A Geometric Description of A Macroeconomic Model With A Center Manifold

A Geometric Description of a
Macroeconomic Model with a Center

Manifold
Pere Gomis-Porqueras
Àlex Haro
Barcelona Economics Working Paper Series

Working Paper nº 364
A geometric description of a macroeconomic model
with a center manifold∗
Pere Gomis-Porqueras† Àlex Haro ‡
Department of Economics Department de Matemàtica Aplicada i Anàlisi

University of Miami Universitat de Barcelona
Abstract
This paper presents a unified framework of different algorithms to numerically compute high order expan-
sions of invariant manifolds associated to a steady state of a dynamical system. The framework is inspired
in the parameterization method of Cabré, Fontich and de la Llave [7], and the semianalytical algorithms
proposed by Simó [13], and those of Gomis-Porqueras and Haro [9]. Within this methodology, one can
compute high order approximations of stable, unstable and center manifolds. In this last case the use of
high order approximations (not just linear) are crucial in understanding the dynamic properties of the
model near the steady state. To illustrate the algorithms we consider a model economy introduced by
Azariadis, Bullard and Smith [6]. Besides its intrinsic importance, this four dimensional macroeconomic
model is an ideal testing ground because it delivers steady states with stable and unstable manifolds (of
dimensions 1 or 2), and each of them has also a one dimensional center manifold. Moreover, the numerical
computations lead to a further theoretical study of the dynamical system completing some of the results
in the original paper.
JEL Classification: E4, E3.

Keywords: Invariant manifold, Center Manifold, Global Dynamics.
∗ The authors would like to thank Xavier Jarque, Dave Kelly, David Kendrick and participants of the 12th Conference on
Computational Economics and Finance held in Limassol for their comments. We also would like to thank Jim Bullard and the
anonymous referee for his helpful comments and suggestions. The source code used to compute the nonlinear manifolds can be
obtained upon request by writing to [email protected].
† Corresponding Author: Address: 5250 University Drive, Coral Gables, FL 33124-6550, U.S.A.; Phone: 1-305-284-
4742; Fax: 1-305-284-2985; E-mail: [email protected].
‡ A.H. gratefully acknowledges the financial support from the Spanish Ministry of Science grant MEC-FEDER
MTM2006-11265; and from the Generalitat of Catalonia through the Barcelona Economics program of CREA and
grant CIRIT SGR2005-01028.
1
1 Introduction
The study of macroeconomic phenomena cannot be completely understood without the analysis of the un-
derlying dynamic process. Many macroeconomic concepts rely on intertemporal trade offs, thus inherently
embracing the concepts and methodology of dynamical systems. In modeling economics we must confront
the fact that agents, whose behavior are generating the dynamics, are themselves observing and trying to
forecast the actual dynamics. In rational expectation models, one tries to derive aggregate behavior from
assumptions on tastes, technologies and market structure that are summarized by a dynamical system. In a
fully specified environment, the optimal allocation of resources is determined by invariant properties of the
resulting dynamical system. These invariant structures organize the motion of the economy as time evolves,
giving us an idea of the possible evolution of the macroeconomic observables describing the economy.
A revolutionary contribution to the theory of dynamical systems was made by Poincaré at the end of the
19th century. Before his time, the main objective was finding explicit functions that solved the laws of motion.
Poincaré’s new proposal was to look at the geometry of solutions instead of explicit solutions. This geometric
or topological approach, largely identified with the qualitative theory of differential/difference equations, aims
at understanding the asymptotic properties of the system. Given that macroeconomic models are described
by difference or differential equations, Poincaré’s approach is ideal. The area in economics that has applied
this approach the most has been economic growth and business cycle theory. For example, bifurcation theory
has been useful in proving the existence both of deterministic cycles1 as in Benhabib and Rustichini [5], and
sunspot equilibria as in Grandmont [11], and Azariadis and Guesnerie [2] among others.
Typically these macroeconomic models seldom have explicit time paths describing the temporal evolution
of macroeconomic observables even though the underlying economies are governed by explicit rules relating
the various observables. Traditionally, only the evolution near the steady state(s) has been analyzed. Un-
fortunately, ignoring non linear properties of these manifolds might not be appropriate when studying some
economic phenomena. For example, Gomis-Porqueras and Haro [10] show how non linear properties of the in-
variant manifolds of an economy with credit market frictions can be consistent with recurrent hyperinflations.
A linear or local description of the associated manifolds are not able to capture this economic phenomena
and other alternative hypothesis, like bounded rationality, are needed to explain recurrent hyperinflations.2
Another instance where is uninformative to ignore non linear aspects of invariant manifolds is when economies
have center manifolds. Under these circumstances, the study of non linear dynamics is essential because linear
approximations cannot give any sort of information regarding the underlying dynamic properties. Thus higher
order terms are required to study the dynamic predictions of the underlying observables.
The goal of this paper is to provide a unified framework for dealing with power series that can be used
to characterize non linear invariant manifolds and normal forms, giving rise to efficient algorithms. The
methods are inspired in the parameterization method of Cabré, Fontich and de la Llave [7], the semianalytical
algorithms proposed by Simó [13], and the algorithms of Gomis-Porqueras and Haro [9]. Since the power series
are just the Taylor series of the functions that parameterize the invariant manifolds or give the normal form
1 Bifurcation theory deals with a parameterized family of dynamical systems and the qualitative variations in the family if we
vary the parameters.
2 Thus, non linearities in the underlying manifolds of an economy are then consistent with some complex economic phenomena.
2
transformations, we also want to emphasize that the power series are convergent in many important cases, like
for the stable and unstable manifolds, or the case of the so called Poincaré-Dulac normal form of a linearly
attracting fixed point.3 Moreover, even in cases where the power series are divergent, like usually happens for
the expansions of center manifolds, they give significant information about the dynamics and the geometry
of the corresponding invariant objects. It is then very useful to compute high order approximations of the
invariant manifolds, to obtain accurate approximations in large regions of phase space of the system. After
all, one can only compute a finite number of terms of the power series expansions, a tedious work that can be
done very efficiently by computers, as long as efficient algorithms have been provided.
To illustrate our methodology, we consider a model economy introduced by Azariadis, Bullard and Smith
[6], ABS model henceforth. This model studies an old topic in monetary economics; i.e, whether the provision
of currency should be an activity left strictly to the government or to private agents. Apart from its intrinsic
importance, this 4-dimensional model is an ideal testing ground for our algorithm, since it has two steady
states (monetary and non monetary) with a 1-dimensional center manifold as well as stable and unstable
manifolds of dimensions 1 or 2. The numerical results lead us to obtain more theoretical results on the ABS
model. In particular, we can obtain close form solutions for the underlying dynamics on the center manifold,
proving that the dynamics is purely oscillatory, contradicting the results in [6]. As a result, we can easily
check the accuracy and performance of our algorithm. Moreover, we also see that the dynamics takes place in
a 3-dimensional manifold, reducing the dimension of the problem. We also show the importance of the slow
manifolds in shaping the predicted time series for the interest rates.4 In particular, we discover a connexion
between the slow stable and the slow unstable manifolds of the two steady states of the model. Thus, there
are economic equilibrium paths going from the monetary steady state and to the non monetary steady state.
This equilibrium then predicts hyperinflationary equilibrium paths in which money is used at all dates but the
price level tends to infinity, generating equilibrium indeterminacy. This new global result is consistent with
Friedman [8] who argued that allowing private provision of close currency substitutes is a recipe for generating
indeterminacy of equilibrium.
The paper is organized as follows. In Section 2 we present our framework that allows us to compute
the invariant manifolds of an economy. The explanations are completed in Appendices A and B. Section 3
introduces the ABS model, presents the theoretical and numerical results. The proofs of the theoretical results
are left to Appendix C. We emphasize that some of the theoretical results have been inspired by the numerical
computations. Finally, we present some conclusions in Section 4.
2 On the computation of invariant manifolds

Coming back to the Poincaré’s foundations of the theory of dynamical systems, this paper emphasizes two
very important goals: (i) the study of the invariant manifolds and (ii) the computation of normal forms. These
3 In this case, we assume that the power series describing the model are also convergent. That is, we assume that the model is
analytic around the steady state. One can also consider cases in which the model is infinitely differentiable, or even just finitely
differentiable.
4 The slow stable (resp. slow unstable) manifold is the invariant manifold inside the stable (resp. unstable) manifold associated
to the biggest (resp. smallest) eigenvalue of modulus lower (resp. greater) than 1.
3
invariant manifolds can be thought as “highways” of the dynamical system that organize the global dynamics
of the economy.5 The theory of normal forms, on the other hand, consists of transforming a dynamical system
to a simpler description which is easier to compute and study. These transformations are given by power
series around the steady state.6 The idea of using power series to approximate invariant manifolds is also very
fruitful, both from the theoretical point of view and for its numerical applications.7
In this section we present the main ideas behind the algorithm that characterizes the associated invariant
manifolds of a dynamical system. We relate the graph and the parameterization methods discussed in Gomis-
Porqueras and Haro [9] as special cases of this general framework.
The underlying idea behind the methods presented in this section is to compute the invariant manifolds non
locally by exploiting an invariance condition that analytically describes them. The first step in achieving our
goal is to obtain a local approximation of the manifold where the fixed point belongs to this local description.
The second step is to define a fundamental domain around the fixed point where the invariance equation is
satisfied with high accuracy. Finally, we iterate this domain expanding the manifold away from the steady
state by successive linear approximations which give the non linear properties of the invariant manifold.
Given that macroeconomic models are described by difference or differential equations, the following
methodology can be applied to all macroeconomic models that are described by a dynamical system. For
illustrative purposes, let us consider a discrete n-dimensional dynamical system given by:
zt+1 = F (zt ) , (1)
where z = (z1 , . . . , zn ) denote coordinates in a open set B ⊂ Rn , the phase space which is the set of states of
the dynamical system, and F : B ⊂ Rn → Rn representing a smooth map that describes the evolution of the
dynamical system.
In general, a set S ⊂ B is said to be invariant under F if z ∈ S implies F (z) ∈ S. The simplest invariant
objects of a dynamical system are the fixed points or steady states and the periodic orbits of F . In the
following sections, we consider different invariant sets given by specific parameterizations.
2.1 Parameterization of invariant manifolds

