Difference Equations for Economists1

preliminary and incomplete

Klaus Neusser

March 21, 2016

1 Klaus
c Neusser
 i

Preface
There are, of course, excellent books on deterministic difference equations:
for example, Elaydi (2005), Agarwal (2000), or Galor (2007). Colonius and
Kliemann (2014) give a presentation from the perspective of dynamical sys-
tems. These books do, however, not go into the specific problems faced in
economics.
The books makes use of linear algebra. Very good introduction to this
topic is presented by the books of Strang (2003) and Meyer (2000).

Bern,
September 2010 Klaus Neusser
ii
Contents

1 Introduction 1
1.1 Notation and Preliminaries . . . . . . . . . . . . . . . . . . . . 1

2 Linear Difference Equations 5

2.1 First Order Difference Equation . . . . . . . . . . . . . . . . . 5
2.2 Steady State and Stability . . . . . . . . . . . . . . . . . . . . 11
2.3 Solutions of First Order Equations . . . . . . . . . . . . . . . 16
2.4 Examples of First Order Equations . . . . . . . . . . . . . . . 19
2.4.1 The simple Cobweb Model . . . . . . . . . . . . . . . . 19
2.4.2 The Solow Growth Model . . . . . . . . . . . . . . . . 20
2.4.3 A Model of Equity Prices . . . . . . . . . . . . . . . . 25
2.5 Difference Equations of Order p . . . . . . . . . . . . . . . . . 32
2.5.1 Homogeneous Difference Equation of Order p . . . . . 33
2.5.2 Nonhomogeneous Equation of Order p . . . . . . . . . 35
2.5.3 Limiting Behavior of Solutions . . . . . . . . . . . . . . 36
2.5.4 Examples . . . . . . . . . . . . . . . . . . . . . . . . . 42

3 Systems of Difference Equations 53

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
3.2 First Order System of Difference Equations . . . . . . . . . . . 55
3.2.1 Homogenous First Order System of Difference Equations 55
3.2.2 Solution Formula for Homogeneous Systems . . . . . . 58
3.2.3 Nonhomogeneous First Order System . . . . . . . . . . 62
3.3 Stability Theory . . . . . . . . . . . . . . . . . . . . . . . . . . 63
3.4 Two-dimensional Systems . . . . . . . . . . . . . . . . . . . . 69
3.5 Boundary Value Problem . . . . . . . . . . . . . . . . . . . . . 78

4 Examples: Linear Systems 85

4.1 Exchange Rate Overshooting . . . . . . . . . . . . . . . . . . . 85
4.1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . 85
4.1.2 Analysis of the Dynamic Properties . . . . . . . . . . . 87

iii
iv CONTENTS

4.1.3 Effects of an Increase in Money Supply . . . . . . . . . 90

4.2 Optimal Growth Model . . . . . . . . . . . . . . . . . . . . . . 92
4.2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . 92
4.2.2 Steady State . . . . . . . . . . . . . . . . . . . . . . . . 93
4.2.3 Discussion of the Linearized System . . . . . . . . . . . 95
4.2.4 Some Policy Experiments . . . . . . . . . . . . . . . . 97
4.3 The New Keynesian Model . . . . . . . . . . . . . . . . . . . . 101

5 Stochastic Difference Equation 103

5.1 Introduction and Assumptions . . . . . . . . . . . . . . . . . . 103
5.2 The univariate case . . . . . . . . . . . . . . . . . . . . . . . . 105
5.2.1 Solution to the homogeneous equation . . . . . . . . . 106
5.2.2 Finding a particular solution . . . . . . . . . . . . . . . 106
5.2.3 Example: Cagan’s model of hyperinflation . . . . . . . 107
5.3 The multivariate case . . . . . . . . . . . . . . . . . . . . . . . 109

A Complex Numbers 115

B Matrix Norm 119

List of Figures

2.1 Cobweb diagram with steady states of the logistic function:

y = 2.5x(1 − x) and x0 = 0.1 . . . . . . . . . . . . . . . . . . . 12
2.2 Price dynamics in the cobweb model . . . . . . . . . . . . . . 20
2.3 Capital Intensity in the Solow Model . . . . . . . . . . . . . . 24
2.4 Impulse response function of the Cagan model with adaptive
expectations taking α = −0.5 and γ = 0.9 . . . . . . . . . . . 30
2.5 Behavior of Xt = λt1 depending on λ ∈ R . . . . . . . . . . . . 38
2.6 Behavior of Xt in case of complex roots . . . . . . . . . . . . . 39
2.7 Stability properties of equation: Xt − φ1 Xt−1 − φ2 Xt−2 = 0 . . 42
2.8 Impulse Response Coefficients of the Multiplier-Accelerator
model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
2.9 Impulse Response after a positive Supply Shock . . . . . . . . 51

3.1 Example of a Phase Diagram . . . . . . . . . . . . . . . . . . 71

3.2 Asymptotically Stable Steady State (λ1 = 0.8, λ2 = 0.5) . . . . 74
3.3 Unstable Steady State (λ1 = 1.2, λ2 = 2) . . . . . . . . . . . . 74
3.4 Saddle Steady State (λ1 = 1.2, λ2 = 0.8) . . . . . . . . . . . . 75
3.5 Repeated Roots with Asymptotically Stable Steady State (λ1 =
λ2 = 0.8) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
3.6 Degenerate Steady State (λ1 = 1, λ2 = 0.8) . . . . . . . . . . . 77
3.7 Repeated eigenvalues one independent eigenvector (λ = 0.8) . 77
3.8 Complex eigenvalues with Stable Steady State (λ1,2 = 0.7 ± 0.2ı) 78
3.9 Complex eigenvalues with Unstable Steady State (λ1,2 = 1±0.5ı) 79
3.10 Complex eigenvalues on the unit circle (λ1,2 = cos (π/4) ±
ı sin (π/4)) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79

4.1 Dornbusch’s Overshooting Model . . . . . . . . . . . . . . . . 90

4.2 Unanticipated Increase in Money Supply Dornbusch’s Over-
shooting Model . . . . . . . . . . . . . . . . . . . . . . . . . . 91
4.3 Phase diagram of the optimal growth model . . . . . . . . . . 94

v
vi LIST OF FIGURES

4.4 Phase diagram of the optimal growth model with distortionary

taxation of capital . . . . . . . . . . . . . . . . . . . . . . . . 98
4.5 Phase diagram of the optimal growth model with government
expenditures . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100

A.1 Representation of a complex number . . . . . . . . . . . . . . 116

List of Definitions

2.1 Definition (Linear Dependence, Linear Independence) . . . . . 6

2.2 Definition (Fundamental Set of Solutions) . . . . . . . . . . . 6
2.3 Definition (Casarotian Matrix) . . . . . . . . . . . . . . . . . . 7
2.4 Definition (Equilibrium Point, Steady State) . . . . . . . . . . 11
2.5 Definition (Stability) . . . . . . . . . . . . . . . . . . . . . . . 13

3.1 Definition (Orbit) . . . . . . . . . . . . . . . . . . . . . . . . . 56

3.6 Definition (Hyperbolic Matrix) . . . . . . . . . . . . . . . . . . 66
3.7 Definition (Saddle Point) . . . . . . . . . . . . . . . . . . . . . 69

vii
viii LIST OF DEFINITIONS
List of Theorems

2.1 Theorem (Dimension of Linear First Order Equation) . . . . . 7

2.2 Theorem (Superposition Principle) . . . . . . . . . . . . . . . 8
2.3 Theorem (Criterion Asymptotic Stability) . . . . . . . . . . . 14
2.4 Theorem (Stability Condition of Nonlinear Equation) . . . . . 14
2.5 Theorem (Fundamental Set for equation of order p) . . . . . . 33
2.6 Theorem (Limiting Behavior of Second Order Equation) . . . 38
2.7 Theorem (Limiting Behavior of Second Order Equation (orig-
inal parameters)) . . . . . . . . . . . . . . . . . . . . . . . . . 40
2.8 Theorem (Stability Conditions of Second Order Equation (orig-
inal parameters)) . . . . . . . . . . . . . . . . . . . . . . . . . 40

3.5 Theorem (Stable Manifold Theorem) . . . . . . . . . . . . . . 67

3.6 Theorem (Hartman–Grobman Theorem) . . . . . . . . . . . . 68
3.8 Theorem (Blanchard-Kahn) . . . . . . . . . . . . . . . . . . . 82

ix
Chapter 1

Introduction

1.1 Notation and Preliminaries

A difference equation or dynamical system describes the evolution of some
variable over time. The value of this variable in period t is denoted by Xt .
The time index t takes on discrete values and typically runs over all integer
numbers Z, e.g. t = . . . , −2, −1, 0, 1, 2, . . . By interpreting t as the time
index, we have automatically introduced the notion of past, present and
future. A difference equation is then nothing but a rule or a function which
instructs how to compute the value of the variable of interest in period t
given past values of that variable and time. The system may be initialized
at some point t0 which in most cases is taken to be t0 = 0. In this case t
runs over all natural numbers, i.e. t ∈ N0 .
In its most general form a difference equation can be written as

F (Xt , Xt−1 , Xt−2 , . . . , Xt−p , t) = 0 (1.1)

where F is a given function. The variable Xt is called the endogenous or

dependent variable and is an n-vector, i.e Xt ∈ Rn , n ≥ 1. In dynamical
system theory, Xt is called the state of the system and Rn its state space. n
is called the dimension of the system. The difference between the largest and
the smallest time index of the dependent variable explicitly involved is called
the order of the difference equation. In the formulation (1.1), this is p with
p ≥ 1. In the difference equation above the time index appears explicitly as
an argument of the function F . In this case one speaks of a nonautonomous
or a time-variant equation. If time is not a separate argument and enters
only as a time index of the dependent variable, the equation is said to be
autonomous or time-invariant. In many applications, time-variance enters
the difference equation by replacing the time index in equation (1.1) by some

1
2 CHAPTER 1. INTRODUCTION

variable Zt ∈ Rk with k ≥ 1. This variable is called the exogenous or

independent variable.
With the exception of chapter 5, we will always assume that it is possible
to solve equation (1.1) uniquely for Xt :

Xt = f (Xt−1 , Xt−2 , . . . , Xt−p , t) (1.2)

The difference equation is called normal in this case.

Given some starting values x0 , x−1 , . . . , x−p+1 for X0 , X−1 , . . . , X−p+1 , the
difference equation (1.2) uniquely determines all subsequent values of Xt , t =
1, 2, . . . , by iteratively inserting into equation (1.2).
A solution to the difference equation is a function X : Z −→ Rn such
that X(t) fulfills the difference equation (1.1), e.g.

F (X(t), X(t−1), X(t−2), . . . , X(t−p), t) = 0 holds for all t ∈ Z. (1.3)

The aim of the analysis is to assess the existence and uniqueness of a solution
to a given difference equation; and, in the case of many solutions, to charac-
terize the set of all solutions. In the normal linear case of dimension n = 1,
for example, the set of solutions turns out to be a vector space of dimension
p.
One way to pin down a particular solution is to require that the solution
must satisfy some boundary conditions. In this case, we speak of a boundary
value problem. The simplest way to specify boundary conditions is to require
that the solution X(t) must be equal to p prescribed values x1 , . . . , xp , called
initial conditions, at given time indices t1 , . . . , tp :

X(t1 ) = x1 , . . . , X(tp ) = xp . (1.4)

In this case, we speak also of an initial value problem. To highlight the

dependency of the solution on the initial values, we may write explicitly
X(t, x1 , . . . , xp ) where it is understood that the initial values are specified for
time zero.
In economics, particularly in dealing with rational expectations models,
boundary conditions also arise from the condition that limt→∞ X(t) must be
finite or equal to a prescribed value, typically zero. Such terminal condi-
tions often arise as transversality conditions in optimal control problems. In
this setting, the task is to find appropriate initial values such that the corre-
sponding solution satisfies the given terminal conditions. If the initial values
can be pinned down uniquely, the rational expectations model is said to be
determinate. If there is a multitude of solutions, the rational expectations
model is said to be indeterminate.
1.1. NOTATION AND PRELIMINARIES 3

Difference equations, like (1.2), transform one sequence (X) = (Xt ) into
another one. The difference equation therefore defines a function or better
an operator on the set of all sequences, denoted by RZ , into itself. Defining
addition and scalar multiplication in the obvious way, RZ forms a vector or
linear space over the real numbers. For many economic applications, it makes
sense to concentrate on the set of bounded sequences, i.e. on sequences (Xt )
for which there exists a real number M such that kXt k ≤ M for all t ∈ Z.
The set of bounded sequences is usually denoted by `∞ . It can be endowed
with a norm k.k∞ in the following way:

k(X)k∞ = sup{kXt k, t ∈ Z}

where kXt k denotes the Euclidean norm of Xt in Rn . If n = 1, kXt k = |Xt |.

It can be shown that `∞ endowed with the metric induced by this norm is
a linear complete metric space or a Banach space (see, for example, Naylor
and Sell, 1982).
For the analysis of difference equations it is useful to introduce the lag
operator or back shift operator denoted L. This operator shifts the time index
one period into the past. The sequence (Xt ) is then transformed by the lag
operator into a new sequence (Yt ) = L(Xt ) such that Yt = Xt−1 . In order
to ease the notation we write LXt instead of L(Xt ). It is easy to see that
the lag operator is a linear operator on RZ . Given any two sequences (Xt )
and (Yt ) and any two real numbers a and b, L(aXt + bYt ) = aLXt + bLYt =
aXt−1 + bYt−1 . For any natural number p ∈ N, Lp is defined as the p-times
application of L:
Lp Xt = |LL{z
· · · L} Xt = Xt−p
p-times
p
For p = 0, L = I the identity operator. Thus, the action of a lag polynomial
of order p, Φ(L) = I − Φ1 L − Φ2 L2 − . . . − Φp Lp on Xt is

Φ(L)Xt = (I − Φ1 L − Φ2 L2 − . . . − Φp Lp )Xt
= Xt − Φ1 Xt−1 − Φ2 Xt−2 − . . . − Φp Xt−p

where Φj , j = 1, . . . , p, are n × n matrices. The minus signs are arbitrary.

The polynomial is written in this way as this is exactly the form in which it
will appear later on.
4 CHAPTER 1. INTRODUCTION
Chapter 2

Linear Difference Equations

with Constant Coefficients

This chapter is entirely devoted to the analysis of linear nonhomogeneous

difference equations of dimension one (n = 1) and order p ≥ 1 with constant
coefficients:

Xt = φ1 Xt−1 + φ2 Xt−2 + · · · + φp Xt−p + Zt , φp 6= 0, (2.1)

where φ1 , . . . , φp are given constant real numbers. The variable Zt represents

the non-autonomous part of the equation which influences the evolution of
Xt over time. Its values are given from outside the system. Thus, Zt is called
exogenous, independent or forcing variable.

2.1 First Order Difference Equation

As a starting point and motivation of the analysis consider the simplest case,
namely the first order (p = 1) linear nonhomogeneous equation:

Xt = φXt−1 + Zt , φ 6= 0. (2.2)

To this nonhomogeneous equation corresponds a first order linear homoge-

nous equation:
Xt = φXt−1 . (2.3)

Starting in period 0 at some arbitrary initial value X0 = x0 , all subsequent

values can be recursively computed by iteratively inserting into the difference

5
6 CHAPTER 2. LINEAR DIFFERENCE EQUATIONS

equation (2.3)
X1 = φx0
X2 = φX1 = φ2 x0
...
Xt = φXt−1 = φt x0 .
This suggests to take
Xt = φt c (2.4)
as the general solution of the first order linear homogenous difference equa-
tion (2.3). Actually, equation (2.4) provides a whole family of solutions
indexed by the parameter c ∈ R. To each value of c, there corresponds a
trajectory (Xt ) = (φt c). In order to highlight this dependency, we may write
the solutions as Xt (c). Note that the trajectories of two different solutions
Xt (c1 ) and Xt (c2 ), c1 6= c2 , cannot cross.
The parameter c can be pinned down by using a single boundary condition.
A simple form of such a boundary condition requires, for example, that Xt
takes a particular value xt0 in some period t0 . Thus, we require that Xt0 = xt0
in period t0 . In this case we speak of an initial value problem. The value of
c can then be retrieved by solving the equation xt0 = φt0 c for c. This leads
x
to the solution: c = φtt00 . In many instances we are given the value at t0 = 0
so that c = x0 .
Suppose that we are given two solutions of the homogenous equation,
(1) (2)
(Xt ) and (Xt ). Then it is easy to verify that any linear combination of
(1) (2)
the two solutions, a1 (Xt ) + a2 (Xt ), a1 , a2 ∈ R, is also a solution. This
implies that the set of all solutions to the homogenous equation forms a
linear space. In order to find out the dimension of this linear space and its
algebraic structure, it is necessary to introduce the following three important
definitions.
Definition 2.1 (Linear Dependence, Linear Independence). The r sequences
(X (1) ), (X (2) ), . . . , (X (r) ) with r ≥ 2 are said to be linearly dependent for
t ≥ t0 if there exist constants a1 , a2 , . . . , ar ∈ R, not all zero, such that
(1) (2) (r)
a1 Xt + a2 Xt + · · · + ar Xt =0 ∀t ≥ t0 .
This definition is equivalent to saying that there exists a nontrivial linear
combination of the solutions which is zero. If the solutions are not linearly
dependent, they are said to be linearly independent.
Definition 2.2 (Fundamental Set of Solutions). A set of r linearly inde-
pendent solutions of the homogenous equation is called a fundamental set of
solutions.
2.1. FIRST ORDER DIFFERENCE EQUATION 7

Definition 2.3 (Casarotian Matrix). The Casarotian matrix C(t) of (X (1) ),

(X (2) ), . . . , (X (r) ) with r ≥ 1 is defined as
 (1) (2) (r) 
Xt Xt . . . Xt
 X (1) (2) (r)
Xt+1 . . . Xt+1 
 t+1
C(t) =  . .

. .
. . . .
.
 . . . . 
(1) (2) (r)
Xt+r−1 Xt+r−1 . . . Xt+r−1
These definitions allow us to the tackle the issue of the dimension of
the linear space given by all solutions to the homogenous first order linear
difference equation.
Theorem 2.1 (Dimension of Linear First Order Equation). The set of solu-
tions to the homogenous first order linear difference equation (2.3) is a linear
space of dimension one.
Proof. Suppose we are given two linearly independent solution (X (1) ) and
(X (2) ). Then according to Definition 2.1, for all constants a1 and a2 , not
both equal to zero,
(1) (2)
a1 X t + a2 X t 6= 0
(1) (2)
a1 Xt+1 + a2 Xt+1 6= 0.
(1) (1) (2) (2)
Inserting in the second inequality φXt for Xt+1 and φXt for Xt+1 leads to
(1) (2)
a1 X t + a2 X t 6= 0
(1) (2)
a1 φXt + a2 φXt 6= 0

or equivalently
! 
(1) (2)
Xt Xt a1
(1) (2) 6= 0
φXt φXt a2

Since this must hold for any a1 , a2 , not both equal to zero, the determinant
of the Casarotian matrix (see Definition 2.3)
!
(1) (2)
Xt Xt
C(t) = (1) (2)
φXt φXt
(1) (2) (1) (2)
must be nonzero. However, det C(t) = φXt Xt − φXt Xt = 0. This
is a contradiction to the initial assumption. Thus, there can only be one
independent solution.
8 CHAPTER 2. LINEAR DIFFERENCE EQUATIONS

The only independent solution is therefore given by (2.4). In Section 2.5

we will give a general proof and show that the dimension of the linear space
generated by the solutions to the homogenous equation of order p is p.
Consider now two solutions of the nonhomogeneous difference equation
(1) (2) (1) (2)
(2.2), (Xt ) and (Xt ), then, as can be easily verified, (Xt ) − (Xt ) sat-
isfies the homogenous equation (2.3). This fact is called the superposition
(1) (2)
principle.1 The superposition principle implies that Xt − Xt = φt c which
leads to the following theorem.

Theorem 2.2 (Superposition Principle). Every solution, (Xt ), of the first

order nonhomogeneous linear difference equation (2.2) can be represented as
(g)
the sum of the general solution of homogenous equation (2.3), (Xt ), and a
(p)
particular solution to the nonhomogeneous equation, (Xt ):
(g) (p)
Xt = Xt + Xt . (2.5)

The proof of this theorem is easily established and is left as an exercise to

(g)
the reader. In the case of a first order equation Xt = φt c. The Superposition
Principle then implies that the solution of the first order equation is given
by:
(p)
Xt = φt c + Xt .
Below we will discuss how to obtain a particular solution. We will also
see subsequently that this principle extends to higher order equations and
linear systems. The Superposition Principle thus delivers a general recipe for
solving linear difference equations in three steps:

1. Find the general solution of the homogeneous equation (X (g) ). This is

usually a technical issue that can be resolved mechanically.

2. Find a particular solution to the nonhomogeneous equation (X (p) ).

This step is usually more involved and requires additional (economic)
arguments. For example, we might argue that if the forcing variable Zt
remains bounded then also Xt should remain bounded.

3. The Superposition Principle (see Theorem 2.2) then delivers the gen-
eral solution of the nonhomogeneous equation as the sum of (X (g) ) and
(X (p) ). However, this solution still depends through (X (g) ) on some
1
The superposition principle means that the net response of Xt caused by two or
more stimuli is the sum of the responses which would have been caused by each stimulus
individually. In the first order case one stimulus comes from the general solution to the
homogeneous equation, the other from the particular solution to the nonhomogeneous
equation.
2.1. FIRST ORDER DIFFERENCE EQUATION 9

constants. To pin down the solution uniquely and therefore solving

the boundary value problem requires additional conditions. These con-
ditions can come in the form of initial values (starting values) or in
the form of requirements that the solution must obey some qualitative
feature. A typical feature in this context is boundedness, a condition
which usually can be given an economic underpinning.

Before continuing with the theoretical analysis consider the following ba-
sic example.

Amortization of a Loan
One of the simplest settings in economics where a difference equation arises
naturally, is compound interest calculation. Take, for example, the evolution
of debt. Denote by Dt the debt outstanding at the beginning of period t,
then the debt in the subsequent period t + 1, Dt+1 , is obtained by the simple
accounting rule:

Dt+1 = Dt + rDt − Zt = (1 + r)Dt − Zt (2.6)

where rDt is the interest accruing at the end of period t. Here we are using for
simplicity a constant interest rate r. The debt contract is serviced by paying
the amount Zt at the end of period t. This payment typically includes a
payment for the interest and a repayment of the principal. Equation (2.6)
constitutes a linear nonhomogeneous first order difference equation with φ =
1 + r.
Given the initial debt at the beginning of period 0, D0 , the amount of
debt outstanding in subsequent periods can be computed recursively using
the accounting rule (2.6):

D1 = (1 + r)D0 − Z0
D2 = (1 + r)D1 − Z1 = (1 + r)2 D0 − (1 + r)Z0 − Z1
...
Dt+1 = (1 + r)t+1 D0 − Zt − (1 + r)Zt−1 − · · · − (1 + r)t Z0
Xt
t+1
= (1 + r) D0 − (1 + r)i Zt−i
i=0

t+1
Note
Pt how Dt+1i is determined as the sum of two parts: (1 + r) D0 and
− i=0 (1 + r) Zt−i . The first expression thereby corresponds to the general
solution of the homogeneous equation and the second one to a particular
10 CHAPTER 2. LINEAR DIFFERENCE EQUATIONS

solution of the nonhomogeneous equation in accordance with Theorem 2.2.2

As the initial value of the debt is given, this value naturally pins down the
parameter c to equal D0 .
If the repayments Zt are constant over time and equal to Z, as is often
the case, we can bring Z outside the summation sign and use the formula for
geometric sums to obtain:

Z
Dt+1 = (1 + r)t+1 D0 − (1 + r)t+1 − 1 .
r
Suppose that the debt must be completely repaid by the beginning of period
T + 1, then the corresponding constant period payment Z can be calculated
by setting DT +1 = 0 in the above equation and solving for Z.3 This gives:
r
Z= D0 .
1 − (1 + r)−T −1

Note that the payment Z required to pay back the debt diminishes with
T. If T approaches infinity Z equals rD0 . In this case the payment is just
equal to interest accruing in each period so that there is no repayment of the
principal. In this case the debt is never paid back and equals the initial debt
D0 in each period. If the payment Z exceeds rD0 , the debt is repaid in a
finite amount of time.
Suppose that instead of requiring that the debt must be zero at some
point in time (including infinity), we impose the condition that the present
discounted value of the debt must be non-positive as T goes to infinity:

DT +1
lim ≤ 0. (2.7)
T →∞ (1 + r)T +1

This condition is referred to as the no Ponzi game (NPG) condition in eco-

nomics. A Ponzi game is a scheme where all principal repayments and interest
payments are rolled over perpetually by issuing new debt.4 If the above limit
is positive, the borrower would be able to extract resources (in present value
terms) from the lenders (See O’Connell and Zeldes (1988) and the literature
cited therein for an assessment of the significance of the NPG condition in
2
The reader is invited to check that the second expression is really a solution to the
nonhomogeneous equation.
3
The repayments are, of course, only constant as long as Dt > 0. Once the debt is paid
back fully, payments cease and Z = 0 from then on.
4
Charles Ponzi was an Italian immigrant who promised to pay exorbitant returns to
investors out of an ever-increasing pool of deposits. A historic account of Ponzi games can
be found in Kindleberger (1978).
2.2. STEADY STATE AND STABILITY 11

economics). Given the difference equation for the evolution of debt, the NPG
condition with constant payment per period is equivalent to:
DT +1
Z Z
lim T +1
= lim D0 − 1 − (1 + r)−T −1 = D0 − ≤ 0
T →∞ (1 + r) T →∞ r r
which implies that Z ≥ rD0 . Thus, the NPG condition holds if the constant
repayments Z are at least as great as the interest.

2.2 Steady State and Stability

Usually, we are not only interested in describing the evolution of the depen-
dent variable over time, but we also want to know some qualitative properties
of the solution. It is appropriate to formulate the relevant stability concepts
not just for linear but also for nonlinear difference equations. Thus, consider
the nonautonomous n-dimensional first order, possibly nonlinear, difference
equation Xt = f (Xt−1 , t) ∈ Rn , t ∈ N. Then, we can give the following
definition of an equilibrium point, fixed point or steady state.
Definition 2.4 (Equilibrium Point, Steady State). A point X ∗ ∈ Rn in the
domain of f is called an equilibrium point, a steady state, or a fixed point if
it is a fixed point of the function f (X, t), i.e. if X ∗ satisfies the equation
X ∗ = f (X ∗ , t), for all t ∈ N. (2.8)
In the case of a first order autonomous equation of dimension one, it is
convenient to represent the location of equilibrium points and the dynamics
of (Xt ) as a graph in the (Xt , Xt+1 )-plane. For this purpose, draw first the
graph of the function y = f (x) in the (Xt , Xt+1 ) plane. Then, draw the graph
of the identity function y = x which is just a line through the origin having
an angle of 450 with the x-axis. The equilibrium points are the points where
the 450 -line intersects with the graph of the function y = f (x). Starting
at some initial value X0 = x0 , the evolution of Xt is then represented in
the (Xt , Xt+1 )-plane by the following sequence of points: (0, x0 ), (x0 , f (x0 )),
(f (x0 ), f (f (x0 ))), (f (f (x0 )), f (f (f (x0 )))), . . . . Connecting these points by
line segments gives the so-called stair step or Cobweb diagram.
Take as an example the logistic function with µ > 0:
(
µx(1 − x), if 0 ≤ x ≤ 1;
f (x) = (2.9)
0, otherwise.

