Stochastic Economic Growth:

An Operator-Theoretic Approach

John Stachurski

Submitted in fulfillment of the requirements

of the degree of Doctor of Philosophy

Department of Economics

The University of Melbourne

July 9, 2002
 1

Abstract: For many years the trend in macroeconomics has been towards
models which are both explicitly stochastic and explicitly dynamic. With
these models, researchers seek to replicate and explain observable properties
of the major economic time series. One manifestation of this trend towards
stochastic dynamic modeling has been increasing use of the inherently dy-
namic models developed in the field of economic growth. The latter have
proved to be suitable not only for the study of growth and development, but
also for that of many other areas within macroeconomics, such as business
cycles, fiscal policy and public finance.

This thesis is a re-examination of the stochastic dynamics arising from some

well-known models of economic growth. The focus is particularly on ergodic
properties and the existence of stable equilibria, where equilibrium here is
defined in the usual stochastic growth sense (i.e., as the stationary distribu-
tion of a Markov process). Brief consideration is also given to asymptotic
statistical properties of economic time series, such as law of large numbers
and central limit tendencies.

On one hand, the thesis has been motivated by the availability of new and un-
exploited techniques for studying the kinds of Markovian systems generated
by these models. The techniques in question are operator-theoretic, with
a particular focus on integral Markov semigroups in the function space L1 .
They are particularly well suited to analysis of Markov chains on unbounded
state space.

On the other hand, motivation also comes from the demand side: new con-
ditions for evaluating the stability of stochastic dynamic models are valuable
 2

to economists who are not familiar with recent mathematical innovations. In

this connection, the thesis has sought to provide sufficient conditions that are
easy to verify in applications and admit standard models in the econometric
tradition.

A highlight of the thesis is a set of new sufficient conditions for the stability of
perturbed dynamical systems on the nonnegative half-ray R+ . By introduc-
ing a Lyapunov criterion, a set of general conditions is found which includes
existing work from the mathematical literature as a special case.
 3

This is to certify that

• the thesis comprises only my original work towards the PhD,

• due acknowledgement has been made in the text to all other material
used,

• the thesis is less than 100,000 words in length, exclusive of tables, maps,
bibliographies and appendices.
 4

Acknowledgments: To begin with, the author is deeply grateful to his two

supervisors, John Creedy and Rabee Tourky. In particular, the author thanks
John for his patience and support, and for teaching him how to write in a
style which is hopefully both succinct and readable. The author thanks Rabee
for his enthusiasm and support for abstract research, as well as his insights
into writing theory papers. Other people who have contributed directly to the
content of this thesis are Peter Bardsley, Cuong Le Van, Andrzej Lasota,
Steven Dowrick and two anonymous referees at Journal of Economic Theory
and Economic Theory respectively.

The author thanks the Economic Theory Center, University of Melbourne for
generous financial support and other assistance. The author is grateful to the
following three institutions for their hospitality during research visits: Cen-
tre de Recherche en Mathématiques, Statistique et Economie Mathématique,
Université Paris 1, Panthéon-Sorbonne; CentER, Tilburg University; and the
Institute of Mathematics, Silesian University, Katowice.

Less direct but equally important contributors to this thesis were the author’s
family, in particular Mum, Dad, Marta, Terry, Nic and Gwen; as well as
friends Aaron Bilcock, Simon Hosking, Kirdan Lees and Murdoch MacPhee.
Thank you for your love and support. Finally, the author acknowledges his
great debt to Miss Cleo Fleming, a.k.a. the lemming, who, on top of ev-
erything else, managed to sit patiently through many long and excited late-
night diatribes about integral Markov semigroups and the axioms of Hausdorff
space.
Contents

1 Introduction 9

1.1 Overview of the Thesis . . . . . . . . . . . . . . . . . . . . . . 9

1.1.1 Structure of the Thesis . . . . . . . . . . . . . . . . . . 10

1.1.2 Sufficient Conditions . . . . . . . . . . . . . . . . . . . 12

1.2 Existing Literature . . . . . . . . . . . . . . . . . . . . . . . . 13

2 Equilibria of Markovian Models 15

2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15

2.1.1 Growth and Markov Chains . . . . . . . . . . . . . . . 17

2.1.2 Outline of the Chapter . . . . . . . . . . . . . . . . . . 18

2.2 Overview of the Method . . . . . . . . . . . . . . . . . . . . . 19

2.2.1 Outline . . . . . . . . . . . . . . . . . . . . . . . . . . 19

2.2.2 Discrete Dynamical Systems . . . . . . . . . . . . . . . 20

2.2.3 Perturbed Dynamical Systems . . . . . . . . . . . . . . 22

5
CONTENTS 6

2.3 Markov Operators . . . . . . . . . . . . . . . . . . . . . . . . . 25

2.3.1 Operators on Measures . . . . . . . . . . . . . . . . . . 26

2.3.2 The L1 Method . . . . . . . . . . . . . . . . . . . . . . 29

2.4 General Fixed Point Results . . . . . . . . . . . . . . . . . . . 32

2.4.1 Lagrange Stability . . . . . . . . . . . . . . . . . . . . 33

2.4.2 Strongly Contractive Systems . . . . . . . . . . . . . . 35

2.5 Applications to Markov Operators . . . . . . . . . . . . . . . . 38

2.5.1 Contractive Markov Operators . . . . . . . . . . . . . . 38

2.5.2 Lagrange Stability of Markov Operators . . . . . . . . 40

2.5.3 Weak Precompactness . . . . . . . . . . . . . . . . . . 41

3 Optimal Growth with Unbounded Shock 43

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43

3.2 Formulation of the Problem . . . . . . . . . . . . . . . . . . . 45

3.2.1 Assumptions . . . . . . . . . . . . . . . . . . . . . . . . 46

3.2.2 Technology . . . . . . . . . . . . . . . . . . . . . . . . 47

3.2.3 The Optimal Policy . . . . . . . . . . . . . . . . . . . . 47

3.3 Statement of Results . . . . . . . . . . . . . . . . . . . . . . . 49

3.3.1 Examples . . . . . . . . . . . . . . . . . . . . . . . . . 51

3.3.2 Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . 51
CONTENTS 7

3.4 Semidynamical Systems . . . . . . . . . . . . . . . . . . . . . 53

3.5 Proof of the Main Theorem . . . . . . . . . . . . . . . . . . . 54

3.5.1 Proof of Lagrange Stability . . . . . . . . . . . . . . . 54

3.5.2 Proof of Theorem 3.2 . . . . . . . . . . . . . . . . . . . 63

4 Systems with Multiplicative Noise 64

4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64

4.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65

4.3 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68

4.3.1 Overlapping Generations . . . . . . . . . . . . . . . . . 68

4.3.2 Stability in a Model with Externalities . . . . . . . . . 71

4.3.3 Existing Conditions . . . . . . . . . . . . . . . . . . . . 75

4.4 Formulation of the Problem . . . . . . . . . . . . . . . . . . . 76

4.5 Proofs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78

5 Asymptotic Distributions 84

5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84

5.2 The Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86

5.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88

5.4 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89

5.5 Proofs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
CONTENTS 8

6 Linearization of Stochastic Models 95

6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95

6.2 Formulation of the Problem . . . . . . . . . . . . . . . . . . . 97

6.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98

6.4 Application . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99

6.5 Proofs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102

A Existing Conditions 115

Chapter 1

Introduction

1.1 Overview of the Thesis

The stochastic growth model, which links production to savings, and savings
in turn to investment and production, has become a benchmark model in
economics, not just for theoretical and empirical growth problems, but also
for areas ranging from business cycle theory and finance to public sector eco-
nomics and development. This thesis presents four essays which examine the
dynamic properties of a collection of well-known growth models. The pri-
mary contribution is new sufficient conditions for existence, uniqueness and
stability of stochastic equilibria. A small amount of supplementary material
considers related dynamic properties, such as the law of large numbers and
the central limit theorem.

The theme which unites the essays is our focus on operator-theoretic meth-
ods. Growth models are viewed as Markov chains, which in turn are trans-

9
CHAPTER 1. INTRODUCTION 10

lated into deterministic linear operators sending a space of signed measures

into itself. These maps—the so-called Markov operators—are defined such
that their iteration on some initial condition generates the time-indexed se-
quence of marginal distributions for the state variables in the growth model.
The primary objective of the studies is to obtain conditions such that this
sequence of distributions converges to a unique invariant limit, the latter
being independent of the economy’s initial conditions.

Many of the arguments are based on a framework for proving global sta-
bility of Markov operators suggested by the Polish mathematician Andrzej
Lasota [38]. Lasota implements a straightforward but relatively unknown
fixed point technique, which combines an interesting contraction condition
with a compactness notion called Lagrange stability.

1.1.1 Structure of the Thesis

Following this introduction, the thesis is divided into five main chapters
(Chapters 2–6). The first of these (Chapter 2) is an exposition of the math-
ematical framework outlined in the previous paragraph. The objective is to
set out the key ideas and definitions, in order to avoid having to repeat them
in each of the remaining theory chapters. Some of the results in this prelim-
inary chapter are—to the best of the author’s knowledge—new, and the key
proofs of existing results are also new.

The preliminary chapter is followed by four essays on stochastic growth the-

ory. The first essay (Chapter 3, based on the manuscript “Stochastic Optimal
Growth with Unbounded Shock,” forthcoming in the Journal of Economic
CHAPTER 1. INTRODUCTION 11

Theory, doi: 10.1006/jeth.2001.2842) treats the by-now classic problem of

Brock and Mirman [10]. Our study extends the existing analysis to a new
class of productivity shocks—including those common to econometric mod-
eling. It is hoped that this extension will therefore be helpful in further
integrating theoretical and empirical research which is based on the Brock-
Mirman framework.

Chapter 4—perhaps the core contribution of the thesis—sets out a new

framework for determining existence, uniqueness and stability of equilibria in
economic systems evolving on the nonnegative real numbers. While this es-
say is in essence an applied mathematics paper that has applications to many
areas other than growth theory (and indeed other sciences), the research was
motivated by the author’s investigations into the stochastic growth problem,
particularly the Brock-Mirman problem that was the subject of the first es-
say. Moreover, it has obvious applications to a range of one-sector growth
models, whether they be of the Ramsey-Brock-Mirman, overlapping genera-
tions or Solow-Swan formulation.

When the distributions of the state variables in a Markov chain converge over
time to a unique limiting distribution, one can perhaps anticipate that if, for
example, one takes a sufficiently long time series from that model and calcu-
lates the average value, then that value should be close to the mean of the
limiting distribution. The third essay (Chapter 5, based on the manuscript
“Stochastic Growth: Asymptotic Distributions,” forthcoming in the journal
Economic Theory) focuses on these kinds of econometric questions. More
precisely, we seek sufficient conditions for globally stable stochastic growth
CHAPTER 1. INTRODUCTION 12

models to also have law of large numbers and central limit theorem proper-
ties. Again our treatment admits the standard econometric shocks. Instead
of restricting the support of shocks, we require a rather strong “average con-
traction” condition on technology.

Finally, the fourth essay (Chapter 6) discusses linearization of random sys-

tems. Linearization is a common approach to nonlinear systems, as the
dynamics of linear stochastic models are well-understood. However, this ap-
proach uses a potentially flawed logic; there is no general result which links
the dynamic properties of the derived linear model with those of the true
(nonlinear) model. In this chapter, operator-theoretic techniques are used to
justify linearization in the important special case of log-linear models.

1.1.2 Sufficient Conditions

As a general comment regarding sufficient conditions for mathematical prop-

erties, there are two main criteria against which these conditions must be
assessed. The first is generality—to what extent do the sufficient conditions
characterize the set of models from within a given class which have the prop-
erty in question. The second criterion is the ease with which the conditions
can be verified from the primitives of a particular model.

These two criteria are in a sense conflicting. For example, a very general
sufficient condition for a given model to have property P is property P .
Obviously the condition is sufficient (P implies P ). Moreover, the condition
characterizes the class of models that have property P (not P implies not P ).
CHAPTER 1. INTRODUCTION 13

Yet it is clear that this sufficient condition does not give the researcher who
seeks to verify property P in a specific model much additional information.

The objective of the author has been to balance the two criteria. On one
hand, the conditions contained here often extend existing work. On the other
hand, it is hoped that the conditions will prove easy for applied researchers
to verify, being stated in a mathematical language which is familiar to the
non-theorist.

1.2 Existing Literature

Extensive literature reviews are included at the beginning of each essay. How-
ever, the thesis as a whole was particularly influenced by the following work.

On the economic side, the most important reference was without doubt the
classic stochastic optimal growth paper of Brock and Mirman [10]. In addi-
tion, Mirman published a number of papers around the same time which all
focused on the dynamic properties of stochastic neoclassical economies with
convex (i.e., decreasing returns) technology [44, 45].

In attempting to extend these initial contributions, current theorists are be-

stowed with Gerschenkron’s advantage of backwardness. It is sometimes dif-
ficult to fully acknowledge—or even remain conscious of—the insights that
were laid out in front of us for the first time by those earlier authors.

On the mathematical side, the greatest influence on this thesis has been
the work of Andrzej Lasota, a functional analyst at the Polish Academy of
CHAPTER 1. INTRODUCTION 14

Science. In particular, the author has made extensive use of Lasota’s invari-
ant principle for Markov semigroups [38], as well as various ideas from his
monograph on the L1 approach to dynamic systems with M. C. Mackey [40].
Chapter 2

Equilibria of Markovian Models

2.1 Introduction

Increasingly, modern economics in general and growth theory in particular is

implemented within the framework of stochastic dynamic systems. Physical
laws, equilibrium constraints and restrictions on the behavior of agents jointly
determine evolution of endogenous state variable x ∈ X according to some
transition rule
xt+1 = S(xt , zt , εt ), t = 0, 1, . . . , (2.1)

where S is an arbitrary function, (zt ) is a sequence of exogenous forcing

variables and (εt ) is uncorrelated noise.

For some models, either zt is constant or the endogenous variables can be

conveniently redefined such that the system is autonomous:

xt+1 = T (xt , εt ), t = 0, 1, . . . (2.2)

15
CHAPTER 2. EQUILIBRIA OF MARKOVIAN MODELS 16

Assume that this is the case.

Of primary concern is whether the autonomous system (2.2) is in some sense

stationary, in which case one can anticipate convergence of the sequence of
distributions (ϕt ) associated with the sequence of random variables (xt ) to
some unique limiting distribution ϕ∗ . The latter is then interpreted as the
long-run equilibrium of the economy (2.2). Typically, comparative dynamics
(policy simulation) will be performed by analyzing the relationship between
its moments and the underlying structural parameters contained in the func-
tion T and the distribution of the shock ε.

When T is linear on real vector space, (2.2) is the standard autoregression

(AR) model. Conditions for stationarity are familiar from elementary time
series analysis [25]. When the map is nonlinear, dynamic behavior is po-
tentially more complicated. General conditions for existence of unique and
stable equilibria are not known.

In this case, a common approach in the applied literature is to linearize

(2.2) using a first order Taylor expansion or similar technique, and then
examine the stability properties of the resulting AR model. However, it
is by no means clear that stability properties obtained for the AR model
have any homeomorphic implications for the behavior of the true model
(2.2). In other words, it is not in general legitimate to infer stability of (2.2)
from stability of the corresponding linear form. Moreover, linearization may
eliminate important features of the model.1
1
For example, Durlauf and Quah [16] find evidence to the effect that standard lin-
earization procedures applied to Solow-Ramsey growth models fail to extract nonlinear
CHAPTER 2. EQUILIBRIA OF MARKOVIAN MODELS 17

A more correct method is to examine the Markov chain generated by (2.2),

and determine whether appropriate conditions for stability of Markovian sys-
tems are satisfied. Early examples are Brock and Mirman [10], Mirman
[44, 45], Green and Majumdar [24], Brock and Majumdar [9] and Razin and
Yahav [51]. An excellent survey of sufficient conditions is provided by Fu-
tia [21]. Stokey, Lucas and Prescott [58, Chapter 13] outline ways to verify
these and related conditions for common economic models. Prescott and
Hopenhayn [28] develop new sufficient conditions using only monotonicity
and a mixing condition. Bhattacharya and Majumdar [7] obtain exponential
convergence in the Kolmogorov metric for real-valued systems that satisfy a
“splitting” condition.

In this thesis we treat precisely the same problem, but use a different set of
techniques that have evolved recently in the applied mathematics, physics
and biology literature. In this preliminary chapter, a rather detailed intro-
duction to these mathematical techniques is given.

2.1.1 Growth and Markov Chains

As the thesis treats only sequences of uncorrelated shocks, growth models in

the form of (2.2) can be reduced to Markov chains. While there are many ap-
proaches to the study of Markov chains, we find the operator-theoretic frame-
work most applicable to solving the relevant economic problems. Within this
framework our focus is primarily on the L1 method.
local increasing returns dynamics that are critical to understanding the evolution of the
cross-country income distribution.
CHAPTER 2. EQUILIBRIA OF MARKOVIAN MODELS 18

Much of the early L1 theory is due to Hopf [29]. The monograph of Foguel
[20] contains an extensive survey of asymptotic results. Lasota and Mackey
[40] use L1 techniques to study perturbed and chaotic systems. Operator-
theoretic treatment of Markov processes begins with Krylov and Bogolioubov
[35]. See also Kakutani and Yoshida [33]. Recently the most active work
within the L1 framework is that being conducted by a team of Polish math-
ematicians led by A. Lasota [38].2

Recently, L1 methods have been applied to the study of particle energy in an

ideal gas [38], fluctuations in the brightness of the Milky Way [40], propaga-
tion of annual plants with seed-bank [30], and cell growth in a proliferating
cell population [39, 61, 60, 41, 38].

2.1.2 Outline of the Chapter

Section 2.2 gives a heuristic overview of how perturbed dynamical systems

in general and stochastic growth models in particular can be rewritten as
continuous linear operators called Markov operators. In Section 2.3 this
discussion is formalized and related to the existing economic literature.

Performing the transformation into Markov operators converts the equilib-

rium problem into one of finding fixed points for such operators. Section
2.4 discusses useful fixed point arguments for linear operators on metric and
topological vector spaces. Section 2.5 shows how the general fixed point
results from Section 2.4 can be applied to certain types of Markov operators.
2
See also Lasota and Mackey [40] and Horbacz [30].
CHAPTER 2. EQUILIBRIA OF MARKOVIAN MODELS 19

2.2 Overview of the Method

This section gives a broad and heuristic overview of the mathematical tech-
niques used in the thesis. The discussion is intended to be intuitive rather
than rigorous. In particular, no definition of integral is given, and mathe-
matical objects are referred to without attempting to prove that they do in
fact exist. Formal arguments begin in Section 2.3.

2.2.1 Outline

For dynamic economic models, an equilibrium (or steady state) is defined to

be a point in the state space that is stationary under the period-to-period
transition rule. If such a point is obtained, then no further change is ob-
served in the system. As well as this invariance property, equilibria may be
attractive for points in the surrounding state space, which is to say that the
transition rule moves nearby points closer to the equilibrium.

