ArticlePDF Available

Existence and stability of stationary solutions of nonlinear difference equations under random perturbations

Authors:
  • Shanghai Univeristy of Finance and Economics

Abstract

Existence and stability of stationary solutions of nonlinear random difference equa-tions are studied in this note. Firstly, we give the weak conditions that guarantee the stability of Lypanunov exponents under small random perturbations. Secondly, we find out the conditions under which the ratio of the random norm and the stan-dard Euclidean norm has deterministic bounds. Based on these new results, we provide easy-to-use conditions that guarantee the existence and almost sure stabil-ity of a stationary solution. In addition, we also prove that the stationary solution converges with probability one to the fixed point of the corresponding deterministic system as the noise intensity tends to zero.
Existence and stability of stationary solutions of
nonlinear difference equations under random
perturbations
Mei Zhu
, Duo Wang and Maozheng Guo
LMAM, School of Mathematical Sciences, Peking University, Beijing 100871, P. R. China
Abstract
Existence and stability of stationary solutions of nonlinear random difference equa-
tions are studied in this note. Firstly, we give the weak conditions that guarantee
the stability of Lypanunov exponents un der small random perturbations. Secondly,
we find out the conditions under which the ratio of the random norm and the stan-
dard Euclidean norm has deterministic bounds. Based on these new results, we
provide easy-to-use conditions that guarantee the existence and almost sure stabil-
ity of a stationary solution. In addition, we also prove that the stationary solution
converges with probability one to the fixed point of the corresponding deterministic
system as the noise intensity tends to zero.
Keywords: difference equation; random perturbation; stationary process; Lya-
punov exponents; random norm
AMS Subject C lassification: 39A11; 91B02; 37N35; 37N40
Corresponding author. Email: meizhu66@g mail.com
1
1 Introduction
Random difference equations have been studied in many contexts, particularly in mod-
eling physical and economic systems subject to exogenous random perturbations. For the
random models, one of the basic problems is the long-term behavior of solutions, in partic-
ular the existence and stability of stationary solutions of the systems. Theoretical study
in this direction is still far from maturity. Many complex practical models described by
nonlinear random difference equations have to be analyzed by numerical simulations (see
[4, 5, 8]).
Generally, r andom difference equations are tr eated by studying invariant measures or
stationary measures. Ara´ujo [1], Futia [7] and Zmarrou & Homburg [14] provided different
conditions to guarantee the existence of invariant measures from different perspectives.
Kifer [10] further studied the connections of properties of invariant measures in the ra ndom
case with those in the deterministic case.
Furthermore, the existence of stationary solutions for random difference equations
is also studied, which implies the existence of invariant measures. Arnold [2] gave the
stationary processes of hyperbolic affine random dynamical systems. Tong [12] studied
the existence of stationary solutions for a special class of random difference equations
where the Jacobian matrices in the linear parts are deterministic. For more general
nonlinear ra ndom difference equations, Arnold & Boxler [3] provided a theorem that
additive noise turns a hyperbolic fixed point into a stationary solution. However, the
conditions of the theorem ar e related to Lyapunov exponents and random norms, which
make it difficult to apply directly theorem to practical models. Therefore, we try to
develop a new mathematical tool to build up a bridge between the theoretical study and
practical applications.
In this paper we provide easy-to-use conditions that guarantee the existence, stability
and convergence of stationary solutions. Firstly, we give weaker conditions to generate
the stability of Lyapunov exponents under small random perturbations by using some
ideas of [13]. In fact, Young [13] studied a more general case. However, some practical
models (see [15]) do not satisfy the conditions given by [13]. In order to deal with the spe-
cial cases related to the practical models, we provide some weaker conditions than those
given by [13]. Consistent with [13], in this paper we also concentrate on two-dimensional
random difference equations. Secondly, to our knowledge, for the first time we give the
conditions to guarantee that the ratio of the random norm and the standard Euclidean
2
norm has deterministic bounds, which plays a crucial role in obtaining the theorem and
propositions of this paper. By using the relationship between the random norm and the
standard Euclidean norm and some ideas of [6], we find out a deterministic neighbor-
hood in which the Lipschitz constant in the case of random norm is deterministic and
sufficiently small. Based on these new results and the idea of [3], we provide easy-to-use
conditions that g uar antee the existence of stationary solutions of nonlinear random dif-
ference equations. In addition, we further obta in that the stationary solutions a re locally
asymptotically stable and converge to the corresponding fixed po ints of the deterministic
system as the noise intensity tends to zero both in the almost sure sense, which implies
that the stationary measures also converge weakly to the Dirac measure supported on
the corresponding deterministic fixed point as the intensity of noise tends to zero. These
properties are significant for studies of practical models.
In order to obtain the above results, here we make two basic assumptions. On the one
hand we assume that the perturbations are bounded. This assumption is motiva t ed by
some pra ctical backgrounds, for example, dividends in financial markets var y randomly
and in a bounded range. See also [14] for reasons o f considering bounded noise. On the
other hand we assume that the random map is C
2
not only with respect to state variables
but also with respect t o random perturbations. In fact, many pra ctical models are all
analytic with respect to all va r ia bles and parameters as well as random perturbations.
Araujo [1] a nd Zmarrou & Homburg [14] have the similar assumptions.
The rest of this paper is set out as follows. Section 2 introduces the preliminaries and
the main theorem. Section 3 provides three lemmas. Section 4 proves the main theorem.
Section 5 presents several properties of stationary solutions. Section 6 gives a conclusion.
