See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/4878621

Expected Utility in Models with Chaos

Article · November 2007

Source: RePEc

CITATIONS READS

4 33

3 authors, including:

Judy Kennedy Brian E Raines

Lamar University Baylor University
64 PUBLICATIONS   830 CITATIONS    26 PUBLICATIONS   293 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Tent maps View project

Shadowing and omega limit sets View project

All content following this page was uploaded by Brian E Raines on 29 May 2014.

The user has requested enhancement of the downloaded file.

 Expected Utility in Models with Chaos
Judy Kennedy∗ Brian Raines† David R. Stockman‡
November 2006

Abstract
In this paper, we provide a framework for calculating expected utility in models with
chaotic equilibria and consequently a framework for ranking chaos. Suppose that an
economic model’s equilibria correspond to orbits generated by a chaotic dynamical
system f : X → X where X is a compact metric infinite space and f is continuous. The
map f could represent the forward dynamics xt+1 = f (xt ) or the backward dynamics
xt = f (xt+1 ). If f represents the forward/backward dynamics, the set of equilibria
forms a direct/inverse limit space. We use an f -invariant measure on X to induce
a measure on the direct/inverse limit space and show that this induced measure is
σ-invariant where σ is the shift operator. Moreover, we show that if the f -invariant
measure is a natural invariant measure, then the induced measure on the direct/inverse
limit space will also be a natural invariant measure. We utilize this framework in the
cash-in-advance model of money where f is the backward map to calculate the expected
utility when equilibria are chaotic.
Keywords: chaos, inverse limits, direct limits, natural invariant measure, SRB measure,
SLYRB measure, Birkhoff ergodic theorem, cash-in-advance.
JEL: C6, E3, and E4.

[PRELIMINARY AND INCOMPLETE]

∗
Department of Mathematical Sciences, University of Delaware, Newark, DE 19716.
†
Department of Mathematics, Baylor University, Waco, TX
‡
Department of Economics, University of Delaware, Newark, DE 19716. Stockman would like to thank
the Lerner College of Business and Economics for its generous summer research support.

1
1 Introduction
In this paper, we provide a framework for calculating expected utility in models with chaotic
equilibria and consequently a framework for ranking chaos. Suppose that an economic
model’s equilibria correspond to orbits generated by a chaotic dynamical system f : X → X
where X is a compact metric infinite space and f is continuous.1 The map f could represent
the forward dynamics xt+1 = f (xt ) or the backward dynamics xt = f (xt+1 ). If f represents
the forward/backward dynamics, the set of equilibria forms a direct/inverse limit space.
The direct/inverse limit space is a subset of X ∞ . The direct limit space consists of the
forward orbits of f and the inverse limit space consists of all the backward orbits of f . We see
that if every forward (backward) orbit of f : X → X is an equilibrium in the model then the
set of equilibria is a direct (inverse) limit space. Furthermore, the dynamics of f (or f −1 ) are
being captured by the shift map σ on the direct (inverse) limit space. If f is a non-invertible
backward map, we say that the model (or dynamical system) has backward dynamics, i.e,
the relationship describing the equilibrium dynamics is multi-valued going forward in time,
but is single-valued going backward in time. Two such models that may have backward
dynamics include the the overlapping generations (OLG) model and the cash-in-advance
(CIA) model.2,3
We would like to use a probability measure that respects the dynamical system and gives
us the actual probability of seeing certain Borel sets of X and the direct/inverse limit space.
These types of measures are called natural invariant measures.4 An f -invariant measure µ
has the property that µ[A] = µ[f −1 (A)] for every measurable set A. The natural invariant
measures µ is an f -invariant measure that respects the dynamics of f : X → X in the
following sense. If S ⊂ X and almost all points in X have 40% of their respective orbits in S
then µ assigns S measure 0.4. We use an f -invariant measure on X to induce a measure on
the direct/inverse limit space. We show that this induced measure is σ-invariant. Moreover,
1
Throughout this paper, by chaotic, we mean in the sense of Devaney (2003).
2
See Grandmont (1985) for the OLG model and Michener and Ravikumar (1998) for the CIA model.
3
Inverse limits is a relatively new approach to analyzing dynamic economic models with backward dy-
namics. Medio and Raines (2006, 2005) use inverse limits to analyze the long-run behavior of an OLG model
with backward dynamics. Kennedy et al. (2004, 2005) investigate the topological structure of the inverse
limit space associated with the CIA model of Lucas and Stokey (1987). Kennedy and Stockman (2006)
utilize the inverse limit space to show that a multi-valued dynamical system with backward dynamics is
chaotic going forward in time if and only if it is chaotic going backward in time.
4
Natural invariant measures are also referred to as SRB or SLYRB measures. See Hunt et al. (2002) for
more on these measures.

2
we show that if the f -invariant measure is a natural invariant measure, then the induced
measure on the direct/inverse limit space will also be a natural invariant measure.
One of the uses of the induced measure on the direct/inverse limit space is to perform
integration over the direct/inverse limit space in a way that is dynamically meaningful. We
show that one can integrate continuous real-valued functions on the direct/inverse limit
space using this induced measure and offer a method for calculating such integrals. To be
more concrete, suppose that Y is the direct/inverse limit space from a DGE model with
a representative agent and let µ be the natural invariant measure for f . Then the utility
function of the representative agent can be viewed as a function W : Y → R given by
∞
X
W (x) := β t U (xt ).
t=0

It would be of interest to calculate the expected utility:

Z
E[W (x)] := W (x)dm(x),
Y

where m is the induced measure on Y by µ. Note that our utility function is essentially
inducing a ranking on direct/inverse limit spaces and consequently providing a means for
ranking chaos. In the context of the cash-in-advance model, this could useful in ranking
monetary policies (money growth rates) all of which lead to chaotic equilibria and assessing
the welfare consequences of changing the money growth rate.
The paper is organized as follows. In section 2 we construct our induced measure on the
direct limit space, show that it is a natural invariant measure with respect to the shift map.
We also establish that one can integrate on these spaces. direct limit spaces. In section 3
we carry out this analysis for the more difficult inverse limit case. The formidable problem
associated with numerically calculating these integrals is discussed in Section 4 along with a
solution (Birkhoff’s ergodic theorem). An application of these tools to the cash-in-advance
model is in Section 5. We conclude in Section 6.

2 Measure and Integration on Direct Limit Spaces

Let (X, d) be a compact metric space and f : X → X continuous. We define the direct limit
space by
lim(X, f ) := {(x0 , x1 , x2 , . . .) ∈ X ∞ : xn = f (xn−1 ), n ∈ N}. (1)
−→

3
We see the direct limit space is simply set of forward orbits of f . Let D := lim(X, f ). We
−→
induce a metric on D using our metric d on X by
∞
X d(xi , yi )
ρ(x, y) := .
i=0
2i

Let B(X) denote the Borel sets of X. If µ is a measure of X with the property that
µ[f −1 (S)] = µ[S] for every closed set S, then µ is called an f -invariant measure (or invariant
measure for f ). Suppose µ is an f -invariant measure defined on the B(X). We wish to define
a measure m on B(D) that is σ-invariant where σ is the shift map on D. Furthermore, if µ is
a natural invariant measure, we would like the induced measure m to be a natural invariant
measure as well. We also want to be able to integrate continuous functions h : D → R, so
that we can compute expected utility over D

2.1 Invariant Measures

Theorem 1. Suppose f : X → X is continuous and X is a compact metric space. Let D :=
lim(X, f ) with induced metric ρ. Then X and D are homeomorphic with a homeomorphism
−→
given by H : X → D defined by H(x) := (x, f (x), f 2 (x), . . .) with H −1 ≡ π0 : D → X defined
by π0 ((x0 , x1 , x2 , . . .)) := x0 . Furthermore, f : X → X and the shift map σ : D → D are
conjugate with f = H −1 ◦ σ ◦ H.

This theorem implies that (D, ρ) is a compact metric space as well. Note that since H is
a homeomorphism, we can use H to induce a measure m on B(D) given by

m[B] := µ[H −1 (B)] ≡ µ[π0 (B)]. (2)

The next theorem establishes that an invariant measures induces an invariant measure for a
conjugate dynamical system.

Theorem 2. Suppose that X and Y are compact metric spaces, f : X → X and g : Y → Y

are continuous and conjugate with conjugacy given by h : X → Y . Then if µf is an f -
invariant measure on the Borel sets B(X) then the induced measure on the Borel sets B(Y )
given by µg [B] := µf [h−1 (B)] is g-invariant.

Proof. Let B be a closed set in Y . Note that since h is a homeomorphism we have µg [h(A)] =
µf [A] for A ∈ B(X) as well. Since g −1 (B) ≡ h ◦ f −1 ◦ h−1 (B) we have

µg [B] = µf [h−1 (B)] = µf [f −1 ◦ h−1 (B)] = µg [h ◦ f −1 ◦ h−1 (B)] = µg [g −1 (B)].

4
Corollary 1. Suppose f : X → X is continuous, X is compact, µ is a Borel f -invariant
measure, D is the direct limit space and σ : D → D the shift map. Then the induced measure
on B(D) given by m[B] := µ[π0 (B)] is σ-invariant.

A Borel measure is a measure µ defined on the collection S of all Borel sets such that
µ(C) < ∞ for every compact set C in X. Our induced measure m is clearly Borel.

2.2 Natural Invariant Measures

Suppose X is a compact metric space with Lebesgue measure λ, f : X → X is continuous,
x ∈ X, and S ⊂ X. Define the fraction of the orbit of x lying in S by

#{f i (x) ∈ S : 1 ≤ i ≤ n}
G(x, S) = lim ,
n→∞ n
provided this limit exists. For B ⊂ X and r > 0, define B r := {x ∈ X : d(x, y) < r for some
y ∈ B}. The natural measure generated by the map f (or the f -measure) is defined by

µf (S) = limG(x, S r ),
r→0

provided that for λ-almost every x this limit exists and is the same.5 If λ is Lebesgue
measure on X and h : X → Y is a homeomorphism, then can we call λ̃ defined on the Borel
sets of Y by λ̃[B] := λ[h−1 (B)] a Lebesgue measure. since our induced measure space is
homeomorphic to a Lebesgue space. The next theorem establishes that a natural invariant
measures induces a natural invariant measure for a conjugate dynamical system.

Theorem 3. Suppose we have (X, f, B(X), µ) and (Y, g, B(Y ), m) where X and Y are com-
pact metric spaces, f : X → X and g : Y → Y are continuous, f and g are conjugate with
conjugacy h : X → Y , µ is an f -invariant Borel measure and m is the Borel g-invariant
measure induced on B(Y ) by h, i.e., m[B] = µ[h−1 (B)]. If µ is a natural invariant measure,
then m is a natural invariant measure.

Proof. Let λ be Lebesgue measure on B(X) and λ̃ be Lebesgue measure on B(Y ) given by
λ̃[B] := λ[h−1 (B)]. Since µ is a natural invariant measure, there exists a measurable set PX
5
Natural invariant measures are also referred to as SRB or SLYRB measures. See Hunt et al. (2002) for
more on these measures.

5
with Lebesgue measure zero where limr→0 G(x, B r ) exists and is the same for x ∈ X \ PX .
Let S ⊂ Y be closed. We define

mg [S] := lim G(y, S r ).

r→0

We will show that this limit exists λ̃-almost every y in D and equals m[S]. Let PY := h(PX ).
Then h−1 (PY ) = PX and λ̃(PY ) = λ[h−1 (PY )] = λ[PX ] = 0. Let y ∈ Y \ PY and x = h−1 (y).
N.B. y ∈ Y \ PY iff x = h−1 (y) ∈ X \ PX . We also have x ∈ h−1 (S r ) iff h(x) ∈ S r which
implies f j (x) ∈ h−1 (S r ) iff h ◦ f j (x) ∈ S r which is equivalent to h ◦ f j ◦ h−1 (y) ∈ S r or
g j (y) ∈ S r . This gives

#{f j (x) ∈ h−1 (S r ) : 1 ≤ j ≤ n} = #{g j (y) ∈ S r : 1 ≤ j ≤ n}.

This implies that G(y, S r ) exists and is the same for all y ∈ Y \ PY . Furthermore, G(y, S r ) =
µ[h−1 (S r )]. So we have

mg [S] := lim G(y, S r ) = lim µ[h−1 (S r )] = µ[h−1 (S)] = m[S].

r→0 r→0

Corollary 2. Suppose f : X → X is continuous, X is compact, µ is a Borel f -invariant

measure, D is the direct limit space and σ : D → D the shift map. If µ is a natural invariant
measure on B(X), then the induced measure m is a natural invariant measure on B(D).

2.3 Integration
Let X be a compact metric space with f : X → X continuous, µ a natural invariant measure
and D := lim(X, f ). Let g : D → R be a continuous function. We now show how to integrate
−→
g on D by approximating g with a sequence of simple functions defined on tower sets. This
sequence of simple functions will converge in measure to g and so we can approximate the
integral of g by integrating each of these simple functions. The following lemma follows from
the uniform continuity of f n for each n ∈ N and how we induced a metric on D from the
metric d on X.

Lemma 1. Let f : X → X be a continuous function on a compact metric space X and

D := lim(X, f ). Let ² > 0. Then there exists a δ > 0 such that for every x ∈ X, π0−1 [Bδ (x)]
−→
has diameter less than ².

6
 The first step in constructing any simple function on D is to choose a partition {Ei }ni=1
1
of the space D. Let k be a positive integer and define a positive number δk ≤ 2k
such that
if x ∈ X, then π0−1 [Bδk (x)] has diameter less than 1
k
for each x ∈ X (the existence of such
values is guaranteed by Lemma 1). Also, define a partition Pk = {Pk,1 , Pk,2 , . . . , Pk,Lk } of X
with the following properties:

(1) mesh(Pk ) < δk , and

(2) Pk refines Pk−1 .

Define also a collection of sample points Qk = {xk,1 , xk,2 , . . . , xk,Lk } ⊂ X such that for
each 1 ≤ i ≤ Lk , xk,i ∈ Pk,i .
Next we translate these sets and sample points into the direct limit space. Let

bk = {Pbk,i }Lk := {π0−1 (Pk,i )}Lk ,

P i=1 i=1

and let
bk = {xk,i : 1 ≤ i ≤ Lk },
Q
bk
where each xk,i is some member of π0−1 (xk,i ). For each integer k, this gives us a partition P
of the direct limit space with the following properties:

bk ) < 1 ,
(1) mesh(P k

bk refines P
(2) P bk−1 , and

bk ] = µ[Pk ].
(3) m[P

Also, we have chosen a representative point xk,i out of each element Pbk,i . We can now
define our sequence of simple functions. For each x ∈ D and each positive integer k, let
Lk
X
gk (x) = g(xk,i )χPbk,i (x).
i=1

In order to demonstrate that g is integrable, we need to show that the sequence {gk }∞
k=1 is

mean fundamental and that it converges in measure to g.

Lemma 2. The sequence {gk }∞

k=1 is mean fundamental.

7
Proof. Let ² > 0, and let δ > 0 so that g(Bδ (x)) has diameter less than ² for all x ∈ D.
bk ) < δ. Let x ∈ D
Choose k, r to be positive integers with r > k large enough so that mesh(P
with x not in the boundary of Pbk,i or Pbr,i for all i. Notice that all of these boundaries are
sets of the form π0−1 (S), where S is the boundary of a member of the partition Pt (t = r
or t = k), so µ(S) = 0 = m(π0−1 (S)). Hence, by restricting our attention to points of the
inverse limit space which are not in any of these boundaries, we are only omitting a set of
measure zero.
Consider ¯ ¯
¯XLr Lk
X ¯
¯ ¯
|gr (x) − gk (x)| = ¯ g(xr,i )χPbr,i (x) − g(xk,i )χPbk,i (x)¯ .
¯ ¯
i=1 i=1

Choose q, s to be positive integers such that x ∈ Pbr,q ⊆ Pbk,s . Then, since x is not on
the boundary of any of the partition elements, gr (x) = g(xr,q ) and gk (x) = g(xk,s ). Since
mesh(Pbk ) < 1 , |xr,q − xk,s | < δ. Thus,
k

|gs (x) − gk (x)| = |g(xr,q ) − g(xk,s )| < ².

Hence, Z Z
ρ(gr , gk ) = |gr − gk | dm < ² dm
R
and since ² dm = ² · m[D] = ², we have established that ρ(gr , gk ) → 0 as r, k → ∞.

Lemma 3. The sequence of simple functions {gk }∞

k=1 converges in measure to g.

Proof. Let ² > 0, and again choose δ > 0 so that g(Dδ (x)) has diameter less than ² for all
x ∈ D. Choose M large enough so that if k > M , then mesh(P bk ) < δ.
bk , and let r
Pick k ≥ M . Choose x so that x is not on the boundary of any element of P
be a positive integer such that x ∈ Pbk,r . Then

|gk (x) − g(x)| = |g(xk,r ) − g(x)|

and since both x and xk,r are in Pbk,r , d(x, xk,r ) < δ. Thus, |g(xk,r ) − g(x)| < ², which
implies that for all x not on the boundary of any element of P bk , |gk (x) − g(x)| < ². Thus,
m[{x : |gk (x) − g(x)| ≥ ²}] = 0 for all k ≥ M , and lim m[{x : |gk (x) − g(x)| ≥ ²}] = 0.
k→∞

3 Measure and Integration on Inverse Limit Spaces

Given a compact metric space X, a continuous map f : X → X, and a measure µ defined
on X which is invariant relative to f , we wish to define a measure m on Y := lim(X, f )
←−

8
invariant relative to the induced shift homeomorphisms F and σ on Y . If the measure µ
is, in addition, a natural measure on X, we would like the induced measure m on Y to be
natural also. We want to be able to integrate continuous functions h : Y → R, so that we
can compute expected values for our utility functions defined on Y .

3.1 Invariant Measures

If X is a locally compact metric space, denote by C the collection of all compact subsets of
X, by S the σ-ring generated by C, by U the collection of all open sets in X. A member of
the collection S is called a Borel set, and S is the collection of Borel sets in X.
A content is a nonnegative, finite, monotone, additive, and subadditive set function
defined on the class C of all compact sets of a locally compact metric space X.
A Borel measure is a measure µ defined on the collection S of all Borel sets such that
µ(C) < ∞ for every C ∈ C.
It is straightforward to generates a regular Borel measure from a content on the compact
subsets of a locally compact space X, see (Halmos, 1974, Section 53, p. 231). Our goal
now is to define a content on the compact subsets of Y := lim(X, f ), where X is a compact
←−
metric space and f : X → X is continuous, and then use that content to generate a measure
on Y . The measure we obtain on Y is not new, see Bochner (1955) or Choksi (1958), but
establishing that it is σ-invariant or a natural invariant measure is new. We include the
construction of this measure for completeness and because in this section we give much of
the notation we will use in later proofs.
Suppose then that X is a compact metric space and f : X → X is continuous. Let
Y := lim(X, f ). We use boldface letters, i.e., x to denote a point of Y , with xn denoting the
←−
nth coordinate of x. Also, πn denotes the nth projection on Y , so πn (x) = xn .
Let n be a nonnegative integer, and let B be a compact subset of Y . Define the tower
sets Bn for B as follows: define Bn := {x ∈ Y : πn (x) := xn ∈ πn (B)}. Note that the
compact subsets of Y are the closed subsets of Y , and the following proposition shows that
B = ∩∞
n=0 Bn and B0 ⊇ B1 ⊇ · · · ⊇ B. If A ⊂ Y , lim(πn (A), f | (πn+1 (A)) is also a subset of
←−
Y . The following are useful properties of inverse limit spaces.

Proposition 1 (Ingram (2000)). Suppose X is a compact metric space and f : X → X is

continuous. Then if A is a closed, nonempty subset of the inverse limit space Y := lim(X, f ),
←−
lim(πn (A), f |πn+1 (A)) is a closed, nonempty subset of Y and A = lim(πn (A), f |πn+1 (A)).
←− ←−

9
Lemma 4 (Ingram and Mahavier (2004)). Suppose X is a compact metric space, f :
X → X is continuous, and Y := lim(X, f ). Let ² > 0. Then there is a positive number δ
←−
and a positive integer n such that for every x ∈ X, πn−1 (Dδ (x)) has diameter less than ².

Suppose that µ is an invariant measure on X. Now define the function Γ on the compact
subsets of Y by first declaring that

Γ[Bn ] = µ[πn (B)]

where Bn is a tower set for the compact set B as defined above. Then define

Γ[B] = lim Γ[Bn ].

n→∞

The function Γ is a content on the compact sets of Y as we prove below:

Lemma 5. The set function Γ is a content on the compact sets of Y := lim(X, f ).

←−

Proof. It is immediate from the definition that Γ is nonnegative, finite and monotone as long
as µ is. To see that Γ is additive, let K and R be disjoint compact subsets of Y . Then there
is an integer N such that for all m > N , πm (K) ∩ πm (R) = ∅. Thus,

Γ[Km ∪ Rm ] = µ[πm (K) ∪ πm (R)] = µ[πm (K)] + µ[πm (R)]

= Γ[Km ] + Γ[Rm ]

for all m > N . Hence, Γ[K ∪ R] = limn→∞ Γ[Km ∪ Rm ] = limn→∞ Γ[Km ] + limn→∞ Γ[Rm ] =
Γ[K] + Γ[R], and Γ is additive. The proof that Γ is subadditive is similar.

A measure µ on a space Z is regular provided that it is both inner regular and outer
regular. A measure µ is outer regular provided that µ[E] = inf{µ[U ] : E ⊂ U and U is
open}. A measure µ is inner regular provided that µ[E] = sup{µ[C] : C ⊂ E and C is
compact}.
A content Γ on the compact sets C is regular provided

Γ[C] = inf{Γ[D] : C ⊂ D◦ ⊂ D ∈ C}.

Let U be an open set and define the inner content Γ∗ induced by the content Γ by

Γ∗ [U ] = sup{Γ[C] : U ⊇ C and C ∈ C}.

10
 Suppose X is a metric space and E ⊂ X. We say that E is σ-compact provided there
is a collection {Ci }∞ ∞
i=1 of compact sets such that E = ∪i=1 Ci . We say that E is σ-bounded

provided there is a collection {Ci }∞ ∞

i=1 of compact sets such that E ⊂ ∪i=1 Ci . Given a

σ-bounded set E we define

µ∗ [E] = inf{Γ∗ [U ] : E ⊂ U and U is open},

and we call µ∗ the outer measure induced by the content Γ.

Lemma 6. With the assumptions as above, Γ is a regular content on Y provided µ is a

regular, invariant measure on X.

Proof. Let C ∈ C, the collection of all compact subsets of Y . Let ² > 0 and choose a positive
integer n such that |Γ[Cn ] − Γ[C]| < ²/2, where the Cn ’s are the tower sets for C. Since µ
is a regular measure on X, there is a compact subset Dn of X such that Dn◦ ⊇ πn (Cn ) and
|µ[Dn ] − µ[πn (Cn )]| < ²/2. Let D = πn−1 (Dn ). Since πn is continuous and D is compact,
C ⊂ Cn ⊂ πn−1 (Dn◦ ) ⊂ D◦ ⊂ D. Also, |Γ[C] − [Γ[D]| ≤ |Γ[C] − Γ[Cn ]| + |Γ[Cn ] − Γ[D]| ≤
²/2 + |µ[πn (Cn ) − µ[Dn ]| = ². Thus, Γ is regular.

Theorem 4 (Halmos (1974), Theorem E, p. 234). If µ∗ is the outer measure induced

by a content Γ, then the set function defined for every Borel set E by µ[E] := µ∗ [E] is a
regular Borel measure.

Using this measure µ from these definitions to calculate might be tricky. However, since
Γ is a regular content we can use the following result:

Theorem 5 (Halmos (1974), Theorem A, p. 235). If µ is the Borel measure induced

by a regular content λ, then µ[C] = λ[C] for each compact set C.

Thus, these definitions and theorems lead us from the f -invariant measure µ on X to a
content Γ on Y, and finally to a regular Borel measure m on Y , and from this last theorem
we see that m behaves exactly like Γ on the compact sets. We primarily work with compact
subsets K of Y because of the fact that m[K] = Γ[K]. However, as the next few lemmas
show, if K ⊆ Y is Borel and it projects to a Borel subset of X then we can say something
about m[K].

Lemma 7. Let K ⊂ Y be Borel . Then m[K] ≤ µ[πn (K)] for all nonnegative integers n
such that πn [K] is Borel.

11
Proof. If K is compact, this follows immediately from the definition of Γ and the fact that
Γ[K] = m[K].
Next assume that K is open. Then

m[K] = sup{m[C] : C ⊆ K, C is compact}.

Since each m[C] ≤ µ[πn (C)], it follows that

m[K] = sup{m[C] : C ⊆ K, C is compact}

≤ sup{µ[πn (C)] : C ⊆ K, C is compact}.

Since πn is continuous, πn (C) ⊂ πn (K). Thus,

{πn (C) : C ⊆ K, C is compact} ⊆ {R : R ⊆ πn (K), R is compact},

and since m[C] ≤ µ[πn (C)], we have

m[K] ≤ sup{µ[πn (C)] : C ⊆ K, C is compact}

≤ sup{µ[R] : R ⊆ πn (K), R is compact}.

But since µ is also a regular Borel measure and πn [K] is Borel,

µ[πn (K)] = sup{µ[R] : R ⊆ πn (K), R is compact}.

Hence, m[K] ≤ µ[πn (K)] for each n.

Finally let K be any Borel set and recall that

m[K] = inf{m[U ] : U ⊇ K, U is open in Y }.

Let V ⊇ πn (K) be open in X. Then πn−1 (V ) ⊇ K is also open in Y . Thus,

{πn−1 (V ) : V ⊇ πn (K), V is open in X} ⊆

{U : U ⊇ K, U is open in Y },

and

inf{m[πn−1 (V )] : V ⊇ πn (K), V is open in X}

≥ inf{m[U ] : U ⊇ K, U is open in Y } = m[K].

12
But by the previous case

µ[V ] = µ[πn ◦ πn−1 (V )] ≥ m[πn−1 (V )].

Hence,

µ[πn (K)] = inf{µ[V ] : V ⊇ πn (K), V is open in X}

≥ inf{m[πn−1 (V )] : V ⊇ πn (K), V is open in X}
≥ m[K].

Lemma 8. Let K ⊂ X be a Borel set. Then, for any nonnegative integer n, µ[K] =
m[πn−1 (K)].

Proof. Let n be a nonnegative integer. Since Γ and m agree on the compact sets of Y , we
need only consider open sets and then Borel sets in the proof of this lemma.
Let U be an open subset of X, and denote πn−1 (U ) by U b . By the previous lemma,
µ[U ] ≥ m[Ub ]. Let {Li }∞ be a collection of compact subsets of X such that Li ⊂ U for
i=1
b ] = sup{m[K] : K ⊆ U
each i and µ[U ] = sup{µ[Li ] : i is a positive integer}. Then m[U b,
K is compact} ≥ sup{m[π −1 (Li ) : i is a positive integer}, because {K : K ⊆ U b , K is
n

compact} ⊇ {πn−1 (Li ) : i is a positive integer}. Since

sup{m[πn−1 (Li )] : i is a positive integer} = sup{Γ[πn−1 (Li )] : i is a positive integer}

= sup{µ[Li ] : i is a positive integer} = µ[U ],

b ] ≥ µ[U ]. Thus, m[U

we see that m[U b ] = µ[U ].
Let K ⊆ X be a Borel set. Let K b = π −1 (K). By the previous lemma, m[K]
b ≤ µ[K].
n

Let {Zi }∞
i=1 be a collection of compact subsets of X with Zi ⊆ K for all positive integers i

and with µ[K] = sup{µ[Ui ] : i is a positive integer}. For each i, let Zbi = π −1 (Zi ). By the
n

previous paragraph, µ[Zi ] = m[Zbi ] for all positive integers i, and since {R : R ⊆ K
b , R is
compact in Y } ⊇ {Zi : i is a positive integer},

b R is compact in Y }
m[K] = sup{m[R] : R ⊆ K,
≥ sup{m[Zbi ] : i is a positive integer}.

But µ[K] = sup{µ[Zi ] : i is a positive integer} = sup{m[Zbi ] : i is a positive integer}. Thus,

b = µ[K].
m[K]

13
3.2 Natural Invariant Measures
Suppose X is a compact metric space, f : X → X is continuous, x0 is a point in X, and S
is a subset of X. Define the fraction of the orbit of x0 lying in S by

#{f i (x0 ) ∈ S : 1 ≤ i ≤ n}
G(x0 , S) = lim ,
n→∞ n
provided this limit exists.
Suppose X is a compact metric space, S is a subset of X, and r is a positive number.
Define Dr (S) = {x ∈ X : d(x, y) < r for some y ∈ S}.
Suppose X is a compact metric space with a Lebesgue measure λ, f : X → X is
continuous, x0 is a point in X, and S is a closed subset of X. The natural measure generated
by the map f (or the f -measure) is defined by

µf (S) = limG(x0 , Dr (S))

r→0

provided that for λ-almost every x0 this limit exists and is the same. If µ is a measure of X
with the property that µ[f −1 (S)] = µ[S] for every closed set S, then µ is called an invariant
measure for f .
What makes natural invariant measures so nice is that, when one exists, it really does
measure the dynamics of f in the sense that if almost all points have 60% of their respective
orbits in S, then this measure assigns a value of 0.6 to S. Furthermore, this measure is
designed especially for chaotic maps: for non-chaotic maps they are not appropriate and are
often atomic. Figuring out whether the natural invariant measure exists for a map f can
be difficult in general, but in at least one situation it not only does exist, it is also easy to
compute.
A map f : [0, 1] → [0, 1] is Markov if there is a collection {A1 , A2 , . . . , Ak } of subintervals
of [0, 1] such that the intervals are allowed to intersect only at their endpoints (if at all) and
the image of each Ai under f is exactly a union of some of the subintervals.

Theorem 6 (Alligood et al. (1996), Theorem 6.5). Let f : [0, 1] → [0, 1] be piecewise
smooth and piecewise expanding. Then an invariant measure µf exists for f , and the associ-
ated density is bounded, i.e., there is a constant c such that µf [[a, b]] ≤ c |b − a|. If f is also
piecewise linear, onto, and Markov, then the natural invariant measure, if it exists, may be
computed as follows:

14
 1. Let {A1 , A2 , . . . , Ak } be intervals intersecting only at their endpoints such that ∪ki=1 Ai =
[0, 1], and such that f |Ai has constant slope greater than 1 on each interval Ai . Also
assume that the image of each Ai under f is exactly a union of some of the subintervals
Ai .

2. Define the Z-matrix of f to be a k × k matrix whose (i, j) entry is the reciprocal of the
absolute value of the slope of the map from Aj to Ai , or zero if Aj does not map across
Ai . The Z-matrix has no eigenvalues greater than 1 in magnitude, and 1 is always an
eigenvalue.

3. Any eigenvector x = (x1 , x2 , . . . , xk ) corresponds to an invariant measure of f , when

properly normalized so that the total measure is 1. The normalization is done by
dividing x by x1 L1 + x2 L2 + · · · + xk Lk , where Li is the length of Ai . The ith component
of the resulting eigenvector gives the height of the density pi on Ai . Let p(x) be the
function on [0, 1] defined by


 p1 , x ∈ A◦1

 p2 , x ∈ A◦2
p(x) = .. ..

 . .

 p ,
k x ∈ A◦k

with the values of p(x) at the endpoints of the Ai s not important.

4. There may be several different invariant measures for f if the space of eigenvectors
associated to the eigenvalue 1 has dimension greater than one. An eigenvalue is called
simple of this is not the case. If 1 is a simple eigenvector and all the other eigenvalues
of the Z-matrix are strictly smaller than 1 in magnitude, then the resulting measure is
a natural measure.

5. An invariant measure constructed this way then is computed for each closed subinterval
[a, b] of [0, 1] by Z
µ[[a, b]] = p(x) dx.
[a,b]

Note that the µ-measure of [0, 1] is 1, so the invariant measure is a probability measure.

For the next few propositions, lemmas, and theorems we assume that X is a compact
metric space, f : X → X is continuous, and µ is an invariant measure on X with respect

15
to f such that µ is regular and µ(O) > 0 for each nonempty open set O in X. We suppose
further that Y = lim(X, f ), and m denotes the invariant measure induced by µ.
←−

Theorem 7. The induced measure, m, is F -invariant.

Proof. Let K ⊆ Y be closed. Then since Y is compact, K is compact. By definition, if

x = (x0 , x1 . . . ) ∈ Y then F −1 (x) = (x1 , x2 , . . . ). Thus πn+1 [K] = πn ◦ F −1 [K]. So if
Kn = πn−1 ◦ πn [K] is the nth tower set for K, it is the n − 1’st tower set for F −1 [K]. By
definition, m[K] = lim m[Kn ] = lim m[Kn−1 ] = m[F −1 (K)]. Thus m is F -invariant.
n→∞ n−1→∞

Proposition 2. If U is an open nonempty subset of Y , m(U ) > 0.

Proof. Suppose not, i.e., suppose U is a nonempty open set in Y and m(U ) = 0. Applying
Lemma 4, for each ² > 0, there are a positive number δ and a positive integer n such
that for each x in X, diam(πn−1 (Dδ (x)) < ². Therefore we can find x ∈ X, an integer n,
and a positive number δ such that πn−1 (Dδ (x)) ⊂ U . Then m[πn−1 (Dδ (x))] = 0. However,
m[πn−1 (Dδ (x))] = µ[Dδ (x)], by Lemma 8 and this means that m[πn−1 (Dδ (x))] = µ[Dδ (x)] > 0.
This is a contradiction to µ being regular.

Proposition 3. If Z is a measurable subset of X such that µ(Z) = 0, then πn−1 (Z) has
empty interior for each nonnegative integer n and m[πn−1 (Z)] = 0.

Proof. That m[πn−1 (Z)] = 0 follows from Lemma 8. That πn−1 (Z) has empty interior follows
from Proposition 2.

Proposition 4. Suppose that x ∈ Y , O is open in X, and k is a nonnegative integer. Then

#{F i (x) ∈ πk−1 (O) : 0 ≤ i ≤ n} #{f i (xk ) ∈ O : 0 ≤ i ≤ n}

lim = lim ,
n→∞ n n→∞ n
provided one of these limits exists.

Proof. Note that for x = (x0 , x1 , . . .) ∈ Y , F i (x) = (f i (x0 ), f i (x1 ), . . .). Thus, F i (x) ∈
πk−1 (O) if and only if f i (xk ) ∈ O. It follows that #{F i (x) ∈ πk−1 (O) : 0 ≤ i ≤ n} =
#{f i (xk ) ∈ O : 0 ≤ i ≤ n} for each n, and the result follows.

16
 Notation: For x ∈ Y , A a measurable subset of Y , let
i
e A) = lim #{F (x) ∈ A : 0 ≤ i ≤ n} ,
G(x,
n→∞ n
provided this limit exists. For x ∈ X, A a measurable subset of X, let

#{f i (x) ∈ A : 0 ≤ i ≤ n}
G(x, A) = lim ,
n→∞ n
provided this limit exists.

Lemma 9. Assume that X is a compact space with a Lebesgue measure λ. Also assume that
µ is a natural invariant measure on X. If A is a closed tower set in Y , then there is some non-
negative integer r such that A = {y ∈ Y : yr ∈ πr (A)} and µ[πr (A)] = m[A]. Furthermore,
e πr−1 (D² (πr (A)))) = lim G(xr , D² (πr (A))) =
for almost every x ∈ Y , µ[πr (A))] = lim G(x,
²→0 ²→0
m[A].

Proof. Suppose ² > 0. Since µ is a natural measure, there is a measurable set Z of measure
0 in X such that if x ∈
/ Z, then for every closed set S in X, µ[S] = lim G(x, D² (S)). By
²→0
−1
Proposition 4, ZY := ∪∞
k=0 πk (S) is a measurable set of measure 0 in Y .

Suppose then that x ∈ Y \ZY . Then for each k, πk (x) = xk ∈

/ Z. Thus, xm ∈
/ Z. Then,
by the previous proposition, for each ε > 0 and nonnegative integer n,

#{f i (xr ) ∈ D² (πr (A)) : 0 ≤ i ≤ n} = #{F i (x) ∈ πr−1 (D² (πr (A))) : 0 ≤ i ≤ n}.

Thus,

#{f i (xr ) ∈ D² (πr (A)) : 0 ≤ i ≤ n}

G(xr , D² (πr (A))) = lim
n→∞ n
e −1
= G(x, πr (D² (πr (A))))
#{F i (x) ∈ πr−1 (D² (πr (A))) : 0 ≤ i ≤ n}
= lim .
n→∞ n
e π −1 (D² (πr (A)))) = G(xr , D² (πr (A))) for
Since µ[πr (A)] = lim G(xr , D² (πr (A))) and G(x, r
²→0
each ² > 0. Because A is a tower set, the result follows.

Lemma 10. Assume that X is a compact space with a Lebesgue measure λ. Also assume that
e D² (A)) ≤ m(A)
µ is a natural invariant measure on X. Let A ⊆ Y be closed. Then lim G(x,
²→0
for every x not contained in a set Z with πn [Z] having Lebesgue measure zero for each n ∈ N.

17
Proof. For each r ∈ N let Ar be the rth tower set for A, i.e. Ar = πr−1 ◦ πr (A) = {x ∈
Y : xr ∈ πr (A)}. Let Zr ⊆ X be a set of µ-measure zero such that if x 6∈ Zr then
e π −1 (Dδ (πr (Ar )))) =
lim G(x, Dδ (πr (Ar ))) = µ[πr (Ar )]. By Lemma 9, for all x 6∈ π −1 (Zr ), lim G(x,
r r
δ→0 δ→0
m[Ar ].
S∞
Let x 6∈ t=1 πt−1 (Zt ) = Z, and let r ∈ N. For all sufficiently small ² > 0 there is a δ² > 0
such that D² (A) ⊆ πr−1 (Dδ² (πr (A))). Fix ² > 0 small enough and δ² > 0 as above.
Define
© ª
Ln = # F i (x) ∈ D² (A) : 0 ≤ i ≤ n

and
© ª
Mn = # F i (x) ∈ πr−1 (Dδ² (πr (A))) : 0 ≤ i ≤ n

Since D² (A) ⊆ πr−1 (Dδ² (πr (A))) we see that 0 ≤ Ln ≤ Mn . Obviously, 0 ≤ Lnn ≤ Mnn , and by
Proposition 4 lim Mn = G(x,e πr−1 (Dδ² (πr (A)))) exists. So 0 ≤ lim Ln ≤ lim Mn , and thus
n→∞ n n→∞ n n→∞ n
0≤ lim Ln e D² (A)) exists and is less than or equal to G(x,
= G(x, e π −1 (Dδ² (πr (A)))).
r
n→∞ n
Let
e D² (A))
N² = G(x,

and
e πr−1 (Dδ (πr (A))))
Pδ = G(x,

By above, for all sufficiently small ² there is a δ² with N² ≤ Pδ² . As ² → 0 we can choose δ²
so that δ² → 0. By Lemma 9, m[Ar ] = lim Pδ , and thus we can pass to a subset of the δ’s
δ→0
m[Ar ] = lim Pδ² ≥ lim N² . Thus, for each r ∈ N,
²→0 ²→0

e D² (A))
m[Ar ] ≥ lim G(x,
²→0

e D² (A)) for all x 6∈ Z.

This implies that m[A] ≥ lim G(x,
²→0

Theorem 8. Assume that X is a compact space with a Lebesgue measure λ. Also assume
that µ is a natural invariant measure on X. Then for every closed set A ⊆ Y there is a set
Z ⊆ Y such that πn [Z] has Lebesgue measure zero for each n ∈ N, and

e D² (A))
m[A] = lim G(x,
²→0

for all x 6∈ Z.

18
Proof. Suppose A is a closed subset of Y , and, for each nonnegative integer n, An denotes
the nth tower set for A. There is a measurable set Zn ⊆ X of measure zero such that if
e π −1 (D² (πn (A)))). Let x 6∈ S π −1 [Zn ] = Z. Let
∞
x ∈ Y \πn−1 [Zn ], then m[An ] = lim G(x, n n
²→0 n=1
n, r, k ∈ N and define
e D1/n (A))
Qn = G(x,

and
e π −1 (D1/k (πr (A))))
Pr,k = G(x, r

Notice that if we fix r ∈ N then m(Ar ) = lim Pr,k and m(A) = lim m(Ar ) = lim lim Pr,k .
k→∞ r→∞ r→∞ k→∞
By Lemma 4, for each n ∈ N, there is some rn , kn ∈ N so that πr−1
n
(D1/kn (πrn (A))) ⊆ D1/n (A).
For this n, rn , kn ∈ N we have

e π −1 (D1/kn (πrn (A)))) ≤ G(x,

Prn ,kn = G(x, e D1/n (A) = Qn
rn

As n → ∞, rn , kn → ∞. So

m(A) = lim Prn ,kn ≤ lim Qn

n→∞ n→∞

e D² (A)) we see that

Since lim Qn = lim G(x,
n→∞ ²→0

e D² (A))
m(A) ≤ lim G(x,
²→0

Combining this with the previous lemma,

e D² (A))
m(A) = lim G(x,
²→0

for every x ∈ Y \ Z. We showed in Theorem 7 that m is F -invariant.

In this case we say that m is a natural invariant measure on lim(X, f ).

←−

3.3 Integration
Let X be a compact metric space with f : X → X continuous and µ a natural invariant
measure. Let g : lim(X, f ) → R be a continuous function. We now show how to integrate g
←−
on lim(X, f ) by approximating g with a sequence of simple functions defined on tower sets.
←−
This sequence of simple functions will converge in measure to g and so we can approximate
the integral of g by integrating each of these simple functions.

19
 Let Y = lim(X, f ). The first step in constructing any simple function on Y is to choose
←−
a partition {Ei }ni=1 of the space Y : Let k be a positive integer and define a positive number
1 1
δk ≤ 2k
and an integer nk such that if x ∈ X, then πn−1
k
(Dδk (x)) has diameter less than k
for each x ∈ X (the existence of such values is guaranteed by Lemma 4). Also, define a
partition Pk = {Pk,1 , Pk,2 , . . . , Pk,Lk } of X with the following properties:

(1) mesh(Pk ) < δk , and

(2) f nk −nk−1 (Pk ) refines Pk−1 .

Define also a collection of sample points Qk = {xk,1 , xk,2 , . . . , xk,Lk } ⊂ X such that for
each 1 ≤ i ≤ Lk , xk,i ∈ Pk,i .
Next we translate these sets and sample points into the inverse limit space:
Let
bk = {Pbk,i }Lk := {πn−1 (Pk,i )}Lk ,
P i=1 k i=1

and let
bk = {xk,i : 1 ≤ i ≤ Lk },
Q

where each xk,i is some member of πn−1 bk

(xk,i ). For each integer k, this gives us a partition P
k

of the inverse limit space with the following properties:

bk ) < 1 ,
(1) mesh(P k

bk refines P
(2) P bk−1 , and

bk ] = µ[Pk ].
(3) m[P

Also, we have chosen a representative point xk,i out of each element Pbk,i . We can now
define our sequence of simple functions. For each x ∈Y and each positive integer k, let
Lk
X
gk (x) = g(xk,i )χPbk,i (x).
i=1

In order to demonstrate that g is integrable, we need to show that the sequence {gk }∞
k=1 is

mean fundamental and that it converges in measure to g.

Lemma 11. The sequence {gk }∞

k=1 is mean fundamental.

20
Proof. Let ² > 0, and let δ > 0 so that g(Dδ (x)) has diameter less than ² for all x ∈ Y .
bk ) < δ. Let x ∈ Y
Choose k, r to be positive integers with r > k large enough so that mesh(P
with x not in the boundary of Pbk,i or Pbr,i for all i. Notice that all of these boundaries are
sets of the form πt−1 (S), where S is the boundary of a member of the partition Pt (t = r
or t = k), so µ(S) = 0 = m(πt−1 (S)). Hence, by restricting our attention to points of the
inverse limit space which are not in any of these boundaries, we are only omitting a set of
measure zero.
Consider ¯ ¯
¯XLr Lk
X ¯
¯ ¯
|gr (x) − gk (x)| = ¯ g(xr,i )χPbr,i (x) − g(xk,i )χPbk,i (x)¯ .
¯ ¯
i=1 i=1

Choose q, s to be positive integers such that x ∈ Pbr,q ⊆ Pbk,s . Then, since x is not on
the boundary of any of the partition elements, gr (x) = g(xr,q ) and gk (x) = g(xk,s ). Since
mesh(Pbk ) < 1 , |xr,q − xk,s | < δ. Thus,
k

|gs (x) − gk (x)| = |g(xr,q ) − g(xk,s )| < ².

Hence, Z Z
ρ(gr , gk ) = |gr − gk | dm < ² dm
R
and since ² dm = ²(m(Y )) = ², we have established that ρ(gr , gk ) → 0 as r, k → ∞.

Lemma 12. The sequence of simple functions {gk }∞

k=1 converges in measure to g.

Proof. Let ² > 0, and again choose δ > 0 so that g(Dδ (x)) has diameter less than ² for all
x ∈ Y . Choose M large enough so that if k > M , then mesh(Pbk ) < δ.
bk , and let r
Pick k ≥ M . Choose x so that x is not on the boundary of any element of P
be a positive integer such that x ∈ Pbk,r . Then

|gk (x) − g(x)| = |g(xk,r ) − g(x)|

and since both x and xk,r are in Pbk,r , d(x, xk,r ) < δ. Thus, |g(xk,r ) − g(x)| < ², which
implies that for all x not on the boundary of any element of P bk , |gk (x) − g(x)| < ². Thus,
m[{x : |gk (x) − g(x)| ≥ ²}] = 0 for all k ≥ M , and lim m[{x : |gk (x) − g(x)| ≥ ²}] = 0.
k→∞

21
4 Computational Issues
4.1 Inverse Limit Case
Let f : X → X be a continuous map on a compact metric space X. Let µ be an invariant
measure of f and m be the induced invariant measure on the inverse limit space Z :=
lim(X, f ). Let W : Z → R be a continuous real-valued function on Z defined by
←−

∞
X
W (x) := β t U (xt ),
t=0

where U : X → R is continuous. The proof used to construct the measure on the inverse
limit space in Section 3 suggests the following algorithm for approximating the integral
Z
W (x)m(dx).
Z

First, we truncate the infinite sum at some value T . Gridding the state space at time T into
N + 1 equally spaced points partitions the state space at time T in N adjacent intervals.
Each of these intervals can be used to partition the inverse limit space according to

{f T (Ij )} × {f T −1 (Ij )} × · · · × {f (Ij )} × {Ij } × {f −1 (Ij )} · · · .

Each of these “tunnels” in the inverse limit space has measure µ(Ij ). Let Hj := {f T (Ij )} ×
{f T −1 (Ij )}×· · ·×{f (Ij )}×{Ij } denote a truncated tunnel. Then our integral is approximated
by
Z T X
X N
W (x)m(dx) ≈ β t U (xjt )µ(Ij ),
Z t=0 j=1

where {xj0 , xj1 , . . . , xjT } ∈ Hj . One issue that needs to be addressed is how sensitive is
the value of the approximation to the length T , the number (and size) of subintervals and
the selections of the point {xj0 , xj1 , . . . , xjT } ∈ Hj . The selection point is potentially very
problematic due the sensitivity of initial conditions inherent in the map f . Ideally, one would
want the projection of tunnel πi (Hj ) to be small for each i. However, for a given number
of subintervals, one may be able to select T large enough so the f j (Ik ) essentially equals
the entire state space X. The problem of sensitive dependence on initial conditions puts
our two objectives at odds with each other: as we increase T to make the tail of the utility
function small we must also increase N to ensure f j (Ik ) is small for all j = 0, 1, . . . , T . Our
initial results using this algorithm with FORTRAN were very discouraging. The number of

22
N needed to keep f j (Ij ) small first required using special subroutines to conduct arbitrary
precision computations (the default using FORTRAN type “quad real” was not precise
enough). Second, the time required to compute the integral using these special subroutines
made this approach impractical (estimated to take years using a Pentium IV desktop).
Fortunately, Birkhoff’s ergodic theorem can be used to speed things up immensely.
The integral of the truncated sum is given by
T Z
X
β t U (f T −t (x))µ(dx).
t=0 I

Since µ is an f -invariant measure, the Birkhoff ergodic theorem tells us

Z Z
T −t
U (f (x))µ(dx) = U (f (x))µ(dx).
X X

Consequently, our integral is given by

Z Z
1
W (x)m(dx) = U (x)µ(dx).
Z 1−β X

4.2 Direct Limit Case

Let f : X → X be a continuous map on a compact metric space X. Let µ be an invariant
measure of f and m be the induced invariant measure on the direct limit space D :=
lim(X, f ). Let W : D → R be a continuous real-valued function on D defined by
−→

∞
X
W (x) := β t U (xt ),
t=0

where U : X → R is continuous. The proof used to construct the measure on the inverse
limit space in Section 3 suggests the following algorithm for approximating the integral
Z
W (x)m(dx).
Z

Gridding the state space at time 0 into N + 1 equally spaced points partitions the state
space at time 0 in N adjacent intervals. Each of these intervals can be used to partition the
inverse limit space according to

Hj := {Ij } × {f (Ij )} × {f 2 (Ij )} · · · .

23
Each of these “tunnels” in the inverse limit space has measure m(Hj ) = µ(Ij ). Let xj :=
(xj , f (xj ), f 2 (xj ), . . .) ∈ Hj . Then our integral is approximated by
Z N
X N X
X ∞
j
W (x)m(dx) ≈ W (x )m(Hj ) = β t U (f t (xj ))µ(Ij )
Z j=1 j=1 t=0

This implies that

Z ∞ Z
X
W (x)m(dx) = β t U (f t (x))µ(dx).
Z t=0 X

The integral of the truncated sum is given by

∞ Z
X
β t U (f t (x))µ(dx).
t=0 I

Since µ is an f -invariant measure, the Birkhoff ergodic theorem tells us

Z Z
U (x)µ(dx) = U (f n (x))µ(dx).
X X

Consequently, our integral is given by

Z Z
1
W (x)m(dx) = U (x)µ(dx).
Z 1−β X

5 Ramsey Meets Chaos in a Cash-in-Advance Model

5.1 Optimal Policy with Multiple Equilibria
The framework in this paper for calculating expected utility can be used to bridge two
important literatures in macroeconomic theory: multiple equilibria and optimal policy. Dy-
namic general equilibrium (DGE) models have become a standard framework for both the
positive and normative evaluation of policy. In the optimal monetary/fiscal policy literature
one considers a mapping from a policy space (e.g., money growth rate or set of taxes) to
outcomes (e.g. allocations from a competitive equilibrium). If the mapping from policies to
outcomes in the DGE model is single-valued, then one can induce a ranking on the policy
space in a very natural way. For instance, suppose Θ is the policy space and for each θ ∈ Θ,
there is a unique competitive equilibrium E given by E = M (θ). If U the utility function
of the household defined over the space of competitive equilibria, then one can use the func-
tion W (θ) := U (M (θ)) to define a ranking on Θ. In addition to perhaps locating the most
preferred or optimal policy θ∗ , such a ranking can be used to measure the welfare gains of

24
switching from some policy θ to another policy θ0 . There is a large literature that takes this
approach to evaluating polices starting with the work of Ramsey (1927).
However, when H is not single-valued this method of ranking polices will not work, and
it is not clear what one should do since there is more than one equilibrium associated with a
particular policy. There are many ways in which M may be multi-valued. For example, the
model may exhibit local indeterminacy in which for a given policy θ there exists a continuum
of equilibria all converging to the steady state equilibrium. However, one may also have a
multi-valued H due to global properties of the model as well. Our framework can be applied
to the class of economic models with equilibria that correspond to orbits generated by a
chaotic dynamical system f : X → X where X is a compact metric infinite space and f
is continuous. Thus there is both a large number of equilibria and a large and complicated
variety as well. Our framework is designed for this type of multi-valued H. Note that if
f represents the backward map, the indeterminacy in the model is greater in the following
sense. If f represents the forward map, there is a unique equilibrium associated with each
x ∈ X. However, if f is the backward map, there is at least one equilibrium (and perhaps
an infinite number of equilibria) associated with with each x ∈ X.

5.2 Cash-in-Advance Model

The model is the standard endowment CIA model of Lucas and Stokey (1987). We closely
follow the exposition of Michener and Ravikumar (1998), hereafter [MR]. Since our intent is
only to apply our techniques to calculate expected utility in a model with backward dynam-
ics and chaos, we will focus on a particular family of utility functions and parameterizations
used in [MR].6 It is an endowment economy with both cash and credit goods. There is a rep-
resentative agent and a government. The government consumes nothing and sets monetary
policy using a money growth rule.
The household has preferences over sequences of the cash good (c1t ) and credit good (c2t )
represented by a utility function of the form
∞
X
β t U (c1t , c2t ), (3)
t=0

with the discount factor 0 < β < 1. The utility function is assumed to take the following
6
See [MR] for more details and a more general framework.

25
form:
c1−σ
1 c1−γ
U (c1 , c2 ) := + 2 ,
1−σ 1−γ
with σ > 0 and γ > 0. To purchase the cash good c1t at time t the household must have cash
mt . This cash is carried forward from t − 1 and in this sense the household is required to
have cash in advance of purchasing the cash good. The credit good c2t does not require cash,
but can be bought on credit. The household has an endowment y each period that can be
transformed into the cash and credit goods according to c1t + c2t = y. Since this technology
allows the cash good to be substituted for the credit good one-for-one, both goods must sell
for the same price pt in equilibrium and the endowment must be worth this price per unit
as well.
The household seeks to maximize (3) by choice of {c1t , c2t , mt+1 }∞
t=0 subject to the con-

straints c1t , c2t , mt+1 ≥ 0,

pt c1t ≤ mt , (4)
mt+1 ≤ pt y + (mt − pt c1t ) + θMt − pt c2t , (5)

taking as given m0 and {pt , Mt }∞

t=0 . The money supply {Mt } is controlled by the government

and follows a constant growth path Mt+1 = (1 + θ)Mt where θ is the growth rate and M0 > 0
given. Each period the household receives a transfer of cash from the government in the
amount θMt .
A perfect foresight equilibrium is defined in the usual way as a collection of sequences
{c1t , c2t , mt }∞ ∞
t=0 and {Mt , pt }t=0 satisfying the following. (1) The money supply follows the

stated policy rule: Mt+1 = (1 + θ)Mt . (2) Markets clear: mt = Mt and c1t + c2t = y. (3) The
solution to the household optimization problem is given by {c1t , c2t , mt+1 }∞
t=0 .

Since this utility function satisfies Assumption 1 in [MR], the solution to this problem
will be interior (strictly positive). The necessary first-order conditions from the household’s
problem imply that
U2 (c1t , c2t )/pt = βU1 (c1t+1 , c2t+1 )/pt+1 , (6)

where Ui is the partial derivative of U with respect to the ith argument. This condition
reflects that at the optimum, the household must be indifferent between spending a little more
on the credit good (giving a marginal benefit U2 (c1t , c2t )/pt ) versus savings the money and
purchasing the cash good in the next period (giving a marginal benefit βU2 (c1t+1 , c2t+1 )/pt+1 ).

26
 Let xt := mt /pt denote the level of real money balances. Using the equilibrium conditions
that Mt = mt and c2t = y − c1t , equation (6) implies
β
xt U2 (c1t , y − c1t ) = xt+1 U1 (c1t+1 , y − c1t+1 ). (7)
1+θ
Let c be the unique solution to U1 (x, y − x) = U2 (x, y − x). If the cash-in-advance constraint
(4) binds, then c1t = xt . If not, then the Lagrange multiplier µt = 0 and c1t = c. It then
follows that c1t = min[xt , c] for all t. Using this relationship we can eliminate c1t and c1t+1
from (7) to get a difference equation in x alone:
β
xt U2 (min[xt , c], y − min[xt , c]) = xt+1 U1 (min[xt+1 , c], y − min[xt+1 , c])
1+θ
or
B(xt ) = A(xt+1 ), (8)

where

B(x) := xU2 (min[x, c], y − min[x, c]),

β
A(x) := xU1 (min[x, c], y − min[x, c]).
1+θ
Whether or not the model has backward dynamics depends on whether or not A(·) is in-
vertible. In one parameterization, [MR] set β = 0.98, σ = 0.5, γ = 4.5, y = 2 and consider
θ equal to 0, 0.5 and 1.0. In this case the function A is not invertible and there exists an
invariant set [xl , xh ] such that the the backward map has a three cycle. The backward map
for this parameterization (with θ = 0) is in Figure 1. We see that for this parameterization,
the CIA model has backward dynamics.
Our function W : lim(I, f ) → R is given by
←−
∞
X
W (x) := β t U (xt , y − xt ).
t=0

To construct the natural f -invariant measure, we approximate µ via a histogram using a

sample trajectory of f for some x ∈ [x, x] : {x, f (x), f 2 (x), . . .}. This mimics the “rain gauge”
description of the natural invariant measure described in Alligood et al. (1996). Figure 2
contains an approximation of µ. This histogram uses 104 bins and a sample trajectory of
length 108 . Given this approximation to the natural invariant measure, the utility function
U and the discount factor β it is now straight-forward to approximate our integral
Z Z
1
W (x)µ(dx) = U (x)µ(dx) ≈ 83.3285573.
X 1−β I

27
 Figure 1: Backward map f : [xl , xh ] → [xl , xh ] from the cash-in-advance model.

1.02

1.015

1.01

1.005

f(x)
0.995

0.99

0.99 0.995 1 1.005 1.01 1.015 1.02

As mentioned in the introduction, our integral allows us to rank inverse limit spaces
according to expected utility (a very natural ranking from the model). To give some sense
of how this might be used to evaluate different monetary policies, imagine that for money
growth rates θ ∈ Θ := [θ, θ], the backward map f is chaotic. However, not all chaotic maps
are the same in terms of utility. One way of framing the question through a Ramsey lens,
is within this subclass of possible monetary policies Θ, which money growth rate gives the
greatest expected utility? We see that for θ ∈ Θ, we have a different backward map fθ ,
natural invariant measure µθ , invariant state space Iθ , inverse limit space Zθ := lim(Iθ , fθ )
←−
and induced measure mθ . We then have an indirect utility function given by
Z Z
1
V (θ) := W (x)mθ (dx) ≡ U (x)µθ (dx).
Zθ 1 − β Iθ
To be more concrete, suppose that the monetary authority is only considering money growth
rates in Θ := [0, 0.1]. Which θ ∈ Θ should the monetary authority choose to maximize
expected utility? Figure 3 contains a plot of the indirect utility function V : Θ → R. We
see that a lower money growth rate is preferred to higher money growth rate (θ = 0 is the
most preferred). This ranking is qualitatively similar to the ranking when considering only
steady state equilibria.7 However, Figure 4 illustrates that considering only the steady state
7
A priori, this need not have been the case. What is driving the utility results in this example is the
distribution (and its support) for different money growth rates θ.

28
 Figure 2: Histogram with 104 bins using a sample trajectory of length 108

−4
x 10
6

0
0.985 0.99 0.995 1 1.005 1.01 1.015 1.02 1.025

equilibria would underestimate the welfare costs of higher money growth rates.

6 Conclusion
In this paper, we provide a framework for calculating expected utility in models with chaotic
equilibria and consequently a framework for ranking chaos. Our framework is quite general
and applies to any DGE model where the set of equilibria correspond to the orbits generated
by a chaotic dynamical system f : X → X where f is continuous and X is compact. We
have illustrated how this framework can be used to bring together two important literatures
in macroeconomic theory: multiple equilibria and optimal policy.
In future research we hope to extend our framework to DGE models with multiple equi-
libria where the underlying dynamical system does not admit a natural invariant measure.
Such DGE models would include large and important class of models, namely those with
local indeterminacy.

29
 Figure 3: Indirect utility function V : Θ → R.

83.36

83.34

83.32

83.3
Expected Utility

83.28

83.26

83.24

83.22

83.2

83.18

83.16
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1
Money Growth Rate

Figure 4: Comparison of expected utility under chaos versus utility in the steady-state
equilibrium.

83.36

83.34
Chaotic Equilibria
Steady State
83.32

83.3

83.28
Utility

83.26

83.24

83.22

83.2

83.18

83.16
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1
Money Growth Rate

30
 View publication stats

References
Alligood, K. T., Sauer, T. D., Yorke, J. A., 1996. Chaos: an introduction to dynamical
systems. Springer, New York, NY.
Bochner, S., 1955. Harmonic analysis and the theory of probability. University of California
Press, Berkeley and Los Angeles.
Choksi, J. R., 1958. Inverse limits of measure spaces. Proc. London Math. Soc. (3) 8, 321–342.
Devaney, R. L., 2003. An introduction to chaotic dynamical systems, 2nd Edition. Westview
Press, Boulder, Colorado.
Grandmont, J.-M., 1985. On endogenous competitive business cycles. Econometrica 53, 995–
1045.
Halmos, P. R., 1974. Measure Theory. Springer-Verlag, New York, NY.
Hunt, B. R., Kennedy, J. A., Li, T.-Y., Nusse, H. E., 2002. SLYRB measures: natural
invariant measures for chaotic systems. Physica D 170, 50–71.
Ingram, W. T., 2000. Inverse limits, Aportaciones Matemáticas, Investigación 15. Sociedad
Matemática Mexicana, México.
Ingram, W. T., Mahavier, W. S., 2004. Interesting dynamics and inverse limits in a family
of one-dimensional maps. Amer. Math. Monthly 111 (3), 198–215.
Kennedy, J., Stockman, D. R., Yorke, J., 2005. Inverse limits and models with backward
dynamics. preprint, 1–37.
Kennedy, J. A., Stockman, D. R., 2006. Chaotic equilibria in models with backward dynam-
ics. preprint, 1–23.
Kennedy, J. A., Stockman, D. R., Yorke, J. A., 2004. Inverse limits and an implicitly defined
difference equation from economics. Topology and Its Applications forthcoming.
Lucas, R. E., Stokey, N. L., 1987. Money and interest in a cash-in-advance economy. Econo-
metrica 55, 491–513.
Medio, A., Raines, B., 2005. Inverse limit spaces arising from problems in economics.
preprint, 1–19.
Medio, A., Raines, B., 2006. Backward dynamics in economics: the inverse limit approach.
Journal of Economic Dynamics and Control, forthcoming, 1–39.
Michener, R., Ravikumar, B., 1998. Chaotic dynamics in a cash-in-advance economy. Journal
of Economic Dynamics and Control 22, 1117–1137.
Ramsey, F. P., 1927. A contribution to the theory of taxation. The Economic Journal
37 (145), 47–61.

31