This provides an example with two steady states. The steady states are
determined by the equation: X ∗ = µX ∗ (1 − X ∗ ). This equation has two
12 CHAPTER 2. LINEAR DIFFERENCE EQUATIONS

0.9

0.8 steady
state 450−line
0.7

0.6
Xt+1

0.5

0.4
f(x)
0.3

0.2

0.1 steady
state
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
X
t

Figure 2.1: Cobweb diagram with steady states of the logistic function: y =
2.5x(1 − x) and x0 = 0.1

solutions which give the corresponding two steady states: X ∗ = 0 and X ∗ =

µ−1
µ
. Figure 2.1 shows the two steady states and the evolution of Xt starting
at X0 = x0 = 0.1 and taking µ = 2.5.
Another example consists of a linear first order nonhomogeneous differ-
ence equation with independent or forcing variable Zt constant over time, i.e.
Zt = Z. It is easy to compute the steady state in this simple case:
Z
X ∗ = φX ∗ + Z ⇒ X∗ = for φ 6= 1. (2.10)
1−φ
For φ = 1, there exists no equilibrium point, unless Z = 0 in which case
every point is an equilibrium point. As the steady state fulfills the difference
equation, it is a valid candidate for a particular solution. Thus, the general
solution in this case is:
Z
Xt = φ t c + X ∗ = φ t c + for φ 6= 1.
1−φ
Sometimes it is convenient to rewrite the nonhomogeneous equation as a
homogeneous equation in terms of deviations from steady state, Xt − X ∗ :5

Xt − X ∗ = φ(Xt−1 − X ∗ ).
5
In economics, equations are often log-linearized leading to log-deviations from steady
states. In which case, Xt − X ∗ can be interpreted as percentage deviations from steady
states.
2.2. STEADY STATE AND STABILITY 13

A major objective of the study of difference equations is to analyze its be-

havior near an equilibrium point. This topic is called stability theory. In the
context of linear difference equations the following basic concepts of stability
are sufficient.

Definition 2.5 (Stability). An equilibrium point X ∗ is called

• stable if for all ε > 0, there exists δε > 0 such that

|X0 − X ∗ | < δε implies |Xt − X ∗ | < ε for all t > 0. (2.11)

If X ∗ is not stable, it is called unstable.

• An equilibrium point X ∗ is called attracting if there exists η > 0 such

that
|X0 − X ∗ | < η implies lim Xt = X ∗ . (2.12)
t→∞
∗
If η = ∞, X is called globally attracting.

• The point X ∗ is asymptotically stable or is an asymptotically stable

equilibrium point 6 if it is stable and attracting. If η = ∞, X ∗ is called
globally asymptotically stable.

• An equilibrium point X ∗ is called exponentially stable if there exists

δ > 0, M > 0, and η ∈ (0, 1) such that for the solution Xt (x0 ) we have

|Xt (x0 ) − X ∗ | ≤ M η t |x0 − X ∗ |, whenever |x0 − X ∗ | < δ

• A solution Xt (x0 ) is called bounded if there exists a positive constant

M < ∞ such that

|Xt (x0 )| ≤ M, for all t.

Thereby the constant M may depend on x0 .

Note that X ∗ can be attracting, but unstable as shown in the following

example taken from Sedaghat (1997) and Elaydi (2005, 181–182). Consider
the difference equation

−2Xt−1 , for Xt−1 < µ;
Xt = f (Xt−1 ) =
0, otherwise,
6
In economics this is sometimes called stable
14 CHAPTER 2. LINEAR DIFFERENCE EQUATIONS

where µ > 0 is some given threshold. It is obvious that X ∗ = 0 is a fixed

point. The solution to this difference equation is

(−2)t x0 , if (−2)t−1 x0 < µ;
Xt =
0, if (−2)t−1 x0 ≥ µ,
where X0 = x0 is some starting value. If x0 ≥ µ, then Xt = 0 for all t ≥ 0.
If x0 < µ, then f (Xτ ) ≥ µ for some τ > 0. Thus, Xt = 0 for t ≥ τ . The
fixed point X ∗ = 0 is therefore attracting, even globally attracting. However,
X ∗ = 0 is unstable because points x0 6= 0, but arbitrarily close to zero, are
mapped to points further away until they exceed the threshold µ.
It can be shown that such a situation can only arise because f is not
continuous. In particular, if f is a continuous function on the real line a
fixed point cannot be simultaneously attracting and unstable (see Sedaghat,
1997; Elaydi, 2005).
A useful criterion for asymptotic stability of fixed points in a situation
where f is continuous, but not necessarily differentiable is provided by Elaydi
(2005, 182 and Appendix C).
Theorem 2.3 (Criterion Asymptotic Stability). A fixed point X ∗ of a con-
tinuous function f is asymptotically stable if and only if there exists an
open interval (a, b) containing X ∗ such that f 2 (x) > x for a < x < X ∗ and
f 2 (x) < x for X ∗ < x < b.
Proof. See Elaydi (2005).
The local stability of a fixed point can be studied by linearizing the non-
linear equation at the fixed point. In particular, the following theorem holds:
Theorem 2.4 (Stability Condition of Nonlinear Equation). Let X ∗ be an
equilibrium point of the nonlinear autonomous difference equation
Xt+1 = f (Xt )
where f is continuously differentiable at X ∗ . Then
1. If |f 0 (X ∗ )| < 1, then X ∗ is an asymptotically stable equilibrium point.
2. If |f 0 (X ∗ )| > 1, then X ∗ is unstable.
Proof. The proof follows Elaydi (2005, 27–28). Suppose that |f 0 (X ∗ )| ≤
M < 1. Then, because of the continuity of the derivative, there exists an
interval J = (X ∗ − γ, X ∗ + γ), γ > 0, such that |f 0 (X)| ≤ M < 1 for all
X ∈ J. For X0 ∈ J,
|X1 − X ∗ | = |f (X0 ) − f (X ∗ )|
2.2. STEADY STATE AND STABILITY 15

The mean value theorem then implies that there exists ξ, X0 < ξ < X ∗ , such
that
|f (X0 ) − f (X ∗ )| = |f 0 (ξ)| |X0 − X ∗ | .

Hence we have
|X1 − X ∗ | ≤ M |X0 − X ∗ | .

This shows that X1 is closer to X ∗ and is thus also in J because M < 1. By

induction we therefore conclude that

|Xt − X ∗ | ≤ M t |X0 − X ∗ | .

For any ε > 0, let δε = min{γ, ε} then |X0 − X ∗ | < δε implies |Xt − X ∗ | < ε
for all t ≥ 0. X ∗ is therefore a stable equilibrium point. In addition, X ∗
is attractive because limt→∞ |Xt − X ∗ | = 0. Thus, X ∗ is asymptotically
stable.

Example: Newton’s method Suppose we want to determine the solution

to the equation g(x) = 0 and suppose further that there is no analytic solution
so that we must solve the equation numerically. A well-known method is
the so-called Newton–Raphson method. Given some guess Xt , the method
consists in considering the linearized version g(X) + g 0 (Xt )(X − Xt ) = 0.
Solving this equation gives an approximate solution X = Xt − gg(X t)
0 (X ) . Taking
t
this solution as the new starting point Xt+1 results in a difference equation
Xt+1 = f (Xt ) = Xt − gg(X t)
0 (X ) . The steady state of this difference equation is
t
then a solution of the original equation, provided g 0 (X ∗ ) 6= 0.
In order to study the stability of the difference equation, we evaluate
0 ∗ 2 ∗ )g 00 (X ∗ ) 
|f 0 (X ∗ )| = 1 − (g (X ))(g0−g(X  = 0 because g(X ∗ ) = 0. Then, by

(X ∗ ))2
Theorem 2.4, X ∗ is asymptotically stable. This implies that limt→∞ Xt = X ∗
provided X0 is chosen close enough to X ∗ .
Note that the case |f 0 (X ∗ )| = 1 is not treated by this theorem. It involves
a more detailed analysis which involves higher order derivatives (see Elaydi,
2005, 29–32).

Remark 2.1. In the mathematical literature a function with |f 0 (X)| =6 1,

like in Theorem 2.4, is called hyperbolic. The general multivariate case is
treated in Section 3.3 under the heading of the Hartman–Groman theorem.
16 CHAPTER 2. LINEAR DIFFERENCE EQUATIONS

2.3 Solutions of First Order Linear Differ-

ence Equations
This section discusses a more systematic way of finding a particular solution
to the first order linear difference equation (2.2). For this purpose insert
recursively equation (2.2) into itself:

Xt = φXt−1 + Zt
Xt = φ(φXt−2 + Zt−1 ) + Zt = φ2 Xt−2 + φZt−1 + Zt
...
Xt = φt X0 + φt−1 Z1 + φt−2 Z2 + · · · + φZt−1 + Zt
t−1
X
= φt X0 + φj Zt−j
j=0

Taking the absolute value of the difference between Xt and the second term
of the right hand side of the equation leads to:
 t−1

 X   
 Xt − φj Zt−j  = φt X0  = |φt | |X0 |
 
 
j=0

When there is a starting period as in the example of the amortization of a

loan, say period 0 without loss of generality, we stop the backwards iteration
(p) Pt−1
at this period and take Xt = j=0 φj Zt−j as the particular solution. How-
ever in many instances there is no natural starting period so that it makes
sense to continue the above iteration into the infinite past. Given that |φt |
vanishes as t → ∞ if |φ| < 1, this suggests to consider
∞
X
(b)
Xt = φj Zt−j (2.13)
j=0

as a particular solution to the equation (2.2). The superscript (b) indicates

that the solution was obtained by iterating the difference equation backward
in time. For this to be a meaningful choice, the infinite sum must be well–
defined. This is, for example, the case if (Zt ) is a bounded sequence, i.e. if
(Z) ∈ `∞ . In particular, if Zt is constant and equal to Z, the above particular
solution becomes
∞
(b)
X Z
Xt = φj Z = , |φ| < 1,
j=0
1−φ
2.3. SOLUTIONS OF FIRST ORDER EQUATIONS 17

which is just the steady state solution described in equation (2.10) of section
2.2.
The requirement that Zt remains bounded can, for example, be violated
if Zt itself satisfies the homogenous difference equation Zt = ψZt−1 which
implies that Zt = ψ t c for some c 6= 0. Inserting this into equation (2.13)
then leads to ∞ ∞  j
(b)
X
j t−j t
X φ
Xt = φ ψ c=ψ c.
j=0 j=0
ψ

The infinite sum converges only if |φ / ψ| < 1. This shows that besides
the stability condition |φ| < 1, some additional requirements with respect
(b)
to the sequence of the exogenous variable are necessary to render Xt in
equation (2.13) a meaningful particular solution. Usually, we assume that
(Zt ) is bounded, i.e. that (Zt ) ∈ `∞ .
Consider next the case |φ| > 1. In this situation the above iteration is
(b)
no longer successful because Xt in equation (2.13) is not well–defined even
when Zt is constant.7 A way out of this problem is to consider the iteration
forward in time instead of backward in time:

Xt = φ−1 Xt+1 − φ−1 Zt+1

= φ−1 φ−1 Xt+2 − φ−1 Zt+2 − φ−1 Zt+1 = φ−2 Xt+2 − φ−2 Zt+2 − φ−1 Zt+1

...
Xh
−h −1
= φ Xt+h − φ φ−j+1 Zt+j for h ≥ 1.
j=1

Taking the absolute value of the difference between Xt and the second term
on the right hand side of the equation leads to:
 
 X h  
Xt + φ−1 φ−j+1 Zt+j  = φ−h Xt+h  = |φ−h | |Xt+h |.
  
 
j=1

As the economy is expected to live forever, there is no end period and the
forward iteration can be carried out indefinitely into the future. Because
|φ| > 1, the right hand side of the equation converges to zero as h → ∞,
provided that Xt+h remains bounded. This suggests the following particular
solution: ∞
(f )
X
−1
Xt = −φ φ−j+1 Zt+j , |φ| > 1, (2.14)
j=1

7
Note however that, if Zt satisfies itself a homogeneous difference equation of the form
Zt = ψZt−1 , the criterion for convergence is, as in the example above, |φ / ψ| < 1.
18 CHAPTER 2. LINEAR DIFFERENCE EQUATIONS

where the superscript (f ) indicates that the solution was obtained by iter-
ating the difference equation forward in time. For this to be a meaningful
choice, the infinite sum must be well-defined. This will be guaranteed if, for
example, Zt remains bounded, i.e. if (Z) ∈ `∞ .
In the case |φ| = 1 neither the backward nor the forward iteration strategy
leads to a sensible solution even when Zt is constant and equal to Z 6=
0. Either an equilibrium point does not exist as in the case φ = 1 or the
equilibrium point exists as is the case for φ = −1, but Xt oscillates forever
between X0 and −X0 + Z so that the equilibrium point is unstable. Most of
the time, we restrict ourself to the case of hyperbolic situations and exclude
the case |φ| = 1.
To summarize, assuming that (Zt ) is bounded, the first order linear dif-
ference equation (2.2) led us to consider the following two representations of
the general solutions:
∞
X
t (b) (b)
X t = φ cb + Xt , whereby Xt = φj Zt−j
j=0
∞
X
(f ) (f ) −1
t
X t = φ cf + Xt , whereby Xt = −φ φ−j+1 Zt+j
j=1

(b)
Note that these equations imply that cb = X0 − X0 , respectively that cf =
(f )
X0 − X0 . Depending on the value of φ, we can distinguish the following
three cases:
|φ| < 1: the backward solution is asymptotically stable in the sense that Xt
(b) (b)
approaches Xt as t → ∞. Any deviation of Xt from Xt vanishes
over time, irrespective of the value chosen for cb . The forward solution,
(f )
usually, makes no sense because Xt does not remain bounded even if
the forcing variable Zt is constant over time.

|φ| > 1: both solutions have an explosive behavior due to the term φt . Even
(b) (f )
small deviations from either Xt or Xt will grow without bounds.
There is, however, one and only one solution which remains bounded.
It is given by cf = 0 which implies that Xt always equals its equilibrium
(f )
value Xt .

|φ| = 1: neither the backward nor the forward solution converge for constant
Zt 6= 0.
Which solution is appropriate depends on the nature of the economic prob-
lem at hand. In particular, the choice of the boundary condition requires
2.4. EXAMPLES OF FIRST ORDER EQUATIONS 19

some additional thoughts and cannot be determined on general grounds. As

the exercises below demonstrate, the nature of the expectations formation
mechanism is sometimes decisive.

2.4 Examples of First Order Linear Differ-

ence Equations
2.4.1 The simple Cobweb Model
The Cobweb model, originally introduced by Moore (1914) to analyze the
cyclical behavior of agricultural markets, was one of the first dynamic mod-
els in economics. It inspired an enormous empirical as well as theoretical
literature. Its analysis culminated in the introduction of rational expecta-
tions by Muth (1961). The model, in its simplest form, analyzes the short-run
price fluctuations in a single market where, in each period, the price level is
determined to equate demand and supply denoted by Dt and St , respectively.
The good exchanged on this market is not storable and is produced with a
fixed production lag of one period. The supply decision of producers in pe-
riod t − 1 is based on the price they expect to get for their product in period
t. Denoting the logarithm of the price level in period t by pt and assuming
a negatively sloped demand curve and a positively sloped supply curve, the
simple Cobweb model can be summarized by the following four equations:8

Dt = −βpt , β>0 (demand)

St = γpet + ut , γ>0 (supply)
St = Dt (market clearing)
pet = pt−1 (expectations formation)

where ut denotes a supply shock. In agricultural markets ut typically repre-

sents weather conditions.
Given the naive expectations formation, pet = pt−1 , the model can be
solved to yield a linear first order difference equation in pt :
γ ut
pt = − pt−1 − = φpt−1 + Zt (2.15)
β β

where φ = − βγ and Zt = − uβt . Due to the negative value of φ, the price oscil-
lates: high prices tend to be followed by low prices which are again followed
by high prices. These price oscillations translate into corresponding quantity
8
The logarithm of the price level is taken to ensure a positive price level.
20 CHAPTER 2. LINEAR DIFFERENCE EQUATIONS

1.4

45°−line 1
1.2
p1
0.9
45°−line
1 0.8
P1
0.7
0.8 steady
0.6
Pt+1

steady state

Pt+1
state 0.5
0.6
−(γ/β)p + 1 0.4
t
0.4
0.3
P2
0.2 −(γ/β)pt + 1
0.2
p2 0.1

0 0
0 0.2 0.4 0.6 0.8 1 1.2 1.4 0 0.2 0.4 0.6 0.8 1
P0 P1 p0
Pt Pt p1

(a) convergence (β = 1 and γ = 0.8) (b) oscillating (β = 1 and γ = 1)

1
45°−line
0.9

0.8

p1 0.7
0.6 steady
state
Pt+1

0.5

0.4 −(γ/β)p + 1
t

p2 0.3
0.2

0.1

0
0 0.2 0.4 0.6 p 0.8 1
p0
Pt 1

(c) exploding (β = 1 and γ = 1.1)

Figure 2.2: Price dynamics in the cobweb model

oscillations. If ut is independent of time and equal to u, the equilibrium price

of the Cobweb model can be computed as follows:
γ u −u
p∗ = − p∗ − ⇒ p∗ = (2.16)
β β β+γ
The Figure 2.2 depicts several possible cases depending on the relative
slopes of supply and demand. In the first panel φ = −0.8 so that we have
an asymptotically stable equilibrium. Starting at p0 , the price approaches
the steady state by oscillating around it. In the second panel φ = −1 so
that independently of the starting value, the price oscillates forever between
p0 and p1 . In the third panel, we have an unstable equilibrium. Starting at
p0 6= p∗ , pt diverges.

2.4.2 The Solow Growth Model

Although this monograph only deals with linear difference equations, the lo-
cal behavior of nonlinear difference equations can be studied by linearizing
2.4. EXAMPLES OF FIRST ORDER EQUATIONS 21

the difference equation around the steady state. We will exemplify this tech-
nique by studying the famous Solow (see Solow (1956)) growth model. A
simple version of this model describes a closed economy with no technical
progress. Output in period t, denoted by Yt , is produced with two essen-
tial production factors: capital, Kt , and labor, Lt . Production possibilities
of this economy in period t are described by a neoclassical production func-
tion Yt = F (Kt , Lt ). This production function is defined on the nonnegative
orthant of R2 and is characterized by the following properties:
• F is twice continuously differentiable;
∂F (K,L) ∂F (K,L)
• strictly positive marginal products, i.e. ∂K
> 0 and ∂L
> 0;
∂ 2 F (K,L) ∂ 2 F (K,L)
• diminishing marginal products, i.e. ∂K 2
< 0 and ∂L2
< 0;
• F has constant returns-to-scale, i.e. F (λK, λL) = λF (K, L) for all
λ > 0;
• F satisfies the Inada conditions:
∂F (K, L) ∂F (K, L)
lim = 0, lim =0
K→∞ ∂K L→∞ ∂L
∂F (K, L) ∂F (K, L)
lim = ∞, lim =∞
K→0 ∂K L→0 ∂L
The Inada conditions are usually not listed among the properties of a neoclas-
sical production function, however, they turn out to be necessary to guar-
antee a strictly positive steady state. The classic example for a produc-
tion function with these properties is the Cobb-Douglas production function:
F (K, L) = AK (1−α) Lα , A > 0, 0 < α < 1. The above properties have two
important implications summarized by the following lemmata.
Lemma 2.1 (Essential Inputs). Let F be a neoclassical production func-
tion as described above then both inputs are essential, i.e. F (K, 0) = 0 and
F (0, L) = 0.
Proof. Suppose that Y → ∞ as K → ∞ then L’Hôpital’s rule together with
the Inada conditions imply
Y ∂Y /∂K ∂Y
lim = lim = lim = 0.
K→∞ K K→∞ 1 K→∞ ∂K
If on the other hand, Y remains finite when K → ∞, we immediately also
get
Y
lim = 0.
K→∞ K
22 CHAPTER 2. LINEAR DIFFERENCE EQUATIONS

The constant returns to scale assumption then implies that, for L > 0 fixed,
Y
lim = lim F (1, L/K) = F (1, 0) = 0.
K→∞ K K→∞

Using the constant returns to scale assumption again, we derive

F (K, 0) = KF (1, 0) = 0.
Thus, capital is essential. The proof that L is essential is analogous.
Lemma 2.2. Let F be a neoclassical production function as described above
then output goes to infinity if either input goes to infinity, holding the other
input fixed.
Proof. Omitting the time subscripts, define k as the capital intensity, i.e.
k = K/L. The assumption of constant returns to scale then implies that
F (K, L) = LF (K/L, 1) = Lf (k) = Kf (k)/k
where f (k) = F (k, 1). Holding K > 0 fixed, L’Hôpital’s rule implies
f (k)
lim F (K, L) = K lim = K lim f 0 (k) = ∞
L→∞ k→0 k k→0

where we have used the result that capital is essential (see the previous
Lemma), i.e. that f (0) = 0. The last equality is a consequence of the Inada
conditions. The proof for limK→∞ F (K, L) = ∞, L > 0 fixed, is analogous.

Output can be used either for consumption, Ct , or investment, It :

Yt = Ct + It . (2.17)
The economy saves a constant fraction s ∈ (0, 1) of the output. Because
saving equals investment in a closed economy, we have
It = sYt . (2.18)
Investment adds to the existing capital stock which depreciates in each period
at a constant rate δ ∈ (0, 1):
Kt+1 = (1 − δ)Kt + It = (1 − δ)Kt + sYt = (1 − δ)Kt + sF (Kt , Lt ) (2.19)
Whereas capital is a reproducible factor of production, labor is a fixed factor
of production which is assumed to grow at the exogenously given constant
rate µ > 0:
Lt+1 = (1 + µ)Lt , L0 > 0 given. (2.20)
2.4. EXAMPLES OF FIRST ORDER EQUATIONS 23

Starting in period 0 with some positive capital K0 > 0, the system con-
sisting of the two difference equations (2.19) and (2.20) completely describes
the evolution of the economy over time. A first inspection of the two equa-
tions immediately reveals that both labor and capital tend to infinity. Indeed,
as µ > 0 labor grows without bound implying according to Lemma 2.2 that
output also grows without bound. This is not very revealing if one is looking
for steady states and is interested in a stability analysis. In such a situation
it is often advisable to look at the ratio of the two variables, in our case at
K/L. This has two main advantages. First, the dimension of the system
is reduced to one and, more importantly, the singularity at infinity is, at
least in the linear case, eliminated.9 Second, these ratios often have a clear
economic meaning making the economic interpretation of the results more
comprehensible.
We apply this device to the Solow model as described by equations (2.19)
and (2.20). Thus, dividing equation (2.19) by Lt+1 and making use of the
constant returns to scale assumption results in the fundamental equation of
the Solow model:
Kt+1 1−δ s
kt+1 = = kt + f (kt ) = g(kt ) (2.21)
Lt+1 1+µ 1+µ
 
where kt = K Lt
t
is known as the capital intensity and f (kt ) = F Kt
Lt
, 1 .
The economy starts in period zero with an initial capital intensity k0 > 0.
The nonlinear first order difference equation (2.21) together with the initial
condition uniquely determines the evolution of the capital intensity over time,
and consequently of all other variables in the model. Note that the concavity
of F is inherited by f and thus by g so that we have g 0 > 0 and g 00 < 0.
Moreover, limk→0 g(k) = 0 and limk→∞ g(k) = ∞.
Proposition 2.1. Given the assumptions of the Solow model, the funda-
mental Solow equation (2.21) has two steady states k ∗ = 0 and k ∗ > 0.
Proof. The steady states must satisfy the nonlinear equation:
k ∗ = g(k ∗ ).

This equation implies

s
k∗ = f (k ∗ ). (2.22)
µ+δ
9
Technically speaking, this induces a new difference equation on the projective space.
In the two dimensional case, the projective space is defined as the set of rays through the
origin. As each ray crosses the unit circle twice, an equivalent definition is given as the
unit circle where opposite points are not distinguished. See Colonius and Kliemann (2014,
chapter 4) for details.
24 CHAPTER 2. LINEAR DIFFERENCE EQUATIONS

0.35

steady state
0.3

k*
0.25
capital intensity

linearized g(k) function

0.2

g(k)
0.15

45−degree line
0.1

0.05

0
0 0.05 0.1 0.15 0.2 0.25 k* 0.3
capital intensity

Figure 2.3: Capital Intensity in the Solow Model

As f (0) = 0, k ∗ = 0 is a steady state. From Figure 2.3 it becomes clear that

the Inada conditions guarantee the existence of a unique strictly positive
steady state k ∗ > 0. In particular, the properties g(0) = 0, limk→0 g 0 (k) =
∞, limk→∞ g 0 (k) = 1+µ1−δ
< 1, and g 0 (k) > 0 ensure that the function g is
sufficiently steep at the origin and becomes eventually flat enough to cross
the 45-degree line once from above.
The steady state k ∗ = 0 is of no economic significance. The asymptotic
stability of k ∗ > 0 is easily established by observing that (kt ) is monotonically
increasing for k0 ∈ (0, k ∗ ) and monotonically decreasing for k0 > k ∗ . Thus,
(kt ) converges monotonically to k ∗ independently of the initial value k0 > 0.
This shows that k ∗ > 0 is attracting. Monotonicity also implies stability
because for all ε > 0, taking δε = ε, |k0 − k ∗ | < δε implies |kt − k ∗ | < ε for
all t ≥ 0. Thus, k ∗ > 0 is stable and therefore asymptotically stable.
This fact can also be established by invoking Theorem 2.4. To do so, we
linearize equation (2.21) around the steady state k ∗ > 0. This amounts to
take a first order Taylor approximation:

∂g(k)
kt+1 ≈ k ∗ + (kt − k ∗ )

 (2.23)
∂k k=k∗

We can therefore study the local behavior of the nonlinear difference equa-
tion (2.21) around the steady state k ∗ > 0 by investigating the properties of
2.4. EXAMPLES OF FIRST ORDER EQUATIONS 25

the first order homogenous difference equation:

kt+1 − k ∗ = φ (kt − k ∗ ) (2.24)


∂g(k) 
where 0 < φ = ∂k  .
k=k∗

∂g(k) 
Proposition 2.2. 0 < φ = ∂k 
< 1.
k=k∗

Proof. Note that g 0 (k) > 0 for all k > 0. Concavity of g implies that g(k) −
k ∗ ≤ g 0 (k ∗ )(k − k ∗ ) for all k > 0. Take k < k ∗ , then g(k) > k. Thus,
g(k) − k ∗ < g 0 (k ∗ )(g(k) − k ∗ ) < 0 so that g 0 (k ∗ ) < 1.
Starting in period zero with an initial capital intensity k0 > 0, the solution
to this initial value problem is:

kt = k ∗ + φt (k0 − k ∗ )

As 0 < φ < 1, the steady state k ∗ is asymptotically stable.

2.4.3 A Model of Equity Prices

Consider an economy where investors have two assets at their disposal. The
first one is a riskless bank deposit which pays a constant interest rate r > 0
in each period. The second one is a common share which gives the owner the
right to a known dividend stream per share. The problem is to determine the
share price pt as a function of the future dividend stream (dt+h )h=0,1,... and the
interest rate r. As we abstract from uncertainty in this example, arbitrage
ensures that the return on both investments must be equal. Given that the
return on the investment in the share consists of the dividend payment dt
plus the expected price change pet+1 − pt , this arbitrage condition yields:

dt + pet+1 − pt
r= ⇔ pet+1 = (1 + r)pt − dt (2.25)
pt
where pet+1 denotes the price expected to prevail in the next period. Assuming
that expectations of the investors are rational which is equivalent to assum-
ing perfect foresight in the context of no uncertainty, the above arbitrage
equation turns into a simple first order difference equation:

pt+1 = (1 + r)pt − dt (2.26)

with φ = 1 + r and Zt = −dt−1 . Note that we are given no initial condition.

