a rewritten version of the Euler equation.
The result on the determinants of the price of capital is useful for economic in-
tuition and received a lot of attention in the literature. But can we say more about
the real price of capital? The answer is yes and it comes from the resource constraint
(8.14). This constraint can be understood as a goods market clearing condition. The
supply of goods Yt equals demand resulting from gross investment Kt+1 Kt + Kt
and consumption. The price of one unit of the capital good therefore equals the price
of one unit of the consumption and the output good, provided that investment takes
place, i.e. It > 0: Hence,
vt = pt: : (8.21)
The real price of capital vt =pt is just equal to one. Not surprisingly, after all, capital
goods and consumption goods are traded on the same market.
155
When we want to understand what this model tells us about the evolution of
consumption, we can look at a modi…ed version of (8.16) by inserting the interest
rate (8.12) with the marginal product of capital from (8.15) and the price expression
(8.21),
@Yt+1
u0 (Ct ) = Et u0 (Ct+1 ) 1 + :
@Kt+1
This is the standard Euler equation (see e.g. (8.8)) that predicts how real consumption
evolves over time, given the real interest rate and the discount factor :
Together with (8.14), we have a system in two equations that determine Ct and
Kt (given appropriate boundary conditions). The price pt and thereby the value vt
can not be determined (which is of course a consequence of Walras’law). The relative
price is trivially unity from (8.21), vt =pt = 1: Hence, the predictions concerning real
variables do not change when a numeraire good is not chosen.9
An endowment economy
Consider an individual that can save in one asset and whose budget constraint is
given by (8.13). Let this household behave optimally such that optimal consumption
follows (8.17).
Now change the capital accumulation equation (8.14) such that - for whatever
reasons - K is constant and let also, for simplicity, depreciation be zero, = 0. Then,
output is given according to (8.11) by Yt = At K L1 ; i.e. it follows some exogenous
stochastic process, depending on the realization of At . This is the exogenous endow-
ment of the economy for each period t: Further, consumption equals output in each
period, Ct = Yt :
Inserting output into the Euler equation (8.17) gives
vt vt+1 @Yt+1
u0 (Yt ) = Et u0 (Yt+1 ) (1 ) +
pt pt+1 @Kt+1
156
8.5 Asset pricing with many assets (*)
8.5.1 A many-sector economy
Consider an economy consisiting of many sectors i. Each sector produces a good Y (i)
using capital and labour,
Yt (i) = A (i) K (i) L (i)1 a
:
Let TFP A (i) be uncertain. Capital is sector speci…c and pays a sector speci…c return
@Yt (i)
vt+1 (i) vt (i) + @Kt (i)
rt (i) = : (8.22)
vt (i)
This return takes physical depreciation into account.
subject to
at+1 = (1 + rt ) at + wt pt ct (8.23)
n
where wealth is given by at = i=1 vit kit ; the interest rate is is the weighted sum of
asset i speci…c interest rates,
n
rt = i=1 it rit : (8.24)
The weights it vit kit =at are given by the share of wealth held in asset i at t: They
sum to unity,
n n
i=1 it =1, t (1) = 1 i=2 it :
In order to solve this maximization problem, we follow the three dynamic program-
ming steps.
DP1: Bellman equation and …rst-order conditions
The Bellman equation reads
V (at ) = fu (ct ) + Et V (at+1 )g (8.25)
and …rst-order conditions are for consumption
u0 (ct ) = Et V 0 (at+1 ) pt (8.26)
and for shares t (i)
drt
Et V 0 (at+1 ) at = 0 , Et V 0 (at+1 ) [rt (1) rt (i)] = 0 (8.27)
d t (i)
In order to get an expression that allows us to link the value of an asset to future
dividend payments as in (8.16), we rewrite the foc as follows.
157
DP2: Evolution of the costate variable
First compute the derivative of the maximized Bellman equation with respect to
assets at : With the envelope theorem, this gives
As (8.27) is identical to
and also
Et V 0 (at+1 ) (1 + rt (i)) = Et V 0 (at+1 ) (1 + rt (1)) : (8.31)
Hence, with the derivative of Bellman equation we obtained this alternative version
(8.31) for the …rst-order condition (8.27) for shares.
Now let´s assume, there is one riskless asset, and call this asset ”asset number 1”.
Then (8.30) reads
V 0 (at ) = (1 + rt (1)) Et V 0 (at+1 ) (8.32)
where the interest rate was pulled in front of the expectation operator as rt (1) is
certain. This is the evolution of the shadow price v 0 (at ) which we need (in addition
to (8.31)) in what follows.
With (8.26) and (8.32) we then get a relationship between marginal utility from
consumption and the marginal value of wealth,10
u0 (ct )
V 0 (at ) = (1 + rt (1)) : (8.33)
pt
10
Without this relationship, a ”nice” Euler equation as (8.34) could not be derived. Hence, the
assumption of a risk less asset is crucial.
158
Inserting this twice into (8.32) yields
u0 (ct ) u0 (ct+1 )
(1 + rt (1)) = (1 + rt (1)) Et (1 + rt+1 (1)) ,
pt pt+1
u0 (ct ) u0 (ct+1 )
= Et (8.34)
pt (1 + rt (1)) 1 pt+1
With the usual interpretation (compare (8.7) in ch. 8.2), this is the Euler equation
for savings in the risky asset.
The …rst-order condition (8.31) for assets becomes with (8.33)
u0 (ct+1 ) u0 (ct+1 )
Et (1 + rt (i)) = Et (1 + rt (1)) (8.35)
pt+1 pt+1
where we used again that rt (1) is certain t. This expression can also be written as
u0 (ct+1 ) u0 (ct+1 )
Et = Et :
(1 + rt+1 (1)) 1 pt+1 (1 + rt+1 (i)) 1 pt+1
The Euler equation (8.34) can therefore alternatively be written as
u0 (ct ) u0 (ct+1 )
= Et (8.36)
pt (1 + rt+1 (i)) 1 pt+1
This is the Euler equation expressing returns in terms of the risky asset i: Comparing
(8.34) and (8.36) shows the indi¤erence between marginally saving in asset i or the
risken asset.
159
Noting that Et Et+1 = Et and inserting very often, one obtains
Hence, the value of a …rm is the discounted sum of future divided payments plus
u0 (cj+1 )=pj+1
a bubble term. The interesting aspect is that the discount factor sj=t u0 (cj )=pj
depends on preferences of households.
Note that the expectations operator always refers to knowledge available at the be-
ginning of the planning horizon, i.e. to t = 0:
Now compute the …rst-order condition with respect to as+1 : This is unusual as we
directly choose the state variable which is usually understood to be indirectly chosen
by the control variable. Clearly, however, this is just a convenient trick and we really,
by choosing as+1 ; a state variable, we choose cs . The derivative with respect to as+1
yields
E0 s u0 (cs ) = E0 s+1 u0 (cs+1 ) (1 + rs+1 ):
160
This almost looks like the standard optimal consumption rule. The di¤erence consists
in the expectations operator being present on both sides. This is not surprising as we
optimally chose as+1 (i.e. cs ), knowing only the state of the system in t = 0: If we now
assume we are in s; our expectations would be based on knowledge in s and we could
replace E0 by Es : We would then obtain Es s u0 (cs ) = s u0 (cs ) for the left-hand side
and our optimality rule reads
u0 (cs ) = Es u0 (cs+1 ) (1 + rs+1 ):
This is the rule we know from Bellman approaches, provided e.g. in (8.7).
In period 0; the individual uses labour income w0 to pay for consumption goods c0 ; to
buy m0 units of the long-term asset and for “normal”assets a0 . In periods 1 to T 1;
the individual uses her assets at 1 plus returns r on assets and her wage income wt to
…nance consumption and buy again assets at (or keep those of the previous period).
In the …nal period T; the long-term asset has a price of pT and is sold. Wealth of the
previous period plus interest plus labour income wT are further sources of income to
pay for consumption cT .
This maximization problem can most easily be solved by inserting consumption
levels for each period into the objective function. The objective function then reads
T 1 t
u (w0 m0 p0 a0 ) + E0 t=1 u (wt + (1 + r)at 1 at )
T
+ E0 u(m0 pT + (1 + r)aT 1 + wT )
and the control variables are wealth holdings at in periods t = 0; :::; T 1 and the
number of assets m0 bought in period zero.
Let us now look at the …rst-order conditions. The …rst-order condition for wealth
in period zero is
u0 (c0 ) = (1 + r) E0 u0 (c1 ) : (8.37)
Wealth holdings in any period t > 0 are optimally chosen according to
E0 t u0 (ct ) = E0 t+1
(1 + r)u0 (ct+1 ) , E0 u0 (ct ) = (1 + r) E0 u0 (ct+1 ) : (8.38)
161
We can insert (8.38) into the …rst-period condition (8.37) su¢ ciently often and …nd
u0 (c0 ) = (1 + r)2 2
E0 u0 (c2 ) = ::: = (1 + r)T T
E0 u0 (cT ) (8.39)
When we insert combined …rst-order conditions (8.39) for wealth holdings into the
…rst-order condition (8.40) for assets, we obtain
T T
p0 (1 + r)T E0 u0 (cT ) = E0 u0 (cT )pT ,
u0 (cT )
p0 = (1 + r) T E0 pT :
E0 u0 (cT )
which is an equation where we nicely see the analogy to the two period example in
1 u0 (c1 )
ch. 7.7.1. Instead of pj0 = 1+r E Eu 0 (c ) j1 in (7.46) where we discount by one period
1
p
only and evaluate returns at expected marginal utility in period 1, we discount by T
periods and evaluate returns at marginal utility in period T:
This equation also o¤ers a lesson for life when we assume risk-neutrality for sim-
plicity: if the payo¤ pT of a long-term asset is not high enough such that the current
price is higher than the present value of the payo¤, p0 > (1 + r) T pT , then the long-
term asset is simply dominated by short-term investments that pay a return of r
per period. Optimal behaviour would imply not to buy the long-term asset and just
put wealth in “normal” assets. This should be kept in mind when talking next time
to your insurance agent who tries to sell you life-insurance or private pension plans.
Just ask for the present value of the payo¤s and compare them to the present value
of what you pay into the savings plan.
Sticky prices are a fact of life. In macro economic models, they are either assumed
or result from some adjustment-cost setup. Here is a simpli…ed way to derive sluggish
price adjustment. This example is inspired by Ireland (2004 NBER WP), going back
to Rotemberg (1982 JPE).
The objective function of the …rm is to maximize its present value de…ned by the
sum of discounted expected pro…ts,
t
1 1
Vt = Et =t :
1+r
=p x w l (p ; p 1)
162
where (p ; p 1 ) are price adjustment costs. These are similar in spirit to the
adjustment costs presented in ch. 5.5.1. We will later use a speci…cation given by
2
p p 1
(p ; p 1) = : (8.41)
2 p 1
This speci…cation only captures the essential mechanism that is required to make
prices sticky, i.e. the fact that the price change is squared (in fact, as with all
adjustment cost mechanisms, any power larger than 1 should do the job). More care
about economic implications needs to be taken when a reasonable model is to be
speci…ed.
The …rm uses a technology
x =A l :
We assume that there is a certain demand elasticity " for the …rm’s output. This
can re‡ect a monopolistic competition setup. The …rm can choose its output x at
each point in time freely by hiring the corresponding amount of labour l : Labour
productivity A or other quantities can be uncertain.
A solution
The second and third line present a rewritten way that allows to better see the in-
tertemporal structure of the maximization problem. We now maximize this objective
function by choosing output xt for today. (Output levels in the future are chosen at
a later stage.)
The …rst-order condition is
d [pt xt ] wt d (pt ; pt 1 ) d (pt+1 ; pt )
Et Et =0,
dxt At dxt dxt
d [pt xt ] wt d (pt ; pt 1 ) d (pt+1 ; pt )
= + + Et :
dxt At dxt dxt
It has certain well-known components and some new ones. If there were no adjustment
costs, i.e. (:) = 0; the intertemporal problem would become a static one and the
usual condition would equate marginal revenue d [pt xt ] =dxt with marginal cost wt =At .
With adjustment cost, however, a change in output today not only a¤ects adjustment
cost today d (pdx
t ;pt 1 )
t
but also (expected) adjustment cost Et d (pdx
t+1 ;pt )
t
tomorrow. As
163
all variables with index t are assumed to be known in t; expectations are formed only
with respect to adjustment cost tomorrow in t + 1:
Specifying the adjustment cost function as in (8.41) and computing marginal rev-
enue using the demand elasticity "t gives
" # " #
2 2
d [pt xt ] wt d pt pt 1 d pt+1 pt
= + + Et ,
dxt At dxt 2 pt 1 dxt 2 pt
dpt 1 wt pt pt 1 1 dpt pt+1 pt d pt+1
xt + pt = 1 + t pt = + + Et
dxt At pt 1 pt 1 dxt pt dxt pt
wt pt pt 1 pt pt+1 pt pt+1 dpt
= + "t 1 Et ,
At pt 1 xt pt 1 pt p2t dxt
wt pt pt 1 pt pt+1 pt pt+1 pt 1
1 + t 1 pt = + "t 1 Et " ,
At pt 1 xt pt 1 pt p2t xt t
wt pt pt 1 pt 1 pt+1 pt
1 + t 1 pt = + "t 1 "t 1 Et pt+1
At pt 1 xt pt 1 pt xt pt
dxt pt
where "t dpt xt
is the demand elasticity for the …rm’s good. Again, = 0 would
give the standard static optimality condition 1 + t 1 pt = wt =At where the price is
a markup over marginal cost. With adjustment cost, prices change only slowly.
1 1
t = Et =t t : (8.42)
(1 + r)
We are today in t; time extends until in…nity and the time between today and in…nity
is denoted by : There is no particular reason why the planning horizon is in…nity in
contrast to ch. 5.5.1. The optimality condition for employment we will stress here is
identical to a …nite horizon problem.
Instantaneous pro…ts are given by the di¤erence between revenue in t; which is
identical to output F (Lt ) with an output price normalized to unity, labour cost wt Lt
and adjustment cost (Lt Lt 1 ) ;
Costs induced by the adjustment of the number of employees between the previous
period and today are captured by (:). Usually, one assumes costs both from hiring
and from …ring individuals, i.e. both for an increase in the labour force, Lt Lt 1 >
0; and from a decrease, Lt Lt 1 < 0: A simple functional form for (:) which
captures this idea is a quadratic form, i.e. (Lt Lt 1 ) = 2 (Lt Lt 1 )2 , where is
a constant.
164
Uncertainty for a …rm can come from many sources: Uncertain demand, uncer-
tainty about the production process, uncertainty in labour cost or other sources. As
we express pro…ts in units of the output good and assume that the real wage wt ,
i.e. the amount of output goods to be paid to labour, is uncertain. Adjustment cost
(Lt Lt 1 ) are certain, i.e. the …rm knows by how many units of output pro…ts
reduce when emplyoment changes by Lt Lt 1 :
As in static models of the …rm, the control variable of the …rm is employment Lt :
In contrast to static models, however, employment decisions today in t have not only
e¤ects on employment today but also on employment tommorrow as the employment
decision in t a¤ects adjustment costs in t + 1: There is therefore an intertemporal link
the …rm needs to take into account which is not present in static models of the …rm.
Solution
This maximization problem can be solved directly by inserting (8.43) into the
objective function (8.42). One can then choose optimal employment for some point
in time t s < 1 after having split the objective function into several subperiods -
as e.g. in the previous chapter 8.7.1. The solution reads (to be shown in exercise 6)
0
0 0 (Lt+1 Lt )
F (Lt ) = wt + (Lt Lt 1 ) Et :
1+r
When employment Lt is chosen in t; there is only uncertainty about Lt+1 : The current
wage wt (and all other deterministic quantities as well) are known with certainty. Lt+1
is uncertain, however, from the perspective of today as the wage in + 1 is unknown
and L +1 will be have to be adjusted accordingly in + 1: Hence, expectations ap-
ply only to the adjustment-cost term which refers to adjustment cost which occur in
period t + 1: Economically speaking, given employment Lt 1 in the previous period,
employment in t is chosen such that marginal productivity of labour equals labour
cost adjusted for current and expected future adjustment cost. Expected future ad-
justment cost are discounted by the interest rate r to obtain its present value.
