Macroeconomics Toolbox - Excelente PDF
Macroeconomics Toolbox - Excelente PDF
Thomas Steger
May 4, 2017
Contents
1 Introduction 3
2 Dynamic Optimization 8
2.1 Method of Lagrange Multipliers . . . . . . . . . . . . . . . . . . . . . . 8
2.1.1 Simple example . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.1.2 Standard dynamic problem . . . . . . . . . . . . . . . . . . . . . 9
2.1.3 Application: Investment under capital adjustment costs . . . . . 11
2.2 Dynamic Programming . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.2.1 Simple example . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.2.2 More general treatment . . . . . . . . . . . . . . . . . . . . . . . 18
2.3 Control Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.3.1 Method of Lagrange Multipliers - static problems . . . . . . . . 21
2.3.2 Maximum Principle . . . . . . . . . . . . . . . . . . . . . . . . . 23
0
2.4 Application (1): Education à la Lucas . . . . . . . . . . . . . . . . . . . 27
2.5 Application (2): TVC in the Ramsey model . . . . . . . . . . . . . . . 29
2.6 Digression: No-Ponzi-game condition . . . . . . . . . . . . . . . . . . . 31
3 Dynamic systems 33
3.1 Di¤erential equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.1.1 First-order, linear di¤erential equations (constant coe¢ cient and
term) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.1.2 Second-order, linear di¤erential equations (constant coe¢ cient
and term) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
3.1.3 Two-dimensional, linear di¤erential equation system . . . . . . . 38
3.2 Di¤erence equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
3.2.1 First-order, linear di¤erence equation (constant coe¢ cient and
term) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
3.2.2 Second-order, linear di¤erence equation (constant coe¢ cient and
term) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
3.2.3 Two-dimensional system of linear di¤erence equations . . . . . . 45
3.3 Linearization techniques . . . . . . . . . . . . . . . . . . . . . . . . . . 46
3.3.1 First-order Taylor series approximation . . . . . . . . . . . . . . 46
3.3.2 Proportional deviations from steady state . . . . . . . . . . . . 48
3.3.3 Application 1: Solow model - speed of convergence . . . . . . . 49
3.3.4 Application 2: PDV of wage income . . . . . . . . . . . . . . . . 50
3.3.5 Application 3: Ramsey model - determining c(0) . . . . . . . . . 51
3.3.6 Application 4: Solow model - internal rate of return . . . . . . . 52
3.4 Miscellaneous . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
3.4.1 First-order, linear di¤erential equations - variable coe¢ cients and
variable term . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
3.4.2 Application 1: Wealth in a small open economy . . . . . . . . . 56
3.4.3 Application 2: Asset prices . . . . . . . . . . . . . . . . . . . . . 57
3.4.4 Normalization of variables . . . . . . . . . . . . . . . . . . . . . 60
1
4 Production and Technology 62
4.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
4.2 General technology (non-parametric) . . . . . . . . . . . . . . . . . . . 63
4.2.1 Homogenous functions . . . . . . . . . . . . . . . . . . . . . . . 63
4.2.2 Euler’s theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
4.2.3 Intensive form and competitive factor prices . . . . . . . . . . . 64
4.2.4 Elasticity of substitution . . . . . . . . . . . . . . . . . . . . . . 65
4.3 Parametric technology: Cobb-Douglas . . . . . . . . . . . . . . . . . . . 66
4.3.1 Properties and implications . . . . . . . . . . . . . . . . . . . . 67
4.4 Parametric technology: CES . . . . . . . . . . . . . . . . . . . . . . . . 69
4.4.1 Properties and implications . . . . . . . . . . . . . . . . . . . . 70
4.4.2 Complementarity: clari…cation on terminology . . . . . . . . . . 73
4.4.3 Application: competitive CES sector . . . . . . . . . . . . . . . 74
4.5 Two methodological issues . . . . . . . . . . . . . . . . . . . . . . . . . 76
2
1 Introduction
The Macroeconomics Toolbox accompanies the course Advanced Macroeconomics at
Leipzig University. It provides a concise summary of those concepts and techniques
that are required to study topics such as: (i) Consumption and Savings: Dynasties and
OLG, (ii) Endogenous Growth, (iii) Investment under Capital Adjustment Costs, (iv)
Real Business Cycles, (v) Price Rigidities: New Keynesians, (vi) Budget De…cits and
Fiscal Policy, (vii) Money and the Financial Sector.
This introduction reviews some important methodological concepts. To start with,
consider a set of well-known methodological statements regarding the use of (economic)
theory.
A map on the scale of one to one would be useless. Joan Robinson (Robinson, 1962,
p. 33).
We prefer simpler theories to more complex ones because their empirical content is
greater, and because they are better testable. In other words, a simple theory applies
to more cases than a more complex one, and is thus more easily falsi…able. Karl Popper
(1992)
Simplifying assumptions. The quotations above all make the following point: If
(economic) theory should be instructive, we need to make use of simplifying assump-
tions. Simplifying assumptions reduce the complexity of the real world and allow us
to learn something about, say, the basic mechanisms which give rise to empirically
observed correlations.
3
Figure 1: Simplifying assumptions
4
di¤er with respect to preferences and incomes. Firms may di¤er with respect to produc-
tion technologies. Building macroeconomic models which capture these heterogeneities
becomes complicated, if not impossible. Fortunately, there is a modeling trick to avoid
these complications.
To illustrate, let us focus on goods demand. Assume that there are, say, 100 con-
sumers (indexed by i 2 f1; :::; 100g) who may di¤er with respect to income (yi ). Ag-
gregate goods demand then results from the aggregation over individual goods demand
P100
schedules denoted as xi (p; yi ), i.e. i=1 xi (p; yi ). However, the same goods demand
function may result from the optimization problem of 100 identical households, i.e.
100 x(p; y), where x(p; y) is the demanded quantity of good x at price p and income
y. If this is the case and if we are interested in aggregated outcomes, then we can
obviously simplify the analysis by assuming that the model economy is populated by a
large number (100 in the example above) of identical households, which is labelled the
representative household (cf. Mas-Colell, Whinston and Green, 1995, Chapter 4D).
The following quote (Acemoglu, 2009, pp. 149/150) may be further instructive:
It is clear that all phenomena which are essentially based on interpersonal hetero-
geneity (e.g. distributional aspects) cannot be analyzed within this framework. The
representative agent model is used instead to analyze central macroeconomic phenom-
ena such as intertemporal consumption decisions, capital accumulation, and growth.
5
Heterogeneous agent economy. Caselli and Ventura (2000) consider the follow-
ing Ramsey model (stated below in the form of the dynamic optimization problem of
households):
Z 1 1
cj (t) + j g(t) 1 t
max e dt (1)
fcj (t)g 0 1
s.t. a_ j (t) = r(t)aj (t) + w(t)hj (t) cj (t); aj (0) = given,
where cj (t) denotes consumption of agent j at time t, aj (t) …nancial wealth of agent
j, hj (t) the amount of human capital of agent j, r(t) the real interest rate, w(t) the
wage rate per unit of human capital, j an agent speci…c taste parameter, g(t) a public
good, and ; > 0 constant preference parameters, respectively. Caselli and Ventura
(2000, p. 912) state that
This economy admits a representative consumer, in the usual sense that the sum of
all consumers behave exactly as if the economy contained a single consumer with the
average asset holdings, skills, and taste parameters.
Mass one of identical agents. Often we employ the simplifying assumption that
the economy is populated by a "large number" of representative agents (households,
…rms or consumer-producer households). As an example, consider an economy with N
identical …nal output …rms. Every …rm is indexed by i 2 f1; :::; N g. The …rm index i
represents a discrete variable (i 2 N). Let xi denote output of …rm i. Total output is
P
then given by X = N i=1 xi . Since all …rms are completely identical, it is plausible to
assume that, in equilibrium, every …rm produces the same amount, i.e. xi = x for all
i 2 f1; :::; N g. Total output may then be expressed as X = N x.
A widely employed ’modeling trick’ is as follows. Assume that …rms are indexed
by i 2 [0; ::; 1]. The …rm index i now represents a continuous variable (i 2 R+ ). This
assumption is often expressed by phrases like "there is mass one of identical …rms"
or "there is a continuum of length one of identical …rms". Let x(i) denote output of
R1
…rm i. Total output is then given by X = i=0 x(i)di . Since all …rms are alike, we
6
can assume that x(i) = x for all i 2 [0; ::; 1]. Aggregate output is then given by X =
R1
i=0
x(i)di = 1x. Hence, individual quantities coincide with aggregate quantities.
demand functions are homogeneous of degree zero, i.e. dji ( p1 ; :::; pn ) = dji (p1 ; :::; pn )
for all 0.
A Walrasian equilibrium (general equilibrium) is de…ned as a price vector (p1 ; :::; pn )
that satis…es the following set of n market clearing equations
X
m X
m
dji (p1 ; :::; pn ) = eji 8 i 2 f1; :::; ng: (2)
j=1 j=1
There are n unknowns (p1 ; :::; pn ) and n equations. The system does not, however,
exhibit a unique solution. It is easily shown that if a solution exists, then there is
a multiplicity of solutions. Assume that (~
p1 ; :::; p~n ) satis…es the system of equations
given by (2). By the property of homogeneity of degree zero, one can multiply this
price vector by any positive constant ( p~1 ; :::; p~n ) such that real goods demand dji (:)
doesn’t change. Hence the price vector ( p~1 ; :::; p~n ) does equally satisfy the above
stated set of market clearing conditions.
Notice that this implies that the n prices (p1 ; :::; pn ) are indetermined, but all ratios
1
of any two prices are determined. For instance, if one sets = pn
, one receives a
solution vector of relative prices ( ppn1 ; :::; 1). One usually states that only the set of
relative prices is determined in general equilibrium. To determine the set of absolute
prices, we can set one of the n prices equal to a constant. Let’s choose pn = 1. Then all
prices are denominated in terms of the numeraire good, which is good n if we choose
pn = 1. In the real world the numeraire good is called ’money’(McKenzie, 2008).
7
2 Dynamic Optimization
Dynamic optimization plays a prominent role in macroeconomic theory. Examples
comprise the determination of an optimal consumption plan (Ramsey model, Diamond
model, Auerbach-Kotliko¤ model) and the determination of an optimal investment
plan (neoclassical …rm under capital adjustment costs). There are three standard ap-
proaches to solve dynamic optimization problems: (i) method of Lagrange multipliers,
(ii) dynamic programming and (iii) maximum principle.
X
2
0 1 2
max u(c0 ) + u(c1 ) + u(c2 ) (3)
fc0 ;c1 ;c2 ;s0 ;s1 g
t=0
s.t. w0 = s0 + c0 (4)
w1 + (1 + r)s0 = s1 + c1 (5)
(1 + r)s1 = c2 (6)
where ct denotes consumption at period t 2 f0; 1; 2g, st savings, wt the wage rate,
0 < < 1 the discount factor, and r the constant interest rate. The problem is to
maximize the objective function (3) with respect to c0 , c1 , c2 , s0 , s1 subject to the
constraints (4), (5), (6). We focus on interior solutions. This problem can be solved
by the method of Lagrange multipliers. The corresponding (present-value) Lagrangian
function reads
0 1 2
L : = u(c0 ) + u(c1 ) + u(c2 )
8
where ( 0 , 1, 2) denote present-value shadow prices. Di¤erentiating L w.r.t. the
unknowns yields
@L 0 0
= u (c0 ) 0 =0
@c0
@L 1 0
= u (c1 ) 1 =0
@c1
@L 2 0
= u (c2 ) 2 =0
@c2
@L
= 0 + 1 (1 + r) = 0
@s0
@L
= 1 + 2 (1 + r) = 0
@s1
and 2. Hence, there are 8 unknowns and 8 equations. Solving the set of equations
yields the optimal time path of consumption fc0 ; c1 ; c2 g and savings decisions fs0 ; s1 g.
Notice that the last static constraint (6) requires that entire wealth in the last period
is to be consumed.
where E0 denotes the expected value given the information at time 0 and "t an i.i.d.
random variable with E("t ) = 0 and V ("t ) = ". Notice that either the constraint
xt 0 (economy without credit market) or the No-Ponzi-Game Condition (NPGC)
1 t
limt!1 1+r
xt 0 (economy with credit market) are assumed to hold.
At t = 0 the agent decides on c0 , taking x0 as given. The unknowns at this stage
1
This section closely follows Chow (1997, Chapter 2) and Heer and Mauß
ner (2009, Chapter 1.3.2).
9
are c0 and x1 . The associated Lagrangian function reads2
(1 )
X X
1
t t
L := E0 u(ct ) t [xt+1 f (xt ; ct ) "t+1 ]
t=0 t=0
0 1 0 1
L = E0 u(c0 ) + u(c1 ) 0 [x1 f (x0 ; c0 ) "1 ] 1 [x2 f (x1 ; c1 ) "2 ] + :::
@L @L
The FOCs @c0
= 0 and @x1
= 0 read as follows
@L @f (x0 ; c0 )
= E0 fu0 (c0 ) + 0 g=0
@c0 @c0
@L 0 1 @f (x1 ; c1 )
= E0 f 0 + 1 g=0
@x1 @x1
At t = 1 the agent decides on c1 , taking x1 as given. The unknowns at this stage
are c1 and x2 . The Lagrangian function for this problem is given by
X
1 X
1
t 1 t 1
L := E1 f u(ct ) t [xt+1 f (xt ; ct ) "t+1 ]g
t=1 t=1
0 1 0 1
L = E1 f u(c1 ) + u(c2 ) 1 [x2 f (x1 ; c1 ) "2 ] 2 [x3 f (x2 ; c2 ) "3 ] + :::g
@L @L
The FOCs @c1
= 0 and @x2
= 0 read as follows
@L @f (x1 ; c1 )
= E1 fu0 (c1 ) + 1 g=0
@c1 @c1
2
The de…nition of L di¤ers from Chow (1997, Chapter 2), where it reads (the above is, however,
in line with our RBC examples in lecture notes):
(1 1
)
X X
t t+1
L := E0 u(ct ) t+1 [xt+1 f (xt ; ct ) "t+1 ]
t=0 t=0
This does not make a di¤erence! It merely means a relabeling of shadow prices (number of shadow
prices remains the same, de…nition of shadow prices changes).
10
@L 1 @f (x2 ; c2 )
= E1 f 1 + 2 g=0
@x2 @x2
In general at period t, the FOCs may be expressed as follows
@f (xt ; ct )
Et fu0 (ct ) + t g=0 (10)
@ct
t @f (xt+1 ; ct+1 )
Et f t + t+1 g=0 (11)
@xt+1
In addition, the dynamic constraint xt+1 = f (xt ; ct ) + "t+1 and the transversality
t
condition (TVC) limt!1 Et ( t xt ) = 0 belong to the set of FOCs. Notice that this
solution strategy does not amount to choose {c0 ; c1 ; c2 ; :::} all at once in an open loop
policy, rather ct is chosen sequentially given the information xt at time t an a closed-
loop policy. Because xt is in the information set when ct is to be determined, the
expectations in (10) and (11) are Et and not E0 .
Setup Consider a …rm that produces a homogenous …nal output good Yt under perfect
competition. The output technology reads
Yt = A (Kt ) (Lt )1
where Kt is the amount of capital at time t 2 f0; 1; :::g, Lt the amount of labor,
and A > 0, 0 < < 1. The …rm’s planning horizon is in…nity. Expanding the stock of
capital is subject to investment costs (capital adjustment costs). Installing the amount
of It additional capital goods (gross investment) requires
It
ICt = It 1 + (12)
Kt
with 0 and 1 units of …nal output. Notice that the usual assumption is =0
11
(standard Ramsey model). Let
denote the …rm’s cash ‡ow (or the capitalist’s / entrepreneur’s residual income). We
assume that the …rm maximizes the PDV of cash ‡ow, i.e.
X
1
1
t
max CFt (14)
fIt ;Lt g
t=0
1+r
s.t. Kt+1 = It + (1 )Kt ; (12) ; (13); K0 = given
where wt denotes the wage rate, r the …xed interest rate, and 0 the capital
depreciation rate.
Assignment: Determine the …rm’s investment demand (i.e. the demand for …nal
output devoted to capital investment).
Remark: The investment demand results from the solution of the above stated
dynamic optimization problem and will be a function It = It (qt ), where qt denotes the
shadow price of installed capital goods.
Solution The Lagrangian function and the associated …rst order conditions read
8 h i 9
X
1
1
t < A (Kt ) (Lt )1 wt Lt It 1 + It =
Kt
L=
t=0
1+r : +qt [It + (1 ) Kt Kt+1 ] ;
@L
= (1 )A (Kt ) (Lt ) wt = 0
@Lt
12
1
@L It It 1
= 1+ It + qt = 0
@It Kt Kt Kt
It It
) 1 + qt = 0
Kt Kt
It
) 1 ( + 1) + qt = 0
Kt
1
qt 1
) It = Kt
( + 1)
@L
From @Kt+1
= 0 one gets
" #
+1
1 1 It+1
qt + A (Kt+1 ) (Lt+1 )1 + + qt+1 (1 ) = 0 (15)
1+r Kt+1
Reasoning for above stated qt+1 = f (qt ) (see also the lecture notes on RBC theory)
0
1 I0
L = A (K0 ) (L0 )1 w0 L0 I0 1 + + q0 (I0 + (1 ) K0 K1 )
1+r K0
1
1 I1
+ A (K1 ) (L1 )1 w1 L1 I1 1 + + q1 (I1 + (1 ) K1 K2 )
1+r K1
+:::
" #
1
@L 1 1 I1 I1
= q0 + A (K1 ) (L1 )1 I1 + q1 (1 ) =0
@K1 1+r K1 (K1 )2
" #
+1
1 1 I1
) q0 + A (K1 ) (L1 )1 + + q1 (1 ) =0
1+r K1
1
qt 1 Kt+1
Steady state From Kt+1 = ( +1)
Kt + (1 ) Kt and Kt
= 1 (steady state)
one gets the steady state value of qt to read
1
qt 1
1= + (1 ) ) q~ = ( + 1) + 1
( + 1)
1
qt 1
Notice that this implies that, by using It = ( +1)
Kt and q~ = ( + 1) + 1,
13
I
the investment ratio K
in the steady state is
1
SS
I q~ 1
= =
K ( + 1)
Interesting observations: (i) In steady state q = 1 if = 0 and larger than one for
> 0. (ii) Steady state q does not depend on r and A. (iii) As investment technology
becomes more e¢ cient, i.e. falls, the long run value of installed capital declines.
These observations should be discussed.
Dynamic responses The following graph shows the dynamic responses of K, q and
I following a TFP shock of 10 percent, i.e. 0 = 0:1. It is assumed that TFP follows
At+1 = (At ) e t with t white noise. The case = 1 is distinguished from = 0:95.
14
The total value of a …rm (shareholder value) is
Z 1
I r( t)
Vt = Y wL I 1+ e d
t K
qA K = V
The variable qA is called average q. The variable q(t) is called marginal q. It shows
by how much the value of a …rm increases when capital increases by one unit, i.e.
@V
q(t) =
@K
Extensions / assignments
15
where 0 < < 1 denotes a corporate income tax rate, 0 < sK < 1 an investment
sK
subsidy (design of …scal incentive: tax allowances), and s~K := 1
may be
labelled the behaviorally relevant subsidy rate.
At+1 = (At ) e t
where 0 < < 1 and "t white noise. (Figure 2 already captures this case.)
Consider an individual living for T > 0 periods. Intertemporal utility is given by4
X
T
t
U (c0 ; :::; cT ) = u(ct )
t=0
where Rt is the (probably stochastic) gross rate of return. The optimal consumption
plan {c0 ; c1 ; :::; cT } can be determined by backward induction.
At T 1 the individual has wealth aT 1 and the maximum utility may be described
as follows
VT 1 (aT 1 ) = maxfu(cT 1) + u((aT 1 c )R )g (16)
cT 1 | {zT 1 T }1
cT =aT
4
This example is taken from Varian (1992, Chapter 19.3).
16
The value function VT 1 (aT 1 ) shows an indirect utility, i.e. the maximum attainable
utility given wealth aT 1. The FOC for the problem on the RHS is
This is just like (16) but with "second-period" utility replaced by indirect utility
VT 1 (aT 1 ); this equation is often referred to as the Bellman equation. The FOC
for the problem on the RHS is
17
Vt = Vt (at ), optimal ct follows from the …rst-order conditions u0 (ct ) = EVt+1
0
(at+1 )Rt .
Finally, the time path at for all t 2 f0; :::; T g results from at+1 = (at ct )Rt with a0
given.
As an illustration, consider an agent living for 11 periods who is endowed with
initial wealth a0 = 1. The subsequent …gure displays the time paths for wealth (at )
and consumption (ct ) assuming di¤erent constellations with regard to R and .
The state variable xt cannot be controlled directly but is under indirect control
by choosing ct . An optimal program is a sequence fct ; ct+1 ; :::g that solves the above
stated problem. The value of an optimal program is denoted by
5
This section closely follows Wälde (2010, Chapter 3.3).
18
The value function V (xt ) shows the maximum level of life-time utility Ut given the
amount of the state variable xt .
Step #1: Bellman equation and FOCs. Since the objective function Ut is ad-
ditively separable, it may be written as Ut = u(ct ) + Ut+1 and hence the maximization
problem may be stated as
s.t. xt+1 = f (xt ; ct ). Equ. (19) is known as the Bellman equation. In this equation,
the problem with potentially in…nitely many control variables was broken down into
many problems with one control variable. Notice that, compared to Ut = u(ct )+ Ut+1 ,
Ut+1 is replaced by V (xt+1 ). We assume that the optimal program for tomorrow is
solved and we worry about the maximization problem for today only.
The FOC for the problem on the RHS of (19), noting xt+1 = f (xt ; ct ), is
Increasing consumption ct has bene…ts and costs. The bene…t consist in higher
utility today, which is re‡ected by marginal utility u0 (ct ). The cost comes from lower
overall utility - the value function V (:) - tomorrow. The reduction in overall utility
amounts to the change in xt+1 , i.e. the derivative fc , times the marginal value of xt+1 ,
i.e. V 0 (xt+1 ). As the cost arises only tomorrow, this is discounted at rate .
Equ. (20), together with xt+1 = f (xt ; ct ), implicitly de…nes ct = ct (xt ). As we know
very little about the properties of V (:) at this stage, however, we need to go through
two further steps to eliminate V (:) from this FOC and obtain a condition that uses
only functions of which properties like signs of …rst and second derivatives are known.
Step #2: Evolution of costate variable. At …rst we set up the maximized
Bellman equation. This is obtained by replacing the control variable in the Bellman
equation by the optimal level of the control variable according to the FOC ct = ct (xt ),
19
which gives
V (xt ) = u(ct (xt )) + V [f (xt ; ct (xt ))]
| {z }
xt+1
This equation is a di¤erence equation in the costate variable, the derivative of the
value function w.r.t. the state variable, i.e. V 0 (xt ). The costate variable is also called
the shadow price of the state variable xt . It says how much an additional unit of the
state variable (e.g. wealth) is valued: As V (xt ) gives the value of optimal behavior
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, equ. (21) describes how the shadow
price of the state variable changes over time when the agent behaves optimally.
u0 (ct 1 )
Step #3: Euler equation. Noting the FOC (20) one may write V 0 (xt ) = fc
u0 (ct )
and V 0 (xt+1 ) = fc
. Inserting into equ. (21) gives
20
2.3 Control Theory
As a preparation for the heuristic proof of the maximum principle, we repeat the
method of Lagrange multipliers.6 Consider the following constrained maximization
problem (equality constraint)
We assume that a local maximum exists at x = (x1 ; x2 ) and derive the conditions
that describe this solution. Totally di¤erentiating the constraint g(x1 ; x2 ) = b gives
@g @g
dg = dx1 + dx2 = 0,
@x1 @x2
dx2 @g=@x1
= .
dx1 @g=@x2
dh @g=@x1
This relation implies an equation of the form x2 = h(x1 ) with dx1
= @g=@x2
. Using
x2 = h(x1 ) one can now state the above constrained optimization problem in two
variables as an unconstrained optimization problem in one variable
@F @F dh
+ = 0.