Instead, the purpose is to find a starting price (initial value).
26 CHAPTER 2. LINEAR DIFFERENCE EQUATIONS

Whereas the general solution to the homogeneous equation is easily found

(g)
to be Xt = φt cf , for some cf ∈ R, the search for an appropriate particular
solution to the nonhomogeneous equation requires additional considerations.
Because φ = 1+r > 1, we can disregard the backward solution as the infinite
sum will not converge for a constant P dividend stream. Thus, we turn to the
forward solution Xt = −φ−1 ∞
(f ) −j+1
j=1 φ Z t+j (see equation (2.14)). We
therefore envision the following general solution to the difference equation
(2.26):  
(f ) (f ) (f )
pt = (1 + r)t cf + Xt = p0 − X0 (1 + r)t + Xt (2.27)

where Xt = (1 + r)−1 ∞
(f ) P −j
j=0 (1 + r) dt+j . Note that the forward solution is
only well-defined if the infinite sum converges. A sufficient condition for this
to happen is the existence of a finite index j0 such that |dt+j /(1 + r)j | < M j ,
for j > j0 and some M < 1. This is guaranteed, in particular, by a constant
dividend stream dt+j = d, for all j = 0, 1, 2, . . .
The term (1 + r)t cf is usually called the bubble term because its behavior
(f )
is unrelated to the dividend stream; whereas the term Xt is referred to as
the fundamentals because it is supposed to reflect the “intrinsic value” of the
share.
Remember that we want to figure out the price of a share. Take period
(f )
0 to be the current period and suppose that cf = p0 − X0 > 0. This
means that the current stock price is higher than what can be justified by
the future dividend stream. According to the arbitrage equation (2.25) this
high price (compared to the dividend stream) can only be justified by an
appropriate capital gain, i.e. an appropriate expected price increase in the
next period. This makes the price in the next period even more different
from the fundamentals which must be justified by an even greater capital
gain in the following period, and so on. In the end, the bubble term takes
over and the share price becomes almost unrelated to the dividend stream.
This situation is, however, not sustainable in the long run.10 Therefore, the
(f )
only reasonable current share price p0 is X0 which implies that cf = 0. This
effectively eliminates the bubble term and is actually the only nonexplosive
solution. Thus, we have a unique (determinate) rational equilibrium solution.
This solution is
∞
X
(f ) −1
pt = Xt = (1 + r) (1 + r)−j dt+j (2.28)
j=0

Thus, the price of a share always equals the present discounted value of the
corresponding dividend stream. Such a solution is reasonable in a situation
10 (f )
A similar argument applies to the case cf = p0 − X0 < 0.
2.4. EXAMPLES OF FIRST ORDER EQUATIONS 27

with no uncertainty and no information problems. Note this solution implies

that the price immediately responds to any change in the expected dividend
stream. The effect of a change in dt+h , h = 0, 1, 2, . . . on pt is given by

∂pt
= (1 + r)−h−1 h = 0, 1, . . .
∂dt+h

Thus, the effect diminishes the further the change takes place in the fu-
ture. Consider now a permanent change in dividends, i.e. a change where all
dividends increase by some constant amount 4d. The corresponding price
change 4pt equals:
∞
−1
X 4d
4pt = (1 + r) (1 + r)−j 4d =
j=0
r

Similarly a proportional increase of all dividends would lead to the same

proportional increase in the share price. It also shows that relatively small
permanent changes in the dividends can lead to large fluctuations in the
share price. These “comparative” exercises demonstrate that the rational
expectations solution which eliminates the bubble term makes sense.

Cagan’s Model of Hyperinflation

In periods of hyperinflation the price level rises by more than 50 percent a
month. As these periods are usually rather short lived, they can serve as a
laboratory for the study of the relation between money supply and the price
level because other factors like changes in real output can be ignored. The
model also serves to illustrate the implications of alternative expectations
mechanisms, in particular the difference between adaptive and rational ex-
pectations. Denoting by mt the logarithm of the money stock in period t and
by pt the logarithm of the price level in period t, the model first proposed by
Cagan (1956) consists of the following three equations.11

mdt − pt = α(pet+1 − pt ), α<0 (money demand)

mst = mdt = mt (money supply)
pet+1 − pt = γ(pt − pt−1 ), γ>0 (adaptive expectations)

The first equation is a money demand equation in logarithmic form. It relates

the logged demand for the real money stock, mdt − pt , where the superscript
d stands for demand, to the rate of inflation expected to prevail in period
11
See also the analysis in Sargent (1987).
28 CHAPTER 2. LINEAR DIFFERENCE EQUATIONS

t+1, pet+1 −pt , where the superscript e stands for expectation. This relation is
negative because households and firms want to hold less money if they expect
the real value of money to deteriorate in the next period due to high inflation
rates. Thus, α < 0. In this model, the central bank perfectly controls the
money stock and sets it independently of the development of the price level.
The model treats the logarithm of the supply of the money stock, mst , where
the superscript s stands for supply, as exogenous. The money stock injected
in the economy is completely absorbed by the economy so that in each point
in time the supply of money equals the demand of money. Combining the
first two equations, i.e. replacing mdt by mt in the first equation, leads to a
portfolio equilibrium condition. As we will see, the behavior of the model
depends crucially on the way in which expectations are formed. Following the
original contribution by Cagan, we postulated that expectations are formed
adaptively, i.e. agents form their expectations by extrapolating past inflation.
The third equation postulates a very simple adaptive expectation formation
scheme: inflation expected to prevail in the next period is just proportional
to the current inflation. Thereby the proportionality factor γ is assumed
to be positive, meaning that expected inflation increases if current inflation
increases. Combining all three equations of the model and solving for pt , we
arrive at the following linear nonhomogeneous first order difference equation:
αγ 1
pt = pt−1 + mt = φpt−1 + Zt (2.29)
1 + αγ 1 + αγ
αγ 1
where φ = 1+αγ and Zt = 1+αγ mt .
From our previous discussion we know that the general solution of this
difference equation is given as the sum of the general solution to the ho-
(p)
mogenous equation and a particular solution, pt , to the nonhomogeneous
equation:
(p)
pt = φt c + pt
One particular solution can be found by recursively inserting into equation
(2.29):
p1 = φp0 + Z1
p2 = φp1 + Z2 = φ2 p0 + φZ1 + Z2
...
pt = φt p0 + φt−1 Z1 + φt−2 Z2 + · · · + φZt−1 + Zt
t−1
X
= φt p0 + φi Zt−i
i=0

This is again an illustration of the superposition principle. The logged price

in period t, pt , is just the sum of two components. The first one is a function
2.4. EXAMPLES OF FIRST ORDER EQUATIONS 29

of p0 whereas the second one is a weighted sum of past logged money stocks.
In economics there is no natural starting period so that one may iterate the
above equation further, thereby going back into infinite remote past:
∞
X
pt = lim φi pt−i + φi Zt−i
i→∞
i=0

From a mathematical point of view this expression only makes sense if the
limit of the infinite sum exists. Thus, additional assumptions are required.
Suppose that logged money remained constant, i.e. mt = m < ∞ for all
t, then the logic of the model suggests that the logged priceP leveli should
remain finite as well. In mathematical terms this means that ∞ i→∞ φ should
converge. This is, however, a geometric sum so that convergence is achieved
if and only if  
 αγ 
|φ| =  <1 (2.30)
1 + αγ 
Assuming that this stability condition holds, the general solution of the dif-
ference equation (2.29) implied by the Cagan model is:
∞
X
t
pt = φ c + φi Zt−i (2.31)
i=0

where the constant c can be computed from an initial value condition. Such
an initial condition arises naturally because the formation of adaptive expec-
tations requires the knowledge of the price from the previous period which
can then serve as an initial condition.
The stability condition therefore has important consequences. First, irre-
spective of the value of c, the first term of the solution (the general solution
to the homogenous equation), φt c, becomes less and less important as time
unfolds. Thus, for a large enough t, the logged price level willPbe dominated
by the particular solution to the nonhomogeneous equation, ∞ i
i=0 φ Zt−i . In
this infinite sum, the more recent values of the money stock are more impor-
tant for the determination of the price level. The importance of past money
stocks diminishes as one goes further back into the past. Third, suppose
that money stock is increased by a constant percentage point, ∆m, in every
period, then the effect on the logged price level, ∆pt is given by
∞  
X
i 1 1 1
∆pt = φ ∆m = ∆m = ∆m.
i=0
1 + αγ 1 − φ 1 + αγ

Thus, the price level moves up by the same percentage point. Such a once-
and-for-all change is termed a permanent change. In contrast a transitory
30 CHAPTER 2. LINEAR DIFFERENCE EQUATIONS

1.5

0.5

−0.5

−1

−1.5

−2
0 2 4 6 8 10 12 14 16 18 20
h

Figure 2.4: Impulse response function of the Cagan model with adaptive
expectations taking α = −0.5 and γ = 0.9

change is a change which occurs only once. The effect of a transitory change
of mt by ∆m in period t on the logged price level in period t + h for some
h ≥ 0 is given by
 h
h 1 αγ 1
∆pt+h = φ ∆m = ∆m
1 + αγ 1 + αγ 1 + αγ

The values ∆p t+h

∆m
seen as a function of h ≥ 0 are called the impulse response
function. It gives the reaction of the logged price level over time to a transi-
tory change of the logged money stock. As is clear from the above formula,
the stability condition implies that the effect on the logged price level dies
out exponentially over time. Usually the impulse response function is plotted
as a function of h as in Figure 2.4.
The character of the model changes drastically if rational expectations are
assumed instead of adaptive expectations. In the context of a deterministic
model this amounts to assuming perfect foresight. Thus, the third equation
of the model is replaced by
pet+1 = pt+1 (2.32)
With this change the new difference equation becomes:
α−1 mt
pt+1 = pt + = φpt + Zt (2.33)
α α
2.4. EXAMPLES OF FIRST ORDER EQUATIONS 31

with Zt = mt /α. As φ = α−1 α

> 1, the stability condition is violated. One
can nevertheless find a meaningful particular solution of the nonhomogeneous
equation by iterating the difference equation forwards in time instead of
backwards:

pt = φ−1 pt+1 − φ−1 Zt

= φ−1 φ−1 pt+2 − φ−1 Zt+1 − φ−1 Zt = φ−2 pt+2 − φ−2 Zt+1 − φ−1 Zt

...
Xh−1
−h −1
= φ pt+h − φ φ−i Zt+i for h > 0
i=0

The logged price level in period t, pt , now depends on some expected logged
price level in the future, pt+h , and on the development of logged money
expected to be realized in the future. Because the economy is expected to
live forever, this forward iteration is carried on into the infinite future to
yield:
∞
X
−h −1
pt = lim φ pt+h − φ φ−i Zt+i
h→∞
i=0

As 0 < φ−1 < 1, the limit and the infinite sum are well defined, provided
that the logged money stock remains bounded. Under the assumption that
the logged money stock is expected to remain bounded, the economic logic
of the model suggests that the logged price level should remain bounded as
well. This suggests the following particular solution to the nonhomogeneous
equation:
X∞
(p)
pt = −φ−1 φ−i Zt+i
i=0

by the superposition principle the general solution of the nonhomogeneous

difference equation (2.33) is:

∞
X
(p) −1
t
pt = φ c + pt t
=φ c−φ φ−i Zt+i (2.34)
i=0

Due to the term φt c, the logged price level grows exponentially without bound
although the logged money stock may be expected to remain bounded, unless
(p)
c = 0. Thus, setting c = 0 or equivalently p0 = p0 guarantees a nonexplosive
rational expectations equilibrium.
32 CHAPTER 2. LINEAR DIFFERENCE EQUATIONS

To summarize, the Cagan model suggests the following two solutions:

∞
X
t (b) (b)
p t = φ cb + pt , whereby pt = φi Zt−i
i=0
∞
X
(f ) (f )
pt = φt cf + pt , whereby pt = −φ−1 φ−i Zt+i
i=0

Which of the two solutions is appropriate depends on the value of φ. If

|φ| < 1 only the first solution delivers sensible paths for pt , i.e. paths which
do not explode for bounded values of Zt . However, we have a whole family
of paths parameterized by the constant cb . Only when we chose a particular
initial value for pt0 for some t0 , or equivalently a value for cb , will the price
path be uniquely determined. In this sense, we can say that the price level
is indeterminate. In the case |φ| > 1 which is implied by the assumption
of rational expectations, only the second solution is meaningful because it
delivers a well-defined particular solution for bounded Zt ’s. However, the
general solution to the homogenous equation implies an exploding price level
except for cf = 0. Thus, there is only one non-exploding solution in this case:
(f )
pt = pt . The price level therefore equals in each period its steady state level.
The assumption of rational expectations together with the assumption that
a bounded forcing variable should lead to a bounded price path pinned down
a unique solution. Thus, the price level is determined without the need of
an initial condition.

2.5 Difference Equations of Order p

We now turn to the general case represented by equation (2.1). As can
(1) (2)
be easily verified if (Xt ) and (Xt ) are two particular solutions of the
(1) (2)
nonhomogeneous equation, (Xt ) − (Xt ) is a solution to the homogeneous
equation:

Xt = φ1 Xt−1 + φ2 Xt−2 + · · · + φp Xt−p , φp 6= 0. (2.35)

Thus, the superposition principle stated in Theorem 2.2 also holds in the
general case: the general solution to the nonhomogeneous equations can
be represented as the sum of the general solution to the homogeneous and
a particular solution to the nonhomogeneous equation. Thus, we begin the
analysis of the general case by an investigation of the homogeneous equation.
2.5. DIFFERENCE EQUATIONS OF ORDER P 33

2.5.1 Homogeneous Difference Equation of Order p

In order to find the general solution of the homogeneous equation, we guess
that it will be of the same form as in the first order case, i.e. of the form
cλt , c 6= 0. Inserting this guess into the homogeneous equation (2.35), we get:

cλt = φ1 cλt−1 + φ2 cλt−2 + · · · + φp cλt−p

1
which after cancelling out c, dividing by λt and substituting z for λ
leads to:

1 − φ1 z − φ2 z 2 − · · · − φp z p = 0 (2.36)

This equation is called the characteristic equation of the homogeneous equa-

tion (2.35). Thus, in order for cλt to be a solution to the homogeneous
equation z = λ1 must be a root to the characteristic equation (2.36). These
roots are called the characteristic roots. Note that the assumption φp 6= 0
implies that none of the characteristic roots is equal to zero.
From the Fundamental Theorem of Algebra we know that there are p,
possibly complex, roots to the characteristic equation. Denote these roots
by z1 , . . . , zp and their corresponding λ0 s by λ1 , . . . , λp . To facilitate the
discussion consider first the standard case where all p roots are distinct.

distinct roots
In this case we have the following theorem.

Theorem 2.5 (Fundamental Set for equation of order p). If all the roots of
the characteristic equation are distinct, the set {λt1 , . . . , λtp } forms a funda-
mental set of solutions.

Proof. It suffices to show that det C(t) 6= 0 where C(t) is the Casarotian
matrix of {λt1 , . . . , λtp }.
 
λt1 λt2 ... λtp
 λt+1
1 λt+1
2 ... λt+1
p

det C(t) = det 
 
.. .. .. ..
.

 . . . 
λt+p−1
1 λ2t+p−1 . . . λt+p−1
p
 
1 1 ... 1
 λ1 λ2 . . . λ p 
t t t
= λ1 λ2 . . . λp det  .. .. .. 
 
..
 . . . . 
p−1 p−1 p−1
λ1 λ2 . . . λp
34 CHAPTER 2. LINEAR DIFFERENCE EQUATIONS

This second
Q matrix is called the Vandermonde matrix whose determinant
equals 1≤i<j≤p (λj − λi ) which is different from zero because the roots are
distinct. Thus, det C(t) 6= 0, because the roots are also different from zero.

The above Theorem thus implies that the general solution to the homo-
(g)
geneous equation Xt is given by
(g)
Xt = c1 λt1 + c2 λt2 + · · · + cp λtp . (2.37)

Using the same technique as in the proof of Theorem 2.1, it is easy to demon-
strate that the set of solutions forms a linear space of dimension p.

multiple roots
When the roots of the characteristic equation are not distinct, the situa-
tion becomes more complicated. Denote the r distinct roots by z1 , · · · , zr ,
r < p, and their corresponding multiplicities by m1 , · · · , mr . Writing the
homogeneous difference equation in terms of the lag operator leads to

(1 − φ1 L − · · · − φp Lp ) Xt
= (1 − λ1 L)m1 (1 − λ2 L)m2 · · · (1 − λr L)mr Xt = 0 (2.38)

where λi , 1 ≤ i ≤ r equals z1i . In order to find the general solution, we will

proceed in several steps. First note if ψt is a solution to

(1 − λi L)mi ψt = 0 (2.39)

it is also a solution to (2.38). Second, Gi = {λti , tλti , t2 λti , · · · , tmi −1 λti } is

a fundamental set of solutions for equation (2.39). Before we prove this
statement in Lemma 2.4, we need the following lemma.

Lemma 2.3. For all k ≥ 1

(1 − L)k ts = 0, 0≤s<k

Proof. The application of the operator 1 − L on ts leads to a polynomial of

degree s − 1, because the term ts cancels in (1 − L)ts = ts − (t − 1)s and only
terms of degree smaller than s remain. Applying 1 − L again reduces the
degree of the polynomial again by one. Finally, (1−L)s ts leads to a constant.
Thus, (1 − L)s+1 ts = 0. This proves the lemma because further applications
of 1 − L will again result in zero.
2.5. DIFFERENCE EQUATIONS OF ORDER P 35

Lemma 2.4. The set Gi = {λti , tλti , t2 λti , · · · , tmi −1 λti } represents a funda-
mental set of solutions to the equation (2.39).
Proof. Take s, 1 ≤ s ≤ mi − 1, then
(1 − λi L)mi (ts λti ) = λti (1 − L)mi (ts ) = 0
because (1−L)mi ts = 0 according to Lemma 2.3.12 Therefore ts λti is a solution
to (2.39). The set Gi is linearly independent because the set {1, t, t2 , · · · , tmi −1 }
is linearly independent.
It is then easily seen that G = ri=1 Gi is a fundamental set of solutions
S
to the equation (2.38). Thus, the general solution can be written as
r
X
ci0 + ci1 t + ci2 t2 + · · · + ci,mi −1 tmi −1 λti .


Xt = (2.40)
i=1

Again, the set of solutions forms a linear space of order p.

2.5.2 Nonhomogeneous Equation of Order p

As in the case of homogeneous difference equations of order one, the set
of all solutions forms a linear space. The dimension of this space is given
by the order of the difference equation, i.e. by p. Consider two solutions,
(X (1) ) and (X (2) ), of the nonhomogeneous equation. It is easy to verify that
(X (1) ) − (X (2) ) is then a solution to the homogeneous equation. This implies
that the superposition principle also applies to nonhomogeneous equations of
order p greater than one. Thus the general solution of the nonhomogeneous
equation can be written as before as
(g) (p)
X t = Xt + X t
(g)
where Xt is the general solution to the homogeneous equation given by
(p)
equation (2.40) and Xt is a particular solution to the nonhomogeneous
equation.
In the search for a particular solution, the same ideas as in first order case
can be used. If the nonhomogeneous part is constant, i.e. Zt = Z, the steady
state, if it exists, qualifies for a particular solution to the nonhomogeneous
equation. If the nonhomogeneous part depends on time, a particular solution
can be found by iterating the equation backwards and/or forwards depending
on the location of the roots. This will become clear by analyzing the examples
in section 2.5.4.
12
Here we made use of the relation P (L)(λt g(t)) = λt P ((λ−1 L))g(t) where P (L) is a lag
polynomial and g is any discrete function.
36 CHAPTER 2. LINEAR DIFFERENCE EQUATIONS

2.5.3 Limiting Behavior of Solutions

Before analyzing concrete examples, we turn to the qualitative behavior of
the solutions. In particular, we will explore the stability properties of the
steady states and the limiting behavior of the solutions, i.e. the behavior
when time goes to infinity. The analysis can be reduced to the discussion of
second order homogenous equations:

Xt − φ1 Xt−1 − φ2 Xt−2 = 0, φ2 6= 0. (2.41)

Higher order equations will add no new qualitative features. Assuming that
1 − φ1 − φ2 6= 0, the unique fixed point of this homogenous equation is 0. The
corresponding characteristic equation is given by the quadratic equation:

1 − φ1 z − φ2 z 2 = 0.

The solutions of this equation are given by the familiar formula:

p
φ1 ± φ21 + 4φ2
z1,2 = − .
2φ2

Or in terms of λ = z1 :
p
φ1 ± φ21 + 4φ2
λ1,2 = . (2.42)
2
To understand the qualitative behavior of Xt , we distinguish three cases:

λ1 and λ2 are real and distinct: The general solution is given by

"  t #
λ2
Xt = c1 λt1 + c2 λt2 = λt1 c1 + c2
λ1
 t
Suppose without loss of generality that |λ1 | > |λ2 | so that λλ21 → 0
as t → ∞. This implies that the behavior of Xt is asymptotically
governed by the larger root λ1 :

lim Xt = lim c1 λt1

t→∞ t→∞

Depending on the value of λ1 , six different cases emerge:

1. λ1 > 1: c1 λt1 diverges to ∞ as t → ∞. The fixed point zero is

unstable.
2.5. DIFFERENCE EQUATIONS OF ORDER P 37

2. λ1 = 1: c1 λt1 remains constant and Xt approaches c1 asymptoti-

cally. Starting the system with initial conditions X1 = X0 = x, x
arbitrary, which is equivalent to c1 = x and c2 = 0, Xt will remain
at this value x forever.
3. 0 < λ1 < 1: c1 λt1 decreases monotonically to zero. Zero is an
asymptotically stable fixed point.
4. −1 < λ1 < 0: c1 λt1 oscillates around zero, alternating in sign, but
converges to zero. Zero is again an asymptotically stable fixed
point.
5. λ1 = −1: c1 λt1 alternates between the values c1 and −c1 . Thus,
the sequence (Xt ) will have two accumulation points c1 and −c1 .
6. λ1 < −1: c1 λt1 alternates in sign, but diverges in absolute value
to ∞. The fixed point zero is unstable.

The behavior of Xt in all six cases is illustrated in Figure 2.5.

equal roots λ = λ1 = λ2 : According to (2.40) the solution is given by: Xt =

(c1 + c2 t) λt . Clearly, if λ ≥ 1, Xt diverges monotonically; or, if λ ≤ −1,
Xt diverges alternating signs. For |λ| < 1, the solution converges to
zero, because limt→∞ tλt = 0.

complex roots: The two roots appear as complex conjugate pairs and may
be written as λ1 = α+ıβ and λ2 = α−ıβ with β 6= 0. In terms of polar
coordinates the two roots may alternatively
p be written as λ1 = reıθ ,
respectively λ2 = re−ıθ , where r = α2 + β 2 and θ = tan−1 αβ . The

solution is then given by

Xt = c1 λt1 + c2 λt2 = c1 (α + ıβ)t + c2 (α − ıβ)t

= c1 rt eıθt + c2 rt e−ıθt
= rt [c1 (cos(θt) + ı sin(θt)) + c2 (cos(θt) − ı sin(θt))]
= rt [(c1 + c2 ) cos(θt) + ı(c1 − c2 ) sin(θt)]

Since Xt must be a real number, c1 + c2 must also be real whereas

c1 − c2 must be purely imaginary. This implies that c1 and c2 must be
complex conjugate. In terms of polar coordinates they can be written
as c1 = ρeıω and c2 = ρe−ıω for some ρ and some ω. Inserting into the
above equation finally gives:

Xt = ρrt eı(θt+ω) + e−ı(θt+ω)

 

= 2ρrt cos(θt + ω)
38 CHAPTER 2. LINEAR DIFFERENCE EQUATIONS

3 3 3

2 2 2

1 1 1

Xt

Xt

Xt
0 0 0

−1 −1 −1

−2 −2 −2

−3 −3 −3
0 5 10 0 5 10 0 5 10
case 1: λ1 = 1.1 case 2: λ1 = 1 case 3: λ1 = 0.8

3 3 3

2 2 2

1 1 1
Xt

Xt

Xt
0 0 0

−1 −1 −1

−2 −2 −2

−3 −3 −3
0 5 10 0 5 10 0 5 10
case 4: λ1 = −0.8 case 5: λ1 = −1 case 6: λ1 = −1.1

Figure 2.5: Behavior of Xt = λt1 depending on λ ∈ R

The solution therefore clearly oscillates because the cosine function

oscillates. Depending on the location of the conjugate roots three cases
must be distinguished:

1. r > 1: both roots are outside the unit circle (i.e. the circle of
radius one and centered in the point (0, 0)). Xt oscillates, but
with ever increasing amplitude. The fixed point zero is unstable.
2. r = 1: both roots are on the unit circle. Xt oscillates, but with
constant amplitude.
3. r < 1: both roots are inside the unit circle. The solution oscillates,
but with monotonically decreasing amplitude and converges to
zero as t → ∞. The fixed point zero is asymptotically stable.

Figure 2.6 illustrates the three cases.

We can summarize the above discussion in the following theorem.
Theorem 2.6 (Limiting Behavior of Second Order Equation). The following
statements hold in the case of linear homogenous difference equation of order
two (equation (2.41)):
(i) All solutions oscillate around zero if and only if the equation has no
positive real characteristic root.
2.5. DIFFERENCE EQUATIONS OF ORDER P 39

20

0
Xt

−20

−40
0 1 2 3 4 5 6 7 8 9 10
case 1: roots outside unit circle (λ = 1+i, λ = 1−i)
1 2

1
t

0
X

−1

0 1 2 3 4 5 6 7 8 9 10
case 2: roots on unit circle (λ1 = (sqrt(2)/2)(1+i), λ2 = (sqrt(2)/2)(1−i))

1
t

0
X

−1

0 1 2 3 4 5 6 7 8 9 10
case 3: roots in unit circle (λ = 0.5(1+i), λ = 0.5(1−i))
1 2

Figure 2.6: Behavior of Xt in case of complex roots

(ii) All solutions converge to zero (i.e. zero is an asymptotically stable

steady state) if and only if max{|λ1 |, |λ2 |} < 1.

Although the limiting behavior of Xt is most easily understood in terms

of the roots of the characteristic equation, it is sometimes more convenient
to analyze the properties of the difference equation in terms of the original
parameters φ1 and φ2 . Consider for this purpose a nonhomogeneous second
order difference equation where the nonhomogeneous part is just a constant
equal to Z:
Xt = φ1 Xt−1 + φ2 Xt−2 + Z, Z 6= 0. (2.43)

Zero is no longer an equilibrium point. Instead, the new equilibrium point

X ∗ can be found by solving the equation:

Z
X ∗ = φ1 X ∗ + φ2 X ∗ + Z ⇒ X∗ =
1 − φ1 − φ2

Note that an equilibrium only exists if 1 − φ1 − φ2 6= 0. This condition is

equivalent to the condition that 1 cannot be a root. As the steady state
qualifies for a particular solution of equation (2.43), the general solution is
given by
(g)
Xt = X ∗ + Xt
40 CHAPTER 2. LINEAR DIFFERENCE EQUATIONS

(g)
Thus, Xt converges to its equilibrium if and only if Xt converges to zero as
(g)
t → ∞. Moreover, the solution oscillates around X ∗ if and only if Xt oscil-
lates around zero. Based on the theorem just above, the following theorems
hold.
Theorem 2.7 (Limiting Behavior of Second Order Equation (original pa-
rameters)). Assuming 1 − φ1 − φ2 6= 0, the following statements hold.
(i) All solutions of the nonhomogeneous equation (2.43) oscillate around
the equilibrium point X ∗ if and only if the characteristic equation has
no positive real characteristic root.
(ii) All solutions to the nonhomogeneous equation (2.43) converge to X ∗
(i.e. X ∗ is asymptotically stable) if and only if max{|λ1 |, |λ2 |} < 1.
Theorem 2.8 (Stability Conditions of Second Order Equation (original pa-
rameters)). The equilibrium point X ∗ is asymptotically stable (i.e. all solu-
tions converge to X ∗ ) if and only if the following three conditions are satisfied:
(i) 1 − φ1 − φ2 > 0
(ii) 1 + φ1 − φ2 > 0
(iii) 1 + φ2 > 0
Proof. Assume that X ∗ is an asymptotically stable equilibrium point. Ac-
cording to the previous Theorem (2.6), this means that both λ1 and λ2 must
be smaller than one in absolute value. According to equation (2.42) this
implies that
   
 φ + pφ2 + 4φ   φ − pφ2 + 4φ 
 1 1 2  1 1 2
|λ1 | =   < 1 and |λ2 | =  <1.
 2   2 

Two cases have to be distinguished.

real roots: φ21 + 4φ2 > 0: This implies the set of inequalities:
q
−2 < φ1 + φ21 + 4φ2 < 2
q
−2 < φ1 − φ21 + 4φ2 < 2

or, equivalently,
q
−2 − φ1 < φ21 + 4φ2 < 2 − φ1
q
−2 − φ1 < − φ21 + 4φ2 < 2 − φ1
2.5. DIFFERENCE EQUATIONS OF ORDER P 41

Squaring the second inequality in the first line implies: φ21 + 4φ2 <
4 − 4φ1 + φ21 which leads to condition (i). Similarly, squaring the first
inequality in the second line yields: 4 + 4φ1 + φ21 > φ21 + 4φ2 which
results in condition (ii). The assumption |λ1 | < 1 and |λ2 | < 1 imply
that |λ1 λ2 | = | − φ2 | < 1 which gives condition (iii).
complex roots: φ21 + 4φ2 < 0: This implies that 0 < φ21 < −4φ2 . There-
fore
4(1 − φ1 − φ2 ) > 4 − 4φ1 + φ21 = (2 − φ1 )2 > 0
which is equivalent to condition (i). Similarly,
4(1 + φ1 − φ2 ) > 4 + 4φ1 + φ21 = (2 + φ1 )2 > 0
which is equivalent to condition (ii). In order to obtain condition (iii),
note that the two complex conjugate roots are given by
φ1 ı φ1 ı
q q
2
λ1 = + φ1 + 4φ2 and λ2 = − φ21 + 4φ2 .
2 2 2 2
Because |λ1 | < 1 and |λ2 | < 1 by assumption, we have that |λ1 λ2 | =
| − φ2 | < 1 which is condition (iii).
Assume now that the three conditions are satisfied. They immediately imply
that −2 < φ1 < 2 and that −1 < φ2 < 1. If the roots are real then
p p
−2 + φ21 + 4φ2 φ1 + φ21 + 4φ2
−1 < < λ1 =
2 p 2
φ1 + φ21 + 4 − 4φ1
<
q 2
φ1 + (2 − φ1 )2
=
2
φ1 − φ1 + 2
= =1
2
Similarly,
p p
2− φ21 + 4φ2 φ1 − φ21 + 4φ2
1> > λ2 =
2 p 2
2
φ1 − φ1 + 4 + 4φ1
>
q 2
φ1 − (φ1 + 2)2
=
2
φ1 − φ1 − 2
= = −1
2
42 CHAPTER 2. LINEAR DIFFERENCE EQUATIONS

2
explosive explosive
oscillations growth
1
asymptotically
stable

φ2
0

-1
explosive oscillations

-2
parabola: φ 21 + 4φ 2 = 0

-3
-3 -2 -1 0 1 2 3
φ1

Figure 2.7: Stability properties of equation: Xt − φ1 Xt−1 − φ2 Xt−2 = 0

If the roots are complex, λ1 and λ2 are complex conjugate numbers. Their
φ2 −(φ2 +4φ )
squared modulus then equals λ1 λ2 = 1 41 2 = −φ2 . As −φ2 < 1, the
modulus of both λ1 and λ2 is smaller than one.