When we specify the adjustment cost function as a quadratic function, (Lt Lt 1 ) =
2
[Lt Lt 1 ]2 ; we obtain
[Lt+1 Lt ]
F 0 (Lt ) = wt + [Lt Lt 1 ] Et :
1+r
If there were no adjustment cost, i.e. = 0; we would have F 0 (Lt ) = wt : Employment
would be chosen such that marginal productivity equals the real wage. This con…rms
the initial statement that the intertemporal problem of the …rm arises purely from
the adjustment costs. Without adjustment cost, i.e. with = 0; the …rm has the
“standard”instantaneous, period-speci…c optimality condition.
165
some future point in time at some exogenous level. This allows then to easily (at least
numerically) compute the optimal path for all points in time before this …nal point.
Let T be the …nal period of life in our model, i.e. set aT +1 = 0 (or some other
level e.g. the deterministic steady state level). Then, from the budget constraint, we
can deduce consumption in T;
aT +1 = (1 + rT ) aT + wT cT , cT = (1 + rT ) aT + wT :
u0 (cT 1) = ET 1 [1 + rT ] u0 (cT ) :
aT = (1 + rT 1) aT 1 + wT 1 cT 1;
166
Exercises Chapter 8
Applied Intertemporal Optimization
Discrete time in…nite horizon models
under uncertainty
1. Central planner
Consider an economy where output is produced by
Yt = At Kt Lt1
Kt+1 = (1 ) Kt + Y t Ct
3. Closed-form solution
1
Solve this model for the utility function u (C) = C 1 1 and for = 1: Solve it
for a more general case (Benhabib and Rustichini, 1994).
4. Habit formation
Assume instantaneous utility depends not only on current consumption but also
on habits (see e.g. Abel, 1990). Let the utility function therefore look like
1 t
Ut = Et =t u (c ; v ) ;
at+1 = (1 + rt ) at + wt pt ct
167
(a) Assume the individual lives in a deterministic world and derive a rule for
an optimal consumption path where the e¤ect of habits are explicitly taken
into account. Specify habits by v = c 1 :
(b) Let there be uncertainty with respect to future prices. At a point in time
t, all variables indexed by t are known. What is the optimal consumption
rule when habits are treated in a parametric way?
(c) Choose a plausible instantaneous utility function and discuss the implica-
tions for optimal consumption given habits v = c 1 .
5. Risk-neutral valuation
Under which conditions is there a risk neutral valuation relationship for contin-
gent claims in models with many periods?
7. Solving by inserting
Solve the problem from ch. 8.7 in a slightly extended version, i.e. with prices pt .
t
Maximize E0 1 t=0 u (ct ) by choosing a time path fct g for consumption subject
to at+1 = (1 + rt ) at + wt pt ct :
Lt+1 = (1 s) Lt + Vt ;
= p Y (L ) w L V :
(a) Assume a deterministic world. Let the …rm choose the number of vacancies
optimally. Use a Lagrangian to derive the optimality condition. Assume
that there is an interior solution. Why is this an assumption that might
not always be satis…ed from the perspective of a single …rm?
168
(b) Let us now assume that there is uncertain demand which translates into
uncertain prices p which are exogenous to the …rm. Solve the optimal
choice of the …rm by inserting all equations into the objective function.
Maximize by choosing the state variable and explain also in words what
you do. Give an interpretation of the optimality condition. What does it
imply for the optimal choice of V ?
169
Part IV
Stochastic models in continuous
time
The choice between working in discrete or continuous time is partly driven by previ-
ous choices: If the literature is mainly in discrete time, students will …nd it helpful
to work in discrete time as well. This dominance of discrete time seems to hold
for macroeconomics. On the other hand, there are very good books that can help
as introductions (and go well beyond an introduction) to work with uncertainty in
continuous time. An example is Turnovsky (1997, 2000).
Whatever the tradition in the literature, continuous time models have the huge
advantage that they are analytically generally more tractable, once some initial invest-
ment into new methods is digested. One example is, as some papers in the literature
have shown (e.g. Wälde, 2005 or Posch and Wälde, 2005), that continuous time mod-
els with uncertainty can be analyzed in simple phase diagrams as in deterministic
continuous time setups. In other …elds, continuous time uncertainty is dominating,
as e.g. in …nance. See ch. 9.5 and 10.6 on further reading for references from many
…elds.
This …nal part of the book presents tools that allow to work with uncertainty in
continuous time models. It is probably the most innovative part of this book as many
results from recent research directly ‡ow into it.
170
ministic) dynamic models: Static models describe one equilibrium, dynamic models
describe a sequence of equilibria. A random variable has “when looked at once”(e.g.
when throwing a die once) one realization. A stochastic process describes a sequence
of random variables and therefore “when looked at once”, describes a sequence of
realizations. More formally, we have the following
Let us look at an example for a stochastic process. We start from the normal
distribution of ch. 7.2.2 whose mean and variance are given by and 2 and its
p 1 1 z 2
density function is f (z) = 2 2 e 2 ( ) : Now de…ne a normally distributed
random variable Z (t) that has a mean and a variance that is a function of some t,
i.e. instead of 2 , write 2 t: Hence, the random variables we just de…ned has as density
p 1 1 zp
2
De…nition 9 (Ross, 1996, ch. 8.8): A process X (t) is stationary if X (t1 ) ; :::; X (tn )
and X (t1 + s) ; :::; X (tn + s) have the same joint distribution for all n and s:
An implication of this de…nition, which might help to get some “feeling” for this
de…nition, is that a stationary process X (t) implies that, being in t = 0, X (t1 ) and
X (t2 ) have the same distribution for all t2 > t1 > 0: A weaker concept of stationarity
only requires the …rst two moments of X (t1 ) and X (t2 ) (and a condition on the
covariance) to be satis…ed.
De…nition 10 (Ross, 1996) A process X (t) is weakly stationary if the …rst two mo-
ments are the same for all t and the covariance between X (t2 ) and X (t1 ) depends
only on t2 t1 ;
2
E0 X (t) = ; V arX (t) = ; Cov (X (t2 ) ; X (t1 )) = f (t2 t1 ) ;
2
where and are constants and f (:) is some function.
The probably best known stochastic process, especially from …nance, is Brownian
motion. It is sometimes called Wiener process after the mathematician Wiener who
provided the following de…nition.
171
The …rst condition z (0) = 0 is a normalization. Any z (t) that starts at, say, z0
can be rede…ned as z (t) z0 : The second condition says that for t4 > t3 t2 > t1
the increment z (t4 ) z (t3 ) ; which is a random variable, is independent of previous
increments, say z (t2 ) z (t1 ). ’Independent increments’implies that Brownian motion
is a Markov process: Assuming that we are in t3 today, the distribution of z (t4 )
depends only on z (t3 ) ; i.e. on the current state, and not on previous states like
z (t1 ). Increments are said to be stationary if, according to the above de…nition of
stationarity, the stochastic process X (t) z (t) z (t s) where s is a constant
has the same distribution for any t. The third condition is …nally the heart of the
de…nition - z (t) is normally distributed. The variance increases linearly in time; the
Wiener process is therefore nonstationary.
Let us now de…ne a stochastic process which plays also a major role in economics.
De…nition 12 Poisson process (adapted following Ross 1993, p. 210)
A stochastic process q (t) is a Poisson process with arrival rate if (i) q (0) = 0; (ii)
the process has independent increments and (iii) the increment q ( ) q (t) in any
interval of length t (the number of “jumps”) is Poisson distributed with mean
[ t] ; i.e. q ( ) q (t) ~Poisson( [ t]) :
A Poisson process (and other related processes) are also sometimes called “count-
ing processes”as q (t) counts how often a jump has occured, i.e. how often something
has happend.
There is a close similarity in the …rst two points of this de…nition with the de…nition
of Brownian motion. The third point here means more precisely that the probability
that the process increases n times between t and > t is given by
[ t] ( [ t])n
P [q ( ) q (t) = n] = e ; n = 0; 1; ::: (9.1)
n!
We know this probability from the de…nition of the Poisson distribution in ch. 7.2.1.
This is probably where the Poisson process got its name from. Hence, one could think
of as many stochastic processes as there are distributions, de…ning each process by
the distribution of its increments.
The most common way how Poisson processes are presented is by looking at the
distribution of the increment q ( ) q (t) over a very small time intervall [t; ] : The
increment q ( ) q (t) for very close to t is usually expressed by dq (t) : A stochastic
process q (t) is then a Poisson process if its increment dq (t) is driven by
n
dq (t) = 01 with prob. 1 dt
with prob. dt ; (9.2)
where the parameter is again called the arrival rate. A high then means that the
process jumps more often than with a low .11
These stochastic processes (and other processes) can now be combined in various
ways to construct more complex processes. These more complex processes can nicely
be represented by stochastic di¤erential equations (SDEs).
11
Note that the probabilities given in (9.2) are an approximation to the ones in (9.1) for t = dt:
We will return to this in ch. 9.4.2, see Poisson process II.
172
9.1.2 Stochastic di¤erential equations
The most frequently used SDEs include Brownian motion as the source of uncertainty.
These SDEs are used to model e.g. the evolution of asset prices or budget constraints
of households. Other examples include SDEs with Poisson uncertainty used explicitly
e.g. in the natural volatility literature, in …nance, international macro or in other
contexts mentioned above. Finally and more recently, Levy processes are used in
…nance as they allow for a much wider choice of properties of distributions of asset
returns than e.g. Brownian motion. We will now get to know examples for each type.
For all Brownian motions that will follow, we will assume, unless explicitly stated
otherwise, that increments have a standard normal distribution, i.e. Et [z ( ) z (t)] =
0 and vart [z ( ) z (t)] = t: It is therefore su¢ cient, consistent with most papers in
the literature and many mathematical textbooks, to work with a normalization of
in de…nition 11 of Brownian motion to 1:
Brownian motion with drift
This is one of the simplest SDEs. It reads
dx (t) = adt + bdz (t) : (9.3)
2
The constant a can be called drift rate, b is sometimes refered to as the variance rate
of x (t) : In fact, ch. 9.4.4 shows that the expected increase of x (t) is determined by
a only (and not by b). In contrast, the variance of x ( ) for some future > t is only
determined by b. The drift rate a is multiplied by dt; a “short” time intervall, the
variance parameter b is multiplied by dz (t) ; the increment of the Brownian motion
process z (t) over a small time intervall. This SDE (and all the others following later)
therefore consists of a deterministic part (the dt-term) and a stochastic part (the
dz-term).
An intuition for this di¤erential equation can be most easily gained by undertaking
a comparison with a deterministic di¤erential equation. If we neglected the Wiener
process for a moment (set b = 0), divide by dt and rename the variable to y; we obtain
the simple ordinary di¤erential equation
y_ (t) = a (9.4)
whose solution is y (t) = y0 + at: When we draw this solution and also the above SDE
for three di¤erent realizations of z (t), we obtain the following …gure.
173
Figure 38 The solution of the deterministic di¤erential equation (9.4) and three re-
alizations of the related stochastic di¤erential equation (9.3) Realization B M _ m ulti.n b
Here, one also refers to a (:) as the drift rate and to b2 (:) as the instantaneous variance
rate. Note that these functions can be stochastic themselves. In addition to arguments
x (t) and time, Brownian motion z (t) can be included in these arguments. Thinking
of (9.5) as a budget constraint of a household, an example could be that wage income
or the interest rate depend on the current realization of the economy’s fundamental
source of uncertainty, which is z (t) :
Di¤erential equations can of course also be constructed that are driven by a Pois-
son process. A very simple example is
A realization of this path for x (0) = x0 is in the following …gure and can be understood
very easily. As long as no jump occurs, i.e. as long as dq = 0; the variable x (t) follows
dx (t) = adt which means linear growth, x (t) = x0 + at: This is plotted as the thin
line. When q jumps, i.e. dq = 1; x (t) increases by b : Writing dx (t) = x~ (t) x (t) ;
where x~ (t) is the level of x immediately after the jump, and letting the jump be “very
fast”such that dt = 0 during the jump, we have x~ (t) x (t) = b 1; where the 1 stems
from dq (t) = 1: Hence,
x~ (t) = x (t) + b: (9.7)
Clearly, the point in times when a jump occurs are random. A tilde (~) will always
denote in what follows (and in various papers in the literature) the value of a quantity
immediately after a jump.
174
Figure 39 An example of a Poisson process with drift (thick line) and a deterministic
di¤erential equation (thin line)
Processes are usually called geometric when they describe the rate of change of some
RV x (t) ; i.e. dx (t) =x (t) is not a function of x (t) : In this example, the deterministic
part shows that x (t) grows at the rate of a (:) in a deterministic way and jumps by
b (:) percent, when q (t) jumps. Note that in contrast to a Brownian motion SDE,
a (:) here is not the average growth rate of x (t) (see below on expectations).
Geometric Poisson processes as here are sometimes used to describe the evolution
of asset prices in a simple way. There is some deterministic growth component a (:)
12
The technical background for the t notation is the fact that the process x (t) is a so called
cádlág process ("continu a droite, limites a gauche"). That is, the paths of x (t) are continuous from
the right with left limits. The left limit is denoted by x (t ) lims"t x (s). See Sennewald (2006) or
Sennewald and Wälde (2006) for further details and references to the mathematical literature.
175
and some stochastic component b (:) : When the latter is positive, this could re‡ect
new technologies in the economy. When the latter is negative, this equation could be
used to model negative shocks like oil-price shocks or natural disasters.
Aggregate uncertainty and random jumps
An interesting extension of a Poisson di¤erential equation consists in making the
amplitude of the jump random. Taking a simple di¤erential equation with Poisson
uncertainty as starting point, dA (t) = bA (t) dq (t) ; where b is a constant, we can
now assume that b (t) is governed by some distribution, i.e.
2
dA (t) = b (t) A (t) dq (t) ; where b (t) ~ ; : (9.9)
Assume that A (t) is total factor productivity in an economy. Then, A (t) does not
change as long as dq (t) = 0: When q (t) jumps, A (t) changes by b (t) ; i.e. dA (t)
A~ (t) A (t) = b (t) A (t) ; which we can rewrite as
A~ (t) = (1 + b (t)) A (t) ; 8t where q (t) jumps.
This equation says that whenever a jump occurs, A (t) increases by b (t) percent,
i.e. by the realization of the random variable b (t) : Obviously, the realization of b (t)
matters only for points in time where q (t) jumps.
Note that (9.9) is the stochastic di¤erential equation representation of the evolu-
tion of the states of the economy in the Pissarides-type matching model of Shimer
(2005), where aggregate uncertainty, A (t) here, follows from a Poisson process. The
presentation in Shimer’s paper is “A shock hits the economy according to a Pois-
son process with arrival rate , at which point a new pair (p0 ; s0 ) is drawn from a
state dependent distribution.”(p. 34). Note also that using (9.9) and assuming large
families such that there is no uncertainty from labour income left on the household
level would allow to analyze the e¤ects of saving and thereby capital accumulation
over the business cycle in a closed-economy model with risk-averse households. The
background for the saving decision would be ch. 10.1.
176
This version is obtain by …rst rewriting (9.5) as dx (s) = a (x; s) ds +b (x; s) dz (s) ;
i.e by simply changing the time index from t to s (and dropping zR(s) and writting
x instead of x (s) to shorten notation). Applying then the integral t on both sides
gives (9.10).
This implies, inter alia, a “di¤erentiation rule”
Z Z
d a(x; s)ds + b(x; s)dz (s) = d [x ( ) x (t)] = dx ( )
t t
= a(x; )d + b(x; )dz ( ) :
Poisson processes
Now consider a generalized version of the SDE in (9.6), with again replacing t by
s, dx (s) = a (xR(s) ; q (s)) ds +b (x (s) ; q (s)) dq (s) : The integral representation reads,
after applying t to both sides,
Z Z
x ( ) x (t) = a (x (s) ; q (s)) ds + b (x (s) ; q (s)) dq (s) :
t t
177
We can now insert equations describing the evolution of TFP and capital, dA (t) and
dK (t) ; and take into account that employment L is constant. This gives
Dividing by dt would give a di¤erential equation that describes the growth of Y; i.e.