@x1 @x2 dx1
6
This section follows Intriligator (1971, Chapter 3.2)
21
dh @g=@x1
Substituting dx1
by @g=@x2
gives
@F @F=@x2 @g
= 0. (23)
@x1 @g=@x2 @x1
@F @F=@x2 @g
= 0. (24)
@x2 @g=@x2 @x2
@F=@x2
De…ning = @g=@x2
the FOC may be expressed as
@F @g
=0 8 i 2 f1; 2g. (25)
@xi @xi
@F=@x1 @g=@x1
= . (26)
@F=@x2 @g=@x2
This is the usual FOC for the initial constrained problem with two variables (the slope
of the contour of the objective function must equal the slope of the constraint).
Now, the set of necessary conditions (25) plus the original constraint g(x1 ; x2 ) = b
can be obtained as the conditions for a stationary point of the function
L = F (x1 ; x2 ) [b g(x1 ; x2 )] .
The conditions for a stationary point (i.e. a point where the gradient is zero) are
given by
@L @F @g
= = 0 8 i 2 f1; 2g
@xi @xi @xi
@L
=b g(x1 ; x2 ) = 0
@
The function L(:) is the Lagrangian function and the variable is known as La-
7 @F @F=@x1 @g @F=@x1 @F=@x2
By symmetry @x 2 @g=@x1 @x2 = 0. In addition, (23) can be expressed as @g=@x1 = @g=@x2 .
Substituting this into the preceding equation gives (24).
22
grange multiplier. Finally, it should be noted that the variable measures the sensi-
tivity of the optimal value of objective function F = F (x1 ; x2 ) to variations in the
@F
constant b, i.e. = @b
(proof: Intriligator, 1971, pp. 36-37).
ZT
max J = I(x; u; t)dt + F (xT ; T )
fu(t)g
t0
s.t. x_ = f (x; u; t)
@H
=0 8 t 2 [t0 ; T ] (28)
@u
_ = @H
8 t 2 [t0 ; T ] (29)
@x
@F
(T ) = (30)
@xT
@H
plus x_ = f (x ; u ; t). @u
denotes a vector of partial derivatives of the Hamiltonian
w.r.t. u. The set of FOCs (28) describes, at each point in time t 2 [t0 ; T ], r equations.
8
This section follows Intriligator (1971, Chapter 14). More textbook treatments can be found
in Barro and Sala-i-Martin (2005), Gandolfo (1996, Chapter 22.1), Silberberg (1990, Chapter 18.2),
Feichtinger and Hartl (1986, Chapter 2), Chiang (1999, Chapter 7), Chiang and Wainwright (2005,
Chapter 20).
9
To simplify, it is assumed that the control trajectory u(t) is unconstrained.
23
The set of FOCs (29) and x_ = f (x ; u ; t) plus boundary conditions - x(t0 ) = x0
@F
and (T ) = @xT
- are called canonical equations. These constitute a system of 2n
di¤erential equations with n initial and n boundary conditions.10
Heuristic proof. Notice at …rst that the Maximum Principle can be considered an
extension of the Method of Lagrange Multipliers to dynamic problems. The expression
to be maximized is the objective functional, as stated in problem (27),
ZT
J= I(x; u; t)dt + F (xT ; T )
t0
f (x; u; t) x_ = 0 8 t 2 [t0 ; T ] :
Recall that the Method of Lagrange Multipliers requires to set up the Lagrangian
function and then to determine the saddle point (maximizing w.r.t. choice variables
and minimizing w.r.t. the Lagrange multipliers) of this function.
Step #1. De…ne the Lagrangian function
ZT ZT
L = I(x; u; t)dt + F (xT ; T ) + [f (x; u; t) _ dt ,
x] (31)
t0 t0
| {z } | {z }
objective function
Lagrange multipliers times constraints
(= mass T t0 constraints)
ZT
= fI(x; u; t) + [f (x; u; t) _ dt + F (xT ; T )
x]g (32)
t0
where (t) = ( 1 (t); 2 (t); :::; n (t)) are called costate variables and represent the dy-
10
In some cases, e.g. the Ramsey model, the control variable(s) u can be eliminated from the
canonical system by employing @H
@u = 0. If this is impossible, the system under study constitutes a
di¤erential-algebraic system.
24
namic equivalent of Lagrange multipliers. Since the constraints and the costate vari-
ables are de…ned over time, the inner product (2nd term on the RHS of (31)) is properly
treated under the integral sign.
For future uses we reformulate the Lagrangian (32) as follows
ZT ZT
L= fI(x; u; t) + f (x; u; t)g dt _ + F (xT ; T )
xdt
t0 t0
RT RT
Noting _ = xjTt0
xdt _ xdt11 one may express the preceding equation as
t0 t0
2 3
ZT ZT
L = fI(x; u; t) + f (x; u; t)g dt 4 xjTt _ xdt5 + F (xT ; T )
0
t0 t0
ZT n o
= I(x; u; t) + f (x; u; t) + _ x dt [ (T )x(T ) (t0 )x(t0 )] + F (xT ; T(33)
):
t0
ZT n o
= H(x; u; ;t) + _ x dt [ (T )x(T ) (t0 )x(t0 )] + F (xT ; T ) , (34)
t0
Step #2. Consider the necessary conditions for a saddle point of (32). A change in
the costate variable trajectory from (t) to (t)+ (t), where (t) is any continuous
11
Rb Rb 0
Recall integration by parts: a f (x)g 0 (x)dx = f (x)g(x)jba a
f (x)g(x)dx. Form the total deriv-
ative of f (x)g(x) to get df (x)g(x) = f 0 (x)g(x)dx + f (x)g 0 (x)dx. Next form the (de…nite) integral
Rb Rb
between a and b, which gives: f (x)g(x)jba = a f 0 (x)g(x)dx + a f (x)g 0 (x)dx.
25
function of time, would change L(:) to
ZT
L= [f (x ; u ; t) x_ ] dt
t0
x_ = f (x ; u ; t) (35)
Obtaining the equations of motions for x as one set of FOCs is completely analogous
to obtaining the constraints in the static problem.
Step #3. Now consider the e¤ect of a change in the control trajectory from u(t)
to u(t) + u(t), associated by a corresponding change in the state trajectory from x(t)
to x(t) + x(t). Considering (34) this yields
ZT
@H @H _ @F
L= u+ + x dt + (T ) xT (36)
@u @x @xT
t0
For a maximum it is necessary that L = 0 implying that (28), (29) and (30) must
hold. FOCs (28) state that the Hamiltonian is maximized by choice of controls at each
point in time (conditions for an interior solution). Notice also that FOCs (35) can also
be expressed in terms of the Hamiltonian
@H
x_ = 8 t 2 [t0 ; T ] : (37)
@
Comments
1. Su¢ ciency conditions. Provided that either the Hamiltonian is jointly concave
in the control and the state variable (Mangasarian su¢ ciency conditions) or that
the maximized Hamiltonian is concave in the state variable (Arrow su¢ ciency
conditions), the necessary conditions are also su¢ cient; see Kamien and Schwartz
(1981, part II section 3 and section 15).
26
2. Costate variables. The costate variables have the following interpretation:
@J
@x(t0 )
= (t0 ). The (initial) costate variables give the change in the optimal
value of the objective functional due to changes in the corresponding initial state
variables. This is analogous to the interpretation of the static Langrange multi-
pliers.
Setup. Consider an in…nitely lived agent who is endowed with one unit of time each
period and can accumulate human capital by devoting time 0 u 1 to education.
Let h denote the level of human capital and w the wage rate per unit of human capital
and period of time. Income net of education costs (forgone wages) is (1 u)wh. The
agent is assumed to maximize welfare
Z 1
c1 1 t
max e dt (38)
fug 0 1
s.t. c = (1 u)wh (39)
27
read as
[(1 u)wh]1 1
H= + Au h hh
1
@H 1
= [(1 u)wh] ( wh) + Au h =0
@u
_ = @H 1
+ = [(1 u)wh] (1 u)w Au h h +
@h
Steady state. Employ the _ -equation to get
^= w 1
[(1 u)wh] (1 u) Au h h
w Au 1 h 1
Noting = [(1 u)wh]
gives
^ = Au 1 h 1 1
[(1 u)wh] (1 u) Au h h
[(1 u)wh]
1 1
= (1 u) Au h Au h 1 h .
1 1
(1 u) Au
| {zh } h }1 +
|Au {z h =0
= h =u = h
1 u
h h + h =0
u
h
u~ = (41)
h + ( 1) h
~= A~
u 1
h .
h
28
2.5 Application (2): TVC in the Ramsey model
@Y
Let the production technology be denoted by Y = F (K) = AK L1 and let YK := @K
.
The reduced-form dynamic system for the Ramsey model may then be expressed as:
C_ 1
= (YK )
C
K_ = Y K C
K(0) = K0 > 0
t
lim e (t)K(t) = 0 (TVC)
t!1
The phase diagram, shown in Fig. 5, illustrates that there are three steady states
(de…ned by K_ = C_ = 0):
n o
steady state 1: K~ 1 = 0; C~1 = 0
1
1
steady state 2: ~2 =
K A
L; C~2 = F K
~2 ~2
K
+
n 1
o
steady state 3: K~3 = A 1
L; C~3 = 0 .
~ 2 . Steady
Figure 5 displays also three trajectories, starting at some 0 < K0 < K
state 1 is unstable, steady state 2 is saddle-path stable, while steady state 3 is condi-
29
tionally stable, but economically not relevant. The reason for this irrelevance of steady
state 3 is that it violates the transversality condition (TVC), as will be shown in the
following. By making use of the FOC, =C , the TVC can be rewritten as
t K(t)
lim e = 0: (42)
t!1 C(t)
The TVC holds if, and only if, the expression under consideration has asymptotically
a negative growth rate, i.e.
h i
lim ^
+ K(t) ^
C(t) = ^
+ lim K(t) ^ < 0:
lim C(t) (43)
t!1 t!1 t!1
^
Therefore, we need to determine limt!1 K(t) ^
and limt!1 C(t), assuming that the
^
dynamic system converges to steady state 3. Consider limt!1 K(t) at …rst
^ = lim Y (t)
lim K
K(t)
lim
C(t)
= 0: (44)
t!1 t!1 K(t) t!1 K(t)
^ = lim 1
lim C(t) (YK ) : (45)
t!1 t!1
With the help of (44) and (45) we can rewrite the TVC (43) to read
0> ^
+ lim K(t) ^
lim C(t) (46)
t!1 t!1
1
0> lim (YK ) (47)
t!1
30
It can be easily shown that steady state 3 does indeed violate limt!1 [YK ] > 0. To
see this, notice that the K_ = 0 locus is given by C = Y K such that
@C 1
= YK = AK L1 (50)
@K
1
Evaluating @C ~3 =
at K = K A 1
L gives
@K
1
! 1
@C A 1
= A L L1 (51)
@K ~3
K=K
1
A 1
= A (52)
= (53)
= ( 1) < 0 (54)
Taken together, the TVC, assuming that the economy converges to steady state 3,
implies limt!1 [YK ] > 0 . However, steady state 3 lies, by construction, on the
K_ = 0 locus implying that @C
@K
= YK < 0. This contradiction means that steady
state 3 violates the TVC and hence is economically irrelevant.
Charles Ponzi became known 1920s as a swindler in for his money making scheme. He promised clients huge pro…ts
by buying discounted postal reply coupons in other countries and redeeming them at face value in the US as a form of
arbitrage. In reality, Ponzi was paying early investors using the investments of later investors. This type of scheme is
31
now known as a "Ponzi scheme". (Wikipedia, June 3rd 2013)
Consider an agent who lives for a …nite time period and has access to the capital
market. Let a(t) denote …nancial wealth at time t (time may be continuous or discrete).
The No-Ponzi-Game condition (NPGC) then reads as follows
a(T ) 0;
i.e. a(T ) < 0 is excluded. The economic signi…cance is as follows. The No-Ponzi-Game
condition (NPGC) represents an equilibrium constraint that is imposed on every agent.
Everyone must repay his/her debt, i.e. leave the scene without debt at terminal point
in time. In case of an in…nitely lived agent, the NPGC reads as follows (time is
continuous):
rt
lim a(t)e 0;
t!1
rt
i.e. limt!1 a(t)e < 0 is excluded.
Current-value debt under Ponzi game. To see the economic signi…cance,
assume that Mr. Ponzi (and his dynasty) wishes to increase consumption today by
x (with x being measured in monetary units). Consumption expenditures are being
…nanced by borrowing money. Debt repayment as well as interest payments are being
…nanced by increasing indebtedness further. Debt then evolves according to
t 0 1 2 ...
current value debt x x(1 + r) x(1 + r)2 ...
dt = x(1 + r)t if t 2 Z
d(t) = xert if t 2 R
32
ymptotic level of present value debt reads
xert e
lim |{z} rt
= x > 0:
t!1
d(t)
rt rt
lim d(t)e >0 () lim a(t)e < 0:
t!1 t!1
3 Dynamic systems
1. Di¤erential equations (DE) are functional equations. That is, the solution of a
DE y(t)
_ = f [y(t)] with y(0) = y0 is itself a function of time, i.e. y = g(t; y0 ),
that satis…es the law of motion y(t)
_ = f [y(t)] at any point in time.
2. If the DE under study satis…es the Lipschitz condition, then there exists a unique
solution.12 This implies that if a solution can be found, we know that this solution
is the unique solution.
y_ + ay = b, (55)
12
Existence and uniqueness for initial value problems (Intriligator, 1971, p. 469): Consider y_ =
f (y; t) with y(0) = y0 . If f (y; t) is continuos in the region under consideration and satis…es the
Lipschitz condition that for any two values y1 and y2 there exists a …nite positive constant k such
that jf (y1 ; t) f (y2 ; t)j kjy1 y2 j, then there exists a unique solution y(t). Intuitively, a Lipschitz
continuous function is limited in how fast it can change: there exists a de…nite real number such
that, for every pair of points on the graph of this function, the absolute value of the slope of the line
connecting them is not greater than this real number.
13
This section closely follows Chiang and Wainwright (2005, Chapters 15/16)
33
dy
where t 2 R denotes the independent variable (often the time index) and y_ := dt
.
This is a …rst-order DE since only the …rst derivative of y with respect to t occurs.
Moreover, it is linear since both y_ and y appear only in …rst degree and there is no
such term yy.
_
Homogenous case. Assuming b = 0 8 t we have y_ + ay = 0. A solution is readily
1 dy
found by rewriting y dt
= a and integrating both sides w.r.t. t to get
Z Z
1 dy
dt = adt:
y dt
Next we form the inde…nite integral on both sides (notice that the chain rule implies
1 dy d ln y
y dt
= dt
)
Z
d ln y
LHS: dt = ln y + c1
dt
Z
RHS: adt = at + c2
=) ln y = at + c with c := c2 c1
at
y(t) = Ae (general solution)
at
y(t) = y0 e (de…nite solution),
where the de…nite solution results from the determination of the arbitrary constant A
by exploiting a border condition like y(0) = y0 .
Non-homogenous case. Assuming b 6= 0 8 t in (55) we have y_ + ay = b. The
solution consists of the sum of the complementary function yc and the particular inte-
gral yp . The complementary function yc is the general solution of the homogenous DE
(containing arbitrary constants), whereas the particular integral yp is simply any par-
ticular solution of the non-homogenous equation (containing no arbitrary constants).
34
From above we know that
at
yc = Ae
As regards yp , we can try the simplest possible solution, i.e. y = k. The DE then
becomes ay = b (since y_ = 0) such that yp = b=a (assuming a 6= 0).
The general solution of the non-homogenous equation then is
b at
y = yc + yp = + Ae :
a
The de…nite solution is found by exploiting a border condition, y(0) = y0 , to pin down
b
the arbitrary constant of integration A: y = y0 a
e at
+ ab .
y• + a1 y_ + a2 y = b;
35
Hence, the homogenous equation may be expressed as
r2 + a1 r + a2 Aert = 0
r 2 + a1 r + a2 = 0 (if A 6= 0)
As there are two solutions for r, there are two solutions which satisfy the homogenous
DE, namely
y1 = A1 er1 t and y2 = A2 er2 t .
Case 2: repeated real roots. Provided that (a1 )2 = 4a2 we have repeated real
a1
roots r1 = r2 = 2
(= r). Forming yc = A1 ert + A2 ert = A3 ert does not work (recall:
it requires two arbitrary constants to receive a unique solution). Hence, we try another
solution satisfying y•+a1 y_ +a2 y = 0 which is linearly independent of A3 ert . A candidate
36
is y = A4 tert implying (product rule)
r2 t + 2r + a1 (rt + 1) + a2 t A4 ert = 0
Notice that r = a1
2
and (a1 )2 = 4a2 implies (r2 t + 2r) + a1 (rt + 1) + a2 t = 0
such that y = A4 tert does indeed qualify as a solution. Therefore, the complementary
function reads
yc = A3 ert + A4 tert :
r1;2 = h vi;
p
a1 4a2 (a1 )2 14
where h = 2
and v = 2
. With the help of Euler’s formula one can
transform any imaginary exponential function into an equivalent linear combination of
sine and cosine functions: eh iv
= eh (cos vt i sin vt).15 The solution yc = A1 e(h+vi)t +
A2 e(h vi)t
= eht [A1 eivt + A2 e ivt
] can now be expressed as
37
Stability, in general, requires that the real part of every characteristic root is nega-
tive. (This is the concept of unconditional stability, below we will discuss the concept
of conditional stability.)
The system of two linear DE can easily be transformed into a 2nd order DE. Dif-
ferentiating (56) w.r.t. t and substituting y_ 2 according to (57) and y2 according to (56)
yields
a22
y•1 = a11 y_ 1 + a12 y_ 2 = a11 y_ 1 + a12 a21 y1 + (y_ 1 a11 y1 )
a12
= (a11 + a22 ) y_ 1 (a11 a22 a12 a21 ) y1 (58)
38
(case 1 above) such that the solution may be expressed as
y_ 1 a11
The solution y2 is readily obtained by solving (56) for y2 , i.e. y2 = a12
y.
a12 1
y_ 1 a11
y2 = y1 (61)
a12 a12
1 a11
= A1 r1 er1 t + A2 r2 er2 t A1 er1 t + A2 er2 t (62)
a12 a12
r1 r1 t r2 r2 t a11 a11
= A1 e + A2 e A1 er1 t A2 er2 t (63)
a12 a12 a12 a12
r1 a11 r 2 a11
= A1 er1 t + A2 er2 t (64)
a12 a12
If the roots are real and equal (case 2 above), then the solutions reads
(r a11 ) A3 + A4 r a11
y2 = + A4 t ert (66)
a12 a12
Provided that the roots are complex conjugate r1;2 = h vi (case 3 above), the
solution is given by
y1 = eht [A5 cos vt + A6 sin vt]
(h a11 ) A5 + vA6 (h a11 ) A6 vA5
y2 = eht cos vt + sin vt (67)
a12 a12
39
Figure 6: Phase diagramms. Left panel: stable focus; right panel: saddle point.
The time index t can now take only integer values, i.e. t 2 Z.18 As t = 1 the
di¤erence quotient y= t can be stated simply as y.
3.2.1 First-order, linear di¤erence equation (constant coe¢ cient and term)
A …rst-order di¤erence equation is one that contains the …rst di¤erence yt := yt+1 yt .
A …rst-order di¤erence of yt is transformable into a sum of terms involving a one-
period time lag and we de…ne a …rst-order di¤erence equation as one that contains a
one-period lag in the dependent variable. Consider a linear, homogenous, …rst-order
di¤erence equations with constant coe¢ cients (m; n 6= 0)
n
Iterative method. Equation (68) may be expressed as yt+1 = y.
m t
Provided that
y0 is given, the solution can be developed by iteratively applying the preceding rule to
18
This section closely follows Chiang and Wainwright (2005, Chapters 17,18)
40
get
n
y1 = y0
m
n n 2
y2 = y1 = y0
m m
n n 3
y3 = y2 = y0
m m
:::
n t
yt = y0
m
where a and c are constants. The general solution comprises the particular solution yp
- any solution of the complete non-homogenous equation (69) - and the complemen-
tary solution yc - the general solution of yt+1 + ayt = 0. The yp part represents the
intertemporal equilibrium, while yc gives the deviation from this stationary solution.
Consider yc …rst. As a solution for yt+1 + ayt = 0 we try yt = Abt (with A; b 6= 0).
This implies yt+1 = Abt+1 such that yt+1 + ayt = 0 becomes
Hence, for the initial trial solution yt = Abt to work we must set b = a implying
yt = A ( a)t .
Next we turn to yp , which is any solution to the complete equation. As a trial
solution we chose the simplest example, i.e. yt = k (implying yt+1 = k). Hence we
41
have
c
k + ak = c =) k=
1+a
Since this particular k value satis…es (69), the particular integral can be written as
c
yp = ; a 6= 1
1+a
c
If it happens that a = 1, the particular solution yp = 1+a
is not de…ned. In this
case, one should try a solution of the form yt = kt, which gives
c
k(t + 1) + akt = c =) k= =c
t + 1 + at
and hence
yp = ct
c
yt = yc + yp = A ( a)t +
1+a
c c
y0 = A + =) A = y0
1+a 1+a
c c
yt = y0 ( a)t +
1+a 1+a
42
3.2.2 Second-order, linear di¤erence equation (constant coe¢ cient and
term)
2
yt = ( yt ) = (yt+1 yt )
= (yt+2 yt 1 ) (yt+1 yt )
= yt+2 2yt 1 + yt
Particular solution. We try a solution of the form yt = k. Plugging this into (70)
c
gives k + a1 k + a2 k = c or k = 1+a1 +a2
. Assuming a1 + a2 6= 1 yields
c
yp = .
1 + a1 + a2
If, on the other hand, a1 + a2 = 1, then the trial solution should be yt = kt such
c
that k(t + 2) + a1 k(t + 1) + a2 kt = c or k = a1 +2
. In this case one obtains "a moving
equilibrium"
c
yp = t.
a1 + 2
Complementary solution. This is the solution to
43
Abt . The preceding equation then becomes
Assuming that Abt 6= 0 the preceding equation implies b2 +a1 b+a2 = 0. This quadratic
equation in b has two roots
p
a1 a21 4a2
b1;2 = :
2
Both b1 and b2 should appear in the solution yc , because the general solution to
(71) must indeed consist of two linearly independent parts. Three cases must be dis-
tinguished:
1. When a21 > 4a2 , the square root is a real number and b1 and b2 are real and
distinct. The terms bt1 and bt2 are linearly independent and yc can be written as
yc = A1 bt1 + A2 bt2 .
2. When a21 = 4a2 , the square root vanishes and the characteristic roots are repeated
a1
b1 = b2 = 2
. In this case, we have yc = A3 bt + A4 tbt .
3. When a21 < 4a2 , the characteristic roots are conjugate complex: b1;2 = h vi
p
4a2 a21
with h = 2a1 v = 2
. In this case, we get yc = A1 bt1 + A2 bt2 = A1 (h + vi)t +
A2 (h vi)t . Using De Moivre’s theorem, the solution can be transformed into a
p
trigonometric form to read yc = Rt (A5 cos t + A6 sin t) with R = a2 (assumed
positive), A5 = A1 + A2 , and A6 = (A1 A2 ) i.19
44
3.2.3 Two-dimensional system of linear di¤erence equations
The system of two linear di¤erence equations can easily be transformed into a 2nd
order di¤erence equation (cf. to two-dimensional di¤erential equation system). To
sketch the procedure notice that the above system can be transformed into a second-
order di¤erence equation
1 a11
yt = A1 (b1 )t+1 + A2 (b2 )t+1 A1 (b1 )t + A2 (b2 )t : (75)
a12 a12
45
Using (b1 )t+1 = b1 (b1 )t and (b2 )t+1 = b2 (b2 )t we get
1 a11
yt = A1 b1 (b1 )t + A2 b2 (b2 )t A1 (b1 )t + A2 (b2 )t (76)
a12 a12
A1 b1 (b1 )t A2 b2 (b2 )t a11 A1 (b1 )t a11 A2 (b2 )t
= + (77)
a12 a12 a12 a12
b1 a11 b2 a11
= A1 (b1 )t + A2 (b2 )t : (78)
a12 a12
Figure 7: Phase diagramms. Left panel: stable focus; right panel: saddle point.
y_ 1 = f (y1 ; y2 ) (79)
y_ 2 = g (y1 ; y2 ) ; (80)
where f (:) and g (:) are at least twice continuously di¤erentiable functions. The system
is assumed to posses a steady state (~
y1 ,~
y2 ) de…ned by f (~
y1 ; y~2 ) = 0 and g (~
y1 ; y~2 ) = 0.