The three conditions listed above determine a triangle in the φ1 -φ2 -plane
with vertices (−2, −1), (0, 1) and (2, −1). Points inside the triangle imply
an asymptotically stable behavior whereas points outside the triangle lead
to an unstable behavior. The parabola φ21 + 4φ2 = 0 determines the region
of complex roots. Values of φ1 and φ2 above the parabola lead to real roots
whereas values below the parabola lead to complex roots. The situation is
represented in Figure 2.7.

2.5.4 Examples
Multiplier Accelerator model
A classic economic example of a second order difference equation is the
multiplier-accelerator model originally proposed by Samuelson (1939). It
was designed to demonstrate how the interaction of the multiplier and the
accelerator can generate business cycles. The model is one of a closed econ-
omy and consists of a consumption function, an investment function which
2.5. DIFFERENCE EQUATIONS OF ORDER P 43

incorporates the accelerator idea and the income identity:

Ct = α + βYt−1 , 0 < β < 1, α > 0 (consumption)

It = γ(Yt−1 − Yt−2 ), γ>0 (investment)
Yt = Ct + It + Gt , (income identity)

where Ct , It , Yt , and Gt denotes private consumption expenditures, invest-

ment expenditures, income, and government consumption, respectively. The
parameter β is called the marginal propensity of consumption and is assumed
to be between zero and one. The remaining parameters of the model, α and
γ, bear no restriction besides that they have to be positive. Inserting the
consumption and the investment equation into the income identity leads to
the following nonhomogeneous second order difference equation:

Yt = (β + γ)Yt−1 − γYt−2 + (α + Gt ) (2.44)

If government expenditures remain constant over time and equal to G, the

equilibrium point Y ∗ for equation (2.44) can be computed as follows:
α+G
Y ∗ = (β + γ)Y ∗ − γY ∗ + α + G ⇒ Y∗ =
1−β
The stability of this equilibrium point can be investigated by verifying if the
three conditions of Theorem 2.8 are satisfied:
(i) 1 − (β + γ) + γ = 1 − β > 0
(ii) 1 + (β + γ) + γ = 1 + β + 2γ > 0
(iii) 1 − γ > 0
Given the assumptions of the model, the first two conditions are automati-
cally satisfied. The third condition, however, is only valid if the accelerator is
not too strong, i.e. if γ < 1. The steady state Y ∗ is therefore asymptotically
stable if one imposes this additional requirement. Yt oscillates around its
steady state if, according to Theorem 2.6, there is no real positive inverse
root of the characteristic equation. The inverse of the characteristic roots
are given by p
(β + γ) ± (β + γ)2 − 4γ
λ1,2 = .
2
If the roots are real, they are both strictly positive and strictly smaller than
one. Thus, Yt can only oscillate around its steady state if and only if the
roots are complex, i.e. if (β + γ)2 − 4γ < 0. If they are complex, their moduli
are strictly smaller than one.
44 CHAPTER 2. LINEAR DIFFERENCE EQUATIONS

In the general case where government expenditures are not constant, but
vary over time, we apply the method of undetermined coefficients to find
(p)
a particular solution, Yt , to equation (2.44). This method conjectures a
certain type of solution and then tries to pin down a solution by inserting it
into the difference equation. In the particular case at hand, the roots of the
characteristic function are all outside the unit circle. Thus, we conjecture a
particular solution of the form:
∞
X
(p)
Yt =c+ ψi Gt−i
i=0

The coefficients ψj are called impulse responses or dynamic multipliers. They

trace the effect on output of an impulse (stimulus) in government expen-
ditures over time. Thereby a unit impulse is specified as ∆Gt = 1 and
∆Gt−i = 0 for i 6= 0. The effect on output is then
∞
X
(p)
∆Yt+h = ψi ∆Gt+h−i = ψh ∆Gt = ψh , h = 0, 1, 2, . . .
i=0

Inserting this conjectured particular solution into the difference equation

leads to
∞
X ∞
X
c+ ψi Gt−i =c(β + γ) + (β + γ) ψi Gt−1−i
i=0 i=0
∞
X
− cγ − γ ψi Gt−2−i + α + Gt
i=0

Equating the constant terms leads to an equation for c:

α
c (1 − (β + γ) + γ) = α ⇒ c = >α>0
1−β
Equating the terms for Gt−i , i = 0, 1, · · · leads to:
ψ0 = 1
ψ1 = (β + γ)ψ0 ⇒ ψ1 = β + γ
ψ2 = (β + γ)ψ1 − γψ0
···
ψj = (β + γ)ψj−1 − γψj−2 , j≥2
Thus, the coefficients ψj , j ≥ 2, follow the same homogenous second order
difference equation with initial values ψ0 = 1 and ψ1 = β + γ. The solution
can therefore be written as
ψj = d1 λj1 + d2 λj2 .
2.5. DIFFERENCE EQUATIONS OF ORDER P 45

The coefficients d1 and d2 can then be determined from the initial conditions:

ψ0 = 1 = d1 + d2
ψ1 = β + γ = d1 λ1 + d2 λ2

In order to illustrate the behavior of the multiplier-accelerator model, we

discuss several numerical examples.
4 1
β= 5
and γ = 5
In this case both roots are real and equal to
√ (
1 5 0.7236
λ1,2 = ± =
2 10 0.2764

Therefore the impulse response coefficients ψj for j ≥ 0 are given by

ψj = d1 λj1 + d2 λj2 . The constants d1 and d2 can be recovered from the
initial conditions: ψ0 = 1 = d1 + d2 and ψ1 = 1 = d1 λ1 + d2 λ2 . Solving
these two equations for d1 and d2 yields:
1 − λ2
d1 = = 1.6180
λ1 − λ2
λ1 − 1
d2 = = −0.6180
λ1 − λ2
The corresponding impulse response function is plotted in Figure 2.8.
The initial increase of government expenditures by one unit raises out-
put in current and the subsequent period by one unit. Then the effect of
the impulse dies out monotonically. After ten periods the effect almost
vanished.
3
β= 4
and γ = 14 In this case we have a multiple root equal to λ = 0.5. Ac-
cording to equation (2.40) the impulse response coefficients are there-
fore given by ψj = (d0 + d1 t)λt . The constants d0 and d1 can again
be found by solving the equation system: ψ0 = 1 = d0 and ψ1 = 1 =
(d0 + d1 )λ. The solution is given by d0 = 1 and d1 = 1. The corre-
sponding impulse response coefficients are plotted in Figure 2.8. They
resemble very much to those of the previous case. They even die out
more rapidly.
2
β= 3
and γ = 23 In this case the discriminant is negative so that we have
two complex conjugate roots:
1 √ 
λ1,2 = 2±ı 2
3
46 CHAPTER 2. LINEAR DIFFERENCE EQUATIONS

1.4

1.2

0.8

0.6
real roots:
β = 4/5, γ = 1/5
0.4 multiple roots:
β = 3/4, γ = 1/4
0.2

−0.2
complex roots:
−0.4
β = 2/3, γ = 2/3

−0.6
0 2 4 6 8 10 12 14 16 18 20
period

Figure 2.8: Impulse Response Coefficients of the Multiplier-Accelerator

model

The constants can again be found by solving the equation system: ψ0 =

1 = d1 + d2 and ψ1 = 43 = d1 λ1 + d2 λ2 . The solution is given by
√
1 2
d1 = −ı
2 √2
1 2
d2 = +ı
2 2
The corresponding impulse response coefficients are plotted in Fig-
ure 2.8. As expected they clearly show an oscillatory behavior. Due
to the accelerator, the initial impulse is amplified in period one. The
effect is around 1.3. After period one the effect rapidly declines and
becomes even negative in period four. However, in period seven the
effect starts to increase and becomes again positive in period ten. As
is also evident from the Figure, these oscillatory movements die out.

Cobweb model with Inventory

In this example we extend the simple Cobweb model analyzed in subsection
2.4 by allowing the good to be stored (see Sargent, 1987). In addition, we
assume that expectations are rational which in the context of a deterministic
2.5. DIFFERENCE EQUATIONS OF ORDER P 47

model is equivalent to perfect foresight. These extensions will lead to fur-

ther insights into the method of undetermined coefficients introduced in the
previous example. The new set of equations then reads as follows:

Dt = −βpt , β>0 (demand)

St = γpet + ut , γ>0 (supply)
It = α(pet+1 − pt ), α>0 (inventory demand)
St = Dt + (It − It−1 ), (market clearing)
pet = pt , (perfect foresight)

where ut denotes again a supply shock. The inventory demand schedule

incorporates a speculative element because inventories will be built up if
prices are expected to be higher next period. The market clearing equation
shows that the supply which remains unsold is used to build up inventories;
on the other hand demand can not only be served by newly supplied goods,
but can also be fulfilled out of inventories. Combining these equations leads
to the following second order linear difference equation in the price:
γ + 2α + β ut
pt+1 = pt − pt−1 + (2.45)
α α
γ+2α+β
Setting φ = α
, the characteristic equation becomes:

1 − φz + z 2 = 0

This equation implies that the two roots, z1 and z2 , are given by
p
φ ± φ2 − 4
z1,2 =
2
First note that because φ > 2 the roots are real, distinct, and positive.
Second they come in reciprocal pairs as z1 z2 = 1. Thus, one root is smaller
than one whereas the other is necessarily greater than one. Thus, we have
one stable and one explosive root. The solution to the homogenous equation
can therefore be written as

pt = c1 λt + c2 λ−t

where, without loss of generality, z1 = λ < 1 and z2 = λ1 . c1 and c2 are

constants yet to be determined.
Because the agents in this model have rational expectations which implies
that they are forward looking, they will incorporate expected future devel-
opments of the supply shock into their decision. However, past decision are
48 CHAPTER 2. LINEAR DIFFERENCE EQUATIONS

reflected in the inventories carried over last period. Thus, we conjecture that
the solution will have both a forward and a backward looking component.
Thus, we seek for a particular solution of the following form:
∞
X
pt = ψj ut−j
j=−∞

Following the method of undetermined coefficients we insert this guess into

the difference equation to get:
∞ ∞ ∞
X X X ut
ψj ut+1−j = φ ψj ut−j − ψj ut−1−j +
j=−∞ j=−∞ j=−∞
α

Writing this equation in extensive form leads to:

· · · + ψ−1 ut+2 + ψ0 ut+1 + ψ1 ut + ψ2 ut−1 + ψ3 ut−2 + · · ·

= · · · + φψ−2 ut+2 + φψ−1 ut+1 + φψ0 ut + φψ1 ut−1 + φψ2 ut−2 + · · ·
− · · · − ψ−3 ut+2 − ψ−2 ut+1 − ψ−1 ut − ψ0 ut−1 − ψ1 ut−2 − · · ·
ut
+
α
Equating terms for ut−j , j = . . . , −2, −1, 0, 1, 2, . . . gives:

···
ut+2 : ψ−1 = φψ−2 − ψ−3
ut+1 : ψ0 = φψ−1 − ψ−2
1
ut : ψ1 = φψ0 − ψ−1 +
α
ut−1 ψ2 = φψ1 − ψ0
ut−2 : ψ3 = φψ2 − ψ1
···

This shows that the ψj ’s follow homogenous second order difference equa-
tions:

ψj = φψj−1 − ψj−2 j≥1

ψ−j = φψ−j−1 − ψ−j−2 j≥1

The solution to these difference equations are:

ψj = d1 λj + d2 λ−j
ψ−j = e1 λj + e2 λ−j
2.5. DIFFERENCE EQUATIONS OF ORDER P 49

where the constants d1 , d2 , e1 , e2 have yet to be determined. A sensible eco-

nomic solution requires that, if the supply shock has been constant in the
past and is expected to remain constant in the future, the price must be
constant too. Thus, we can eliminate the exploding parts of the above so-
lutions, setting d2 = 0 and e2 = 0. Next observe that both solutions must
coincide for j = 0 which implies that d1 = e1 . Denote this value by d. d can
be determined by observing that the solutions must satisfy the initial value
condition: ψ1 = φψ0 − ψ−1 + α1 . Inserting the solutions for ψ1 , ψ0 , ψ−1 leads
to:
1 α−1
dλ = φd − dλ + ⇒ d=
α λ − λ−1
The general solution to the Cobweb model with inventory represented by the
difference equation (2.45) is therefore given by
∞
α−1 X
pt = c1 λt + c2 λ−t + λ|j| ut−j (2.46)
λ − λ−1 j=−∞

If we impose again the requirement that the price must be finite if the supply
shock has always been constant and is expected to remain constant in the
future, we have to set c1 = 0 and c2 = 0 to get the solution:
∞
α−1 X
pt = λ|j| ut−j (2.47)
λ − λ−1 j=−∞

This implies that {pt } is a bounded sequence, i.e. that (pt ) ∈ `∞ . In this
case the price pt is just a function of all past shocks and all expected future
shocks.
Another way to represent this solution is to express pt as
X∞
pt = λpt−1 − α−1 λ λj ut+j .
j=0

In this expression the double infinite sum is replaced by a single one. This is
due to the fact that the past evolution of supply is now summarized by pt−1
which is supposed to be known in period t. The effect of discounted expected
future supply is just as before.
In order to gain a better understanding of the dynamics, we will analyze
the following numerical example. In this example α = 20 9
and the parameters
β and γ are such that φ = 2.05. This implies that β + γ = 19 . The roots are
then given by λ = 0.8 and λ−1 = 1.25. The bounded solution is then given
by
X∞
pt = − 0.8|j| ut−j
j=−∞
50 CHAPTER 2. LINEAR DIFFERENCE EQUATIONS

Suppose that the supply shock has been constant forever and is expected
to remain constant at u. The above formula then implies that the logged
price level pt equals −9u and that It = 0. Suppose that an unexpected and
transitory positive supply shock of value 1 hits the market in period 0. Then
according to the first panel in Figure 2.9 the price immediately falls by 1.
At the same time inventories rise because prices are expected to move up in
the future due to the transitory nature of the shock. Here we have a typical
price movement: the price falls, but is expected to increase. After the shock
the market adjusts gradually as prices rise to their old level and by running
down inventories.
Consider now a different gedankenexperiment. Suppose that the shock
is not unexpected, but expected to hit the market only in period 5. In
this case, we see a more interesting evolution of prices and inventories. In
period zero when the positive supply shock for period 5 is announced, market
participants expect the price to fall in the future. They therefore want to
get rid of their inventories by trying to selling them already now.13 As a
result, the price and the inventories start to fall already before the supply
shock actually takes place. In period 5 when the supply shock finally hits the
market, market participants expect the price to move up again in the future
which leads to a buildup of inventories. Note that this buildup is done when
the price is low. From period 5 on, the market adjusts like in the previous
case because the supply shock is again assumed to be transitory in nature.

Taylor model
In this example we analyze a simple deterministic version of Taylor’s stag-
gered wage contract model which also has a backward and forward component
(see Taylor (1980) and Ashenfelter and Card (1982)).14 In this model, half
of the wages have to be contracted in each period for two periods. Thus, in
each period half of the wages are renegotiated taking the wages of the other
group as given. Assuming that the two groups are of equal size, wages are
set according to the following rule:

wt = 0.5wt−1 + 0.5wt+1 + h(yt + yt+1 ), h > 0. (2.48)

Thus, wage setting in period t takes into account the wages of contracts still
in force, wt−1 , and the expected wage contract in the next period, wt+1 . As
the two groups are of equal size and power, we weight them equally by 0.5.
In addition wages depend on the state of the economy over the length of the
13
In our example they actually go short as I0 < 0.
14
The model could equally well be applied to analyze staggered price setting behavior.
2.5. DIFFERENCE EQUATIONS OF ORDER P 51

unexpected positive transitory supply shock

logged price level / inventory 0.5
inventories

logged price level

−0.5

−1
0 2 4 6 8 10 12 14 16 18 20
period
expected positive transitory supply shock in period 5
0.5
logged price level / inventory

inventories
0

−0.5
logged price level

−1
0 2 4 6 8 10 12 14 16 18 20
period

Figure 2.9: Impulse Response after a positive Supply Shock

contract, here represented by aggregate demand averaged over the current

and next period. The aggregate wage in period t, Wt , is then simply the
average over all existing individual contract wages in place in period t:
1
Wt = (wt + wt−1 ) (2.49)
2
The model is closed by adding a quantity theoretic aggregate demand equa-
tion relating Wt and yt :
yt = γWt + vt , γ < 0. (2.50)
The negative sign of γ reflects the fact that in the absence of full accom-
modation by the monetary authority, higher average nominal wages reduce
aggregate demand. vt represents a shock to aggregate demand.
Putting equations (2.48), (2.50), and (2.49) together one arrives at a
linear difference equation of order 2:
(1 + hγ)wt+1 − 2(1 − hγ)wt + (1 + hγ)wt−1 = −2h(vt + vt+1 )
or equivalently
wt+1 − φwt + wt−1 = Zt (2.51)
(1−hγ) 2h
with φ = 2 (1+hγ) and Zt = − (1+hγ) (vt + vt+1 ). The characteristic equation
for this difference equation is
1 − φz + z 2 = 0.
52 CHAPTER 2. LINEAR DIFFERENCE EQUATIONS

The symmetric nature of the polynomial coefficients implies that the roots
appear in pairs such that one root is the inverse of the other.15 This means
that one root, say λ1 , is smaller than one whereas the other one is greater
than one, i.e. λ2 = 1/λ1 . To see this note first that the discriminant is equal
to 4 = −hγ > 0. Thus, the roots are real and second that λ1 λ2 = 1. If we
denote λ1 by λ then λ2 = 1/λ and we have φ = λ + λ−1 .
Applying the superposition principle, the solution becomes
(p)
wt = c1 λt + c2 λ−t + wt (2.52)
(p)
where the coefficients c1 and c2 and a particular solution wt have yet to be
determined. In order to eliminate explosive solutions, we set c2 = 0. The
other constant can then be determined by noting that (wt ) is a predetermined
variable such that the wage negotiations in period one take wages from the
(p)
other group negotiated in period zero as given. Thus, c1 = w0 − w0 . To
find the particular solution, set
∞
X
(p)
wt = ψj Zt−j
j=−∞

and insert this solution into the difference equation (2.52) and perform a
comparison of coefficients as in the previous exercise. This leads again to
two homogeneous difference equations for the coefficients (ψj ) and (ψ−j ),
j ≥ 1 with solutions
ψj = d1 λj + d2 λ−j
ψ−j = e1 λj + e2 λ−j
where the coefficients d1 , d2 , e1 and e2 have still to be determined. The elim-
ination of explosive coefficient sequences leads to d2 = e2 = 0. Furthermore,
both solutions must give the same ψ0 so that d1 = e1 . Denote this value by
d, then comparing the coefficients for Zt and noting that φ = λ + λ−1 leads
to:
ψ1 = φψ0 − ψ−1 + 1 ⇐⇒ dλ = φd − dλ + 1.
Therefore
1
d= < 0.
λ − λ−1
The effect of a shock to aggregate demand in period j ≥ 0 is then
∂wt+j 2h
= ψ−j + ψ−j−1 = − d(1 + λ)λj , j = 0, 1, 2, . . .
∂vt 1 + hγ
15
This conclusion extends to contracts longer than two periods (see Ashenfelter and
Card, 1982).
Chapter 3

Systems of Linear Difference

Equations with Constant
Coefficients

3.1 Introduction
This chapter treats systems of linear difference equations. For each variable
X1t , · · · , Xnt , n ≥ 1, we are given a linear nonhomogeneous difference equa-
tion of order p where each variable can, in principle, depend on all other
variables with a lag. Writing each difference equation separately, the system
is given by
(1) (1) (1)
X1t = φ11 X1,t−1 + φ12 X2,t−1 + · · · + φ1n Xn,t−1
(p) (p) (p)
+ · · · + φ11 X1,t−p + φ12 X2,t−p + · · · + φ1n Xn,t−p + Z1t
(1) (1) (1)
X2t = φ21 X1,t−1 + φ22 X2,t−1 + · · · + φ2n Xn,t−1
(p) (p) (p)
+ · · · + φ21 X1,t−p + φ22 X2,t−p + · · · + φ2n Xn,t−p + Z2t
···
(1) (1)
Xnt = φn1 X1,t−1 + φn2 X2,t−1 + · · · + φ(1)
nn Xn,t−1
(p) (p)
+ · · · + φn1 X1,t−p + φn2 X2,t−p + · · · + φ(p)
nn Xn,t−p + Znt

Using matrix notation this equation system can be written more com-
pactly as

Xt = Φ1 Xt−1 + Φ2 Xt−2 + · · · + Φp Xt−p + Zt , Φp 6= 0, (3.1)

where Zt denotes an n-vector of exogenous variables and where Φi , i =

53
54 CHAPTER 3. SYSTEMS OF DIFFERENCE EQUATIONS
 
(k)
1, 2, . . . , p, denote the matrices Φk = φi,j for k = 1, 2, . . . , p. The
i,j=1,2,...,n
solution of this difference equation is based again on the same principles as
in the univariate case (see page 8). Before doing so we show how to reduce
this p-th order system to a first order system.
Any system of order p can be rewritten as a system of order 1. In order to
see this, define a new variable Yt as the stacked vectors Xt , Xt−1 , · · · , Xt−p+1 .
This new variable then satisfies the following first order system:
     
Xt   Xt−1 Zt
 Xt−1  Φ1 Φ2 Φ3 . . . Φp−1 Φp 
 Xt−2   0 
  
 Xt−2   In 0 0 . . . 0 0
  
  Xt−3   0 
Yt =  ..  =  0 In 0 . . . 0 0   ..  +  .. 
      
 .   .. .. .. . . .. ..   .   . 
   . . . . . .    
Xt−p+2  Xt−p+1   0 
0 0 0 . . . In 0
Xt−p+1 Xt−p 0
= ΦYt−1 + Zt (3.2)

0
where Zt is redefined to be Zt0 0 0 . . . 0 0 . In denotes the identity
matrix of dimension n. The matrix Φ is an np × np matrix called the com-
panion matrix of (3.1).1 Thus, multiplying out the equation system (3.2) one
can see that the first equation gives again the original equation (3.1) whereas
the remaining p − 1 equations are just identities. The study of a p-th order
system can therefore always be reduced to a first order system.

Properties of the Companion Matrix in the Univariate Case

The one-dimensional difference equation of order p, equation (2.1), can also
be written in this way as a first order system of dimension p. The companion
matrix is given in this case by
 
φ1 φ2 φ3 · · · φp−1 φp
 1 0 0 ··· 0 0
 
 0 1 0 ··· 0 0
C= . (3.3)

 .. .. .. . . .. .. 
. . . . . .
0 0 0 ··· 1 0

For this companion matrix, we can derive the following properties:

• The companion matrix is nonsingular if and only if φp 6= 0.

1
The literature distinguishes four forms of companion matrices depending on whether
the Φi ’s appear in the first, as in equation (3.2), or last row or first or last column.
3.2. FIRST ORDER SYSTEM OF DIFFERENCE EQUATIONS 55

• The characteristic polynomial of the companion matrix is: p(λ) =

λp − φ1 λp−1 − · · · − φp−1 λ − φp . Thus, the roots of the characteris-
tic polynomial of the companion matrix are just the inverses of the
roots of the characteristic polynomial of the difference equation (2.36).
• The geometric multiplicity of each eigenvalue equals 1 i.e. there is only
one independent eigenvector
0 for each λi . These eigenvectors are of the
p−1 p−2
form λi , λi , · · · , λi , 1 . Thus, in this situation there is no need to
rely on the Jordan canonical form (see Section 3.2.2).

3.2 First Order System of Difference Equa-

tions
The introduction above demonstrated that the first order system of differ-
ence equations encompasses single as well as systems of difference equations
of order p. We therefore reduce our analysis to the first order system of
difference equations:
Xt = ΦXt−1 + Zt , Φ 6= 0, (3.4)
where Xt denotes an n-vector and Φ an n × n matrix. The nonautonomous
part is represented by the n-vector Zt which corresponds to a vector of ex-
ogenous variables. In general, solutions may not exist for negative times. In
fact, when Φ has not full rank, then for points in the range of Φ there exists
X−1 such that X0 = ΦX−1 with X−1 being not unique.2 Thus, for simplicity,
we restrict ourself to the case where Φ is nonsingular, i.e. Φ ∈ GL(n), the
set of invertible real n × n matrices.

3.2.1 Homogenous First Order System of Difference

Equations
As in the one-dimensional case, we start the analysis with the discussion of
the homogeneous equation:
Xt = ΦXt−1 , Φ ∈ GL(n). (3.5)
We immediately see that starting with some initial vector X0 = x0 , all sub-
sequent values of Xt , t > 0, are uniquely determined. In particular,
Xt = Φt x0
2
A singular Φ matrix may be interpreted as a system which encompasses some redun-
dant variables. See also Section 5.3.
56 CHAPTER 3. SYSTEMS OF DIFFERENCE EQUATIONS

To highlight the dependency on the initial condition, we write X(t, x0 ) =

Φt x0 . Looking at the set of points reached from a starting point x0 suggests
the following definition.
Definition 3.1 (Orbit). The set O+ (x0 ) = {X(t, x0 )|t ∈ N0 } is called the
forward orbit of x0 . The set O(x0 ) = {X(t, x0 )|t ∈ Z} is called the whole
orbit, or orbit for short.
Remark 3.1. The invertibility of Φ implies that, if x0 6= y0 , O(x0 )∩O(y0 ) =
∅. Thus, two trajectories do not cross unless they have a common starting
value.
Suppose that we have two solutions to the homogenous system (3.5),
(1) (2)
Xt and Xt . Then, it is clear that any linear combination of these two
(1) (2)
solutions, c1 Xt + c2 Xt , is also a solution. Thus, the set of all solutions to
the homogenous system (3.5) forms a linear space. As in the univariate case,
we analyze the algebraic structure of this space.
As we want this chapter to be self-contained, we repeat the definition for
the linear independence of r solutions.
Definition 3.2. The sequences (X (1) ), (X (2) ), · · · , (X (r) ) with r ≥ 1 are said
to be linearly dependent for t ≥ 0 if there exist constants a1 , a2 , . . . , ar ∈ R,
not all zero, such that
(1) (2) (r)
a1 X t + a2 X t + · · · + ar X t =0 ∀t ≥ t0 .