2 Random difference equation
Now consider nonlinear difference equation
Y
n
= A(ǫε
n
)Y
n1
+ H(ǫε
n
, Y
n1
) + h(ǫε
n
), n Z, (2.1)
where Y
n
R
2
, A(·) C
1
([b, b], Gl(2, R)), H(·, ·) C
2
([b, b] × R
2
, R
2
), h(·) C
2
([b, b], R
2
). Here b is a positive number. ǫε
n
denotes the random perturbation, where
ǫ( [0, 1]) is a constant revealing the corresponding perturbation intensity. As ǫ = 0,
the difference equation (2.1) is deterministic. Suppose that {ε
n
} is independent and
ident ically distributed (i.i.d.) with noise ε
n
independent of {Y
m
, m < n}. Further, let
3
ν denote the probability measure of ε
n
on [b, b]. Let := [b, b]
Z
be the bi-infinite
product of the interval [b, b] with itself. For each ω Ω, let its n’th coordinate (ω)
n
equal some value of ε
n
. That is, ω = (· · · ε
1
, ε
0
, ε
1
· · · ). Further, let σ : be the
shift op erator, i.e., (σω)
n
= (ω)
n+1
.
For convenience of expression, set
Y
n
= A(ǫε
n
)Y
n1
+ H(ǫε
n
, Y
n1
) + h(ǫε
n
)
:= A
ǫ
(ε
n
)Y
n1
+ H
ǫ
(ε
n
, Y
n1
) + h
ǫ
(ε
n
). (2.2)
Furthermore, we assume
A
ǫ
(ε
n
) = A(ǫε
n
) =
a
1
(ǫε
n
) a
2
(ǫε
n
)
a
3
(ǫε
n
) a
4
(ǫε
n
)
:=
a
1
ǫ
(ε
n
) a
2
ǫ
(ε
n
)
a
3
ǫ
(ε
n
) a
4
ǫ
(ε
n
)
.
Define the linear cocycle
Φ
ǫ
(n, ω) :=
A
ǫ
(ε
n
) · · · A
ǫ
(ε
2
)A
ǫ
(ε
1
), for n 1,
id, for n = 0,
where id denotes a 2 × 2 identity matrix. When ǫ = 0, A
0
(ε
n
) = A(0) is a constant
matrix independent of noise. Let A
0
:=
a
1
0
a
2
0
a
3
0
a
4
0
denote A
0
(ε
n
) a nd Φ
0
(n) denote
Φ
0
(n, ω) = A
n
0
.
Let u P
1
denote the direction represented by the unit vector u R
2
, where P
1
is
the projective space on R
2
. For sufficiently small ǫ and t [b, b], define a map A
ǫ
(t)(u)
: P
1
P
1
by A
ǫ
(t)(u) = A
ǫ
(t)u. Then the map A
ǫ
(ε
n
) induces a Markov chain on P
1
.
Further, let E be a Borel subset of P
1
. The transition probabilities ar e defined by
P
ǫ
(u, E) = ν{t [b, b] : A
ǫ
(t)(u) E}.
In the following proof, we also need the definition and some properties of the random
norm. First of all, we give the definition of the random norm. Here we use a definition
similar to t hat in [3].
Assume that Lyapunov exponents {λ
ǫ
i
} of Φ
ǫ
satisfy λ
ǫ
2
λ
ǫ
1
< 0. Set λ
ǫ
s
:= λ
ǫ
1
. Fix a
κ(ǫ) > 0 such that λ
ǫ
s
+ κ(ǫ) < 0. For any Y : R
2
, let
kY (ω)k
ω
:=
X
i=0
e
(λ
ǫ
s
+κ(ǫ))i
kΦ
ǫ
(i, ω)Y (ω)k (2.3)
4
denote the random norm of Y w.r.t the cocycle Φ
ǫ
(n, ω), where k · k is the standard
Euclidean norm in R
2
. The crucial advantage of this random norm is that for all Y :
R
2
kΦ
ǫ
(n, ω)Y (ω)k
σ
n
ω
e
(λ
ǫ
s
+κ)n
kY (ω)k
ω
, n = 0 , 1, 2, · · · . (2.4)
This is because, for n = 0, 1, 2, · · ·
kΦ
ǫ
(n, ω)Y (ω)k
σ
n
ω
e
(λ
ǫ
s
+κ)n
X
i=n
e
(λ
ǫ
s
+κ)(i+n)
kΦ
ǫ
(i + n, ω)Y (ω)k = e
(λ
ǫ
s
+κ)n
kY (ω)k
ω
.
Further, define the random norm for the linear operator Φ
ǫ
(n, ω) by
kΦ
ǫ
(n, ω)k
ω ,σ
n
ω
= sup
kY k
ω
6=0
kΦ
ǫ
(n, ω)Y (ω)k
σ
n
ω
kY (ω)k
ω
.
Hence based on (2.4),
kΦ
ǫ
(n, ω)k
ω ,σ
n
ω
e
(λ
ǫ
s
+κ)n
.
Note that the above random norms for vectors and matrices as well as κ all depend on
the noise intensity ǫ.
Set k|Y k| := ess sup kY (ω)k
ω
.
The main result of this paper is to provide a theorem that guarantees the existence of
stationary solutions to the nonlinear random difference equation (2.1) , i.e. the following
Theorem 2.1. Its proof will be given in Section 4.
Theorem 2.1 Fo r random difference equation (2.1), assume
(i) The two eigenva l ues of A(0) are both n onzero and less than one in modulus,
(ii) For an y given u P
1
, the probability measure P
ǫ
(u, ·) in duced by A
ǫ
(ε
1
)(u) is abso-
lutely continuous with respect to Lebesgue measure m(·) on P
1
,
(iii) H(ǫε
n
, 0) = 0, D
Y
H(ǫε
n
, 0) = 0 for any ǫ [0, 1] and any n Z,
(iv) h(0) = 0.
Then for sufficiently small ǫ, there exists ex a ctly one stationary solution in the neighbor-
hood of the origin whose initial value ξ
ǫ
(ω) satisfies
k|ξ
ǫ
k| Kǫ, (2.5)
where K > 0 i s a constant i ndependent of ǫ.
5
Remark 2.1 Conditions (i), (iii) and (iv) imply that the origin 0 is a stable fixed point
of the deterministic difference equation Y
n
= A(0)Y
n1
+ H(0, Y
n1
) + h(0). In fact, any
random difference equation Y
n
= F (ǫε
n
, Y
n1
) can be tra nsformed into equation (2.1),
where F (·, ·) C
2
([b, b] × R
2
, R
2
), if its correspo nding deterministic difference equation
Y
n
= F (0, Y
n1
) has a fixed point.