In this section we consider invariant sets described by a particular parameterization. The main observation is
that if we have smooth maps Φ : U ⊂ Rd → Rn , and f : U ⊂ Rd → Rd , where d ≤ n and U ⊂ Rd is open, such
that
F (Φ(u)) − Φ(f (u)) = 0 , (2)
then the manifold parameterized by Φ is defined by W = {Φ(u) ∈ Rn | u ∈ U ⊂ Rd }. This manifold is invariant

under the dynamical system described by equation (1).8 The coordinates on this manifold are u =(u1 , . . . , ud ).
That is to say, a point on the manifold W is parameterized by ut , zt =Φ(ut ), which is mapped onto another
point zt+1 =Φ(f (ut )) on the manifold by ut+1 =f (ut ). In other words, the dynamical process on the invariant
manifold W, which is parameterized by Φ, is described by the map f that describes the dynamics of one point
5 These theoretical “highways” have been used to design space missions in order to minimize fuel consumption.
6 See[1] for more details.
7 Both methodologies have been implemented in Celestial Mechanics, see e.g. [13].
8 We also assume that Φ is an immersion, that is the rank of its differential is d for all the points in its domain.
4
in the manifold as time evolves so that: ut+1 =f (ut ). Note that the dynamics of ut is restricted to belong to
W. Thus, we can think of f as a subsystem of F . In the special case when d = n, equation (2) suggests that
f is conjugate to F , i.e. f =Φ−1 ◦F ◦Φ.
In this paper, we consider invariant manifolds attached to a fixed point z∗ of the system given by equation
(1); i.e, F (z∗ )=z∗ . Local information regarding the underlying dynamics is given by the linearization of F
near z∗ which is given by:
vt+1 = Avt , (3)
where A = DF (z∗ ). These vectors v can be thought as small perturbations near the fixed point and equation
(3) describes their evolution.9 But this linear information is not enough for describing the global dynamics
of the system nor even for describing the local dynamics near the fixed point if this has center manifolds
(associated to eigenvalues of modulus 1).
Given an eigenspace W of dimension d, the goal is to characterize the linear and non linear components
of the associated invariant manifold W and describe the dynamics on it. Thus the problem of finding the
invariant manifolds for F is equivalent to solving the functional equation given by (2) for both Φ and f .10
This system of invariance equations is infradetermined; i.e, we have n + d unknowns, the components of the
maps Φ and f , and just n equations, the components of F ◦Φ−Φ◦f . Hence, we cannot expect a unique solution
for equation (2). Thus, we can parameterize the invariant manifold W using two basics techniques: the graph
and the parameterization approach.
In the graph method, one considers a splitting of the z variables (possibly after a linear change of coor-
dinates), say z =(x, y) where x ∈ Rd and y ∈ Rn−d . Then one tries to find a parameterization of the form
Φ(u) = (u, ψ(u)), where ψ : U → Rn−d . Hence, the invariant manifold is described (locally) as a graph given
by y = ψ(x). Notice that if we denote Fx and Fy as the x and y components of F , respectively, the invariance
equation can be rewritten as follows:
ψ(Fx (x, ψ(x))) = Fy (x, ψ(x)) .
Moreover, the dynamics on the manifold is given by f (x) = Fx (x, ψ(x)). Notice also that in this case we have
n − d equations and n − d unknowns (the components of ψ).
In the parameterization method, one fixes f and then solves equation (2) for Φ. In this case, we have n
equations and n unknowns (the components of Φ). But how do we determine f ? The main idea in [7] is to
choose f as simple as possible. For instance, in case that the manifold is attached to a steady state, we could
try to find f just as a linear mapping f (u) = A1 u. The method has the flavor of the Theory of Normal Forms,
described by Poincaré, since one tries to find a normal form for the dynamics on the invariant manifold.11
Notice that normal forms correspond to the particular case d = n.
As we can see, the parameterization method looks for an adapted parameterization of the invariant manifold,
while the graph method represents the manifold as a graph. Thus the graph method is more rigid than the
parameterization method, because one cannot consider (local) manifolds with folds, that is cannot deal with
9 The eigenspaces of A are linear manifolds invariant under equation (3).
10 In [7] several important results on existence of a variety of manifolds attached to fixed points are obtained. Some numerical
implementations and applications to macroeconomic models appear in [9] and [10].
11 See [1] for an exposition of the theory.
5
manifolds that are correspondences. On the other hand, the parameterization method is more flexible because
is able to compute invariant manifolds described by correspondences.
2.1.1 Formal computations
There are several aspects one can consider when studying the invariance equation given by (2). For example,
one can consider the known terms of F and the unknowns terms of Φ and f as multivariate power series. Thus,
as long as one is able to solve the equations formally or algebraically, one is interested in knowing if these
expansions correspond to truly Taylor series of smooth functions or even if the Taylor series are convergent.
Since these power series represent the Taylor series of the corresponding mappings Φ and f , in principle
having as many terms as possible can provide us with very accurate approximations of the invariant manifolds.
Notice that, in general, just linear approximations (that are given by first order Taylor polynomials) give poor
approximations valid in very small neighborhoods of the steady state.
Let us start now with the main ideas and introduce some notation. For the sake of simplicity we assume
that the fixed point z∗ is at the origin; i.e., F (0) = 0. The Taylor series of F around the fixed point is of the
form:
X
F (z) = Az + F [k] (z) , (4)
k≥2
where each term F [k] (z) is an n-vector whose components are homogeneous polynomials of order k in n
variables z = (z1 , . . . , zn ). These homogeneous polynomials are the normalized derivatives of order k given by:
1 k
F [k] (z) = D F (0)(z, . k. ., z) .
k!
In particular, F [1] (z) = DF (0)z = Az. We will also use the notation:
X
F [k] (z) = F` z ` ,
|`|=k
where the coefficients F` are n vectors, and the subindices are n-tuples ` = (`1 , . . . , `n ) of order k = |`| =
`1 + · · · + `n . Here we use the standard multi-index notation z ` = z1`1 . . . zn`n and also denote F [≤k] (zt ) =
Pk [i] 12
i=1 F (zt ) as the truncated Taylor series up to order k, that is the Taylor polynomial of degree k.
Thus, the unknown map Φ is an n-vector of power series in d variables, u = (u1 , . . . , ud ), describing the
invariant manifold and the unknown map f is a d-vector of power series in d variables that describes the
dynamics on the manifold. Both of these power series have zero constant terms which are given by:
X
Φ(u) = P1 u + Φ[k] (u) ,
k≥2
and
X
f (u) = A1 u + f [k] (u) ,
k≥2
where P1 = DΦ(0) is a n × d matrix and A1 = Df (0) is a d × d matrix.

The goal of this algorithm is then to compute recursively the homogeneous polynomials in Φ[k] and f [k]
from the invariance equation given by (2), F (Φ(u)) − Φ(f (u)) = 0, starting from the first order P1 and A1 . At
12 We also use the notation F [<k] as meaning F [≤k−1] .
6
first order, one obtains the equation AP1 = P1 A1 which emphasizes that the linear manifold W generated by
P1 is invariant under the linearization of F , given by A. The matrix A1 represents the linear map restricted
to such linear manifold W . So, the first order equation is a very natural one, since it emphasizes that the
manifold W is tangent to the linear space W .
In the recursive process, one assumes that has already computed all the terms (the derivatives) of the
unknowns Φ and f up to order k − 1, and wants to compute the terms of order k. If one considers the
expansions up to order k of F (Φ(u)) − Φ(f (u)) = 0 one obtains the following:
AΦ[k] (u) − Φ[k] (A1 u) − P1 f [k] (u) − R[≤k] (u) + · · · = 0 , (5)
where R[≤k] (u) are terms up to degree k coming from the known terms of Φ[<k] (u) and f [<k] (u) and the dots
represent terms of higher order. The resulting equation that we have to solve at the step k is a linear system
given by:
AΦ[k] (u) − Φ[k] (A1 u) − P1 f [k] (u) = R[k] (u) , (6)
where the unknowns are the coefficients of the homogeneous polynomials in Φ[k] (u) and f [k] (u).
The linear system given by equation (6) is known in the dynamical systems literature as the homological
equation. We observe that this system is also infradetermined. It is worth mentioning that the solution of (6)
depends only on the properties of the linearizations A and A1 , in particular of their eigenvalues. A first result
on the computation of the expansions corresponding to the invariant manifold is the following proposition.
Proposition 1 Assume that W 1 is an invariant space of A, generated by the columns of the matrix P1 , and
let A1 be the matrix such that AP1 = P1 A1 , λ1 , . . . , λd be the eigenvalues of A1 , and µ1 , . . . µn−d be the rest
of eigenvalues completing the spectrum of A.
• If λm − µj 6= 0 for all |m| ≥ 2 and j = 1, . . . , n − d, then the homological equation can be solved, and we
can compute all the terms of the expansions of Φ(u) and f (u) recursively. The solutions are not unique.
A particular parameterization is of the graph form:
Φ(u) = P1 u + P2 ψ(u) , (7)
where P2 is a n × (n − d) matrix whose columns complete those of P1 to a basis of Rn , and ψ is a

(n − d)-vector of power series.
• Moreover, if λm − λi 6= 0 for all |m| ≥ 2 and i = 1, . . . , d, then the homological equation can be solved
even when choosing f [k] (u) = 0. Then we can compute all terms of the expansions of Φ be choosing
f (u) = A1 u, so that:
F (Φ(u)) = Φ(A1 u) (8)
Proof: See Appendix A.