This definition is equivalent to saying that there exists a nontrivial linear

combination of the solutions which is zero. If the solutions are not linearly
dependent, they are said to be linearly independent.
For given n sequences (X (1) ), (X (2) ), · · · , (X (n) ), we can define the Casaro-
tian matrix as follows:
 (1) (2) (n) 
X1t X1t . . . X1t
X (1) X (2) . . . X (n) 
C(t) =  .2t 2t 2t 

. .
. . . .. 
 . . . . 
(1) (2) (n)
Xnt Xnt . . . Xnt

The Casarotian matrix is closely related to the issue whether or not the
sequences are independent.
Lemma 3.1. If det C(t) of n sequences (X (i) ), 1 ≤ i ≤ n, is different from
zero for at least one t0 ≥ 0, then (X (i) ), 1 ≤ i ≤ n, are linearly independent
for t ≥ 0.
3.2. FIRST ORDER SYSTEM OF DIFFERENCE EQUATIONS 57

Proof. Suppose that (X (i) ), 1 ≤ i ≤ n, are linearly dependent. Thus, there

exists a nonzero vector c such that C(t)c = 0 for all t ≥ 0. In particular,
C(t0 )c = 0. This stands, however, in contradiction with the assumption
det C(t0 ) 6= 0.
Note that the converse
 is not true
 as can be seen from the following
(1) 1 (2) t
example: Xt = and Xt = 2 . These two sequences are linearly
t t
independent, but det C(t) = 0 for all t ≥ 0. The converse of Lemma 3.1 is
true if the sequences are solutions to the homogenous equation (3.5).

Lemma 3.2. If (X (i) ), 1 ≤ i ≤ n, are n linearly independent solutions of

the homogenous system (3.5), then det C(t) 6= 0 for all t ≥ 0.

Proof. Suppose there exists a t0 such det C(t0 ) = 0. This implies that there
(i)
exists a nonzero vector c such that C(t0 )c = ni=1 ci Xt = 0. Because the
P
(i) (i)
Xt are solutions so is the linear combination Yt = ni=1 ci Xt . For this
P
solution Yt0 = 0 thus Yt = 0 for all t because the uniqueness of the solution.
As the solutions are, however, linearly independent c must be equal to 0
which stands in contradiction to c 6= 0.
We can combine the two Lemmas to obtain the following theorem.

Theorem 3.1. The solutions (X (i) ), 1 ≤ i ≤ n, of the homogenous system

(3.5) are linearly independent for t ≥ 0 if and only if there exists t0 ≥ 0 such
that det C(t0 ) 6= 0.
(1) (n)
The above Theorem implies that the n solutions Ut , · · · , Ut of the
homogenous system (3.5) which satisfy the initial conditions
 0
(i) (i) 0 ··· 0 1 0 · · · 0
U0 = e = |{z} , 1 ≤ i ≤ n, (3.6)
i-th element

are linearly independent. Thus, we have at least n linearly independent so-

lutions. Suppose now that we are given any solution to the homogenous
system (3.5), say Xt . Then it is easy to see that we can express Xt as
Pn (i) (i)
Xt = i=1 Xi,0 Ut where Ut is the solution to the homogenous system
(3.5) satisfying the initial condition (3.6). As the solutions are uniquely
determined, we have thus shown that the space of all solutions to the ho-
mogenous system (3.5) is a linear space of dimension n. Thus, any solution
can be written as n
X (i)
Xt = Xi,0 Ut = U(t)X0 (3.7)
i=1
58 CHAPTER 3. SYSTEMS OF DIFFERENCE EQUATIONS

where U(0) = In , the identity matrix of dimension n × n, and U(t) =

(1) (n)
(Ut , . . . , Ut ).
Because each column of the matrix function U(t) is a solution to the ho-
mogenous system (3.5), U(t) satisfies the homogenous linear matrix system:

U(t + 1) = ΦU(t) (3.8)

This leads to the following definitions.

Definition 3.3. Any n×n matrix U(t) which is nonsingular for all t ≥ 0 and
which satisfies the homogenous matrix system (3.8) is called a fundamental
matrix. If in addition the matrix satisfies U(0) = In then it is called a
principal fundamental matrix.

Note that if V(t) is any fundamental matrix then U(t) = V(t)V −1 (0)
is a principal fundamental matrix. Note also if a V(t) is a fundamental
matrix then V(t)C is also a fundamental matrix where C is any nonsingular
matrix. This implies that there are infinitely many fundamental matrices for
a given homogenous matrix system. There is, however, only one principal
fundamental matrix because the matrix difference equation (3.8) uniquely
determines all subsequent matrices once an initial matrix is given. In the
case of a principal fundamental matrix this initial matrix is the identity
matrix. In this monograph we will not pursue the concept of the fundamental
matrix further because it does not payoff in the context of constant coefficient
systems.3 Only note that U(t) = Φt where U(t) is a principal fundamental
matrix. Thus, any solution to the homogenous system (3.5) has the form:

Xt = U(t)c = Φt c (3.9)

where c ∈ Rn is a constant vector.

3.2.2 Solution Formula for Homogeneous Systems

In order to find the explicit solution formulas of the homogenous system (3.5)
and to understand its properties, we therefore need to find an expression
for Φt . Such an expression can be found in terms of the eigenvalues and
eigenvectors of the matrix Φ.4 It is useful to distinguish in this context two
cases.
3
See Elaydi (2005) and Agarwal (2000) for a further elaboration and the relations to
Green’s matrix.
4
Recommended books on linear algebra are among others Meyer (2000) and Strang
(2003).
3.2. FIRST ORDER SYSTEM OF DIFFERENCE EQUATIONS 59

Distinct Eigenvalues
If all the eigenvalues, λ1 , · · · , λn of Φ are distinct, then Φ is diagonalizable,
i.e. similar to a diagonal matrix. Thus, there exists a nonsingular matrix
Q such that Q−1 ΦQ = Λ where Λ = diag(λ1 , · · · , λn ). The columns of Q
consist of the eigenvectors of Φ. With this similarity transformation in mind
it is easy to compute Φt :

Φt = QΛQ−1 QΛQ−1 · · · QΛQ−1 = QΛt Q−1

| {z }
t times
 t  
λ1 0 · · · 0 λt1 0 ··· 0
 0 λ2 · · · 0   0 λt ··· 0
 −1 2  −1
= Q  .. ..  Q = Q  .. ..  Q
 
.. . . .. ..
. . . . . . . .
0 0 · · · λn 0 0 · · · λtn

In the case of distinct eigenvalues, it is easy to proof the following theorem.

Theorem 3.2. If the spectrum of Φ, σ(Φ) = {λ1 , · · · , λn }, consists of n

distinct eigenvalues with corresponding eigenvectors qi , i = 1, . . . , n, then
the set
(i)
Xt = qi λti , i = 1, . . . , n,
represents a fundamental set of solutions to the homogenous system (3.5).

Proof. As qi is an eigenvector of Φ corresponding to λi , we have

(i) (i)
Xt+1 = qi λt+1
i = λi qi λti = Φqi λti = ΦXt , t ≥ 0.
(i)
The third equality follows from ΦQ = QΛ. Thus, the Xt , i = 1, · · · , n
are solutions to the homogenous system (3.5). In addition, we have that the
determinant of the corresponding

Casarotian matrix evaluated at t = 0 is
det C(0) = det q1 , · · · , qn = det Q 6= 0 because Q consists of n linearly
independent eigenvectors and is therefore nonsingular. Thus, according to
Theorem 3.1 these solutions are linearly independent.

Thus, the general solution to the homogenous system (3.5) can be written
as n
X
Xt = ci qi λti , t≥0 (3.10)
i=1

for some constants c1 , · · · , cn .

Another way to understand the result of Theorem 3.2 is to observe that
the similarity transformation actually decomposes the interrelated system in
60 CHAPTER 3. SYSTEMS OF DIFFERENCE EQUATIONS

Xt into n unrelated univariate first order difference equations. This decom-

position is achieved by the variable transformation Yt = Q−1 Xt :

Yt+1 = Q−1 Xt+1 = Q−1 ΦQQ−1 Xt = ΛYt

    
Y1,t+1 λ1 · · · 0 Y1,t
 ..   .. . . . . 
 .  = . . ..   .. 
 
Yn,t+1 0 · · · λn Yn,t

Thus, through this transformation we have obtained n unrelated univariate

first order difference equations in Yi,t , i = 1, . . . , n:

Y1,t+1 = λ1 Y1,t
···
Yn,t+1 = λn Yn,t

These equations can be solved one-by-one by the methods discussed in chap-

ter 2. The general solutions to these univariate first order homogenous equa-
tions are therefore Yi,t = ci λti , i = 1, · · · , n. Transforming the system in Yt
back to the original system by multiplying Yt from the left with Q yields
exactly the solution in equation (3.10).

Repeated Eigenvalues

The situation with repeated eigenvalues is more complicated. Before dealing

with the general case, note that even with repeated eigenvalues the matrix
Φ can be diagonalizable. This is, for example, the case for normal matrices,
i.e. matrices for which ΦΦ0 = Φ0 Φ. Examples of normal matrices include
symmetric matrices (Φ = Φ0 ), skew symmetric matrices (Φ = −Φ0 ) and
unitary or orthogonal matrices (ΦΦ0 = Φ0 Φ = I). More generally, Φ is
diagonalizable if and only if, for each eigenvalue, the algebraic multiplicity
equals the geometric multiplicity (i.e. if each eigenvalue is semisimple).
As Φ is not, in general, diagonalizable, we have to use its Jordan canonical
form.5 For every matrix Φ with distinct eigenvalues σ(Φ) = {λ1 , · · · , λs },
there exists a nonsingular matrix Q such that Φ can be reduced to a matrix J
by a similarity transformation, i.e. Q−1 ΦQ = J. The block diagonal matrix
5
A detailed treatment of the Jordan canonical form can be found, for example, in
Meyer (2000). The current exposition uses the complex Jordan form. There is, however,
an equivalent presentation based on the real Jordan form (see Colonius and Kliemann,
2014).
3.2. FIRST ORDER SYSTEM OF DIFFERENCE EQUATIONS 61

J is of the following form:

 
J (λ1 ) 0 ··· 0
 0 J (λ 2) ··· 0 
J = Q−1 ΦQ =  ..
 
.. .. .. 
 . . . . 
0 0 · · · J (λs )

The Jordan segments J (λi ), i = 1, · · · , s, consist of ti Jordan blocks, Jj (λi ),

j = 1, · · · , ti , where ti is the dimension of the nullspace of Φ − λi I, i.e.
ti = dim N (Φ − λi I). Each Jordan segment J (λi ) has a block diagonal
structure:
 
J1 (λi ) 0 ··· 0
 0 J2 (λi ) · · · 0 
J (λi ) =  .. ..  .
 
.. ..
 . . . . 
0 0 · · · Jti (λi )

The Jordan blocks themselves are of the following form:

 
λi 1 0 ··· 0 0
 0 λi
 1 ··· 0 0 
Jj (λi ) =  ... ... .. .. ..  = λ I + N (3.11)

 . . .
i
0 0 0 · · · λi 1
0 0 0 ··· 0 λi

where
 
0 1 0 ··· 0 0
0
 0 1 ··· 0 0

N =  ... .. .. ..
..  .

 . . ..
0 0 0 · · · 0 1
0 0 0 ··· 0 0

The square matrix N is a nilpotent matrix, i.e. N k = 0 if k is the dimension of

N . The dimension of the largest Jordan block in the Jordan segment J (λi )
is called the index of the eigenvalue λi , denoted by ki = index(λi ).
Given these preliminaries it is now a straightforward task to compute
Φ = QJ t Q−1 = Q diag (J t (λ1 ) , · · · , J t (λs )) Q−1 where the t-th
power of a
t

Jordan segment J t (λi ) is just J t (λi ) = diag J1t (λi ) , · · · , Jtti (λi ) . The t-th
62 CHAPTER 3. SYSTEMS OF DIFFERENCE EQUATIONS

power of a Jordan block Jj (λi ), j = 1, · · · , ti , is given by the expression:

Jj (λi )t = (λi I + N )t
     
t t−1 t t−2 2 t
t
= λi I + λ N+ λ N + ··· + λt−k+1 N k−1
1 i 2 i k−1 i
 t t
t−1 t
t−2 t

t−k+1 
λi 1 λi λ
2
i
· · · k−1
i
λ
t t−1 t
0
 λi t
1
λi · · · k−2 λit−k+2 
 .. .
.. . .. . .. .
..
= . (3.12)


t

t−1 

1
λi 
t
0 0 0 ··· λi
where N is the nilpotent matrix of size corresponding to the Jordan block.

3.2.3 Nonhomogeneous First Order System

Consider now the first order nonhomogeneous system:

Xt = ΦXt−1 + Zt (3.13)

As in the one-dimensional case the superposition principle also holds in the

multivariate case. Suppose that there exist two solutions to the nonho-
(1) (2)
mogeneous system (3.13), Xt and Xt . Then one can easily verify that
(1) (2)
Xt − Xt is a solution to the homogeneous system (3.5). Thus, the super-
(1) (2)
position principle implies that Xt − Xt = Φt c for some vector c.
Theorem 3.3. Every solution (Xt ) to the first order nonhomogeneous sys-
tem (3.13) can be represented as the sum of the general solution to the
(g)
homogeneous system (3.5), (Xt ), and a particular solution to the nonho-
(p)
mogeneous system (3.13), (Xt ):
(g) (p) (p)
Xt = Xt + Xt = Φt c + Xt . (3.14)

A particular solution can be found by iterating the difference equation

backwards:

Xt = ΦXt−1 + Zt
Xt = Φ(ΦXt−2 + Zt−1 ) + Zt = Φ2 Xt−2 + ΦZt−1 + Zt
...
Xt = Φt X0 + Φt−1 Z1 + Φt−2 Z2 + · · · + ΦZt−1 + Zt
t−1
X
t
= Φ X0 + Φj Zt−j
j=0
3.3. STABILITY THEORY 63

This suggests to take

t−1
X
(p)
Xt = Φj Zt−j
j=0

as a particular solution. So that the general solution to the nonhomogeneous

equation becomes
t−1
X
t
Xt = Φ c + Φj Zt−j (3.15)
j=0

For the initial value problem X0 = x0 , c = x0 . In the previous chapter we

have already seen that this backward looking solution is often not sensible
due to the forward looking nature of rational expectations (see the equity
price model and Cagan’s hyperinflation model in 2.4). As we can decompose
the system into n unrelated first order univariate difference equations, the
decision whether to use the forward or the backward solution depends in each
case on the location of the eigenvalues (roots of characteristic polynomial)
with respect to the unit circle. In general, we will have some eigenvalues
smaller than one in absolute value and some eigenvalues larger than one
in absolute value. Thus, we will have a mixture of backward and forward
solutions. For this strategy to work the number of stable roots must match
the number of initial conditions. We will analyze such situations in detail in
section 3.5.

3.3 Stability Theory

In this section we are interested in the asymptotic behavior of Xt as t → ∞.
In general a first order, possibly nonlinear, system of difference equations can
be written as
Xt+1 = f (Xt , t) (3.16)
where f (x, t) is a function from Rn × N ∪ {0} → Rn which is continuous in
x. A fixed point, a steady state, or an equilibrium point of the system (3.16)
is defined as follows.
Definition 3.4. A point X ∗ ∈ Rn in the domain of f is called a fixed point,
steady state or an equilibrium point if

X ∗ = f (X ∗ , t) for all t ≥ 0

In the special case of nonhomogeneous linear system (3.13) with a con-

stant forcing variable Zt = Z, the steady state is defined through the equation
X ∗ = ΦX ∗ + Z. This equation has a unique solution if and only if In − Φ is
64 CHAPTER 3. SYSTEMS OF DIFFERENCE EQUATIONS

nonsingular. This is equivalent to the statement that one is not an eigenvalue

of Φ. In this case the steady state is given by X ∗ = (In − Φ)−1 Z.6 In many
situations it is useful to express the nonhomogeneous difference equation as a
homogenous system by rewriting the variable Xt as a deviation from steady
state:
Xt+1 − X ∗ = Φ (Xt − X ∗ ) . (3.17)
We repeat the definitions of stability given in Section 2.2.7

Definition 3.5. An equilibrium point X ∗ is called:

• stable if for all ε > 0 there exists a δε > 0 such that

kX0 − X ∗ k < δε implies kXt − X ∗ k < ε for all t > 0. (3.18)

If X ∗ is not stable it is called unstable.

• The point X ∗ is called attracting if there exists η > 0 such that

kX0 − X ∗ k < η implies lim Xt = X ∗ . (3.19)

t→∞

If η = ∞, X ∗ is called globally attracting.

• The point X ∗ is asymptotically stable or is an asymptotically stable

equilibrium point 8 if it is stable and attracting. If η = ∞, X ∗ is called
globally asymptotically stable.

• exponentially stable if there exist δ > 0, M > 0, and η ∈ (0, 1) such

that for the solution X(t, x0 ) we have

kX(t, x0 ) − X ∗ k ≤ M η t kx0 − X ∗ k, whenever kx0 − X ∗ k < δ.

• A solution X(t, x0 ) is bounded if there exists a positive constant M < ∞

such
kX(t, x0 )k ≤ M for all t.
Thereby the constant M may depend on x0 .
6
If In − Φ is singular multiple steady steady states may exist or no steady states may
exist.
7
As we are dealing almost only with linear systems refined definitions of stability are
not necessary (see Elaydi (2005) and Agarwal (2000)).
8
In economics this is sometimes called stable.
3.3. STABILITY THEORY 65

In this definition kXk denotes some norm in Rn , typically the Euclidian

norm. The induced matrix norm is denoted by kΦk (see appendix B). It
does not matter which norm is actually used because we are dealing with
norms in finite dimensional spaces only (vectors in Rn , respectively matrices
in Rn×n ). For these spaces all norms are equivalent.

Remark 3.2. Clearly, exponential stability implies asymptotic stability and,

therefore, stability and attractiveness. The reverse is, however, not true for
each of these implications.

The stability of equilibrium points is intimately related to the eigenval-

ues of Φ, especially to the spectral radius of Φ, ρ(Φ). Before proofing the
corresponding theorem, consider the following lemma.

Lemma 3.3. The zero solution of the homogeneous system (3.5) is stable if
and only if there exists M > 0 such that
t
Φ ≤ M for all t ≥ 0. (3.20)

Proof. Suppose that the inequality (3.20) is satisfied then kXt k ≤ M kX0 k.
Thus, for ε > 0, let δ = Mε . Then kX0 k < δ implies kXt k < ε so that the
zero point is stable. Conversely, suppose that the zero point is stable. Then
for all kX0 k < δ

Φ = sup Φt ξ = 1 sup Φt X0 ≤ ε = M
t
kξk≤1 δ kX0 k≤δ δ

where the first inequality is just the definition of the matrix norm correspond-
ing to the norm in Rn . The second equality is a consequence of kX0 k ≤ δ.
The inequality above follows from the assumption that the zero point is a
stable equilibrium.

The condition given in equation (3.20) is equivalent to the condition that all
solutions are bounded.

Theorem 3.4. For the homogeneous system (3.5) the following statements
are true:

(i) The zero solution is stable if and only if ρ(Φ) ≤ 1 and the eigenvalues
on the unit circle are semisimple.

(ii) The zero solution is asymptotically stable if and only if ρ(Φ) < 1. In
this case, the solution is even exponentially stable.
66 CHAPTER 3. SYSTEMS OF DIFFERENCE EQUATIONS

Proof. According to previous lemma 3.3 we have to prove that kΦt k ≤ M for
some M > 0. Using the Jordan canonical form Φ = QJQ−1 this amounts to
kΦt k = kQJ t Q−1 k ≤ M . But this is equivalent to the existence of a M̃ > 0
M̃
such that kJ t k ≤ M̃ . M may then be taken as M = kQkkQ −1 k . Now the power

of J is given by powers of the Jordan blocks. The t-th power of a Jordan

block corresponding to the eigenvalue λi is given according to equation (3.12)
by

Jj (λi )t = (λi I + N )t
     
t t−1 t t−2 2 t
t
= λi I + λ N+ λ N + ··· + λt−k+1 N k−1
1 i 2 i k−1 i
 t t
t−1 t
t−2 t

t−k+1 
λi 1 λi λ
2
i
· · · k−1
λit−k+2 
t t t−1 t
0
 λi 1
λi · · · k−2 λi 
 .. .
.. . .. .. ..
= . .

. .
t


t−1 

1
λ i
0 0 0 ··· λti

The elements in this matrix become unbounded if |λi | > 1. They also become
unbounded if |λi | = 1 and Jj (λi ) is not a 1 × 1 matrix. If, however, for all
eigenvalues with |λi | = 1 the largest Jordan blocks are 1 × 1, then the Jordan
segment corresponding to λi with |λi | = 1, J (λi ), is just a diagonal matrix
with ones in the diagonal and is therefore obviously bounded. The Jordan
segments, J (λi ), corresponding to eigenvalues |λi | < 1, converge to zero, i.e.
limt→∞ J (λi )t = 0 because tk λti → 0 as t → ∞ by l’Hôpital’s rule.
Theorem 3.4 showed that the stability properties crucially depend on
the location of the eigenvalues relative to the unit circle. If there are no
eigenvalues on the unit circle, small perturbations of Φ will not affect the
location of the eigenvalues relative to the unit circle because they depend
continuously on the entries of Φ. Thus, small perturbations preserve the
stability properties. This motivates to bring out matrices with no eigenvalues
on the unit circle in an own definition.
Definition 3.6 (Hyperbolic Matrix). A hyperbolic matrix is a matrix with
no eigenvalues on the unit circle. If A is a hyperbolic matrix, then the
corresponding linear homogeneous difference equation Xt+1 = AXt is also
called hyperbolic.
Consider a hyperbolic matrix Φ with spectrum σ = {λ1 , . . . , λs }, 1 ≤
s ≤ n.9 Then, partition the spectrum into the stable, respectively unstable
9
The spectrum of a matrix is given by the set of its distinct eigenvalues.
3.3. STABILITY THEORY 67

eigenvalues, i.e. σs = {λ ∈ σ : |λ| < 1} and σu = {λ ∈ σ : |λ| > 1}. Denote

by W s the space generated by the eigenvectors (generalized eigenvectors)
corresponding to the eigenvalues in σs and, similarly, by W u the space gen-
erated by the eigenvectors (generalized eigenvectors) corresponding to the
eigenvalues in σu . Given these definitions, we may follow Elaydi (2005) and
state the following theorem.
Theorem 3.5 (Stable Manifold Theorem). Given a linear autonomous hy-
perbolic difference equation (3.5) the following properties hold:
(i) If X(t, x0 ) is a solution of (3.5) with x0 ∈ W s , then X(t, x0 ) ∈ W s for
all t. Moreover,
lim X(t, x0 ) = 0.
t→∞

(ii) If X(t, x0 ) is a solution of (3.5) with x0 ∈ W u , then X(t, x0 ) ∈ W u for

all t. Moreover,
lim X(t, x0 ) = 0.
t→−∞

Proof. The proof follows Elaydi (2005, theorem 4.14). Let X(t, x0 ) be a
solution with x0 ∈ W s . The definition of the (generalized) eigenvector corre-
sponding to some eigenvalue λ ∈ σ implies that ΦEλ = Eλ where Eλ denotes
the space generated by the (generalized) eigenvectors corresponding to λ.
Hence ΦW s = W s and s s
PrX(t, x0 ) ∈ W for all t ≥ 0. Given x0 ∈ W , we can
represent x0 as x0 = j=1 αj ej where r is the number of eigenvalues, distinct
or not, strictly smaller than one and ej denotes the (generalized) eigenvec-
tors corresponding to those eigenvalues. Let J = Q−1 ΦQ be the Jordan form
of Φ. Next, partition the Jordan matrix into blocks of stable, respectively
unstable eigenvalues accordingly:
 
Js 0
J=
0 Ju

where Js and Ju are r × r, respectively (n − r) × (n − r) matrices. Therefore,

r
X
X(t, x0 ) = Φt x0 = QJ t Q−1 x0 = QJ t Q−1 αj ej
j=1
 t  r r  t 
Js 0 X −1
X Js 0
=Q αj Q ej = Q αj Q−1 ej
0 Jut 0 Jut
j=1 j=1
r  
X Jst 0
=Q αj Q−1 ej .
0 0
j=1
68 CHAPTER 3. SYSTEMS OF DIFFERENCE EQUATIONS

The last equality follows from the fact that the Q−1 ej ’s are of the form
(ξ1j , . . . , ξrj , 0, . . . , 0)0 which is a consequence of the ej ’s being (generalized)
eigenvectors. This implies that X(t, x0 ) → 0 as t → ∞ because Jst → 0 as
t → ∞.
The proof of (ii) is analogous.
Remark 3.3. Theorem 3.5 may be refined by relaxing the assumption of
an hyperbolic matrix and can be sharpened with respect to its conclusions
(Colonius and Kliemann, 2014, theorem 1.5.8).

Linearization
As in the case of one dimensional difference equations (see Theorem 2.4), we
can analyze the stability properties of nonlinear systems via a linear approx-
imation around the steady state. Consider for this purpose the nonlinear
homogeneous system Xt+1 = F (Xt ) with F : Rn −→ Rn and fixed point
X ∗ . Suppose that F is continuously differentiable in an open neighborhood
of X ∗ , then the linearized difference equation is given by
Xt+1 − X ∗ = A(Xt − X ∗ )
where A is the derivative of F evaluated at X ∗ , i.e. A = DF (X ∗ ), the matrix
of partial derivatives. If A is a hyperbolic matrix, then we say that X ∗ is a
hyperbolic fixed point.
The Stable Manifold Theorem 3.5 can be extended, at least locally, to
nonlinear difference equations. This is the subject of the Hartman–Grobman
theorem (see Elaydi (2005) and Robinson (1999, section 5.6)).
Theorem 3.6 (Hartman–Grobman Theorem). If F is C r –diffeomorphism10
with hyperbolic fixed point X ∗ , then there exist neighborhoods V of X ∗ and
W of 0 and a homeomorphism H : W −→ V such that F (H(X)) = H(AX).
Proof. Robinson (See 1999, sections 5.6 and 5.7)
Remark 3.4. The Theorem says that F is topologically equivalent in a
neighborhood of the fixed point X ∗ to the linear map induced by the deriva-
tive evaluated at the fixed point. This fact can be represented graphically as
follows
A
W −−−→ W
 
 
Hy yH
F
V −−−→ V
10
This means that the function is r-times continuously differentiable and that it is a
homeomorphism (i.e. one-to-one and onto, continuous with continuous inverse).
3.4. TWO-DIMENSIONAL SYSTEMS 69

Thus, if a fixed point X ∗ is hyperbolic, we can analyze its stability properties

by investigating the induced linear homogeneous difference equation and, in
particular, the eigenvalues of the matrix of partial derivatives.

An interesting special case, often encountered in economics, is obtained

when the system expands in some directions, but contracts in others. In such
a case the fixed point is called saddle point.

Definition 3.7 (Saddle Point). The zero solution of the hyperbolic linear
difference equation Xt+1 = AXt is called a saddle point if there exist at least
two eigenvalues of A, λu and λs , such that |λu | > 1 and |λs | < 1.

3.4 Two-dimensional Systems

Many theoretical economic models are reduced and investigated as two-
dimensional systems. Thus, we devote this section to the analysis of such
systems. In case the system is of dimension n = 2 a necessary and sufficient
condition for asymptotic stability is given in the following theorem.

Theorem 3.7. The homogeneous two-dimensional system has an asymptot-

ically stable solution if and only if

|tr(Φ)| < 1 + det Φ < 2. (3.21)

Proof. The characteristic polynomial, P(λ), of a 2 × 2 matrix is P(λ) =

λ2 − tr(Φ)λ
√ + det Φ. The roots of this quadratic polynomial are thus λ1,2 =
tr(Φ)± tr2 (Φ)−4 det Φ
2
.
Suppose that the zero point is asymptotically stable then
|λ1,2 | < 1. But, in the case of real roots, this is equivalent to the following
two inequalities:
q
−2 − tr(Φ) < tr2 (Φ) − 4 det Φ < 2 − tr(Φ)
q
−2 − tr(Φ) < − tr2 (Φ) − 4 det Φ < 2 − tr(Φ).

Squaring the second inequality in the first line and simplifying gives: tr(Φ) <
1 + det Φ. Squaring the first inequality in the second line gives −1 − det Φ <
tr(Φ). Combining both results gives the first part of the stability condition
(3.21). The second part follows from the observation that det Φ = λ1 λ2 and
the assumption that |λ1,2 | < 1. If the roots are complex, they are conjugate
complex, so that the second part of the stability (3.21) results from det Φ =
70 CHAPTER 3. SYSTEMS OF DIFFERENCE EQUATIONS

λ1 λ2 = |λ1 | |λ2 | < 1. The first part follows from tr2 (Φ) − 4 det Φ < 0 which
is equivalent to 0 < tr2 (Φ) < 4 det Φ. This can be used to show that
4 (1 + det Φ − tr(Φ)) > 4 + tr2 (Φ) − 4tr(Φ) = (2 − tr(Φ))2 > 0.
which is the required inequality.
Conversely, if the stability condition (3.21) is satisfied and if the roots are
real, we have
p p
−2 + tr2 (Φ) − 4 det Φ tr(Φ) + tr2 (Φ) − 4 det Φ
−1 < < λ1 =
2 p 2
2
tr(Φ) + tr (Φ) + 4 − 4tr(Φ)
<
q 2
tr(Φ) + (2 − tr(Φ))2
= < 1.
2
Similarly, for λ2 . If the roots are complex, they are conjugate complex and we
2 2
have |λ1 |2 = |λ2 |2 = λ1 λ2 = tr (Φ)−tr 4(Φ)+4 det Φ = det Φ < 1. This completes
the proof.

Phase Diagrams of Two-Dimensional Systems

An additional advantage of two-dimensional systems are that their qualitative
properties can be easily analyzed by a phase diagram. Consider for this
purpose the homogenous first order system (3.5) written as a two equation
system:11     
X1,t+1 φ11 φ12 X1,t
Xt+1 = = ΦXt = ; (3.22)
X2,t+1 φ21 φ22 X2,t
or equivalently as
X1,t+1 = φ11 X1,t + φ12 X2,t (3.23)
X2,t+1 = φ21 X1,t + φ22 X2,t . (3.24)
It is clear that (0, 0)0 is an equilibrium for this system. In order to understand
the dynamics of the system, we can draw two lines in a (X1 , X2 )-diagram.
The first line is given by all points such that the first variable does not change,
i.e. where X1,t+1 = X1,t . From equation (3.23), these points are represented
by a line with equation (φ11 − 1)X1,t + φ12 X2,t = 0. Similarly, the points
where the second variable does not change is, from equation (3.24), the line
with equation φ21 X1,t + (φ22 − 1)X2,t = 0. These two lines divide the R × R-
plane into four regions I, II, III, and IV as in figure 3.1. In this example both
lines have positive slopes.
11
We may also interpret the systems as written in deviations from steady state.
3.4. TWO-DIMENSIONAL SYSTEMS 71

X2 t
schedule: X1 t+1
, - X1 t = 0

region I region II

schedule:
steady X2 t+1
, - X2 t = 0
state

X1 t

region III
region IV

Figure 3.1: Example of a Phase Diagram

72 CHAPTER 3. SYSTEMS OF DIFFERENCE EQUATIONS

The dynamics of the system in each of the four regions can be figured
out from the signs of the coefficients as follows. Suppose we start at a point
on the X1,t+1 − X1,t = (φ11 − 1)X1,t + φ12 X2,t = 0 schedule then we know
that the first variable does not change. Now increase X2,t a little bit. This
moves us into region I or IV. If φ12 is positive, this implies that X1,t+1 −
X1,t > 0 so that the first variable increases. Thus, we know that above the
X1,t+1 − X1,t = 0 line, X1,t increases and that below this line X1,t decreases.
We can indicate this result in figure 3.1 by arrows from left to right in regions
I and IV and arrows from right to left in regions II and III. If φ12 is negative,
we obtain, of course, the opposite result. Similarly, consider the schedule
X2,t+1 − X2,t = φ21 X1,t + (φ22 − 1)X2,t = 0 where the second variable does
not change. Consider now an increase in X1,t . This moves us into region III
or IV. If the coefficient φ21 is positive, this implies that X2,t+1 − X2,t > 0 so
that the second variable must increase. As before we can infer that below the
X2,t+1 − X2,t = 0 line X2,t increases whereas above this line X2,t decreases.
We can again indicate this behavior by arrows: upward arrows in regions
III and IV and downward arrows in regions I and II. If the sign of φ21 is
negative, the opposite result is obtained. This type of analysis gives us a
phase diagram as in figure 3.1. This diagram shows us the dynamics of the
system starting in every possible point. In this example, the arrows indicate
that wherever we are, we will move closer to the steady state. But this
corresponds exactly to the definition of an asymptotically stable equilibrium
point (see definition 3.18). Thus, we conclude from this diagram that the
steady state is an asymptotically stable equilibrium point.

Of course the situation depicted in figure 3.1 is not the only possible
one. In order to analyze all possible situations which can arise in a two-
dimensional system, we make a variable transformation using the Jordan
canonical form of Φ. If Φ has the Jordan canonical form Φ = QJQ−1 then
we make the variable transformation Yt = Q−1 Xt . This results in a new first
order homogenous difference equation system:

Yt+1 = Q−1 Xt+1 = Q−1 ΦXt = Q−1 ΦQQ−1 Xt = JYt (3.25)

3.4. TWO-DIMENSIONAL SYSTEMS 73

where J has one of the following three forms:

 
λ1 0
J= distinct or repeated semisimple real eigenvalues
0 λ2
 
λ 1
J= repeated eigenvalue with one independent eigenvector
0 λ
 
α β
J= complex eigenvalues: λ1,2 = α ± ıβ
−β α

Note that the steady state is not affected by this variable transformation. It
is still the point (0, 0)0 . Let us treat these three cases separately.

case 1: distinct or repeated semisimple real eigenvalues The variable

transformation has effectively decoupled the two-dimensional system
into two separate homogenous first order difference equations with so-
lutions: Y1,t = λt1 y1,0 and Y2,t = λt2 y2,0 where y1,0 and y2,0 are given
initial values. From the previous discussion we know that the steady
state is asymptotically stable if and only if both eigenvalues are smaller
than one in absolute value. Such a situation is plotted in figure 3.2. The
arrows indicate that for every starting point the system will converge
towards the equilibrium point. As an example we have plotted four
trajectories starting at the points (1, 1), (1, −1), (−1, 1), and(−1, −1),
respectively.
Figure 3.3 displays a situation where the equilibrium point is unstable.
Indeed both eigenvalues are larger than one and the trajectories quickly
diverge in the directions indicated by the arrows. In the figure, we have
plotted again four trajectories with the same starting values as in the
previous example.
Figure 3.4 shows an interesting configuration which is often encountered
in economic models, especially in those which involve rational expecta-
tions. We have one eigenvalue smaller than one in absolute value and
one eigenvalue larger than one in absolute value, i.e. |λ1 | > 1 > |λ2 |.
This implies that the system is expanding in the direction of the eigen-
vector corresponding to λ1 , but is contracting in the direction of the
eigenvector corresponding to λ2 . This configuration of the eigenvalues
leads to a saddle point equilibrium (see Definition 3.7). Although the
steady state is unstable, as almost all trajectories diverge, there are
some initial values for which the system converges to the steady state.
In figure 3.4 all trajectories starting on the y-axis converge to the steady
74 CHAPTER 3. SYSTEMS OF DIFFERENCE EQUATIONS

0.8

0.6

0.4

0.2
Y2,t

−0.2
steady
state
−0.4

−0.6

−0.8

−1
−1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1
Y1,t

Figure 3.2: Asymptotically Stable Steady State (λ1 = 0.8, λ2 = 0.5)

1000

500
Y2,t

steady
state
−500

−1000

−6 −4 −2 0 2 4 6
Y1,t

Figure 3.3: Unstable Steady State (λ1 = 1.2, λ2 = 2)

3.4. TWO-DIMENSIONAL SYSTEMS 75

0.8

0.6

0.4

0.2
2,t

0
Y

−0.2
steady
−0.4 state

−0.6

−0.8

−1
−6 −4 −2 0 2 4 6
Y
1,t

Figure 3.4: Saddle Steady State (λ1 = 1.2, λ2 = 0.8)

state. Thus, given an initial value y20 for Y2,t , the requirement that the
solution must be bounded uniquely determines an initial condition for
Y1,t too, which in this reduced setting is just Y1,0 = y10 = 0. Thus,
the solution is given by Y1,t = 0 and Y2,t = λt2 y2,0 for some initial value
y2,0 . Note that the saddle path, in contrast to the other paths, is a
straight line through the origin. This property is carried over when
the system is transformed back to its original variables. In fact, the
solution becomes X1,t = q12 λt2 y2,0 and X2,t = q22 λt2 y2,0 where (q12 , q22 )0
is the eigenvector corresponding to λ2 . Thus, the ratio of X1,t and X2,t
equals q12 /q22 constant.12 As saddle point equilibria are very promi-
nent in economics, we investigate this case in depth in Section 3.5. In
particular, we will go beyond the two-dimensional systems and analyze
the role of initial values.
When there are multiple eigenvalues with two independent eigenvec-
tors, Φ can again be reduced by a similarity transformation to a di-
agonal matrix. The trajectories are then straight lines leading to the
origin if the eigenvalue is smaller than one as in figure 3.5, and straight
lines leading away from the origin if the eigenvalue is larger than one
in absolute terms.
12
If q22 = 0, we take the ratio X2,t /X1,t which again defines a straight line through the
origin.
76 CHAPTER 3. SYSTEMS OF DIFFERENCE EQUATIONS

0.8

0.6

0.4

0.2
2,t

0
Y

−0.2
steady
−0.4 state

−0.6

−0.8

−1
−1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1
Y1,t

Figure 3.5: Repeated Roots with Asymptotically Stable Steady State (λ1 =
λ2 = 0.8)

When one eigenvalue equals one whereas the second eigenvalue is smaller
than one in absolute value, we arrive at a degenerate situation. Whereas
Y1,t remains at its starting value y10 , Y2,t converges to zero so that the
system converges to (y10 , 0)0 as is exemplified by figure 3.6. Only when
y10 = 0 will there be a convergence to the steady state (0, 0). However,
(0, 0) is not the only steady state because the definition of an eigen-
value implies that A − I is singular. Thus, there exists X ∗ 6= 0 such
that X ∗ = ΦX ∗ .
In case that the first eigenvalue equals minus one, there is no conver-
gence as Y1,t will oscillate between y10 and −y10 .

case 2: repeated eigenvalues with one independent eigenvector In this

case Φ can no longer be reduced to a diagonal matrix by a similarity
transformation. As J is no longer a diagonal matrix, the locus of all
points where Y1,t does not change is no longer the x-axis, but is given
by the line with equation (λ − 1)Y1,t + Y2,t = 0. Figure 3.7 displays
this case with an eigenvalue of 0.8 which implies an asymptotically sta-
ble steady state. Note that, given our four starting points, the system
moves first away from the equilibrium point until it hits the schedule
where Y1,t does not change, then it changes direction and runs into the
steady state.
3.4. TWO-DIMENSIONAL SYSTEMS 77

1.5

0.5
Y2,t

steady
state
−0.5

−1

−1.5
−1.5 −1 −0.5 0 0.5 1 1.5
Y1,t

Figure 3.6: Degenerate Steady State (λ1 = 1, λ2 = 0.8)

0.5

steady
state
Y2,t

schedule:
Y1,t+1−Y1,t = 0
−0.5

−1

−2.5 −2 −1.5 −1 −0.5 0 0.5 1 1.5 2 2.5

Y1,t

Figure 3.7: Repeated eigenvalues one independent eigenvector (λ = 0.8)

78 CHAPTER 3. SYSTEMS OF DIFFERENCE EQUATIONS

schedule:
Y1,t+1−Y1,t = 0
1

0.8

0.6

0.4

0.2

0
schedule:
Y2,t+1−Y2,t = 0
−0.2

−0.4

−0.6

−0.8

−1

−1 −0.8−0.6−0.4−0.2 0 0.2 0.4 0.6 0.8 1

Figure 3.8: Complex eigenvalues with Stable Steady State (λ1,2 = 0.7 ± 0.2ı)

case 3: complex eigenvalues If the two conjugate complex eigenvalues

are λ1,2 = α ± ıβ then Φ is similar to the matrix
   
α β cos ω sin ω
= |λ| .
−β α − sin ω cos ω
p
where |λ| = α2 + β 2 and ω = tan−1 αβ . Figure 3.8 shows the dy-

namics in the case of eigenvalues inside the unit circle. One can clearly
discern the oscillatory behavior and the convergence to the steady state.
Figure 3.9 displays a situation with an unstable steady state where all
trajectories move away from the steady state. Finally figure 3.10 dis-
plays a degenerate case where the eigenvalues are on the unit circle. In
such a situation the system moves around its steady state in a circle.
The starting values are (0.25, 0.25), (0.5, 0.5), (1, 1), and (1.5, 1.5).

3.5 Boundary Value Problems under Ratio-

nal Expectations
In this section we discuss the general boundary value problem under rational
expectations. In these models there are, typically, not enough initial values to
3.5. BOUNDARY VALUE PROBLEM 79

10

−5

schedule: schedule:
Y2,t+1−Y2,t = 0 Y1,t+1−Y1,t = 0
−10

−10 −5 0 5 10

Figure 3.9: Complex eigenvalues with Unstable Steady State (λ1,2 = 1±0.5ı)

2.5

1.5

0.5

−0.5

−1

−1.5

−2

−2.5
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2 2.5

Figure 3.10: Complex eigenvalues on the unit circle (λ1,2 = cos (π/4) ±
ı sin (π/4))
80 CHAPTER 3. SYSTEMS OF DIFFERENCE EQUATIONS

pin down a unique solution. Thus, one has to resort to additional restrictions.
These restrictions come from the assumption that the solution must remain
bounded. In the well-behaved scenario, this will give just enough additional
initial values to determine a unique solution and the model is said to be
determinate (see also Section 1.1). Geometrically, this solution has the form
of a saddle path and the steady state is a saddle point.
A concise treatment of such models in the context of stochastic difference
equations was first given by Blanchard and Kahn (1980). Klein (2000) and
Sims (2001) provide further insights and solution approaches (see Chapter 5
for further details). As the stochastic setting delivers similar conclusions
with respect to uniqueness, we adopt the Blanchard-Kahn setup to the de-
terministic case by replacing rational expectations by perfect foresight. This
framework delivers first order linear nonautonomous difference equations:

Xt = ΦXt−1 + Zt , t ∈ N ∪ {0} (3.26)

where, as before, Φ is a n × n invertible real matrix, i.e. Φ ∈ GL(n). In

addition, we are given k initial values with 0 < k < n: 13

x0 = RX0 (3.27)

where R is a (k × n)-matrix of rank k. The simplest case is the one where

initial values are given for the first k variables:

X10 = x10 , . . . , Xk0 = xk0 ,

or in matrix form
 
X10
     ... 
x10 1 ... 0 0 ... 0  
X
 
x0 =  ...  =  ... . . . ... ... . . . ...   k0  = (Ik , 0k×n−k )X0
    
Xk+1,0 
xk0 0 ... 1 0 ... 0   ... 


Xn0

The application of the superposition principle gives the general solution

of the difference equation (3.26) as the sum of the general solution of the
(g)
homogeneous equation Xt and a particular solution of the nonhomogeneous
13
The case k = 0 can be treated in a similar manner (See the example in section 4.3).
If k = n there are just enough initial conditions and the solution procedure of Section 3.2
can be directly applied.
3.5. BOUNDARY VALUE PROBLEM 81

(p)
equation Xt :
(g) (p)
Xt = X t + Xt
(p)
= QΛt Q−1 c + Xt
(p)
where the n-vector c is yet to be determined. Taking t = 0, X0 = c + X0 .
Given a particular solution, the initial values given by equation (3.27) thus
determine c only up to n − k degrees of freedom so that we are lacking n − k
additional boundary conditions.
In order to solve the boundary value problem, we partition the Jordan
form of Φ according to the moduli of the eigenvalues. Let Φ = Q−1 JQ
where J is the Jordan form of Φ and where the columns of Q consist of the
corresponding eigenvectors (generalized eigenvectors). We assume that the Φ
is hyperbolic, i.e. has no eigenvalues on the unit circle. Define the number of
possibly multiple eigenvalues strictly smaller than one by n1 and the number
of eigenvalues strictly larger than one by n2 so that n = n1 + n2 . Assume
that the eigenvalues are ordered in terms of their moduli, then define the
matrices Λ1 and Λ2 as follows:
   
J1 0 0 0
Λ1 = and Λ2 =
0 0 0 J2
where J1 consists of the Jordan segments corresponding to the eigenvalues
smaller than one and J2 to the segments corresponding to the eigenvalues
greater than one. We focus on the case where the zero solution is a saddle
point, meaning that there exists at least two eigenvalues λ1 and λ2 such that
|λ1 | < 1 and |λ2 | > 1 (see Definition 3.7).14 With this notation, we propose
the following particular solution:
∞
X ∞
X
(p)
Xt = QΛj1 Q−1 Zt−j − QΛ−j −1
2 Q Zt+j (3.28)
j=0 j=1

The reader is invited to verify that this is indeed a solution to equation (3.26).
The solution proposed in equation (3.28) has the property that “variables”
corresponding to eigenvalues smaller than one are iterated backwards whereas
those corresponding to eigenvalues larger than one are iterated forwards. This
(p)
ensures that {Xt } remains bounded whenever {Zt } is. The general solution
therefore is of the form
∞
X X∞
t −1 t −1
Xt = QΛ1 Q c + QΛ2 Q c + j −1
QΛ1 Q Zt−j − QΛ−j −1
2 Q Zt+j . (3.29)
j=0 j=1

14
If n1 or n2 equal zeros then the terms corresponding to the matrices Λ1 and Λ2 are
omitted.
82 CHAPTER 3. SYSTEMS OF DIFFERENCE EQUATIONS

The final step consists in finding the constant c. There are two types of
restrictions: the first one are the initial values given by equation (3.27); the
second one are the requirement that we are only interested in non-exploding
solutions. Whereas the first type delivers k restrictions because rank(R) =
k, the second type delivers n2 restrictions. Thus, c must be determined
according to the following equation system:
(p)
initial values: R c = x0 − RX0
no explosive solutions: Q(2) c = Q(21) c1 + Q(22) c2 = 0 (3.30)
 (11)   
−1 Q Q(12) (21) ..
where Q = (2)
, Q = Q .Q (22)
and c = (c01 , c02 )0 and
Q(21) Q(22)
where the partitioning of Q−1 and c conforms to the partitioning of the
eigenvalues. Note that Q(2) is a n2 × n matrix. Depending on whether the
number of independent restrictions is greater, smaller or equal to n, several
situations arise.
Theorem 3.8 (Blanchard-Kahn). Let Φ ∈ GL(n) be hyperbolic then the
nonhomogeneous difference equation (3.26) with initial values given by (3.27)
has a unique nonexplosive solution if and only if
 
R
rank = n. (3.31)
Q(2)

The solution is given by equation (3.29) with c uniquely determined from (3.30).
A necessary condition is that k = rank(R) = n1 : the number of eigenval-
ues smaller than one, n1 , must be equal to the number of independent initial
conditions, k = rank(R).
In many instances the values of the first k variables at given time, say
time zero, are given. This is the case when there are exactly k predetermined
variables in the system. The initial values then fix the first k values of c:
(p) (p) (p)
c1 = x10 − X10 where x10 and X10 denote the first k values of x0 and X0 .
Corollary 3.1. If R = (Ik , 0k×(n−k) ) with k = rank(R) = n1 , the necessary
and sufficient condition (3.31) reduces to Q(22) is invertible. Given that c1 is
fixed by the initial conditions, c2 = −(Q(22) )−1 Q(21) c1 .
If k = rank(R) > n1 then there are too many restrictions and there is no
nonexplosive solution. We may, however, soften the boundedness condition
and accept explosive solutions in this situation.
If k = rank(R) < n1 there are not enough initial conditions so that it
is not possible to pin down c uniquely. We thus have an infinite amount
3.5. BOUNDARY VALUE PROBLEM 83

of solutions and we call such a situation indeterminate. The multiplicity of

equilibria indeterminacy offers the possibility of sunspot equilibria.15 Sunspot
equilibria have been introduced by Cass and Shell (1983), Azariadis (1981),
and Azariadis and Guesnerie (1986) (see also Azariadis (1993) and Farmer
(1993)).

15
Sunspot equilibria explore the idea that extraneous beliefs about the state of nature
influence economic activity. The disturbing feature of sunspot equilibria is that economic
activity may change across states although nothing fundamental has changed.
84 CHAPTER 3. SYSTEMS OF DIFFERENCE EQUATIONS
Chapter 4

Examples of Linear
Deterministic Systems of
Difference Equations

4.1 Exchange Rate Overshooting

4.1.1 Introduction
A classic example for a system with one predetermined and one so-called
”jump”-variable is the exchange rate overshooting model by Dornbusch (Dorn-
busch (1976)).1 The model describes the behavior of the price level and the
exchange rate in a small open economy. It consists of an IS equation, a price
adjustment equation, and a LM equation. In addition, the uncovered interest
rate parity (UIP) is assumed together with rational expectations:

ytd = δ(et + p∗ − pt ) − σ(rt − pt+1 + pt ), (IS)

pt+1 − pt = α(ytd − y), (price adjustment)
m − pt = φy − λrt , (LM)
rt = r∗ + et+1 − et , (UIP)

where the parameters δ, σ, α, φ, and λ are all positive. The variables are
all expressed in logarithms. The IS-equation represents the dependence of
aggregate demand ytd on the relative price of foreign to home goods and on
the real interest rate. A devaluation of the exchange rate2 , an increase in the
1
Rogoff (Rogoff (2002)) provides an appraisal of this influential paper.
2
The exchange rate et is quoted as the price of a unit of foreign currency in terms of
the domestic currency. An increase in et therefore corresponds to a devaluation of the
home currency.

85
86 CHAPTER 4. EXAMPLES: LINEAR SYSTEMS

foreign price level, p∗ , or a decrease in the domestic price level pt all leads to an
increase in aggregate demand. On the other hand, an increase in the domestic
nominal interest rate, rt , or a decrease in the expected inflation rate, pt+1 −pt ,
lead to a reduction in aggregate demand. A crucial feature of the Dornbusch
model is that prices are sticky and adjust only slowly. In particular, the price
adjustment pt+1 − pt is proportional to the deviation of aggregate demand
from potential output y. If aggregate demand is higher than potential output,
prices increase whereas, if aggregate demand is below potential output, prices
decrease. In particular, the price level is treated as predetermined variable
whose value is fixed in the current period. The LM-equation represents
the equilibrium on the money market. The demand for real balances, m −
pt , depends positively on potential output3 and negatively on the domestic
nominal interest rate. The model is closed by assuming that the uncovered
interest parity holds where r∗ denotes the foreign nominal interest rate. In
contrast to the price level, the exchange rate is not predetermined. It can
immediately adjust within the current period to any shock that may occur.
For simplicity, the exogenous variables y, m, r∗ , and p∗ are assumed to remain
constant.
This system can be reduced to a two-dimensional system in the exchange
rate and the price level:
1
et+1 − et = (φy + pt − m) − r∗ (4.1)
λ
α h σ i
pt+1 − pt = δ (et + p∗ − pt ) − y − (φy − m + pt ) (4.2)
1 − ασ λ
The first equation was obtained by combining the LM-equation with the
UIP. The second equation was obtained by inserting the IS-equation into
the price adjustment equation, replacing the nominal interest rate using the
LM-equation and then solving for pt+1 − pt .
The steady state of this system is obtained by setting et = ess and pt = pss
for all t and solving for this two variables:

pss = λr∗ + m − φy (4.3)

1
ess = pss − p∗ + (y + σr∗ ) (4.4)
δ
In the steady state, UIP and the price adjustment imply rt = r∗ and ytd = y.
The system can be further reduced by writing it in terms of deviations from
3
This represents a simplification because money demand should depend on aggregate
demand, ytd , and not on potential output, y. However, we adopt this simplified version for
expositional purposes.
4.1. EXCHANGE RATE OVERSHOOTING 87

steady state:
 1

 ss
 1  ss

et+1 − e λ
σ et − e
=  αδ α(δ + λ ) 
pt+1 − pss 1− pt − pss
1 − ασ 1 − ασ
 ss

e −e
= Φ t (4.5)
pt − pss

4.1.2 Analysis of the Dynamic Properties

The dynamic behavior of the system (4.5) depends on the eigenvalues of Φ.
Denote the characteristic polynomial of Φ by P(µ) and the two corresponding
eigenvalues by µ1 and µ2 , then

P(µ) = (µ − µ1 )(µ − µ2 ) = µ2 − tr(Φ)µ + det Φ

Without additional assumptions on the parameters, it is impossible to obtain

further insights into the qualitative behavior of the system. We suppose that
the price adjustment is sufficiently slow. Specifically, we assume that
4
0 < ασ < 1 and α< ,
(2 + 1/λ)(δ + 2σ)
we obtain:
α  σ
trΦ = µ1 + µ2 = 2 − δ+ <2
1 − ασ
 λ 
α σ δ
det Φ = µ1 µ2 = 1 − δ+ + <1
1 − ασ λ λ
 2
2 α  σ 4αδ
4 = (trΦ) − 4 det Φ = δ+ + >0
1 − ασ λ λ(1 − ασ)
αδ
P(1) = (1 − µ1 )(1 − µ2 ) = − <0
λ(1 − ασ)
  
α σ δ
P(−1) = (1 + µ1 )(1 + µ2 ) = 4 − 2 δ+ + >0
1 − ασ λ λ
where 4 denotes the discriminant of the quadratic equation. The above
inequalities have the following implications for the two eigenvalues:
• 4 > 0 implies that the eigenvalues are real;

• P(1) < 0 implies that they lie on opposite sides of 1;

• trΦ < 2 implies that the sum of the eigenvalues is less than 2;
88 CHAPTER 4. EXAMPLES: LINEAR SYSTEMS

• P(−1) > 0 finally implies that one eigenvalue, say µ1 , is larger than 1
whereas the second eigenvalue, µ2 , lies between −1 and 1.
Because the eigenvalues are distinct, we can diagonalize Φ as Φ = QΛQ−1
where Λ is a diagonal matrix with µ1 and µ2 on its diagonal. The column of
the matrix Q consist of the eigenvectors of Φ. Multiplying the system (4.5)
by Q−1 , we obtain the transformed system:
   ss
  ss

êt+1 −1 et+1 − e −1 −1 et − e
= Q = Q ΦQQ
p̂t+1 pt+1 − pss pt − pss
  
µ1 0 êt
= (4.6)
0 µ2 p̂t

Through this change of variables we have obtained a decoupled system: the

original two-dimensional system is decomposed into two one-dimensional ho-
mogenous difference equations. These two equations have the general solu-
tion:

êt = c1 µt1
p̂t = c2 µt2

where c1 and c2 are two constants yet to be determined. Transforming the

solution of the decoupled systems back into the original variables yields:

et = ess + c1 q11 µt1 + c2 q12 µt2 (4.7)

pt = pss + c1 q21 µt1 + c2 q22 µt2 (4.8)

where Q = (qij )i,j=1,2 . Because µ1 > 1 this represents an unstable system.