46
Applying a …rst-order Taylor series approximation gives
@f (y1 ; y2 ) @f (y1 ; y2 )
y_ 1 = f (y1 ; y2 )jy1 =~y1 + (y1 y~1 ) + (y2 y~2 ) (81)
y2 =~
y2 @y1 y1 =~
y1 @y2 y1 =~
y1
y2 =~
y2 y2 =~
y2
@g (y1 ; y2 ) @g (y1 ; y2 )
y_ 2 = g (y1 ; y2 )jy1 =~y1 + (y1 y~1 ) + (y2 y~2 ) :
y2 =~
y2 @y1 y1 =~
y1 @y2 y1 =~
y1
y2 =~
y2 y2 =~
y2
@f (~
y1 ;~
y2 ) @f (y1 ;y2 )
Simplifying the notation by writing @y1
instead of @y1 y1 =~
y1
, noting that
y2 =~
y2
the …rst terms on the RHS equal zero and using y_ 1 = (y1 y~1 ) and y_ 2 = (y2 y~2 ) one
receives
@f (~
y1 ; y~2 ) @f (~
y1 ; y~2 )
(y1 y~1 ) = (y1 y~1 ) + (y2 y~2 ) (82)
@y1 @y2
@g (~
y1 ; y~2 ) @g (~
y1 ; y~2 )
(y2 y~2 ) = (y1 y~1 ) + (y2 y~2 ) : (83)
@y1 @y2
This is a system of linear and homogenous di¤erential equations in the variables
y1 y~1 and y2 y~2 . The coe¢0cient matrix is given
1 by the Jacobian matrix of the
@f (~
y1 ;~
y2 ) @f (~
y1 ;~
y2 )
original non-linear system J = @ @y1 @y2 A. The solution to such a system
@g(~y1 ;~
y2 ) @g(~y1 ;~
y2 )
@y1 @y2
may be expressed as (cf. to Section 3.1.3)
47
3.3.2 Proportional deviations from steady state
Consider a non-linear function y = f (x; z) with f (:) at least once continuously di¤eren-
tiable. It may be useful to transform this non-linear function into a relation y^ = g(^
x; z^)
y y~
with g(:) being a linear function and y^ = y~
etc. denote proportional deviations from
the steady state denoted y~, x~ and z~. Linearizing f (x; z) by means of a …rst-order Taylor
approximation around the steady state gives
@f (x; z) @f (x; z)
f (x; z) = f (~
x; z~) + (x x~) + (z z~)
@x x=~
x @z x=~
x
z=~
z z=~
z
@f (x; z) @f (x; z)
= f (~
x; z~) + x~ x^ + z~ z^
@x x=~
x @z x=~
x
z=~
z z=~
z
y = y~ + 1(y y~)
= y~ + y^ y~
@f (x;z) @f (~
x;~
z)
Simplifying the notation by writing @x x=~
x
= @x
, noting that y~ = f (~
x; z~)
z=~
z
and setting the LHS equal to the RHS again, we obtain
@f (~
x; z~) @f (~
x; z~)
y~ + y^ y~ = f (~
x; z~) + x~ x^ + z~ z^
@x @z
@f (~
x; z~) x~ @f (~
x; z~) z~
y^ = x^ + z^:
@x y~ @z y~
Example 1. Consider y = ax z . Applying the above sketched procedure gives20
1 1
y^y~ = a x~ z~ x~ x^ + a x~ z~ z~ z^
20
Alternatively, one can divide y = ax y by y~ = a x
~ y~ to get
y x z
=
y~ x
~ z~
ln y ln y~ = (ln x ln x
~) + (ln z ln z~)
y^ = x^ + z^:
48
y^ = x^ + z^:
1 1
y^y~ = a x~ (x x~) + b z~ (z z~)
1 1
= a x~ x~ x^ + b z~ z~ z^
x~ z~
y^ = a x^ + b z^:
y~ y~
The speed at which an economy converges to its steady state can be measured by the
rate of convergence (ROC). The ROC of any variable x(t) is de…ned by21
x(t)
_
x (t) := :
x(t) x~
If x > 0 ( x < 0), variable x converges to (diverges from) the steady state x~.
Consider a potentially non-linear di¤erential equation, x(t)
_ = f [x(t)], which is assumed
to possess a stationary equilibrium de…ned by f (~
x) = 0. Linearization of f (x) by means
of a …rst-order Taylor approximation around x~ gives
x(t)
_ x) + f 0 (x)jx=~x (x
= f (~ x~) .
x(t)
_
x := = f 0 (x)jx=~x :
x(t) x~
49
Linearizing around k~ gives
h i
k_ = sk~ 1
( + n + g) k k~
" 1 #
s 1
= s ( + n + g) k k~
+n+g
= ( 1) ( + n + g) k k~ :
k_
Hence, the (local) speed of convergence is given by k := k k ~ = (1 ) ( +n+g).
Consider the application in Section 2.4 (education à la Lucas). Assume the agents starts
~ To determine the agent’s (human) wealth (cf. consumption and
with some h0 < h.
permanent consumption hypothesis) the PDV value of wage income, we can linearize
R1
_
the h-equation around the steady state and then apply the formula 0 xe bt dt = xb .
1
1
First-order approximating h_ = Au h ~ ~
h h around h, noting h =
A~
u
h
, yields
h_ = A~ ~
u h 1
h h ~
h
h ~
= A~
u h h h
A~
u
= ( 1) h h ~1
h
h ~ =(
h 1) h h ~
h with h(0) = h0
~ + h0
h=h ~ e(
h 1) ht
Wage income at each instant of time is w(1 u~)h. The PDV value of wage income
50
therefore reads
Z 1 Z 1 h i
w(1 u~)he rt
dt = w(1 u~) h~ + h0 ~ e(
h 1) ht
e rt
dt
0 0
Z 1
= w(1 ~
u~)he rt
+ w(1 u~) h0 ~ e[(
h 1) h r]t
dt
0
~ w(1 u~) h0 ~
h
w(1 u~)h
= +
r (1 ) h +r
1
c( k )
c_ =
k_ = k k c
~ 1 ~ 2
where a11 = k
, a12 = c ( 1)k
, a21 = 1 and a22 = k~ 1
: The solution
may be expressed as
51
0 1
~
k 1 c ( ~
1)k 2
@ A, v11
and v12
are the eigenvectors associated with r1
v21 v22
1 k~ 1
and r2 . It can be shown that one eigenvalue is positive, while the other one is negative,
say r1 > 0 (unstable root) and r2 < 0 (stable root). To determine the two "arbi-
trary constants of integration" B1 and B2 , let us consider the solution to the linearized
system evaluated at t = 0 (notice that e0 = 1)
Remember that B1 is associated with the unstable root (r1 > 0) and B2 is associated
with the stable root (r2 < 0). The impact of the unstable eigenvalue on the dynamic
evolution can only be eliminated by setting c(0) such that B1 ! Technically this means
we set B1 = 0 and solve the above system for c(0) and B2 .
Let the instantaneous net bene…t of a project be denoted by N (t).22 Assume that N (t)
converges at constant rate , i.e. we can write
t
N (t) = N1 + (N0 N1 ) e :
The net present value (NPV) of a project that generates an in…nite ‡ow N (t) may
then be expressed as
Z 1
t rt
NPV = N1 + (N0 N1 ) e e dt
0
Z 1
rt t rt
= N1 e + (N0 N1 ) e e dt
0
N1 N1 N0
= ;
r +r
22
I thank Sjak Smulders for pointing my attention to this application.
52
where r denotes the discount rate. The internal rate of return is de…ned by N P V = 0
such that
N1 N1 N0 N1
=0 ) r= .
r +r N0
Recall the Solow model, described by
t
y = y~1 + (~
y0 y~1 ) e
s1 1
with = (1 ) ; y~1 =
s0
Assume, the economy is initially in a steady state y~0 = 1
. The government
may ask for the rate of return of an investment program that aims at increasing the
gross investment rate. More speci…cally, what is the rate of return of a permanent
increase in the gross saving rate from s0 to s1 ? To answer this question, consider the
associated net bene…t
bene…t costs
z }| { z}|{
N (t) = y y~0 s0 s^y
= (1 s0 s^) y y~0
s1 s0
where s^ := s0
. The initial net bene…t, obtained by setting y = y~0 , is given
by N0 = s0 s^y~0 (initial investment costs). The …nal net bene…t, obtained by setting
y~1 y~0
y = y~1 , reads N1 = (1 s0 s^) y~1 y~0 . Let y^ = y~0
, then the rate of return may be
expressed as
53
3.4 Miscellaneous
1 dy
Homogenous case. Consider y_ + u(t)y = 0, which may be expressed as y dt
= u(t).
Integrating both sides gives
Z Z
1 dy d ln y
LHS: dt = dt = ln y + c
y dt dt
Z Z
RHS: u(t)dt = u(t)dt
Z
=) ln y = c u(t)dt
R R
=) y = e ce u(t)dt
= Ae u(t)dt
which implies
_
y_ = A(t)e bt
A(t)be bt
1
Plugging the RHS into y_ + by = a0
g(t) yields
_ bt 1
A(t)e A(t)be bt + bA(t)e bt = g(t)
| {z } | {z } a0
y_ y
_ bt 1 _ = 1 g(t)ebt
A(t)e = g(t) =) A(t)
a0 a0
23
We follow Gandolfo (1997, Chapter 12.2.6), who applies the method of variation of parameters.
54
Integrating both sides gives
Z
1
A(t) = g(t)ebt dt + B
a0
R R R
y = A(t)e u(t)dt
=) _
y_ = A(t)e u(t)dt
u(t)A(t)e u(t)dt
R R R
_
A(t)e u(t)dt
u(t)A(t)e u(t)dt
+ u(t)A(t)e u(t)dt
= w(t)
| {z } | {z }
y_ y
R R
_
A(t)e u(t)dt
= w(t) =) _ = w(t)e
A(t) u(t)dt
R
u(t)dt
and plugging the RHS into the trial solution y = A(t)e gives (see also Chiang,
1984, pp. 487/488) Z R R
u(t)dt u(t)dt
y= w(t)e dt + B e
where B is an arbitrary constant. Notice that this solution, again, consists of two
additive components.
55
3.4.2 Application 1: Wealth in a small open economy
Consider the open economy Ramsey model and assume that r = . The dynamics of
per capita wealth a(t) is governed by a(t)
_ (r n)a(t) = w c(t) with w and r constant
r
and c(t) may be expressed as c(t) = c(0)e( )t . The solution a(t) can be inferred from
the solution of …rst-order, linear di¤erential equations with variable coe¢ cients and
variable term
Z R R
u(t)dt u(t)dt
y(t)
_ + u(t)y(t) = w(t) has solution
! y(t) = w(t)e dt + B e
Z R R
(r n)dt (r n)dt
a(t)
_ (r n)a(t) = w c(t) has solution
! a(t) = (w c(t)) e dt + A1 e :
(86)
R r
Noting e (r n)dt
=e (r n)t
for r and n constant and c = c(0)e( )t one gets
Z Z
r
a(t) = we (r n)t
dt c(0)e( )t e (r n)t
dt + A1 e(r n)t
r (1 )r + n R
Rewrite (r n) = and use xeat dt = a1 xeat + c to get
w 1 r
a(t) = e (r n)t + A1 c(0)e( )t e (r n)t + A2 A3 e(r n)t
(r n) b
w 1 r
= e (r n)t c(0)e( )t e (r n)t + B e(r n)t (87)
r n b
(1 )r + n
where b := < 0. This equation contains two unknowns, namely B :=
A 1 + A2 A3 and c(0). To determine these two unknowns we exploit two boundary
(r n)t
conditions, namely a(0) = a0 and lim a(t)e = 0 (resulting from NPGC together
t!1
r
with TVC). The latter, assuming r n > 0 and (r n) < 0, gives
(r n)t
lim a(t)e =B=0
t!1
56
Turning to c(0), notice that B = 0 together with a(0) = a0 implies
w 1
a0 = c(0) (88)
r n b
w
c(0) = b a0 + (89)
r n
Remark #1: Notice that the PDV of consumption may be expressed as
Z 1
r 1 1 1
c(0)e( )t e
1 1
(r n)t
dt = c(0)e b 1 c(0)e b 0 = c(0)
0 b b b
R1 r
Hence (88) requires that 0
c(0)e( )t e (r n)t
dt = a0 + w
r n
, i.e. the PDV of con-
sumption must equal total wealth.