Y_ (t) :
The objective of the subsequent sections is to provide rules on how to compute
di¤erentials, of which dY (A (t) ; K (t) ; L) is an example, in setups where K (t) or
A (t) are described by stochastic DEs and not ordinary DEs as just used in this
example.
(dx)2 = a2 (:) (dt)2 + 2a (:) b (:) dtdz + b2 (:) (dz)2 = b2 (:) dt;
where the last equality used the “rules”from (9.12). Unfortunately, these rules have
a pretty complex stochastic background and an explanation would go far beyond the
objective of these notes. For those interested in the details, see the further-reading
chapter 9.5 on “general background”. The di¤erential of F (t; x) then reads
1
dF = Ft dt + Fx a (:) dt + Fx b (:) dz + Fxx b2 (:) dt
2
1
= Ft + Fx a (:) + Fxx b2 (:) dt + Fx b (:) dz: (9.13)
2
When we compare this di¤erential with the “normal”one, we recognize familiar terms:
The partial derivatives times deterministic changes, Ft +Fx a (:) ; would appear also in
circumstances where x follows a deterministic evolution. Put di¤erently, for b (:) = 0
in (9.5), the di¤erential dF reduces to fFt + Fx a (:)g dt: Brownian motion therefore
a¤ects the di¤erential dF in two ways: First, the stochastic term dz is added and sec-
ond, maybe more “surprisingly”, the deterministic part of dF is also a¤ected through
the quadratic term containing the second derivative Fxx :
178
The lemma for many stochastic processes
This was the simple case of one stochastic process. Now consider the case of many
stochastic processes. Think of the price of many stocks traded on the stock market.
We then have the following
or in matrix notation
dx = udt + vdz(t)
where
0 1 0 1 0 1 0 1
x1 u1 v11 ::: v1m dz1
B C B .. C ; v = B .. C ; dz = B .. C :
x = @ ... A ; u = @ . A @
..
. . A @ . A
xn un vn1 ::: vnm dzm
Consider further a function F (t; x) from [0; 1[ Rn to R with time t and the n
processes in x as arguments. Then
n 1 n n
dF (t; x) = Ft dt + i=1 Fxi dxi + i=1 j=1 Fxi xj [dxi dxj ] (9.14)
2
where, as an extension to (9.12),
Let us now consider an example for a function F (t; x; y) of two stochastic processes.
As an example, assume that x is described by a generalized Brownian motion similar
to (9.5), dx = a (x; y; t) dt + b (x; y; t) dzx and the stochastic process y is described by
dy = c (x; y; t) dt + g (x; y; t) dzy : Ito’s Lemma (9.14) gives the di¤erential dF as
1
dF = Ft dt + Fx dx + Fy dy + Fxx (dx)2 + 2Fxy dxdy + Fyy (dy)2 (9.16)
2
Given the rule in (9.15), the squares and the product in (9.16) are
(dx)2 = b2 (t; x; y) dt; (dy)2 = g 2 (t; x; y) dt; dxdy = xy b (:) g (:) dt;
179
where xy is the correlation coe¢ cient of the two processes. More precisely, it is the
correlation coe¢ cient of the two normally distributed random variables that underlie
the Wiener processes. The di¤erential (9.16) therefore reads
Note that this di¤erential is almost simply the sum of the di¤erentials of each
stochastic process independently. The only term that is added is the term that
contains the correlation coe¢ cient. In other words, if the two stochastic processes
were independent, the di¤erential of a function of several stochastic processes equals
the sum of the di¤erential of each stochastic process individually.
where the second equality used (9.12). The di¤erential of F (t; x) therefore reads from
(9.14)
m 1 m 2
dF (t; x) = Ft dt + Fx [u1 dt + i=1 vi dzi ] + Fxx [ i=1 vi dzi ]
2
1 m 2 m
= fFt + Fx u1 g dt + Fxx [ i=1 vi dzi ] + Fx i=1 vi dzi :
2
2
Computing the [ m
i=1 vi dzi ] term requires to take potential correlations into account.
For any two uncorrelated increments dzi and dzj ; dzi dzj would from (9.15) be zero.
When they are correlated, dzi dzj = ij dt which includes the case of dzi dzi = dt:
Lemma 5 Let there be a stochastic process x (t) driven by Poisson uncertainty q (t)
described by
dx (t) = a (:) dt + b (:) dq (t) :
180
What was stressed before for Brownian motion is valid here as well: The functions
a (:) and b (:) in the deterministic and stochastic part of this SDE can have as argu-
ments any combinations of q (t) ; x (t) and t or can be simple constants. Consider the
function F (t; x) : The di¤erential of this function is
This rule is very intuitive: The di¤erential of a function is given by the “normal
terms” and by a “jump term”. The “normal terms” include the partial derivatives
with respect to time t and x times changes per unit of time (1 for the …rst argument
and a (:) for x) times dt. Whenever the process q increases, x increases by the b (:) :
The “jump term” therefore captures that the function F (:) jumps from F (t; x) to
F (t; x~) = F (t; x + b (:)).
Lemma 6 Let there be two independent Poisson processes qx and qy driving two
stochastic processes x (t) and y (t) ;
Again, this “di¤erentiation rule” consists of the “normal” terms and the “jump
terms”. As the function F (:) depends on two arguments, the normal term contains
two drift components, Fx a (:) and Fy c (:) and the jump term contains the e¤ect of
jumps in qx and in qy : Note that the dt term does not contain the time derivative
Ft (x; y) as in this example F (x; y) is assumed not to be a function of time and
therefore Ft (x; y) = 0: In applications where F (:) is a function of time, the Ft (:)
would of course have to be taken into consideration.
Let us now consider a case that is frequently encountered in economic models
when there is one economy-wide source of uncertainty, say new technologies arrive or
commodity price shocks occur according to some Poisson process, and many variables
in this economy (e.g. all relative prices) are a¤ected simultaneously by this one shock.
The CVF in situations of this type reads
where uncertainty stems from the same q for both variables. Consider the function
F (x; y) : The di¤erential of this function is
181
One nice feature about di¤erentiation rules for Poisson processes is their very
intuitive structure. When there are two independent Poisson processes as in (9.19),
the change in F is given by either F (x + b (:) ; y) F (x; y) or by F (x; y + g (:))
F (x; y) ; depending on whether one or the other Poisson process jumps. When both
arguments x and y are a¤ected by the same Poisson process, the change in F is given
by F (x + b (:) ; y + g (:)) F (x; y) ; i.e. the level of F after a simultaneous change of
both x and y minus the pre-jump level F (x; y).
We now present the most general case. Let there be n stochastic processes xi (t)
and de…ne the vector x (t) = (x1 (t) ; :::; xn (t))T : Let stochastic processes be described
by n SDEs
where ij (:) stands for ij (t; x (t)) : Each stochastic process xi (t) is driven by the
same m Poisson processes. The impact of Poisson process qj on xi (t) is captured by
ij (:) : Note the similarity to the setup for the Brownian motion case in (9.14).
Proposition 8 (Sennewald 2006 and Sennewald and Wälde 2006) Let there be n sto-
chastic processes described by (9.20). For a once continuously di¤erentiable function
F (t; x), the process F (t; x) obeys
n
dF (t; x (t)) = fFt (:) + i=1 Fxi (:) i (:)g dt
m
+ j=1 F t; x (t) + j (:) F (t; x (t)) dqj , (9.21)
where ft and fxi , i = 1; : : : ; n, denote the partial derivatives of f with respect to t and
T
xi , respectively, and j stands for the n-dimensional vector function 1j ; : : : ; nj .
182
Lemma 8 Let there be a variable x which is described by
and where uncertainty stems from Brownian motion z and a Poisson process q. Con-
sider the function F (t; x) : The di¤erential of this function is
1
dF (t; x) = Ft + Fx a (:) + Fxx b2 (:) dt+Fx b (:) dz +fF (t; x + g (:)) F (t; x)g dq:
2
(9.23)
Note that this lemma is just a “combination”of Ito’s Lemma (9.13) and the CVF
for a Poisson process from (9.18). For an arrival rate of zero, i.e. for dq = 0 at all
times, (9.23) is identical to (9.13). For b (:) = 0; (9.23) is identical to (9.18).
The starting point is the price S of an asset which evolves according to a geometric
process
dS
= dt + dz: (9.24)
S
Uncertainty is modelled by the increment dz of Brownian motion. We assume that
the economic environment is such (inter alia short selling is possible, there are no
transaction cost) that the price of the option is given by a function F (:) having as
arguments only the price of the asset and time, F (t; S (t)): The di¤erential of the
price of the option is then given from (9.11) by
1
dF = Ft dt + FS dS + FSS (dS)2 : (9.25)
2
As the square of dS is given by (dS)2 = 2
S 2 dt; the di¤erential reads
1 2 dF
dF = Ft + SFS + S 2 FSS dt + SFS dz , F dt + F dz (9.26)
2 F
where the last step de…ned
1 2
Ft + SFS + 2
S 2 FSS SFS
F = ; F = : (9.27)
F F
183
Absence of arbitrage
Now comes the trick - the no-arbitrage consideration. Consider a portfolio that
consists of N1 units of the asset itself, N2 options and N3 units of some riskless assets,
say wealth in a savings account. The price of such a portfolio is then given by
P = N1 S + N2 F + N3 M;
where M is the price of one unit of the riskless asset. The proportional change of the
price of this portfolio can be expressed as (holding the Ni s constant, otherwise Ito’s
Lemma would have to be used)
dP N1 S dS N2 F dF N3 M dM
dP = N1 dS + N2 dF + N3 dM , = + + :
P P S P F P M
De…ning shares of the portfolio held in these three assets by 1 N1 S=P and 2
N2 S=P; inserting option and stock price evolutions from (9.24) and (9.25) and letting
the riskless asset M pay a constant return of r, we obtain
Now assume someone chooses weights such that the portfolio no longer bears any risk
1 + 2 F = 0: (9.29)
The return of such a portfolio with these weights must then of course be identical to
the return of the riskless interest asset, i.e. identical to r;
dP=dt r F r
= 1 [ r] + 2 [ F r] + rj = =r, = :
P riskless
1 2 F
F
If the return of the riskless portfolio did not equal the return of the riskless interest
rates, there would be arbitrage possibilities. This approach is therefore called no-
arbitrage pricing.
Finally, inserting F and F from (9.27) yields the celebrated di¤erential equation
that determines the evolution of the price of the option,
1 2
r Ft + SFS + S 2 FSS rF 1 2
= 2
, S 2 FSS + rSFS rF + Ft = 0: (9.30)
SFS 2
Clearly, this equation does not to say what the price F of the option actually is. It
only says how it changes over time and in reaction to S. But as we will see in ch.
9.3.1, this equation can actually be solved explicitly for the price of the option. Note
also that we did not make any assumption so far about what type of option we talk
about.
184
9.2.6 Application - Deriving a budget constraint
Most maximization problems require a constraint. For a household, this is usually
the budget constraint. It is shown here how the structure of the budget constraint
depends on the economic environment the household …nds itself in and how the CVF
needs to be applied here.
Let wealth at time t be given by the number n (t) of stocks a household owns times
their price v (t), a (t) = n (t) v (t). Let the price follow a process that is exogenous to
the household (but potentially endogenous in general equilibrium),
where and are constants. For we require > 1 to avoid that the price can
become zero or negative. Hence, the price grows with the continuous rate and at
discrete random times it jumps by percent. The random times are modeled by the
jump times of a Poisson process q (t) with arrival rate , which is the “probability”
that in the current period a price jump occurs. The expected (or average) growth
rate is then given by + (see ch. 9.4.4).
Let the household earn dividend payments, (t) per unit of asset it owns, and
labour income, w (t). Assume furthermore that it spends p (t) c (t) on consumption,
where c (t) denotes the consumption quantity and p (t) the price of one unit of the
consumption good. When buying stocks is the only way of saving, the number of
stocks held by the household changes in a deterministic way according to
When savings n (t) (t) + w (t) p (t) c (t) are positive, the number of stocks held by
the household increases by savings divided by the price of one stock. When savings
are negative, the number of stocks decreases.
The change of the household’s wealth, i.e. the household’s budget constraint, is
then given by applying the CVF to a (t) = n (t) v (t). The appropriate CVF comes
from (9.19) where only one of the two di¤erential equations shows the increment of
the Poisson process explicitly. With F (x; y) = xy, we obtain
This is a very intuitive budget constraint: As long as the asset price does not jump,
i.e., dq (t) = 0, the household’s wealth increases by current savings, r (t) a (t) + w (t)
185
p (t) c (t), where the interest rate, r (t), consists of dividend payments in terms of the
asset price plus the deterministic growth rate of the asset price. If a price jump occurs,
i.e., dq (t) = 1, wealth jumps, as the price, by percent, which is the stochastic part
of the overall interest-rate. Altogether, the average interest rate amounts to r (t) +
(see ch. 9.4.4).
186
To simplify an economic interpretation set Y0 = z (0) = 0: Output is then given by
Y (t) = (At + z (t)) K: This says that with a constant factor input K; output in t
is determined by a deterministic and a stochastic part. The deterministic part At
implies linear (i.e. not exponential as is usually assumed) growth, the stochastic part
z (t) implies deviations from the trend. As z (t) is Brownian motion, the sum of
the deterministic and stochastic part can become negative. This is an undesirable
property of this approach.
To see that Y (t) is in fact a solution of the above di¤erential-representation, just
apply Ito’s Lemma and recover (9.33).
where g and are constants. What does this alternative speci…cation imply?
Solving the SDE yields A (t) = A0 + gt + z (t) (which can again be checked by
applying Ito’s lemma). Output is therefore given by
Let us now assume that TFP follows geometric Brownian motion process,
where again g and are constants. Let output continue to be given by Y (t) = A (t) K:
The solution for TFP, provided an initial condition A (0) = A0 ; is given by
1
A (t) = A0 e(g )t+
2 z(t)
2 : (9.35)
At any point in time t; the TFP level depends on time t and the current level of the
stochastic process z (t) : This shows that TFP at each point t in time is random and
thereby unkown from the perspective of t = 0: Hence, a SDE and its solution describe
the deterministic evolution of a distribution over time. One could therefore plot a
picture of A (t) which in principle would look like the evolution of the distribution in
ch. 7.4.1.
Interestingly, and this is due to the geometric speci…cation in (9.34) and impor-
tant for representing technologies in general, TFP can not become negative. While
Brownian motion z (t) can take any value between minus and plus in…nity, the term
1 2
e(g 2 )t+ z(t) is always positive. With an AK speci…cation for output, output is
187
1 2
always positive, Y (t) = A0 e(g 2 )t+ z(t) K. In fact, it can be shown that output and
TFP are lognormally distributed. Hence, the speci…cation of TFP with geometric
Brownian motion provides an alternative to the di¤erential-representation in (9.33)
which avoids the possibility of negative output.
The level of TFP at some future point in time t is determined by a deterministic
part, g 12 2 t; and by a stochastic part, z (t) : Apparently, the stochastic nature
of TFP has an e¤ect on the deterministic term. The structure (the factor 1=2 and the
quadratic term 2 ) reminds of the role the stochastic disturbance plays in Ito’s lemma.
There as well (see e.g. (9.13)), the stochastic disturbance a¤ects the deterministic
component of the di¤erential. As we will see later, however, this does not a¤ect
expected growth. As we will see in (9.53), expected output grows at the rate g (and
is thereby independent of the variance parameter ).