As time evolves the unstable eigenvalue will eventually dominate. In order to
avoid this explosive behavior, we set c1 equal to zero. The second constant
c2 can be determined from the boundary condition associated with the pre-
determined variable, in our case the price level. Suppose the system starts
in period zero and we are given a value p0 ssfor the price level in this period,
then according to equation (4.8) c2 = (p0q−p22
)
, provided q22 6= 0. Combining
all these elements with the equations (4.7) and (4.8) leads to the equation
for the saddle path:
q12
et = ess + (pt − pss ) (4.9)
q22
Next, we show that the saddle path is downward sloping, i.e. that qq22 12
<
0. This can be established by investigating the defining equations for the
eigenvector corresponding to the second eigenvalue µ2 . They are given by
(φ11 −µ2 )q12 +φ12 q22 = 0 and φ21 q12 +(φ22 −µ2 )q22 = 0. Because (φ11 −µ2 ) > 0
4.1. EXCHANGE RATE OVERSHOOTING 89

and φ12 > 0, q12 and q22 must be of opposite sign. The same conclusion is
reached using the second equation. This reasoning also shows that q22 6= 0.
Suppose that q22 = 0 then q12 must also be zero because (φ11 − µ2 ) > 0.
This, however, contradicts the assumption that (q12 , q22 )0 is an eigenvector.
Note that although the eigenvector is not uniquely determined, its direction
and thus the slope of the saddle path is.
The Dornbusch model is most easily analyzed in terms of a phase diagram
representing the price level and the exchange rate as in figure 4.1. The graph
consists of two schedules: pt+1 − pt = 0 and et+1 − et = 0. Their intersection
determines the steady state denoted by S. These two schedules correspond
to the equations (4.4) and (4.3). The et+1 − et = 0 schedule does not depend
on the exchange rate and is therefore horizontal intersecting the price axis
at pss . Above this schedule the exchange rate depreciates whereas below
this schedule the exchange appreciates, according to equation (4.5). This
is indicated by arrows pointing to the right, respectively to the left. The
pt+1 − pt = 0 schedule is upward sloping. To its left, prices are decreasing
whereas to its right prices are increasing, according to equation (4.5). The
two schedules divide the e-p-quadrant into four regions: I, II, III, and IV. In
each region the movement of e and p is indicated by arrows. In the Dornbusch
model the price level is sticky and considered to be a predetermined variable.
Suppose that in period 0 its level is given by p0 . The exchange rate in this
period is not given, but endogenous and has to be determined by the model.
Suppose that the exchange rate in period 0 is at a level corresponding to
point A. This point is to the left of the pt+1 − pt = 0 schedule and above
the et+1 − et = 0 schedule and therefore in region I. This implies that the
price level has to fall and the exchange rate to increase. The path of e and
p will continue in this direction until they hit the et+1 − et = 0 schedule.
At this time the system enters region IV and the direction is changed: both
the price level and the exchange rate decrease. They will so forever. We are
therefore on an unstable path. Consider now an exchange rate in period 0
corresponding to point B. Like A, this point is also in region I so that the
exchange rate increases and the price level decreases. However, in contrast to
the previous case, the path starting in B will hit the pt+1 −pt = 0 schedule and
move into region II. In this region, both the price level and the exchange rate
increase forever. Again this cannot be a stable path. Thus, there must be an
exchange rate smaller than the one corresponding to point B, but higher than
the one corresponding to point A, which sets the system on a path leading
to the steady steady. This is exactly the exchange rate which corresponds
to the saddle path given by equation (4.9). In this way the exchange rate in
period 0 is pinned down uniquely by the requirement that the path of (et , pt )0
converges.
90 CHAPTER 4. EXAMPLES: LINEAR SYSTEMS

schedule:
p pt+1 - pt = 0
I
saddle path
A B
p0
II

S
ss schedule:
p et+1 - et = 0
IV

III
saddle path

e
ss e

Figure 4.1: Dornbusch’s Overshooting Model

4.1.3 Effects of an Increase in Money Supply

The phase diagram is also very convenient in analyzing the effects of changes
in the exogenous variables. Consider, for example, an unanticipated perma-
nent increase in money supply. This moves, according to equation (4.3), the
et+1 − et = 0 schedule up and, according to (4.4), the pt+1 − pt = 0 to the left
as shown in figure 4.2. The steady state therefore jumps from Sold to Snew
and the new saddle path goes through Snew . Suppose that the price level
was initial at pss
old . As the price level cannot react to the new situation it will
remain initially at the old steady state level. The exchange rate, however,
can adapt immediately and jumps to e0 such that the system is on the new
saddle path. As this value lies typically above the new steady level, we say
that the exchange rate overshoots. The reason for this ”excess” depreciation
of the exchange rate is the stickiness of the price level. In the short-run,
the exchange rate carries all the burden of the adjustment. As time evolves
the system moves along its saddle path to its new steady state. During this
transition the price level increases and the exchange rate appreciates. Thus,
the immediate reaction of the economy is a depreciation of the exchange rate
coupled with an expected appreciation.
4.1. EXCHANGE RATE OVERSHOOTING 91

p schedule:
pt+1 - pt = 0
saddle path

Snew
ss
pnew
schedule:
et+1 - et = 0
ss
pold
saddle path

ss ss
eold enew e0 e

Figure 4.2: Unanticipated Increase in Money Supply Dornbusch’s Overshoot-

ing Model
92 CHAPTER 4. EXAMPLES: LINEAR SYSTEMS

4.2 Optimal Growth Model

4.2.1 Introduction
In contrast to the previously discussed Solow model (see section 2.4.2), the
optimal growth model seeks to determine the saving-consumption decision
optimally.4 Consider for this purpose a planer who seeks the optimize the
discounted utility stream from per capita consumption (ct ). If Ct and Lt
denote aggregate consumption and aggregate labor input in period t then per
capita consumption is given by ct = Ct /Lt . As before labor input increases
at the exogenously given rate µ > 0, i.e. Lt+1 = (1 + µ)Lt . The planner is
assumed to maximize the following Bentham type objective function:
∞
X
V (c0 , c1 , · · · ) = β t Lt U (ct )
t=0
∞
X
= L0 (β(1 + µ))t U (ct ), 0 < β(1 + µ) < 1. (4.10)
t=0

The constant β is called the subjective discount rate. The period utility func-
tion U : R+ → R is continuously differentiable, increasing, strictly concave,
and, in order to avoid corner solutions, fulfills limc→0 U 0 (c) = ∞.5
The rest of the specification is exactly the same as for the Solow model
(see section 2.4.2): Output is produced according to a neoclassical aggregate
production satisfying the Inada conditions. Recognizing that investment in
period t, It equals It = Kt+1 − (1 − δ)Kt the national accounting identity
becomes:
Ct + It = Ct + Kt+1 − (1 − δ)Kt = F (Kt , Lt )

or in per capita terms

ct + (1 + µ)kt+1 − (1 − δ)kt = f (kt )

respectively,

ct + (1 + µ)kt+1 = f (kt ) + (1 − δ)kt = h(kt ). (4.11)

4
The optimal growth model originates in the work of Ramsey (1928). It has been
widely analyzed and stands at the heart of the Real Business Cycle approach. An extensive
treatment of this model together with additional references can be found in Stokey and
Lucas Jr. (1989).
5
Instead of a Bentham P∞type objective function one could also work with the conventional
one: V (c0 , c1 , · · · ) = t=0 β t U (ct ) with 0 < β < 1. This modification will, however, not
change the qualitative implications of the model.
4.2. OPTIMAL GROWTH MODEL 93

The first order condition for the optimum is given by the Euler-equation,
sometimes also called the Keynes-Ramsey rule:
U 0 (ct ) = βh0 (kt+1 )U 0 (ct+1 ) (4.12)
Thus, the Euler-equation equates the marginal rate of transformation, 1/h0 (kt+1 ),
to the marginal rate of substitution, βU 0 (ct+1 )/U 0 (ct ).
The equation system consisting of the transition equation (4.11) and the
Euler-equation (4.12) constitutes a nonlinear difference equation system. The
analysis of this system proceeds in the usual manner. First, we compute
the steady state(s). Then we linearize the system around the steady state.
This gives a linear homogeneous difference equation in terms of deviations
from steady state. We find the solution of this difference equation using
the superposition principle. Finally, we select, if possible, one solution using
initial conditions and boundedness arguments.

4.2.2 Steady State

The steady state (k ∗ , c∗ ) is found by setting c∗ = ct = ct+1 and k ∗ = kt = kt+1
in equations (4.11) and (4.12). This results in the nonlinear equation system:
∆k = 0 : c∗ = f (k ∗ ) − (δ + µ)k ∗ (4.13)
∆c = 0 : βh0 (k ∗ ) = β(f 0 (k ∗ ) + 1 − δ) = 1. (4.14)
The ∆c = 0 equation is independent of c and of the shape of the utility
function U . This equation therefore determines k ∗ . The first equation viewed
as a function c of k has an inverted U-shape in the (k, c)-plane as can be
deducted from the following reasoning:
• k = 0 implies c = 0. The derivative dc/dk = f 0 (k) − (δ + µ) evaluated
at k = 0 is strictly positive because of the Inada conditions.
• The function reaches a maximum at k ∗∗ determined by dc/dk = f 0 (k ∗∗ )−
(δ + µ) = 0. Because β(1 + µ) < 1 by assumption, k ∗ < k ∗∗ . k ∗∗ is
called the modified golden rule capital stock. It is larger than the op-
timal capital stock because of discounting.
• For k > k ∗∗ the ∆k = 0 schedule declines monotonically and crosses
the x-axis at k max . This is the maximal value of capital sustainable in
the long-run. It is achieved when the consumption is reduced to zero.
Thus, k max is determined from the equation f (k max ) = (δ + µ)k max .
The shape of both schedules is plotted in figure 4.3 as the blue lines. They
cross at point E which corresponds to the unique nonzero steady state of the
system.
94 CHAPTER 4. EXAMPLES: LINEAR SYSTEMS

saddle path
c Dc=0

c
* E
Dk=0

c0

0 k0 k
*
k
**
k
max
k
Figure 4.3: Phase diagram of the optimal growth model
4.2. OPTIMAL GROWTH MODEL 95

4.2.3 Discussion of the Linearized System

The equations (4.11) and (4.12) constitute a two-dimensional system of non-
linear difference equations of order one or, after inserting h(kt ) − (1 + µ)kt+1
for ct and h(kt+1 ) − (1 + µ)kt+2 for ct+1 , a single nonlinear difference equation
of order two. We therefore need two boundary conditions to pin down a so-
lution uniquely. One condition is given by the initial value of the per capita
capital stock k0 > 0.
In order to study the dynamics of the system in detail, rewrite the Euler
equation (4.12) as

U 0 (ct+1 ) − U 0 (ct ) = [1 − βh0 (kt+1 )]U 0 (ct+1 ). (4.15)

From this equation we can deduce that ct+1 ≥ ct when kt+1 < k ∗ and vice
versa. Thus, to left of the (∆c = 0)-schedule consumption rises whereas to
the right consumption falls. Similarly, the transition equation (4.11) implies
that kt+1 ≥ kt when ct is lower than the corresponding c∗ implied by the
(∆k = 0)-schedule. Thus, the two schedules divide the nonnegative orthant
of the (k, c)-plane in four regions. The dynamics in these four regions is
indicated in figure 4.3 by orthogonal arrows. In the region to the left of
the (∆c = 0)-schedule and above the (∆k = 0)-schedule, i.e. the north-
west region, consumption would increase whereas the capital intensity would
decrease. This dynamics would continue until the c-axis is hit. When this
happens, the economy has no capital left and therefore produces nothing but
consumes a positive amount. Such a situation is clearly infeasible. Paths with
this property have therefore to be excluded. Starting at k0 , there is, however,
one path, the saddle path, where the forces which lead to explosive paths,
respectively infeasible paths, just offset each other and lead the economy to
the steady state. This is the red line in figure 4.3.
The algebraic analysis requires the linearization of the nonlinear equation
system (4.11) and (4.12). This leads to:
kt+1 − k ∗
  −1
kt − k ∗
   
1+µ 0 β −1
=
βU 0 (c∗ )h00 (k ∗ ) U 00 (c∗ ) ct+1 − c∗ 0 U 00 (c∗ ) ct − c∗
where we used the fact that βh0 (k ∗ ) = 1. Given that µ > 0 and U 00 (c∗ ) < 0,
the matrix on the left hand side is invertible. This then leads to the following
linear first order homogenous system:
kt+1 − k ∗ β −1 kt − k ∗
    
1 −1
= −1 −1
ct+1 − c∗ 1 + µ RA (c∗ )h00 (k ∗ ) −βRA (c∗ )h00 (k ∗ ) + (1 + µ) ct − c∗
kt − k ∗
 
=Φ (4.16)
ct − c∗
96 CHAPTER 4. EXAMPLES: LINEAR SYSTEMS

where RA (c∗ ) equals −U 00 (c∗ )/U 0 (c∗ ) > 0, the absolute risk aversion coeffi-
cient evaluated at the steady state.6 As discussed in section 3.3 and 3.4, the
dynamics of the system is determined by the eigenvalues of Φ. Denote the
characteristic polynomial of Φ by P(λ) and the corresponding eigenvalues by
λ1 and λ2 , then we have

P(λ) = (λ − λ1 )(λ − λ2 ) = λ2 − tr(Φ)λ + det Φ

with
−1 ∗ 00 ∗
trΦ = λ1 + λ2 = 1 + [β(1 + µ)]−1 − βRA (c )h (k )/(1 + µ) > 2
1
det Φ = λ1 λ2 = >1
β(1 + µ)
4 = (trΦ)2 − 4 det Φ = [1 − (β(1 + µ))−1 ]2
−1 ∗ 00 ∗  −1 ∗ 00 ∗ 
βRA (c )h (k ) βRA (c )h (k ) 2
+ −2− >0
1+µ 1+µ β(1 + µ)
−1 ∗ 00 ∗
βRA (c )h (k )
P(1) = 1 − trΦ + det Φ = <0
1+µ
where 4 denotes the discriminant of the quadratic equation and where we
used the fact that h00 < 0. The above inequalities have the following impli-
cations for the two eigenvalues:

• 4 > 0 implies that the eigenvalues are real and distinct;

• trΦ > 2 implies that at least one eigenvalue is greater than 1;

• P(1) < 0 then implies that they lie on opposite sides of 1;

• P(0) = det Φ > 1 finally implies that one eigenvalue, say λ1 , is larger
than 1 whereas the second eigenvalue, λ2 , lies between 0 and 1.

Because the eigenvalues are distinct, we can diagonalize Φ as Φ = QΛQ−1

where Λ is a diagonal matrix with λ1 and λ2 on its diagonal. We take
λ1 > 1 > λ2 > 0. The column of the matrix Q = (qij ) consist of the
eigenvectors of Φ. Thus, the solution can be written as

kt − k ∗ = c1 q11 λt1 + c2 q12 λt2 (4.17)

ct − c∗ = c1 q21 λt1 + c2 q22 λt2 . (4.18)
6
This concept is intimately related to the intertemporal elasticity of substitution in
1−α
this context. In particular, for the U (c) = c 1−α−1 the coefficient of relative risk aversion
RR = c RA = α is just the inverse of the intertemporal elasticity of substitution.
4.2. OPTIMAL GROWTH MODEL 97

Because λ1 > 1 this represents an unstable system. As time evolves the

unstable eigenvalue will eventually dominate. In order to avoid this explosive
behavior, we set c1 equal to zero. The second constant c2 can be determined
from the boundary condition associated with the predetermined variable.
Suppose the system starts in period zero and we are given a value k0 for the ∗)
capital intensity in this period, then according to equation (4.17) c2 = (k0q−k
12
,
provided q12 6= 0. Combining all these elements with the equations (4.17)
and (4.18) leads to the equation for the saddle path:
q22
ct = c∗ + (kt − k ∗ ) (4.19)
q12

Next, we show that the saddle path is upward sloping, i.e. that qq12 22
> 0 as
shown by the red line in Figure 4.3. This can be verified by manipulating the
defining equations for the eigenvector corresponding to the second eigenvalue
λ2 . They are given by (φ11 −λ2 )q12 +φ12 q22 = 0 and φ21 q12 +(φ22 −λ2 )q22 = 0.
Because (φ11 −λ2 ) > 0 and φ12 < 0, q12 and q22 are of the same sign. The same
conclusion is reached using the second equation. This argument also shows
that q12 6= 0. Suppose that q12 = 0 then q22 must also be zero because φ12 =
−1
1+µ
< 0. This, however, contradicts the assumption that (q12 , q22 )0 is an
eigenvector. Note that although the eigenvector is not uniquely determined,
its direction and thus the slope of saddle path is.

4.2.4 Some Policy Experiments

In the following we discuss two policies. The first one analyzes the introduc-
tion of a tax on the return to capital. The second one investigates the effects
of an increase in government expenditures.

Taxation of Capital
Suppose that the government levies a proportional tax on the gross return
to capital. For simplicity, we assume that the revenues from the tax are just
wasted. Therefore only the Euler equation (4.12) is affected. The new Euler
equation then becomes

U 0 (ct ) = β(1 − τ )h0 (kt+1 )U 0 (ct+1 ) (4.20)

where τ is the tax rate with 0 < τ < 1. The transition equation for capital
is not altered. The new ∆c = 0 schedule then is

∆c = 0 : β(1 − τ )h0 (k ∗ ) = β(1 − τ )(f 0 (k ∗ ) + 1 − δ) = 1.

98 CHAPTER 4. EXAMPLES: LINEAR SYSTEMS

new Dc=0
c schedule saddle path corresponding
to new steady state
c0

*
c old Eold Dk=0

*
c
Enew

0
*
k new k0 k
max
k
Figure 4.4: Phase diagram of the optimal growth model with distortionary
taxation of capital

This implies that an increase in the tax rate τ will lower the steady state
capital stock. The tax is thus distortionary.

The evolution of consumption and capital can best be understood by

examining the corresponding phase diagram in Figure 4.4. Starting from
the initial steady state Eold , the introduction of the tax implies that the
∗
new steady state Enew involves a lower steady state capital stock knew . This
lower capital stock is achieved by raising consumption. On impact, consump-
tion jumps from c∗old to c0 such that the point (k0 , c0 ) is on the saddle path
corresponding to the new steady state. The capital stock remains initially
unaffected because it is predetermined. Over time both consumption and
capital start to decrease along the new saddle path to reach the new steady
state.
4.2. OPTIMAL GROWTH MODEL 99

Increase in Government Expenditures

Next we consider a permanent increase in government expenditures. These

expenditures are deducted from output before the investment/consumption
decision by the households takes place. Such a policy will affect only the
resource constraint, i.e. the national accounting, but leaves the Euler equation
unchanged. Thus, the steady state capital stock remains unchanged by this
policy. Given that this policy is not announced in advance, consumption
adjusts immediately by falling to its new steady state value. There is no
dynamics. Only the distribution of consumption between private and public
use is affected.
Denoting government expenditures by g, the national accounting iden-
tity (4.11) becomes

ct + g + (1 + µ)kt+1 = f (kt ) + (1 − δ)kt = h(kt ). (4.21)

The ∆k = 0 schedule changes accordingly:

∆k = 0 : c∗ = f (k ∗ ) − (δ + µ)k ∗ − g.

Thus, starting initially with no government expenditures, the schedule is

shifted down by g as shown in figure 4.5. Consumption reacts immediately
and drops from its old steady state value c∗0 to its new one c∗1 .
In contrast to the previously discussed policies which were all unantici-
pated and permanent, we analyze a transitory increase in government expen-
ditures which is unanticipated by the time of announcement. More precisely,
in period t0 the government introduces unexpectedly expenditures by an
amount g > 0, but announces credibly at the same time that it will discon-
tinue this policy in period t1 and return to g = 0. The dynamics of this
policy is shown in figure 4.5. On impact at time t0 , consumption will drop
from c∗0 to c0 > c∗1 . Thus, the reduction in consumption is smaller than in
the permanent case. This puts the system on an unstable path shown in
green in figure 4.5. On this path, capital is continuously lowered whereas
consumption increases. At time t1 when the policy is reverted as expected,
this unstable path hits the saddle path corresponding to the old steady state.
Capital then reaches its lowest value k1 . From then on, the economy moves
along the old saddle path back to the old equilibrium E0 : (c∗0 , k ∗ ). Capital
then starts to grow again. In the long-run the economy moves back to the
old steady state.7
7
See Judd (1985) for further details of this type of analysis in a continuous framework.
100 CHAPTER 4. EXAMPLES: LINEAR SYSTEMS

c Dc=0 saddle path

unstable path
* E0 old: Dk=0
c0
c1 g
c0
* new: Dk=0
c1
E1

0 k1 k
*
k
Figure 4.5: Phase diagram of the optimal growth model with government
expenditures
4.3. THE NEW KEYNESIAN MODEL 101

4.3 The New Keynesian Model

In this section we study a simple version of the New Keynesian macroeco-
nomic model as it has become popular recently. A detailed description of the
model and its microfoundations can be found in Woodford (2003) and Galı́
(2008). Here we follow the exposition by Galı́ (2011). The model consists of
the following three equations:
1
yt = yt+1 − (it − πt+1 ), (IS-equation)
σ
πt = βπt+1 + κyt + ut , (forward-looking Phillips-curve)
it = φπt , (Taylor-rule)

where yt , πt , and it denote income, inflation and the nominal interest rate, all
measured as deviations from the steady state. ut is an exogenous cost-push
shock. Furthermore, we assume that σ > 0, κ > 0, and 0 < β < 1. In
addition, we take an aggressive central bank, i.e. φ > 1.
This system can be solved for (yt+1 , πt+1 )0 by inserting the Taylor-rule
and Phillips-curve into the IS-equation:
      
πt+1 1 1 −κ πt −ut /β
Xt+1 = = +
yt+1 β (φβ − 1)/σ β + κ/σ yt ut /(σβ)
= ΦXt + Zt+1 (4.22)

Denote the characteristic polynomial of Φ by P(λ) and the corresponding

eigenvalues by λ1 and λ2 , then we have

P(λ) = (λ − λ1 )(λ − λ2 ) = λ2 − tr(Φ)λ + det Φ

with
1 κ
trΦ = λ1 + λ2 = 1 + + >2
β σβ
1 κφ
det Φ = λ1 λ2 = + >1
β σβ
 2  
2 1 κ κ 2
4 = (trΦ) − 4 det Φ = 1 − + + 2 + − 4φ
β σβ σβ β
κ
P(1) = (1 − λ1 )(1 − λ2 ) = (φ − 1) > 0 if φ > 1
σβ

where 4 denotes the discriminant of the quadratic equation. Depending on

φ, the roots of P(λ) may be complex. We therefore distinguish two cases.
102 CHAPTER 4. EXAMPLES: LINEAR SYSTEMS

First assume that φ is high such that 4 < 0. In this case we have two complex
conjugate roots. Because det Φ > 1, they are located outside the unit circle.8
Alternatively assume that φ is small enough such that 4 > 0. In this case
both eigenvalues are real. Using the assumption φ > 1, P(1) > 0. Thus,
both roots are either greater or smaller than one. They cannot be smaller
than one because trΦ > 2. Thus, in both cases we reach the conclusion
that the eigenvalues are outside the unit circle. As both variables are non-
predetermined, the boundedness condition (3.30), Qc = 0 which is equivalent
to c = 0, then determines the unique solution:
∞  −j   
X λ1 0 −1 ut−1+j /β
Xt = Q Q
j=1
0 λ−j2 −ut−1+j /(σβ)

where the columns of Q consist of the eigenvectors corresponding to λ1 and

λ2 .
Suppose now that the central bank fixes the path of the interest rate. The
interest rate then becomes an exogenous variable and the system changes to:
      
πt+1 1 1 −κ πt −ut /β
Xt+1 = = + ∗
yt+1 β −1/σ β + κ/σ yt it /σ + ut /(σβ)
= ΦX
e t + Zt+1 (4.23)
where i∗t is the exogenous path of the interest rate. The trace, determinant,
and discriminant, 4 e of the characteristic polynomial of Φ
e become:

e = λ1 + λ2 = 1 + 1 + κ > 2
trΦ
β σβ
1
det Φe = λ1 λ2 = > 1
β
 2  
2 1 κ κ 2
4 = (trΦ) − 4 det Φ = 1 −
e + +2+ >0
β σβ σβ β
κ
P(1) = (1 − λ1 )(1 − λ2 ) = − < 0.
σβ
The discriminant is now unambiguously positive so that both eigenvalues are
real. Moreover, P(1) < 0 so that one eigenvalue is smaller than one and the
other bigger than one. Thus, the boundedness condition does not determine
a unique solution. Instead there is a continuum of solutions indexed by c1
and we are faced with the case of indeterminacy. The implications of this
indeterminacy for monetary policy and possible remedies are discussed in
Galı́ (2011).
8
Another way to reach this conclusion is by observing that the real part of the roots is
trΦ
2 > 1.
Chapter 5

Linear Stochastic Expectational

Difference Equations

5.1 Introduction and Assumptions

In the following we analyze linear stochastic expectational difference equa-
tions in {Xt } of the form:

Φ1 Et Xt+1 = Φ0 Xt + Zt , t = 0, 1, 2 . . . , (5.1)

where {Xt } and {Zt } are real-valued n-dimensional stochastic processes de-
fined on some probability space (Ω, F, P). The random variables Xt and
Zt are measurable with respect to the σ-algebra Ft = σ{(Xs , Zs ) : s ≤ t}.
This makes the sequence {Ft } a filtration adapted to {Xt } and {Zt }. In eco-
nomics Ft is also called the information set. Et then denotes the conditional
expectation with respect to Ft . As expectations are based on current and
past Xt ’s and Zt ’s only and not on extraneous variables, we have eliminated
the possibility of sunspot solutions.
Expectational difference equations of the type (5.1) arise typically in the
context of rational expectations models. Starting with the seminal paper by
Blanchard and Kahn (1980), an extensive literature developed which analyzes
the existence and nature of its solutions. The most influential papers, at least
for the present exposition, are Gourieroux et al. (1982), Klein (2000), Sims
(2001), among others. There is no loss of generality involved by confining
the analysis to first order equations as higher order equations can be reduced
to first order ones by inflating the dimension of the process (see Binder and
Peseran, 1994).
In the following, Φ1 is not necessarily invertible. Thus, we allow for the
possibility that some equations do not involve expectational terms. The

103
104 CHAPTER 5. STOCHASTIC DIFFERENCE EQUATION

stochastic theory developed below is therefore more general than its deter-
ministic counterpart.
To make the problem tangible, we consider only a certain class of so-
lutions. In particular, we require that the exogenous input process {Zt } is
bounded in Lp , p > 1.1 This means that supt kZt kp < ∞. This assumption
implicitly restricts the solution processes to be bounded as well.

Assumption 5.1. We restrict the class of processes {Zt } to processes bounded

in Lp , p > 1. Thus, supt EkZt kp < ∞.

Remark 5.1. The class of stationary processes is the prime example of such
processes as the expected value and the variance remain constant. In many
applications, {Zt } is specified as an ARMA-process.2

Throughout the analysis we follow King and Watson (1998) and assume
that the linear matrix pencil Φ1 z + Φ0 is regular:3

Assumption 5.2. The linear matrix pencil Φ1 z + Φ0 is regular, i.e. there

exists at least one z ∈ C such that det(Φ1 z + Φ0 ) 6= 0.

Remark 5.2. If the matrix pencil Φ1 z + Φ0 would not be regular, there

would exist a polynomial vector P in the forward operator FXt = Et Xt+1
such that P (F)Zt = 0 for arbitrary processes {Zt }. However, this cannot be
the case. Note also that the assumption allows either Φ1 or Φ0 or both to
be singular. In that it generalizes the common assumption Φ0 and Φ1 to be
invertible.

As was already pointed out by Blanchard and Kahn (1980), the notion of
a predetermined variable or process is key for understanding the nature of the
solution.4 Following Klein (p.1412, 2000), we adopt the following definition.

Definition 5.1. A stochastic process {Kt } adapted to the filtration {Ft } is

a predetermined process if
1

1/p
Lp denotes the space of random variables such that kXkp = kXkp dP
R
< ∞
where inside the integral k.k is the Euclidian norm in Rn . If p = 2, we obtain the square-
integrable random variables. In this case Lp becomes a Hilbert space.
2
ARMA-processes are stationary processes which fulfill a stochastic difference equation
of the form Zt − a1 Zt−1 − . . . − ap Zt−p = Ut + b1 Ut−1 + . . . + bq Ut−q , ap bq 6= 0, with {Ut }
being white noise.
3
See Gantmacher (1959) for an extensive discussion of matrix pencils and the general-
ized eigenvalue problem of finding λ ∈ C such det(Φ1 z + Φ0 ) = 0.
4
Sims (2001) provides an alternative approach which does not rely on an a priori division
of the variables into predetermined and non-predetermined ones.
5.2. THE UNIVARIATE CASE 105

(i) The process of expectational errors {ηt } with ηt+1 = Kt+1 − Et Kt+1 is
an exogenous martingale difference process;
(ii) K0 ∈ F0 is exogenously given.
This definition is more general than the one given in Blanchard and Kahn
(1980) who require that Kt+1 = Et Kt+1 .
Finally, note that the superposition principle still applies in this context.
(1) (2)
Suppose that there exists two solutions {Xt } and {Xt } then {Xt } =
(1) (2)
{Xt − Xt } satisfies the homogeneous stochastic expectational difference
equation
Φ1 Et Xt+1 = Φ0 Xt .
Thus, the general solution of equation (5.1) is
(g) (p)
X t = Xt + Xt
(g)
where Xt denotes the general solution to the homogeneous equation and
(p)
Xt a particular solution to the nonhomogeneous equation.

5.2 The univariate case

Before turning to the details of the multivariate case, we will lay out the main
principles by examining a simple univariate example. Consider as a prototype
the Cagan model (Cagan, 1956) which we also analyzed in its deterministic
form in Section 2.4.3. As in equation (2.33), the current logged price level,
here denoted by Xt , is determined by the expectation of its value tomorrow
and some exogenous bounded process {Zt } which is a simple transformation
of money supply:
Et Xt+1 = φXt + Zt , φ 6= 0. (5.2)
The superposition principle implies that we can find a solution in two steps.
First, find the general solution to the homogeneous equation
Et Xt+1 = φXt . (5.3)
Then, second, find a particular solution to the nonhomogeneous equation.
Note that the equation (5.2) can be rewritten as a first order autoregres-
sive scheme:
Xt+1 = Et Xt+1 + (Xt+1 − Et Xt+1 ) = φXt + Zt + ηt+1 (5.4)
where the expectational errors ηt+1 = Xt+1 − Et Xt+1 form a martingale
difference sequence.5 Although this process has no serial correlation, it is
not necessarily white noise because its variance may change over time.
5
This is the representation preferred by Sims (2001).
106 CHAPTER 5. STOCHASTIC DIFFERENCE EQUATION

5.2.1 Solution to the homogeneous equation

In order to find the general solution to the homogeneous equation (5.3) note
that {Mt } defined as Mt = φ−t Xt is a martingale with respect to {Ft }:

Et Mt+1 = Et φ−t−1 Xt+1 = φ−t−1 Et Xt+1 = φ−t−1 φXt = Mt .