Remark #2: Equ. (87) together with B = 0 and (89) gives
c(0)
z }| {
w w
1 ( r )t
a(t) = + b a0 + e
r n r n b
w r w
= a0 + e( )t
r n r n
Capital market equilibrium and implied asset prices. Consider a world with
two assets: Land yields a rent of vZ and has price pZ . The second asset (bond issued
by the government or …rms) yields a real rate of return of r and its price is normalized
to one. In equilibrium arbitrage opportunities vanish such that
p_Z + vZ = rpZ .
It can be shown that the di¤erential equation p_Z (t) r(t)pZ (t) = vZ (t) has the
57
following forward-looking solution
Z1
R( ) R( )
pZ (t) = vZ ( ) e d + lim pZ ( ) e :
|!1 {z }
|=t {z } bubble
fundamental price
R
with R( ) := t
r(s)ds.24 Notice that vZ ( ) walks along the time axes ( ) starting at
R( )
= t, running until = 1. At each point in time , vZ ( ) is weighted by e ,
where the cumulative discount rate R( ) is the sum of all instantaneous interest rates
r(s) between s = t (the starting date) and s = (the terminal date). Notice that
is the terminal date for the calculation of R( ) and, at the same time, denotes the
current point in time w.r.t. vZ ( ). Remarks:
vZ
2. Assume that vZ and r are constant. Then, as long as r > 0, we have pZ = r
.
As r falls, pZ rises. And limr!0 pZ = 1. Summers (2014) has stressed that
r declined over time. Notice also that pZ = vrZ does not hold for negative r:
Z1
pZ (t) = vZ e r( t) d = 1r vZ e r1 1
v e r0 = 1 for r < 0.25
r Z
=t
The math. Consider an economy with two assets: an asset that yields a payo¤ each
period and a bond paying an interest rate. Let ( ) denote the payo¤ at time 2 [t; T ]
and let r(s) denote the interest rate at time s 2 [t; T ]. The present value of the payo¤
R 1 r(s)ds
in period = 1, from the perspective of period t, then is ( 1 )e t , where
R 1 r(s)ds
e t represents the cumulative discount factor between t and 1; in case where
Z R R
24
Recall that y(t)
_ + u(t)y(t) = w(t)has solution
! y(t) = w(t)e u(t)dt dt + B e u(t)dt .
25
Reminder : r~ = ( s 1) in Solow model without population growth and technical change. That
is, dynamic ine¢ ciency s > implies r~ < 0.
58
r( t)
r(s) = r, the discount factor simpli…es to e 1
. The fundamentally warranted price
of an asset, at time t, that entitles the holder to a stream of earnings ( ) 8 2 [t; T ]
is then given by Z T R
r(s)ds
v(t) = ( )e t d . (90)
t
We now di¤erentiate this equation w.r.t. time to receive the no-arbitrage condition in
the capital market26
=0
zZ }| {
t
RT r(s)ds @t Z T R
r(s)ds
r(s)ds @T @ ( )e t
v_ = (T )e t (t)e t + d
@t
|{z} @t
|{z} t @t
=0 =1
The integrand of the last expression on the RHS can be expressed as27
R
r(s)ds R R
Z
@ ( )e t @ r(s)ds r(s)ds @
= ( ) e t = ( )e t r(s)ds
@t @t @t t
To evaluate the last expression on the RHS we apply the Leibniz rule again
0 1
Z Z
@ B @ @t @r(s) C
r(s)ds = @r( ) r(t) + dsA = r(t)
@t t @t
|{z} @t
|{z} t | @t
{z }
=0 =1 =0
59
under consideration (say a "share") provides a gain of (t) + v_ (instantaneous pro…t
plus change in asset price). This must equal the reward that can be earned by investing
the amount v(t) into bonds which pays r(t)v(t).
Remark. Di¤erentiating equ. (90) w.r.t. time yields the no-arbitrage condi-
tion (91). However, solving equ. (91), which represents a …rst-order, linear, non-
homogenous DE with variable coe¢ cient, yields equ. (90) provided that the so-called
transversality condition holds (Brunnermeier, Bubbles, New Palgrave of Economics,
2008, p. 4)28
R
r(s)ds
lim v( )e t = 0:
!1
Basic concepts. The dynamic systems under consideration typically exhibit perma-
nent growth. For several reasons, the variables are transformed so that the dynamic
system possesses a stationary point (most important reason: stability analysis). A
widely employed technique is scale adjustment (or normalization of variables). Assume
R
28
(CHECK) Multiplying both sides of v_ r( )v( ) = ( ) by e t r(s)ds and integrate forward
to get
Z1 R Z1 R
r(s)ds
e t [v_ r( )v( )] d = e t r(s)ds ( )d (92)
t t
h R i1 Z1 R
r(s)ds r(s)ds
e t v( ) = e t ( )d (93)
t
t
The integration on the LHS can be understood by noting that (Leibniz rule of di¤erentiation)
R
@ h R
r(s)ds
i @e t
r(s)ds R
r(s)ds
e t v( ) = v( ) + e t v(
_ )
@ @ R R
r(s)ds r(s)ds
= r( )e t v( ) + e t v(
_ )
Hence, we have
R
Z1 R
r(s)ds r(s)ds
lim e t v( ) v(t) = e t ( )d (94)
!1
t
R
r(s)ds
If the so-called TVC holds lim !1 e t v( ) = 0 one gets
Z1 R
r(s)ds
v(t) = e t ( )d (95)
t
60
that X(t) grows in the long run at rate g, i.e.
_
X(t)
lim = g:
t!1 X(t)
X(t)
x(t) :=
egt
where x~ denotes a constant. With these concepts at hand the steady state growth path
(SSGP) can be expressed as follows
~
X(t) = x~egt
which shows that x~ represents the level of the SSGP (possibly a¤ected by market
distortions, public policy etc.). The equilibrium growth path (EGP) can now be written
as
X(t) = x(t) egt
|{z} :
|{z}
transitional dynamics component steady state growth component
This demonstrates that the EGP X(t) is obviously composed of two components: (i)
a transitional dynamics component x(t) and (ii) a steady state growth component egt .
Dynamic system in normalized variables. Consider the following dynamic
system (boundary conditions suppressed)
_
X(t) = F (X; Y )
Y_ (t) = G(X; Y )
_
X(t) Y (t)
with limt!1 X(t)
= gX and limt!1 Y (t)
= gY . Assuming that the original dy-
61
namic system ful…lls some not very restrictive requirements (especially linear semi-
homogeneity), there exists an associated (and autonomous) dynamic system in scale-
X Y
adjusted variables x := egX t
and y := egY t
of the following shape (Trimborn, 2006,
Section 2.3)
x_ = F (x; y) x gX
y_ = G(x; y) y gY
4.1 Motivation
In textbooks or in research papers one often …nds statements like the following:
w = f (k) f 0 (k)k
Y
r=
K
Y
w = (1 )
L
62
autarky. Assume that capital market integration induces an increase in the inter-
est rate r by 1 percent (e.g., from 3 percent to 3.03 percent). Labor is immobile.
As a consequence, GDP (Y ) decreases by 1
percent.
The following pages provide the required concepts and techniques to make sure that
statement like these are fully understood.
We often assume that the representative …rm has access to a production technology of
the form
Y = F (K; L)
where Y denotes …nal output (value added), K 0 physical capital, and L 0 the
amount of labor. It is most satisfactory to think of both outputs and inputs as ‡ows:
quantities per unit of time. The K appearing in the production function should be
seen as the number of machine hours per year and L should be interpreted as the ‡ow
of labor services per period of time (K and L can also be interpreted as the stock of
capital and labor). The function F : R2+ ! R+ is twice continuously di¤erentiable
("C 2 ") with FK > 0, FKK < 0, FL > 0, FLL < 0, and FKL = FLK > 0, where
@F (:)
FX := @X
. Moreover, the function F (K; L) is homogenous of degree 2 R+ , i.e.
Y = F ( K; L) with 2 R+
63
4.2.2 Euler’s theorem
FK K + FL L = F (K; L)
@F (:) @ K @F (:) @ L 1
+ = F (K; L)
@( K) | @{z } @( L) @
| {z }
outer derivativeinner derivative
@F (:) @F (:) 1
K+ L= F (K; L)
@( K) @( L)
Setting = 1 yields
FK K + FL L = F (K; L)
Assuming that Y = F (K; L) is linearly homogenous, i.e. exhibits CRS, and de…ning
Y K
y := L
(output per unit of labor) and k := L
(capital per unit of labor), the technology
F (K; L) may be expressed in intensive form to read
Y K
= F ( ; 1) =) y = f (k)
L L
The marginal product of capital may then be expressed as FK (:) = f 0 (k). Proof :
Y F (:)
From L
= L
= f (k) we have
F (:) = f (k)L
1
FK (:) = f 0 (k) L = f 0 (k)
| {z } L
|{z}
outer derivative
inner derivative
The marginal product of labor may be expressed as FL (:) = f (k) f 0 (k)k. Proof :
64
From Y = FK (:)K + FL (:)L (Euler’s theorem) we have
y = FK (:)k + FL (:)
r| {z
+} = f 0 (k)
user cost of capital
w = f (k) f 0 (k)k
L L L
d( K )= K d ln( K )
" := = 0
d( FFKL )= FFKL d ln( FFKL )
L
The EoS indicates by how much (in percentage terms) the ratio K
changes, if the
dL FK
slope of an isoquant is changed (recall: dY = FK dK + FL dL = 0 implies dK
= FL
,
dL
where dK
denotes the MRS). The EoS is a dimensionless measure of the substitutability
between the factors under consideration. The economic importance of this concept
r FK
becomes clearer by recalling that, under cost minimization, w
= FL
. In most economic
applications it is assumed that the EoS is constant.
r L
Notice that " 0: an increase in w
induces an increase in K
. When the relative
price of capital ( wr ) increases, capital intensity K
L
declines (i.e. L
K
rises).30
The EoS may look like a fairly abstract concept. However, it can readily be reasoned
that the value of the EoS may have important economic implications. Two examples
illustrate this point. First, the EoS between man-made capital and natural resources
dx d ln x
29
Notice that d ln x = x . This results from dx = x1 .
30 d( K K
L )= L
Sometimes the EoS is de…ned as follows: " := F F 0 (e.g. Silberberg, 1990, p. 287).
d( FK )= FK
L L
65
Figure 1: Figure 7: Substitution of input factors
is decisive for the feasibility of sustained economic growth under limited supply of a
natural resource. Second, the value of the EoS between K and L is decisive for the
impact of an increase in K on the competitive wage rate.
Y = AK L1
where Y denotes …nal output (value added), A > 0 total factor productivity, K 0
physical capital, L 0 the amount of labor, and 0 1. At a more general
level, a Cobb-Douglas technology is of the following form Y = i Xi. Notice that the
i
P
Cobb-Douglas representation leads to a log-linear form, i.e. ln Y = i i ln Xi.
66
4.3.1 Properties and implications
A( K) ( L)1 = Y
1 Y
r + = AK L1 =
K
Y
w = (1 )AK L = (1 )
L
1
r + = Ak
w = (1 )Ak
dY =Y dY K Y K
Y;K := = = =
dK=K dK Y KY
L L L
d( K )= K d ln( K )
" := = =1
d( FFKL )= FFKL d ln( YYKL )
Proof
1 1
Solve Y = AK L1 for L to yield L = Y 1 A1 K1 (the isoquant with
production level Y ).
1 1
@L 1
The slope is given by @K
= 1
Y 1 A1 K1 .
67
Replacing Y by AK L1 gives
@L L L 1 @L
= ) =
@K 1 K K @K
@L YK
Notice that @K
= YL
(from dY = 0) and hence we have
L 1 YK
=
K YL
Taking the natural logarithm on both sides and forming the derivative w.r.t.
YK
YL
yields
L
d ln( K )
" := =1
d ln( YYKL )
Y Y
K + (1 ) L=Y
K L
dY dK dL
= + (1 ) (96)
Y K L
Y
Proof : Form the total derivative, dY = K
dK + (1 ) YL dL, and then divide
both sides by Y .
p = B (r + ) w1
68
1 (1 )
with B := A (1 ) . Proof. From the …rst-order conditions for a pro…t
maximum one obtains the factor demand schedules to read
Y p Y
r+ =p ) K=
K r+
Y p(1 )Y
w = p(1 ) ) L=
L w
Inserting into the production technology gives
1
p Y p(1 )Y
Y =A
r+ w
1
1 1
Y =A (p Y ) (p(1 )Y )1
r+ w
1 (1 )
p=A (1 ) (r + ) w1
1
where 0 < a < 1, 1< 1 and 6= 0. The EoS is given by " = 1
. Three
cases should be distinguished:
Moreover, provided that 0 < " < 1 ( 1 < < 0) there is weak substitutability. For
1 < " < 1 (0 < < 1) there is strong substitutability. Notice that Y = aL +(1 a)K
69
1
does also represent a CES function with " = 1
(yet this version does not exhibit
CRS).
1
1. The EoS is given by " = 1
. Proof :
1
1 1 a
Solving (97) w.r.t. L gives L = a
Y a
K (the isoquant with pro-
duction level Y ).
1
1
@L 1 1 1 a 1 a 1
= Y K ( 1) K
@K a a a
1
1
@L 1 1 a 1 a 1 a 1
= L + K K ( 1) K
@K a a a
1
@L 1
11 a 1 1 1 a 1 L 1 a
= (L ) K = L K =
@K a a K a
1 1
L L a @L
Solving for K
yields K
= 1 a
1
@K
1
.
70
@L YK
Noting @K
= YL
leads to
1
L YK 1
=c
K YL
1
a
with c := 1 a
1
.
Now take the natural logarithm on both sides and form the derivative of
L YK
ln K
w.r.t. ln YL
to get
L
L 1 YK d ln( K ) 1
ln( ) = ln(c) + ln( ) ) "= =
K 1 YL d ln( YYKL ) 1
1 1 1
[a( L) +(1 a)( K) ] = [a L +(1 a) K ] = [aL +(1 a)K ] = Y
Y
1 z }| { 1 1
Y
YK = [aL + (1 a)K ] 1 (1 a) K 1 = (1 a) (99)
| {z }| inner derivative
{z } K
outer derivative
71
Y
1 z }| { 1 1
Y
YL = [aL + (1 a)K ] 1 a | L
1
{z } = a (100)
| {z }inner derivative L
outer derivative
1 1
Y Y
YK K + YL L = (1 a) K +a L
K L
= (1 a)Y 1 K + aY 1
L
= Y1 Y =Y
h 1 1
i 1
Y =A K + (1 )L (101)
where A > 0, 0 < < 1 and 0. Let r + denote the user cost of capital and
w the wage rate. The equilibrium goods price p is then given by a CES index of
factor prices according to
1
p=A 1
(r + )1 + (1 ) w1 1
@Y h 1 1
i 1
1 1 1
1
= A K + (1 )L K
@K 1
( 1) 1 1 1 1
Notice 1
1= 1
= 1
and 1= = . Hence, we get
@Y h 1 1
i 1
1 1
=A K + (1 )L K
@K
Moreover
h 1 1
i 1
1 1 1
K + (1 )L =Y A
72
Therefore 1
@Y 1 1 1 1 Y
= AY A K =A
@K K
Similarly
1
@Y 1 1 1 1 Y
= AY A (1 )L =A (1 )
@L L
First-order conditions read as follows
1 1
!
1 Y pA 1
r + = pA =) K= Y =A p Y
K r+ r+
1 1
!
1 Y pA (1 ) 1 (1 )
w = pA (1 ) =) L= Y =A p Y
L w w
" 1 1 # 1
1 1 (1 )
Y =A A p Y + (1 ) A p Y
r+ w
Eliminating Y and solving for p gives the goods price in competitive equilibrium
" 1 1 # 1
(1 ) 1
Y = A + (1 ) A p Y
r+ w
" 1 1 # 1
(1 )
p = A + (1 )
r+ w
" #11
1 1
1 r+ w
p = A + (1 )
(1 )
1
= A 1
(r + )1 + (1 ) w1 1
Often one reads statements like "due to complementarity between K and L a higher
stock of K increases the marginal product of L". There are in fact (at least) two uses
of the phrase "complementarity".
73
(1) Complementarity applies (only) under a Leontief technology Y = minfK; Lg
(strict complementarity).31
(2) Complementarity applies if one input factor’s marginal product depends posi-
tively on the level of other input factor(s). Hence, complementarity requires that the
underlying technology is not additively separable.
Example 1
Y = aK + bL; a; b > 0
Example 2
Y = aK + bL ; 0< <1
Example 3
1
Y = (aK + bL ) ; < 1; 6= 0
Consider a sector composed of a continuum of length one of identical …rms, which have
access to the following CES technology
Z N
1
X= x(i) di ;
0
31
Starting from an e¢ cient ratio of input factors, Y = minfK; Lg implies that an increase in one
input factor increases the marginal product of the other input factor from zero to a positive constant.
74
where x(i) denotes the amount of intermediate good of type i 2 [0; :::; N ], N the
’number’of intermediate goods, and 0 < < 1. Pro…t of the typical X-…rm is
Z N
= PX
|{z} p(i)x(i)di
0
sales revenue | {z }
production costs
where P denotes the price of good X and p(i) the price of intermediate good i, respec-
tively. The demand schedule for any xi , i.e. arg max , is given by
xi
1
1
p(i) 1 x(i)
x(i) = X () p(i) = P
P X
The ideal price index of intermediate goods prices, which allows us to write P X =
RN
0
p(i)x(i)di, is obtained as follows. Substituting x(i) by the RHS of the demand
schedule yields
Z N
1
p(i) 1
PX = p(i) di X
0 P
Solving for P gives
Z N
1
P = p(i) 1 di P 1
0
Z N
1
= p(i) 1 di
0
The price index times a quantity index, P X, equals total expenditures for intermedi-
RN
ates goods 0 p(i)x(i)di. Notice that P represents an ideal price index of intermediate
goods prices and represents the price of good X. See also the remarks in index numbers
at the end of this section.
RN
It can easily be checked that P satis…es P X = 0
p(i)x(i)di. At …rst, notice that
an equilibrium must be characterized by a symmetric allocation of x(i) (an asymmetric
75
x(i)-allocation cannot be e¢ cient and the …rst theorem of welfare applies) and hence
Z N
1
x(i)=x8i 1
X= x(i) di =) X = N x(i)
0
Now assume, for the time being, that the price index P is given by
Z N
1
P = p(i) 1 di
0
1
1 1
Under symmetry this implies P = N p(i). Hence, one may write (notice: + =
1)
1 1
PX = N p(i)N x(i) = N p(i)x(i)
which implies that the equilibrium condition = 0 does indeed hold true.
"...the purpose of any function’s parameters...is to help describe the relationship be-
tween the function’s dependent variable and its independent variables, including their
dimension units. Therefore, a function’s parameter dimensions (and values) simply are
whatever they need to be (including roots and powers) to describe the relationship
fully and accurately— including that the left and right side dimensions must match.
Variables must have (understandable) dimensions. Parameters may or may not have
76
dimensions... In the Cobb-Douglas production function, parameter A has dimensions
that force the function’s left and right side dimensions to match. More generally, pa-
rameter A has dimensions that the function needs to describe its assumed relationship
between output and input variables..." (Folsom and Gonzalez, Quarterly Journal of
Austrian Economics, 2005)
Heterogeneous capital. Under what conditions can a consistent meaning be
given to the quantity of capital? Suppose we have a production function which relates
the output of a single commodity to inputs of labor (assumed homogenous) and the
services of several kinds of capital goods. When if ever can the various capital inputs
be summed up in a single index-…gure, so that the production function can be "col-
lapsed" to give output as a function of inputs of labor and "capital-in-general"? This
is sometimes possible, but only in a very narrow class of cases.
Formally, suppose we have a production function Q = F (L; C1 ; C2 ) where Q is a
single output, L is labor, and C1 and C2 are inputs of the services of two distinct kinds
of capital equipment. The question is: when can we write, identically
That is to say, when can we collapse the production function from one having three
variables to one having only two? If this can be done, we would seem to have a right to
call K an index of the quantity of capital. For the purposes of production any pattern
of inputs C1 , and C2 are equivalent so long as they yield the same value of the index
K.
Two examples: (i) Consider the Cobb-Douglas production function Q = Lu C1v C2w .
v w
It is obviously collapsible since we can rewrite it Q = Lu K v+w with K := C1v+w C2v+w .
2
(ii) Consider the production function Q = (L0:5 + C10:5 + C20:5 ) . It can then be col-
2 2
lapsed into Q = (L0:5 + K 0:5 ) with K := (C10:5 + C20:5 ) . (Solow, RES, 1955-56)
Index numbers. The index number problem may be stated as follows. Suppose we
have price data p := (p1 ; : : : ; pN ) and quantity data q := (q1 ; : : : ; qN ) on N commodities.
77
The index number problem is to …nd numbers P and numbers Q such that
X
N
P Q = pq = p n qn
n=1
The elasticity of marginal utility w.r.t. consumption (a measure for the curvature
of the utility function) is
78
The last equality uses u0 (c) = c and u00 (c) = c 1
.
u00 (c) c
The Arrow-Pratt measure of relative risk aversion reads u0 (c)
. Hence, the CIES
utility function exhibits a constant, i.e. independent of the level of consumption,
degree of relative risk aversion.
1 c1
The term " 1 " is motivated as follows. The function 1
is not de…ned for
= 1. In contrast (102) is de…ned for = 1 (it boils down to an expression of
the form 00 , which can be evaluated by applying L’Hopital’s rule) and given by
d(c1 1)
1 c1
u(c) = ln c. Proof : lim c 1 1
= lim d
d(1 ) = lim 1
ln c
= ln c.32
!1 !1 d !1
The primary reason for the frequent use of the CIES utility function lies in the
fact that this utility function is compatible with steady state growth (cf. the
Keynes-Ramsey rule).
79
where 0 is the discount rate and t 2 f0; :::; T g is the discrete time index. If the
time index is continuous, i.e. t 2 [0; :::; T ], intertemporal welfare reads as follows
Z T
t
U= e u[c(t)]dt
t=0
The relation between these two formulations can be seen as follows. Consider the
case where an amount of Ae is being invested into an asset that pays a (nominal)
interest rate of r per period of length 1. Assuming that the interest is compounded
once a period, the value of this asset after t periods is
0 1t
V = A @1 + r
|{z}
A
interest rate for period of length 1
If the interest is compounded m times per period, assuming that the payo¤ is
reinvested immediately, the value of this asset after t years reads33
0 1mt
B r C
B C
V = A B1 + C
@ m
|{z} A
1
interest rate for subperiod of length m
r mt
V = lim A 1 + = Aert
m!1 m
80
If m is …nite, then t is discrete. For instance, assuming m = 2, one gets t 2
f1=2; 1; :::g, while for m = 4, we have t 2 f1=4; 2=4; 3=4; 1; :::g. If, on the other hand,
m ! 1, then t becomes continuous (i.e. t 2 Q+ ).
The case of discounting is, of course, closely related to the cases considered above.
t
The present value of a payment of V taking place in t periods is A = V (1 + r) (with
rt
interest compounding once a period) and A = V e (with instantaneous interest
compounding).
Why should we care about the intertemporal elasticity of substitution (IES)? Answer:
The IES is one important determinant which a¤ects the growth rate of consumption
according to the KRR: c^ = 1 (r ) where 1
is the IES, as will be shown below.
CIES preferences. The present value of an in…nite utility stream is given by
R1
U = t=0 u[c(t)]e t dt where 0 is the discount rate per period of length dt. Assume
that instantaneous utility is of the CIES type such that
Z 1
c(t)1 1 t
U= e dt (103)
t=0 1
c(i)
The IES indicates by how much (in percentage terms) the ratio c(j)
changes, if the
slope of an indi¤erence curve is changed (recall: dU [c(i); c(j)] = Uc(i) dc(i)+Uc(j) dc(j) =
dc(i) Uc(j) dc(i)
0 implies dc(j)
= Uc(i)
, where dc(j)
denotes the MRS). The IES is a dimensionless
measure of the substitutability between the consumption in period i and consumption
in period j.
81
Figure 9: Intertemporal consumption substitution
i
The IES can be evaluated as follows. First, notice that Uc(i) = e c(i) and
j
Uc(j) = e c(j) such that
Uc(j) j
c(j) e c(i) (j i)
= i
= e
Uc(i) c(i) e c(j)
c(i)
Taking the natural logarithm and forming the derivative w.r.t. ln c(j)
yields
Uc(j)
Uc(j) c(i) d ln Uc(i)
ln = ln (j i) =) =
Uc(i) c(j) d ln c(i)
c(j)
c(i)
d ln( c(j) ) 1
IES = Uc(j)
=
d ln Uc(i)
Notice that this result holds for any two c(i) and c(j).
CARA preferences. The IES between any two consumption points t = i and
82
t = j is 2 3
U 1 U
d Uc(j) c(j)
Uc(i)
IES := 4 5
c(i)
c(i) c(i)
:
d c(j) c(j)
Uc(j) c(i)
Forming the derivative of Uc(i)
w.r.t. c(j)
yields
U
d Uc(j)
c(i) (j i) (c(j) c(i))
c(i)
= c(j)e e :
d c(j)
(j i) (c(j) c(i))
(j i) (c(j) c(i)) 1 e e
IES = c(j)e e c(i)
c(j)
(j i) (c(j) c(i))
1 e e c(j) 1
= (j i) e (c(j) c(i))
= :
c(j)e c(i) c(i)
R1 h i h i
t 1 1 1 0
The last equality uses 0
u(~
c)e dt = F (1) F (0) = u(~
c) e u(~
c) e =
u(~
c) t
, where F (t) is the antiderivative (inde…nite integral) of f (t) = u(~
c)e .
If consumption grows at a constant (geometric) rate of growth, i.e. c = c~egt , one
gets
Z 1 Z 1
cegt )1
(~ 1 t c~1 1
U = e dt = e((1 )g )t
e t
dt
0 1 0 1 1
c~1 1
=
(1 )((1 )g ) (1 )
83
It is usually assumed that (1 )g < 0. Otherwise intertemporal welfare would not
be bounded from above implying that usual dynamic optimization techniques cannot
be applied.
Notice that in the case of log utility, u(c) = ln(c), life-time utility in the stationary
steady state (c = c~) would be given by
Z 1
t ln(~
c)
U= ln(~
c)e dt = : (105)
0
If the steady state growth rate of consumption is positive, then life-time utility
reads as follows Z 1
ln(c0 ) + g
U= ln(c0 eg t )e t
dt = 2
: (106)
0
Remark:
Z 1 Z 1
gt t t
U = ln(c0 e )e dt = [ln(c0 ) + gt] e dt (107)
0 0
Z 1 Z 1
t t ln(~
c) g ln(c0 ) + g
= ln(c0 )e dt + gte dt = + 2
= 2
: (108)
0 0
84
R1 (cA (t))1 1 t
and UA = 0 1
e dt enables one to calculate U = UA UB , i.e. the
absolute change in life-time utility resulting from the policy measure under study.
Since this information is not very instructive, one usually transforms U into the
c
associated c0
, i.e. the (hypothetical) permanent, percentage variation in consumption
which is equivalent to U . In Figure 4, cH (t) shows the hypothetical, time-invariant
cH (t) cB (t)
level of consumption which leads to the life-time utility UA such that cB (t)
gives
the (hypothetical) permanent, percentage variation in consumption which is equivalent
to U.
c
The permanent, percentage variation in consumption c0
which is equivalent to the
given change in intertemporal welfare U can be computed as follows. We proceed in
two steps.
First, we calculate the hypothetical, time-invariant level of consumption, denoted
as cH , which delivers the life-time utility UA . This results from (recall (1 )g < ):
Z 1
(cH egt )1 1 t (cH )1 1
UA = e dt =
0 1 (1 ) [(1 )g ] (1 )
85
Solving this equation for cH gives
1
( (1 )g) 1
cH = UA (1 )( (1 )g) + >0
In case g = 0 (i.e. the steady state growth rate is zero) the preceding result simpli…es
to read:35
1
cH = [UA (1 ) + 1] 1 >0
c cH cB
=
c0 cB
c
One can also determine c0
directly, which is conceptually a bit more complicated.
The starting point here is the following equation:
Z 1 1 Z 1 1
((c0 + c)egt ) 1 t (c0 egt ) 1 t
U= e dt e dt
0 1 0 1
| {z } | {z }
welfare under alternative scenario (policy change) welfare under baseline scenario (no policy change)
(109)
The change in intertemporal welfare U is represented to result from a one-time
permanent shift in consumption by c. In other words, c indicates the hypothetical
permanent increase in consumption (the BGP is shifted upwards by c) which is
equivalent to U ! Notice that the unknown in (109) is c.
Z 1 1 Z 1
((c0 + c)egt ) t 1 t
U = e dt e dt
0 1 0 1
Z 1 1 Z 1
(c0 egt ) t 1
e dt + e t dt
0 1 0 1
35 c1H 1
By noting equ. (104) this result can also be determined directly from UA = (1 ) .
86
Z Z
((c0 + c))1
1 1
c01
U = e(g(1 ) )t
dt e(g(1 ) )t
dt
0 1 0 1
((c0 + c))1 1 c10 1
=
1 (g(1 ) ) 1 (g(1 ) )
!
1 1
((c0 + c)) c0 1
=
1 1 (g(1 ) )
1
c (c10 + U( 1)(g(1 ) )) 1
= 1
c0 c0
c
c0
gives the permanent, proportional (per time period) increase of consumption
which is equivalent to U . See the study on changes of the R&D subsidy based on the
Jones (1995) model and the …le ’optimal_capital_subsidization.nb’.
Consider a small open Ramsey economy with r = . The constant level of consumption,
c~, reads36
w
c~ = a0 + ;
r
Life-time utility is then given by
Z 1
c1 1 t c~1 1
e dt = .
0 1 (1 )
Assume that the wage w increases once and for all according to A 1. The
36
See section 3.4.2 with r = and n = 0.
87
associated welfare gain may then be expressed by calculating that satis…es
0 11
0 11 c~(alternative)
c~(baseline) z }| {
z }| { B Aw C
B w C B C
B a0 + C 1 B a0 + C 1
@ r A @ r A
=
(1 ) (1 )
w Aw
a0 + = a0 +
r r
Aw
a0 + Awr
1+ a0 r
= =
a0 + wr 1+ w
a0 r
We are interested in (a0 ). Notice that lima0 !0 (a0 ) = A and lima0 !1 (a0 ) = 1.
Hence, this setup implies that the welfare gain of a once and for all increase in TFP
is independent of and declines with a0 . However, this setup may be too simple to
gain deeper insights. The assumed constancy of r and the implied constancy of c may
appear too restrictive.
88
Mc Graw-Hill, New York. (the respective chapter contains a nice summary on control
theory)
Intriligator, Michael D. (1971), Mathematical Optimization and Economic Theory,
Prentice-Hall, Inc., Englewood Cli¤s, N.J. ("despite its age" an excellent book on
optimisation methods)
Seierstad, Atle and K. Sydsaeter (1987) Optimal Control Theory with Economic
Applications, North Holland.
89
6.4 Stochastic growth
References
[1] Acemoglu Daron (2009), Introduction to Modern Economic Growth, Princeton
University Press, Princeton, London.
[2] Acemoglu, Daron, Philippe Aghion, Leonardo Bursztyn, and David Hemous. 2012.
"The Environment and Directed Technical Change.", American Economic Review,
102(1): 131-66.
[3] Barro, Robert J. and Xavier Sala-i-Martin (2005), Economic Growth, 2nd Edition,
The MIT Press, Cambridge, London.
[4] Blanchard, Oliver Jean and Fischer, Stanley (1989), Lectures on Macroeconomics,
The MIT press, Cambrdige, Oxfordshire.
[5] Caselli,Francesco and Jaume Ventura, (2000), "A Representative Consumer The-
ory of Distribution", The American Economic Review, Vol. 90, No. 4, pp. 909-926.
90
[6] Chiang, A. C. (1999), Elements of Dynamic Optimization, McGraw Hill, New
York.
[7] Chiang, Alpha C. and Kevin Wainwright (2005), Fundamental Methods of Math-
ematical Economics, McGraw Hill, New York.
[9] Diewert, W.E., Index Numbers, in: Durlauf, Steven N. and Lawrence E. Blume
(eds.), The New Palgrave Dictionary of Economics, Second Edition, Volume 4,
Palgrave Macmillan, 2008, pp. 190-214.
[10] Einstein, Albert (1934), "On the Method of Theoretical Physics", Philsophy of
Science, Vol. 1, No. 2, pp. 163-169.
[11] Feichtinger, Gustav and Richard F. Hartl (1986), Optimale Kontrolle ökonomis-
cher Prozesse, Anwendung des Maximumprinzips in den Wirtschaftswis-
senschaften, de Gruyter, Berlin.
[12] Folsom, Roger N, and Rodolfo A. Gonzalez (2005), "Dimensions and Economics:
Some Answers", The Quarterly Journal of Austrian Economics, Vol 8. No.4, pp
45-65.
[16] Jones, Charles I. (1995), "R & D-Based Models of Economic Growth", The Journal
of Political Economy, Vol. 103 Issue 4, pp. 759-784.
91
[17] Kamien, Morton I. and Nancy L. Schwartz (1981), Dynamic Optimization, The
Calculus of Variations and Optimal Control in Economics and Management,
North-Holland, New York.
[18] Mas-Colell, Andreu, Whinston, Michael D. And Jerry R. Green (1995), Microeco-
nomic Theory, Oxford University Press, New York.
[19] McKenzie, Lionel W., General Equilibrium, in: Durlauf, Steven N. and Lawrence
E. Blume (eds.), The New Palgrave Dictionary of Economics, Second Edition,
Volume 3, Palgrave Macmillan, 2008, pp. 586-604.
[20] Popper, Karl, 1992, The Logic of Scienti…c Discovery, pp. 121-132.
[21] Robinson, J. (1962), Essays in the Theory of Economic Growth, London: Macmil-
lan.
[23] Solow, Robert M. 1955–1956. “The Production Function and the Theory of Cap-
ital.”Review of Economic Studies 23 (2): 101–108.
[24] Solow, Robert (1956), "A contribution to the Theory of Economic Growth", The
Quarterly Journal of Economics, Vol. 70, No. 1 , pp. 65-94.
[25] Solow, Robert (2000), Growth Theory: An Exposition, 2nd Edition, Oxford Uni-
versity Press, New York.
[26] Summers, Lawrence H., 2014, “Re‡ections on the ‘New Secular Stagnation Hy-
pothesis”’. In Secular Stagnation: Facts, Causes and Cures, ed. by Coen Teulings
and Richard Baldwin, A VoxEU.org Book.
[27] Sydsæter, Knut, Arne Strøm and Peter Berck, Economists’Mathematical Manual,
Springer Verlag, 2005.
92
[29] Waelde, Klaus, Applied Intertemporal Optimization, 2014
(http://www.waelde.com/aio.html)
93