One can verify that (9.35) is a solution to (9.34) by using Ito’s lemma. To do so, we
bring (9.34) into a form which allows us to apply the formulas which we got to know
in ch. 9.2.2. De…ne x (t) g 21 2 t + z (t) and A (t) F (x (t)) = A0 ex(t) : As a
consequence, the di¤erential for x (t) is a nice SDE, dx (t) = g 21 2 dt + dz (t) :
As this SDE is of the form as in (9.5), we can use Ito’s lemma from (9.13) and …nd
1 2 1 2
dA (t) = dF (x (t)) = Fx (x (t)) g + Fxx (x (t)) dt + Fx (x (t)) dz:
2 2
1 1
dA (t) = A0 ex(t) g 2
+ A0 ex(t) 2
dt + A0 ex(t) dz ,
2 2
1 2 1 2
dA (t) =A (t) = g + dt + dz = gdt + dz;
2 2
where the “i¤” reinserted A (t) = A0 ex(t) and divided by A (t) : As A (t) satis…es the
orignianl SDE (9.34), A (t) is a solution of (9.34).
Option pricing
Let us come back to the Black-Scholes formula for option pricing. The SDE
derived above in (9.30) describes the evolution of the price F (t; S (t)) ; where t is
time and S (t) the price of the underlying asset at t. We now look at a European
call option, i.e. an option which gives the right to buy an asset at some …xed point
in time T; the maturity date of the option. The …xed exercise or strike price of the
option, i.e. the price at which the asset can be bought is denoted by E:
Clearly, at any point in time t when the price of the asset is zero, the value of
the option is zero as well. This is the …rst boundary condition for our SDE (9.30).
When the option can be exercised at T and the price of the asset is S; the value of
the option is zero if the exercise price E exceeds the price of the asset and S E if
not. This is the second boundary condition.
188
In the latter case where S E > 0, the owner of the option would in fact buy the
asset.
Given these two boundary conditions, the SDE has the solution (originating from
physics, see Black and Scholes, 1973, p. 644)
r[T t]
F (t; S (t)) = S (t) (d1 ) Ee (d2 )
where
Z y
1 u2
(y) = p e 2 du;
2 1
r2
ln S(t)
E
+ r 2
(T t) p
d1 = p ; d2 = d1 T t:
T t
The expression F (t; S (t)) gives the price of an option at a point in time t T where
the price of the asset is S (t) : It is a function of the cumulative standard normal
distribution (y) : For any path of S (t) ; time up to maturity T a¤ects the option
price through d1 ; d2 and directly in the second term of the above di¤erence. More
interpretation is o¤ered by many …nance textbooks.
with boundary value x (0) = x0 : W i (t) is a Brownian motion. The correlation coe¢ -
cient of its increments with the increments of W j (t) is ij : Its solution is
0 1
Rt 1
B x0 + (s) fa (s) Bb (s)g ds C
x (t) = B
(t) @ 0 C (9.37)
Pm Rt
1
A
+ (s) bi (s) dW i (s)
i=1 0
where
Rt 1
P
m t
R
(A(s) 2 BB (s) )ds+ Bi (s)dW i (s)
(t) = e 0 i=1 0 ;
m m
BB (s) = i=1 j=1 ij Bi (s) Bj (s) ;
m m
Bb (s) = i=1 j=1 ij Bi (s) bi (s) :
Arnold (1974, ch. 8.4) provides the solution for independent Brownian motions.
This solution is an “educated guess”. To obtain some intuition for (9.37), we can
consider the case of certainty. For Bi (t) = bi (t) = 0; this is an linear ODE and the
solution should correspond to the results we know from ch. 4.3.2. For the general
case, we now show that (9.37) indeed satis…es (9.36).
189
Proof. In order to use Ito’s Lemma, de…ne
Z t Z t
1 Pm
y (t) A (s) BB (s) ds + Bi (s) dW i (s)
0 2 i=1 0
such that
(t) = ey(t) (9.38)
and de…ne
Zt Zt
1 P
m
1
z(t) x0 + (s) fa (s) Bb (s)g ds + (s) bi (s) dW i (s)
i=1
0 0
where
1 P
m
dy(t) = A (t) BB (t) dt + Bi (t) dW i (t)
2 i=1
1 P
m
1
dz(t) = (t) fa (t) Bb (t)g dt + (t) bi (t) dW i (t)
i=1
1 m i
dx(t) = x(t) A (t) BB (t) dt + i=1 Bi (t) dW (t)
2
m i
+ fa (t) Bb (t)g dt + i=1 bi (t) dW (t)
1 m m m m
+ x(t) i=1 j=1 ij Bi (t) Bj (t) dt + 2 i=1 j=1 ij Bi (t) bi (t) dt
2
190
Rearranging gives
Imagine that TFP follows a deterministic trend and occasionally makes a discrete
jump. This is capture by a geometric description as in (9.34) only that Brownian
motion is replaced by a Poisson process,
In contrast to the solution (9.35) for the Brownian motion case, uncertainty does not
a¤ect the deterministic part here. As before, TFP follows a deterministic growth
component and a stochastic component, [q (t) q (0)] ln (1 + ). The latter makes
future TFP uncertaint from the perspective of today.
The claim that A (t) is a solution can be proven by applying the appropriate CVF.
This will be done for the next, more general, example.
A budget constraint
191
De…ning (s) w (s) c (s) ; the backward solution of (9.45) with initial condition
a (0) = a0 reads Z t
y(t) y(s)
a (t) = e a0 + e (s) ds (9.46)
0
where y (t) is Z t
y (t) = r (u) du + [q (t) q (0)] ln (1 + ) : (9.47)
0
Note that the solution in (9.46) has the same structure as the solution to a determin-
istic version of the di¤erential equation (9.45) (which we would obtain for = 0).
Put di¤erently, the stochastic component in (9.45) only a¤ects the discount factor
y (t) : This is not surprising - in a way - as uncertainty is proportional to a (t) and the
factor can be seen as the stochastic component of the interest payments on a (t) :
We have just stated that (9.46) is a solution. This should therefore be verifyed.
To this end, de…ne Z t
y(s)
z(t) a0 + e (s)ds; (9.48)
0
and write the solution (9.46) as a(t) = ey(t) z(t) where from (9.47) and (9.48),
y(t)
dy(t) = r(t)dt + ln(1 + )dq(t); dz(t) = e (t)dt: (9.49)
We have thereby de…ned a function a (t) = F (y (t) ; z (t)) where the SDEs describing
the evolution of y (t) and z (t) are given in (9.49). The CVF (9.19) then says
da(t) = ey(t) z(t)r(t)dt + ey(t) e y(t) (t)dt + ey(t)+ln(1+ ) z(t) ey(t) z(t) dq
= fr (t) a (t) + (t)g dt + a (t) dq:
This is the original di¤erential equation. Hence, a (t) is a solution for (9.45).
In stochastic worlds, there is also a link between dynamic and intertemporal bud-
get constraints, just as in deterministic setups as e.g. in ch. 4.3.4. We …rst present
here a budget constraint for a …nite planning horizon and then generalize the result.
For the …nite horizon case, we can rewrite (9.46) as
Z t Z t
y(s) y(t)
e c (s) ds + e a (t) = a0 + e y(s) w (s) ds:
0 0
192
discounting takes place at the realized stochastic interest rate y (s) - no expectations
are formed.
In order to obtain an in…nite horizon intertemporal budget constraint, the solution
(9.46) should be written more generally - after replacing t by and 0 by t - as
Z
y( )
a ( ) = e at + ey( ) y(s) (w (s) c (s)) ds (9.50)
t
where y( ) is Z
y( ) = r (u) du + ln (1 + ) [q ( ) q (t)] : (9.51)
t
y( )
Multiplying (9.50) by e and rearranging gives
Z
y( )
a( )e + e y(s) (c (s) w (s))ds = a(t):
t
y( )
Letting go to in…nity, assuming a no-Ponzi game condition lim !1 a( )e =0
and rearranging again yields
Z 1 Z 1
y(s)
e c (s) ds = a(t) + e y(s) w (s) ds:
t t
A switch process
Here are two funny processes which have an interesting simple solution. Consider
the initial value problem,
dy = 2 (y y) dq; y 0 = y + x0 :
Its solution is y (t) = y + ( 1)q(t) x0 and y (t) oscillates now between x (t) oscillates
between y x0 and y + x0 :
This could be nicely used for models with wage uncertainty, where wages are
sometimes high and sometimes low. An example would be a matching model where
labour income is w (i.e. y + x0 ) when employed and b (i.e. y x0 ) when unemployed.
The di¤erence between labour income levels is then 2x0 . The drawback is that the
probability of …nding a job is identical to the probability of losing it.
193
9.4 Expectation values
9.4.1 The idea
What is the idea behind expectations of stochastic processes? When thinking of a
stochastic process X (t), either a simple one like Brownian motion or a Poisson process
or more complex ones described by SDEs, it is useful to keep in mind that one can
think of the value X ( ) the stochastic process takes at some future point in time as
a “normal” random variable. At each future point in time X ( ) is characterized by
some mean, some variance, by a range etc. This is illustrated in the following …gure.
The …gure shows four realizations of the stochastic process X (t) : The starting
point is today in t = 0 and the mean growth of this process is illustrated by the
dotted line. When we think of possible realizations of X ( ) for = :4 from the
perspective of t = 0; we can imagine a vertical line that cuts through possible paths
at = :4: With su¢ ciently many realizations of these paths, we would be able to
make inferences about the distributional properties of X (:4) : As the process used
for this simulation is a geometric Brownian motion process as in (9.34), we would
…nd that X (:4) is lognormally distributed as depicted above. Clearly, given that we
have a precise mathematical description of our process, we do not need to estimate
distributional properties for X ( ) ; as suggested by this …gure, but we can explicitly
compute them.
What this section therefore does is to provide tools that allow to determine proper-
ties of the distribution of a stochastic process for future points in time. Put di¤erently,
understanding a stochastic process means understanding the evolution of its distri-
butional properties. We will …rst start by looking at relatively straightforward ways
to compute means. Subsequently, we provide some martingale results which allow
us to then understand a more general approach to computing means and also higher
moments.
194
9.4.2 Simple results
Brownian motion I
Consider a Brownian motion process Z ( ) ; > 0: Let Z (0) = 0. From the
de…nition of Brownian motion in ch. 9.1.1, we know that Z ( ) is normally distributed
with mean 0 and variance 2 : Let us assume we are in t today and Z (t) 6= 0. What
is the mean and variance of Z ( ) for > t?
We construct a stochastic process Z ( ) Z (t) which at = t is equal to zero. Now
imagining that t is equal to zero, Z ( ) Z (t) is a Brownian motion process as de…ned
in ch. 9.1.1. Therefore, Et [Z ( ) Z (t)] = 0 and vart [Z ( ) Z (t)] = 2 [ t] :
This says that the increments of Brownian motion are normally distributed with mean
zero and variance 2 [ t] : Hence, the reply to our question is
2
Et Z ( ) = Z (t) ; vart Z ( ) = [ t] :
For our convention 1 which we follow here as stated at the beginning of ch. 9.1.2,
the variance would equal vart Z ( ) = t:
Brownian motion II
Consider again the geometric Brownian process dA (t) =A (t) = gdt + dz (t) de-
1 2
scribing the growth of TFP in (9.34). Given the solution A (t) = A0 e(g 2 )t+ z(t)
from (9.35), we would now like to know what the expected TFP level A (t) in t from
the perspective of t = 0 is. To this end, apply the expectations operator E0 and …nd
1 1
E0 A (t) = A0 E0 e(g )t+ = A0 e(g )t E e
2 z(t) 2
z(t)
2 2
0 ; (9.52)
1 2
where the second equality exploited the fact that e(g 2 )t is non-random. As z (t)
is standard normally distributed, z (t) ~N (0; t), z (t) is normally distributed with
N (0; 2 t) : As a consequence (compare ch. 7.3.3, especially eq. (7.6)), e z(t) is log-
1 2
normally distributed with mean e 2 t : Hence,
1
E0 A (t) = A0 e(g )t e 21
2 2t
2 = A0 egt : (9.53)
The expected level of TFP growing according to a geometric Brownian motion process
grows at the drift rate g of the stochastic process. The variance parameter does
not a¤ect expected growth.
Poisson processes I
The expected value in t of q ( ) ; where > t; and its variance are (cf. e.g. Ross,
1993, p. 210 and 241)
Et q ( ) = q (t) + [ t] ; >t (9.54)
vart q ( ) = [ t] :
The expected value Et q ( ) directly follows from the de…nition of a Poisson process
in ch. 9.1.1, see (9.1). There are simple generalizations of the Poisson process where
the variance di¤eres from the mean (see ch. 9.4.4).
195
Poisson processes II
Let q (t) be a Poisson process with arrival rate : What is the probability that it
jumped exactly once between t and ?
[ t] ( [ t])n
Given that by def. 12, e n!
is the probability that q jumped n times by
; i.e. q ( ) = n; this probability is given by P (q ( ) = q (t) + 1) = e [ t] [ t] :
Note that this is a function which is non-monotonic in time :
Following the same structure as with Brownian motion, we now look at the geo-
metric Poisson process describing TFP growth (9.43). The solution was given in
(9.44) which reads, slightly generalized with the perspective of t and initial condition
At ; A ( ) = At eg[ t]+[q( ) q(t)] ln(1+ ) : What is the expected TFP level in ?
196
Applying the expectation operator gives
where, as in (9.52), we split the exponential growth term into its deterministic and
stochastic part. To proceed, we nee the following
Lemma 9 (taken from Posch and Wälde, 2006) Assume that we are in t and form
expectations about future arrivals of the Poisson process. The expected value of ckq( ) ,
conditional on t where q (t) is known, is
k
Et (ckq( ) ) = ckq(t) e(c 1) ( t)
; > t; c; k = const:
Note that for integer k, these are the raw moments of cq( ) :
Proof. We can trivially rewrite ckq( ) = ckq(t) ck[q( ) q(t)] : At time t; we know
the realization of q(t) and therefore Et ckq( ) = ckq(t) Et ck[q( ) q(t)] : Computing this
expectation requires the probability that a Poisson process jumps n times between t
and : Formally,
( t)
kn e ( ( t))n 1 e (ck (
( t)
t))n
Et ck[q( ) q(t)]
= 1
n=0 c = n=0
n! n!
( t) (ck 1) ( t) k
(ck 1) ( t) 1 e (c ( t))n
=e n=0
n!
ck ( t) k
k 1 e (c ( t))n k
= e(c 1) ( t)
n=0 = e(c 1) ( t)
;
n!
n ck ( t) k n
where e n!( ) is the probability of q ( ) = n and 1 n=0
e (c ( t))
n!
= 1 denotes
the summation of the probability function over the whole support of the Poisson
distribution which was used in the last step. For a generalization of this lemma, see
ex. 7.
To apply this lemma to our case Et eq( ) ln(1+ ) ; we set c = e and k = ln (1 + ) :
ln(1+ ) 1) [
Hence, Et eq( ) ln(1+ ) = eln(1+ )q(t) e(e t]
= eq(t) ln(1+ ) e [ t] = eq(t) ln(1+ )+ [ t]
This is an intuitive result: The growth rate of the expected TFP level is driven both by
the deterministic growth component of the SDE (9.43) for TFP and by the stochastic
part. The growth rate of expected TFP is higher, the higher the determinist part,
the g; the more often the Poisson process jumps on average (a higher arrival rate )
and the higher the jump (the higher ).
This con…rms formally what was already visible in and informally discussed after
…g. 39 of ch. 9.1.2: A Poisson process as a source of uncertainty in a SDE implies
that average growth is not just determined by the deterministic part of the SDE (as
is the case when Brownian motion constitutes the disturbance term) but also by the
Poisson process itself. For a positive ; average growth is higher, for a negative ; it
is lower.
197
9.4.3 Martingales
Martingale is an impressive word for a simple concept. Here is a simpli…ed de…nition
which is su¢ cient for our purposes. More complete de…nitions (in a technical sense)
are in Øksendal (1998) or Ross (1996).
Et x ( ) = x (t) ; t: (9.56)
As the expected value of x (t) is x (t) ; Et x (t) = x (t) ; this can easily be rewritten
as Et [x ( ) x (t)] = 0: This is identical to saying that x (t) is a martingale if the
expected value of its increments between now t and somewhere at in the future
is zero. This de…nition and the de…nition of Brownian motion imply that Brownian
motion is a martingale.
In what follows, we will use the martingale properties of certain processes rela-
tively frequently, e.g. when computing expected growth rates. Here are now some
fundamental examples for martingales.
Brownian motion
First, look at Brownian motion where we have a central result useful for many
applications
R where expectations and other moments are computed. It states that
t
f (z (s) ; s) dz (s), where f (:) is some function and z (s) is Brownian motion, is a
martingale (Corollary 3.2.6, Oksendal, 1998, p. 33),
Z
Et f (z (s) ; s) dz (s) = 0: (9.57)
t
Poisson uncertainty
A similar fundamental result for Poisson processes exists. We will use in what
follows the martingale property of various expressions containing R Poisson uncer-
tainty.
R These expressions are identical to or special cases of t f (q (s) ; s) dq (s)
t
f (q (s) ; s) ds; of which Garcia and Griego (1994, theorem 5.3) have shown that
it is a martingale indeed,
Z Z
Et f (q (s) ; s) dq (s) f (q (s) ; s) ds = 0: (9.58)
t t
198
9.4.4 A more general approach to computing moments
When we want to understand moments of some stochastic process, we can proceed in
two ways. Either, a SDE is expressed in its integral version, expectations operators
are applied and the resulting deterministic di¤erential equation is solved. Or, the
SDE is solved directly and then expectation operators are applied. We already saw
examples for the second approach in ch. 9.4.2. We will now follow the …rst way as
this is generally the more ‡exible one. We …rst start with examples for Brownian
motion processes and then look at cases with Poisson uncertainty.
We will now return to our …rst SDE in (9.3), dx (t) = adt + bdz (t) ; and want to
understand why a is called the drift term and b2 the variance term. To this end, we
compute the mean and variance of x ( ) for > t:
We start by expressing (9.3) in its integral representation as in ch. 9.1.3. This
gives
Z Z
x ( ) x(t) = ads + bdz(s) = a [ t] + b [z( ) z (t)] : (9.59)
t t
where the second step used the fact that the expected increment of Brownian motion
is zero. As the expected value Et x( ) is a deterministic variable, we can compute
the usual time derivative and …nd how the expected value of x ( ), being today in t,
changes the further the point lies in the future, dEt x( )=d = a: This makes clear
why a is referred to as the drift rate of the random variable x (:) :
Let us now analyse the variance of x ( ) where we also start from (9.59). The
variance can from (7.4) be written as vart x( ) = Et x2 ( ) [Et x( )]2 : In contrast
to the term in (7.4) we need to condition the variance on t: If we computed the
variance of x ( ) from the perspective of some earlier t; the variance would di¤er -
as will become very clear from the expression we will see in a second. Applying the
expression from (7.4) also shows that we can look at any x ( ) as a “normal”random
variable: Whether x ( ) is described by a stochastic process or by some standard
description of a random variable, an x ( ) for any …x future point in time has some
distribution with corresponding moments. This allows us to use standard rules for
computing moments. Computing …rst Et x2 ( ) gives by inserting (9.59)
199
where the last equality used again that the expected increment of Brownian motion is
zero. As [Et x( )]2 = [x (t) + a [ t]]2 from (9.60), inserting (9.61) into the variance
expression gives vart x( ) = b2 Et [z( ) z (t)]2 :
Computing the mean of the second moment gives
vart x( ) = b2 [ t] :
This equation shows why b is the variance parameter for x (:) and why b2 is called
its variance rate. The expression also makes clear why it is so important to state the
current point in time, t in our case. If we were further in the past or is further in
the future, the variance would be larger.
Imagine an individual owns wealth a that is allocated to N assets such that shares
in wealth are given by i ai =a: The price of each asset follows a certain pricing rule,
say geometric Brownian motion, and let’s assume that total wealth of the household,
neglecting labour income and consumption expenditure, evolves according to
N
da (t) = a (t) rdt + i=1 i i dzi (t) ;
N
where r i=1 i ri : This is in fact the budget constraint (with w c = 0) which
will be derived and used in ch. 10.4 on capital asset pricing. Note that Brownian
motions zi are correlated, i.e. dzi dzj = ij dt as in (9.15). What is the expected return
and the variance of holding such a portfolio, taking i and interest rates and variance
parameters as given?
Using the same approach as in the previous example we …nd that the expected
return is simply given by r: This is due to the fact that the mean of the BMs zi (t)
are zero. The variance of a (t) is - coming soon -
Let us now compute the expected return of wealth when the evolution of wealth is
described by the budget constraint in (9.32), da (t) = fr (t) a (t) + w (t) p (t) c (t)g dt+
a (t) dq (t). When we want to do so, we …rst need to be precise about what we mean
by the expected return. We de…ne it as the growth rate of the mean of wealth when
consumption expenditure and labour income are identical, i.e. when total wealth
changes only due to capital income.
200
Using this de…nition, we …rst compute the expected wealth level at some future
point in time : Expressing this equation in its integral representation as in ch. 9.1.3
gives
Z Z
a ( ) a (t) = fr (s) a (s) + w (s) p (s) c (s)g ds + a (s) dq (s) :
t t
The second equality used an ingredient of the de…nition of the expected return, i.e.
w (s) = p (s) c (s) ; that r (s) is independent of a (s) and the martingale result of (9.58).
When we de…ne the mean of a (s) from the perspective of t by (s) Et a (s) ; this
equation reads
Z Z
( ) a (t) = Et r (s) (s) ds + (s) ds:
t t
201
Disentangling the mean and variance of a Poisson process
Consider the SDE dv (t) =v (t) = dt + dq (t) where the arrival rate of q (t) is
given by = : The mean of v ( ) is Et v ( ) = v (t) exp( + )( t) and hthereby indepen-
i
2 2
[ t]
dent of : The variance, however, is given by vart v1 ( ) = [Et v ( )] exp 1 .
A mean preserving spread can thus be achieved by an increase of : This increases
the randomly occurring jump and reduces the arrival rate = - the mean remains
unchanged, the variance increases.13
Let us now consider some stochastic process X (t) described by a di¤erential equa-
tion and ask what we know about expected values of X ( ) ; where lies in the future,
i.e. > t: We take as …rst example a stochastic process similar to (9.8). We take as
an example the speci…cation for total factor productivity A,
dA (t)
= dt + 1 dq1 (t) 2 dq2 (t) ; (9.63)
A (t)
where and i are positive constants and the arrival rates of the processes are given
by some constant i > 0. This equation says that TFP increases in a deterministic
way at the constant rate (note that the left hand side of this di¤erential equation
gives the growth rate of A (t)) and jumps at random points in time. Jumps can be
positive when dq1 = 1 and TFP increases by the factor 1 ; i.e. it increases by 1 %;
or negative when dq2 = 1; i.e. TFP decreases by 2 %:
The integral version of (9.63) reads (see ch. 9.1.3)
Z Z Z
A ( ) A (t) = A (s) ds + 1 A (s) dq1 (s) 2 A (s) dq2 (s)
t t t
Z Z Z
= A (s) ds + 1 A (s) dq1 (s) 2 A (s) dq2 (s) : (9.64)
t t t
202
where the second equality used the martingale result from ch. 9.4.3, i.e. the expression
in (9.58). Pulling the expectations operator into the integral gives14
Z Z Z
Et A ( ) A (t) = Et A (s) ds + 1 1 Et A (s) ds 2 2 Et A (s) ds:
t t t
which we can then di¤erentiate with respect to time by applying the rule for di¤er-
entiating integrals from (4.9),
m
_ 1 ( ) = m1 ( ) + 1 1 m1 ( ) 2 2 m1 ( )
=( + 1 1 2 2 ) m1 ( ) : (9.66)
We now immediately see that TFP does not increase in expected terms, more
precisely Et A ( ) = A (t) ; if + 1 1 = 2 2 : Economically speaking, if the increase
in TFP through the deterministic component and the stochastic component 1
is “destroyed” on average by the second stochastic component 2 ; TFP does not
increase.
There are many textbooks on di¤erential equations, Ito calculus, change of vari-
able formulas and related aspects in mathematics. One that is widely used is Øk-
sendal (1998). A more technical approach is presented in Protter (1995). A classic
mathematical reference is Gihman and Skorohod (1972). A special focus with a de-
tailed formal analysis of SDEs with Poisson processes can be found in Garcia and
Griego (1994). They also provide solutions of SDEs and the background for com-
puting moments of stochastic processes. Further solutions, applied to option pricing,
are provided by Das and Foresi (1996). The CVF for the combined Poisson-di¤usion
setup in lemma 8 is a special case of the expression in Sennewald (2006) which in
turn is based on Øksendal and Sulem (2005). Øksendal and Sulem present CVFs for
more general Levyprocesses of which the SDE (9.22) is a very simple special case.
203
A very readable introduction to stochastic processes is by Ross (1993, 1996). He
writes in the introduction to the …rst edition of his 1996 book that his text is “a
nonmeasure theoretic introduction to stochastic processes”. This makes this book
highly accessible for economists.
An introduction with many examples from economics is Dixit and Pindyck (1994).
See also Turnovsky (1997, 2000) for many applications. Brownian motion is treated
extensively in Chang (2004).
The CVF for Poisson processes is most easily accessible in Sennewald (2006) or
Sennewald and Wälde (2006). Sennewald (2006) provides the mathematical proofs,
Sennewald and Wälde (2006) has a focus on applications.
There is a tradition in economics starting with Eaton (1981) where output is rep-
resented by a stochastic di¤erential equation as presented in ch. 9.3.1. This and
similar representations of technologies are used by Epaulart and Pommeret (2003),
Pommeret and Smith (2005), Turnovsky and Smith (2006), Turnovsky (2000), Chat-
terjee, Giuliano and Turnovsky (2004), and many others. It is well known (see e.g.
footnote 4 in Grinols and Turnovsky (1998)) that this implies the possibility of nega-
tive Y . An alternative where standard Cobb-Douglas technologies are used and TFP
is described by a SDE to this approach is presented in Wälde (2005) or Nolte and
Wälde (2007).
The Poisson process is widely used in …nance (early references are Merton, 1969,
1971) and labour economics (in matching and search models, see e.g. Pissarides,
2000, Burdett and Mortenson, 1998 or Moscarini, 2005). See also the literature
on the real options approach to investment (McDonald and Siegel, 1986, Dixit and
Pindyck, 1994, Chen and Funke 2005 or Guo et al., 2005). It is also used in growth
models (e.g. quality ladder models à la Aghion and Howitt, 1992 or Grossman and
Helpman, 1991), in analyzes of business cycles in the natural volatility tradition (e.g.
Wälde, 2005), contract theory (e.g. Guriev and Kvasov, 2005), in the search approach
to monetary economics (e.g. Kiyotaki and Wright, 1993 and subsequent work) and
many other areas. Further examples include Toche (2005), Steger (2005), Laitner and
Stolyarov (2004), Farzin et al. (1998), Hassett and Metcalf (1999), Thompson and
Waldo (1994), Palokangas (2003, 2005) and Venegas-Martínez (2001).
One Poisson shock a¤ecting many variables (as stressed in lemma 7) was used by
Aghion and Howitt (1992) in their famous growth model. When deriving the budget
constraint in Appendix 1 of Wälde (1999a), it is taken into consideration that a jump
in the technology level a¤ects both the capital stock directly as well as its price. Other
examples include the natural volatility papers by Wälde (2002, 2005) and Posch and
Wälde (2006).
204
Exercises Chapter 9
Applied Intertemporal Optimization
Stochastic di¤erential equations and rules
for di¤erentiating
Let GDP in a small open economy (with internationally given constant prices
px and py ) be given by
205
the number of shares held by the household and vi is the value of the share.
Assume that the value of the share evolves according to
dvi = i vi dt + i vi dqi :
Assume further that the number of shares held by the household changes ac-
cording to
( 1 n1 + 2 n2 + w e)
dni = i dt;
vi
where i is the share of savings used for buying stock i:
5. Option pricing
Assume the price of an asset follows dS=S = dt + dz + dq (as in Merton,
1976), where z is Brownian motion and q is a Poisson process. This is a gen-
eralization of (9.24) where = 0: How does the di¤erential equation look like
that determines the price of an option on this asset?
6. Martingales
(a) The weather tommorow will be just the same as today. Is this a martin-
gale?
(b) Let z (s) be Brownian motion. Show that Y (t) de…ned by
Z t Z
1 t 2
Y (t) exp f (s) dz (s) f (s) ds (9.67)
0 2 0
is a martingale.
(c) Show that X (t) de…ned by
Z t Z t
1
X (t) exp f (s) dz (s) f 2 (s) ds
0 2 0
is also a martingale.
7. Expectations - Poisson
Assume that you are in t and form expectations about future arrivals of the
Poisson process q (t). Prove the following statement by using lemma 9: The
number of expected arrivals in the time interval [ ; s] equals the number of
expected arrivals in a time interval of the length s for any s > t,
k
Et (ck[q( ) q(s)]
) = E(ck[q( ) q(s)]
) = e(c 1) ( s)
; > s > t; c; k = const:
Hint: If you want to cheat, look at the appendix to Posch and Wälde (2006). It
is available e.g. at www.waelde.com/publications.
206
8. Expected values
Show that the growth rate of the mean of x (t) described by the geometric
Brownian motion
dx (t) = ax (t) dt + bx (t) dz (t)
is given by a:
(a) Do so by using the integral version of this SDE and compute the increase
of the expected value of x (t).
(b) Do the same as in (a) but solve …rst the SDE and compute expectations
by using this solution.
(c) Compute the covariance of x ( ) and x (s) for >s t:
9. Expected returns
Consider the budget constraint
(a) What is the expected return for wealth? Why does this expression di¤er
from (9.62)?
(b) What is the variance of wealth?
dx (t) = [a (t) x (t)] dt + B1 (t) x (t) dW1 (t) + b2 (t) dW2 (t)
207
13. Dynamic and intertemporal budget constraints - Brownian motion
Consider the dynamic budget constraint
da ( ) = (r ( ) a ( ) + w ( ) p ( ) c ( )) dt + a ( ) dz ( ) ;
(a) Show that the intertemporal version of this budget constraint, after making
a standard economic assumption, can be written as
Z 1 Z 1
'( )
e p ( ) c ( ) d = at + e '( ) w ( ) d
t t
(b) Now assume we are willing to assume that intertemporal budget constraints
need to hold in expectations only and not in realizations: we require only
that agents balance at each instant expected consumption expenditure to
current …nancial wealth plus expected labour income. Show that the in-
tertemporal budget constraint then simpli…es to
Z 1 R Z 1 R
r(s)ds
Et e t p ( ) c ( ) d = at + Et e t r(s)ds w ( ) d ;
t t
208
10 Dynamic programming
We now return to our main concern: How to solve maximization problems. We
…rst look at optimal behaviour under Poisson uncertainty where we analyse cases for
uncertain asset prices and labour income. We then switch to Brownian motion and
look at capital asset pricing as an application.
where the interest rate r (t) + (t) =v (t) was de…ned as the deterministic rate of
change of the price of the asset (compare the equation for the evolution of assets
in (9.31)) plus dividend payments (t) =v (t) : We treat the price p here as a constant
(see the exercises for an extension). Following the tilde notation from (9.7), we can
express wealth a~ (t) after a jump by
a
~ (t) = (1 + ) a (t) : (10.3)
The budget constraint of this individual re‡ects standard economic ideas about
budget constraints under uncertainty. As visible in the derivation in ch. 9.2.6, the
uncertainty for this households stems from uncertainty about the evolution of the
price (re‡ected in ) of the asset she saves in. No statement was made about the
evolution of the wage w (t) : Hence, we take w (t) here as parametric, i.e. if there are
stochastic changes, they all come as a surprise and are therefore not anticipated. The
household does take into consideration, however, the uncertainty resulting from the
209
evolution of the price v (t) of the asset. In addition to the deterministic growth rate
of v (t) ; v (t) changes in a stochastic way by jumping occasional by percent (again,
see (9.31)). The returns to wealth a (t) are therefore uncertain and are composed of
the “usual”r (t) a (t) term and the stochastic a (t) term.
The …rst tool we need to derive rules for optimal behavior is the Bellman equation.
De…ning the optimal program as V (a) maxfc( )g U (t) subject to a constraint, this
equation is given by (see Sennewald and Wälde, 2006, or Sennewald, 2006)
1
V (a (t)) = max u (c (t)) + Et dV (a (t)) : (10.4)
c(t) dt
The Bellman equation has this basic form for “most”maximization problems in con-
tinuous time. It can therefore be taken as the starting point for other maximization
problems as well, independently e.g. of whether uncertainty is driven by Poisson
processes, Brownian motion or Levy processes. We will see examples for related
problems later (see ch. 10.1.4, ch. 10.2.2 or ch. 10.3.2) and discuss then how to ad-
just certain features of this “general”Bellman equation. In this equation, the variable
a (t) represents the state variable, in our case wealth of the individual.
Given the general form of the Bellman equation in (10.4), we need to compute the
di¤erential dV (a (t)) : Given the evolution of a (t) in (10.2) and the CVF from (9.18),
we …nd
In contrast to the CVF notation in e.g. (9.18), we use here and in what follows
simple derivative signs like V 0 (a) as often as possible in contrast to e.g. Va (a) :
This is possible as long as the functions, like the value function V (a) here, have one
argument only. Forming expectations about dV (a (t)) is easy and they are given by
The …rst term, the “dt-term” is known in t: The current state a (t) and all prices
are known and the shadow price V 0 (a) is therefore also known. As a consequence,
210
expectations need to be applied only to the “dq-term”. The …rst part of the “dq-term”,
the expression V ((1 + ) a) V (a) is also known in t as again a (t), parameters and
the function V are all non-stochastic. We therefore only have to compute expectations
about dq: From (9.54), we know that Et [q ( ) q (t)] = [ t] : Now replace q ( )
q (t) by dq and t by dt and …nd Et dq = dt: The Bellman equation therefore reads
Note that forming expectations the way just used is somewhat “‡oppy”. Doing it
in the more stringent way introduced in ch. 9.4.4 would, however, lead to identical
results.
The …rst-order condition is
As always, (current) utility from an additional unit of consumption u0 (c) must equal
(future) utility from an additional unit of wealth V 0 (a), multiplied by the price p of
the consumption good, i.e. by the number of units of wealth for which one can buy
one unit of the consumption good.
In order to understand the evolution of the marginal value V 0 (a) of the optimal
program, i.e. the evolution of the costate variable, we need to (i) compute the partial
derivative of the maximized Bellman equation with respect to assets and (ii) compute
the di¤erential dV 0 (a) by using the CVF and insert the partial derivative into this
expression. These two steps corresponds to the two steps in DP2 in the deterministic
continuous time setup of ch. 6.1.
(i) In the …rst step, we state the maximized Bellman equation from (10.5) as the
Bellman equation where controls are replaced by their optimal values,
We then compute again the derivative with respect to the state - as in discrete time
and deterministic continuous time setups - as this gives us an expression for the
shadow price V 0 (a). In contrast to the previous emphasis on Ito’s Lemmas and CVFs,
we can use for this step standard rules from algebra as we compute the derivative for
a given state a - the state variable is held constant and we want to understand the
derivative of the function V (a) with respect to a: We do not compute the di¤erential
of V (a) and ask how the value function changes as a function of a change in a.
Therefore, using the envelope theorem,
211
(ii) In the second step, the di¤erential of the shadow price V 0 (a) is computed,
given again the evolution of a (t) in (10.2),
Finally, replacing V 00 (a) [ra + w pc] in (10.8) by the same expression from (10.7)
gives
dV 0 (a) = f( r) V 0 (a) [V 0 (~
a) [1 + ] V 0 (a)]g dt + fV 0 (~
a) V 0 (a)g dq:
Finally, we can replace the marginal value by marginal utility from the …rst-
order condition (10.6). In this step, we employ that p is constant and therefore
dV 0 (a) = p 1 du0 (c). Hence, the Keynes-Ramsey rule describing the evolution of
marginal utility reads
Note that the constant price p dropped out. This rule shows how marginal utility
changes in a deterministic and stochastic way.
If we want to know more about the evolution of consumption, we can use the CVF
formula as follows. Let f (:) be the inverse function for u0 ; i.e. f (u0 (c)) = c; and
apply the CVF to f (u0 (c)) : This gives
212
‡uctuations in a natural volatility setup (see e.g. Matsuyama, 1999, Francois and
Lloyd-Ellis, 2003 or Wälde, 1999, 2005). It corresponds to its deterministic pendant
in (5.30) in ch. 5.5.3: By setting = 0 here (implying dq = 0), noting that we treated
the price as a constant and dividing by dt, we obtain (5.30).
Let us now assume for the instantaneous utility function the widely used constant
relative risk aversion (CRRA) instantaneous utility function,
c ( )1 1
u (c ( )) = ; > 0: (10.11)
1
Then, the Keynes-Ramsey rule becomes
dc n h c io c~
= r + (1 + ) 1 dt + 1 dq: (10.12)
c c~ c
The left-hand side gives the proportional change of consumption times ; the inverse
1
of the intertemporal elasticity of substitution : This corresponds to c=c
_ in the
deterministic Keynes-Ramsey rule in e.g. (5.31). Growth of consumption depends
on the right-hand side in a deterministic way on the usual di¤erence between the
interest rate and time preference rate plus the “ term”which captures the impact
of uncertainty. When we want to understand the meaning of this term, we need to
…nd out whether consumption jumps up or down, following a jump of the Poisson
process. When is positive, the household holds more wealth and consumption
should increase. Hence, the ratio c=~ c is smaller than unity and the sign of the bracket
term (1 + ) cc~ 1 is qualitatively unclear. If it is positive, consumption growth is
faster in a world where wealth occasionally jumps by percent.
The dq-term gives discrete changes in the case of a jump in q: It is tautological,
however: When q jumps and dq = 1 and dt = 0 for this small instant of the jump,
(10.12) says that dc=c on the left-hand side is given by f~ c=c 1g on the right hand
side. As the left hand side is by de…nition of dc given by [~ c c] =c; both sides are
identical. Hence, the level of c~ after the jump needs to be determined in an alternative
way.
Consider a household that is endowed with some initial wealth a0 > 0. At each
instant, the household can invest its wealth a (t) in both a risky and a safe asset. The
213
share of wealth the household holds in the risky asset is denoted by (t). The price
v1 (t) of one unit of the risky asset obeys the SDE
where r1 2 R and > 0. That is, the price of the risky asset grows at each instant with
a …xed rate r1 and at random points in time it jumps by percent. The randomness
comes from the well-known Poisson process q (t) with arrival rate . The price v2 (t)
of one unit of the safe asset is assumed to follow
where r2 0. Let the household receive a …xed wage income w and spend c (t) 0
on consumption. Then, in analogy to ch. 9.2.6, the household’s budget constraint
reads
We allow wealth to become negative but we could assume that debt is always covered
by the household’s lifetime labour income discounted with the safe interest rate r2 ,
i.e. a (t) > w=r2 .
Let intertemporal preferences of households be identical to the previous maximiza-
tion problem - see (10.1). The instantaneous utility function is again characterized
by CRRA as in (10.11), u(c( )) = (c1 1) = (1 ) : The control variables of the
household are the nonnegative consumption stream fc (t)g and the share f (t)g held
in the risky asset. To avoid a trivial investment problem, we assume
That is, the guaranteed return of the risky asset, r1 , is lower than the return of the
riskless asset, r2 , whereas, on the other hand, the expected return of the risky asset,
r1 + , shall be greater than r2 . If r1 was larger than r2 , the risky asset would
dominate the riskless one and noone would want to hold positive amounts of the
riskless asset. If r2 exceeded r1 + ; the riskless asset would dominate.
Again, the …rst step of the solution of this maximization problem requires a Bell-
man equation. De…ne the value function V again as15 V (a (t)) maxfc( )g U (t) :
The basic Bellman equation is taken from (10.4). When computing the di¤eren-
tial dV (a (t)) and taking into account that there are now two control variables, the
Bellman equation reads
214
where a~ (1 + ) a denotes the post-jump wealth if at wealth a a jump in the risky
asset price occurs. The …rst-order conditions which any optimal paths must satisfy
are given by
u0 (c) = V 0 (a) (10.18)
and
V 0 (a) (r1 r2 ) a + V 0 (~
a) a = 0: (10.19)
While the …rst …rst-order condition equates as always marginal utility with the shadow
price, the second …rst-order condition determines optimal investment of wealth into
assets 1 and 2; i.e. the optimal share : The latter …rst-order condition contains a
deterministic and a stochastic term and households hold their optimal share if these
two components just add up to zero. Assume, consistent with (10.16), that r1 < r2 : If
we were in a deterministic world, i.e. = 0; households would then only hold asset 2
as its return is higher. In a stochastic world, the lower instantaneous return on asset
1 is compensated by the fact that, as (10.13) shows, the price of this asset jumps up
occasionally by the percentage : Lower instantaneous returns r1 paid at each instant
are therefore compensated for by large occasional positive jumps.
As this …rst-order condition also shows, returns and jumps per se do not matter:
The di¤erence r1 r2 in returns is multiplied by the shadow price V 0 (a) of capital and
the e¤ect of the jump size times its frequence, ; is multiplied by the shadow price
0
V (~a) of capital after a jump. What matters for the household decision is therefore
the impact of holding wealth in one or the other asset on the overall value from
behaving optimally, i.e. on the value function V (a). The channels through which
asset returns a¤ect the value function is …rst the impact on wealth and second the
impact of wealth on the value function i.e. the shadow price of wealth.
We can now immediately see why this more complex maximization problem yields
simpler solutions: Replacing in equation (10.19) V 0 (a) with u0 (c) according to (10.18)
yields for a 6= 0
u0 (~
c) r2 r 1
0
= ; (10.20)
u (c)
where c~ denotes the optimal consumption choice for a
~. Hence, the ratio for optimal
consumption after and before a jump is constant. If we assume e.g. a CES utility
function as in (10.11), this jump is given by
1=
c~
= : (10.21)
c r2 r1
No such result on relative consumption before and after the jump is available for the
maximization problem without a choice between a risky and a riskless asset.
Since by assumption (10.16) the term on the right-hand side is greater than 1, this
equation shows that consumption jumps upwards if a jump in the risky asset price
occurs. This result is not surprising since, if the risky asset price jumps upwards, so
does the household’s wealth.
215
In the next step, we compute the evolution of V 0 (a (t)), the shadow price of wealth.
Assume that V is twice continuously di¤erentiable. Then, due to budget constraint
(10.15), the CVF from (9.18) yields
dV 0 (a) = f[ r1 + (1 ) r2 ] a + w cg V 00 (a) dt
0 0
+ fV (~a) V (a)g dq (t) : (10.22)
Di¤erentiating the maximized Bellman equation yields under application of the en-
velope theorem
V 0 (a) = f[ r1 + (1 ) r2 ] a + w cg V 00 (a)
+ f r1 + [1 ] r2 g V 0 (a) + fV 0 (~
a) [1 + ] V 0 (a)g :
Rearranging gives
f[ r1 + (1 ) r2 ] a + w cg V 00 (a)
=f [ r1 + (1 ) r2 ]g V 0 (a) fV 0 (~
a) [1 + ] V 0 (a)g :
f [ r1 + (1 ) r2 ]g V 0 (a)
dV 0 (a) = 0 dt + fV 0 (~
a) V 0 (a)g dq (t) :
f[1 + ] V (~ a) V 0 (a)g
Replacing V 0 (a) by u0 (c) following the …rst-order condition (10.18) for optimal
consumption, we obtain
f [ r1 + (1 ) r2 ]g u0 (c)
du0 (c) = 0 dt + fu0 (~
c) u0 (c)g dq (t) :
f[1 + ] u (~ c) u0 (c)g
Now applying the CVF again to f (x) = (u0 ) 1 (x) and using (10.20) leads to the
Keynes-Ramsey rule for general utility functions u,
216
For the CRRA utility function as in (10.11), the elimination of c~ becomes even
simpler and we obtain with (10.21)
( )
1=
dc (t) 1 r2 r1
= r2 1 dt + 1 dq (t) :
c (t) r 2 r1
The optimal change in consumption can thus be expressed in terms of well-known pa-
rameters. As long as the price hof the risky asset does not
i jump, optimal consumption
r2 r1
grows constantly by the rate r2 1 = . The higher the risk-free
interest rate, r2 , and the lower the guaranteed interest rate of the risky asset, r1 , the
discrete growth rate, , the probability of a price jump, , the time preference rate,
, and the risk aversion parameter, , the higher becomes the consumption growth
rate. If the risky asset price jumps, consumption jumps as well to its new higher level
c (t) = [( ) = (r2 r1 )]1= c (t). Here the growth rate depends positively on , , and
r1 , whereas r2 and have negative in‡uence.
s
dq = 1 c(a)
N
e
c
s
c
N
a0 a1 a2 a
217
Figure 42 Consumption c as a function of wealth a
218
10.1.6 Expected growth
Let us now try to understand the impact of uncertainty on expected growth. In order
to compute expected growth of consumption from realized growth rates (10.12), we
rewrite this equation as
c (t)
dc (t) = r (t) + (1 + ) 1 c (t) dt + f~
c (t) c (t)g dq:
c~ (t)
Expressing it in its integral version as in (9.64), we obtain for > t;
Z
c (s)
[c ( ) c (t)] = r (s) + (1 + ) 1 c (s) ds
t c~ (s)
Z
+ f~
c (s) c (s)g dq (s) :
t
Using again the martingale resultR from ch. 9.4.3 as already in R(9.65), i.e. the ex-
pression in (9.58), we replace Et t f~ c (s) c (s)g dq (s) by Et t f~
c (s) c (s)g ds;
i.e.
Z
1 c (s)
Et c ( ) c (t) = Et r (s) + (1 + ) 1 c (s) ds
t c~ (s)
Z
+ Et f~
c (s) c (s)g ds:
t
219
10.2 Matching on the labour market: where value functions
come from
Value functions are widely used in matching models. Examples are unemployment
with frictions models of the Mortensen-Pissarides type or shirking models of unem-
ployment (Shapiro and Stiglitz, 1984). These value functions can be understood very
easily on an intuitive level, but they really come from a maximization problem of
households. In order to understand when value functions as the ones used in the just
mentioned examples can be used (e.g. under the assumption of no saving, or being
in a steady state), we now derive value functions in a general way and then derive
special cases used in the literature.
10.2.1 A household
Let wealth a of a household evolve according to
da = fra + z cg dt:
Wealth increases per unit of time dt by the amount da which depends on current
savings ra + z c: Labour income is denoted by z which includes income w when
employed and unemployment bene…ts b when unemployed, z = w; b: Labour income
follows a stochastic Poisson di¤erential equation as there is job creation and job
destruction. In addition, we assume technological progress that implies a constant
growth rate g of labour income. Hence, we can write
dz = gzdt dqw + dqb ;
where w b: Job destruction takes place at an exogenous state-dependent
arrival rate s (z) : The corresponding Poisson process counts how often our household
moved from employment into unemployment is qw : Job creation takes place at an
exogenous rate (z) which is related to the matching function presented in (5.42).
The Poisson process related to the matching process is denoted by qb : It counts how
often a household leaves her “b-status”, i.e. how often she found a job. As an
individual cannot loose her job when she does not have one and as …nding a job
makes (in this setup) no sense for someone who has a job, both arrival rates are state
dependent. As an example, when an individual is employed, (w) = 0; when she is
unemployed, s (b) = 0:
z w b
(z) 0
s (z) s 0
R1 [ t]
Let the individual maximize expected utility Et t e u (c ( )) d ; where in-
1
stantaneous utility is of the CES type, u (c) = c 1 1 with > 0:
220
10.2.2 The Bellman equation and value functions
The state space is described by a and z: The Bellman equation has the same structure
as in (10.4). The adjustments that need to be made here follow from the fact that we
have two state variables instead of one. Hence, the basic structure from (10.4) adopted
to our problem reads V (a; z) = maxfc(t)g u (c) + dt1 Et dV (a; z) : The change of
V (a; z) is, given the evolution of a and z from above and the CVF from (9.21),
dV (a; z) = fVa [ra + z c] + Vz gzg dt
+ fV (a; z ) V (a; z)g dqw + fV (a; z + ) V (a; z)g dqb :
Forming expectations, remembering that Et dqw = s (z) dt and Et dqb = (z) dt; and
“dividing”by dt gives the Bellman equation
u (c) + Va [ra + z c] + Vz gz
V (a; z) = max :
c +s (z) [V (a; z ) V (a; z)] + (z) [V (a; z + )
V (a; z)]
(10.23)
The value functions in the literature are all special cases of this general Bellman
equation.