This suggests that the general solution to the homogeneous equation is of

the form
(g)
Xt = φt Mt (5.5)
where {Mt } is any martingale defined with respect to {Ft }. Mt plays the
same role as the constant c in the deterministic case. Given Assumption 5.1,
we consider only solutions which are bounded in Lp , p > 1 i.e. for which
(g)
supt kXt kp < ∞. By the Martingale Convergence Theorem (see Hall and
Heyde, 1980), there exists a random variable M ∈ Lp such that Mt =
E(M |Ft ) = Et M . Moreover, kMt − M kp converges to zero for t → ∞.
As in the deterministic case, we can distinguish several cases depending
on the value of φ.

|φ| > 1: An implication of the Martingale Convergence Theorem is that EkMt k

(g)
converges to EkM k < ∞. Thus, for {Xt } to remain bounded Mt must
be equal to zero for all t and hence the general solution of the homo-
geneous equation vanishes.
(g)
φ = 1: The general solution is Xt = Mt which converges to M .

φ = −1: No convergent solution exists.

|φ| < 1: In this case, the representation (5.4) implies that the solution fol-
lows an autoregressive process of order one. This representation admits
a causal representation with respect to the
P expectational errors. This
representation is given by Xt = φt X0 + t−1 j=0 φj
η t−j and is bounded.
However, if {Xt } is not predetermined, there is no starting value X0
and we are faced with a situation of indeterminacy because any mar-
tingale difference sequence defined with respect to Ft would satisfy the
difference equation.

5.2.2 Finding a particular solution

A particular solution can be found by iterating the difference equation for-
ward in time and applying the law of iterated expectations. After k iterations
5.2. THE UNIVARIATE CASE 107

one obtains:

Xt = φ−1 Et Xt+1 − φ−1 Zt

= φ−1 Et φ−1 Et+1 Xt+2 − φ−1 Zt+1 − φ−1 Zt

= φ−2 Et Xt+2 − φ−1 Zt − φ−2 Et Zt+1

= ...
k
X
−k−1 −1
=φ Et Xt+k+1 − φ φ−j Et Zt+j
j=0

As we are looking for solutions which remain bounded, this suggests to take
∞
X
−1
Xt = −φ φ−j Et Zt+j (5.6)
j=0

as a particular solution if |φ| > 1. As {Zt } is bounded the expression (5.6)

qualifies as a candidate for the particular solution when |φ| > 1. Note that
{Xt } will be bounded (stationary) for any bounded (stationary) process {Zt }.
If |φ| < 1, the forward iteration may still make sense if Et kZt+j k goes
to zero quick enough. Take as example the case where {Zt } follows an au-
toregressive process of order one, i.e. Zt = ρZt−1 + ut with ut ∼ WN(0, σ 2 )
and |ρ| < 1. This specification implies that Et Zt+j = ρj Zt . Thus, as long as
|φ−1 ρ| < 1, the forward solution will exist and will be bounded (stationary).
However, this will not be true for every |ρ| < 1.
In general, if |φ| < 1, we may consider the equivalent representation (5.4)
instead. Iterating this equation backward, we obtain:
t−1
X
t
Xt = φ X0 + φj (Zt−j + ηt−j ).
j=0

This solution, however, makes only sense when {Xt } is predetermined so that
X0 is given.

5.2.3 Example: Cagan’s model of hyperinflation

Let us illustrate our findings by taking up again Cagan’s model of hyperin-
flation with rational expectations (see section 2.4). This model leads to the
following difference equation in the logged price level {pt }:

α−1 mt
Et pt+1 = pt + , α < 0, (5.7)
α α
108 CHAPTER 5. STOCHASTIC DIFFERENCE EQUATION

where {mt } is now a stochastic process and where the expected price level is
replaced by the conditional expectation of the logged price level.
Given that α < 0, the coefficient of pt , φ = (α − 1)/α, is positive and
strictly greater than one. Therefore the only bounded (stationary) solution
(g)
to the homogeneous equation is {Xt } = 0 and a particular solution can be
found by forward iteration. Thus the solution is:
∞  j
1 X α
pt = Et mt+j . (5.8)
1 − α j=0 α − 1

Suppose that the money supply follows an autoregressive process of order

one:
mt = amt−1 + εt , |a| < 1 and εt ∼ IID(0, σ 2 ).
The conditional expectation Et mt+j then equals aj mt . Inserting this into
equation (5.8) leads to
∞  j
1 X α 1
pt = aj mt = mt .
1 − α j=0 α−1 1 − α(1 − a)

This shows that the relation between the price level and money supply de-
pends on the conduct of monetary policy, i.e. it depends on the autoregressive
coefficient a. Thus whenever the monetary authority changes its rule, it af-
fects the relation between pt and mt . This cross-equation restriction is viewed
by Hansen and Sargent (1980) to be the hallmark of rational expectations.
It also illustrates that a simple regression of pt on mt can not be considered a
structural equation, i.e. cannot uncover the true structural coefficients (α in
our case), and is therefore subject to the so-called Lucas-critique (see Lucas
(1976)).
A similar conclusion is reached if money supply follows a moving average
process of order one instead of an autoregressive process of order one:

mt = εt + θεt−1 , |θ| < 1 and εt ∼ WN(0, σ 2 ).

As |θ| < 1, the process for mt is invertible and we can express εt as

εt = mt − θmt−1 + θ2 mt−2 − . . .

This then leads to the following conclusions:

mt+1 = εt+1 + θεt = εt+1 + θ(mt − θmt−1 + θ2 mt−2 − . . .)

Et mt+1 = θ(mt − θmt−1 + θ2 mt−2 − . . .)
5.3. THE MULTIVARIATE CASE 109

and

Et mt+2 = Et εt+2 + θEt εt+1 = 0.

The particular solution then becomes
1 αθ
pt = mt + (mt − θmt−1 + θ2 mt−2 − . . .)
1−α (α − 1)(1 − α)
αθ2 αθ3
 
αθ mt
= 1+ + m t−1 − mt−2 + . . .
α − 1 1 − α (1 − α)2 (1 − α)2
Adding the corresponding expression for θpt−1 finally gives:
 
αθ mt mt−1
pt + θpt−1 = 1 + +θ
α−1 1−α 1−α
This shows that {pt } follows an autoregressive moving average process of
order (1, 1), i.e. an ARMA(1,1) process, with respect to {mt }. The same
remarks as before also apply in this case. Note that the AR-polynomial of
pt and the MA-polynomial of mt coincide. This remains true for general
ARMA-specifications for {mt } (Gourieroux et al., 1982).

5.3 The multivariate case

Following the bulk of the literature, we try to decouple the system into a non-
explosive (bounded) and an explosive (unbounded) part. Suppose that there
(k)
are nk predetermined variables assembled in {Xt } then we can partition
the vector Xt as !
(k)
Xt
Xt = (d) .
Xt
The analysis proceeds by first examining the case where Φ1 is invertible.
This is the specification investigated initially by Blanchard and Kahn (1980).

Φ1 invertible
The invertibility of Φ1 implies that we can rewrite equation (5.1) as
Et Xt+1 = ΦXt + Z̃t , t = 0, 1, 2 . . .
where Φ = Φ−1 −1
1 Φ0 and Z̃t = Φ1 Zt . Let us further assume that Φ is diagonal-
izable with Φ = QΛQ−1 , Λ diagonal. As in the discussion of the deterministic
case in Section 3.5, we partition Λ as
 
Λ1 0
Λ=
0 Λ2
110 CHAPTER 5. STOCHASTIC DIFFERENCE EQUATION

such that the eigenvalues in Λ1 are strictly inside the unit circle whereas those
in Λ2 are strictly outside the unit circle. We disregard the case of eigenvalues
on the unit circle.
We make the following assumption with respect to the dimension of Λ1
and Λ2 .

Assumption 5.3. The dimension of Λ1 is nk . Thus, there are exactly as

many eigenvalues inside the unit circle as there are predetermined variables.
Hence, there are as many non-predetermined variables as there are eigenval-
(k) (1) (d) (2)
ues outside the unit circle. This implies that Xt = Xt and Xt = Xt .

We will discuss later what happens if this condition is violated. Parti-

tioning Q and Xt accordingly leads to
!  ! !
(1)    (11)  (1) (1)
Et Xt+1 Q11 Q12 Λ1 0 Q Q(12) Xt Z̃t
(2) = +
Et Xt+1 Q21 Q22 0 Λ2 Q(21) Q(22) Xt
(2) (2)
Z̃t
 (11) (12)

Q Q
where Q−1 = . Multiplying this equation from the left by Q−1
Q(21) Q(22)
leads to the decoupled system
!  ! !
(1)  (1) (1)
Et Yt+1 Λ1 0 Yt Z̄t
(2) = (1) + (2)
Et Yt+1 0 Λ2 Yt Z̄t

where Q−1 Xt = Yt and Q−1 Z̃t = Z̄t , i.e.

(1) (1) (2)
Yt = Q(11) Xt + Q(12) Xt
(2) (1) (2)
Yt = Q(21) Xt + Q(22) Xt
(1) (1) (2)
Z̄t = Q(11) Z̃t + Q(12) Z̃t
(2) (1) (2)
Z̄t = Q(21) Z̃t + Q(22) Z̃t .

Following the logic of the discussion in Section 5.2, the unique bounded
(2)
solution for {Yt } is
∞
X ∞
X  
(2)
Λ−j
(2)
Yt = −Λ−1
2
−1
2 Et Z̄t+j = −Λ2 Λ−j
2 Et Q(21) (1)
Z̃ t+j + Q(22) (2)
Z̃ t+j
j=0 j=0
X∞
= −Λ−1 Λ−j


2 2 Q(21) Q(22) Et Z̃t+j . (5.9)
j=0
5.3. THE MULTIVARIATE CASE 111

Turn next to the first part of the decoupled equation. Note that the
(1)
predetermined variable Xt satisfies the identity:
   
(1) (1) (1) (1) (2) (2)
Xt+1 − Et Xt+1 = Q11 Yt+1 − Et Yt+1 + Q12 Yt+1 − Et Yt+1 = ηt+1 .
(1)
Inserting this equation into the equation for Yt+1 , we get
 
(1) (1) (1) (1)
Yt+1 = Et Yt+1 + Yt+1 − Et Yt+1
 
(1) (2) (2)
= Λ1 Yt + Z̄t (1) + Q−1
11 η t+1 − Q −1
11 Q 12 Yt+1 − Et Y t+1
(1)
= Λ1 Yt + Z̄t (1) + Q−1 −1
11 ηt+1 − Q11 Q12 εt+1 (5.10)
(2) (2)
where εt+1 = Yt+1 − Et Yt+1 is an exogenous martingale difference process.
Thus, equation (5.10) is a first order autoregressive scheme with starting
value given by  
(1) (k) (2)
Y0 = Q−1 11 X 0 − Q 12 Y 0 .
Equations (5.10) and (5.9) determine the solution for Yt . This step in the
derivation is only valid if Q11 is invertible. Otherwise, we could not determine
(1) (1)
the initial values of Yt from those of Xt and there would be a lack of initial
values for Yt .6 Hence, Assumption 5.3 is not sufficient for the uniqueness of
the solution. In addition, we need the following assumption.
Assumption 5.4. Q11 is nonsingular.
Finally, the solution for Yt can be turned back into a solution for Xt by
multiplying Yt by Q.
We can get further insights into the nature of the solution by assuming
that {Z̃t } is a causal autoregressive process of order one:
Z̃t+1 = AZ̃t + ut+1 , ut ∼ WN(0, σ 2 ) and kAk < 1
where {ut } is exogenous. This specification implies that Et Z̃t+j = Aj Z̃t ,
j = 1, 2, . . . Inserting this into equation (5.9) we find that
∞
X
(2)
−Λ−1 Λ−j Q(21) Q(22) Aj Z̃t = M Z̃t


Yt = 2 2
j=0

−1
P∞ −j (21) (22)

j (1)
where M = −Λ2 j=0 Λ 2 Q Q A . The solution to Yt then can
be written as
(1) (1)
Yt+1 = Λ1 Yt + Z̄t (1) + Q−1 −1
11 ηt+1 − Q11 Q12 M ut+1 .
6
See Klein (2000, section 5.3.1) and King and Watson (2002) for details and examples.
112 CHAPTER 5. STOCHASTIC DIFFERENCE EQUATION

Finally, the initial condition can be computed as

 
(1) (k)
Y0 = Q−1 11 X 0 − Q12 M Z̃0 .

Before turning to the general case several remarks are in order.

Remark 5.3. The above derivation remains valid even if the matrix Φ is not
diagonalizable. In this case, we will have to work with the Jordan canonical
form instead (see Section 3.2.2).

Remark 5.4. The derivation excluded the possibility of roots on the unit
circle.

Φ1 singular
In many practical applications, Φ1 is not invertible so that the procedure just
outlined is not immediately applicable. This, for example, is the case when
a particular equation contains no expectations at all which translates into a
corresponding row of zeros in Φ1 . One way to deal with this problem is to
take a generalized inverse of Φ1 and proceed as explained above.
The most appropriate type of generalized inverse in the context of dif-
ference equations is the Drazin-inverse (see Campbell and Meyer, 1979, for
a comprehensive exposition). This generalized inverse can be obtained for
any n × n matrix A in the following manner. Denote by IndA the smallest
nonnegative integer k such that rankAk = rankAk+1 . This number is called
the index of A. Then the following Theorem holds (see Theorem 7.2.1 in
Campbell and Meyer, 1979).

Theorem 5.1. Let A be an n × n matrix with Ind(A) = k > 0, then there

exists a nonsingular matrix P such that
 
C 0
A=P P −1
0 N

where C is nonsingular and N is nilpotent of index k (i.e. N k = 0). The

Drazin-inverse AD is then given by
 −1 
C 0
D
A =P P −1 .
0 0

With this Theorem in mind, we can now decouple the system in two
parts. The first part will be similar to the case when Φ1 is invertible. The
second one will correspond to the singular part and will in some sense solve
5.3. THE MULTIVARIATE CASE 113

out the expectations. Multiply for this purpose equation (5.1) from the left
by (zΦ1 + Φ0 )−1 where assumption 5.2 guarantees that the inverse exists for
some number z. Thus, we get

(zΦ1 + Φ0 )−1 Φ1 Et Xt+1 = (zΦ1 + Φ0 )−1 Φ0 Xt

b 1 Et Xt+1 = (zΦ1 + Φ0 )−1 (zΦ1 + Φ0 − zΦ1 )Xt
Φ
b 1 Et Xt+1 = (In − z Φ
Φ b 1 )Xt

where Φb 1 = (zΦ1 + Φ0 )−1 Φ1 . The application of Theorem 5.1 to Φ

b 1 then
leads to the decoupled system
    
C 0 C 0
Et Xt+1 = In − z
e Xet
0 N 0 N

et denotes P −1 Xt . This leads to the following two equations:

where X
(1) et(1)
CEt Xet+1 = (In1 − zC)X
(2) et(2)
N Et X
et+1 = (In2 − zN )X

where X et has been partitioned appropriately. As C is invertible, the first

difference equation can be treated as outlined above. By shifting the time
index, the second equation can be written as
e (2) = (In2 − zN )X
N Et X e (2) .
t+k t+k−1

Applying the law of iterated expectations and multiplying the equation from
the left by N k−1 gives
e (2) = (In2 − zN )2 N k−2 Et X
0 = N k Et X e (2)
t+k t+k−2
3 k−3
= (In2 − zN ) N Et X e (2)
t+k−3
= ...
et(2)
= (In2 − zN )k Et X
et(2) .
= (In2 − zN )k X

Because (In2 − zN )k is invertible, the only solution to the above equation is

et(2) = 0.
X
114 CHAPTER 5. STOCHASTIC DIFFERENCE EQUATION
Appendix A

Complex Numbers

As the simple quadratic equation x2 + 1 = 0 has no solution in the field

of real numbers, R, it is necessary to envisage the larger field of complex
numbers C. A complex number z is an ordered pair (a, b) of real numbers
where ordered means that we regard (a, b) and (b, a) as distinct if a 6= b. We
endow the set of complex numbers by an addition and a multiplication. Let
x = (a, b) and y = (c, d) be two complex numbers, then we have the following
definitions:

addition: x + y = (a, b) + (c, d) = (a + c, b + d)

multiplication: xy = (a, b)(c, d) = (ac − bd, ad + bc).

These two operations will turn C into a field where (0, 0) and (1, 0) play
the role of 0 and 1.1 The real numbers R are embedded into C because we
identify any a ∈ R with (a, 0) ∈ C.
The number ı = (0, 1) is of special interest. It solves the equation x2 +1 =
0, i.e. ı2 = −1. The other solution being −ı = (0, −1). Thus any complex
number (a, b) may be written as (a, b) = a + ıb where a, b are arbitrary real
numbers.2

1
Substraction and division can be defined accordingly:

subtraction: (a, b) − (c, d) = (a − c, b − d)

(ac + bd, bc − ad)
division: (a, b)/(c, d) = , c2 + d2 6= 0.
(c2 + d2 )

2
A more detailed introduction of complex numbers can be found in Rudin (1976) or
any other mathematics textbook.

115
116 APPENDIX A. COMPLEX NUMBERS

Representation of complex numbers

2

z=a+ib
b
i
1

r
imaginary part

−1 θ a
0
1

−1
−i
−b
unit circle: z=a−ib
a2 + b2 = 1

−2
−2 −1 0 1 2
real part

Figure A.1: Representation of a complex number

An element z in this field can be represented in two ways:

z = a + ıb Cartesian coordinates
ıθ
= re = r(cos θ + ı sin θ) polar coordinates.

In the representation in Cartesian coordinates a = Re(z) = <(z) is called

the real part whereas b = Im(z) = =(z) is called the imaginary part of z.
A complex number z can be viewed as a point in the two-dimensional
Cartesian coordinate system with coordinates (a, b). This geometric inter-
pretation is represented in Figure A.1.
√ The absolute value or modulus of z, denoted by |z|, is given by r =
a2 + b2 . Thus the absolute value is nothing but the distance of z viewed
as a point in the complex plane (the two-dimensional Cartesian coordinate
system) to the origin (see Figure A.1). θ denotes the angle to the positive real
axis (x-axis) measured in radians. It is denoted by θ = arg z. It holds that
tan θ = ab . Finally, the conjugate of z, denoted by z̄, is defined by z̄ = a − ıb.
Setting r = 1 and θ = π, gives the following famous formula:

eıπ + 1 = (cos π + ı sin π) + 1 = −1 + 1 = 0.

 117

This formula relates the most famous numbers in mathematics.

From the definition of complex numbers in polar coordinates, we get
immediately the following implications:

eıθ + e−ıθ a
cos θ = = ,
2 r
eıθ − e−ıθ b
sin θ = = .
2ı r
Further implications are de Moivre’s formula and the Pythagoras’ theorem
(see Figure A.1):

n
de Moivre’s formula reıθ = rn eınθ = rn (cos nθ + ı sin nθ)
Pythagoras’ theorem 1 = eıθ e−ıθ = (cos θ + ı sin θ)(cos θ − ı sin θ)
= cos2 θ + sin2 θ

From Pythagoras’ theorem it follows that r2 = a2 + b2 . The representation

in polar coordinates allows to derive many trigonometric formulas.
Consider the polynomial Φ(z) = φ0 −φ1 z −φ2 z 2 −. . .−φp z p of order p ≥ 1
with φ0 = 1.3 The fundamental theorem of algebra then states that every
polynomial of order p ≥ 1 has exactly p roots in the field of complex numbers.
Thus, the field of complex numbers is algebraically complete. Denote these
roots by λ1 , . . . , λp , allowing that some roots may appear several times. The
polynomial can then be factorized as follows:

Φ(z) = 1 − λ−1 1 − λ−1 −1

1 z 2 z . . . 1 − λp z .

This expression is well-defined because the assumption of a nonzero constant

(φ0 = 1 6= 0) excludes the possibility of roots equal to zero. If we assume
that the coefficients φj , j = 0, . . . , p, are real numbers, the roots appear in
conjugate pairs. Thus if z = a + ıb, b 6= 0, is a root then z̄ = a − ıb is also a
root.

3
The notation with “−φj z j ” instead of “φj z j ” was chosen to conform to the notation
of AR-models.
118 APPENDIX A. COMPLEX NUMBERS
Appendix B

Matrix Norm

Consider a vector x in Rn and a matrix A ∈ Rn×n . If kxk is any vector norm

on Rn then we may consider the induced matrix norm:

kAk = max kAxk.

kxk=1

Thus the induced matrix norm is the maximum amount a vector on the unit
sphere can be stretched. The matrix norm induced by the Euclidian vector
norm is given by: p
kAk = max kAxk = ρ(A0 A)
kxk=1

where ρ(A0 A) is the spectral radius of A0 A, i.e. ρ(A0 A) = max{|λ| : λ is an eigenvalue of A0 A}.
Another convenient matrix norm is the Frobenius norm, sometimes also
called the Hilbert-Schmidt or the Schur norm. It is defined as follows:
n
X n
X
kAk2 = |aij |2 = tr(A0 A) = λi
i,j i

where λi are the eigenvalues of A0 A. Thus the Frobenius norm stakes the
columns of A into a long n2 -dimensional vector and takes its Euclidian norm.
The matrix norm has the following properties:

kAk ≥ 0 and kAk = 0 ⇔ A = 0.

kαAk = |α|kAk for all α ∈ R.
kA + Bk ≤ kAk + kBk for all A, B ∈ Rn×n .
kABk ≤ kAkkBk for all A, B ∈ Rn×n .

The last property is called submultiplicativity.

Because all norms are equivalent, it does not really matter which one we
will use. For more details see Meyer (2000).

119
120 APPENDIX B. MATRIX NORM
Bibliography

R. P. Agarwal. Difference Equations and Inequalities. Pure and Applied

Mathematics. Marcel Dekker, New York, second edition, 2000.

O. Ashenfelter and D. Card. Time series representations of economic vari-

ables and alternative models of the labour market. Review of Economics
and Statistics, 49:761–782, 1982.

C. Azariadis. Self-fulfilling prophecies. Journal of Economic Theory, 25:

380–396, 1981.

C. Azariadis. Intertemporal Macroeconomics. Blackwell Publishers, Cam-

bridge, Massachusetts, 1993.

C. Azariadis and R. Guesnerie. Sunspots and cycles. Review of Economic

Studies, 53:725–736, 1986.

M. Binder and H. Peseran. Multivariate rational expectations models: A

review of some results. In Handbook of Applied Econometrics: Macroeco-
nomics, pages 139–187. Basil Blackwell, Oxford, 1994.

O. J. Blanchard and C. M. Kahn. The solution of linear difference models

under rational expectations. Econometrica, 48(5):1305–1311, 1980.

P. Cagan. The monetary dynamics of hyperinflation. In M. Friedman, edi-

tor, Studies in the Quatity Theory of Money, pages 23–117. University of
Chicago Press, Chicago, 1956.

S. L. Campbell and C. D. Meyer, Jr. Generalized Inverses of Linear Trans-

formations. Dover Books on Advanced Mathematics. Dover Publications,
New York, 1979.

D. Cass and K. Shell. Do sunspots matter. Journal of Political Economy,

91:193–227, 1983.

121
122 BIBLIOGRAPHY

F. Colonius and W. Kliemann. Dynamical Systems and Linear Algebra, vol-

ume 158 of Graduate Studies in Mathematics. American Mathematical
Society, Providence, Rhode Island, 2014.

R. Dornbusch. Expectations and exchange rate dynamics. Journal of Political

Economy, 84:1161–76, 1976.

S. N. Elaydi. An Introduction to Difference Equations. Springer, New York,

third edition, 2005.

R. E. A. Farmer. The Macroeconomics of Self-Fulfilling Prophecies. The MIT

Press, Cambridge, Massachusetts, 1993.

J. Galı́. Monetary Policy, Inflation, and the Business Cycle: An Introduction

to the New Keynesian Framework. Princeton University Press, Princeton,
New Jersey, 2008.

J. Galı́. Are central banks’ projections meaningful? Journal of Monetary

Economics, 58:537–550, 2011.

O. Galor. Discrete Dynamical Systems. Springer, Berlin, 2007.

F. R. Gantmacher. Matrix Theory, volume 2. Chelsea Publishing Company,

New York, 1959.

C. Gourieroux, J. Laffont, and A. Monfort. Rational expectations in dynamic

linear models: Analysis of the solutions. Econometrica, 50:409–425, 1982.

P. Hall and C. C. Heyde. Martingale Limit Theory and its Applications.

Academic Press, San Diego, 1980.

L. P. Hansen and T. J. Sargent. Formulating and estimating dynamic linear

rational expectations models. Journal of Economic Dynamics and Control,
2:7–46, 1980.

K. L. Judd. Short-run analysis of fiscal policy in a simple perfect foresight

model. Journal of Political Economy, 93:298–319, 1985.

C. P. Kindleberger. Maniacs, Panics, and Crashes. The Macmillan Press,

London, 1978.

R. G. King and M. W. Watson. The solution of singular linear difference

systems under rational expectations. International Economic Review, 39:
1015–1026, 1998.
BIBLIOGRAPHY 123

R. G. King and M. W. Watson. System reduction and solution algorithms

for singular linear difference systems under rational expectations. Compu-
tational Economics, 20:57–86, 2002.

P. Klein. Using the generalized Schur form to solve a multivariate linear

rational expectations model. Journal of Economic Dynamics and Control,
24:1405–1423, 2000.

R. E. Lucas, Jr. Econometric policy evaluation: A critique. Carnegie-

Rochester Conference Series on Public Policy, 1:19–46, 1976.

C. D. Meyer. Matrix Analysis and Applied Linear Algebra. SIAM, Philadel-

phia, 2000.

H. Moore. Economic Cycles: Their Law and Cause. MacMillan, New York,
1914.

J. F. Muth. Rational expectations and the theory of price movements. Econo-

metrica, 29(3):315–335, 1961.

A. W. Naylor and G. R. Sell. Linear Operator Theory in Engeneering and

Science. Springer-Verlag, New York, 1982.

S. A. O’Connell and S. P. Zeldes. Rational ponzi games. International

Economic Review, 29(3):431–450, 1988.

F. P. Ramsey. A mathematical theory of saving. Economic Journal, 38:

543–559, 1928.

R. C. Robinson. Dynamical Systems: Stability, Symbolic Dynamics, and

Chaos. CRC Press, Boca Raton, Florida, 2nd edition, 1999.

K. Rogoff. Dornbusch’s overshooting model after twenty-five years. IMF

Staff Papers, 49, 2002.

W. Rudin. Principles of Mathematical Analysis. McGraw-Hill, New York,

3rd edition, 1976.

P. Samuelson. Interactions between multiplier analysis and the principle of

acceleration. Review of Economic Studies, 21:75–78, 1939.

T. J. Sargent. Macroeconomic Theory. Academic Press, Orlando, Florida,

2nd edition, 1987.
124 BIBLIOGRAPHY

H. Sedaghat. The impossibility of unstable, globally attracting fixed points

for continuous mappings of the line. American Mathematical Monthly, 104:
356–358, 1997.

C. A. Sims. Solving linear rational expecattions models. Computational

Economics, 20:1–20, 2001.

R. M. Solow. A contribution to the theory of economic growth. Quarterly

Journal of Economics, 70(1):65–94, February 1956.

N. L. Stokey and R. E. Lucas Jr. Recursive Methods on Economic Dynamics.

Harvard University Press, Cambridge, Massachusetts, 1989.

G. Strang. Introduction to Linear Algebra. Wellesley-Cambridge Press, Cam-

bridge: Massachusetts, third edition, 2003.

J. B. Taylor. Aggregate demand and staggered contracts. Journla of Political

Economy, 88:1–23, 1980.

M. Woodford. Interest and Prices: Foundations of a Theory of Monetary

Policy. Princeton University Press, Princeton, New Jersey, 2003.