Remark 2.2 If the measure ν is absolutely continuous with respect to Leb esgue measure
on [b, b] and a rctan
a
3
ǫ
(t) cos θ+ a
4
ǫ
(t) sin θ
a
1
ǫ
(t) cos θ+ a
2
ǫ
(t) sin θ
is monotone or piecewise monotone with respect to
t for any given θ [
π
2
,
π
2
], then the condition (ii) in Theorem 2.1 holds.
Remark 2.3 Since ǫ = 0 is a degenerate case where 0 is a trivial solution, in the following
we consider the case that ǫ > 0 only.
3 Several lemmas
In order to prove Theorem 2.1, at first we provide several lemmas.
Lemma 3.1 Let A(·) : [b, b] Gl(2, R) be C
1
and A(0) be invertible. Assume that the
transition probability P
ǫ
(u, ·) is absolutely continuous with respect to Lebesgue measure
m(·) on P
1
for any given u P
1
. Then the Lyapunov exponents λ
ǫ
i
λ
0
i
as ǫ 0,
i = 1, 2.
Proof. In the case that λ
0
1
> λ
0
2
, there are two unit eigenvectors {u
i
} corresponding
to {λ
0
i
} respectively for the matrix A
0
= A(0). Let θ
i
denote the angle corresponding to
the vector u
i
. Since P
1
=
S
1
/{θ, θ + π}, then we can identify P
1
with S
1
/{θ, θ + π} and
use θ instead of ¯u. Further, let U (S
1
/{θ, θ + π}) denote the space of Borel probability
measures on S
1
/{θ, θ + π}. Define the operator L
ǫ
on U (S
1
/{θ, θ + π}) by
L
ǫ
(·)(E) =
Z
P
1
P
ǫ
(θ, E) d(·).
Then a probability measure µ on S
1
/{θ, θ + π} is an invariant measure if and only if for
any Borel subset E in S
1
/{θ, θ + π}
L
ǫ
(µ)(E) = µ(E).
Invariant measures on compact space S
1
/{θ, θ + π} always exist and are absolutely con-
tinuous with respect to Lebesgue measure, because we assume that P
ǫ
(θ, ·) is absolutely
continuous with respect to Lebesgue measure for any θ.
6
Following t he similar argument as in [13], the maximal Lypunov exponent λ
ǫ
1
can be
expressed by
λ
ǫ
1
=
Z
P
1
×[b,b]
log kA(t
)uk d(µ
ǫ
× ν
ǫ
),
where µ
ǫ
is any invariant measure on P
1
and ν
ǫ
is the measure induced by t
= ǫt on
[ǫb, ǫb]. Hence in order to prove the lemma it is sufficient to prove that the invariant
measure µ
ǫ
converges weakly to δ
u
1
as ǫ 0.
If we define a sequence of probability measures from any probability measure µ by
1
n
n1
X
k=0
(L
k
ǫ
µ), µ U (S
1
/{θ, θ + π}), n = 1, 2, 3, · · · ,
then any (weak) limit of this sequence is an invariant measure, reversely any invariant
measure can be constructed in this way. See for example, [2].
Thus to show the lemma, it suffices to show that there is a small neighborhood N of
θ
2
in S
1
/{θ, θ + π} such that µ
ǫ
(N ) 0 as ǫ 0, i.e.,
lim
ǫ0
lim
n→∞
1
n
n1
X
k=0
(L
k
ǫ
µ)(N ) = 0, (3.1)
for some µ U (S
1
/{θ, θ + π}).
Since A
0
is an invertible matrix, then (a
1
0
cos θ + a
2
0
sin θ)
2
+ (a
3
0
cos θ + a
4
0
sin θ)
2
6= 0
for any θ S
1
. Thus based on
A
0
cos θ
sin θ
=
a
1
0
cos θ + a
2
0
sin θ
a
3
0
cos θ + a
4
0
sin θ
,
A
0
(θ) = arctan
a
3
0
cos θ + a
4
0
sin θ
a
1
0
cos θ + a
2
0
sin θ
or
π
2
, θ S
1
/{θ, θ + π}.
Note that in the space S
1
/{θ, θ + π}, θ and π + θ are identical.
Hence
A
0
(θ) =
a
1
0
a
4
0
a
2
0
a
3
0
(a
1
0
cos θ + a
2
0
sin θ)
2
+ (a
3
0
cos θ + a
4
0
sin θ)
2
.
If θ = θ
2
, then
a
1
0
cos θ
2
+ a
2
0
sin θ
2
= ±e
λ
0
2
cos θ
2
, a
3
0
cos θ
2
+ a
4
0
sin θ
2
= ±e
λ
0
2
sin θ
2
.
Hence
A
0
(θ
2
) =
±e
λ
0
1
e
λ
0
2
e
2λ
0
2
cos
2
θ
2
+ e
2λ
0
2
sin
2
θ
2
= ±e
λ
0
1
λ
0
2
.
7
Thus we can choose a positive constant δ
1
(<
1
4
|θ
2
θ
1
|) such that
|A
0
(θ)| e
λ
0
1
λ
0
2
ǫ
0
> 1, θ [θ
2
δ
1
, θ
2
+ δ
1
], (3.2)
where ǫ
0
is a positive number satisfying ǫ
0
<
λ
0
1
λ
0
2
2
. Take N = (θ
2
δ
1
, θ
2
+ δ
1
). Let
ρ
ǫ
(θ) denote the density of P
ǫ
(θ, ·) for given θ. In the following we first show that supp
ρ
ǫ
(θ
2
± δ
1
) is disjoint from the δ
1
-neighborhood of θ
2
for sufficiently small ǫ. Namely the
case 2 in the proof of Theorem 1 in [13] never occurs under our situation.
If the two eigenvalues of A
0
are both p ositive, then A
0
(θ
2
) = e
λ
0
1
λ
0
2
> 1 . Thus based
on A
0
(θ
2
) = θ
2
, for the above sufficiently small positive constant δ
1
,
A
0
(θ
2
δ
1
) θ
2
< δ
1
, A
0
(θ
2
+ δ
1
) θ
2
> δ
1
.
That is,
θ
:= A
0
(θ
2
δ
1
) < θ
2
δ
1
:= θ
, θ
+
:= A
0
(θ
2
+ δ
1
) > θ
2
+ δ
1
:= θ
+
.
Further, since A(t
) is continuous with respect to t
and A
0
is invertible, then ǫ
1
(0, 1],
such that as ǫ < ǫ
1
,
a
1
ǫ
(t)a
4
ǫ
(t) a
2
ǫ
(t)a
3
ǫ
(t) 6= 0, t [b, b].
Thus A(ǫt)(θ) is well defined. Moreover, for θ = θ
, θ
+
, θ
1
, there exists ǫ
2
(0, ǫ
1
] such
that for any ǫ < ǫ
2
and t [b, b],
|A(ǫt)(θ
) A
0
(θ
)| <
1
4
(θ
θ
),
|A(ǫt)(θ
+
) A
0
(θ
+
)| <
1
4
(θ
+
θ
+
),
|A(ǫt)(θ
1
) A
0
(θ
1
)| < δ
1
.
Note that A
0
(θ
1
) = θ
1
. Therefore, supp ρ
ǫ
(θ
2
± δ
1
) and supp ρ
ǫ
(θ
1
) are disjoint from
the δ
1
-neighborhood of θ
2
as ǫ < ǫ
2
, respectively. If the two eigenvalues of A
0
are both
negative or have different signs, the same result still holds following the similar argument.
Let µ be any Borel probability measure on S
1
/{θ, θ+π} which is absolutely continuous
with respect to Lebesgue measure. Let η
0
denote the density of µ on S
1
/{θ, θ + π}, and
let η
i
denote the density of L
i
ǫ
µ on S
1
/{θ, θ + π}, i = 1, 2, · · ·. Note that the densities η
0
and η
1
have the relation
η
1
(θ) =
Z
[b,b]
η
0

A
ǫ
(t)
1
(θ)
·

A
ǫ
(t)
1
(θ)
(t), (3.3)
8
where
A
ǫ
(t)
1
(θ) is the inverse mapping of A
ǫ
(t)(θ) with respect to θ for given t. In
the case that ǫ < ǫ
2
, since supp ρ
ǫ
(θ
2
± δ
1
) and supp ρ
ǫ
(θ
1
) are disjoint from the δ
1
-
neighborhood of θ
2
and A
ǫ
(t)(θ) is strictly monotone with respect to θ for any g iven t,
then for any θ 6∈ [θ
2
δ
1
, θ
2
+δ
1
], A
ǫ
(t)(θ) 6∈ [θ
2
δ
1
, θ
2
+δ
1
]. That is, for θ [θ
2
δ
1
, θ
2
+δ
1
],
(A
ǫ
(t))
1
(θ) [θ
2
δ
1
, θ
2
+ δ
1
].
Furthermore, for any ǫ < ǫ
2
and t [b, b], based on
A
ǫ
(t)(θ)
θ
6= 0, (A(t
))
1
(θ) is continu-
ously differentiable with respect to t
:= ǫt and θ. Thus ((A(t
))
1
)
(θ) is continuous with
respect to t
and θ. Hence there exists ǫ
3
(0 < ǫ
3
< ǫ
2
), such that as 0 < ǫ < ǫ
3
, for any
t [b, b],
max
θ∈N
|((A(ǫt))
1
)
(θ)| e
ǫ
0
max
θ∈N
|((A
0
)
1
)
(θ)| e
λ
0
1
+λ
0
2
+2ǫ
0
, (3.4)
where the second inequality is based on (3.2).
Let eη
i
:= max{η
i
(θ) : |θ θ
2
| δ
1
}, i = 0, 1, · · · . Without loss of generality, we
assume eη
0
< and then eη
i
is well defined for i 1. Hence based on (3.3) and (3.4),
eη
1
eη
0
· e
λ
0
1
+λ
0
2
+2ǫ
0
,
where 0 < e
λ
0
1
+λ
0
2
+2ǫ
0
< e
λ
0
1
+λ
0
2
+λ
0
1
λ
0
2
= 1 .
Recursively,
eη
k
eη
0
· e
k(λ
0
1
+λ
0
2
+2ǫ
0
)
, k = 1, 2, · · · .
Therefore,
1
n
n1
X
k=0
(L
k
ǫ
µ)N
2δ
1
n
n1
X
k=0
eη
k
2δ
1
n
eη
0
1
1 e
λ
0
1
+λ
0
2
+2ǫ
0
0, as n ,
where the last convergence is uniform with respect to ǫ (0, ǫ
3
). That is to say that we
have the equation (3.1). Hence this lemma holds in the case that λ
0
1
> λ
0
2
.
The proof for the case that λ
0
1
= λ
0
2
is easy, and hence we omit it here.
Lemma 3.2 If all the assumptions of Lemm a 3.1 are satisfied and λ
0
2
λ
0
1
< 0, then for
sufficiently small ǫ, the random norm and the standard norm have the f ollowing relation
kY (ω)k kY (ω)k
ω
BkY (ω)k, ω , (3.5)
where B is a positive con s tant independent of ǫ.
9
Proof. At first, according to the definition of the random norm (2.3),
kY (ω)k kY (ω)k
ω
. (3.6)
On the other hand, since the Lypunov exponents λ
0
2
λ
0
1
< 0, then the sp ectral radius
of the matrix A
0
denoted by ρ(A
0
) satisfies ρ(A
0
) < 1. Hence there is a norm k · k
a
in
R
2
and a corresponding operator norm k · k
a
for matrices such that kA
0
k
a
< 1, see for
example, [11]. Let 0 < α < β be such t hat
αkY k kY k
a
βkY k, Y R
2
.
Moreover, based on the assumptions, t
ǫ
[ǫb, ǫb] such that for any t [b, b]
kA
ǫ
(t)k
a
= k A(ǫt)k
a
= kA(t
)k
a
max
t
[ǫb,ǫb]
kA(t
)k
a
= kA(t
ǫ
)k
a
= e
M
ǫ
,
where M
ǫ
= log kA(t
ǫ
)k
a
. In particular, for ǫ = 0, M
0
= log kA
0
k
a
< 0. Thus, for any
i = 1, 2, · · · ,
kΦ
ǫ
(i, ω)k
a
kA
ǫ
(ε
i
)k
a
· · · k A
ǫ
(ε
2
)k
a
kA
ǫ
(ε
1
)k
a
e
iM
ǫ
, ω .
Hence
kY (ω)k
ω
X
i=0
e
i(λ
ǫ
s
+κ(ǫ))
kΦ
ǫ
(i, ω)Y (ω)k
1
α
X
i=0
e
i(λ
ǫ
s
+κ(ǫ))
kΦ
ǫ
(i, ω)Y (ω)k
a
β
α
X
i=0
e
i(M
ǫ
λ
ǫ
s
κ(ǫ))
kY (ω)k.
Furthermore, since M
ǫ
M
0
and λ
ǫ
s
λ
0
s
as ǫ 0, then ǫ
4
(0, ǫ
3
] such that for any
ǫ < ǫ
4
, we have
|λ
ǫ
s
λ
0
s
| <
1
4
M
0
, |(M
ǫ
λ
ǫ
s
) (M
0
λ
0
s
)| <
1
4
M
0
.
Take κ κ(ǫ) =
1
2
M
0
λ
0
s
(> M
0
λ
0
s
0). Then
λ
ǫ
s
+ κ(ǫ) < λ
0
s
1
4
M
0
+
1
2
M
0
λ
0
s
=
1
4
M
0
< 0,
M
ǫ
λ
ǫ
s
κ(ǫ) < M
0
λ
0
s
1
4
M
0
1
2
M
0
+ λ
0
s
=
1
4
M
0
< 0 .
Thus as ǫ < ǫ
4
,
kY (ω)k
ω
β
α
X
i=0
e
i(M
ǫ
λ
ǫ
s
κ(ǫ))
kY (ω)k
β
α
X
i=0
e
iM
0
4
kY (ω)k = BkY (ω)k, (3.7)
10
where B =
β
α
P
i=0
e
iM
0
4
=
β
α
·
1
1e
M
0
4
.
Therefore, for sufficiently small ǫ, combination of (3.6) with (3.7) gives
kY (ω)k kY (ω)k
ω
BkY (ω)k.
Note that in the following κ =
1
2
M
0
λ
0
s
. Further, we set Λ = 1 e
M
0
4
( (0, 1)).
Lemma 3.3 Under the conditions of Lemma 3. 2, let H : [b, b] × R
2
R
2
be a C
2
function with H
ǫ
(ε
1
, 0) = H(ǫε
1
, 0) = 0 and D
Y
H
ǫ
(ε
1
, 0) = D
Y
H(ǫε
1
, 0) = 0 for any
ǫ [0, 1]. Then for sufficiently small ǫ, there exists a cons tant r (0, 1] independent of ǫ
and ω such that for any kY
i
(ω)k r 1, i = 1, 2
kH
ǫ
(ε
1
, Y
1
(ω)) H
ǫ
(ε
1
, Y
2
(ω))k
σω
LkY
1
(ω) Y
2
(ω)k
ω
, ω ,
where L =
Λ
2
> 0.
Proof. Since H(·, ·) is a C
2
function, for any kY
1
(ω)k 1, kY
2
(ω)k 1, we have
kD
Y
H(ǫε
1
, Y
1
(ω)) D
Y
H(ǫε
1
, Y
2
(ω))k JkY
1
(ω) Y
2
(ω)k < , ǫ [0, 1], ε
1
[b, b],
where J is a positive constant independent of ǫ and ω as well as Y
i
, i = 1, 2 . Take r (0, 1].
Since D
Y
H(ǫε
1
, , 0) = 0, then for any kY (ω)k r,
kD
Y
H(ǫε
1
, Y (ω))k = kD
Y
H(ǫε
1
, Y (ω)) D
Y
H(ǫε
1
, , 0)k JkY (ω)k Jr.
Hence for any kY
i
(ω)k r 1, i = 1, 2,
kH
ǫ
(ε
1
, Y
1
(ω)) H
ǫ
(ε
1
, Y
2
(ω))k sup
kY (ω)k≤r
kD
Y
H(ǫε
1
, Y (ω))kkY
1
(ω) Y
2
(ω)k
JrkY
1
(ω) Y
2
(ω)k. (3.8)
From Lemma 3.2, as ǫ < ǫ
4
, then
kH
ǫ
(ε
1
, Y
1
(ω)) H
ǫ
(ε
1
, Y
2
(ω))k
σω
BkH
ǫ
(ε
1
, Y
1
(ω)) H
ǫ
(ε
1
, Y
2
(ω))k
BJrkY
1
(ω) Y
2
(ω)k
BJrkY
1
(ω) Y
2
(ω)k
ω
.
Set
r = min{
L
BJ
, 1} = min{
Λ
2BJ
, 1}.
Therefore, for any kY
i
(ω)k r, i = 1, 2, as ǫ < ǫ
4
, we have
kH
ǫ
(ε
1
, Y
1
(ω)) H
ǫ
(ε
1
, Y
2
(ω))k
σω
BJrkY
1
(ω) Y
2
(ω)k
ω
LkY
1
(ω) Y
2
(ω)k
ω
.
11
4 Proof of Theorem 2. 1
Proof. We first consider a special case and then provide the proof in a general case.
Step 1: If H 0, then (2.1) is an ane random difference equation. In this case, for
sufficiently small ǫ, the var ia t io n of constant s formula gives the solution with its initial
value
ξ
ǫ
(ω) :=
0
X
i=−∞
Φ
ǫ
(i, σ
i
ω)h
ǫ
(ε
i
). (4.1)
The expression makes sense, for
k|ξ
ǫ
k| = ess sup kξ
ǫ
(ω)k
ω
ess sup
0
X
i=−∞
kΦ
ǫ
(i, σ
i
ω)k
σ
i
ω ,ω
kh
ǫ
(ε
i
)k
σ
i
ω
ess sup
0
X
i=−∞
e
i(λ
ǫ
s
+κ)
kh
ǫ
(ε
i
)k
σ
i
ω
1
1 e
λ
ǫ
s
+κ
k|h
ǫ
k|
1
Λ
k|h
ǫ
k|, (4.2)
where the last inequality holds because that 0 < 1 e
1
4
M
0
= Λ < 1 e
λ
ǫ
s
+κ
as ǫ < ǫ
4
.
Note that all ess sup kh
ǫ
(ε
i
)k
σ
i
ω
, i Z are equal, we denote them by k|h
ǫ
k|.
Based on the assumptions that h(·) C
2
and h(0) = 0, ǫ
5
(0, ǫ
4
] such that as
ǫ < ǫ
5
,
kh
ǫ
(ε
0
)k = kh(ǫε
0
)k
C
0
b
|ǫε
0
| C
0
ǫ,
where C
0
is a positive constant independent of ǫ.
Furthermore,
kh
ǫ
(ε
0
)k
ω
= k h(ǫε
0
)k
ω
Bkh(ǫε
0
)k BC
0
ǫ. (4.3)
Hence
k|ξ
ǫ
k|
k|h
ǫ
k|
Λ
BC
0
ǫ
Λ
< .
On the other hand, according to the cocycle property,
A
ǫ
(ε
1
)ξ
ǫ
(ω) + h
ǫ
(ε
1
) =
1
X
i=−∞
Φ
ǫ
(1 i, σ
i
ω)h
ǫ
(ε
i
) = ξ
ǫ
(σω).
12
Hence t he stochastic process with initial value ξ
ǫ
(ω) is a stationary process. Namely,
{ξ
ǫ
n
(ω)} := {ξ
ǫ
(σ
n
ω)} is a stationary solution to (2.1).
Furthermore, we will show that the stationary solution is unique in the neighborhood
of the orig in. If ζ
ǫ
1
, ζ
ǫ
2
are both stationary solutions of (2.1), then ζ
ǫ
:= ζ
ǫ
1
ζ
ǫ
2
is also one
stationary solution and satisfies ζ
ǫ
n+1
= A
ǫ
(ε
n+1
)ζ
ǫ
n
. However, based on Lemma 3.1 and
Lemma 3.2, the linearized system ζ
ǫ
n+1
= A
ǫ
(ε
n+1
)ζ
ǫ
n
is hyperbolic. Hence ζ
ǫ
0. That
is, the stationary solution is unique.
Step 2: Consider now the full equation (2.1).
Just as [3], we now define a similar space having similar properties. Let
L
:= {Y : R
2
measurable with k|Y k| < ∞}.
Then L
is a Banach space. Thus for r > 0 in Lemma 3.3,
D
r
:= {Y L
: k|Y k| r}.
is also complete. In the following we will prove the existence of a stationary solution
using the Banach fixed point theorem. Hence we first define an operator and then test
the conditions o f the theorem.
(1) We now define a nonlinear operator T
ǫ
: D
r
L
by
T
ǫ
Y (ω) =
0
X
i=−∞
Φ
ǫ
(i, σ
i
ω)
e
h
ǫ
(σ
i
ω), (4.4)
where
e
h
ǫ
(ω) = H
ǫ
(ε
0
, Y (σ
1
ω)) + h
ǫ
(ε
0
).
The definition makes sense since, according to Lemma 3.3 and (4.3),
k
e
h
ǫ
(ω)k
ω
kH
ǫ
(ε
0
, Y (σ
1
ω))k
ω
+ kh
ǫ
(ε
0
)k
ω
Lr + BC
0
ǫ.
That is,
k|
e
h
ǫ
k| = ess sup k
e
h
ǫ
(ω)k
ω
Lr + BC
0
ǫ. (4.5)
Furthermore, the comparison of (4.1) with (4.4) shows that T
ǫ
Y is just the initial value
of the unique stationary solution of the affine equation Y
n
= A
ǫ
(ε
n
)Y
n1
+
e
h
ǫ
(ε
n
). Hence
from ( 4.2),
k|T
ǫ
Y k|
1
1 e
λ
ǫ
s
+κ
k|
e
h
ǫ
k|
1
Λ
k|
e
h
ǫ
k|. (4.6)
13
(2) T
ǫ
maps D
r
to D
r
.
At first, ǫ
6
=
rΛ
2BC
0
> 0, such that as 0 < ǫ < ǫ
6
, Lr + BC
0
ǫ rΛ. Furthermore,
based on (4.5) and (4.6), as 0 < ǫ < ǫ
7
:= min{ǫ
5
, ǫ
6
}, we have
k|T
ǫ
Y k|
Lr + BC
0
ǫ
Λ
rΛ
Λ
= r.
(3) T
ǫ
is contracting on D
r
.
From Lemma 3.3 and the property of the random norm (2.4), fo r any Y
1
, Y
2
D
r
, we
have
kT
ǫ
Y
1
(ω) T
ǫ
Y
2
(ω)k
ω
= k
0
X
i=−∞
Φ
ǫ
(i, σ
i
ω)(H
ǫ
(ε
i
, Y
1
(σ
i1
ω)) H
ǫ
(ε
i
, Y
2
(σ
i1
ω)))k
ω
0
X
i=−∞
kΦ
ǫ
(i, σ
i
ω)k
σ
i
ω ,ω
kH
ǫ
(ε
i
, Y
1
(σ
i1
ω)) H
ǫ
(ε
i
, Y
2
(σ
i1
ω))k
σ
i
ω
0
X
i=−∞
e
i(λ
ǫ
s
+κ)
LkY
1
(σ
i1
ω) Y
2
(σ
i1
ω)k
σ
i1
ω
1
2
k|Y
1
Y
2
k|.
That is,
k|T
ǫ
Y
1
T
ǫ
Y
2
k| = ess sup kT
ǫ
Y
1
(ω) T
ǫ
Y
2
(ω)k
ω
1
2
k|Y
1
Y
2
k|.
Therefore, by the Banach fixed point theorem, (2.1) has a stationary solution whose
initial value ξ
ǫ
(ω) satisfies kξ
ǫ
(ω)k kξ
ǫ
(ω)k
ω
k|ξ
ǫ
k| r for almost all ω (and is unique
with this property).
Furthermore, for the stationa ry solution with initial value ξ
ǫ
(ω), based on Lemma 3.3
and (4.6),
kξ
ǫ
(ω)k
ω
= kT
ǫ
ξ
ǫ
(ω)k
ω
k|T
ǫ
ξ
ǫ
k|
1
Λ
k|H
ǫ
(ε
0
, ξ
ǫ
(σ
1
ω)) + h
ǫ
(ε
0
)k|
1
Λ
(k|H
ǫ
(ε
0
, ξ
ǫ
(σ
1
ω)) H
ǫ
(ε
0
, 0)k| + k|h
ǫ
k|)
1
Λ
(Lk|ξ
ǫ
k| + k |h
ǫ
k|).
That is,
k|ξ
ǫ
k| = ess sup kξ
ǫ
(ω)k
ω
1
Λ
(Lk|ξ
ǫ
k| + k|h
ǫ
k|).
14
Hence for almost all ω,
kξ
ǫ
(ω)k kξ
ǫ
(ω)k
ω
k|ξ
ǫ
k|
k|h
ǫ
k|
Λ L
=
2k|h
ǫ
k|
Λ
.
Moreover based on (4.3 ) , for almost all ω,
kξ
ǫ
(ω)k k|ξ
ǫ
k|
2k|h
ǫ
k|
Λ
2BC
0
ǫ
Λ
. (4.7)
Take K =
2BC
0
Λ
. Then the proof of Theorem 2.1 is complete.
5 Properties
Proposition 5.1 The stationary solution obtained in Theorem 2.1 is measurable.
Proof. We start with ξ
ǫ,0
0 (which is obviously measurable) and recursively define
ξ
ǫ,k+1
(ω) := T
ǫ
ξ
ǫ,k
(ω), k = 0, 1, 2 · · · .
Based on the assumptions of theorem 2.1, A(·) is C
1
, H(·, ·), h(·) are C
2
, and hence they
are all measurable. Thus according to the definition of T
ǫ
in (4.4), ξ
ǫ,1
(ω) := T
ǫ
ξ
ǫ,0
(ω) is
measurable. Hence recursively ξ
ǫ,k
(ω) is measurable. In addition, the proof of (2) of step
2 in section 4 also shows that ξ
ǫ,k
(ω) D
r
.
By the Banach fixed point theorem on D
r
,
lim
k→∞
ξ
ǫ,k
= ξ
ǫ
.
That is, for any eε > 0, K(eε) > 0 such that k K(eε),
k|ξ
ǫ,k
ξ
ǫ
k| < eε.
Furthermore, due to the relation of norms, as k > K(eε),
kξ
ǫ,k
(ω) ξ
ǫ
(ω)k kξ
ǫ,k
(ω) ξ
ǫ
(ω)k
ω
k|ξ
ǫ,k
ξ
ǫ
k| < eε.
Hence for almost all ω,
lim
k→∞
ξ
ǫ,k
(ω) = ξ
ǫ
(ω).
Given that the fixed point ξ
ǫ
(ω) is a pointwise limit of measurable functions, then it is
measurable.
Note that the proof follows Lemma 7.3.4 in [2].
15
Proposition 5.2 As ǫ 0, the s tationa ry solution obtained in Theorem 2.1 converges
almost surely to zero, which is an isola ted fixed poin t in the deterministic case.
Proof. Since according to the assumptions of Theorem 2.1, h
0
= h(0) = 0 and H
0
(ε
n
, 0) =
H(0, 0) = 0, then 0 is a fixed point for the deterministic difference equation, namely the
equation ( 2.1) with ǫ = 0. Based on (4.7), fo r almost all ω,
kξ
ǫ
(ω)k kξ
ǫ
(ω)k
ω
2BC
0
ǫ
Λ
0, as ǫ 0.
That is, the random fixed point converges almost surely to the deterministic fixed point
as the noise intensity tends to zero.
Note that in general study of stochastic stability, only weak convergence can be derived.
Proposition 5.3 The stationary solution obtained in Theorem 2.1 is locally exponentially
stable with probability one.
Proof. If ζ
n
(ω) is any other solution of (2.1) on Z
+
with initial va lue k|ζ
0
k| r, then
ζ
n+1
= A
ǫ
(ε
n+1
)ζ
n
+ H
ǫ
(ε
n+1
, ζ
n
) + h
ǫ
(ε
n+1
). (5.1)
Since kζ
0
(ω)k
ω
r, based on the above equation (5.1) and 1e
λ
ǫ
s
+κ
= 1e
λ
ǫ
s
+
1
2
M
0
λ
0
s
> Λ,
as ǫ < ǫ
7
,
kζ
1
k
σω
kA
ǫ
(ε
1
)k
ω ,σ
1
ω
kζ
0
k
ω
+ kH
ǫ
(ε
1
, ζ
0
)k
σω
+ kh
ǫ
(ε
1
)k
σω
re
λ
ǫ
s
+κ
+ Lr + BC
0
ǫ r.
We recursively obtain kζ
n
k
σ
n
ω
r. Hence
k|ζ
n
k| r.
On the other hand, ξ
ǫ
n
is the stationary solution of model (2.1) with initial value ξ
ǫ
0
(ω) =
ξ
ǫ
(ω). That is,
ξ
ǫ
(σ
n+1
ω) := ξ
ǫ
n+1
= A
ǫ
(ε
n+1
)ξ
ǫ
n
+ H
ǫ
(ε
n+1
, ξ
ǫ
n
) + h
ǫ
(ε
n+1
). (5.2)
Subtracting (5.2) from (5.1) yields the equation
ζ
n+1
ξ
ǫ
n+1
= A
ǫ
(ε
n+1
)(ζ
n
ξ
ǫ
n
) + H
ǫ
(ε
n+1
, ζ
n
) H
ǫ
(ε
n+1
, ξ
ǫ
n
).
Set ζ
n+1
:= ζ
n+1
ξ
ǫ
n+1
, n = 0, 1, 2, · · · . Then ζ
n+1
satisfies the following equation:
ζ
n+1
= A
ǫ
(ε
n+1
)ζ
n
+ H
ǫ
(ε
n+1
, ζ
n
+ ξ
ǫ
n
) H
ǫ
(ε
n+1
, ξ
ǫ
n
). (5.3)
16
Since kζ
n
+ ξ
ǫ
n
k
σ
n
ω
= kζ
n
k
σ
n
ω
k|ζ
n
k| r and kξ
ǫ
n
k
σ
n
ω
k|ξ
ǫ
k| r, then kH
ǫ
(ε
n+1
, ζ
n
+
ξ
ǫ
n
) H
ǫ
(ε
n+1
, ξ
ǫ
n
)k
σ
n+1
ω
Lkζ
n
k
σ
n
ω
, based on Lemma 3.3.
Thus
kζ
n+1
k
σ
n+1
ω
(e
λ
ǫ
s
+κ
+ L)kζ
n
k
σ
n
ω
(e
λ
ǫ
s
+κ
+ L)
n+1
kζ
0
k
ω
= γ
n+1
kζ
0
k
ω
,
where γ = e
λ
ǫ
s
+κ
+ L 1 Λ +
Λ
2
= 1
Λ
2
(0, 1).
For any eε > 0, without loss of generality, we assume eε < r 1. Take
e
δ =
eε
B
. Since B > 1,
e
δ < 1. As kζ
0
ξ
ǫ
0
k <
e
δ < 1, for almost all ω and any n = 0, 1, 2, · · ·
kζ
n
ξ
ǫ
n
k kζ
n
ξ
ǫ
n
k
σ
n
ω
γ
n
kζ
0
ξ
ǫ
0
k
ω
< B
e
δ = eε.
Hence the stationary solution with the initial value ξ
ǫ
(ω) is locally stable. Furthermore,
for almost all ω
kζ
n
ξ
ǫ
n
k kζ
n
ξ
ǫ
n
k
σ
n
ω
< γ
n
kζ
0
ξ
ǫ
0
k
ω
< γ
n
eε < γ
n
, for any n N.
Therefore, the stationary solution with the initial value ξ
ǫ
(ω) is locally exponentially
stable with pro bability one.
6 Conclus i on and future direc tions
In this paper we provide some new easy-to-use conditions that guara ntee the existence
and stability of nonlinear random difference equations with a view of application. First,
we obtain the stability of Lyapunov exponents for a special case under weaker conditions,
namely Lemma 3 .1 . Second, we derive a new r elationship between the random norm a nd
the standard norm, i.e., the ratio of the random norm and the standard norm has deter-
ministic bounds, namely Lemma 3.2. However, in general theory of random dynamical
systems, the bound of the ratio of the random norm a nd the standard norm is random
depending on sample paths (see [2]). This result plays a crucial role in obtaining a de-
terministic neighborhood in which the Lipschitz constant in the case of random norm
is deterministic and sufficiently small, namely Lemma 3 .3 . But in the general theory o f
random dynamical systems, the neighborhoo d is random depending on sample paths (see
[2, 6]). Eventually, by integrat ing the lemmas we provide a theorem that guar antees the
existence of stationary solutions under easy-to-use conditions, as stated in Theorem 2.1.
In addition, we also obtain several prop erties including measurability, local exponent sta-
bility with probability one and the relationship between the stationary solution and the
17
corresponding deterministic fixed point. Under new conditions, the general result of local
asymptotical stability in probability given by [2] is improved. Moreover, the result that
the stationary solution converges to the deterministic fixed point with probability o ne as
the noise intensity tends to zero also implies that the stationary measure converges to the
Dirac measure suppo r t ed on the deterministic fixed point with the noise intensity tending
to zero.
That is, for nonlinear random difference equations, we not only establish the easy-to-
use conditions to guarantee the existence of stationary solutions but also study stability
and relation o f the stationary solution and its corresponding deterministic fixed point.
These properties are imp ortant for practical models. For example, in [9] the original
random model is studied by using an approximate deterministic model which is given
by setting the random dividend in the random model as a constant, its mean. With
the theorem and propositions in this paper, it can be proved that there exists a unique
stable stationary solution in a small neighborhood of the fixed point of the corresponding
deterministic difference equation provided the perturbation intensity is sufficiently small
(see [15 ]). Thus from the dynamic point of view, the dynamic prop erties are very close for
the random model and the corresponding deterministic model near the fixed point since
random dividends in real markets are generally small. That is, the method they used in [9]
is meaningful at least for the dynamics of the model near the fixed point. Hence in some
sense this paper builds up a bridge between some theoretical researches of mathematics
and their applications to practical models.
In future research, firstly, it is possible to provide new easy-to-use conditions that
guarantee the existence and stability stationary solutions of high dimensional nonlinear
random difference equations with a view of application. Secondly, it is significant to
generalize the i.i.d. noise to more general cases. Finally, it is also interesting to study
various kinds of bifurcations based on the obtained results fr om a dynamical point of
view.
Acknowledgement s
Financial support f r om the NSF of China (No. 10571003, 10871005) is acknowledged.
18
References
[1] V. Ara´ujo, Attrac tors and time avera g es for ra ndom maps, Ann. Inst. H. