A Guide to Dynamic Systems∗

Markus Epp

1 Recap
Variable names and notions
• A variable that is given at the beginning of a period and that can be affected
indirectly, if at all, is called a stock- or state variable. Examples: capital/assets,
technology, ...

• A variable that is determined within a period is called a flow- or choice variable

(sometimes also: jump variable/jumper). Examples: consumption, output, ...

• A variable that is determined by modeled behavior of economic agents is called

endogenous [choice/state variable]. Example: consumption in the Ramsey-Cass-
Koopmans model.

• A variable that is driven by other processes is called exogenous [choice/state vari-

able]. Example: consumption in the Solow model.

Growth Denitions
Let x (t) be the value of variable x at time t. With slight abuse of notation, we will also
use x (t) ≡ xt .

• Time is continuous, iff t ∈ R. All functions h(t) have domain R (h : R → R).

• Time is discrete, iff t ∈ Z. All functions g(t) have domain Z (g : Z → R).

∗ This guide is supposed to serve as a tool for solving the questions in the tutorials and exams to Ad-
vanced Macroeconomics I. Note that the content is not directly/in itself matter of the exam. Though,
concepts may be required to answer exam questions. Some formulas will be provided in the exam,
if needed (equations indicated by a ). For questions regarding copyright and distribution feel free
to contact Markus Epp via [Link]@[Link].

1
Dynamic Optimization Advanced Macroeconomics

• A growth rate gx (t) of xt between two periods (time distance h) is defined as

∆xt ( x − xt )/h
gx (t) = = t+h . (1)
xt xt
The polar cases h = 1 and h → 0 yield the respective definitions for growth rates
in discrete and in continuous time.
∂xt
ẋt
• A growth rate in continuous time is thus defined by gx (t) = xt = xt where
∂t
ẋt is
∂xt x (t+h)− x (t)
the derivative of xt with respect to time, i.e. ẋt ≡ ∂t = lim h .
h →0

• Given a functional form of xt , the growth rate can be calculated by taking the
derivative of the log of a variable with respect to time:

∂ log( xt ) 1 ∂xt ẋt

= = = gx (t) (2)
∂t xt ∂t xt

• To determine an average growth rate from equally distanced data { x1 , x2 , ..., xn },

the geometric mean has to be employed:
s
n √
x̄ geom = n
∏ xi = n
x1 · x2 · · · x n ()
i =1

Dierential equations
• An equation that relates a variable to its derivatives is called a differential equa-
tion (DE). The initial value problem includes such a DE:

ẋ (t) = f ( x (t), t), x (t0 ) = x0 . (3)

• The order of a differential equation is defined by the highest derivative included

(ẋ, ẍ, ...).

• A linear first-order differential equation is a differential equation like (48) but

satisfying that f ( x (t), t) is in the class of linear functions:

f ( x ( t ), t ) = a ( t ) x ( t ) + b ( t ). (4)

• Solution existence, uniqueness as well as reliable method (variation of constants,

separation of variables, integrating factor, ...) will in general depend on the class
of differential equation (linearity, homogeneity, exactness, ...).

• For first-order linear differential equations there always exists a unique - possibly
unstable - solution.

2
Dynamic Optimization Advanced Macroeconomics

• A general solution to (3) is an equation x (t) = g̃( A, t), where g̃(·) is known and A
is an arbitrary starting value.

• A special or particular solution to (3) is an equation x (t) = g(t), where g(·) is

known because the starting value at t0 (i.e. x (t0 )) is given.

• A steady state is a point in time where a variable x does not change any more
without interventions, e.g. in (3): ẋ (t) = f ( x (t), t) = 0. Steady state values of
variables are often indicated by dropping time indices and stars and: x ∗ .

• A steady state is stable if deviations from the steady state (in the neighborhood
e) imply that x (t) converges back to x ∗ :

∂ ẋ (t) 
<0 (5, )
∂x (t)  x(t)= x∗

Dierence equations
• An equation that relates a variable y to its past (and future) values is called a
difference equation, e.g.:
y t = f ( y t +1 , y t −1 ) (6)

is a second order difference equation.

• Each component of a difference equation, yt− j , j ∈ N is called “iterate”. Past

values are called “lags”. Equation (6) e.g. is a difference equation of three iterates,
i.e. yt+1 depends on two lagged values/has two lags.

• The order of a difference equation is defined by the number of included lagged

values (difference equation (6) has two lags as yt and yt−1 implicitly lag yt+1 ).

• Interpreting yt as a vector, a system of difference equations is obtained.

• Linear difference equations are difference equations of the form

n
yt = b + ∑ a j yt− j (7)
j =1

where a j and b are parameters.

• For linear (systems of) first-order difference equations there always exists a unique
solution (possibly unstable).

• For zero constant(s) b, the (system of) difference equation(s) is called homoge-
neous.

3
Dynamic Optimization Advanced Macroeconomics

2 Dynamic Optimization in Continuous Time

Consider the class of dynamic optimization problems with infinite horizon of the fol-
lowing form:

Z∞
max e− ϕt f [ x (t), u(t)]dt (8)
x,u
0

s.t. ẋ (t) = g[ x (t), u(t)] (9)

x0 = x (0) > 0 . (10)

x (t) represents a (vector of-) state variable(s) with shorthand xt or just x. Accordingly,
u(t) represents the (vector of-) choice variable(s) with shorthand ut or just u. In this
formulation, x and u are paths of x (t) and u(t), respectively1 .
The solution to this problem can be found by employing optimal control theory. This
amounts to setting up an auxiliary function, the Hamiltonian-function. There are two
mathematically consistent approaches to reformulate the problem: the present value
(PV) approach and the current value (CV) approach.
The idea of these two approaches is as follows: searching solution paths for u(t) and
x (t) so that the integral in (8) is maximized is computationally tedious. Instead, we
solve the problem by looking at smaller subproblems - exploiting Bellman’s principle
of optimality2 - and combining these partial solutions thereafter. In particular, we look
at the value added by u(t) and x (t) at any point in time to the target function f (·). More-
over, we relate this value added to its “externalities”, i.e. its effect on the constraint(s)
g(·). The latter is done by introducing a co-state variable3 . The principle of optimality
implies that if the solution is optimal in any point in time (taking into account all possi-
ble effects of one point in time to another) it is also optimal on the entire time horizon.
The approaches (PV) and (CV) are explained in more detail in the following.

1A path is defined as: x ≡ { x (t)}∞

t =0 .
2 Formore information, see “Dynamic Programming” [Bellman (1957), Princeton University Press].
3 The manuscript “Dynamic optimization in continuous-time economic models (a guide for the per-

plexed)” [Obstfeld (1992), University of California at Berkeley] provides an excellent explanation of

this idea.

4
Dynamic Optimization Advanced Macroeconomics

2.1 PV-approach
Hamiltonian function:
H PV [ x (t), u(t), λ(t)] = e− ϕt f [ x (t), u(t)] + λ(t) g[ x (t), u(t)]
The solution to problem (8) s.t. to (9) and (10) is then characterized by:
∂HtPV ∂ f ( xt , ut ) ∂g( xt , ut )
=0 ⇔ e− ϕt + λt =0 (11)
∂ut ∂ut ∂ut
∂HtPV ∂ f ( xt , ut ) ∂g( xt , ut )
= −λ̇t ⇔ e− ϕt + λt = −λ̇t (12)
∂xt ∂xt ∂xt
∂HtPV
= ẋt ⇔ g( xt , ut ) = ẋt , (13)
∂λt
the transversality condition
lim λ( T ) x ( T ) = 0 (14)
T →∞
as well as the initial condition
x (0) = x0 . (15)
Usually, the strategy to determine the solution is to take the derivative of (11) with
respect to time t and to insert the result, say (11’), together with its primitive (11) into
(12). Whether there is a closed form solution (analytical solution you can obtain using
pen and paper) will in general depend on the structure of f (·) and g(·).

2.2 CV-approach
Hamiltonian function:
HCV [ x (t), u(t), µ(t)] = f [ x (t), u(t)] + µ(t) g[ x (t), u(t)]
Note that HCV (t) = e ϕ H PV (t). The solution to problem (8) s.t. to (9) and (10) is then
characterized by:
∂HtPV ∂ f ( xt , ut ) ∂g( xt , ut )
=0 ⇔ + µt =0 (16)
∂ut ∂ut ∂ut
∂HtPV ∂ f ( xt , ut ) ∂g( xt , ut )
= ϕµt − µ̇t ⇔ + µt = ϕµt − µ̇t (17)
∂xt ∂xt ∂xt
∂HtPV
= ẋt ⇔ g( xt , ut ) = ẋt (18)
∂µt
and the transversality condition
lim e− ϕT µ( T ) x ( T ) = 0 (19)
T →∞
as well as the initial condition
x (0) = x0 (20)
The same strategy as in the PV-approach can in principle be employed to obtain a
solution.

5
Dynamic Optimization Advanced Macroeconomics

Remarks
PV-approach vs. CV-approach
In the PV-approach, the value added at period t is discounted back to its present-value,
i.e. the value at t = 0. Hence, all per-period choices are linked over the very first
period. In contrast, the CV-approach is set up in relative time and links all periods
implicitly by relating the first order conditions to each other. It is easily seen, that the
approaches are consistent since µ(t) = e ϕt λ(t) implies that (11)-(14) and (16)-(19) are
equivalent.

2.3 Example: A cake-eating problem

An infinitely living agent owns a cake of size k0 at t = 0. Let k (t) be the size of this cake
p
at t. Suppose the agent derives utility from eating cake according to: u(c(t)) = c(t)
where c(t) represents the amount of cake eaten up in t. So the agent faces the following
problem:
Z∞ q
−ρt
max e c(t)dt (21)
k ( t ), c ( t )
0

s.t. k̇ (t) = −c(t) (22)

k 0 = k (0) > 0 . (23)

Which amount of cake should the agent eat at t?

This problem seems to be a variation of the problem in (8)-(10) with k (t)=
ˆ x ( t ), c ( t ) =
ˆ u(t)
and ϕ= ˆ ρ. It must also be the case that the functions f (·) and g(·) are defined by:
p
f [ x (t), u(t)] = u(t) = f [u(t)] and g[ x (t), u(t)] = −u(t) = g[u(t)]. The Hamiltonian
can therefore be set up as:
q
H PV (t) = e−ρt c(t) − λ(t)[c(t)] (24)

so that the First-order conditions are given by

∂HtPV 1
=0 ⇔ e−ρt √ − λt = 0 (25)
∂ct 2 ct
∂HtPV
= −λ̇t ⇔ 0 = −λ̇t (26)
∂k t
∂HtPV
= k̇ t ⇔ −ct = k̇ t (27)
∂λt
Taking the derivative of (25) with respect to t yields

ρe−ρt e−ρt
λ̇t = − 1
− 3 ċt
2ct2 4ct2

6
Dynamic Optimization Advanced Macroeconomics

which implies together with (26) that

ρe−ρt e−ρt ċt ρ 1 ċt

0=− 1
− 3 ⇔ 0=− − ⇔ gc = −2ρ
2 2 2 4 ct
2ct 4ct

Hence, the time path of consumption can be stated as follows (reassure yourself):

ct = c0 e−2ρt

with the constant c0 = c(0) yet to be determined4 . The time path of consumption is
depicted in Figure 1. It is left to the reader to verify that the transversality condition
holds for these paths.

Consumption path
0.15 = 0.02, k 0 = 1
= 0.08, k 0 = 1

0.1
c(t)

0.05

0
0 10 20 30 40 50 60 70 80 90 100
t

Figure 1: c(t) depicted for two values of ρ.

4 To determine c we exploit knowledge of k0 . To do so, we derive the time path of cake k t by integrating
0
k̇ t = −c0 e − 2ρt . Then we evaluate k t at t = 0 and reformulate for c0 to obtain c0 = 2ρk0 .

7
Dynamic Optimization Advanced Macroeconomics

3 Dynamic Optimization in Discrete Time

Consider the class of dynamic optimization problems with infinite horizon of the fol-
lowing form:
∞
max ∑ βt f [ xt , ut ] (28)
x,u
t =0

s.t. x t +1 ≤ g [ x t , u t ] , ∀ t (29)
x0 = x (0) > 0 . (30)

xt represents a (vector of-) state variable(s) determined in t. Accordingly, ut represents

the (vector of-) choice variable(s) chosen at t ∈ N. In this formulation, x and u are
paths of xt and ut , respectively5 . The solution to (28)-(30) are paths x and u.
A natural starting point to find the solution to this problem is setting up the Lagrangean.
Alternatively, the problem can be solved by setting up the Bellman equation. These two
approaches are explained in more detail in the following.

3.1 Lagrangean
Setting up the Lagrangean as follows:
∞
Lt = ∑ βt ( f [xt , ut ] + λt [ g(xt , ut ) − xt+1 ]) (31)
t =0

yields the first-order conditions

∂Lt
 
t ∂ f ( xt , ut ) ∂g( xt , ut )
=0 ⇔ β + λt =0 (32)
∂ut ∂ut ∂ut
∂Lt
 
t +1 ∂ f ( x t +1 , u t +1 ) ∂g( xt+1 , ut+1 )
=0 ⇔ β − λ t +1 − βt λt = 0 (33)
∂xt+1 ∂xt+1 ∂xt+1
∂Lt
=0 ⇔ g( xt , ut ) = x t +1 (34)
∂µt
which characterize the solution together with analogous versions of transversality con-
dition (14) and initial condition (15).
Note that (32)-(34) posit an infinity of conditions, as they hold for all t ∈ [0, ∞). The
natural strategy to solve this problem is to eliminate the Lagrangean multiplier by re-
arranging and combining (32) and (33). Here we can rewrite (32)-(34) as:

∂ f ( xt , ut ) ∂g( xt , ut ) −1
 
λt = − (35)
∂ut ∂ut
 
∂ f ( x t +1 , u t +1 ) ∂g( xt+1 , ut+1 )
λt = β + λ t +1 (36)
∂xt+1 ∂xt+1
x t +1 = g ( x t , u t ) (37)
5 Here a path is defined as: x ≡ { xt }∞
t =0 , t ∈ N

8
Dynamic Optimization Advanced Macroeconomics

so that we obtain the following system:

 
∂ f ( xt ,ut ) ∂ f ( xt+1 ,ut+1 )
∂ f ( x t +1 , u t +1 ) ∂ut+1 ∂g( xt+1 , ut+1 ) 
− ∂g(∂u t
xt ,ut )
= β − ∂g( xt+1 ,ut+1 )
(38)
∂xt+1 ∂xt+1
∂ut ∂ut+1

x t +1 = g ( x t , u t ). (39)

(38) is called an Euler-equation. Note that (39) pins down the path for ut when the path
of xt is given. The system (38)-(39) does not in general exhibit closed form solutions
unless f (·) and g(·) have some specific forms. Hence, finding x can be difficult and is
sometimes conducted using numerical methods.

3.2 Bellman equation

As a result of (29) (or by anticipating (34)), u is implied by x. Hence we can delay the
search for u by finding just x. This is especially convenient when the constraint(s) (29)
can be easily rewritten as ut ≤ h( xt+1 , xt )6 . Define f˜( xt , xt+1 ) = f ( xt , h( xt+1 , xt )) so
that in this case, we look for a solution to the recursive problem:

V ( xt ) = max { f˜[ xt , xt+1 ] + βV ( xt+1 )} (40)

x t +1 ∈ Γ ( x t )

x0 = x (0) > 0 . (41)

and Γ( xt ) is the set of feasible choices of xt+1 so that xt+1 ≤ g( xt , h( xt+1 , xt )).
The solution is characterized by the first-order condition to the right-hand side of (40)
and an Envelope condition (somestimes called Benveniste-Scheinkman formula):

∂ f˜[ xt , xt+1 ] ∂V ( xt+1 )

+β =0 (42)
∂xt+1 ∂xt+1
∂V ( xt ) ∂ f˜[ xt , xt+1 ]
= (43)
∂xt ∂xt
Dating up (43) to t + 1 and combining the result with (42), we obtain a single difference
equation to find x:
∂ f˜[ xt , xt+1 ] ∂ f˜[ xt+1 , xt+2 ]
+β =0 (44)
∂xt+1 ∂xt+1

3.3 Example: "AK"-production

An infinitely living agent owns capital stock k t at t = {0, 1, ...} and has initial endow-
ment of capital k0 > 0. Capital can be consumed or used in production, where k t units
of capital produce Ak t units of goods. δk t units of capital decay per period. The agent

6A sufficient condition for analytical tractability is that g[ xt , ut ] is linear in ut .

9
Dynamic Optimization Advanced Macroeconomics

p
derives utility from consumption according to u(c(t)) = c(t) and discounts future
utility at factor β ∈ (0, 1). The agent then faces the following problem:
∞ q
max ∑ β c(t)
t
(45)
k,c t=0

s.t. k t+1 ≤ Ak t − ct + (1 − δ)k t , ∀t (46)

k0 > 0 . (47)

Which amount of cake should the household eat at t?

3.3.1 Using the Lagrangean

√
This problem is a variation of the problem in (28)-(30) with k t =
ˆ xt , ct =
ˆ ut . f [ xt , ut ]=
ˆ ut
ˆ Axt − ut + (1 − δ) xt . It holds ∀t that
and g[ xt , ut ]=
∂ f ( xt , ut ) 1
= √
∂ut 2 ut
∂ f ( xt , ut )
=0
∂xt
∂g( xt , ut )
= −1
∂ut
∂g( xt , ut )
= 1−δ+ A
∂xt
so that the version of the Euler-equation (38) for the problem (45)-(47) is given by:
1 β (1 − δ + A )
√ = √ (48)
ct c t +1
or, rewritten:
ct+1 = [ β(1 − δ + A)]2 ct (48’)
Note also, that the resource constraint is binding: k t+1 = Ak t − ct + (1 − δ)k t . Equation
(48’) implies that ct grows/shrinks at rate [ β(1 − δ + A)]2 − 1. By guess and verify7 ,
one can show that this is also the growth rate of k t . The paths are thus:

k t = [ β(1 − δ + A)]2t k0
ct = [ β(1 − δ + A)]2t c0

where c0 can be determined from the t = 0 resource constraint8 . Figure 2 depicts the
paths of consumption and capital under different assumptions on A.
7 Intuitively, capital needs to grow/shrink at the same rate, as to maintain that ct neither surpasses
capital in size (making the optimal choice of ct eventually unfeasible) nor that capital is accumulated
too fast (higher consumption at any point in time feasible). Thus we guess first that k t+1 = χk t .
Inserting this guess into the budget constraint and reformulating yields: ct = (1 − δ + A − χ)k t .
c k
Relating the resource constraint at t and t + 1 yields tc+t 1 = tk+t 1 = [ β(1 − δ + A)]2 which verifies the
guess.
8 Inserting k1 = [ β(1 − δ + A)]2 k0 into c0 = (1 − δ + A)k0 − k1 yields a formula for c0 .

10
Dynamic Optimization Advanced Macroeconomics

3.3.2 Using the Bellman equation

Stating the Bellman equation as
q 
V (k t ) = max (1 − δ + A)k t − k t+1 + βV (k t+1 ) (49)
k t +1 ∈ Γ ( k t )

x0 = x (0) > 0 . (50)

ˆ 1 − δ + A) xt − xt+1 , f˜( xt , xt+1 ) =

p
implies that h( xt+1 , xt )=( (1 − δ + A ) x t − x t +1 ,
and Γ( xt ) = [0, (1 − δ + A) xt ]. Hence, (44) for the present model can be written as

1 β (1 − δ + A )
p = p (51)
2 (1 − δ + A ) k t − k t +1 2 (1 − δ + A ) k t +1 − k t +2

or, equivalently:
1 β (1 − δ + A )
√ = √
ct c t +1
which is (48) from above.

Path of consumption and capital: = 0.5, = 0.05, A = 0.1

1
ct
0.8 k
t

0.6
ct

0.4

0.2

0
1 1.5 2 2.5 3 3.5 4 4.5 5
t
= 0.5, = 0.05, A = 2
25

20

15
ct

10
ct
5
k
t
0
1 1.5 2 2.5 3 3.5 4 4.5 5
t

Figure 2: ct and k t depicted for two values of A, given k0 = 1.

11
Dynamic Optimization Advanced Macroeconomics

4 Analytical Methods
Growth Accounting
Suppose we are interested in the growth rate of a variable xt but only observe yt and
zt . If xt , yt and zt are related (in a specific fashion), we can figure out the growth rate
of xt . Suppose:
xtα
zt = β
(52)
yt
In general, we would need to take the derivative of (52) w.r.t. time and divide the result
by (52) again (applying product/chain rules), which is tedious:
 
xα  
∂ tβ ∂ x α y− β
yt t t −β − β −1
żt = = = αxtα−1 ẋt yt − βyt ẏt xtα
∂t ∂t
−β −β
żt xtα ẋt yt yt ẏt xtα (48) zt ẋt zt ẏt
⇔ gz = = α −β = α −β
zt zt xt zt yt zt xt zt yt
1
⇔ gz = αgx − βgy ⇒ gx = ( gz + βgy ))
α
For multiplicative associations as in relationship (52), exploiting (46) is attractive. Tak-
ing logs and then the derivative with respect to time is much faster:
 
xα
∂ log tβ
∂ log(zt ) yt
=
∂t ∂t
∂ (α log xt − β log yt )
⇔ gz =
∂t
∂ log xt ∂ log yt
⇔ gz = α −β = αgx − βgy ⇒ ...
∂t ∂t
Example: A Cobb-Douglas production function states that output is produced accord-
ing to:
β
Yt = At K α Lt , (53)

where usually β = 1 − α (constant returns to scale) is assumed. Hence, observations

on Yt , Kt and Lt can be used to estimate growth of factor productivity At from:

g A = gY − αgK − βg L

α and β are factor income shares (or production factor elasticities) which can be esti-
mated from Yt , Kt and Lt and wages wt as well as rate of return on capital rt :

rt Kt wt Lt
α= , β= .
Yt Yt
Under the assumption of constant returns to scale, the Euler-Theorem implies β = 1 − α.

12
Dynamic Optimization Advanced Macroeconomics

Guess and verify - Method of undetermined coecients

A method to solve a wide class of (dynamic) problems is to guess a solution and to
verify the solution by applying the method of undetermined coefficients. For some
problems obeying some regularity conditions, this method proves extremely powerful
and fast. The method is illustrated for differential and difference equations in the fol-
lowing.
Suppose you face the canonical problem to trace out a path of xt given the difference
equation:
xt+1 = axt + bxt−1 , x0 > 0 (54)
where a, b are constants.
Guess: let’s assume (because we have some idea about the appearance of the solution)
that (54) implies the law of motion for xt :

xt+1 = λxt (55)

so that x T = λ T x0 (λ is yet undetermined, giving this method its name). To find λ,

insert (55) into (54) to obtain:
 
b b
x t +1 = axt + xt = a+ xt (56)
λ λ
Note that we have verified our guess already: it is indeed the case, that one solution to
(54) is a linear law of motion as stated in (55): a comparison implies that equations (55)
and (56) coincide whenever:
b
λ = a+ ⇔ λ2 − aλ − b = 0 (57)
λ
Hence, values of λ that solve this equation also solve (54). This quadratic form (which
we will get to know as the characteristic polynomial) has roots
√
a ± a2 + 4b
λ1,2 = (58)
2
both of which are possible solutions to (54).
Correspondingly, an initial value problem like:

ẋ (t) − ax (t) = 0, x (0) = x0 = 1 (59)

can be solved by guessing a functional form for x (t), e.g.:

x (t) = c exp( at) (60)

Inserting (60) and the derivative of (60) into (59) verifies the guess. The constant c and
therefore the particular solution can be determined using x0 = 1:

x (0) = c exp(0) = c ⇒ x (t) = x0 exp( at) = exp( at). (61)

13
Dynamic Optimization Advanced Macroeconomics

Solving dierence equations

Characteristic equation/polynomial
For homogeneous linear difference equations (systems of the form (7) where b = 0), a
solution candidate can be found by finding the root(s) of the characteristic polynomial,
i.e. solving the characteristic equation9
n
λn − ∑ a j λn− j = 0 (62, )
j =1

for the characteristic roots λi , i = 1, ..., n. The solution is then of the form
n
xt = ∑ c j λtn− j (63, )
j =1

where the constants c j can be traced back by using known values of xt . Specifically, this
implies that for a difference equation of order n with k ≤ n distinct characteristic roots
within the unit circle (|λi | < 1 is true for k distinct roots), a unique solution requires
that at least k values of xt have to be known. In other words, xt needs to be known at k
different points in time.
Note: the absolute values of the distinct roots λi can be ordered such that |λ1 | < |λ2 | <
... < |λk |. Let D ∈ N be the number of distinct roots within the unit circle (i.e. |λi | < 1,
∀i ≤ D) and K ∈ N the number of known values of xt , i.e. the number of initial
conditions. The following holds:

• A unique solution exists whenever D = K.

• An infinity of solutions exists whenever D > K.

• The system is over-identified whenever 0 < D < K.

• No (non-trivial) solution exists whenever 0 = D < K.

Example: no text yet.

Non-homogeneous dierence equations

Non-homogeneous difference equations (b > 0) can be easily transformed into homo-
geneous difference equations by (1) subtracting the difference equation by its first lag
or by (2) exploiting, if existent, the steady state condition.
9 Note that this implies for difference equations of second and higher orders (n > 1), that the roots may
not be real. Complex solutions are often ruled out by the assumptions on the model’s parameters.
In case of multiplicity of solutions, the so-called "minimal state variable criterion" (MSV) can be
used to find a single, bubble-free solution. For more information, go to: “On non-uniqueness in
rational expectations models: An attempt at perspective” [McCallum (1983), Journal of Monetary
Economics].

14
Dynamic Optimization Advanced Macroeconomics

(1) Subtract yt−1 from yt , i.e. exploit that (7) holds for all t:
n n
y t − y t −1 = b − b + ∑ a j y t − j − ∑ a j y t − j −1
j =1 j =1
n
∑ aj


y t − y t −1 = y t − j − y t − j −1
j =1

Defining a dummy xt− j ≡ yt− j − yt− j−1 , the latter is again a homogeneous differ-
ence equation:
n
xt = ∑ a j xt− j . (64)
j =1

Alternatively, reformulations of (62) yield the homogeneous difference equation

form (51) in original iterates of y in constants c j ≡ a j−1 − a j , j = 1, 2, ..., n (where
a0 ≡ 1):
n
yt = ∑ c j yt− j . (65)
j =1

(2) Often it is known that the system (7) exhibits a unique steady state: from some
point in time t it holds for all past iterates that yt− j = y∗ , where j ∈ {0, 1, 2, ..., n}
so that also for future iterates yt+ j = y∗ , j ∈ {0, 1, 2, ..., n}. The steady state value
can be computed from
b
y∗ = . (66)
1 − ∑nj=1 a j
Defining a variable xt− j ≡ yt− j − y∗ , the difference equation can be rewritten in
deviations from the steady state10 :
n
xt = ∑ a j xt− j (67)
j =1

10 This follows from subtracting y∗ on both sides and applying a zero trick:
n n n
yt − y∗ = ∑ a j yt− j + b − ∑ a j y∗ + ∑ a j y∗ − y∗
j =1 j =1 j =1
n n
⇔ yt − y∗ = ∑ aj yt− j − y∗ − y∗ + y∗ ∑ aj + b


j =1 j =1
!
n n
⇔ xt = ∑ a j xt− j − y∗ 1− ∑ aj +b
j =1 j =1
!
n n
b
⇔ xt = ∑ a j xt− j − n 1− ∑ aj +b
j =1
1 − ∑ j =1 a j j =1
n
⇔ xt = ∑ a j xt− j
j =1

15
Dynamic Optimization Advanced Macroeconomics

Log-linearization
No text yet, to follow opportunely.

16