In the case of stochastic models, a state cannot be stationary in the same

sense as those in deterministic systems, given that shocks continue to disturb
activity in each period. Instead, a steady state must be viewed as a situation
where the probabilistic laws that govern the state variables cease to change
over time [24]. For stochastic economies the notion of stable equilibrium
can be approached as follows. Since the path of the economy is a stochastic
process, the state at any time in the future can be known only up to a
probability distribution. Hence the state space is re-interpreted to be the
collection of all density functions on the original space. Densities can be
CHAPTER 2. EQUILIBRIA OF MARKOVIAN MODELS 20

identified with points on the unit sphere in the space of integrable functions.
(The set of densities coincides with the intersection of the positive cone and
the boundary of the unit sphere.) Thus any stable stochastic equilibrium
can be viewed as a point on this infinite dimensional sphere to which nearby
points are attracted as time evolves.

In this sense, deterministic and stochastic equilibria can be thought of as

differing not conceptually but rather in the nature (in particular, in the di-
mension) of the space in which they are located. Here the above identification
of stochastic equilibria with attractors on the unit sphere of the space of inte-
grable functions is exploited to obtain sufficient conditions for the existence
of stable equilibria in a range of stochastic growth models.

2.2.2 Discrete Dynamical Systems

Consider first a deterministic abstract system characterized at each time t

by a vector of state variables xt taking values in state space X. Evolution is
governed by a first-order difference equation

xt+1 = T xt , T : X → X. (2.3)

The map T encodes the structure of the economic system, which is in turn
determined by the primitives of the model, such as preferences, technology
and market conditions. A realization or trajectory for the system is a se-
quence (T n x) in X generated by iterating the map T on initial state x.3 An
equilibrium is a fixed point of T on X.
3
Here T n x is defined recursively by T n x = T (T n−1 x), T 1 x = T x.
CHAPTER 2. EQUILIBRIA OF MARKOVIAN MODELS 21

More generally, a semidynamical system is a pair (X, T ), where X is a topo-

logical space and T is a continuous mapping of X into itself.4 An equilibrium
or steady state of (X, T ) is a fixed point of T on X (i.e., a point x∗ ∈ X such
that T x∗ = x∗ ). Fixed points are said to be stationary or invariant under
T . Similar terminology also applies to sets. If T A ⊂ A, then A is said to be
invariant under T . For fixed point x∗ of T on X, the stable set ST (x∗ ) of x∗
is that subset of X which is convergent to x∗ under iteration of T :5

ST (x∗ ) = {x ∈ X : T n x → x∗ (n → ∞)}.

The point x∗ is said to be stable, or an attractor, whenever there exists a set

G open in X such that x∗ ∈ G and ST (x∗ ) ⊃ G. In particular:

Definition 2.1. Semidynamical system (X, T ) is said to be globally asymp-

totically stable (or just globally stable) if there exists a unique fixed point x∗
and ST (x∗ ) = X.

Figure 2.1 shows motion induced by iteration of an arbitrary map T , X = R2 .

Continued iteration generates a sequence in the plane.
4
The system is called dynamical if, in addition, the mapping T is invertible with con-
tinuous inverse (i.e. is a homeomorphism).
5
In the interests of generality, X has been defined to be a topological space. However,
the reader may think of it as having sufficient structure to make convergence of sequences
a significant concept. For example, we may imagine throughout that X is a Hausdorff
space.
CHAPTER 2. EQUILIBRIA OF MARKOVIAN MODELS 22

xt+1 = T xt

xt+2
xt

Figure 2.1: Deterministic system in R2 .

2.2.3 Perturbed Dynamical Systems

Suppose now that the system (2.3) is perturbed at each transition from state
xt to state xt+1 by serially uncorrelated, X-valued shock εt with distribution
given by density ψ:

xt+1 = T (xt , εt ), εt ∼ ψ, T : X × X → X. (2.4)

For each fixed xt ∈ X, xt+1 is a random variable with distribution uniquely

determined by the value of xt , the density ψ and the map T . Let the density
of this conditional distribution be p(xt , ·). That is,
Z
p : X × X → R, Prob(xt+1 ∈ B|xt ) = p(xt , xt+1 )dxt+1 , (2.5)
B

where Prob(xt+1 ∈ B|xt ) is the probability that the state vector is in B ⊂ X

at time t + 1 given its current location at xt . Figure 2.2 shows a perturbed
system with additive shock in state space R2 . The circles represent con-
tour lines for the conditional density p(xt , ·). The bold arrows are sample
realizations of the process.

The formulation (2.5) is convenient for calculation of the unconditional dis-

tribution of the state vector at each point in time. Suppose that the uncon-
CHAPTER 2. EQUILIBRIA OF MARKOVIAN MODELS 23

T xt
xt+1 = T xt + εt

T xt+1
xt xt+2

Figure 2.2: The perturbed system.

ditional (marginal) distribution of xt is known, and is given by density ϕt .

In this case,

Z
ϕt+1 (xt+1 ) = p(xt , xt+1 )ϕt (xt )dxt (2.6)

defines the unconditional density of the state at time t + 1. The intuition is

that the integral sums the probability p(xt , xt+1 ) of traveling to xt+1 from xt
for all xt ∈ X, weighted at each point by the likelihood ϕt (xt ) of xt occurring
as the current state. The recursion (2.6) provides a way to calculate the
entire sequence of densities (ϕt ) that represent the marginal distributions for
the stochastic process (xt ) from any initial density ϕ0 (x0 ∼ ϕ0 ).

In analyzing the behavior of the sequence (ϕt ), one possibility is to use stan-
dard techniques from the classical theory of Markov processes.6 However, it
is also possible to frame the same problem as a semidynamical system. The
idea is to re-interpret the state space to be the collection of all densities on
X. Call this set D. The other half of the pair is the operator (call it P ) that
6
See, for example, Shiryaev [52, Chapter 8].
CHAPTER 2. EQUILIBRIA OF MARKOVIAN MODELS 24

associates current-period with next-period densities through the integration

defined in (2.6).

In this notation, (2.6) can be rewritten as

ϕt+1 = P ϕt , P : D → D. (2.7)

But the recursion (2.7) is now in exactly the same formula as the determin-
istic system (2.3), which means that similar techniques can be applied to its
analysis. This translation of the perturbed system (2.4) into a deterministic
map on the space of density functions is called the L1 approach to Markov
processes. Evolution of the economy is characterized by a sequence of densi-
ties generated by iterating P on some initial density ϕ0 . An equilibrium is a
fixed point of the semidynamical system (D, P ). The economy has a unique,
globally stable equilibrium whenever (D, P ) is asymptotically stable in the
sense of Definition 2.1, 21.

These definitions are consistent with those used in previous studies.7 How-
ever, the space of possible states D and hence equilibria has been constructed
to include only those distributions that can be represented by density func-
tions. Thus probability mass cannot be concentrated at a point. In partic-
ular, this means that the model does not include the deterministic system
as a special case; the distribution of the disturbance term ε must be non-
degenerate.
7
The operator P is analogous to T ∗ in Brock and Majumdar [9, Eq. (4.3)], Futia [21,
p. 380], and Stokey et al. [58, Eq. (2), p. 213], and to T in Hopenhayn and Prescott [28,
p. 1392].
CHAPTER 2. EQUILIBRIA OF MARKOVIAN MODELS 25

2.3 Markov Operators

A more formal discussion of perturbed dynamical systems and Markov oper-

ators is now given. To begin, let X be a topological space, let B be the Borel
sets of X and let µ be a fixed σ-finite measure on (X, B). In what follows,
density functions will be defined in terms of their integral with respect to µ.
Integration where the measure is not made explicit is taken with respect to
R
µ; integration using the symbol without subscript is taken over the whole
space X.

Let M be the normed vector lattice of finite signed Borel measures on X

with standard partial order and total variation norm.8 Let P be the elements
ν ∈ M such that ν ≥ 0 and ν(X) = kνk = 1. The subset P will be called the
distributions on X.

Further, let L1 (µ) be the space of µ-integrable real functions on the mea-
surable space (X, B). As usual, L1 (µ) is interpreted as a Banach lattice of
equivalence classes; functions equal off a µ-null set are identified. A density
R
function on X is an element ϕ ∈ L1 (µ) such that ϕ ≥ 0 and ϕ = kϕk = 1.
The set of all density functions is denoted D(µ).

The sets M and L1 (µ) are related in that L1 (µ) is isometrically and lattice
isomorphic to that subset of M (call it Mµ ) which is absolutely continuous
with respect to µ. The relevant isomorphism is Radon-Nikodým (RN) differ-
entiation. In what follows the two sets L1 (µ) and Mµ are identified; we do
8
In other words, M is the class of countably additive real functions on B. For a
definition of the total variation norm, see, for example, Stokey et al. [58, Section 11.3], or
Futia [21, p. 380].
CHAPTER 2. EQUILIBRIA OF MARKOVIAN MODELS 26

not distinguish between them in the presentation.

2.3.1 Operators on Measures

Consider again the economic model defined by (2.4). Random outcomes

(states of nature) are implemented as follows. For some measurable space
(Ω, F), and for some fixed probability measure P on (Ω, F), a state of nature
is selected from Ω according to P, and mapped into X by random variable
ε : Ω → X. The random variable defines a distribution Ψ ∈ P associating
event B ∈ B with the real number P[ε−1 (B)] ∈ [0, 1].

In what follows, by a stochastic dynamic economy on X is meant a pair

(T, Ψ), where T is a map from X × X into X, and, given current state value
xt ∈ X, a shock εt ∈ X is selected independently from Ψ, and the next
period state is realized as

xt+1 = T (xt , εt ).

Dynamics of Markovian models are usually described in terms of transition

kernels [21, Definition 1.1]. Let 1B : X → {0, 1} be the characteristic function
for B ∈ B.9 The economy (T, Ψ) determines a Markov process on X with
transition kernel N,
Z
N : X × B 3 (x, B) 7→ 1B [T (x, z)]Ψ(dz) ∈ [0, 1]. (2.8)

The value N(x, B) should be interpreted as the conditional probability that

the next period state is in Borel set B, given that the current state is equal
to x. A Markov process is fully characterized by its transition kernel.
9
Thus 1B (x) = 1 when x ∈ B and zero otherwise.
CHAPTER 2. EQUILIBRIA OF MARKOVIAN MODELS 27

We seek to derive using N a recursion that links successive marginal distri-

butions of the state variables. Let B be any Borel set, and let νt ∈ P be
the marginal distribution for the random variable xt .10 By the law of total
probability, if νt+1 is the distribution for xt+1 , then
Z
νt+1 (B) = N(x, B)νt (dx). (2.9)

Intuitively, the probability that the state variable is in B next period is the
sum of the probabilities that it travels to B from x across all x ∈ X, weighted
by the probability νt (dx) that x occurs as the current state.

Following Futia [21], Stokey et al. [58] and other authors, the relationship
(2.9) is redefined in terms of operators. Suppose we define an operator
P : M 3 ν 7→ P ν ∈ M by
Z
P ν(B) = N(x, B)ν(dx). (2.10)

It follows from (2.9) and (2.10) that if νt is the distribution for the current
state xt , then P νt is the distribution for the next period state xt+1 .

Repeated iteration of P on a fixed distribution ν is equivalent to moving

forward in time. If P t is defined by P t = P ◦ P t−1 and P 1 = P , and if ν
is the current marginal distribution for the state variable, then P t ν is the
distribution t periods hence.

Evidently P is linear and also positive, in the sense that it maps the positive
cone of M (i.e., the finite measures on B) into itself. In addition, P P ⊂ P,
10
The distribution for the entire stochastic process (xt )t≥0 can be constructed uniquely
from the transition kernel and an initial value x0 [52, Theorem II.9.2]. The real number
νt (B) is the probability that this distribution assigns to the event xt ∈ B and xs ∈ X for
all other s 6= t.
CHAPTER 2. EQUILIBRIA OF MARKOVIAN MODELS 28

because if ν ∈ P, then ν(X) = 1, and hence P ν(X) =

R
N (x, X)ν(dx) =
ν(X) = 1.

At this point we introduce the notion of a Markov operator. Our definition

is a generalization of Hopf [29, Definition 2.1].

Definition 2.2. Let U be an ordered normed space. Let D ⊂ U be the

intersection of the positive cone and the boundary of the unit sphere in U . A
Markov operator on U is any linear operator Q : U → U such that QD ⊂ D.11

It is well-known that every positive linear operator from a Banach lattice

into itself is automatically continuous [1, Theorem 8.6]. Evidently Markov
operators are positive. Thus, when U is a Banach lattice, every Markov
operator is automatically norm-continuous on all of U by positivity. In this
case, it is clear that (D, Q) forms a semidynamical system in the sense of the
definition given in Section 2.2.2, page 20.

By the above discussion, the operator P defined in (2.10) is a Markov opera-

tor on M. The following notion of global stability corresponds to our earlier
definition of stability for semidynamical systems (Definition 2.1, p. 21) as
applied to (P, P ).

Definition 2.3. Let (T, Ψ) be a perturbed dynamical system. Let P be the

corresponding Markov operator. An equilibrium or steady state for (T, Ψ) is
a distribution ν ∗ ∈ P such that P ν ∗ = ν ∗ . An equilibrium ν ∗ is called unique
if there exists no other fixed point of P in the space P, and globally stable if
P t ν → ν ∗ in the total variation norm as t → ∞ for every ν ∈ P.
11
Markov operators in our sense are often called stochastic operators in the literature
on positive operators on AL and AM spaces.
CHAPTER 2. EQUILIBRIA OF MARKOVIAN MODELS 29

This equilibrium concept is entirely standard.12 The real number ν ∗ (B) gives
the probability of being in Borel set B in this period, assuming that the
current distribution is ν ∗ . The value P ν ∗ (B) gives the probability of being
in B next period, given that the current distribution is ν ∗ . The equality of
ν ∗ (B) and P ν ∗ (B) for each B ∈ B therefore implies that the probability law
ν ∗ , once reached, is the law that governs the economy in the next and indeed
all subsequent periods.

Note, however, that stability is defined here in terms of the norm topology.
Existing techniques typically obtain only weak or weak-star stability.13

2.3.2 The L1 Method

The framework introduced so far essentially follows Brock and Majumdar [9],
Futia [21], Stokey et al. [58] and other previous work in economics. However,
in this section we diverge slightly, approaching Markov chains generated by
(2.4) using the L1 method [29]; stochastic processes are studied by analyzing
evolution of density functions which represent the marginal distributions of
current and future state variables. The advantage is that we can exploit a
very useful technique for studying Markov chains in L1 due to Lasota [38].

Embedding the Markov problem in the function space L1 (µ) requires that
12
See, for example, Brock and Mirman [10, p. 492], Futia [21, p. 377] or Stokey, Lucas
and Prescott [58, pp. 317–8].
13
Using coarser topologies is not a free lunch. After all, in every infinite dimensional
normed space U , there exists a sequence of points all with norm one such that the sequence
converges to the zero element in the weak topology induced by the norm dual of U .
CHAPTER 2. EQUILIBRIA OF MARKOVIAN MODELS 30

the set of transition probabilities given in (2.8) can be represented by density

functions. For the moment, let us simply assume that this is the case:

Assumption 2.1. For all x ∈ X, the distribution B 7→ N(x, B) is absolutely

continuous with respect to µ.

In the chapters that follow this assumption will be verified for different con-
ditions on T and Ψ.

Denote the density that represents B 7→ N(x, B) by y 7→ p(x, y). Heuris-

tically, the number p(x, y)dy is the probability of traveling from state x to
state y in one step. In this paper, p is called the density kernel corresponding
to (T, Ψ).

Let Assumption 2.1 hold, implying the existence of p. Using p, the Markov
operator P corresponding to (T, Ψ) can now be reinterpreted as a linear
self-mapping on the function space L1 (µ). Specifically, if h ∈ L1 (µ), then
Z
P h(y) = p(x, y)h(x)dx. (2.11)

The two definitions (2.10) and (2.11) of P are equivalent for the absolutely
continuous measures Mµ ⊂ M when these measures and their RN derivatives
in L1 (µ) are identified. That is, if h ∈ L1 (µ) is the RN derivative of λ ∈ Mµ ,
then P h defined by (2.11) is the RN derivative of P λ defined by (2.10). To
CHAPTER 2. EQUILIBRIA OF MARKOVIAN MODELS 31

see this, pick any B ∈ B. An application of Fubini’s theorem gives

Z Z Z
P h(y)dy = p(x, y)h(x)dxdy
B
ZB Z
= p(x, y)dyh(x)dx
B
Z
= N(x, B)h(x)dx
Z
= N(x, B)λ(dx)

= P λ(B).

Formally, the semidynamical systems defined by (Mµ , P ), where Mµ is the

absolutely continuous measures and P is the Markov operator on measures;
and (L1 (µ), P ), where P is the Markov operator on functions; are topologi-
cally conjugate, in that they commute with the homeomorphism defined by
Radon-Nikodým differentiation. Topologically conjugate dynamical systems
have identical dynamic properties.

Note that P D(µ) ⊂ D(µ), as can be shown directly using Fubini’s theorem.
Thus P is indeed a Markov operator (on L1 ) in the sense of Definition 2.2,
page 28. As before, if ϕ is the current marginal density for the state variable,
then P t ϕ is that of the state t periods hence. Evolution of such a sequence
of density functions is illustrated in Figure 2.3 for the case X = R.

For systems evolving in L1 (µ), we redefine the equilibrium notion slightly.

Definition 2.4. Let (T, Ψ) be a stochastic dynamic economy on X satisfying

Assumption 2.1. Let p be the associated density kernel, and let P be the
Markov operator defined by (2.11). An equilibrium or steady state for (T, Ψ)
is a density ϕ∗ on X such that P ϕ∗ = ϕ∗ . An equilibrium ϕ∗ is called unique
CHAPTER 2. EQUILIBRIA OF MARKOVIAN MODELS 32

P P
ϕt ϕt+1 ϕt+2

Figure 2.3: Evolution of densities on R.

if there exists no other fixed point of P in the space D(µ), and globally stable
if P t ϕ → ϕ∗ in the L1 (µ) metric as t → ∞ for every ϕ ∈ D(µ).

2.4 General Fixed Point Results

Thus equilibria in Markovian models are defined to be fixed points in a

particular semidynamical system (recall that the latter is a continuous self-
mappings on topological space; c.f. Section 2.2.2, p. 20), where in the
present case the self-mapping is the Markov operator P corresponding to the
economy in question, and the domain is either the space of distributions P,
or the density functions D(µ). (The topology of the domain is that generated
by the usual norm on the linear span of these spaces.)
CHAPTER 2. EQUILIBRIA OF MARKOVIAN MODELS 33

This section considers some fixed point and stability results for arbitrary
semidynamical systems. The key notions we will use are Lagrange stability
and strong contractiveness.

It is shown that a semidynamical system which is Lagrange stable has at

least one equilibrium, and that a semidynamical system which is strongly
contractive has at most one equilibrium. Moreover, for a semidynamical sys-
tem with both properties, the unique equilibrium is globally asymptotically
stable.

While the ideas in this section draw heavily on Lasota [38], the approach and
the proofs are presented here for the first time.

2.4.1 Lagrange Stability

Lagrange stability has been used extensively in the study of nonlinear differ-
ential equations and iterated function systems. Lagrange’s original stability
work was on the N -body problem of planetary motion. He showed that a
first-order approximation of the system does not grow without bounds. The
concept of Lagrange stability retains this meaning.

Recall that a subset A of topological space U is called precompact if A has

compact closure.

Definition 2.5. Let U be a topological space and let T be a continuous

self-mapping on U . Semidynamical system (U, T ) is called Lagrange stable
if the trajectory of x (i.e., the set of points T n x, n ∈ N) is precompact for
every x ∈ U .
CHAPTER 2. EQUILIBRIA OF MARKOVIAN MODELS 34

Remark 2.1. In finite dimensional space, precompactness is equivalent to

boundedness by the Heine-Borel theorem. Thus for such a space Lagrange
stability corresponds to the idea that none of the possible trajectories for the
state variables grow without bounds.

A fixed point result for Lagrange stable systems is now stated. An alter-
native proof based on spectral decomposition can be found in Lasota and
Mackey [40, Proposition 5.4.1], although the notation and formulation is
somewhat different. Here a new proof is offered based on an infinite dimen-
sional Brouwer (i.e., Schauder) fixed point theorem.14

Theorem 2.1. Let V be a Banach space, and let U be a nonempty, closed

and convex subset of V . Let T be a linear and continuous self-mapping on V
such that T U ⊂ U . If (U, T ) is Lagrange stable, then T has a fixed point in
U.

Proof. Take any x ∈ U . Define γ(x) to be the set of all points T n x, n ∈ N; let
γ̂(x) be its convex hull; and let cl(γ̂(x)) be the closure of the latter. Since the
convex hull of a precompact set in V is again precompact [1, Theorem 5.20],
it follows that γ̂(x) is precompact. Since the closure of a precompact set is
compact, cl(γ̂(x)) must be compact. Using the linearity of T , if a ∈ γ̂(x),
then evidently T a is again in γ̂(x), or T γ̂(x) ⊂ γ̂(x). But then T cl(γ̂(x)) ⊂
cl(γ̂(x)). The reason is that if A is any set with T A ⊂ A and A0 is the
closure of A, then T A0 ⊂ A0 , because a0 ∈ A0 implies the existence of a
14
Schauder: Every self-mapping invariant on a nonempty compact convex subset of a
locally convex Hausdorff space has at least one fixed point in that set.
CHAPTER 2. EQUILIBRIA OF MARKOVIAN MODELS 35

sequence (an ) ⊂ A, an → a0 , whence T a0 = T lim an = lim T an , which, as the

limit of a sequence in A, must again be in A0 . Hence T A0 ⊂ A0 .

Thus T is invariant on nonempty convex compact set cl(γ̂(x)). But then T

has a fixed point in cl(γ̂(x)) [32, Theorem 4.3.10]. Finally, since cl(γ̂(x)) ⊂ U
by the assumption that U is closed and convex, the fixed point must also be
in U .

Remark 2.2. Note that Lagrange stability is quite a bit stronger than we
actually require in the proof. Lagrange stability has been used here because
of its relationship with subsequent results.

2.4.2 Strongly Contractive Systems

Next strong contractiveness and its relationship to Lagrange stability is dis-

cussed.

In many fields of economics, Banach’s contraction principle is used to locate

equilibria and solve dynamic programs.15 Let (U, %) be a metric space and
let T : U → U . The map T is called a Banach contraction if there exists an
α < 1 such that
%(T x, T y) ≤ α%(x, y), ∀x, y ∈ U. (2.12)

Banach’s contraction principle is equivalent to the statement that if U is

complete and T satisfies (2.12), then semidynamical system (U, T ) is asymp-
totically stable [32, Theorem 4.1.1].
15
See, for example, Stokey et al. [58, Lemma 11.11 and Section 17.2].
CHAPTER 2. EQUILIBRIA OF MARKOVIAN MODELS 36

Unfortunately, for the semidynamical systems generated by stochastic growth

models, in general (2.12) either does not hold or is difficult to verify. In
contrast, the slightly weaker condition (2.14) below will prove both useful
and easy to verify in applications.

Definition 2.6. Let topological space U be metrizable under distance %, and

let T be a continuous self-mapping on U . Semidynamical system (U, T ) is
called contractive (or nonexpansive) if

%(T x, T y) ≤ %(x, y), ∀x, y ∈ U. (2.13)

The system is called strongly contractive if, in addition,

%(T x, T y) < %(x, y), ∀x, y ∈ U, x 6= y. (2.14)

Evidently (2.12) =⇒ (2.14) =⇒ (2.13). Like contractiveness in the sense of

Banach, strong contractiveness implies uniqueness of equilibrium. (Suppose
otherwise. In particular, let distinct points x and y be stationary under T .
Then both d(x, y) = d(T x, T y) and d(T x, T y) < d(x, y). Contradiction.)
However, strong contractiveness does not guarantee existence.16 Neverthe-
less, existence and stability can be obtained if strong contractiveness is sup-
plemented by compactness of U :

Lemma 2.1. Let (U, T ) be a semidynamical system, where U is a metrizable

space. If (U, T ) is strongly contractive and U is compact, then (U, T ) is
globally asymptotically stable.
16
For example, consider U = R+ , T : x 7→ x + e−x .
CHAPTER 2. EQUILIBRIA OF MARKOVIAN MODELS 37

Remark 2.3. Strictness of the inequality in (2.14) is necessary for both

uniqueness and existence. For example, existence fails if U is the boundary
of the unit sphere in R2 , and T x = −x.

Proof. This result is already known [32, Theorem 4.1.6, Corollary 1], but we
include a proof for completeness. The stability part of the proof may be new.

Uniqueness has already been proved above. Consider the problem of exis-
tence. Define r : U → R by r(x) = %(T x, x). Evidently r is continuous. Since
U is compact, r has a minimizer x∗ . But then T x∗ = x∗ , because otherwise
T x∗ minimizes r on U .

Regarding stability, pick any x ∈ U , and define αn = %(T n x, x∗ ). Since (αn )

is monotone decreasing and nonnegative it has a limit α. If α = 0 then
we are done. Suppose otherwise. By compactness, (T n x) has a convergent
subsequence T nk x → x̄ ∈ U . Evidently %(x̄, x∗ ) = α > 0, so x̄ and x∗ are
distinct. But then

%(T x̄, T x∗ ) = %(T lim T nk x, x∗ )

k

= lim %(T T nk x, x∗ )
k

= lim αnk +1 = α,
k

which contradicts (2.14). This argument proves convergence to the fixed

point.

We have proved that contractiveness of the operator and compactness of the

space together imply existence, uniqueness and global stability of equilibrium.
In the problems generated by Markov operators, however, compactness of
CHAPTER 2. EQUILIBRIA OF MARKOVIAN MODELS 38

the underlying space will not in general be satisfied. An alternative (weaker)

criterion is Lagrange stability of the operator. In particular,

Theorem 2.2. Let X be a metrizable space, let U be a nonempty closed

subset of X and let T : X → X be a continuous function invariant on U .
If (U, T ) is both Lagrange stable and strongly contractive, then it is globally
asymptotically stable.

Proof. Fix x ∈ U . Define Γ(x) to be the closure of {T n x : n ∈ N}.

Since (U, T ) is Lagrange stable, Γ(x) is a compact subset of X. Moreover,
T Γ(x) ⊂ Γ(x). Therefore (Γ(x), T ) is itself a strongly contractive semidy-
namical system on a compact set, and, by Lemma 2.1, has a unique fixed
point x∗ ∈ Γ(x) with T n x → x∗ . The point x∗ is in U because U is closed
and hence Γ(x) ⊂ U . Moreover, (U, T ) has at most one fixed point by strong
contractiveness. Therefore x∗ does not depend on x. The result follows.

2.5 Applications to Markov Operators

The objective of this section is to identify when Markov operators generated

by stochastic dynamic economies might satisfy the conditions of the above
fixed point theorems. These insights are entirely due to Lasota [38].

2.5.1 Contractive Markov Operators

Regarding contractiveness of Markov operators,

CHAPTER 2. EQUILIBRIA OF MARKOVIAN MODELS 39

Lemma 2.2. If P is a Markov operator on L1 (µ), then (L1 (µ), P ) is con-

tractive in the sense of (2.13).

Proof. Fix f ∈ L1 (µ). Define f + = max(f, 0) and f − = max(−f, 0). By

linearity and positivity,

|P f (x)| = |P f + (x) − P f − (x)| ≤ P f + (x) + P f − (x) = P |f (x)|.

Integration obtains
Z Z
kP f k = |P f |dµ ≤ P |f |dµ = kf k.

An application of linearity yields (2.13).

An important sufficient condition for strong contractiveness of Markov oper-

ators on D(µ)—which has been emphasized by A. Lasota—is as follows.

Lemma 2.3. Let a perturbed dynamical system (T, Ψ) satisfying Assumption

2.1 be given, let p be the associated density kernel, and let P : L1 (µ) → L1 (µ)
be the Markov operator defined from p by (2.11). If p > 0 on X × X, then P
is strongly contractive on D(µ) with respect to the metric induced by the L1
norm.

Proof. Pick any two densities ϕ 6= ϕ0 . Evidently the function ϕ − ϕ0 is both

strictly positive on a set of positive measure and strictly negative on a set
of positive measure. Pick any y ∈ X. Since p(x, y) > 0, it follows that
x 7→ p(x, y)[ϕ(x) − ϕ0 (x)] is also strictly positive on a set of positive measure
and strictly negative on a set of positive measure. Therefore, by the strict
CHAPTER 2. EQUILIBRIA OF MARKOVIAN MODELS 40

triangle inequality,

kP ϕ − P ϕ0 k = kP (ϕ − ϕ0 )k
Z Z
= p(x, y)[ϕ(x) − ϕ0 (x)]dx dy
Z Z
< p(x, y)[ϕ(x) − ϕ0 (x)] dx dy
Z Z
= p(x, y)|ϕ(x) − ϕ0 (x)|dx dy
Z Z
= p(x, y)dy|ϕ(x) − ϕ0 (x)|dx

= kϕ − ϕ0 k,

which proves (2.14).

2.5.2 Lagrange Stability of Markov Operators

The condition for strong contractiveness of P on D(µ) is therefore relatively

straightforward. The other half of our primary stability condition, Theorem
2.2, is Lagrange stability. Here we discuss some strategies for establishing
this property.

Lasota [38, Theorem 4.1] has made the important insight that in the case of
integral Markov operators such as (2.11), it is sufficient to prove that {P t ϕ :
t ≥ 0} is weakly precompact for every ϕ ∈ D(µ). The reason is that integral
Markov operators map weakly precompact subsets of L1 (µ) into strongly
precompact subsets.17 Therefore, if {P t ϕ : t ≥ 0} is weakly precompact,
then {P t ϕ : t ≥ 1} is strongly precompact. But then {P t ϕ : t ≥ 0} is also
strongly precompact.
17
This result appears to be due to A. Krasnosielski.
CHAPTER 2. EQUILIBRIA OF MARKOVIAN MODELS 41

In fact Lasota [38, Proposition 3.4] has used a Cantor diagonal argument to
show that weak precompactness of {P t ϕ : t ≥ 0} need only be established
for a collection of ϕ such that the norm closure of the collection contains
D(µ).

Therefore,

Proposition 2.1 (Lasota). If P : D(µ) → D(µ) is an integral Markov op-

erator such as (2.11), then the semidynamical system (D(µ), P ) is Lagrange
stable if and only if the set of functions {P t ϕ : t ≥ 0} is weakly precompact
for every ϕ ∈ D, D a norm-dense subset of D(µ).

To sum up our discussion so far,

Proposition 2.2. Let (T, Ψ) be a stochastic dynamic economy evolving on

state space X. Suppose that Assumption 2.1 holds, and let p be the associated
density kernel. Suppose further that p > 0 on X × X. If there exists a set
D ⊂ D(µ) such that D is norm-dense in D(µ) and {P t ϕ : t ≥ 0} is weakly
precompact for each ϕ ∈ D, then (T, Ψ) has a unique and globally stable
equilibrium in the sense of Definition 2.4.

Proof. The proof is immediate from Theorem 2.2, Lemma 2.3 and Proposi-
tion 2.1.

2.5.3 Weak Precompactness

Before closing, we briefly recall a well-known characterization of weak pre-

compactness in L1 due to Dunford and Pettis [15, Theorem 3.2.1]. Given any
CHAPTER 2. EQUILIBRIA OF MARKOVIAN MODELS 42

σ-finite measure space (X, B, µ), a collection of functions {fα }α∈Λ in L1 (µ)
is a weakly precompact subset of that space if and only if it is norm bounded
and the following two conditions hold:

(i) ∀ε > 0, ∃ δ > 0 such that if A ∈ B and µ(A) < δ, then

Z
|fα |dµ < ε, ∀α ∈ Λ; and
A

(ii) ∀ε > 0, ∃ G ∈ B such that µ(G) < ∞ and

Z
|fα |dµ < ε, ∀α ∈ Λ.
X\G

Evidently it is sufficient to verify that these conditions are satisfied for all
but a finite number of the collection {fα }.
Chapter 3

Optimal Growth with

Unbounded Shock

3.1 Introduction

This first essay studies equilibria in the stochastic optimal growth economy
of Brock and Mirman [10], but without the assumption that the shock which
perturbs production is realized within a bounded interval. It provides suffi-
cient conditions for existence, uniqueness and stability of equilibria in terms
of the primitives of the one-sector model, namely the utility function u, the
per capita production function f and the distribution ψ of the disturbance
term ε.

The original work of Brock and Mirman extends the deterministic optimal
growth problem of Ramsey [50], Cass [11], Koopmans [34] and others to a
stochastic setting. With regard to equilibria, they show that the existence,

43
CHAPTER 3. OPTIMAL GROWTH WITH UNBOUNDED SHOCK 44

uniqueness and stability results of the deterministic case are also realized
in a stochastic model under similar assumptions on preferences and produc-
tion technology. In their analysis, the productivity shock is restricted to a
bounded interval of the real line.1

The problem of characterizing equilibria and long-run behavior in Brock-

Mirman economies with bounded shock has subsequently been studied by
Mirman [45], Mirman and Zilcha [46], Brock and Majumdar [9], Razin and
Yahav [51], Donaldson and Mehra [13], Majumdar and Zilcha [43], Stokey et
al. [58], Hopenhayn and Prescott [28] and Amir [4]. The analogous problem
for the overlapping generations model with bounded shock has been studied
by Laitner [36] and Wang [62]. The related question of ergodicity in moments
for the Solow-Swan model with a shock that is unbounded above but cannot
be arbitrarily small is investigated in Binder and Pesaran [8]. Evstigneev and
Flåm [17] and Amir and Evstigneev [5] investigate the asymptotic distribu-
tions of aggregate rewards accumulated along equilibrium and optimal paths.
More general studies of stochastic equilibria in economics include Futia [21]
and Duffie et al. [14].

Stochastic growth with unbounded shock is treated in Mirman [44], who

provides an existence result and proves that the equilibrium measure is not
concentrated at zero. His treatment leaves room for further analysis, however,
as his sufficient conditions pertain to a class of consumption policies that
may or may not be optimal [44, A1–A3, p. 275]. In other words, the savings
1
Such a shock is said to have compact support. For simplicity these shocks are re-
ferred to as “bounded”. Shocks where no restrictions are placed on the support are called
“unbounded”.
CHAPTER 3. OPTIMAL GROWTH WITH UNBOUNDED SHOCK 45

rate is exogenously given, and the conditions are not stated in terms of the
primitives u, f and ψ. Further, the problems of uniqueness and stability
are not treated. In the present essay, conditions for existence, uniqueness
and stability are obtained in terms of the triple (u, f, ψ) and the restrictions
imposed by optimizing behavior.

The mathematical techniques used in the essay are based on recent innova-
tions in the theory of perturbed dynamical systems. The two key concepts
are Lagrange stability and strong contractiveness.

In addition to identifying and characterizing equilibria in Brock-Mirman

economies with unbounded shock, the essay also makes the following con-
tributions. First, the L1 approach is introduced to stochastic growth theory.
Second, the notions of strong contractiveness and Lagrange stability are de-
veloped in the context of stochastic optimal growth.

The structure of the essay is as follows. Section 3.2 formulates the stochastic
optimal growth problem. Section 3.3 states the main result. The proof is
then developed over Sections 3.4–3.5.

3.2 Formulation of the Problem

This section contains a formulation of the stochastic optimal growth problem

studied by Brock and Mirman [10]. The symbols R+ and R++ denote the
nonnegative and positive reals respectively. Given any topological space X,
B(X) denotes the Borel sets of X. All sets of real numbers introduced in the
essay are assumed to be Borel sets, and all real functions are Borel functions.
CHAPTER 3. OPTIMAL GROWTH WITH UNBOUNDED SHOCK 46

Lebesgue measure is denoted by µ. Unless otherwise stated, integration is

with respect to µ.

The accumulation problem evolves as follows. At the start of period t the

(representative) agent receives income xt . In response a level of consumption
ct ≤ xt is chosen, yielding current utility u(ct ). The remainder is invested
in production, returning in the following period output xt+1 = f (xt − ct )εt .
Here f is the production function and ε is a nonnegative random variable.2
The process then repeats.

3.2.1 Assumptions

The functions u and f satisfy the usual assumptions.

Assumption 3.1. The production function f : R+ → R+ is zero at zero,

strictly increasing, strictly concave, differentiable and satisfies the Inada con-
ditions limx↓0 f 0 (x) = ∞ and limx↑∞ f 0 (x) = 0.

Assumption 3.2. The utility function u : R+ → R is strictly increasing,

strictly concave, differentiable and satisfies limx↓0 u0 (x) = ∞.

The shock is permitted to be unbounded.

Assumption 3.3. The shocks to production are uncorrelated and identically

distributed. The distribution of ε is represented by density ψ. The shock
2
Following Stokey et al. [58] and Hopenhayn and Prescott [28], it is assumed that the
disturbance term ε is multiplicative. Brock and Mirman use the more general formulation
xt+1 = f (xt − ct , εt ). See Amir [4] for an even more general technology.
CHAPTER 3. OPTIMAL GROWTH WITH UNBOUNDED SHOCK 47

has finite mean E(ε). In addition, ε satisfies E(1/ε) < 1. The shock is less
R1
than one with positive probability, i.e. 0 ψ(x)dx 6= 0.

3.2.2 Technology

The conditional density for next-period output given income x and consump-
tion c is, by a change of variable argument,
 
y 1
y 7→ ψ . (3.1)
f (x − c) f (x − c)

Given that f (0) = 0, (3.1) is not defined when consumption is equal to in-
come. In this case (when c = x), next-period income is zero with probability
one. Such a probability cannot be represented by a density. Consequently,
the fully specified technology associating savings x − c to next-period income
will be defined by probability B 7→ Q(x, c; B), where
Z  
y 1
Q(x, c; B) = ψ dy.
B f (x − c) f (x − c)

when c < x, and by the probability concentrated at zero when c = x. Thus

Q(x, c; B) is the probability that next-period output is in B given that cur-
rent income is x and consumption is c ∈ [0, x].

3.2.3 The Optimal Policy

Future utility is discounted geometrically at rate β ∈ (0, 1). The agent selects
a sequence (ct ) to solve " ∞
#
X
max E β t u(ct ) (3.2)
t=0
CHAPTER 3. OPTIMAL GROWTH WITH UNBOUNDED SHOCK 48

subject to the feasibility constraint 0 ≤ ct+1 ≤ f (xt − ct )εt .

The meaning of the expectations operator in (3.2) is not immediately clear. A

more formal statement of the problem is that the agent seeks a control policy
g : R+ 3 xt 7→ ct ∈ R+ that is feasible (i.e., g(x) ∈ [0, x]) and maximizes
v(x, g), where " ∞
#
X
v(x, g) = Egx β t u(g(xt )) .
t=0

Here Egx signals integration with respect to the (well-defined and unique)
Markovian distribution over infinite-dimensional sequence space RN+ gener-
ated by Markov transition kernel Q(x, g(x); dy).3

The value function V for the problem is defined at x as the supremum of

v(x, g) over the set of all feasible policies. A feasible policy g ∗ is called optimal
if v(x, g ∗ ) = V (x) for all x.

The following results are well-known.

Theorem 3.1. Let u, f and ψ satisfy Assumptions 3.1–3.3. The following

results hold.

1. The value function V is finite and satisfies the Bellman equation

 Z 
V (x) = max u(c) + β V (y)Q(x, c; dy) .
0≤c≤x

2. There exists a unique optimal policy g and

Z
V (x) = u(g(x)) + β V (y)Q(x, g(x); dy).

3
See, for example, Hernández-Lerma and Lasserre [26].
CHAPTER 3. OPTIMAL GROWTH WITH UNBOUNDED SHOCK 49

3. The value function is nondecreasing, concave and differentiable with

V 0 (x) = u0 (g(x)).

4. If g is an optimal policy, then 0 < g(x) < x, ∀x > 0, and both x 7→ g(x)
and x 7→ x − g(x) are nondecreasing (savings and consumption both
increase with income).

Proof. See, for example, Mirman and Zilcha [46, p. 331–2]. (For a formal dis-
cussion of Markov control programs with unbounded reward see Hernández-
Lerma and Lasserre [27, Chapter 8].) Here parts 1–3 of the theorem =⇒ part
4.

Substitution of the optimal control into the production relation yields the
closed-loop law of motion

xt+1 = f (xt − g(xt ))εt . (3.3)

3.3 Statement of Results

It is now possible to state our main result, which gives sufficient conditions
for existence, uniqueness and stability of equilibria in the stochastic growth
model of the previous section. It shows that the original results of Brock
and Mirman also hold for many of the standard (unbounded) shocks used in
mathematical statistics.

Theorem 3.2. Let u, f and ψ satisfy Assumptions 3.1–3.3. The following

statements are true.
CHAPTER 3. OPTIMAL GROWTH WITH UNBOUNDED SHOCK 50

1. The economy (u, f, ψ) has at least one (nonzero) equilibrium.

2. If, in addition, ψ is everywhere positive, then the equilibrium is unique

and globally stable.

The proof is developed in stages through the remaining sections. The ap-
proach is to represent the economy (u, f, ψ) as a semidynamical system and
then apply the concepts of Lagrange stability and strong contractiveness.

In Section 3.4 it is shown that (u, f, ψ) can be represented as a semidynamical

system formed by a Markov operator on the space of density functions. If
it can be established under Assumptions 3.1–3.3 that this semidynamical
system generated by (u, f, ψ) is Lagrange stable, then Theorem 2.1, page
34, can be used to demonstrate the existence of at least one equilibrium. If,
in addition, it can be shown that positivity of ψ in part 2 of Theorem 3.2
implies strong contractiveness, then by Theorem 2.2, page 38, the system
is also asymptotically stable, which is to say that there exists a unique and
globally stable equilibrium.4 Lagrange stability and strong contractiveness
are established in Section 3.5, completing the proof of the Theorem 3.2.

The proof of Lagrange stability (Proposition 3.1, p. 57) constitutes the main
technical contribution of the essay. As expected, the Inada conditions and
the concavity of the program are crucial to the proof.
4
See Definition 2.4, page 31.
CHAPTER 3. OPTIMAL GROWTH WITH UNBOUNDED SHOCK 51

3.3.1 Examples

Let f and u satisfy Assumptions 3.1 and 3.2 respectively, and let the density
ψ of ε be lognormal. In other words, ln ε is normally distributed with mean
m and variance σ 2 , σ > 0. Since E(1/ε) = exp(σ 2 /2 − m), E(1/ε) < 1 if
σ 2 /2 < m. In this case, all of the components of Assumption 3.3 are also
satisfied, and (u, f, ψ) has at least one equilibrium. In addition, the density
function is everywhere positive. It follows from part 2 of the theorem that
the equilibrium is unique and globally stable.

In fact the same result holds for any lognormal shock. To see this, let m and
σ be arbitrary, σ > 0, and let θ be a constant strictly larger than E(1/ε).
If ε∗ = θε, f ∗ = (1/θ)f and ψ ∗ is the distribution of ε∗ , then (u, f ∗ , ψ ∗ )
satisfies Assumptions 3.1–3.3 and all of the conditions of the theorem. Hence
(u, f ∗ , ψ ∗ ) has a unique, globally stable equilibrium. But

1
f ∗ (x − c)ε∗ = f (x − c)θε = f (x − c)ε,
θ

so (u, f, ψ) and (u, f ∗ , ψ ∗ ) are identical.5 It follows that (u, f, ψ) also has a
unique, globally stable equilibrium.

3.3.2 Remarks

The restriction
Z ∞
1
E(1/ε) = ψ(x)dx < 1 (3.4)
0 x
5
More formally, both economies have the same density kernel.
CHAPTER 3. OPTIMAL GROWTH WITH UNBOUNDED SHOCK 52

used in the theorem has a simple interpretation. Previous work has assumed
that ε is realized in a compact interval [a, b], 0 < a ≤ b < ∞. Here, in con-
trast, the shock may be arbitrarily large or arbitrarily close to zero. Equation
(3.4) implies that ε is “unlikely” to be very close to zero, or, in other words,
that the left-hand tail of the density ψ is relatively small. To see this, define
for nonnegative summable function V and for density h on R++ the (possibly
infinite) number
Z
E(V |h) = V (x)h(x)dx, (3.5)
R++

as well as the set Ga = {x ∈ R++ : V (x) < a}. Evidently,

Z
E(V |h) ≥ V (x)h(x)dx,
R++ \Ga

implying
E(V |h)
Z
h(x)dx ≤ . (3.6)
R++ \Ga a
This is in fact a version of Chebychev’s inequality. Substituting I −1 : x 7→ x−1
for V , ψ for h and 1/r for a gives
Z r
ψ(x)dx ≤ rE(1/ε).
0

Thus (3.4) is a restriction on the left-hand tail of ψ.

It has also been assumed that E(ε) is finite. This is a restriction on the
right-hand tail. To see this, substitute I : x 7→ x for V and ψ for h to obtain
Z ∞
E(ε)
ψ(x)dx ≤ . (3.7)
a a

These restrictions on the tails of ψ can be thought of as a generalization of

the assumption that ψ is zero below a and above b made in previous studies.
CHAPTER 3. OPTIMAL GROWTH WITH UNBOUNDED SHOCK 53

(As a caveat to the claim that the restrictions on ψ are a generalization

of boundedness, recall that in this paper—in contrast to the majority of
previous work—the shock must be non-degenerate and representable by a
density function.)

The assumption on positivity of ψ in part 2 of the theorem is akin to the

“communication” assumptions used in traditional Markov chain theory [52,
Chapter 8].

3.4 Semidynamical Systems

In this section it is shown that (u, f, ψ) can be interpreted as a semidynamical

system defined by a Markov operator on the space of density functions.

Following the notation in Chapter 2, let L1 (µ) be the Banach lattice of µ-

integrable real functions on R++ , and let D(µ) be the set of h ∈ L1 (µ) such
that h ≥ 0 and khk = 1.

The density kernel, Markov operator and semidynamical system associated

with the Brock-Mirman economy (u, f, ψ) are derived from the law of motion
(3.3).

By a change of variable argument, the conditional density for next-period

income given that current income equals x is
 
y 1
y 7→ ψ . (3.8)
f (x − g(x)) f (x − g(x))

As a function of both x and y, (3.8) defines a density kernel for measure

space (R++ , B(R++ ), µ) in the sense of Chapter 2. Denote by Q the Markov
CHAPTER 3. OPTIMAL GROWTH WITH UNBOUNDED SHOCK 54

operator associated with (3.8) by (2.11). The semidynamical system for the
Brock-Mirman process is then (D(µ), Q). If initial income x0 is distributed
according to ϕ0 , then time t income is distributed according to Qt ϕ0 .

A plot of (3.8) is shown in Figure 3.1 for the parameterization f : x 7→ x1/2 ,

u : x 7→ ln x, ε lognormal. The origin is the corner of the graph furthest from
the viewer. For each xt , a density function runs parallel to the xt+1 axis. The
density governs the likelihood that income per head takes values along that
axis, given that the current state is xt .6

3.5 Proof of the Main Theorem

The proof of Theorem 3.2 proceeds as follows. After some preliminary results,
Lagrange stability of the semidynamical system associated with (u, f, ψ) is
established. Next, strong contractiveness of the economy is established using
the additional hypothesis of positivity of ψ. The proof is then completed in
Section 3.5.2.

3.5.1 Proof of Lagrange Stability

This first lemma is required in the proof of Lagrange stability of the semidy-
namical system (D(µ), Q) associated with (u, f, ψ).
6
For a kernel estimated nonparametrically from actual growth data see Quah [48, Fig-
ures 5 and 6].
CHAPTER 3. OPTIMAL GROWTH WITH UNBOUNDED SHOCK 55

xt
xt+1

Figure 3.1: Density kernel (3.8).

CHAPTER 3. OPTIMAL GROWTH WITH UNBOUNDED SHOCK 56

Lemma 3.1. Let (u, f, ψ) satisfy Assumptions 3.1–3.3. If g is an optimal

policy, then there exists an x0 > 0 such that

f (x − g(x)) ≥ x whenever x ∈ (0, x0 ].

Proof. The first order condition of Theorem 3.1, part (1) is

Z ∞
0
u (g(x)) = β V 0 (f (x − g(x))z)f 0 (x − g(x))zψ(z)dz.
0

Using the envelope relation of Theorem 3.1, part (3) obtains

Z ∞
0
V (x) = β V 0 (f (x − g(x))z)f 0 (x − g(x))zψ(z)dz
Z0 1
≥ β V 0 (f (x − g(x))z)f 0 (x − g(x))zψ(z)dz
Z0 1
≥ β V 0 (f (x − g(x)))f 0 (x − g(x))zψ(z)dz,
0

where the first inequality follows from the fact that V is nondecreasing and
the second from the fact that V is concave. Thus,
Z 1
0 0 0
V (x) ≥ V (f (x − g(x)))f (x − g(x))M, M = β zψ(z)dz.
0

The constant M is positive by Assumption 3.3. Assumption 3.1 and the

monotonicity of x 7→ x − g(x) then imply the existence of an x0 > 0 such
that f 0 (x − g(x))M ≥ 1 whenever x ∈ (0, x0 ]. Therefore,

V 0 (x) ≥ V 0 (f (x − g(x))) on (0, x0 ].

The result now follows from the concavity of V .

The proof of the following proposition draws extensively on methods devel-

oped by Lasota [38] and Horbacz [30].
CHAPTER 3. OPTIMAL GROWTH WITH UNBOUNDED SHOCK 57

Proposition 3.1. If (u, f, ψ) satisfies Assumptions 3.1–3.3, then the asso-

ciated semidynamical system (D(µ), Q) is Lagrange stable.

Proof. By Proposition 2.1 on page 41, it is sufficient to find a D ⊂ D(µ) such

that D is dense in D(µ) and the set {Qn h} is weakly precompact for every
h in D.

Let D be the collection of densities h that satisfy

Z ∞ Z ∞
1
xh(x)dx < ∞ and h(x)dx < ∞. (3.9)
0 0 x

We claim that D has the desired properties. To see that D is dense in the
densities, fix ϕ ∈ D(µ) and define h0k = 1(1/k,k) ϕ. Since kh0k k ↑ 1 by the
monotone convergence theorem, it follows that for some K ∈ N, kh0k k > 0
whenever k ≥ K. For all such k define

hk = kh0k k−1 h0k .

It can be established that hk satisfies (3.9) for each k. In addition, hk is a

density by construction, and hk → ϕ pointwise. But then hk → ϕ in the L1
norm by Scheffé’s lemma [59, Proposition 4.5.14]. Thus D is dense in D(µ).

It remains to show that that if h ∈ D then {Qn h}n≥1 is weakly precompact.

Fix arbitrary h ∈ D. It is sufficient to establish precompactness of {Qn h}n≥N
for some fixed N ∈ N, because appending a finite number of elements to a
(weakly) precompact set does not alter the property of (weak) precompact-
ness. We now show that {Qn h}n≥N is weakly precompact for some N ∈ N
by verifying the Dunford-Pettis conditions given in Section 2.5.3, page 41.
CHAPTER 3. OPTIMAL GROWTH WITH UNBOUNDED SHOCK 58

Boundedness of the collection is satisfied because kQn hk = khk = 1 for all n

by the positive isometry property of Markov operators (Definition 2.2). We
now show that conditions (i) and (ii) of the Dunford-Pettis condition also
hold (c.f. p. 41).

For notational simplicity define q(x) = f (x − g(x)). Use will be made of the
fact that
1 1
E(1/ε) ≤ γ + C (3.10)
q(x) x
for all positive x, where γ and C are nonnegative constants, γ < 1.

To verify (3.10), recall that ∃x0 > 0 such that q(x) ≥ x when x ≤ x0 by
Lemma 3.1. Choose any γ such that E(1/ε) < γ < 1. Then

1 1
E(1/ε) ≤ γ , ∀x ≤ x0 . (3.11)
q(x) x

Moreover, on [x0 , ∞), monotonicity of f and x 7→ x − g(x) implies that

q(x) ≥ q(x0 ), or
1 1
E(1/ε) ≤ E(1/ε) = C. (3.12)
q(x) q(x0 )
Together, (3.11) and (3.12) imply (3.10).

Let I −1 be the map x 7→ x−1 . Applying in succession (3.5), (3.8), Fubini’s

CHAPTER 3. OPTIMAL GROWTH WITH UNBOUNDED SHOCK 59

theorem, a change of variable argument and (3.10),

Z ∞
−1 n 1 n
E(I |Q h) = Q h(y)dy
0 y
Z ∞ Z ∞   
1 y 1 n−1
= ψ Q h(x)dx dy
0 y 0 q(x) q(x)
Z ∞ Z ∞   
y 1 1
= ψ dy Qn−1 h(x)dx
q(x) q(x) y
Z0 ∞ 0
1
= E(1/ε)Qn−1 h(x)dx
q(x)
Z0 ∞
1
≤ [γ + C]Qn−1 h(x)dx
0 x
= γE(I −1 |Qn−1 h) + C.

Repeating the argument n times,

C
E(I −1 |Qn h) ≤ γ n E(I −1 |h) + ,
1−γ

or, using finiteness of E(I −1 |h),

C
E(I −1 |Qn h) ≤ 1 +
1−γ

when n ≥ K, K suitably large.

An application of the Chebychev argument (3.6) gives

Z r
Qn h(x)dx ≤ rE(I −1 |Qn h)
0

for any positive r. Therefore,

Z r  
n C
Q h(x)dx ≤ r 1 + , n ≥ K. (3.13)
0 1−γ

Now fix any ε > 0. According to the Dunford-Pettis condition part (i), we
require a δ > 0 and a K ∈ N such that n ≥ K implies
Z
Qn h(x)dx < ε
A
CHAPTER 3. OPTIMAL GROWTH WITH UNBOUNDED SHOCK 60

whenever µ(A) ≤ δ. For this purpose, consider the decomposition

Z Z Z
n n
Q h(x)dx = Q h(x)dx + Qn h(x)dx. (3.14)
A A∩(0,r) A∩(r,∞)

Using (3.13) gives

Z Z r
n ε
Q h(x)dx ≤ Qn h(x)dx < . (3.15)
A∩(0,r) 0 2

when r > 0 is chosen to be sufficiently small and n ≥ K.

Take r as given and consider the second term in (3.14).

Z Z  Z ∞  
n y 1 n−1
Q h(x)dx = ψ Q h(x)dx dy
A∩(r,∞) A∩(r,∞) 0 q(x) q(x)
Z ∞ Z   
y 1
= ψ dy Qn−1 h(x)dx.
0 A∩(r,∞) q(x) q(x)
Z ∞ "Z #
= ψ(z)dz Qn−1 h(x)dx.
A∩(r,∞)
0 q(x)

The term in brackets can be written as

Z ∞
G(x) = 1 A (z)ψ(z)dz.
r q(x)
q(x)

By (3.7) it is possible to choose α > 0 so small that

Z ∞
ε
ψ(z)dz < .
r
q(α)
2

Evidently,
ε
G(x) < whenever x ≤ α.
2
Now consider the case where x > α. Select δ 0 > 0 such that
Z
0 ε
µ(B) < δ =⇒ ψ(z)dz < .
B 2
CHAPTER 3. OPTIMAL GROWTH WITH UNBOUNDED SHOCK 61

Existence of such a δ 0 follows from absolute continuity of the integral measure

with respect to µ [59, Ex. 2.8.15]. Define δ = q(α)δ 0 . Then x > α and
µ(A) < δ implies
Z
ε
G(x) ≤ ψ(z)dz <
A
q(x)
2
because µ(A/q(x)) = µ(A)/q(x) ≤ µ(A)/q(α) < δ/q(α) = δ 0 . Thus µ(A) < δ
implies G(x) < ε/2 for all x, and hence
Z
ε
Qn h(x)dx < . (3.16)
A∩(r,∞) 2

Combining (3.14), (3.15) and (3.16) gives

Z
Qn h(x)dx < ε
A

when µ(A) < δ and n ≥ K. Thus part (i) of the Dunford-Pettis condition
holds for the collection (Qn h)n≥K .

Next, part (ii) of the condition needs to be checked for the same h. Let I be
the identity map on R++ . We have
Z ∞
n
E(I|Q h) = yQn h(y)dy
Z0 ∞ Z ∞   
y 1 n−1
= y ψ Q h(x)dx dy
0 0 q(x) q(x)
Z ∞ Z ∞
   
y 1
= ψ ydy Qn−1 h(x)dx
q(x) q(x)
Z0 ∞ 0
= E(ε)q(x)Qn−1 h(x)dx
Z0 ∞
≤ E(ε)f (x)Qn−1 h(x)dx.
0

Since E(ε) is finite, it follows from the concavity and Inada conditions in
Assumption 3.1 that x 7→ E(ε)f (x) can be majorized on R++ by an affine
CHAPTER 3. OPTIMAL GROWTH WITH UNBOUNDED SHOCK 62

function with slope less than one. In other words, there exist nonnegative
constants a and b, a < 1, such that E(ε)f (x) ≤ ax + b, ∀x > 0. Therefore
Z ∞ Z ∞
n−1
E(ε)f (x)Q h(x)dx ≤ [ax + b]Qn−1 h(x)dx
0 0

= aE(I|Qn−1 h) + b.

Repeating the argument n times,

b
E(I|Qn h) ≤ an E(I|h) + ,
1−a

or, using finiteness of E(I|h),

b
E(I|Qn h) ≤ 1 +
1−a

when n ≥ M , M suitably large.

By (3.6),
∞
E(I|Qn h)
Z
Qn h(x)dx ≤
r r
for any n and any positive r. Hence
Z ∞  
n 1 b
Q h(x)dx ≤ 1+ , n ≥ M,
r r 1−a

and condition (ii) of the Dunford-Pettis condition holds for {Qn h}n≥M .

Finally, define N = max(K, M ). For such an N , {Qn h}n≥N satisfies both

parts of the Dunford-Pettis condition, completing the proof of Lagrange sta-
bility.

Regarding strong contractiveness of the semidynamical system (D(µ), Q),

the following statement is true.
CHAPTER 3. OPTIMAL GROWTH WITH UNBOUNDED SHOCK 63

Proposition 3.2. If the density ψ of the shock ε is everywhere positive, then

(D(µ), Q) is strongly contractive.

Proof. The result follows immediately from implied positivity of the density
kernel (3.8) and Lemma 2.3, page 39.

3.5.2 Proof of Theorem 3.2

It is now possible to prove Theorem 3.2. Proposition 3.1 shows that if

(u, f, ψ) satisfies Assumptions 3.1–3.3, then the associated semidynamical
system (D(µ), Q) is Lagrange stable. Evidently D(µ) is a closed convex
subset of L1 (µ). Moreover Q is both linear and continuous. Hence all the
conditions of Theorem 2.1, page 34, are satisfied, implying the existence of
an equilibrium density. Since the equilibrium is a density, probability is not
concentrated at zero (i.e. it is a nonzero equilibrium). Regarding part 2 of
Theorem 3.2, if, in addition, ψ is assumed to be everywhere positive, then
(D(µ), Q) is also strongly contractive by Proposition 3.2. Existence, unique-
ness and stability of equilibrium now follow from Theorem 2.2, page 38.
Chapter 4

Systems with Multiplicative

Noise

4.1 Introduction

In this essay we focus on existence, uniqueness and stability of equilibrium

in a specific class of models that arise naturally in economics. In particular,
we assume that the shock ε in (2.4), page 22, is multiplicative, and that the
state space for the endogenous variable xt is the positive half-ray R+ = [0, ∞).
That is,
xt+1 = g(xt )εt , t = 0, 1, . . . , (4.1)

where g : R+ → R+ , and εt ∈ R+ . The importance of these models within

economics stems from inherently nonnegative state variables, such as prices
or physical quantities. A key example is of course stochastic growth theory,

64
CHAPTER 4. SYSTEMS WITH MULTIPLICATIVE NOISE 65

which in turn provides foundations for real business cycle and other macroe-
conomic literature.

While (4.1) excludes a larger model architecture than previous studies of the
equilibrium problem, it is demonstrated that the additional structure can be
exploited to obtain results that have considerable generality within this class.
Further, our approach leads naturally to sufficient conditions stated directly
in terms of the primitives g and ε; such conditions are easy to verify in
applications. Third, the temptation to compactify the state space is resisted,
permitting incorporation of standard econometric shocks. Fourth, equilibria
are realized as fixed points of a contractive linear operator, and are therefore
amenable to approximation by numerical methods.

The stability of (4.1) has previously been studied in the mathematical litera-
ture. In particular, there exists a well-known set of sufficient conditions due
to K. Horbacz [30, Theorem 1]. The results obtained here provide a general
principle which yields the conditions of Horbacz as a special case.

The essay proceeds as follows. Section 4.2 states our results. Section 4.3
gives applications. Section 4.4 formalizes the problem as a preliminary to
the proofs. The proofs are collected in Section 4.5.

4.2 Results

In this section we state new stability results for the economy on R+ defined
by (4.1). The results pertain to the existence, uniqueness and stability of
stochastic equilibria. Our definition of this property is the standard one:
CHAPTER 4. SYSTEMS WITH MULTIPLICATIVE NOISE 66

existence of a unique “stationary” distribution on the state space to which

the marginal distributions of the state variables (xt ) always converge as t →
∞, regardless of the initial state. (Recall Definition 2.3, p. 28.) Formal
definitions of the distribution space and topology of convergence are given in
Section 4.4.

Our basic assumptions on the structure of the model (4.1) are as follows.
First,

Assumption 4.1. The shocks (εt ) are uncorrelated and identically dis-
tributed by density function ψ on R+ .

Second,

Assumption 4.2. The map g is strictly positive almost everywhere on R+ .1

In the remainder of the essay, the model (4.1) is represented by notation

(g, ψ), where ψ is the density of the shock ε.

Our main condition uses the notion of a Lyapunov function on R+ , which we

define to be a continuous, nonnegative function V from R+ into R+ ∪ {∞}
such that V (0) = ∞, V (x) < ∞ for x > 0 and limx→∞ V (x) = ∞.

Condition 4.1. Corresponding to (g, ψ), there exists a Lyapunov function

V on R+ and constants α, C ≥ 0, α < 1, such that
Z
V [g(x)z]ψ(z)dz ≤ αV (x) + C, ∀x ∈ R+ .

1
Thus we accommodate the possibility that g may be zero at a finite number of points.
CHAPTER 4. SYSTEMS WITH MULTIPLICATIVE NOISE 67

The function V in Condition 4.1 is large at 0 and +∞. The condition should
be interpreted as a restriction on the probability that the state variable moves
toward these limits without bound.

Condition 4.2. The density ψ is strictly positive almost everywhere on R+ .2

Most “named” densities on R+ have this property, such as the lognormal,

exponential, χ-squared, gamma, and Weibull densities.

Condition 4.3. For some M < ∞, ψ satisfies ψ(z)z ≤ M , ∀z ∈ R+ .

Condition 4.3 also holds for the lognormal, exponential, χ-squared, gamma
and Weibull distributions. The condition is used here to bound the proba-
bility that ψ assigns to closed intervals in R+ \ {0}.

Theorem 4.1. Let (g, ψ) be an economy on R+ satisfying Assumptions 4.1

and 4.2. If g and ψ also satisfy Conditions 4.1, 4.2 and 4.3, then (g, ψ) has
a unique, globally stable equilibrium.

Alternatively, suppose that

Condition 4.4. The map g is weakly monotone increasing on the nonempty

interval [0, r), and g(x) ≥ b > 0 on [r, ∞).

Theorem 4.2. Let (g, ψ) be an economy on R+ satisfying Assumptions 4.1

and 4.2. If g and ψ also satisfy Conditions 4.1, 4.2 and 4.4, then (g, ψ) has
a unique, globally stable equilibrium.
2
When this is the case, the same distribution for ε can be represented by a density
which is positive everywhere on R+ . Hence in the remainder of the paper we can assume
without loss of generality that ψ(z) > 0, ∀z ∈ R+ .
CHAPTER 4. SYSTEMS WITH MULTIPLICATIVE NOISE 68

The proofs of Theorems 4.1 and 4.2 are given in Section 4.5.

Corollary 4.1. Let (g, ψ) be an economy on R+ satisfying Assumptions 4.1

and 4.2. If g is weakly monotone increasing and, in addition, g and ψ to-
gether satisfy Conditions 4.1 and 4.2, then (g, ψ) has a unique, globally stable
equilibrium.

Proof. Evidently Condition 4.4 is also satisfied if Assumption 4.2 holds and
g is weakly monotone increasing on R+ . Theorem 4.2 then implies the stated
result.

4.3 Applications

We analyze the dynamics of two capital accumulation models using our meth-
ods. One is a standard overlapping generations model, while the other is of
optimal growth with externality-driven increasing returns.

4.3.1 Overlapping Generations

In the deterministic case, dynamics of the overlapping generations model with

productive capital were extensively studied by Galor and Ryder [22]. They
establish convergence to a unique, nontrivial equilibrium under a strength-
ened Inada condition. Here the analysis is extended to the stochastic case,
proving that analogous results hold under the same condition.3
3
Previously the stochastic overlapping generations model was also analyzed by
Wang [62].
CHAPTER 4. SYSTEMS WITH MULTIPLICATIVE NOISE 69

The framework is as follows. Agents live for two periods, working in the
first and living off savings in the second. Savings in the first period forms
capital stock, which in the following period will be combined with the labor
of a new generation of young agents for production under the technology
yt+1 = F (kt , `t )εt , where y is income, k is capital and ` is labor input. For
convenience we assume that labor is constant (`t = `), and set f (k) = F (k, `).
Following Galor and Ryder [22, p. 362], we assume that f : R+ → R+ has
the usual properties f (0) = 0, f ∈ C 2 , f 0 > 0, f 00 < 0, and

lim f 0 (k) = ∞, lim f 0 (k) = 0.

k↓0 k↑∞

In addition, Galor and Ryder [22, Proposition 5, Corollary 1] introduce the

extended Inada condition

lim[−kf 00 (k)] > 1. (GR)

k↓0

The shocks (εt ) are uncorrelated and identically distributed on R+ according

to density ψ. We assume that ψ is strictly positive on R+ .

As Galor and Ryder point out [22, Lemma 1, p. 365], restrictions on the
utility function are necessary to obtain unique self-fulfilling expectations.
Here it is assumed that agents maximize utility

U (ct , c0t+1 ) = ln ct + βE(ln c0t+1 ), β ∈ (0, 1),

subject to the budget constraint c0t+1 = (wt − ct )(1 + rt+1 ), where c (respec-
tively, c0 ) is consumption while young (respectively, old), wt is the wage and
rt is the interest rate. In this case optimization implies a savings rate from
CHAPTER 4. SYSTEMS WITH MULTIPLICATIVE NOISE 70

wage income of β/(1 + β), whence kt+1 = (β/(1 + β))wt . Assuming that
labor is paid its marginal factor product yields the law of motion

β
kt+1 = [f (kt ) − kt f 0 (kt )]εt . (4.2)
1+β

Proposition 4.1. Assume that the Galor-Ryder condition (GR) holds. If,
in addition, E(ε) < ∞ and E(1/ε) < β/(1 + β), then the economy (4.2) has
a unique and globally stable stochastic equilibrium.

Remark 4.1. As in Chapter 3, Section 3.3.2, the bound E(1/ε) < β/(1 + β)
is used to restrict weight in the left-hand tail of ψ, preventing the economy
from collapsing to zero as a result of adverse shocks.

Proof. We verify that (4.2) satisfies the conditions of Corollary 4.1. To this
end, let D = β/(1 + β), let h(k) = f (k) − kf 0 (k) and let g(k) = Dh(k). It
follows from our assumptions on f that the function k 7→ h(k) is zero at zero,
continuously differentiable, strictly increasing and satisfies limk↓0 h0 (k) > 1.
This last fact—which is equivalent to (GR)—implies that

∃ δ > 0 s.t. h(k) ≥ k, ∀k ∈ [0, δ). (4.3)

Evidently Assumptions 4.1 and 4.2 are satisfied. Regarding Condition 4.1,
consider the Lyapunov function defined by V (k) = 1/k + k. We have
Z
1
V [g(k)z]ψ(z)dz = E(1/ε) + E(ε)g(k). (4.4)
g(k)

Consider the first term in the right hand side of (4.4). By (4.3),

1 1
E(1/ε) ≤ α1 , ∀k ∈ [0, δ), (4.5)
g(k) k
CHAPTER 4. SYSTEMS WITH MULTIPLICATIVE NOISE 71

where α1 = E(1/ε)D−1 < 1. In addition, monotonicity of g yields

1 1
E(1/ε) ≤ E(1/ε) , ∀k ∈ [δ, ∞). (4.6)
g(k) g(δ)

Combining (4.5) and (4.6) gives

1 1
E(1/ε) ≤ α1 + C1 , ∀k ∈ R+ . (4.7)
g(k) k

Consider now the second term in the sum (4.4). By the assumptions on f
it is clear that the function k 7→ E(ε)Df (k) can be majorized on R+ by an
affine function k 7→ α2 k + C2 , where α2 and C2 are nonnegative constants,
α2 < 1. Therefore,

E(ε)g(k) ≤ E(ε)Df (k) ≤ α2 k + C2 , ∀k ∈ R+ . (4.8)

Let α = max(α1 , α2 ), and let C = C1 + C2 . Substituting (4.7) and (4.8) into

(4.4) gives
Z
V [g(k)z]ψ(z)dz ≤ α(1/k + k) + C = αV (k) + C. (4.9)

Since α < 1, Condition 4.1 is satisfied.

In addition, Condition 4.2 is satisfied by hypothesis, and k 7→ g(k) is mono-

tone increasing on R+ . Thus all of the conditions of Corollary 4.1 are veri-
fied.

4.3.2 Stability in a Model with Externalities

Consider the following Brock-Mirman optimal growth model with increasing

returns. The production function is Cobb-Douglas with an external effect
CHAPTER 4. SYSTEMS WITH MULTIPLICATIVE NOISE 72

due to the existence of increasing social returns; technology is sensitive to

economy-wide capital aggregates. Thus,

yt+1 = A(kt )ktα `1−α

t εt , (4.10)

where externalities are captured by the function k 7→ A(k). This dependence

is external to individual agents, and A is treated as constant with respect to
private investment. The capital share α satisfies 0 < α < 1.

Regarding the nature of increasing returns, we assume only that

Assumption 4.3. The range of A : R+ → R+ is contained in a closed and

bounded subset of R+ \ {0}.

Macroeconomic models with external effects satisfying Assumption 4.3 in-

clude Azariadis and Drazen [6], Galor and Zeira [23], and Quah [48]. For
example, Azariadis and Drazen consider the model


A1 , if k ≤ kb ;

A(k) =

A2 , if k > kb .


Here 0 < A1 < A2 represent the state of technology, and kb is a fixed “thresh-
old” level of capital per worker.

For convenience, labor supply is normalized to unity. The productivity shocks

εt ∈ R+ are uncorrelated and identically distributed with density ψ.

Let ct be time t consumption. A representative agent seeks to maximize

"∞ #
X
t
E β ln ct , β ∈ (0, 1),
t=0
CHAPTER 4. SYSTEMS WITH MULTIPLICATIVE NOISE 73
kt+1
450

αβA(kt )ktα

k1∗ kb k2∗ kt

Figure 4.1: The map kt 7→ αβA(kt )ktα .

subject to kt+1 + ct ≤ yt . The optimal policy [58, p. 19 and pp. 274–77] is

kt+1 = αβA(kt )ktα εt . (4.11)

A plot of the map k 7→ αβA(k)k α is given in Figure 4.1 for the Azariadis-
Drazen threshold case, in which A is a step function. As drawn, the deter-
ministic version has two equilibria, k1∗ and k2∗ .

Despite the apparent simplicity of (4.11), establishing dynamic stability is

complicated by the dependence of A on kt , which is potentially highly non-
linear. For example, none of the three main sufficient conditions used by
Stokey, Lucas and Prescott [58, Theorem 12.12] are satisfied. Also, in the
deterministic case (when εt is held constant), the model (4.11) may have
multiple local attractors. It is therefore somewhat surprising that

Proposition 4.2. For a class of shocks that include the lognormal distribu-
tions, the economy (4.11) has a unique, globally stable stochastic equilibrium.
CHAPTER 4. SYSTEMS WITH MULTIPLICATIVE NOISE 74

1.4
ϕ0
ϕ25
1.2 ϕ2000

0.8

0.6

0.4

0.2

0
-8 -6 -4 -2 0 2 4

Figure 4.2: Convergence to equilibrium

Proof. We verify that (4.11) satisfies the conditions of Theorem 4.1. Evi-
dently Assumptions 4.1 and 4.2 hold. By the hypotheses of the proposition,
we may assume that ψ satisfies Conditions 4.2 and 4.3, and that E | ln ε| < ∞.
Regarding Condition 4.1, let V (k) = | ln k|. The function V so constructed
is a Lyapunov function on R+ . Moreover,
Z Z
α
V [DA(k)k z]ψ(z)dz = | ln D + ln A(k) + α ln k + ln z|ψ(z)dz

≤ α| ln k| + C

= αV (k) + C,

where C = | ln D| + supk | ln A(k)| + E | ln ε|. Since α < 1 and C < ∞,

Condition 4.1 holds. The proof is complete.

To illustrate this result, Figure 4.2 presents a sequence of densities generated

by iterating the Markov operator P implied by (4.11)—see Section 4.4 for
details on the construction of P —on an arbitrary initial distribution ϕ0 ,
CHAPTER 4. SYSTEMS WITH MULTIPLICATIVE NOISE 75

where the x-axis is the logarithm of capital per head.4 Here ϕ0 can be
thought of as an initial distribution of a “large” number of Azariadis-Drazen
economies.

In the figure, the density ϕ0 is the left-most distribution, with probability

mass shifting rightwards over time. It is interesting to observe that the
nonlinear external effects in the production function (4.10) lead to the cross-
country income distribution developing a bimodal structure which has been
observed in the actual cross-country growth data by, among others, Quah
[48, 49], Jones [31] and Durlauf and Quah [16].

Proposition 4.2 implies that the sequence of densities (ϕn ) converges to a

unique limiting density ϕ∗ . In this case there is little observable change after
t = 2000.

4.3.3 Existing Conditions

Previously a set of conditions for obtaining stability of the model (4.1) was
identified by K. Horbacz [30, Theorem 1]. Her results can be obtained as a
special case of Theorem 4.2.

The statement and proof of the above problem are somewhat tangential to
the economics (as opposed to mathematics) discussed in this chapter. As a
4
The parameters are β = 1, α = 0.5, A1 = 0.5, A2 = 2, kb = 0.6875, ε lognormal, ln ε ∼
N (0, 0.5). The densities are generated using a Monte Carlo simulation and estimated
nonparametrically by the Parzen window method with Gaussian kernel and bandwidth
0.38. Such estimates are known to converge to the true density in L1 norm for large
sample size [12]. Here each generation is represented by 200 sample points.
CHAPTER 4. SYSTEMS WITH MULTIPLICATIVE NOISE 76

result they have been deferred to the appendix.

4.4 Formulation of the Problem

Prior to the proofs, a more formal definition of the model (4.1) is given. To
begin, let R be the real numbers, let B be the Borel sets of R, let R+ =
[0, ∞), and let B+ = B ∩ R+ . The Lebesgue measure is again denoted by µ.
Integration where the measure is not made explicit is taken with respect to
R
µ; integration using the symbol without subscript is taken over R+ .

In what follows we use the notation of Chapter 2, Section 2.3, where the
underlying space X of Chapter 2 now corresponds to the half-ray R+ with
the usual topology.

Let the sequence of shocks (εt ) in (4.1) be uncorrelated and identically dis-
tributed by Ψ ∈ P. We assume that Ψ is absolutely continuous with re-
spect to µ. In this case there exists a unique density ψ ∈ D(µ) satisfying
ψ = Ψ(B) for all B ∈ B+ ; ψ is the Radon-Nikodým (RN) derivative of Ψ
R
B

with respect to µ.

Definition 4.1. Let g : R+ → R+ be a measurable function. In what follows,

a perturbed dynamical system on R+ is defined by a pair (g, ψ), where, given
current state value xt ∈ R+ , a shock εt ∈ R+ is selected independently from
density ψ, and the next period state is realized as in (4.1).

We wish to embed the Markov chain generated by (g, ψ) in the function space
L1 (µ). To do so requires that the transition probabilities B 7→ N(x, B) of
CHAPTER 4. SYSTEMS WITH MULTIPLICATIVE NOISE 77

the Markov chain (see (2.8), p. 26) can be represented by density functions
(recall Assumption 2.1, p. 30, and the subsequent discussion).

If g satisfies Assumption 4.2 and ε is distributed according to density ψ, then

for almost all x, the distribution B 7→ N(x, B) is absolutely continuous with
respect to µ (i.e., Assumption 2.1 is satisfied). To see this, pick any x such
that g(x) > 0, and any E ∈ B such that µ(E) = 0. We need to show that
N(x, E) = 0. This must be the case, because if E is a null set then so is
E/g(x), and hence
Z
N(x, E) = 1E [g(x)z]ψ(z)dz
Z
= ψ(z)dz
E/g(x)

= 0.

In particular, for x such that g(x) > 0,

 
y 1
p(x, y) = ψ , (4.12)
g(x) g(x)

because changing variables shows that for any B ∈ B+ ,

Z Z
p(x, y)dy = 1B [g(x)z]ψ(z)dz = N(x, B).
B

For other x set p(x, ·) equal to any density.5

It follows from this discussion that the density kernel p exists, as does the
Markov operator in L1 (µ) defined by (2.11) on page 30.
5
Density kernels need be defined only up to the complement of a null set—systems
with kernels equal µ × µ-a.e. have identical Markov operators and we do not distinguish
between them in what follows.
CHAPTER 4. SYSTEMS WITH MULTIPLICATIVE NOISE 78

The notion of existence, uniqueness and stability of equilibrium used in the

essay can now be formalized by Definition 2.4, page 31.

4.5 Proofs

The main proof is based on Theorem 2.2, page 38. We show that the semidy-
namical system (D(µ), P ) associated with the economy (g, ψ) is both strongly
contracting (Definition 2.6, p. 36) and Lagrange stable (Definition 2.5, p.
33). As usual, D(µ) is treated as a metric space with the L1 norm distance.

Lemma 4.1. Let (g, ψ) be a perturbed dynamical system satisfying Assump-

tions 4.1 and 4.2, let p be the density kernel defined in Section 4.4, and let P
be the Markov operator associated with p by (2.11). If Condition 4.2 holds,
then the semidynamical system (D(µ), P ) is strongly contracting.

Proof. Since ψ is strictly positive, representation (4.12) implies that the den-
sity kernel p is µ × µ-a.e. strictly positive on X × X. As stated earlier, per-
turbed dynamical systems with kernels that are equal µ × µ-a.e. have iden-
tical Markov operators. If p is strictly positive, then (D(µ), P ) is strongly
contracting by Lemma 2.3, page 39. The result follows.

Next we treat Lagrange stability of the semidynamical system (D(µ), P )

associated with (g, ψ). Recall that to prove Lagrange stability it is sufficient
to find a set D ⊂ D(µ) such that D is norm-dense in D(µ) and the trajectory
of ϕ under P is weakly precompact for each ϕ ∈ D (Proposition 2.1, p. 41).
CHAPTER 4. SYSTEMS WITH MULTIPLICATIVE NOISE 79

Lemma 4.2. Let (g, ψ) be a perturbed dynamical system on R+ satisfying

Assumptions 4.1 and 4.2, and let P be the associated Markov operator. If
Condition 4.1 and either one of Condition 4.3 or 4.4 holds, then there exists
a set D ⊂ L1 (µ) such that D is dense in D(µ) and {P t ϕ : t ≥ 0} is weakly
precompact for each ϕ ∈ D.

Proof. Let V be the Lyapunov function in Condition 4.1. Let D be the set
of all density functions ϕ in L1 (µ) such that
Z
V (x)ϕ(x)dx < ∞. (4.13)

We claim that D has the desired properties.

Pick any density ϕ. To see that there exists a (ϕk ) ⊂ D with ϕk → ϕ,

define first ϕ0k = 1[1/k,k] ϕ. By the monotone convergence theorem, kϕ0k k → 1.
Hence kϕ0k k > 0 for all k greater than some constant K. For all such k define
ϕk = kϕ0k k−1 ϕ0k . Then ϕk ∈ D(µ) for all k ≥ K by construction. Moreover,
ϕk → ϕ pointwise, and hence in L1 norm by Scheffé’s lemma. Finally, ϕk ∈ D
for all k ≥ K, because
Z Z
1
V (x)ϕk (x)dx = 1[1/k,k] (x)V (x)ϕ(x)dx,
kϕ0k k

and V is bounded on compact subsets of R+ \ {0} by continuity.

It remains to show that if ϕ ∈ D, then {P t ϕ : t ≥ 0} is weakly precompact.

Note first that the collection {P t ϕ} is norm-bounded, because P D(µ) ⊂
D(µ). Thus it remains only to verify parts (i) and (ii) of the Dunford-Pettis
condition (Section 2.5.3, p. 41).
CHAPTER 4. SYSTEMS WITH MULTIPLICATIVE NOISE 80

Regarding (i), pick any ε > 0. We exhibit a δ > 0 and an N ∈ N such that
Z
µ(A) < δ =⇒ P t f (x)dx < ε, ∀ t ≥ N.
A

R
Define E(V |g) = V g. By Fubini’s theorem,
Z
t
E(V |P ϕ) = V (y)P t ϕ(y)dy
Z Z 
t−1
= V (y) p(x, y)P ϕ(x)dx dy
Z Z 
= V (y)p(x, y)dy P t−1 ϕ(x)dx.

But
Z Z
V (y)p(x, y)dy = V [g(x)z]ψ(z)dz ≤ αV (x) + C

for all x by Condition 4.1. Therefore,

Z
t
E(V |P ϕ) ≤ [αV (x) + C]P t−1 ϕ(x)dx = αE(V |P t−1 ϕ) + C.

Repeating this argument obtains

C
E(V |P t ϕ) ≤ αn E(V |ϕ) + .
1−α

Since E(V |ϕ) is finite by (4.13), it follows that

C
E(V |P t ϕ) ≤ 1 + , t ≥ N,
1−α

for some N ∈ N.

On the other hand, it can be verified that for arbitrary positive a,

Z
a P t ϕ ≤ E(V |P t ϕ)
R+ \Ga
CHAPTER 4. SYSTEMS WITH MULTIPLICATIVE NOISE 81

when Ga is defined as the set of x ∈ R+ with V (x) ≤ a. Therefore,

Z  
t 1 C
P ϕ≤ 1+ , ∀ t ≥ N, ∀ a > 0. (4.14)
R+ \Ga a 1−α
Choose a so large that
 
1 C ε
1+ ≤ . (4.15)
a 1−α 2

Consider now the decomposition

Z Z Z
t t
P ϕ= P ϕ+ P t ϕ.
A A∩Ga A∩[R+ \Ga ]

Using (4.14) and (4.15) gives

Z Z
t ε
P ϕ≤ P tϕ + , (4.16)
A A∩Ga 2
whenever t ≥ N . Here a is the constant determined in (4.15).

The next step is to bound the first term in the sum on the right hand side
of (4.16), taking the constant a as given, and assuming that at least one of
Condition 4.3 or Condition 4.4 holds.

Assume first that Condition 4.3 holds. Using the expression for the density
kernel given in (4.12), we have
Z
P ϕ(y) = p(x, y)P t−1 ϕ(x)dx
t

Z  
y 1
= ψ P t−1 ϕ(x)dx
g(x) g(x)
Z  
y y 1 t−1
= ψ P ϕ(x)dx
g(x) g(x) y
M
≤ .
y
Therefore,
Z Z Z
t M
P ϕ(y)dy ≤ dy ≤ J(a)dy = J(a)µ(A),
A∩Ga A∩Ga y A
CHAPTER 4. SYSTEMS WITH MULTIPLICATIVE NOISE 82

where the finite number J(a) is the maximum of M/y over the closed and
bounded interval Ga ⊂ R+ \ {0}.

Now pick any positive δ satisfying δ ≤ ε/(J(a)2). For such a δ we have

Z
ε
µ(A) < δ =⇒ P t f (x)dx < .
A∩Ga 2
Combining this with (4.16) proves (i) of the Dunford-Pettis characterization
for the collection {P t ϕ : t ≥ N } when Condition 4.3 holds.

We now establish the same when Condition 4.4 holds, again by bounding
the first term in the sum (4.16). Suppose first that there exists a c with
g(x) ≥ c > 0 for all x in R+ . In this case, because
Z Z Z
t
P ϕ(y)dy = p(x, y)P t−1 ϕ(x)dxdy
A∩Ga A∩G
Z Za 
= p(x, y)dy P t−1 ϕ(x)dx,
A∩Ga

and because
Z Z  
y 1
p(x, y)dy = ψ dy
A∩Ga g(x) g(x)
ZA∩Ga
= ψ(z)dz
A∩Ga
g(x)
Z
≤ ψ(z)dz
A∩Ga
c
Z
≤ ψ(z)dz
A
c

for all x ∈ R+ , it follows that if δ 0 > 0 is chosen such that

Z
0 ε
µ(A) < δ =⇒ ψ(z)dz <
A 2
(existence of such a δ 0 is by absolute continuity of A 7→ A ψ with respect to
R

µ), then
Z Z
ε
p(x, y)dy ≤ ψ(z)dz <
A∩Ga A
c
2
CHAPTER 4. SYSTEMS WITH MULTIPLICATIVE NOISE 83

whenever µ(A) < δ, δ = δ 0 c, and, therefore,

Z
ε
µ(A) < δ =⇒ P tϕ < .
A∩Ga 2
Again, combining this with (4.16) yields (i) of the Dunford-Pettis condition.

Finally, suppose to the contrary that while Condition 4.4 is satisfied, there
exists no c with g(x) ≥ c > 0 for all x ∈ R+ . In this case Condition 4.4
implies that g(x) ↓ 0 as x ↓ 0, and hence there exists a d > 0 such that
Z Z
ε
p(x, y)dy = ψ(z)dz ≤ for almost all x ∈ [0, d), (4.17)
A∩Ga A∩Ga
g(x)
2
owing to the fact that A ∩ Ga is bounded away from 0. For x ≥ d, g(x) ≥
c0 = min[g(d), b] > 0, where b is the positive constant in Condition 4.4.6 In
this case, an argument similar to that given above for the case g(x) ≥ c > 0
implies that
Z Z
ε
p(x, y)dy ≤ ψ(z)dz < (4.18)
A∩Ga A 2
c0

whenever x ∈ [d, ∞) and µ(A) < δ, δ = δ 0 c0 . Combining (4.17) and (4.18)

yields
Z
ε
µ(A) < δ =⇒ P tϕ < .
A∩Ga 2
Once again, (i) of the Dunford-Pettis characterization holds.

It remains to establish that part (ii) of the Dunford-Pettis condition also

holds for the same collection. We have already shown that
Z  
t 1 C
P ϕ≤ 1+
R+ \Ga a 1−α
for all positive a, all t ≥ N . But this inequality is sufficient, because Ga is
always bounded. Hence condition (ii) is also satisfied for {P t ϕ : t ≥ N }.
This completes the proof of the lemma.
6
Here g(d) > 0 by Condition 4.4 and the almost everywhere positivity of g.
Chapter 5

Asymptotic Distributions

5.1 Introduction

As discussed in Chapter 3, dynamic properties of the one-sector stochastic

optimal growth model with concave production technology were first studied
in the well-known paper of Brock and Mirman [10]. Using the same assump-
tions on preferences and technology as used in the deterministic case, they
showed that there exists a unique and globally stable stochastic equilibrium.
In their framework of analysis the influence of the shock is bounded, and the
state eventually converges to a compact “invariant set” in the positive real
numbers.

Stochastic models that have an equilibrium or steady state distribution may

also have asymptotic statistical properties related to the existence of a steady
state distribution, such as convergence of sample averages from time series to
the mean of this limiting distribution, or asymptotic normality of the partial

84
CHAPTER 5. ASYMPTOTIC DISTRIBUTIONS 85

sums. The first question concerns a law of large numbers (LLN) result, while
the latter concerns a central limit theorem (CLT) result.

The importance of these questions can be summarized as follows. If an

LLN condition holds, then it is possible to test a given theoretical model by
comparing the mean of the limiting distribution with a sample average from
a sufficiently large data set generated by the system under study. Conversely,
suppose that an expression is available for the mean of the hypothetical model
in terms of its parameters. Then the implied equality of this expression
and the sample mean calculated from data provides a consistent method for
estimating parameter values. If, in addition, a CLT result is available, then
inference can be drawn as to the likelihood of values in the parameter space.

It known that both the LLN and the CLT result are realized for the general
discrete-time concave stochastic optimal growth model when the shock has
compact support [58, 17, 5, 7]. Central to the proofs is boundedness of the
productivity shock, which, in combination with Inada conditions, allow the
state space to be taken to be compact.1 When the influence of the shock is
not bounded, however, compactness of the state space fails. In this case it is
unclear whether or not LLN and CLT properties continue to hold.

In this essay we begin to address this question by introducing a methodology

and studying some specific parameterizations. The techniques are based on
results recently obtained in an important paper of Loskot and Rudnicki [47],
1
In a related paper, ergodicity in moments for the Solow-Swan model with a shock that
is unbounded above but cannot be arbitrarily small is investigated in Binder and Pesaran
[8].
CHAPTER 5. ASYMPTOTIC DISTRIBUTIONS 86

which studies the LLN and CLT properties of nonlinear dynamical systems
perturbed by uncorrelated noise.2 The question of asymptotic statistical
properties for the general concave one-sector optimal growth model with
unbounded shock is left open. It is hoped that the methodology used here
can be extended to the general problem.

Section 5.2 formulates the problem and gives the major definitions. Section
5.3 provides a general stability result. Section 5.4 gives applications. Section
5.5 gives proofs.

5.2 The Model

As in Chapter 2, we consider a growth model evolving on state space X, where

X is an arbitrary topological space. Specifically, we consider the stochastic
dynamic economy (T, Ψ) of Chapter 2, Section 2.3.1.

As in Chapter 2, Section 2.3, B is the Borel sets of X, M denotes the normed

vector lattice of finite signed Borel measures µ : B → R, and P is the distri-
butions in M. All integrals of real functions defined on X are over the whole
space X unless otherwise stated.

In this chapter, however, we assume throughout that the topology of X is

metrizable. Let % : X × X → R+ be a distance which metrizes the topology
on X.
2
Previously, the techniques of Loskot and Rudnicki have been applied to areas such as
neurodynamics [19] and entropy computation in iterated function systems [53].
CHAPTER 5. ASYMPTOTIC DISTRIBUTIONS 87

To briefly recall the main ideas from Chapter 2, random outcomes are selected
from some measurable space (Ω, F) by probability measure P, and mapped
into X by random variable ε : Ω → X. Corresponding to ε is a finite Borel
measure Ψ ∈ P defined at B ∈ B by Ψ(B) = P[ε−1 (B)]. The measure
Ψ is called the distribution of ε, and, as usual, the distribution satisfies
f [ε(ω)]P(dω) = X f (z)Ψ(dz) for any real B-measurable function f .
R R
Ω

Given a transition rule T mapping X × X into X, and, given current state

value xt ∈ X, a shock εt ∈ X is selected independently from Ψ, and the next
period state is realized as

xt+1 = T (xt , εt ). (5.1)

The notion of equilibrium we use is again the standard one of Chapter 2

(Definition 2.3, p. 28). To repeat, an equilibrium for the economy (T, Ψ) is
a probability measure ϕ ∈ P that satisfies
Z Z 
1B [T (x, z)]Ψ(dz) ϕ(dx) = ϕ(B) (5.2)

for all B ∈ B.3 The equilibrium is unique if there exists no other point in P
satisfying (5.2).

Suppose that the growth model (5.1) has a unique equilibrium ϕ. For the
purposes of this paper, (5.1) is said to satisfy the law of large numbers if, for
any Lipschitz function g : X → R,
N −1 Z
1 X
g(xt ) → g(x)ϕ(dx) (5.3)
N t=0
3
As before, 1B : X → {0, 1} is the characteristic function of B ∈ B.
CHAPTER 5. ASYMPTOTIC DISTRIBUTIONS 88

P-almost surely as N → ∞.4

The economy is said to have the central limit property if, for any g as above,
N −1
1 X
√ g(xt ) → N (m, σ 2 ) (5.4)
N t=0

in distribution, where N (m, σ 2 ) is a normal distribution with mean m =

R
g(x)ϕ(dx) and variance σ 2 ≥ 0.

5.3 Results

Loskot and Rudnicki [47] consider stochastic models that satisfy the following
contraction condition.

Definition 5.1. The pair (T, Ψ), where T is the map in (5.1) and Ψ is the
distribution of the shock ε, is called an average contraction with respect to %
R
if there exists a Borel function λ : X → R such that E(λ) = λ(z)Ψ(dz) < 1
and
%(T (x, z), T (x0 , z)) ≤ λ(z)%(x, x0 ), ∀x, x0 , z ∈ X.

By adapting the results of Loskot and Rudnicki and using additional restric-
tions on the space X, we obtain the following stability condition.

Theorem 5.1. Let (T, Ψ) be the stochastic dynamic economy of Section 5.2.
Let the state space X be both locally compact and σ-compact in the topology
4
A real function g on X is called Lipschitz if there exists a constant λ such that
|g(x) − g(x0 )| ≤ λ%(x, x0 ) for any x, x0 ∈ X.
CHAPTER 5. ASYMPTOTIC DISTRIBUTIONS 89

metrized by %.5 If the growth model defined by the law of motion T and the
distribution of the shock Ψ is an average contraction with respect to %, and
if there exists at least one point x̄ ∈ X such that
Z
%(x̄, T (x̄, z))Ψ(dz) < ∞, (5.5)

then there exists a unique stochastic equilibrium ϕ ∈ P satisfying (5.2), and,

in addition, the model satisfies the law of large numbers property (5.3). If,
moreover, E(λ2 ) < 1 and
Z
[%(x̄, T (x̄, z))]2 Ψ(dz) < ∞ (5.6)

holds, then the model also satisfies the central limit property (5.4).

The proof (Section 5.5) is a straightforward consequence of the results of

Loskot and Rudnicki. The only technical difficulty is to verify that the steady
state notion used by Loskot and Rudnicki is equivalent to the definition (5.2),
which is standard in the economic literature. This can be done under local
and σ-compactness of the state space, as was assumed in the theorem.

5.4 Applications

Let X = (0, ∞). Consider the one-sector optimal growth problem

"∞ #
X
max E β t u(ct ) (5.7)
t=0
s.t. kt+1 = f (kt , εt ) − ct (5.8)
5
Recall that a topological space is called locally compact if every point in the space has
a neighborhood with compact closure, and σ-compact if every open set can be obtained
as the union of a countable number of compact sets.
CHAPTER 5. ASYMPTOTIC DISTRIBUTIONS 90

where ct ∈ X is consumption, kt ∈ X is capital per head, β ∈ (0, 1) is a

discount factor, and f : X × X → X and u : X → R are the production
and utility functions respectively [10, pp. 484–488]. The utility function
u is assumed to satisfy u0 (0) = ∞, which assures interiority of solutions,
and therefore eliminates the possibility of zero savings or consumption. The
shocks εt ∈ X are independent draws from P as before.

The solution to the planning problem, if it exists, is an optimal policy

g : X → X, which associates realized output f (kt , εt ) with optimal current
consumption ct . Optimal consumption g(f (kt , εt )) can then be substituted
into (5.8) to obtain the closed loop law of motion for the system, which is in
the form of (5.1). A unique and well-defined stochastic process is specified
by this law and any initial condition k0 ∈ X. The process so generated is
called an optimal program.

Example 5.1. Consider first the unit-elastic decreasing returns model u(c) =
ln c, f : (k, ε) 7→ Ak α ε, A > 0, α ∈ (0, 1). For such a specification, the
optimal policy consumes a fraction 1 − αβ of realized output Ak α ε, implying
the law of motion
kt+1 = αβAktα εt . (5.9)

Define a binary % on X × X by %(x, y) = | ln x − ln y|. Evidently % is a metric

on X. Moreover, the space (X, %) is isometrically isomorphic to R under the
mapping x 7→ ln x when the latter space is endowed with its usual Euclidean
metric. Hence (X, %) is both locally and σ-compact.

It can be verified that (5.9) is an average contraction on (X, %) for every

random variable ε. If E| ln ε| is finite, then condition (5.5) of Theorem 5.1
CHAPTER 5. ASYMPTOTIC DISTRIBUTIONS 91

holds for x̄ = 1, implying the existence of a unique stochastic equilibrium

ϕ ∈ P, and the LLN result (5.3) for the process (kt )t≥0 .

Evidently E(λ2 ) < 1 also holds. If, in addition, E[(ln ε)2 ] is finite, then (5.6)
is satisfied for x̄ = 1, and the CLT result (5.4) obtains.

Remark 5.1. The conditions E| ln ε| < ∞ and E[(ln ε)2 ] < ∞ can be viewed
as restrictions on the left- and right-hand tails of the distribution. See the
discussion in Stachurski [55, Remark 4.1].

Example 5.2. The second example is from Mirman and Zilcha [46, Example
A, p. 333]. The state space X, the shock ε, the discount factor β, the
productivity parameter A and utility u(c) = ln c are as before. Let α be a
Borel function from X into (0, 1). The production function is (k, ε) 7→ Ak α(ε) .
For such a specification, the law of motion is
Z
α(ε )
kt+1 = ᾱβAkt t , ᾱ = α[ε(ω)]P(dω). (5.10)
Ω

Once again, the system is an average contraction on X under the metric %,

this time using λ(z) = α(z). Note that in this case (5.5) holds for any shock
ε when x̄ = 1. Hence for any strictly positive initial condition k0 , a unique
equilibrium distribution ϕ exists and the LLN condition (5.3) holds.

Moreover, E(λ2 ) < 1, and (5.6) holds for any shock ε when x̄ = 1, implying
that the CLT condition (5.4) also holds.
CHAPTER 5. ASYMPTOTIC DISTRIBUTIONS 92

5.5 Proofs

This section contains the proof of Theorem 5.1. Throughout, Cb denotes

the Banach lattice of continuous bounded real functions on X, and C0 ⊂ Cb
denotes the continuous real functions with compact support. In what follows,
the scalar product notation is used for integration. Thus, for bounded Borel
function f : X → R and µ ∈ M, we write hf, µi for f dµ. Finally, let M+
R

be the nonnegative measures in M (i.e., M+ is the positive cone of M).

For f ∈ C0 and µ ∈ M+ ,
Λ(f ) = hf, µi (5.11)

defines a positive linear functional Λ on C0 . Denote by C0∗ the class of

positive linear functionals on C0 . We will make use of the well-known fact
that the association µ 7→ Λ from the finite Borel measures M+ to the linear
functionals C0∗ defined by (5.11) is one-to-one [2, Theorems 38.3 and 38.4].

Loskot and Rudnicki [47, Theorems 1 and 3] proved that when (T, Ψ) is an
average contraction, (X, %) is complete and separable, and (5.5)–(5.6) holds,
then there exists a unique distribution ϕ ∈ P such that
Z Z  Z
f [T (x, z)]Ψ(dz) ϕ(dx) = f (x)ϕ(dx), ∀f ∈ Cb , (5.12)

and, moreover, the LLN and CLT results (5.3) and (5.4) both hold for ϕ.

Every locally and σ-compact metric space is both complete and separable.
Thus to establish Theorem 5.1 it suffices to verify that the condition (5.12)
characterizes the set of Brock-Mirman equilibria under the hypotheses of the
theorem.
CHAPTER 5. ASYMPTOTIC DISTRIBUTIONS 93

Lemma 5.1. Let (X, %) be a locally and σ-compact metric space, and let ϕ
be any finite Borel measure on X. The measure ϕ satisfies (5.12) if and only
if it also satisfies (5.2).

Proof. In what follows, by a Borel function is meant a real B-measurable

function on X. Define an operator P ∗ from the set of all bounded Borel
functions f : X → R into itself by
Z
∗
(P f )(x) = f [T (x, z)]Ψ(dz), x ∈ X.

Define in addition an operator P from the space of finite measures M+ into

itself by
Z Z
(P µ)(B) = 1B [T (x, z)]Ψ(dz)µ(dx), B ∈ B.

(This is just the Markov operator associated with (T, Ψ).) The operator P
is “adjoint” to P ∗ , in the sense that

hf, P µi = hP ∗ f, µi (5.13)

for every bounded Borel function f and every finite Borel measure µ ∈ M+ .
To see this, pick any B ∈ B. Evidently (5.13) holds when f = 1B . By
linearity of the inner product, (5.13) also holds when f is a step function
taking only finitely many values. This can be extended from step functions
to any bounded nonnegative Borel function by pointwise approximation and
a monotone convergence result in the usual way. Linearity then implies the
result for an arbitrary bounded Borel function f , which can always be written
as the difference between two nonnegative parts.

Assume now that ϕ satisfies (5.12). Then

hP ∗ f, ϕi = hf, ϕi, ∀f ∈ Cb .
CHAPTER 5. ASYMPTOTIC DISTRIBUTIONS 94

Therefore,
hP ∗ f, ϕi = hf, ϕi, ∀f ∈ C0 . (5.14)

Since ϕ is a finite Borel measure and since each f ∈ C0 is a bounded Borel

function, together (5.13) and (5.14) imply that

hf, P ϕi = hf, ϕi, ∀f ∈ C0 . (5.15)

This says precisely that the positive linear functionals on C0 generated by

the two measures P ϕ and ϕ in the manner of (5.11) are identical. Given that
P ϕ and ϕ are finite Borel measures, and that the association (5.11) from M+
to C0∗ is one-to-one, this implies that the representing measures P ϕ and ϕ
are identical. This is equivalent to stating that ϕ satisfies (5.2).

The converse is obvious, and completes the proof of the lemma.

Chapter 6

Linearization of Stochastic
Economic Models

6.1 Introduction

Consider again a stochastic macroeconomic system with the stylized repre-

sentation
xt+1 = T (xt , εt ), t = 0, 1, . . . , (6.1)

where x is a collection of endogenous variables taking values in topological

vector space X, T is an arbitrary function with rng T ⊂ X, and (εt ) is a se-
quence of serially uncorrelated random variables. Given (6.1), the researcher
seeks to characterize dynamics in terms of the sequence (xt ). Ideally, the
distribution of xt will converge to a unique limiting distribution independent
of x0 as t → ∞. As usual, this distribution is defined to be the equilibrium
of the economy, and is a focal point for policy simulation and other analysis.

95
CHAPTER 6. LINEARIZATION OF STOCHASTIC MODELS 96

In the applied literature, a common approach to the analysis of stochastic

dynamics is via linearization. There are two standard methods. One is to
replace εt in (6.1) with its mean, solve the resulting deterministic model
for fixed points, and linearize in their vicinity by Taylor expansion. (For an
exposition, see, e.g., Farmer [18, Sections 2.3.3 and 2.3.4].) The other method
is log-linearization, which is of course only applicable when the underlying
model is log-linear. (See, e.g., Long and Plosser [42].)

However, it must be recalled at all times that the linearized system is auxiliary
to the analysis: it is valuable only to the extent that it provides insight into
the dynamic properties of the true model (6.1). In this connection, we stress
that for stochastic systems such as (6.1), it is not in general legitimate to
infer such properties as existence, uniqueness and stability of equilibrium
from similar properties as they may or may not occur in the linear version.

In this final essay we begin to address this issue by giving a formal justi-
fication for log-linearization of stochastic models. It is shown that for this
particular case parallel existence, uniqueness and stability results hold for
equilibria in the original and linearized models.

Section 6.2 formulates the problem. Section 6.3 states results. Section 6.4
gives an application to the multisector macroeconomic model of Long and
Plosser. Section 6.5 gives proofs.
CHAPTER 6. LINEARIZATION OF STOCHASTIC MODELS 97

6.2 Formulation of the Problem

In this section we briefly recall the definitions of equilibria and stability for
the model (6.1). Let topological space X be the state space for (6.1). That
is, xt ∈ X for all t, and, assuming that εt also takes values in X, the map T
satisfies T : X × X → X. Let B = B(X) be the Borel subsets of X, and let
P = P(X) be the set of probabilistic measures mapping events B ∈ B into
probabilities in [0, 1]. In other words, P is the class of countably additive,
real valued functions ν on B such that ν ≥ 0 and ν(X) = 1. As before, P is
metrized by the total variation norm.

The shocks εt in (6.1) are all drawn independently according to some fixed
distribution Ψ ∈ P. Given an initial condition x0 , a shock ε0 is drawn, and
x1 is realized according to (6.1). The process then repeats. Evidently xt is
an X-valued random variable. We denote the distribution of xt by νt ∈ P.

We use again the following conventions. The symbol 1E denotes the char-
acteristic function of E ⊂ X. The notation x ∼ ν means that random
variable x has distribution ν; for function g, the notation g n means the n-th
composition of g with itself.

By Definition 2.3, page 28, an equilibrium for (6.1) is a distribution ν∗ ∈ P

such that
Z Z 
ν∗ (B) = 1B [T (x, z)]Ψ(dz) ν∗ (dx), ∀ B ∈ B. (6.2)

The equilibrium ν∗ is unique if there exists no other measure in P satisfying

(6.2). The equilibrium is called globally stable if νt → ν∗ in the norm topology
as t → ∞ for every initial condition ν0 (i.e., x0 ∼ ν0 ).
CHAPTER 6. LINEARIZATION OF STOCHASTIC MODELS 98

6.3 Results

Consider now the case where (6.1) is log-linear. Let R++ = (0, ∞). The
general form of a stochastic log-linear (log-affine) system on Rn++ = ×ni=1 R++
is

x1,t+1 = γ1 xa1t11 × · · · × xnt

a1n
ε1t
.. ..
. . (NL)

xn,t+1 = γn xa1tn1 × · · · × xantnn εnt .

Here γ = (γi )ni=1 ∈ Rn++ , and the vector of shocks εt = (εit )ni=1 ∈ Rn++ is
assumed to be serially independent and distributed by Ψ ∈ P(Rn++ ).

Given n-dimensional vector x, it is convenient to use the abbreviations ln x

for (ln xi )ni=1 and exp x for (exp xi )ni=1 . We also use ln B to denote the set
of points ln x, x ∈ B, and exp B for all points exp x, x ∈ B. For example,
ln Rn++ = Rn .

After taking logs, the linear version of (NL) is

xt+1 = γ̂ + Axt + ε̂t , x ∈ Rn , (LV)

where γ̂ = ln γ, ε̂t = ln εt and A is the n × n matrix (aij ).

Our main result is as follows.

Proposition 6.1. The log-linear economy (NL) has a unique equilibrium

ν∗ ∈ P(Rn++ ) if and only if the linear version (LV) has a unique equilibrium
ν̂∗ ∈ P(Rn ). The equilibrium ν∗ for (NL) can be recovered from ν̂∗ by the
identity ν∗ (B) = ν̂∗ (ln B), ∀ B ∈ B(Rn++ ). The equilibrium ν∗ is globally
stable if and only if ν̂∗ is globally stable.
CHAPTER 6. LINEARIZATION OF STOCHASTIC MODELS 99

The proof is given in Section 6.5. The argument is based on establishing

topological conjugacy between the two systems.

6.4 Application

A well-know study using log-linearization is that of Long and Plosser [42,

pp. 52–54]. As an application of Proposition 6.1, in this section we complete
their analysis by verifying that the stability of the linearized system that
they discuss does in fact (under suitable conditions) imply stability of the
original model.

The model is an infinite horizon, representative agent economy with n sec-

tors. Let ct = (cit )ni=1 be time t consumption. Utility is given by u(ct ) =
Pn
i=1 θi ln cit , θi > 0. Production is according to the Cobb-Douglas technol-

ogy
ai1
yi,t+1 = `biti xi1t × · · · × xaint
in
εit , i = 1, . . . n, (6.3)

where yi is output of commodity i, `i is labor allocated to sector i, xij is

the amount of commodity j used in the production of good i, and εi is a
sector-specific shock. The vector of shocks is uncorrelated and identically
distributed over time.1

Production is assumed to be constant returns to scale. In particular,

n
X
bi , aij > 0; bi + aij = 1, i = 1, . . . , n. (6.4)
j=1

1
The objective of Long and Plosser was to generate fluctuations in time series con-
sistent with the business cycle from a general equilibrium framework and without serial
dependence in external noise.
CHAPTER 6. LINEARIZATION OF STOCHASTIC MODELS 100

The economy faces constraints

n
X
cjt + xijt ≤ yjt , j = 1, . . . , n, (6.5a)
i=1
n
X
`i ≤ L, i = 1, . . . , n. (6.5b)
i=1

The representative agent seeks a solution to

"∞ #
X
max E β t u(ct )
t=0

subject to (6.3) and (6.5), where β ∈ (0, 1) is a discount factor. The optimal
controls are

λi
xijt = aij βyjt , (6.6a)
λj
n
X −1
`it = λi bi λj b j L, (6.6b)
j=1

where λ0 = θ0 (1 − βA), A being the matrix (aij ) of output elasticities with

respect to commodity inputs [42, pp. 47–48]. Substitution of (6.6) into (6.3)
gives

a11 a1n
y1,t+1 = γ1 y1t × · · · × ynt ε1t
.. ..
. . (LP)
an1 ann
yn,t+1 = γn y1t × · · · × ynt εnt ,

where γ = (γi )ni=1 is a vector of positive constants.

Following Long and Plosser, we can convert (LP) into a linear form and
observe global stability of the latter. By virtue of Proposition 6.1, this implies
conditions under which (LP) is itself globally stable. Formally,
CHAPTER 6. LINEARIZATION OF STOCHASTIC MODELS 101

Proposition 6.2. Let ε be the vector of sectoral shocks in (LP). If, for some
norm k · k on Rn , the expectation E k ln εk is finite, then the economy (LP)
has a unique, globally stable equilibrium in P(Rn++ ).

Proof. The linear version of (LP) is

yt+1 = γ̂ + Ayt + ε̂t , y ∈ Rn , (6.7)

where, as before, γ̂ = ln γ, ε̂ = ln ε, and A = (aij ) is the matrix of in-

put/output elasticity coefficients. The linear stochastic system (6.7) is well
understood. In particular, it is known that if, for some norm k·k on Rn , E kε̂k
is finite, and, in addition, that the spectral radius of A is less than one, then
(6.7) has a unique and globally stable equilibrium distribution ν̂∗ ∈ P(Rn ).2

The spectral radius of any nonnegative matrix is less than or equal to the
maximum of the row sums [37, Theorem 7.2.1]. In the case of A, these sums
are all strictly less than one by (6.4). It follows that, under the hypothesis
E kε̂k < ∞, (6.7) has a unique, globally stable equilibrium in P(Rn ).

But then E k ln εk < ∞ implies that (LP) has a unique, globally stable
equilibrium in P(Rn++ ) by Proposition 6.1.

Remark 6.1. The hypotheses of Proposition 6.2 are satisfied if, for example,
E | ln εi | is finite for each i. This condition enforces small left- and right-hand
tails on the distributions of the sectoral shocks εi . These small tails prevent
the economy from either collapsing to zero or growing without bound.
2
See, for example, Lasota and Mackey [40, Proposition 12.7.1, Theorem 12.7.2].
CHAPTER 6. LINEARIZATION OF STOCHASTIC MODELS 102

6.5 Proofs

It remains to prove Proposition 6.1. The method is as follows. First, we recall

the notion of topological conjugacy between dynamical systems; conjugate
systems have identical dynamics. Since the notion of conjugacy is defined for
deterministic rather than stochastic systems, our next step is to convert the
log-linear and linear systems (NL) and (LV) into deterministic self-mappings
on spaces of probability measures. Finally, we show that these deterministic
versions are topologically conjugate.

Recall that by a homeomorphism is meant a bijection from one topological

space to another that is continuous and has continuous inverse. Let X and
X̂ be two metrizable spaces. Consider the two dynamical systems

xt+1 = g(xt ), g : X → X, (6.8)

x̂t+1 = ĝ(x̂t ), ĝ : X̂ → X̂. (6.9)

Suppose that, corresponding to g and ĝ, there exists a homeomorphism H

from X into X̂ such that g and ĝ commute with H in the sense that ĝ =
HgH −1 on X. Then (6.8) and (6.9) are said to be topologically conjugate.

In this case, (6.8) has a unique, globally stable equilibrium (i.e., fixed point
x∗ of g on X such that lim g t (x) → x∗ as t → ∞, ∀x ∈ X) if and only if (6.9)
has a unique, globally stable equilibrium. These results are well-known [3,
Section 3.3] and not difficult to verify. For example, if x∗ is a fixed point of
g on X, then x̂∗ = Hx∗ is a fixed point of ĝ on X̂, because g(x∗ ) = x∗ , and
therefore ĝ(x̂∗ ) = HgH −1 Hx∗ = Hg(x∗ ) = Hx∗ = x̂∗ .
CHAPTER 6. LINEARIZATION OF STOCHASTIC MODELS 103

Thus to prove Proposition 6.1, it remains only to rewrite the nonlinear and
linear systems (NL) and (LV) in the form of (6.8) and (6.9), and show that
they commute with a suitable homeomorphism.

To rewrite these systems in the deterministic form of (6.8) and (6.9), we

use the techniques of Chapter 2, Section 2.3. To aid the exposition these
arguments are presented here again, albeit rather tersely.

Consider again the generic system (T, Ψ) on X discussed in Section 6.2. Let
νt ∈ P(X) be the marginal distribution of the random variable xt , and let
νt+1 be that of xt+1 . Then νt and νt+1 are connected by the recursion
Z Z 
νt+1 (B) = 1B [T (x, z)]Ψ(dz) νt (dx), ∀ B ∈ B(X). (6.10)

To repeat, the intuition is that the right hand side of (6.10) sums the proba-
bility of the state moving to B from x in one step over all possible values of
x, weighted by the probability νt (dx) of x occurring as the current state.

Let Q be the associated Markov operator from P(X) into itself:

Z Z 
(Qν)(B) = 1B [T (x, z)]Ψ(dz) ν(dx), B ∈ B(X).

In this form, Q is sometimes called the Foias operator corresponding to (6.1).

Using Q allows (6.10) to be rewritten as νt+1 = Qνt . By recursion, if ν0 is
the initial state for the system (6.1) in the sense that x0 ∼ ν0 , then Qt ν0 is
the distribution for the state at time t, where Qt is defined by Qt = QQt−1 ,
Q1 = Q. In light of (6.2), an equilibrium for the system (6.1) is a distribution
ν∗ ∈ P such that Qν∗ = ν∗ . The equilibrium is globally stable if and only if
Qt ν0 → ν∗ as t → ∞ for all ν0 ∈ P(X) (see Definition 2.3, p. 28).
CHAPTER 6. LINEARIZATION OF STOCHASTIC MODELS 104

Denote by P and P̂ the Foias operators associated with the log-linear system
(NL) and the linear version (LV) respectively. These two systems can now
be represented as

νt+1 = P νt , P : P(Rn++ ) → P(Rn++ ), (6.11)

ν̂t+1 = P̂ ν̂t , P̂ : P(Rn ) → P(Rn ). (6.12)

Note that the pair (6.11) and (6.12) are in the same deterministic form as
(6.8) and (6.9). Thus to complete the proof of Proposition 6.1 we need to
establish a homeomorphism H from P(Rn++ ) onto P(Rn ) such that P and P̂
commute with H, in the sense that P = H −1 P̂ H on P(Rn++ ).

A suitable candidate for a homeomorphism is the map H : P(Rn++ ) → P(Rn )

defined at ν ∈ P(Rn++ ) by

(Hν)(B) = ν(exp B), B ∈ B(Rn ). (6.13)

It is not difficult to verify that H is a one-to-one correspondence from P(Rn++ )

onto P(Rn ), where if ν̂ ∈ P(Rn ), then H −1 ν̂(B) = ν̂(ln B).

Thus H is a bijection. In fact,

Lemma 6.1. The map H defined in (6.13) is a homeomorphism.

Proof. Regarding continuity of H, let νn → ν in P(Rn++ ). It is shown in

Stokey et al. [58, Theorem 11.6] that, for any metric space X, any sequence
(νn ) in P(X) and any ν ∈ P(X), the statement νn → ν in total variation
norm is equivalent to |νn (B) − ν(B)| → 0 uniformly on B(X).
CHAPTER 6. LINEARIZATION OF STOCHASTIC MODELS 105

Fix ε > 0. By hypothesis, there exists an N ∈ N such that n ≥ N implies

|νn (A) − ν(A)| < ε, ∀A ∈ B(Rn++ ).

But then

n ≥ N =⇒ |(Hνn )(B) − (Hν)(B)| = |νn (exp B) − ν(exp B)| < ε

for any B ∈ B(Rn ). Hence Hνn → Hν in P(Rn ). This proves continuity of

H. The proof of continuity of H −1 is similar.

To complete the proof of Proposition 6.1 we need show only that

Lemma 6.2. The relation P = H −1 P̂ H holds on P(Rn++ ).

Proof. Prior to the main proof we briefly recall how to integrate with respect
to the induced measure Hν [1, Theorem 12.46].

Fix ν ∈ P(Rn++ ). If h : Rn → R is any B(Rn )-measurable function that is

summable with respect to Hν ∈ P(Rn ), then Rn++ 3 x 7→ h(ln x) ∈ R is
B(Rn++ )-measurable, and
Z Z
h(x)(Hν)(dx) = h(ln x)ν(dx). (6.14)
Rn Rn
++

Also, note that the system (NL) can be expressed more compactly as

xt+1 = exp(ĉ + A ln xt + ln εt ), x ∈ Rn++ , (6.15)

We now show that

(P̂ Hν)(B) = (HP ν)(B), ∀ν ∈ P(Rn++ ), ∀B ∈ B(Rn ), (6.16)

CHAPTER 6. LINEARIZATION OF STOCHASTIC MODELS 106

which is equivalent to the statement of the lemma. Note first that ε ∼ Ψ ∈

P(Rn++ ) implies ε̂ ∼ HΨ ∈ P(Rn ), because for all B ∈ P(Rn ), Prob[ε̂ ∈ B] =
Prob[ln ε ∈ B] = Prob[ε ∈ exp B] = Ψ(exp B) = HΨ(B).

Hence
Z Z 
(P̂ Hν)(B) = 1B (ĉ + Ax + z)(HΨ)(dz) Hν(dx)
Rn Rn
Z "Z #
= 1B (ĉ + Ax + ln z)Ψ(dz) Hν(dx)
Rn Rn
++
Z "Z #
= 1B (ĉ + A ln x + ln z)Ψ(dz) ν(dx)
Rn
++ Rn
++
Z "Z #
= 1exp B [exp(ĉ + A ln x + ln z)]Ψ(dz) ν(dx),
Rn
++ Rn
++

where we have used (6.14) to change variables.

But the representation (6.15) shows that this is just HP ν(B), which proves
(6.16).
Bibliography

[1] Charalambos D. Aliprantis and Kim C. Border, Infinite dimensional

analysis, 2 ed., Springer-Verlag, Berlin, 1999.

[2] Charalambos D. Aliprantis and Owen Burkinshaw, Principles of real

analysis, 3rd ed., Academic Press, New York, 1998.

[3] Kathleen T. Alligood, Tim D. Sauer, and James A. Yorke, Chaos: An

introduction to dynamical systems, Springer-Verlag, Berlin, 1997.

[4] Rabah Amir, A new look at optimal growth under uncertainity, Journal
of Economic Dynamics and Control 22 (1997), 67–86.

[5] Rabah Amir and Igor V. Evstigneev, A functional central limit theorem
for equilibrium paths of economic dynamics, Journal of Mathematical
Economics 33 (2000), 81–99.

[6] Costas Azariadis and Allan Drazen, Threshold externalities in economic

development, Quarterly Journal of Economics 105 (1990), 501–526.

[7] Rabi N. Bhattacharya and Mukul Majumdar, On a class of stable ran-

dom dynamical systems: Theory and applications, Journal of Economic
Theory 96 (2001), 208–229.

107
BIBLIOGRAPHY 108

[8] Michael Binder and M. Hashem Pesaran, Stochastic growth models and
their econometric implications, Journal of Economic Growth 4 (1999),
139–183.

[9] William A. Brock and M. Majumdar, Global asymptotic stability results

for multisector models of optimal growth under uncertainty when future
utilities are discounted, Journal of Economic Theory 18 (1978), 225–243.

[10] William A. Brock and Leonard Mirman, Optimal economic growth and
uncertainty: The discounted case, Journal of Economic Theory 4 (1972),
no. 3, 479–513.

[11] David Cass, Optimum growth in an aggregative model of capital accu-

mulation, Review of Economic Studies 32 (1965), 233–240.

[12] Luc Devroye and Laszlo Györfi, Nonparametric density estimation: The
L1 view, John Wiley, New York, 1985.

[13] John B. Donaldson and Rajnish Mehra, Stochastic growth with correlated
production shocks, Journal of Economic Theory 29 (1983), 282–312.

[14] Darrel Duffie, John Geanakoplos, Andreu Mas-Collel, and Andrew

McLennan, Stationary Markov equilibria, Econometrica 62 (1994), no. 4,
745–781.

[15] Nelson Dunford and B. J. Pettis, Linear operations on summable func-

tions, Transactions of the American Mathematical Society 47 (1940),
no. 3, 323–392.
BIBLIOGRAPHY 109

[16] Steven Durlauf and Danny Quah, The new empirics of economic growth,
Handbook of Macroeconomics (John B. Taylor and M. Woodford, eds.),
Elsevier Science, North Holland, 1999, pp. 231–304.

[17] Igor V. Evstigneev and Sjur D. Flåm, The turnpike property and the cen-
tral limit theorem in stochastic models of economic dynamics, Statistics
and Control of Stochastic Processes (B. L. Rozovskii Yu M. Kabanov
and A. N. Shiryaev, eds.), World Scientific, London, 1997, pp. 63–101.

[18] Roger E. A. Farmer, The macroeconomics of self-fulfilling prophercies,

MIT Press, Cambridge, 1993.

[19] Jianfeng F. Feng and David Brown, Fixed-point attractor analysis for a
class of neurodynamcis, Neural Computation 10 (1998), no. 1, 189–213.

[20] Shaul R. Foguel, The ergodic theory of Markov processes, Van Nonstrad
mathematical studies, no. 21, Van Nonstrad, New York, 1969.

[21] Carl A. Futia, Invariant distributions and the limiting behavior of

Markovian economic models, Econometrica 50 (1982), no. 2, 377–408.

[22] Oded Galor and Harl E. Ryder, Existence, uniqueness and stability of
equilibrium in an overlapping-generations model with productive capital,
Journal of Economic Theory 49 (1989), 360–375.

[23] Oded Galor and J. Zeira, Income distribution and macroeconomics, Re-
view of Economic Studies 60 (1993), 35–52.

[24] Jerry R. Green and Mukul Majumdar, The nature of stochastic equilib-
ria, Econometrica 43 (1975), 647–660.
BIBLIOGRAPHY 110

[25] James D. Hamilton, Time series analysis, Princeton University Pres,

Princeton, 1994.

[26] Onésimo Hernandez-Lerma and Jean Bernard Lasserre, Discrete-time

Markov control processes: Basic optimality criteria, Springer-Verlag,
1996.

[27] , Further topics in dicrete time Markov control processes,

Springer-Verlag, 1999.

[28] Hugo A. Hopenhayn and Edward C. Prescott, Stochastic monotonicity

and stationary distributions for dynamic economies, Econometrica 60
(1992), no. 6, 1387–1406.

[29] Bernard Hopf, The general temporally discrete Markov process, J. Rat.
Mech. Anal. 3 (1954), 13–45.

[30] Katarzyna Horbacz, Asymptotic stability of dynamical systems with mul-

tiplicative perturbations, Annales Polonici Mathematici L (1989), 209–
218.

[31] Charles I. Jones, On the evolution of the world income distribution.,

Journal of Economic Perspectives 11 (1997), no. 3, 19–36.

[32] Mohan C. Joshi and Ramendra K. Bose, Some topics in nonlinear func-
tional analysis, Wiley Eastern Limited, New Delhi, 1985.

[33] Shizuo Kakutani and Kosaku Yoshida, Operator-theoretical treatment

of Markoff ’s process and mean ergodic theorem, Ann. Math. 42 (1941),
no. 1.
BIBLIOGRAPHY 111

[34] Tjalling Koopmans, On the concept of optimal economic growth, Pontif.

Acad. Sci. 28 (1965), 225–300.

[35] Nikolai Vladimirovich Krylov and Nikolai Nikolaevich Bogolioubov, Sur

les properties en chaine, C. R. Paris 204 (1937), 1386–1388.

[36] John Laitner, The steady states of a stochastic decentralized growth

model, Journal of Economic Theory 24 (1981), 377–392.

[37] Peter Lancaster, Theory of matrices, Academic Press, New York, 1969.

[38] Andrzej Lasota, Invariant principle for discrete time dynamical systems,
Universitatis Iagellonicae Acta Mathematica 31 (1994), 111–127.

[39] Andrzej Lasota and Micheal C. Mackey, Globally asyptotic properties

of proliferating cell populations, Journal of Mathematical Biology 19
(1984), 43–62.

[40] , Chaos, fractals and noise: Stochastic aspects of dynamics,

Springer-Verlag, 1994.

[41] Andrzej Lasota, Micheal C. Mackey, and Joanna Tyrcha, The statistical
dynamics of recurrent biological events, Journal of Mathematical Biology
30 (1992), 775–800.

[42] John B. Long and Charles I. Plosser, Real business cycles, Journal of
Political Economy 91 (1983), no. 1, 39–69.

[43] Mukul Majumdar and Itzhak Zilcha, Optimal growth in a stochastic

environment: Some sensitivity and turnpike results, Journal of Economic
Theory 43 (1987), 116–133.
BIBLIOGRAPHY 112

[44] Leonard J. Mirman, On the existence of steady state measures for one
sector growth models with uncertain technology, International Economic
Review 13 (1972), 271–286.

[45] , The steady state behavior of a class of one sector growth models
with uncertain technology, Journal of Economic Theory 6 (1973), 219–
242.

[46] Leonard J. Mirman and Itzhak Zilcha, On optimal growth under uncer-
tainty, Journal of Economic Theory 11 (1975), 329–339.

[47] Krysztof Loskot and Ryszard Rudnicki, Limit theorems for stochastically
perturbed dynamical systems, Journal of Applied Probability 32 (1995),
459–469.

[48] Danny Quah, Convergence empirics across economies with (some) cap-
ital mobility, Journal of Economic Growth 1 (1996), 95—124.

[49] , Twin peaks: Growth and convergence in models of distribution

dynamics, Economic Journal 106 (1996), no. 437, 1045–1055.

[50] Frank P. Ramsey, A mathematical theory of saving, Economic Journal

38 (1928), 543–559.

[51] Assaf Razin and Joseph A. Yahav, On stochastic models of economic

growth, International Economic Review 20 (1979), no. 3, 599–604.

[52] Albert N. Shiryaev, Probability, 2 ed., Springer-Verlag, New York, 1996.

BIBLIOGRAPHY 113

[53] Wojciech Slomczynski, Jaroslaw Kwapien, and Karol Zyczkowski, En-

tropy computing via integration over fractal measures, Chaos 10 (2000),
no. 1, 180–188.

[54] John Stachurski, Asymptotic stability of a Brock-Mirman economy with

unbounded shock, University of Melbourne Department of Economics
Research Paper 746 (2000).

[55] , Log-linearization of perturbed dynamical systems, with applica-

tions to optimal growth, University of Melbourne Department of Eco-
nomics Research Paper 788 (2001).

[56] , Stochastic growth: Asymptotic distributions, Economic Theory

(Forthcoming 2002).

[57] , Stochastic optimal growth with unbounded shock, Journal of Eco-

nomic Theory (Forthcoming 2002).

[58] Nancy L. Stokey, Robert E. Lucas, and Edward C. Prescott, Recursive

methods in economic dynamics, Harvard University Press, 1989.

[59] John C. Taylor, An introduction to measure and probability, Springer-

Verlag, New York, 1997.

[60] Joanna Tyrcha, Asymptotic stability in a generalized probabilis-

tic/deterministic model of the cell cycle, Journal of Mathematical Bi-
ology 26 (1988), 465–475.
BIBLIOGRAPHY 114

[61] John J. Tyson and Kenneth B. Hannsgen, Cell growth and division: A
deterministic/probabilistic model of the cell cycle, Journal of Mathemat-
ical Biology 23 (1986), 231–246.

[62] Yong Wang, Stationary equilibria in an overlapping generations economy

with stochastic production, Journal of Economic Theory 61 (1993), 423–
435.
Appendix A

Existing Conditions

We now state and prove the problem of Chapter 4, Section 4.3.3.

Let (g, ψ) be a perturbed dynamical system on R+ satisfying Assumptions

4.1 and 4.2. Horbacz [30, Theorem 1] shows that (g, ψ) has a unique and
globally stable equilibrium whenever

(i) The map g is weakly monotone increasing and continuously differen-

tiable on [0, r) 6= ∅, and g(x) ≥ b > 0 on [r, ∞);

(ii) the map g satisfies g(0) = 0 and g 0 (0) > 0;

(iii) there exist a, B ≥ 0 such that g(x) ≤ ax + B for all x ∈ R+ ;

R
(iv) the mean E(ε) = zψ(z)dz is finite and, moreover, E(ε)a < 1;

(v) there exists a λ > 0 such that E[(g 0 (0)ε)−λ ] < 1; and

(vi) the density ψ is everywhere positive on R+ .

115
APPENDIX A. EXISTING CONDITIONS 116

Proposition A.1. Conditions (i)–(vi) of Horbacz are a special case of The-

orem 4.2.

Proof. Evidently Conditions 4.2 and 4.4 of the theorem are satisfied. It
remains to verify Condition 4.1. To this end, let λ be as in (v). If we set
V (0) = ∞ and V (x) = x−λ + x for x > 0, then V is a Lyapunov function on
R+ , and
Z Z Z
−λ
V [g(x)z]ψ(z)dz = [g(x)z] ψ(z)dz + g(x)zψ(z)dz. (A.1)

Consider the first term in the sum (A.1). By (v), there exists a positive
number σ so small that
Z
[(g 0 (0) − σ)z]−λ ψ(z)dz < 1. (A.2)

By (i) and (ii), there exists a δ > 0 such that

g(x) ≥ (g 0 (0) − σ)x whenever x ∈ [0, δ). (A.3)

Combining (A.2) and (A.3) yields a γ < 1 such that

Z
[g(x)z]−λ ψ(z)dz ≤ γx−λ , ∀ x ∈ [0, δ).

Moreover, (i) implies the existence of a c > 0 such that

g(x) ≥ c whenever x ∈ [δ, ∞).

Thus, for all x ∈ R+ , we have the bound

Z
[g(x)z]−λ ψ(z)dz ≤ γx−λ + C0 , (A.4)

where γ < 1 and C0 is a finite constant.

APPENDIX A. EXISTING CONDITIONS 117

Regarding the second term in the sum (A.1), (iii) implies that
Z
g(x)zψ(z)dz ≤ E(ε)ax + C1 , x ∈ R+ , (A.5)

where C1 is a finite constant.

Combining (A.4) and (A.5) gives

Z
V [g(x)z]ψ(z)dz ≤ αV (x) + C, (A.6)

where α = max[E(ε)a, γ] < 1 and C = C0 +C1 < ∞. This confirms Condition

4.1. Hence all of the conditions of Theorem 4.2 are satisfied.