Denote by U V (a; b) the expected present value of (optimal behaviour of a
worker) being unemployed (as in Pissarides, 2000, ch. 1.3) and by W V (a; w) the
expected present value of being employed. As the probability of loosing a job for an
unemployed worker is zero, s (b) = 0; and (b) = ; the Bellman equation (10.23)
reads
U = max fu (c) + Ua [ra + b c] + Uz gz + [W U ]g ;
c
where we used that W = V (a; b + ) : When we assume that agents behave optimally,
i.e. we replace control variables by their optimal values, we obtain the maximized
Bellman equation,
U = u (c) + Ua [ra + b c] + Uz gb + [W U] :
When we now assume that households can not save, i.e. c = ra + b; and that there
is no technological progress, g = 0; we obtain
U = u (ra + b) + [W U] :
Assuming further that households are risk-neutral, i.e. u (c) = c; and that they have
no capital income, i.e. a = 0, consumption is identical to unemployment bene…ts
c = b: If the interest rate equals the time preference rate, we obtain eq. (1.10) in
Pissarides (2000),
rU = b + [W U ] :
221
The notation is as always, uncertainty stems from Brownian R 1motion z: The individual
maximizes a utility function as given in (10.1), U (t) = Et t e [ t] u (c ( )) d . The
value function is de…ned by V (a) = maxfc( )g U (t) : We again follow the three step
scheme for dynamic programming.
The Bellman equation is given for Brownian motion by (10.4) as well. When a
maximization problem other than one where (10.4) is suitable is to be formulated and
solved, …ve adjustments are in principle possible for the Bellman equation. First, the
discount factor might be given by some other factor - e.g. the interest rate r when
the present value of some …rm is maximized. Second, the number of arguments of
the value function needs to be adjusted to the number of state variables. Third, the
number of control variables depends on the problem that is to be solved and, fourth,
the instantaneous utility function is replaced by what is found in the objective function
- which might be e.g. instantaneous pro…ts. Finally and obviously, the di¤erential
dV (:) needs to be computed according to the rules that are appropriate for the
stochastic processes which drive the state variables.
As the di¤erential of the value function, following Ito’s Lemma in (9.13), is given
by
1
dV (a) = V 0 (a) [ra + w pc] + V 00 (a) 2 a2 dt + V 0 (a) adz;
2
forming expectations Et and dividing by dt yields the Bellman equation for our speci…c
problem
1 2 2
V (a) = max u (c (t)) + V 0 (a) [ra + w pc] + V 00 (a) a
c(t) 2
(i) The derivative of the maximized Bellman equation with respect to the state
variable is (using the envelope theorem)
1
V 0 = V 00 [ra + w pc] + V 0 r + V 000 2 a2 + V 00 2
a,
2
1 000 2 2
( r) V 0 = V 00 [ra + w pc] + V a + V 00 2 a: (10.25)
2
Not surprisingly, given that the Bellman equation already contains the second deriva-
tive of the value function, the derivative of the maximized Bellman equation contains
its third derivative V 000 .
222
(ii) Computing the di¤erential of the shadow price of wealth V 0 (a) gives, using
Ito’s Lemma (9.13),
1 2 2
dV 0 = V 00 da + V 000 a dt
2
1
pc] dt + V 000 2 a2 dt + V 00 adz;
= V 00 [ra + w
2
and inserting into the partial derivative (10.25) of the maximized Bellman equation
yields
2
dV 0 = ( r) V 0 dt V 00 adt + V 00 adz
2
= ( r) V 0 V 00 a dt + V 00 adz: (10.26)
As always at the end of DP2, we have a di¤erential equation (or di¤erence equation in
discrete time) which determines the evolution of V 0 (a) ; the shadow price of wealth.
DP3: Inserting …rst-order conditions
Assuming that the evolution of aggregate prices is independent of the evolution of
the marginal value of wealth, we can write the …rst-order condition for consumption
in (10.24) as du0 (c) = pdV 0 + V 0 dp: This follows e.g. from Ito’s Lemma (9.16) with
0
pV 0 = 0: Using (10.26) to replace dV ; we obtain
223
With a CRRA utility function, we can replace the …rst, second and third derivative
of u (c) and …nd the corresponding rule
dc ca
1 hc i2
2 a
= r a+ [ + 1] a dt (10.29)
c c
2 c
ca dp
+ adz + :
c p
where the price of an asset is denoted by pi and the number of shares held by ni . The
total number of assets is given by N . Let us assume that the price of an asset follows
geometric Brownian motion,
dpi
= i dt + i dzi ; (10.30)
pi
where each price is driven by its own drift parameter i and its own variance para-
meter i : Uncertainty results from Brownian motion zi which is also speci…c for each
asset. These parameters are exogenously given to the household but would in a gen-
eral equilibrium setting be determined by properties of e.g. technologies, preferences
and other parameters of the economy.
Households can buy or sell assets by using a share i of their savings,
1 N
dni = i i=1 i ni +w c dt: (10.31)
pi
When savings are positive and a share i is used for asset i; the number of stocks held
in i increases. When savings are negative and i is positive, the number of stocks i
decreases.
The change in the households wealth is given by da = N i=1 d (pi ni ) : The wealth
held in one asset changes according to
N
d (pi ni ) = pi dni + ni dpi = i i=1 i ni +w c dt + ni dpi :
224
The …rst equality uses Ito’s Lemma from (9.16), taking into account that second
derivatives of F (:) = pi ni are zero and that dni in (10.31) is deterministic and there-
fore dpi dni = 0: Using the pricing rule (10.30) and the fact that shares add to unity,
N
i=1 i = 1, the budget constraint of a household therefore reads
N N
da = i=1 i ni +w c dt + i=1 ni pi [ i dt + i dzi ]
N i N
= i=1 ni pi + ni pi i +w c dt + i=1 ni pi i dzi
pi
N i N
= i=1 ai + i +w c dt + i=1 ai i dzi :
pi
Now de…ne i as always as the share of wealth held in asset i; i ai =a: Then, by
de…nition, a = N i=1 i a and shares add up to unity,
N
i=1 i = 1: We rewrite this for
later purposes as
N 1
N = 1 i=1 i (10.32)
De…ne further the interest rate for asset i and the interest rate of the market portfolio
by
i N
ri + i; r i=1 i ri : (10.33)
pi
This gives us the budget constraint,
N N
da = a i=1 i ri +w c dt + a i=1 i i dzi
N
= fra + w cg dt + a i=1 i i dzi : (10.34)
The Bellman equation is given by (10.4), i.e. V (a) = maxc(t); i (t) u (c (t)) + dt1 Et dV (a) .
Hence, we need again the expected change of the value of one unit of wealth. With
one state variable, we simply apply Ito’s Lemma from (9.11) and …nd
1 1 1
Et dV (a) = Et V 0 (a) da + V 00 (a) [da]2 : (10.35)
dt dt 2
In a …rst step required to obtain the explicit version of the Bellman equation,
we compute the square of da: It is given, taking (9.12) into account, by [da]2 =
2
a2 N i=1 i i dzi : When we compute the square of the sum, the expression for the
product of Brownian motions in (9.15) becomes important as correlation coe¢ cients
225
need to be taken into consideration. Denoting the covariances by ij i j ij , we
get
2 2 2 2
= a2 1 1 dt + 1 1 2 2 12 dt + ::: + 2 2 1 1 12 dt + 2 2 dt + ::: + :::
2 N N
=a j=1 i=1 i j ij dt: (10.36)
The second preliminary step for obtaining the Bellman equation uses (10.32) again
and expresses the interest rate from (10.33) as a sum of the interest rate of asset N
(which could but does not need to be a riskless asset) and weighted excess returns
ri rN ;
N 1 N 1 N 1
r= i=1 i ri + 1 i=1 i rN = rN + i=1 i [ri rN ] : (10.38)
The Bellman equation with (10.37) and (10.38) now …nally reads
1
V (a) = max u (c) + V 0 (a) (rN + N 1
i=1 i [ri rN ])a + w c + V 00 (a) [da]2 ;
c(t); i (t) 2
First-order conditions
u0 (c) = V 0 (a) ;
and the …rst-order condition for assets. The …rst-order condition for consumption has
the well-known form.
226
To compute …rst-order conditions for shares i, we compute d f:g =d i for (10.37),
d f:g
= N 1
j=1 j ( ij 2 iN ) + N 1
i=1 i ( ij 2 iN ) +2 iN 2 1 N 1
j=1 i
2
N
d i
N 1 N 1 2
= 2 j=1 j( ij 2 iN ) + iN 1 j=1 i N
N 1 N 1 2
=2 j=1 j ( ij iN ) + 1 j=1 i iN N
N
=2 j=1 i [ ij iN ] : (10.39)
Hence, the derivative of the Bellman equation with respect to i is with (10.38) and
(10.39)
1
V 0 a[ri rN ] + V 00 2a2 N
j=1 i [ ij iN ] = 0 ()
2
V 00 N
ri rN = a j=1 i [ ij iN ] : (10.40)
V0
The interpretation of this optimality rule should take into account that we assumed
that an interior solution exists. This condition therefore says that agents are indi¤er-
ent between the current portfolio and a marginal increase in a share i if the di¤erence
in instantaneous returns, ri rN , is compensated by the covariances of of assets i and
N: Remember that from (10.33), instantaneous returns are certain at each instant.
ri r= iM : (10.41)
2
The ratio iM = M is what is usually called the -factor.
227
10.5 Natural volatility II (*)
Before this book comes to an end, the discussion of natural volatility models in ch.
7.8 is completed in this section. We will present a simpli…ed version of those models
that appear in the literature which are presented in stochastic continuous time setups.
The usefulness of Poisson processes will become clear here. Again, more background
is available at www.wifak.uni-wuerzburg.de/vwl2/nv.
Technologies
Y = AK L1 : (10.42)
where g and b are positive constants and is the exogenous arrival rate of the Poisson
process q. We know from (9.66) in ch. 9.4.4 that the growth rate of the expected
value of A is given by + :
The …nal good can be used for consumption and investment, Y = C + I; which
implies that the prices of all these goods are identical. Choosing Y as the numeraire
good, the price is one for all these goods. Investment increases the stock of production
units K if investment is larger than depreciation, captured by a constant depreciation
rate ;
dK = (Y C K) dt: (10.44)
There are …rms who maximize instantaneous pro…ts. They do not bear any risk
and pay factors r and w; marginal productivities of capital and labour.
Households
228
Wealth of households consists in shares in …rms which are denoted by k: This wealth
changes in a deterministic way (we do not derive it here but it could be done following
the steps in ch. 9.2.6), despite the presence of TFP uncertainty. This is due to two
facts: First, wealth is measured in units of physical capital, i.e. summing k over all
households gives K: As the price of one unit of K equals the price of one unit of the
output good and the latter was chosen as numeraire, the price of one unit of wealth is
non-stochastic. This di¤ers from (9.31) where the price jumps when q jumps. Second,
a jump in q does not a¤ect k directly. This could be the case when new technologies
make part of the old capital stock obsolete. Hence, the constraint of housholds is a
budget constraint which reads
dk = (rk + w c) dt: (10.46)
Optimal behaviour
When computing optimal consumption levels, households take the capital stock
k and the TFP level A as their state variables into account. This setup is there-
fore similar to the deterministic two-state maximization problem in ch. 6.3. Going
through similar steps (concerning e.g. the substituting out of cross derivatives in step
DP2) and taking the speci…c aspects of this stochastic framework into account, yields
optimal consumption following (see exercise 8)
u00 (c) u0 (~
c) u00 (c)
dc = r + 1 dt f~
c cg dq: (10.47)
u0 (c) 0
u (c) u0 (c)
Despite the deterministic constraint (10.46) and due to TFP jumps in (10.43), con-
sumption jumps as well: a dq term shows up in this expression and marginal utility
levels before (u0 (c)) and after the jump (~
uc ), using the notation from (9.7) appear
as well. The reason is straightforward: Whenever there is a discrete increase in the
TFP level, the interest rate and wages jump. Hence, returns for savings or households
change and the household adjusts her consumption level. This is in principle iden-
tical to the behaviour in the deterministic case as illustrated in …g. 27 in ch. 5.5.3.
(Undertaking this here for this stochastic case would be very useful.)
General equilibrium
We are now in the position to start thinking about the evolution of the economy
as a whole. It is described by a system in three equations. TFP follows (10.43). The
capital stock follows
dK = AK L1 C K dt
from (10.42) and (10.44). Aggregate consumption follows
dC C C
= r 1 dt + 1 dq
C C~ C~
from (10.47) and (10.45) and from aggregating over households. This system looks
fairly familiar from deterministic models, the only substantial di¤erence lies in the dq
term and the post-jump consumption levels C: ~
229
The simplest way to get an intuition about how the economy evolves consists in
looking at an example, i.e. by looking at a solution of the above system that holds
for a certain parameter set. We choose as example the parameter set for which the
saving rate is constant, C = (1 s) Y:
t.b.c.
... , we can draw a phase diagram and illustrate cyclical growth.
10.5.3 Summary
There are many, and in most cases much more technical, presentations of dynamic
programming in continuous time under uncertainty. A classic mathematical reference
is Gihman and Skorohod (1972(Gihman and Skorohod 1972)) and a widely used
mathematical textbook is Øksendal (1998); see also Protter (1995). These books are
probably useful only for those wishing to work on the theory of optimization and
not on applications of optimization methods. Optimization with unbounded utility
functions by dynamic programming was studied by Sennewald (2006).
With SDEs we need boundary conditions as well. In the in…nite horizon case, we
would need a transversality condition (TVC). See Smith (1996) for a discussion of
a TVC in a setup with Epstein-Zinn preferences. Sennewald (2006) has TVCs for
Poisson uncertainty.
Applications
Books in …nance that use dynamic programming methods include Du¢ e (1988,
2001) and Björk (2004). Stochastic optimization for Brownian motion is also covered
nicely by Chang (2004).
A maximization problem of the type presented in ch. 10.1 was …rst analyzed in
Wälde (1999, 2005b). This chapter combines these two papers. These two papers
were also jointly used in the Keynes-Ramsey rule appendix to Wälde (2005). It is
also used in Posch and Wälde (2006), Sennewald and Wälde (2006) and elsewhere.
Optimal control in stochastic continuous time setups is used in many applications.
Examples include issues in international macro (Obstfeld, 1994, Turnovsky, 1997,
2000), international …nance and debt crises (Stein, 2006) and many others (coming
soon). A …rm maximization problem with risk-neutrality where R&D increases quality
230
of goods, modelled by a stochastic di¤erential equation with Poisson uncertainty, is
presented and solved by Dinopoulos and Thompson (1998, sect. 2.3).
The Keynes-Ramsey rule in (10.27) was derived in a more or less complex frame-
work by Breeden (1986) in a synthesis of his consumption based capital asset pricing
model, Cox, Ingersoll and Ross (1985) in their continuous time capital asset pricing
model, or Turnovsky (2000) in his textbook. Cf. also Obstfeld (1994).
Closed-form solutions and analytical expressions for value functions are derived
by many authors. This approach was pioneered by Merton (1969, 1971) for Brownian
motion. Chang (2004) devotes an entire chapter (ch. 5) on closed-form solutions for
Brownian motion. For setups with Poisson-uncertainty, Dinopoulos and Thompson
(1998, sect. 2.3), Wälde (1999b) or Sennewald and Wälde (2006, ch. 3.4) derive
closed-form solutions.
Closed-form solutions for Levy processes are available e.g. from Aase (1984),
Framstad, Øksendal and Sulem (2001) and in the textbook by Øksendal and Sulem
(2005).