As we can see from this Proposition, there are just algebraic obstructions to either solving the homological
equation or reducing the dynamics on the manifold to a linear mapping. These kinds of obstructions are
known in the literature as resonances. In our context, there are two types of resonances:
7
• The primary resonances are the indices m and j, with |m| ≥ 2 and j = 1, . . . , n−d, such that λm −µj = 0.
These are the obstructions to the solution of the homological equation, and to the existence of the
invariant manifold W.
• The secondary resonances are the indices m and i, with |m| ≥ 2 and i = 1, . . . , d, such that λm − λi = 0.
These are obstructions to the linearization of the dynamics on the invariant manifold W.
These two types of resonances are very different. On the one hand, the absence of primary resonances let
us easily construct the invariant manifold. In particular, a solution is guaranteed if the eigenvalues of A1 have
modulus smaller than one (stable manifold), greater than one (unstable manifold), or equal to one (center
manifold). So, the methodology covers the standard cases.13
On the other hand, the secondary resonances are the ones that appear in the theory of normal forms.14
These are obstructions to the linearization of the dynamics on the invariant manifold, but not on its con-
struction. Notice that if the eigenvalues λi of A1 are all of them of modulus smaller (bigger) than one, the
number of possible secondary resonances is finite, and we can reduce the dynamics to a polynomial. This is
the Poincaré-Dulac normal form of the dynamics on the invariant manifold.
In summary, what we have shown so far is that in order to fully characterize the invariant manifold we
need both a representation of the manifold itself as well as the dynamics on it. The algorithm presented in
this section leads to a simplification of a finite part of the Taylor expansion of the system at the origin, by
means of suitable change of coordinates. On this simpler system, we perform the analysis to gain as much
information about the dynamics as possible. In this way we can obtain an asymptotic approximation of the
motion of the original system. We are then able to characterize non linear approximations which allows to
draw conclusions about the “true” solutions to the invariant manifold and the dynamics on it.15
2.1.2 Globalization of the invariant manifold
Once we have computed a local approximation of the invariant manifold, we have to extend or globalize it. In
order to do so, we first determine a domain for which the approximation given by the power series is accurate.
Since the proposed algorithm employs Taylor series expansions, we have an idea of the associated error of the
approximations by looking at higher order terms.
Let z = Φ̃(u) be an approximation to the invariant manifold z = Φ(u), say up to order k, and ut = f˜(ut ) an
approximation to the dynamics on the manifold ut+1 = f (ut ). Since the coefficients of the Taylor polynomials
13 But, as is emphasized in [7], there are other cases of interest. One can e.g. consider invariant manifolds corresponding to the
less contracting directions, called slow manifolds, as long as there are no primary resonances.
14 Notice that one could consider the case d = n. So, the theory of normal forms is included in this framework. We emphasize
that in the usual way this theory is presented, see e.g. [1], the recursive procedure is done by using a sequence of transformations,
instead of computing just one single transformation recursively. The methodology presented here is then more suitable for
computer implementation.
15 There are several results in the dynamical systems literature that associate “true” invariant manifolds to the expansions
given above. These results are the classical theorems of the stable/unstable manifolds and the center manifold (see e.g. [12]),
and the more recent results on non-resonant manifolds using the parameterization method [7], that include the stable/unstable
manifolds theorem. See [4] for proofs of how the coefficients increase for the case of the one dimensional center manifold of a
parabolic-hyperbolic steady state.
8
are affected by computer round off errors, we prefer writing Φ̃(u), f˜(u) instead of Φ[≤k] (u), f [≤k] to explicitly
take into account this fact.
Fixing an arbitrary small tolerance ε>0, we define the corresponding domain of validity by:
Dε = {u ∈ Rd | ||F (Φ̃(u)) − Φ̃(f˜(u))|| ≤ ε} ,
where ||·|| is a norm in Rn , for instance the supremum norm.

In order to extend the invariant manifold and capture more information regarding the non linear properties
we iterate the points of this domain. If the manifold we are calculating is attracting, we iterate using F −1 .
On the other hand, whenever it is repelling, we use F . If there are eigenvalues of modulus 1, we have a center
manifold, we have to make an analysis of the higher order terms of the manifold in order to know the stability
properties of the center manifold. In the next section we explicitly address this situation in the simplest
case. We emphasize that, even in cases where the dynamics on the center manifold is attracting or repelling
(although with a small), the slowness of the motion makes more difficult the process of globalization, so high
order local expansions are very useful in enlarging the domain of validity.
2.1.3 Stability analysis on an invariant manifold
The first order terms of a stable or unstable invariant manifold are able to give enough information to determine
if trajectories tend to or escape from the steady state. If the eigenvalues of A1 are all of moduli smaller than 1,
the manifold associated to such a subspace is attracting, and the dynamics on the manifold is asymptotically
stable to the fixed point. An important particular case is when the fixed point is attracting, with all the
eigenvalues of modulus smaller than one, and A1 is associated to the bigger eigenvalue. This corresponds to
the case of a slow manifold that captures the dynamics near the fixed point. Similar considerations can be
done if the eigenvalues of A1 have modulus bigger than one.
In the case that A1 contain eigenvalues of modulus equal to 1, existence of a center manifold, the situation
is more difficult because the first order term says nothing about the stability in the neutral directions. Even if
we want to have a very “rough” approximation of the underlying dynamics, one needs to consider non linear
terms of the invariant manifold. Thus we need to consider higher order terms.
An illuminating example is the case in which A1 is a one dimensional matrix so that A1 = 1, and the
eigenvalues of A2 are all of modulus different from 1. In this case, the graph method and the parameterization
method for computing the center manifold are equivalent, since all the terms are secondary resonances.16
Hence, the dynamics on the center manifold is of the form:
f (u) = u + f2 u2 + f3 u3 + . . .
where fN is the first non-zero term in the expansion of f . Notice that if we just consider up to first order
we have that f (u) = u and all points are fixed (at first order), a non generic situation. Thus, we need to
consider higher order terms to determine the stability of the center manifold. Let us assume that fN is the
first non-zero term of the expansion of f , so it is of the form f (u) = u + fN uN + . . . . In this formulation, the
16 If one uses a more suitable parameterization method, one can get a polynomial dynamics on the center manifold of the form
f (u) = u + a uN + b x2N −1 . This refinement appears in [3] and [4], using normal form techniques that go back to [14].
9
stability of the dynamics on the center manifold relies on the sign of the coefficient fN and the parity of N .
If N is even the manifold is attracting in one branch and repelling in the other branch. If N is odd and if
fN > 0 the center manifold is repelling and if fN < 0 the manifold is attracting.
When the one dimensional center manifold is attached to the eigenvalue −1, then using the graph method
one obtains an expression for the dynamics of the form:17
f (u) = −u + f2 u2 + f3 u3 + . . .
The analysis of the stability of the motion in the center manifold reduces to the previous case by just considering
the composition; that is:
f ◦f = f 2 (u) = u − 2(f3 + f22 )u3 + . . .
For instance, if f3 + f22 > 0 the center manifold is attracting and if f3 + f22 < 0 the center manifold is repelling.
If f3 + f22 = 0 we have to compute higher order terms to know the stability on the center manifold.
In summary, in order to determine how the system is going to evolve near the center manifold higher order
approximations regarding the dynamics on the manifold are always needed. A linear analysis is not going to
be able to determine if orbits near the center manifold are going to tend or move away from it. As a result,
a linear approximation is not going to describe the time series of the underlying macroeconomic observables
near a center manifold.
2.2 Preserved quantities and invariant manifolds

In the previous section we have considered invariant manifolds that are described using a suitable parameteri-
zation. We can also consider manifolds that are defined implicitly by a set of equations. To do so, notice that
given a smooth map H : B ⊂ Rn → Rn−m , and given by: H0 ∈ Rn−m , the level set
ΣH0 = {z ∈ B | H(z) = H0 } (9)
is an m-dimensional manifold if for all each points z ∈ ΣH0 the rank of the differential DH(z) is maximal (i.e.,
n − m).18 Then, just notice that the level set ΣH0 is invariant under the dynamical system (1) iff
H(z) = H0 ⇒ H(F (z)) = H0 .
One way of creating invariant level sets is finding preserved quantities for the dynamical system. We say
that the map H is a preserved quantity for the dynamical system given by (1) iff for all z ∈ B, H(F (z)) = H(z).
As a result, given an initial state z0 of the dynamical system, its motion zt = F t (z0 ) is restricted to the level
set of H0 = H(z0 ).
The existence of preserved quantities is useful since it can reduce the dimension of the problem. The
computation of preserved quantities for a particular system is far from being trivial, and it is very model
dependent. Historically, they have been used in continuous dynamical systems in order to look for explicit
17 If one uses the parameterization method explained in this paper, one obtains an expansion containing only odd terms.
18 This is just an straightforward application of the Implicit Function Theorem.
10
solutions when there are enough preserved quantities.19 Possibly the most important family of differential
equations with preserved quantities are the Hamiltonian systems, defined by a Hamiltonian function that is
the preserved quantity.20
The existence of preserved quantities restricts also the regions of phase space in which asymptotic invariant
manifolds can exist. For instance, the points on the stable and unstable manifolds of a steady state z∗ belong
to the level set of z∗ . This is a consequence (in fact, a particular case) of the following result.
Proposition 2 Let H : B → Rn−m be a smooth map, preserved by the dynamical system (1). Let z∗ ∈ B be
an steady state of (1). Then, for all z ∈ B s.t. lim F t (z) = z∗ or lim F t (z) = z∗ , then z ∈ ΣH∗ where
t→+∞ t→−∞
H∗ = H(z∗ ).
Proof: See Appendix B.

As we can see, finding preserved quantities in a dynamical system can substantially reduce the computa-
tional costs of obtaining the invariant manifolds.
3 Application to an economic model

In order to illustrate the methodology presented in the previous sections, we study ABS economy. The ABS
model addresses the following important monetary questions: (1) Do we “need” the government to provide
money, or can we rely on “the market” to produce a well-functioning monetary arrangement? (2) Does an
efficient monetary system require a mix of private and government money, or should the government be a
monopoly provider of currency and close currency substitutes?
Besides its intrinsic importance, the ABS economy is a four dimensional macroeconomic model and is an
ideal testing ground because it delivers steady states with stable and unstable manifolds (of dimensions 1
or 2), and each of them has also a one dimensional center manifold. Moreover, the economic model can be
transformed so that we can obtain close form solutions for the underlying dynamics on the center manifold.
As a result, we can easily check the accuracy and performance of our algorithm.
Let us first briefly describe the ABS economic environment. The economy is inhabited by heterogeneous
agents with spatial separation and limited communication where privately-issued liabilities may circulate,
either by themselves, or alongside a stock of outside money. These agents in the ABS economy live for three
periods and inhabit in two different locations. All young generations are identical in size and composition.
In each period agents are endowed with a single non storable consumption good, ej for lenders and wj for
borrowers where the index j denotes the age of the endowment good. For simplicity, ABS assume that
e1 = e>0, e2 =e3 =0, w2 = w>0 and w1 =w3 =0, and that w < e. Furthermore, lenders and borrowers care
only about young and middle aged consumptions.
19 For instance, the existence of enough preserved quantities in the Two Body Problem about the motion of two particles which
interact through gravitation, such as the linear momentum, the angular momentum and the energy, is crucial to integrate the
equations and solve the problem and, in particular, derive mathematically the Kepler’s laws of motion of the planets around the
Sun that were found by observation.
20 Hamiltonian systems appear in Geometrical Optics and Analytical Mechanics.
11
In each period, in location 1 there are N/2 young lenders and γN young borrowers, where γ ∈ [1/2, 1). In
location 2 there are N/2 lenders and (1 − γ)N borrowers.21 Furthermore, some lenders in this economy may
move, lenders who are born in location 1 (2) move to location 2 (1) when middle aged and then remain in the
new location until they are old. On the other hand, borrowers spend their entire life in the same location, thus
borrowers and lenders that interact when young will never meet again. The ABS framework also considers
spatial separation and limited communication across locations. These frictions then imply that trade can only
occur between agents who are in direct contact with each other. In particular, young borrowers in any location
would like to acquire resources from young lenders.22
The economic environment in ABS allows then the possibility of having two kind of liabilities: one period
and two period maturities. At each date there will be newly issued liabilities of each type in both locations.
In particular, trade must occur as follows: young borrowers acquire resources at time t from young lenders in
exchange for claims. Lenders take these claims to their next location at t + 1 to the next generation of lenders
in exchange for goods. At t + 2 these now newly middle aged lenders bring the claims to the original issuers.
But how do borrowers redeem these two period claims when old when they have no old age endowment? Well
they must acquire claims on resources at t+1 when they are middle aged. Thus transactions mediated through
circulation liabilities require middle aged borrowers to save in order to redeem their two period circulating
liabilities. These features of the environment force intertemporal trading to be intermediated in part by private
liabilities issued by borrowers. These are taken to another location and exchanged for goods before they are
brought back to the original issuer for redemption.
Summarizing, trade in the ABS model requires that young borrowers issue some long maturity and some
short term liabilities which coexist with a constant stock of fiat money. This initial stock of money, which is
constant over time, is held by middle aged borrowers. Long term maturity liabilities are sold to young lenders
who take them elsewhere and trade them. Short term liabilities are sold to middle aged borrowers who acquire
them as a method of honoring their own liability issues or they can hold fiat currency between periods. Short
maturity liabilities do not circulate. Note that lenders can hold circulating liabilities of the borrowers and in
addition can hold government issued fiat currency. As we can see, fiat money can be viewed as a substitute
for the private liabilities issued by borrowers.23
3.1 The dynamical system

In order to obtain the underlying dynamical system of the ABS economy, one needs to be a bit more specific
regarding tastes as well as market structure in order to derive the aggregate behavior of the economy. The ABS
model assumes logarithmic preferences and perfect competition yielding the following 4-dimensional dynamical
21 Note that by setting γ≥1/2 location 1 is relatively large and the “excess demand” for credit in location 1 is relatively large
at any rate of interest.
22 In the absence of spatial separation and limited communication, the third period would not involve economic activity.
23 For further details we refer the reader to the original ABS article.
12
system:
1 p1,t
R(1,1),t+1 = = (10)
1 + α − αR(2,1),t p1,t+1
1 p2,t
R(2,2),t+1 = = (11)
1 + β − βR(1,2),t p2,t+1
R(2,2),t p1,t
R(1,2),t+1 = R(2,2),t+1 · = (12)
R(2,1),t p2,t+1
R(1,1),t p2,t
R(2,1),t+1 = R(1,1),t+1 · = (13)
R(1,2),t p1,t+1
where pi,t represents the price level for agents in location i at time t, R(i,j),t denotes the gross one period rate
of return for the one period liabilities for agents moving from location i at time t to location j at t + 1, and
e e 1
the parameters α and β are defined as follows: α = 2γw and β = 2(1−γ)w . We recall that 2 ≤ γ < 1 and
w
e > w, so we have α ≤ β and e < 1.
As we can see, the consumption and saving behavior of agents are fully characterized once the different
gross one period rate of returns are known. The evolution of the ABS economy is further restricted by the
fact that the two locations have different prices and there are no-arbitrage opportunities. Thus the rates of
returns on the liabilities are related as follows:
R(1,2),t+1 R(2,2),t R(2,1),t+1 R(1,1),t
= , = . (14)
R(2,2),t+1 R(2,1),t R(1,1),t+1 R(1,2),t
These dynamic relations between rates of return across locations impose further restrictions on the temporal
evolution of the economic observables predicted by the ABS model economy. In order to characterize the
asymptotic properties of the ABS economy, we have to characterize the steady states of the economy and their
associated manifolds.
3.2 Steady State Equilibria

In order to find the fixed points of ABS we have to restrict the gross rates of return so that R(i,j),t =R(i,j),t+1
for a given i, j and for all t. The resulting steady states (R(1,1) , R(2,2) , R(1,2) , R(2,1) ) satisfy the following
relations:
R(1,2) R(2,1) = (R(1,1) )2 = (R(2,2) )2 . (15)
which explicitly relates the gross one period rate of return for agents moving from location 1 at time t to
location 2 at t + 1 to the returns for agents that do not move across locations. In the original paper, ABS
characterized the steady states, and also analyzed their linear stability. We state below their findings in
Propositions 3 and 4, and we present the proofs in Appendix C.
Proposition 3 The ABS economy has two steady states:

(i) A monetary steady state with a common rate return such that
R(1,1) = R(2,2) = R(1,2) = R(2,1) = 1 ,
and a common price level in both locations;
13
(ii) a non monetary steady state with rates of return

1 1 1 1
R(1,1) = R(2,2) = R∗ , R(1,2) = 1+β− , R(2,1) = 1+α− ,
β R∗ α R∗
where R∗ is the only root of the cubic equation
αβR3 + αβR2 − (1 + α + β)R + 1 = 0 (16)
1 w
that is bigger than 1+α . Moreover: e ≤ R∗ < 1.
Proof: See Appendix C.
As we can see, Proposition 3 highlights the fact that the non monetary steady state corresponds to R(1,1) =

1
R(2,2) = R∗ , with max 1+α , we ≤ R∗ < 1. Furthermore, notice that 1+α 1
< we if and only if 1 < we + 2γ
1
.
1 e
Moreover, in the symmetric case γ = 2, α = β = w, the non monetary steady state is given by R(1,1) =
w w 1
R(2,2) = R(1,2) = R(2,1) = e > w+e = 1+α .
The results on the linear stability of both monetary and non monetary steady states are summarized by
the following proposition.
Proposition 4 The steady states of the ABS economy satisfy:
• The four eigenvalues of the Jacobian in the monetary steady state corresponding to R = 1 are real.
Moreover, it has one central eigenvalue λc = −1, one stable eigenvalue λs ∈] − 1, 0[, and two unstable
eigenvalues λu,1 < λu,2 ∈]1, ∞[. Moreover, λu,1 = 1/R∗ .
• The four eigenvalues of the Jacobian in the non monetary steady state corresponding to R = R∗ are real.
Moreover, it has one central eigenvalue λc = −1, two stable eigenvalues λs,1 ∈] − 1, 0[ and λs,2 ∈]0, 1[
and one unstable eigenvalue λu ∈]1, ∞[. Moreover, λs,2 = R∗ .

As we can see, regardless of the specific calibration both monetary and non monetary steady states have
a center manifold since −1 is one of the eigenvalues. Moreover, Proposition 4 emphasizes that the monetary
steady state has a 1-dimensional center manifold, a 1-dimensional stable manifold and a 2-dimensional unstable
manifold. Similarly, the non monetary steady state has a 1-dimensional center manifold, a 2-dimensional stable
manifold and a 1-dimensional unstable manifold. The underlying dynamics of the macroeconomic observables
of the ABS system are affected by the properties of these invariant manifolds. A local analysis will not able to
determine how points near the center manifold evolve over time. In order to do so we need to consider higher
order terms of the associated manifolds.
3.3 Invariant manifolds: formal expansions

Now that we have identified the steady states of the ABS system we would like to characterize the associated
manifolds. The underlying manifolds can give us an idea of the potential predicted time series as we move
away from the steady states. In this section we compute high order approximations of the invariant manifolds
associated to the two steady states of the ABS system.
14
To be more specific, we fix the parameters so that e =1, w =0.3 and γ =0.6 as in the original ABS model.24
The corresponding coordinates of the monetary steady state are zm =(1.000000, 1.000000, 1.000000, 1.000000)
which correspond to the different interest rates. The associated eigenvalues are λs =-0.6679793, λc =-1.000000,
λu,1 =3.203277, λu,2 =5.409147. Notice that if λm
u,1 = λu,2 , then m = 1.450033, that is not an integer number
greater or equal than 2. As a result, there are no secondary resonances on the unstable manifold and we can
reduce the dynamics on such a manifold to a linear dynamics given by the diagonal matrix diag(λu,1 , λu,2 ).
Moreover, apart from the classical fast (or strong) unstable manifold associated to λu,2 , we can also construct
a slow unstable manifold associated to λu,1 . We emphasize that for parameters close to those given above the
dynamics on the unstable manifold is also linearizable, and it also contains fast and strong unstable manifolds.
The coordinates of the non monetary steady state are zn =(0.3121803, 0.3121803, 0.4712136, 0.2068204)
which correspond to the different interest rates. The associated eigenvalues are λs,1 =-0.2085300, λs,2 =0.3121803,
λc =-1.000000, λu = 1.688629. Since there are no resonances between the stable eigenvalues, that is there is
no an integer m ≥ 2 such that λm
s,2 = λs,1 , the dynamics on the stable manifold is linearizable. Moreover,
we can also associate fast stable and slow stable manifolds of dimension 1 to the eigenvalues λs,1 and λs,2 ,
respectively. Notice also that the dynamics on the fast stable manifold converges oscillatorially to the non
monetary steady state.
Both steady states have one dimensional center manifolds, associated to the eigenvalue λc = −1. Thus
orbits in the center manifold oscillate around the corresponding steady state, but we do not know their
asymptotic behavior; i.e, if they converge or not to the steady state. In order to determine the behavior of
the motion on the center manifold, a study of higher order terms is mandatory.
Table 1 displays the coefficient of the Taylor series up to 10th order both for the dynamics and parameter-
ization of the stable manifold of the monetary steady state.25 Table 2 shows the corresponding results for the
center manifold.26
Since we use the parameterization method, the numerical computation produces linear dynamics on the
stable manifold; i.e., f (u) = λs u, which is reflected in the second column of Table 1. In the last two columns
of Table 1 we give an idea of the accuracy of the expansions. The column labeled “domain” corresponds to the
validity domain of the expansion up to order k with a tolerance ε = 10−6 . “Length” represents the length of
the corresponding curve; i.e., the image under Φ[≤k] of the validity domain. Notice that both the interval and
the length grow as we increase k. All computations are obtained in double precision arithmetics. Moreover, the
Taylor series coefficients corresponding to the different manifolds reported in the different Tables are obtained
using the parameterization method.
Table 2 presents the coefficients of the expansions corresponding to the center manifold for the monetary
steady state. Since we use the parameterization method, in principle the expansion of the dynamics of the
center manifold is of the form: f (u) = −u + f [3] u3 + f [5] u5 + . . . , that is only has odd order terms. Our
numerical results yield coefficients f [3] ,f [5] ,..., smaller than 10−18 , below the round-off error of the computer.
24 The dynamical properties associated with this example are not exclusive to these particular parameters, and the methodology
can be applied to other calibrations.
25 Even though we have computed the one and two dimensional invariant manifolds of the monetary and non monetary steady
states up to order 30.

26 We emphasize that all the computation of the expansions have been carried out in less than one second using nowadays
computers, thanks to the high performance of our implementations.
15
[k] [k] [k] [k]
k f [k] Φ1 Φ2 Φ3 Φ4 domain length
0 0.000000e+00 0.000000e+00 0.000000e+00 0.000000e+00 0.000000e+00
1 -6.679793e-01 1.000000e+00 0.000000e+00 0.000000e+00 0.000000e+00 [-1.2500e-03,1.2500e-03] 2.5000e-03
2 0.000000e+00 -3.462864e-02 2.681350e-01 2.602246e-02 9.633629e-03 [-3.2500e-02,3.2500e-02] 6.5004e-02
3 0.000000e+00 7.624036e-02 4.967390e-03 4.820844e-04 1.784698e-04 [-7.9375e-02,7.9375e-02] 1.5888e-01
4 0.000000e+00 -1.249909e-03 2.070675e-02 2.009587e-03 7.439578e-04 [-1.9875e-01,1.9938e-01] 4.0016e-01
5 0.000000e+00 5.838347e-03 7.655094e-04 7.429256e-05 2.750343e-05 [-2.8500e-01,2.8875e-01] 5.7989e-01
6 0.000000e+00 1.206414e-05 1.606155e-03 1.558771e-04 5.770639e-05 [-4.4188e-01,4.4750e-01] 9.1231e-01
7 0.000000e+00 4.490866e-04 8.860896e-05 8.599485e-06 3.183567e-06 [-5.3375e-01,5.4563e-01] 1.1207e+00
8 0.000000e+00 9.247159e-06 1.251258e-04 1.214344e-05 4.495554e-06 [-6.9063e-01,7.0563e-01] 1.4865e+00
9 0.000000e+00 3.469793e-05 9.130467e-06 8.861103e-07 3.280419e-07 [-7.7188e-01,7.9625e-01] 1.6972e+00
10 0.000000e+00 1.353743e-06 9.789055e-06 9.500261e-07 3.517038e-07 [-9.1500e-01,9.4563e-01] 2.0792e+00
Table 1: Coefficients corresponding to the stable manifold of the monetary steady state.
Similar results are obtained when computing the expansions of the center manifold for the non monetary
steady state.
[k] [k] [k] [k]
k f [k] Φ1 Φ2 Φ3 Φ4 domain length
0 0.000000e+00 0.000000e+00 0.000000e+00 0.000000e+00 0.000000e+00
1 -1.000000e-00 0.000000e+00 1.000000e+00 0.000000e+00 0.000000e+00 [0.0000e+00,0.0000e+00] 0.0000e+00
2 0.000000e+00 8.000741e-02 2.479119e-01 -1.022280e-01 -2.534513e-02 [-6.8750e-03,6.8750e-03] 1.3750e-02
3 -1.084202e-19 -1.215280e-02 0.000000e+00 -6.413486e-02 -2.313605e-02 [-3.0000e-02,3.0000e-02] 5.9999e-02
4 0.000000e+00 -2.243779e-02 -4.342071e-02 -5.396960e-02 -1.596927e-02 [-6.2500e-02,6.2500e-02] 1.2499e-01
5 -5.590417e-20 1.490997e-03 0.000000e+00 -2.407681e-02 -8.927910e-03 [-1.0875e-01,1.0938e-01] 2.1809e-01
6 0.000000e+00 9.379552e-03 -6.533255e-04 -2.654547e-02 -7.096582e-03 [-1.4875e-01,1.4875e-01] 2.9733e-01
7 -2.879912e-20 -1.431981e-03 0.000000e+00 -1.460054e-02 -5.535222e-03 [-1.9938e-01,2.0000e-01] 3.9901e-01
8 0.000000e+00 -2.383643e-03 -7.533388e-03 -1.774142e-02 -4.924297e-03 [-2.3875e-01,2.3813e-01] 4.7611e-01
9 -2.202286e-20 -1.753388e-05 0.000000e+00 -9.535741e-03 -3.649439e-03 [-2.8438e-01,2.8563e-01] 5.6893e-01
10 0.000000e+00 1.720851e-03 -2.636964e-03 -1.256799e-02 -3.372125e-03 [-3.1813e-01,3.1688e-01] 6.3316e-01
Table 2: Coefficients corresponding to the center manifold of the monetary steady state.
The numerical analysis then suggests that the dynamics on both center manifolds is of the form f (u) = −u;
i.e, the dynamics is purely oscillatory. Thus, orbits on the center manifold are two periodic.27 This fact is
rigorously proved in Proposition 6.
3.4 The 2-periodic system and dynamics on the center manifolds

The previous numerical results suggest that the dynamics on both center manifolds of the two steady states
are 2-periodic. In other words, there are curves (the center manifolds) of 2 periodic points. In this section we
prove that this insight is correct. Before proving this fact, we also derive some interesting properties of the
ABS model.
27 In their original paper, Azariadis, Bullard and Smith (2001) studied numerically the stability of the dynamics along the center
manifolds for 1000 economies chosen randomly in e = 1, w ∈]0, 1[, γ ∈ [ 12 , 1[ for both steady states. The authors claim that for
all the cases they studied, the dynamics on the center manifold of the monetary is asymptotically stable. In particular, they
claim that trajectories on the center manifold approach the steady state with a oscillatory motion, and the oscillations may only
dampen very slowly over the time. This result is, as we will see, inaccurate. The dynamics on the center manifold is stable, but
not asymptotically stable. In fact, it is purely oscillatory both for the monetary and non monetary steady states.
16
Let us start by considering the 2-period ABS system, the ABS2 system henceforth, which is given by:
1
R(1,1),t+2 = R(1,1),t
, (17)
α
1+α− 1+α−αR(2,1),t · R(1,2),t
1
R(2,2),t+2 = β R(2,2),t
, (18)
1+β− 1+β−βR(1,2),t · R(2,1),t
1 + α − αR(2,1),t R(1,2),t
R(1,2),t+2 = R(2,2),t+2 · · , (19)
1 + β − βR(1,2),t R(1,1),t
1 + β − βR(1,2),t R(2,1),t
R(2,1),t+2 = R(1,1),t+2 · · . (20)
1 + α − αR(2,1)),t R(2,2),t
The fixed points of the ABS2 system correspond to 2-periodic orbits of the ABS system. We now prove
that the ABS2 has a preserved quantity. In other words, the ABS2 system has several 3-dimensional invariant
manifolds which are the levels sets of the preserved quantity. Among them, there is one 3-dimensional manifold
that is in fact invariant under the action of the ABS system.
R
(1,1) (2,2) R
Proposition 5 The 2-period value of Q = R(1,2) R(2,1) is constant along the evolution of the economy. That is
to say, for each positive value of a constant s, the 3-dimensional manifold Σs defined implicitly by:
R(1,1) R(2,2)
=s,
R(1,2) R(2,1)
is invariant under the ABS2 system. Moreover, if for a given time t we have that
R(1,1),t R(2,2),t
=s
R(1,2),t R(2,1),t
then
R(1,1),t+1 R(2,2),t+1 1
= .
R(1,2),t+1 R(2,1),t+1 s
The manifold Σs is mapped onto Σs−1 under the action of ABS system. In particular, Σ1 is invariant
under the ABS system.

R R
Note that Q = R(1,1) (2,2)
(1,2) R(2,1)
is a measure indicating if the arbitrage condition is satisfied in the ABS model or
not. Given the underlying assumptions of the ABS model, economies that satisfy Q 6=1 are not an equilibrium
since the no arbitrage condition does not hold. The only relevant preserved quantity consistent with the ABS
environment is then when Q = 1 which yields a 3-dimensional invariant manifold Σ1 .
The steady states of the ABS2 system correspond to the 2-periodic states of the ABS system that satisfy
the following equations:
1
R(1,1) = R(1,1)
(21)
α
1+α− 1+α−αR(2,1) · R(1,2)
1
R(2,2) = β R(2,2)
(22)
1+β− 1+β−βR(1,2) · R(2,1)
1 + α − αR(2,1) R(1,2)
R(1,2) = R(2,2) · · (23)
1 + β − βR(1,2) R(1,1)
1 + β − βR(1,2) R(2,1)
R(2,1) = R(1,1) · · . (24)
1 + α − αR(2,1)) R(2,2)
17
Note that the equations (23) and (24) are both equivalent to the following equation:

R(1,1) 1 + β − βR(1,2) = R(2,2) 1 + α − αR(2,1) (25)
which reemphasizes the fact there is a relation between the different gross interest rates across locations that
are preserved every two periods. Note then that we have four unknowns and three equations (21), (22) and
(25), that (possibly) define a 1-dimensional object in the 4-dimensional space, formed by the 2-periodic orbits
of the ABS system. In order to determine a solution, we need an extra condition. A natural one is to fix a
level set Σ1 , that is invariant under the ABS2 system. So, on each level set Σ1 we will look for 2-periodic
orbits of the ABS system as suggested by the following proposition.
Proposition 6 The 2-periodic orbits in the manifold Σs are obtained by solving the following equations

αR(2,1) s βR(1,2) s
(1 + α)(1 + β)R(1,2) R(2,1) s = 1 + 1+ , (26)
1 + β − βR(1,2) 1 + α − αR(2,1)
(1 + β)(1 + β − βR(1,2) + αsR(2,1) ) = (1 + α)(1 + α − αR(2,1) + βsR(1,2) ) , (27)
for R(1,2) and R(2,1) . The explicit relations are given by:

1 αR(2,1) s
R(1,1) = 1+ , (28)
1+α 1 + β − βR(1,2)

1 βR(1,2) s
R(2,2) = 1+ . (29)
1+β 1 + α − αR(2,1)
As we can see, by eliminating, for instance, R(2,1) from (26) and (27), we obtain a quartic equation for
R(1,2) . This allows us to have an explicit solution for the gross interest rate for agents moving from location 1
at time t to location 2 at t + 1. Moreover, Proposition 6 shows that the ABS system has curves of 2-periodic
points, parameterized by s. These include of course the monetary and non monetary states, that correspond
to s = 1. This is a non generic feature of the ABS system.28 As a result, the center manifolds of both
steady states are contained in these curves, and the dynamics on the center manifolds is purely oscillatory.
Moreover, the center manifolds are not relevant in the ABS model under no-arbitrage conditions, since they
are transversal to the no-arbitrage manifold Σ1 .
3.5 The reduced system

So far we have shown that by considering the ABS2 system we are able to find a preserved quantity that can
help us simplify the study of its evolution. Moreover, we have seen that the level set Σ1 is invariant under the
ABS system, and contains the monetary and non monetary steady states and their corresponding stable and
29
unstable manifolds. More importantly, given the relationship between rates of return and the no-arbitrage
conditions, we have that the initial states of the economy satisfy the following:
p(1),0 p(2),0
R(1,2),0 = R(2,2),0 , R(2,1),0 = R(1,1),0 ,
p(2),0 p(1),0
28 Notice that this result comes from the fact that the four equations that determine the periodic orbits can be reduced to three
equations.
29 See Proposition 2.
18
for given initial price levels p(1),0 , p(2),0 . Hence,
R(1,1),0 R(2,2),0
=1,
R(1,2),0 R(2,1),0
and the initial states belong to the level set Σ1 , and as a consequence the rates of return evolve on Σ1 , that is
R(1,1),t R(2,2),t
=1,
R(1,2),t R(2,1),t
for all t.
In summary, the evolution of the model under the no-arbitrage conditions takes place in the 3-dimensional
manifold Σ1 , thus reducing the dimension of the system. We refer to this invariant manifold Σ1 as the
no-arbitrage manifold.
As a result, we just need to consider three coordinates to represent the dynamics of the ABS model in Σ1 .
For instance, we can focus on (R(1,1) , R(2,2) , R(1,2) ). Moreover, since the stable manifold of the non monetary
steady state and the unstable manifold of the monetary steady state are 2-dimensional in a 3-dimensional
space, these manifolds separate different types of dynamical behavior of the system.30 In particular, if these
manifolds intersect each other, the intersections would be generically 1-dimensional curves.31 In the following
section we pursue this geometric study.
3.6 Invariant manifolds: globalization

As we have demonstrated in the previous section, the dynamics on the center manifolds is purely periodic.
This particular feature of the ABS model is ideal because one can easily check the accuracy and performance
of our algorithm. In fact, we emphasize that the rigorous results were inspired by the numerical results. In our
numerical computation, we find that all the coefficients corresponding to the dynamics of the center manifold
up to 100th order are zero within the numerical accuracy.
We have also seen that, under no-arbitrage conditions of the economy, the relevant invariant object is the
3-dimensional manifold Σ1 . The center manifolds are transversal to Σ1 and do not affect the dynamics in
Σ1 . Moreover, the monetary and non monetary steady states and their corresponding asymptotic manifolds
(stable and unstable, fast and slow), are contained in Σ1 . Thus, the transition dynamics, which always lie on
Σ1 , is influenced by the stable and unstable manifolds. The dynamics on the 1-dimensional and 2-dimensional
non resonant manifolds (stable, unstable, fast and slow) can be reduced to a linear form, even if the manifolds
themselves are not linear at all.
The analysis of the non linear invariant manifolds of the ABS economy is summarized in Figure 1 and 2.
In Figure 1 we depict all the relevant manifolds of the ABS model inside the 3-dimensional manifold Σ1 , using
the coordinates R(1,1) , R(2,2) and R(1,2) .
The monetary steady state, labeled with m, has a 2 dimensional unstable manifold, colored with red, and a
1 dimensional stable manifold, colored with blue. The slow and fast unstable 1 dimensional manifolds inside
30 This is thanks to the existence of preserved quantities. In general, these manifolds would be 2-dimensional in a 4-dimensional
space.
31 This is, again, due to the existence of preserved quantities. In general, 2-dimensional manifolds in a 4-dimensional space
intersect in O-dimensional points.
19
Figure 1: Invariant manifolds of the ABS model.
the unstable manifold are designated with single and double arrows, respectively. The non monetary steady
state, labeled with n, has a 2 dimensional stable manifold, colored with blue, and a 1 dimensional unstable
manifolds, colored with red. The slow and fast stable 1 dimensional manifolds inside the stable manifold are
designated with single and double arrows, respectively.
As in ABS, the coexistence of publicly and privately issued liabilities is not a source of indeterminacy
only when we are near the monetary steady state. Indeterminacies do arise in a neighborhood of the non
monetary steady state. The global analysis of the invariant manifolds of the ABS economy predicts a new
source for equilibrium indeterminacy previously not found by the original paper. In particular, as we can see
from Figure 1, the unstable manifold of the monetary steady and the stable manifold of the non monetary
steady state intersect in a curve. The points on the curve are known as heteroclinic points. These points move
under the evolution of the economy from the monetary steady state to the non monetary steady state. An
interesting fact of the ABS model is that this heteroclinic connection between the steady states is constituted
by the slow stable manifold of the non monetary steady state and the slow unstable manifold of the monetary
steady state. In other words, the 1-dimensional slow manifolds coincide. This is what is observed in Figure 1.
As a result, the predicted time series are such that the interest rates are going to decrease over time, generating
a new source for equilibrium indeterminacy. To gain more insight, Figure 2 displays the actual time series of
the four rates of return of an heteroclinic point.
The rates of return associated with this heteroclinic point are such that these returns move from those
in the monetary steady state to those in the non monetary steady state. Thus, heteroclinic points predict
hyperinflationary equilibrium paths in which money is used at all dates but the price level tends to infinity.
The predicted time series of an heteroclinic point cannot be generated when performing a local analysis.
20
Figure 2: Time series of the rates of return of an heteroclinic point.
When performing a local analysis, hyperinflationary equilibrium paths can only be generated for a very small
set of initial conditions; i.e., only those near the non monetary steady state. On the other hand, when we
consider a global analysis we observe that there is a much broader set of initial conditions that are consistent
with hyperinflationary equilibrium paths. The existence of these heteroclinic points allows for an economy to
transition from points near a monetary steady state to those in the non monetary steady state, generating
indeterminacy of equilibrium. Moreover, these paths are not oscillatory, thus not generating “excess” volatility.
As opposed to the linear results of ABS economy, a global analysis supports the idea that the use of private
liabilities is conducive to indeterminacy of equilibrium when valued outside liabilities are present. This new
global result is consistent with Friedman [8] who argued that allowing private provision of close currency
substitutes is a recipe for generating indeterminacy of equilibrium.
It is apparent from these numerical explorations that once we contemplate the non linear properties of
the dynamical system, the corresponding phase space can become quite complicated because of the possible
intersections of the stable and unstable manifolds. It is thanks to the non linearities of these manifolds that
we can capture new dynamical phenomena not observed when performing a linear analysis.
4 Conclusions
Macroeconomic models are in many cases described by difference or differential equations explicitly defining
a dynamical system. In this paper we have presented a unified framework that helps us characterize the
associated invariant manifolds of a given dynamical system. In particular, we are able to relate the graph and
the parameterization approach as special cases of this general framework. While in the graph method one
tries to simplify the local representation of the manifold, in the parameterization method one tries to simplify
the local representation of its dynamics.
To illustrate our methodology, we consider a model economy introduced by Azariadis, Bullard and Smith
[6] which studies whether the provision of currency should be an activity left strictly to the government or
to private agents. We analyze the stability properties of this economy that has a center, stable and unstable
manifolds. This model economy is an ideal testing ground for our algorithm because it delivers a center
21
manifold which requires a global analysis. Moreover, the economic model can be transformed so that we can
obtain close form solutions for the underlying dynamics on the center manifold. As a result, we can easily
check the accuracy and performance of our algorithm.
Finally, we implement our algorithm and compute the corresponding high order approximations of the
associated invariant manifolds of the Azariadis, Bullard and Smith [6] model economy. These non linear in-
variant manifolds are able to generate new predicted time series that cannot be detected when performing a
local analysis. In particular, we show the existence of heteroclinic points that predict hyperinflationary equi-
librium paths of economies transitioning from the monetary steady state to the non monetary one, generating
indeterminacy of equilibrium. This new global result is consistent with Friedman [8] who argued that allowing
private provision of close currency substitutes is a recipe for generating indeterminacy of equilibrium.
References
[1] V. I. Arnol0 d. Geometrical methods in the theory of ordinary differential equations, volume 250 of
Grundlehren der Mathematischen Wissenschaften [Fundamental Principles of Mathematical Science].
Springer-Verlag, New York, 1983. Translated from the Russian by Joseph Szücs, Translation edited
by Mark Levi.
[2] C. Azariadis and R. Guesnerie. Sunspots and cycles. Review of Economic Studies, 53:725–737, 1986.
[3] Inmaculada Baldomá, Ernest Fontich, Rafael de la Llave, and Pau Martı́n. The parameterization method
for one-dimensional invariant manifolds of higher dimensional parabolic fixed points. Discrete Contin.
Dyn. Syst., 17(4):835–865, 2007.
[4] Inmaculada Baldomá and Àlex Haro. One dimensional invariant manifolds of gevrey type in real-analytic
maps. Discrete Contin. Dyn. Syst., to appear.
[5] J. Benhabib and A. Rustichini. Growth and indeterminacy in dynamic models with externalities. Econo-
metrica, 62:323–342, 1994.
[6] J. Bullard C. Azariadis and B. Smith. Public and private circulating liabilities. Journal of Economic
Theory, 99:59–116, 2001.
[7] Xavier Cabré, Ernest Fontich, and Rafael de la Llave. The parameterization method for invariant mani-
folds. I. Manifolds associated to non-resonant subspaces. Indiana Univ. Math. J., 52(2):283–328, 2003.
[8] M. Friedman. A Program for Monetary Stability. Fordham University Press. New York, 1960.
[9] P. Gomis-Porqueras and A. Haro. Global dynamics in macroeconomics: An overlapping generations

example. Journal of Economic Dynamics and Control, 27:1941–1959, 2003.
[10] P. Gomis-Porqueras and A. Haro. Global bifurcations, credit rationing and recurrent hyperinflations.
Journal of Economic Dynamics and Control, 31:473–491, 2007.
[11] J-M. Grandmont. On endogenous competitive business cycles. Econometrica, 53:995–1045, 1985.
22
[12] John Guckenheimer and Philip Holmes. Nonlinear oscillations, dynamical systems, and bifurcations of
vector fields, volume 42 of Applied Mathematical Sciences. Springer-Verlag, New York, 1990. Revised and
corrected reprint of the 1983 original.
[13] Carles Simó. On the analytical and numerical approximation of invariant manifolds. In Daniel Benest and
Claude Froeschlé, editors, Les Méthodes Modernes de la Mecánique Céleste (Course given at Goutelas,
France, 1989), pages 285–329, Paris, 1990. Editions Frontières.
[14] Floris Takens. Normal forms for certain singularities of vectorfields. Ann. Inst. Fourier (Grenoble),
23(2):163–195, 1973. Colloque International sur l’Analyse et la Topologie Différentielle (Colloques Inter-
nationaux du Centre National de la Recherche Scientifique, Strasbourg, 1972).
A Invariant manifolds and power series

In this section we will extend the explanations given in Section 2.1.1, including the proof of Proposition 1.
For ease of exposition and without loss of generality let us consider a change of coordinates so that:
• the origin is a fixed point so that z∗ =0 and F (0)= 0;
• the eigenspace W is horizontal, that is the Jacobian at the origin DF (0) is block triangular and is given
by:  
A1 B
A=  ,
0 A2
where A1 , A2 and B are matrices of dimensions d × d, (n − d) × (n − d) and d × (n − d), respectively.
To make the splitting of the phase space more evident, we introduce the notation z =(x, y), where
x =(x1 , . . . , xd ), y =(y1 , . . . , yn−d ), and rewrite the dynamical system as follows:
X [k]
xt+1 = F1 (xt , yt ) = A1 xt + Byt + F1 (xt , yt ) ,
k≥2
X [k]
(30)
yt+1 = F2 (xt , yt ) = A2 yt + F2 (xt , yt ) .
k≥2
By rewriting the dynamical system in this particular form, it is clear that the x-plane W = {y = 0} is a
d-dimensional subspace that is invariant under A.
We are now interested in constructing an invariant manifold W tangent to the x-plane. We will split also
the parameterization zt = Φ(ut ) in components: xt = ϕ(ut ), yt = ψ(ut ). Since the manifold is tangent to the
x-plane, we can choose a parameterization such that Dϕ(0) = Id and Dψ(0) = 0. In these coordinates, the
linearization of the dynamics ut+1 = f (ut ) on the manifold is given by A1 . As a result, the Taylor expansions
of our unknown functions ϕ, ψ and f around u = 0 are of the form
X
x = ϕ(u) = u + ϕ[k] (u)
k≥2
X
y = ψ(u) = ψ [k] (u)
k≥2
X
f (u) = A1 u + f [k] (u) .
k≥2
23
where ϕ(u) =(ϕ1 (u), . . . , ϕd (u)), ψ(u) =(ψ1 (u), . . . , ψn−d (u)), f (u) =(f1 (u), . . . , fd (u)). Again, the terms
[k]
ϕ[k] (u), ψ [k] (u), f [k] (u) denote normalized derivatives of order k, and for instance, ϕ1 (u) is a homogeneous
polynomial of order k.
In order to find the unknowns ϕ, ψ and f , one can compute recursively the terms of their Taylor series.
Notice that the starting point of the recursion is the terms of order 1, which are already known. Let us assume
then that we have already computed all the terms up to order k − 1, ϕ[<k] (u), ψ [<k] (u), f [<k] (u), and we want
to compute the terms of order k, ϕ[k] (u), ψ [k] (u), f [k] (u).
By formal substitution of the Taylor series expansions into equation (2), and truncation up to order k, we
obtain the equation for ϕ[k] (u), ψ [k] (u), f [k] (u): 32
ϕ[k] (A1 u) − A1 ϕ[k] (u) + f [k] (u) − Bψ [k] (u) = r[k] (u) , (31)
ψ [k] (A1 u) − A2 ψ [k] (u) = s[k] (u) , (32)
where r[k] (u) corresponds to the known coefficients of the x coordinate and s[k] (u) corresponds to the known
coefficients of the y coordinate, that is:
h i[k] h i[k]
r[k] (u) = F1 (ϕ[<k] (u), ψ [<k] (u)) − ϕ[<k] (f [<k] (u)) , (33)
h i[k] h i[k]
s[k] (u) = F2 (ϕ[<k] (u), ψ [<k] (u)) − ψ [<k] (f [<k] (u)) . (34)
Hence, at each step k, we have to solve a linear system of equations (31), (32) where the unknowns are
the coefficients of the homogeneous terms, ϕ[k] (u), ψ [k] (u), f [k] (u). These kinds of systems are known as the
homological equations in the dynamical system literature. As we will see, there can be algebraic obstructions
to solve these equations that in some cases can be overcome. It is worth mentioning that this methodology
has the flavor of the theory of normal forms, initiated by Poincaré, and that the algebraic obstructions are
known as resonances.
In our construction, notice that equations (31), (32) are a sort of block triangular system. Thus, we can
first solve equation (32) to compute ψ [k] (u) and then we solve equation (31) to obtain ϕ[k] (u) and f [k] (u).
For the sake of simplicity, we will assume that the matrices A1 and A2 are diagonal (with possibly complex
entries), but most of what is explained below works in the general case that A1 and A2 have complex Jordan
normal form. Thus, we assume that A1 =diag(λ1 , . . . , λd ), A2 =diag(µ1 , . . . , µn−d ).
First, we split (32) into components, so for j = 1, . . . , n − d we have the following relations:
X X
j
ψm (λm − µj )um = sjm um , (35)
|m|=k |m|=k
where we use the multi-index notation for m = (m1 , . . . , md ). As long as the so called non-resonance condition
for all |m| = k, j = 1, . . . , n − d, λm − µj 6= 0 (36)
holds, we can compute the coefficients of the homogeneous terms of ψ j by taking
j sjm
ψm = .
λm − µj
32 By “formal substitution” we mean that we operate algebraically the composition of Taylor polynomials, and we kill all the
terms of order higher than k. Finally, we match the terms of order k.
24
We note that if for some |m| ≥ 2 and j = 1, . . . , n − d, it happens that λm = µj , we have an algebraic
j
obstruction (a resonance) for computing ψm , so in such a case we cannot compute the terms of the expansion
of the manifold (an in general there is not an invariant manifold attached to such a linear subspace!). We
will refer to such an obstruction as a primary resonance. We moreover emphasize that the obstruction only
involves an algebraic relation between the eigenvalues of both A1 and A2 .
Once we have computed the k order coefficients of ψ [k] , we can substitute them in equation (31), and solve
for ϕ[k] and f [k] . For i = 1, . . . , d we have the following relations:
X X
(ϕim (λm − λi ) + fm
i
)um = i m
r̄m u , (37)
|m|=k |m|=k
where r̄[k] (u) = r[k] (u) + Bψ [k] (u). Notice that, in this case, there are non-unique solutions of the equations
(37). This fact has to do with the non-uniqueness of the parameterization of the invariant manifold, even if
the manifold itself is unique. There are basically two different ways of solving the previous system of linear
equations. The first approach, the graph method, tries to simplify the local representation of the manifold.
The second one, the parameterization method, tries to simplify the local representation of the dynamics on
the manifold.
Graph method. This method simplifies the local representation of the invariant manifold. In order to do
i i
so, we set ϕm = 0 and fm = r̄m so that in the parameterization of W has ϕ(u) = u and as a result the
manifold W is (locally) a graph y =ψ(x).
Notice also that the invariance equation can then be rewritten as follows:
ψ(F1 (x, ψ(x))) = F2 (x, ψ(x)) ,
where the dynamics on the manifold is given by:
f (x) = F1 (x, ψ(x)).
Parameterization method. This method simplifies the dynamics on the invariant manifold by choosing a
suitable parameterization of the manifold. In order to do so, we impose the following conditions:
• If λm − λi 6= 0, we set
i
r̄m
ϕim = i
, fm =0;
λm − λi
• If λm − λi = 0, we set
ϕim = 0 , fm
i i
= r̄m .
We emphasize that if λm 6= λi for all |m| ≥ 2 and i = 1, . . . , d, then we can reduce the dynamics on the
manifold to a linear dynamics. The obstructions λm = λi are the secondary resonances, and their existence
avoid the possibility of linearization. Moreover, if the eigenvalues λi of A1 are all of modulus smaller (bigger)
than one, then there is only a finite set of possible secondary resonances, and at least one can reduce the
dynamics to a polynomial dynamics. See [7]. This resembles the Poincaré-Dulac normal form.
25
The set of possibilities is not closed with the two above. The only thing we need is to be able to solve the
homological equations. The strategy can be pushed much more. For instance, if A1 is (1), then all the terms
in the manifold are resonant, an in principle we cannot eliminate any of the terms of the expansion of f . But
if instead of choosing ϕim = 0 at each step we are more sharp, we can be able of eliminating further terms of
f an obtain a polynomial. This strategy has been used in [3] and [4].
B Preserved quantities and asymptotic manifolds

In this section we prove Proposition 2. It follows from a simple continuity argument. So, it is enough to
assume that the preserved quantity H : B → Rn−m is a continuous function.
Assume that z ∈ B is such that its orbit zt = F t (z) converges to the steady state z∗ ∈ B in the future.
That is, assume that lim F t (z) = z∗ . Since H0 = H(z) = H(zt ) for all t, then H0 = lim H(zt ) = H(z∗ ).
t→+∞ t→+∞
The study of the case lim F t (z) = z∗ is analogous.
t→−∞
As a result, all the points that converge asymptotically either in the future or in the past to a steady state,
belong to the same level surface of such a steady state. u
t
C Proofs of the ABS model

Proof of Proposition 3. The steady states (R(1,1) , R(2,2) , R(1,2) , R(2,1) ) of the ABS model satisfy the
equations:

1 1
R(2,1) = α 1+α− R(1,1) , (38)

1 1
R(1,2) = β 1+β− R(2,2) , (39)
R(1,2) R(2,1) = (R(1,1) )2 = (R(2,2) )2 . (40)
Since the rates of returns are all positive, we have that R(1,1) = R(2,2) = R > 0, and from equations (38) and
(39) we obtain that R is a root of the following quartic polynomial
f (R) = αβR4 − (1 + α)(1 + β)R2 + (2 + α + β)R − 1. (41)
Notice that we have to consider the solutions such that R(1,2) and R(2,1) are positive. In particular, R = 1 is
a solution, thus the corresponding steady state is R(1,1) = R(2,2) = R(1,2) = R(2,1) = 1.
Since f (R) = (R − 1)g(R), where
g(R) = αβR3 + αβR2 − (1 + α + β)R + 1, (42)
the rest of the roots solve the equation g(R) = 0. Furthermore, note that g(−1) = 2 + α + β > 0, g(0) = 1 and

1 β 1
g 1+α = − (1+α) 3 < 0. Hence, the three roots of g are real an they are in the intervals ] − ∞, −1[, ]0, 1+α [
1
and ] 1+α , ∞[.
1 1
The root R∗ we are interested is the one that is in the third interval, because R = R∗ > 1+α ≥ 1+β which
w
implies that both R(1,2) and R(2,1) are positive. Furthermore, under the assumption e > w, e ≤
R∗ < 1. This
w
1 w w

follows from the following facts: g e = 4γ(1−γ) e − 1 ≤ 0 and g(1) = 2αβ − α − β = 2αβ 1 − e > 0.
u
t
26
Proof of Proposition 4. In order to study the linear stability at a steady state we first compute the
Jacobian matrix
 
0 0 0 αR2
 
 0 0 βR2 0 
A=  .
 
1
1 1 1 (1+β)− R
−α

 0 βR (1 + β) − R R (1 + β) − R β · (1+α)− R1

 1

1 1 β (1+α)− R
+ α) − R1

αR (1 + α) − R 0 −α · 1
(1+β)− R
R (1
where either R = 1 or R = R∗ depending on the steady state we are examining. The associated eigenvalues
are the roots of the characteristic polynomial, that is given by:
p(λ) = (λ + 1)q(λ) ,
where
q(λ) = λ3 + (1 − (2 + α + β)R)λ2 + αβR4 λ + αβR4 .
Hence, −1 is an eigenvalue of A.
We have to analyze the roots of the cubic polynomial q(λ) for the cases R = 1 and R = R∗ .
If R = 1,
3 2 1 3
q(λ) = λ − (1 + α + β)λ + αβλ + αβ = λ g ,
λ
1
where g is defined in (42). In the proof of Proposition 3, we separated the zeros of g, so λ is in ] − ∞, −1[, or
1 1
in ]0, 1+α [ or in ] 1+α , 1[. Hence, λs ∈] − 1, 0[, λu,1 ∈]1, 1 + α[ and λu,2 ∈]1 + α, ∞[.
If R = R∗ ,
q(λ) = λ3 + (1 − (2 + α + β)R∗ )λ2 + αβR∗4 λ + αβR∗4 .
Then:
1. q(−1) = −(2 + α + β)R∗ < 0;
2. q(0) = αβR∗4 > 0;
1 w

3. q(1) = 2 − (2 + α + β)R∗ + 2αβR∗4 = 2αβ R∗ R∗3 − R∗2 − R∗ + e < 0.
In order to prove the third inequality we proceed as follows. First, notice that g(R∗ ) = 0 (see equation
(42)) implies that
2 − (2 + α + β)R∗ = (α + β)R∗ − 2αβR∗2 − αβR∗3 ,
which proves
1 w
q(1) = R∗ R∗3 − R∗2 − R∗ + .
2αβ e
w
Since e ≤ R∗ < 1, and the function R → R3 − R2 − R is decreasing in [0, 1], we have
w w 3 w 2 w 2 w
R∗3 − R∗2 − R∗ + ≤ − = −1 <0 .
e e e e e
In fact,
q(R∗ ) = R∗2 − (1 + α + β)R∗3 + αβR∗4 + αβR∗5 = R∗2 g(R∗ ) = 0 ,
so λs,2 = R∗ is an eigenvalue of the Jacobian matrix in the non monetary steady state. u
t
27
Proof of Proposition 5. The proof is a simple consequence of the fact that
R(1,2),t+1 R(2,1),t+1 R(1,1),t R(2,2),t
= ,
R(1,1),t+1 R(2,2),t+1 R(1,2),t R(2,1),t
which can be obtained by multiplying the two formulae in equation (14). u

t
Proof of Proposition 6. Since we look for the solutions of (21),(22) and (25) that are on the level set Σ1 ,
we add the extra equation
R(1,1) R(2,2)
=1. (43)
R(1,2) R(2,1)
From (21) we obtain
2
αR(1,1) = R(1,2) (1 + α − αR(2,1) )((1 + α)R(1,1) − 1)
R(2,2) R(1,1) R(1,1)
= ((1 + β − βR(1,2) )((1 + α)R(1,1) − 1) ,
R(2,1) R(2,2)
where in the last equality we use (43). It follows that

1 αR(2,1)
R(1,1) = 1+ ,
1+α 1 + β − βR(1,2)
which is (28). A similar argument proves the formula for R(2,2) , (29).
Finally, equations (26) and (27) of Proposition 6 are obtained by substituting the formulae (28) and (29)
in (25) and (43). u
t
28

A Geometric Description of A Macroeconomic Model With A Center Manifold

Uploaded by

Copyright:

Available Formats

A Geometric Description of A Macroeconomic Model With A Center Manifold

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

A Geometric Description of A Macroeconomic Model With A Center Manifold

Uploaded by

Copyright:

Available Formats

A Geometric Description of a

Macroeconomic Model with a Center

Barcelona Economics Working Paper Series

Department of Economics Department de Matemàtica Aplicada i Anàlisi

JEL Classification: E4, E3.

2 On the computation of invariant manifolds

zt+1 = F (zt ) , (1)

2.1 Parameterization of invariant manifolds

then the manifold parameterized by Φ is defined by W = {Φ(u) ∈ Rn | u ∈ U ⊂ Rd }. This manifold is invariant

ψ(Fx (x, ψ(x))) = Fy (x, ψ(x)) .

2.1.1 Formal computations

where P1 = DΦ(0) is a n × d matrix and A1 = Df (0) is a d × d matrix.

AΦ[k] (u) − Φ[k] (A1 u) − P1 f [k] (u) − R[≤k] (u) + · · · = 0 , (5)

Φ(u) = P1 u + P2 ψ(u) , (7)

where P2 is a n × (n − d) matrix whose columns complete those of P1 to a basis of Rn , and ψ is a

Proof: See Appendix A.

2.1.2 Globalization of the invariant manifold

Dε = {u ∈ Rd | ||F (Φ̃(u)) − Φ̃(f˜(u))|| ≤ ε} ,

where ||·|| is a norm in Rn , for instance the supremum norm.

2.1.3 Stability analysis on an invariant manifold

2.2 Preserved quantities and invariant manifolds

ΣH0 = {z ∈ B | H(z) = H0 } (9)

H(z) = H0 ⇒ H(F (z)) = H0 .

Proof: See Appendix B.

3 Application to an economic model

3.1 The dynamical system

3.2 Steady State Equilibria

Proposition 3 The ABS economy has two steady states:

R(1,1) = R(2,2) = R(1,2) = R(2,1) = 1 ,

and a common price level in both locations;

αβR3 + αβR2 − (1 + α + β)R + 1 = 0 (16)

Proof: See Appendix C.

Proposition 4 The steady states of the ABS economy satisfy:

Proof: See Appendix C.

3.3 Invariant manifolds: formal expansions

states up to order 30.

computers, thanks to the high performance of our implementations.

3.4 The 2-periodic system and dynamics on the center manifolds

Proof: See Appendix C.

3.5 The reduced system

3.6 Invariant manifolds: globalization

intersect in O-dimensional points.

[9] P. Gomis-Porqueras and A. Haro. Global dynamics in macroeconomics: An overlapping generations

A Invariant manifolds and power series

• the origin is a fixed point so that z∗ =0 and F (0)= 0;

for all |m| = k, j = 1, . . . , n − d, λm − µj 6= 0 (36)

holds, we can compute the coefficients of the homogeneous terms of ψ j by taking

ψ(F1 (x, ψ(x))) = F2 (x, ψ(x)) ,

where the dynamics on the manifold is given by:

f (x) = F1 (x, ψ(x)).

B Preserved quantities and asymptotic manifolds

C Proofs of the ABS model

f (R) = αβR4 − (1 + α)(1 + β)R2 + (2 + α + β)R − 1. (41)

g(R) = αβR3 + αβR2 − (1 + α + β)R + 1, (42)

1. q(−1) = −(2 + α + β)R∗ < 0;

2. q(0) = αβR∗4 > 0;

which can be obtained by multiplying the two formulae in equation (14). u

where in the last equality we use (43). It follows that

You might also like