Sometimes, restricting the parameter set of the economy in some intelligent way al-
lows to provide closed-form solutions for very general models. These solutions provide
insights which can not that easily be obtained by numerical analysis. Early examples
are Long and Plosser (1983) and Benhabib and Rustichini (1994) who obtain closed-
form solutions for a discrete-time stochastic setup. In deterministic, continuous time,
Xie (1991, 1994) and Barro, Mankiw and Sala-i-Martin (1995) use this approach as
well. Wälde (2005) and Wälde and Posch (2006) derive closed form solutions for
business cycle models with Poisson uncertainty.
There are various recent papers which use continuous time methods under un-
certainty. For examples from …nance and monetary economics, see DeMarzo and
Urošević (2006), Gabaix et al. (2006), Maenhout (2006), Piazzesi (2005), examples
from risk theory and learning include Kyle et al. (2006) and Keller and Rady (1999),
for industrial organization see e.g. Murto (2004), in spatial economics there is e.g.
Gabaix (1999), the behaviour of households in the presence of durable consumption
goods is analyzed by Bertola et al. (2005). The real options approch to investment
under uncertainty is another larger area. A recent contribution is by Guo et al.
(2005). Further references to papers that use Poisson processes can be found in ch.
9.5.
231
Exercises Chapter 10
Applied Intertemporal Optimization
Dynamic Programming in continuous
time under uncertainty
1. Optimal saving under Poisson uncertainty with two state variables (exam)
Consider the objective function
Z 1
U (t) = Et e [ t] u (c ( )) d
t
where r and w are constant interest and wage rates, q (t) is a Poisson process
with an exogenous arrival rate and is a constant as well. Letting g and
denote constants, assume that the price p (t) of the consumption good follows
(a) Derive a rule which optimally describes the evolution of consumption. De-
riving this rule in the form of marginal utility, i.e. du0 (c (t)) is su¢ cient.
(b) Derive a rule for consumption, i.e. dc (t) = :::
(c) Derive a rule for optimal consumption for = 0 or = 0:
3. Adjustment cost R1
Consider a …rm that maximizes its present value de…ned by = Et t e r[ t] ( ) d :
The …rm’s pro…t is given by the di¤erence between revenues and costs,
2
= px ci ; where output is assumed to be a function of the current cap-
ital stock, x = k : The …rm’s control variable is investment i that determines
its capital stock,
dk = (i k) dt:
232
The …rm operates in an uncertain environment. Output prices and costs for
investment evolve according to
dp=p = p dt + p dzp ; dc=c = c dt + c dzc ;
(a) What is the optimal static employment level l of this …rm for a given
technological level q?
(b) Formulate an intertemporal objective function given by the present value of
the …rms pro…t ‡ows over an in…nite time horizon. Continue to assume that
q is constant and let the …rm choose l optimally from this intertemporal
perspective. Does the result change with respect to (a)?
(c) Using the same objective function as in (b), let the …rm now determine
both l and lq optimally. What are the …rst order conditions? Give an
interpretation in words.
(d) Compute the expected output level for > t, given the optimal employ-
ment levels l and lq : In other words, compute Et x ( ) : Hint: Derive …rst
a stochastic di¤erential equation for output x (t) :
233
(a) Derive the budget constraint of the household. Use (t) v (t) k (t) =a (t)
as the share of wealth held in stocks.
(b) Derive the budget constraint of the household by assuming that q (t) is
Brownian motion.
(c) Now let the household live in a world with three assets (in addition to the
two above, there are assets available on f oreign markets). Assume that
the budget constraint of the household is given by
da (t) = fr (t) a (t) + w (t) p (t) c (t)g dt+ k k (t) a (t) dq (t)+ f f (t) a (t) dqf (t) ;
where
r (t) = k (t) rk + f (t) rf + (1 k (t) f (t)) rb
is the interest rate depending on weights i (t) and constant instantaneous
interest rates rk ; rf and rb : Let q (t) and qf (t) be two Poisson processes.
Given the usual objective function
Z 1
U (t) = Et e [ t] u (c ( )) d ;
t
what is the optimal consumption rule? What can be said about optimal
shares k and f ?
(a) Show that the covariance of asset i with the market portfolio iM is given
by Nj=1 j ij :
N 2
(b) Show that j=1 i iM is the variance of the market portfolio M.
234
7. Standard and non-standard technologies (exam)
Let the social welfare function of a central planner be given by
Z 1
U (t) = Et e [ t] u (C ( )) d :
t
235
11 Notation and variables, references and index
Notation and variables
The notation is as homogenous as possible. Here are the general rules but some
exceptions are possible.
A function f (:) is always presented by using parentheses (:) ; where the paren-
theses give the arguments of functions. Brackets [:] always denote a multiplica-
tion operation
Here is a list of variables and abbreviations which shows that some variables
have multi-tasking abilities, i.e. they have multiple meanings
Greek letters
236
Latin letters
237
References
Aase, K. K. (1984): “Optimum Portfolio Diversi…cation in a General Continuous-
Time Model,”Stochastic Processes and their Applications, 18, 81–98.
Abel, A. B. (1990): “Aggregate Asset Pricing; Asset Prices under Habit Formation
and Catching up the Joneses,”American Economic Review; AEA Papers and Pro-
ceedings, 80(2), 38–42.
Aghion, P., and P. Howitt (1992): “A Model of Growth Through Creative Destruc-
tion,”Econometrica, 60, 323–351.
Araujo, A., and J. A. Scheinkman (1983): “Maximum principle and transversality con-
dition for concave in…nite horizon economic models,”Journal of Economic Theory,
30, 1 –16.
Arrow, K. J.and Kurz, M. (1970): Public investment, the rate of return and optimal
…scal policy. Johns Hopkins University Press, Baltimore.
Benhabib, J., and A. Rustichini (1994): “A note on a new class of solutions to dy-
namic programming problems arising in economic growth,” Journal of Economic
Dynamics and Control, 18, 807–813.
Bertola, G., L. Guiso, and L. Pistaferri (2005): “Uncertainty and Consumer Durables
Adjustment,”Review of Economic Studies, 72, 973U1007.½
Björk, T. (2004): Arbitrage Theory in Continuous Time, 2nd ed. Oxford University
Press.
Black, F., and M. Scholes (1973): “The Pricing of Options and Corporate Liabilites,”
Journal of Political Economy, 81, 637–659.
Bossmann, M., C. Kleiber, and K. Wälde (2007): “Bequests, taxation and the distri-
bution of wealth in a general equilibrium model,” Forthcoming Journal of Public
Economics.
238
Brock, W. A., and A. G. Malliaris (1989): “Di¤erential Equations, Stability and
Chaos in Dynamic Economics,”Advanced Textbooks in Economics.
Burdett, K., and D. T. Mortensen (1998): “Wage Di¤erentials, Employer Size, and
Unemployment,”International Economic Review, 39(2), 257–73.
Chen, Y.-F., and M. Funke (2005): “Non-Wage Labour Costs, Policy Uncertainty
And Labour Demand - A Theoretical Assessment,” Scottish Journal of Political
Economy, 52, 687–709.
Cox, J. C., J. E. Ingersoll, and S. A. Ross (1985): “An intertmporal general Equilib-
rium Model of asset prices,”Econometria, 53(2), 363–.
Danthine, Jean Pierre Donaldson, J. B., and R. Mehra (1992): “the equity premium
and the allocation of income risk,”Journal fo Economics and Control, 16, 509 532.
Das, S. R., and S. Foresi (1996): “Exact Solutions for Bond and Option Prices with
Systematic Jump Risk,”Review of Derivatives Research, 1, 7–24.
de la Croix, D., and P. Michel (2002): A theory of economic growth: dynamics and
policy in overlapping generations. Cambridge University Press.
239
(2001): Dynamic Asset Pricing Theory, 3rd ed. Princeton University Press.
Eaton, J. (1981): “Fiscal Policy, In‡ation and the Accumulation of Risky Capital,”
Review of Economic Studies, 48(153), 435–445.
Epaulard, A., and A. Pommeret (2003): “Recursive Utility, Endogenous Growth, and
the Welfare Cost of Volatility,”Review of Economic Dynamics, 6(2), 672–684.
Evans, Hastings, and Peacock (2000): Statistical Distributions, 3rd edition. Wiley.
Francois, P., and H. Lloyd-Ellis (2003): “Animal Spirits Trough Creative Destruc-
tion,”American Economic Review, 93, 530–550.
Gabszewicz, J., and J. P. Vial (1972): “Oligopoly "à la Cournot" in a general equi-
librium analysis,”Journal of Economic Theory, 4, 381–400.
García, M. A., and R. J. Griego (1994): “An Elementary Theory of Stochastic Dif-
ferential Equations Driven by A Poisson Process,” Commun. Statist. - Stochastic
Models, 10(2), 335–363.
240
Grossman, G. M., and E. Helpman (1991): Innovation and Growth in the Global
Economy. The MIT Press, Cambridge, Massachusetts.
Guo, X., J. Miao, and E. Morelle (2005): “Irreversible investment with regime shifts,”
Journal of Economic Theory, 122, 37U59.½
Guriev, S., and D. Kvasov (2005): “Contracting on Time,” American Economic Re-
view, 95(5), 1369–1385.
Hassett, K. A., and G. E. Metcalf (1999): “Investment with Uncertain Tax Policy:
Does Random Tax Discourage Investment?,”Economic Journal, 109, 372–393.
Heer, B., and A. Maußner (2005): Dynamic General Equilibrium Modelling, Compu-
tational Methods and Applications. Springer, Berlin.
Judd, K. L. (1985): “The Law of Large Numbers with a Continuum of I.I.D. Random
Variables,”Journal of Economic Theory, 35, 19–25.
Kleiber, C., and S. Kotz (2003): Statistical Size Distributions in Economics and
Actuarial Sciences. Wiley.
Kyle, A. S., H. Ou-Yang, and W. Xiong (2006): “Prospect theory and liquidation
½
decisions,”Journal of Economic Theory, 129, 273U288.
241
Laitner, J., and D. Stolyarov (2004): “Aggregate returns to scale and embodied tech-
nical change: theory and measurement using stock market data,”Journal of Mon-
etary Economics, 51, 191–233, pdf.
Leonard, D., and N. V. Long (1992): Optimal Control Theory and Static Optimization
in Economics. Cambridge University Press, New York.
Ljungqvist, L., and T. J. Sargent (2004): Recursive Macroeconomic Theory 2nd ed.
MIT Press.
McDonald, R., and D. Siegel (1986): “The Value of Waiting to Invest,” Quarterly
Journal of Economics, 101, 707–27.
242
Obstfeld, M. (1994): “Risk-Taking, Global Diversi…cation, and Growth,” American
Economic Review, 84, 1310–29.
Øksendal, B., and A. Sulem (2005): Applied Stochastic Control of Jump Di¤usions.
Springer-Verlag, Berlin.
(2005): “International Labor Union policy and Growth with Creative De-
struction,”Review of International Economics, 13, 90–105.
Piazzesi, M. (2005): “Bond Yields and the Federal Reserve,” Journal of Political
Economy, 113(2), 311–344.
Santanu Chatterjee, P. G., and S. J. Turnovsky (2004): “Capital Income Taxes and
Growth in a Stochastic Economy: A Numerical Analysis of the Role of Risk Aver-
sion and Intertemporal Substitution,” Journal of Public Economic Theory, 6(2),
277–310.
243
Sennewald, K., and K. Wälde (2005): “"Ito’s Lemma" and the Bellman equation for
Poisson processes: An applied view,”Journal of Economics, forthcoming.
Shapiro, C., and J. E. Stiglitz (1984): “Equilibrium Unemployment as a Worker
Discipline Device,”American Economic Review, 74, 433–44.
Shell, K. (1969): Applications of Pontryagin’s Maximum Principle to Economics.
Springer, Berlin.
Shimer, R. (2005): “The Cyclical Behavior of Equilibrium Unemployment and Va-
cancies,”American Economic Review, 95, 25–49.
Smith, W. T. (1996): “Feasibility and Transversality Conditions for Models of Port-
folio Choice with Non-Expected Utility in Continuous Time,” Economics Letters,
53, 123–31.
Steger, T. M. (2005): “Stochastic growth under Wiener and Poisson uncertainty,”
Economics Letters, 86(3), 311–316.
Stein, J. L. (2006): Stochastic Optimal Control, International Finance, and Debt
Crises. Oxford University Press.
Stokey, N. L., R. E. Lucas Jr., and E. C. Prescott (1989): Recursive Methods in
Economics Dynamics. Harvard University Press, Cambridge, MA.
Thompson, P., and D. Waldo (1994): “Growth and trusti…ed capitalism,”Journal of
Monetary Economics, 34, 445–462.
Toche, P. (2005): “A tractable model of precautionary saving in continuous time,”
Economics Letters, 87, 267–272.
Turnovsky, S. J. (1997): International macroeconomic dynamics. MIT Press.
(2000): International macroeconomic dynamics Methods of Macroeconomic
Dynamics. MIT Press.
Turnovsky, S. J., and W. T. Smith (2006): “Equilibrium consumption and precaution-
ary savings in a stochastically growing economy,” Journal of Economic Dynamics
and Control, 30(2), 243–278.
Varian, H. (1992): Microeconomic Analysis. Norton, Third Edition.
Venegaz-Martinez, F. (2001): “Temporary Stabilization: A Stochastic Analysis,”
Journal of Economic Dynamics and Control, 25, 1429–1449.
Vissing-Jorgensen, A. (2002): “Limited Asset Market Participation and the Elasticity
of Intertemporal Substittuion,”Journal of Political Economy, 110, 825–853.
Wälde, K. (1996): “Proof of Global Stability, Transitional Dynamics, and Interna-
tional Capital Flows in a Two-Country Model of Innovation and Growth,”Journal
of Economics, 64, 53–84.
244
(1999a): “A Model of Creative Destruction with Undiversi…able Risk and
Optimising Households,”Economic Journal, 109, C156–C171.
245
Index
adjustment cost, 164, 165 Hamiltonian, present value, 95, 97
adjustment costs, 84, 163 heterogeneous agents, 134
Bellman equation, 35, 38, 40, 104, 148, identically and independently distributed
150, 152, 157, 210, 221, 222, 226 (iid), 123, 126, 128
adjusting the basic form, 222 integral representation of di¤erential equa-
basic form, 210 tions, 68, 176, 199, 202, 219
Bellman equation, maximized, 41, 105, 149, interest rate, de…nition, 18
150, 158, 221 investment, gross, 17
De…nition, 36, 211 investment, net, 17
Brownian motion, 170, 171, 173, 178, 186, Ito’s Lemma, 177, 190, 223
195, 198, 221, 237
budget constraint, 25, 31, 37, 49, 69, 74, Keynes-Ramsey rule, 106, 166
76, 86, 148, 150, 157, 191, 192, Leibniz rule, 65
209, 221, 225 level e¤ect
dynamic, 87 dynamic, 41, 88
capital asset pricing (CAP), 139, 157, 224 martingale, 198, 219
central planner, 14, 33, 50, 100, 151, 167
CES utility, 146 natural volatility, 140, 213, 228
closed-form solution, 8, 167, 213, 231 numeraire good, 7, 25, 43, 152, 156, 228
2 period model, 130, 131, 136 numerical solution, 166
Cobb-Douglas utility, 7, 131
consumption level and interest rate, 88 Poisson process, 170, 172, 180, 203, 205,
control variable, 35, 78, 95, 105, 129, 160, 209, 237
161, 221
representative agent assumption, 134
CVF (change of variable formula), 177,
resource constraint, 17
181, 185, 211, 212
risk aversion
depreciation rate, 17 absolute, 87
relative, 90
Euler’s theorem, 15, 16
expectations operator, 124, 129, 131, 150, shadow price, 14, 15, 36, 79, 80, 94, 158
160, 199, 203, 219 state variable, 35, 78, 79, 95, 107, 149,
expected value, 195, 202 160, 210, 222
sticky prices, 162
…rms, intertemporal optimization, 83, 162 switch process, 193
…rst-order condition with economic inter-
pretation, 135, 215 Wiener process, 171, 237
246
12 Exams
sorry - all future exams deleted ... :-( ;-)
247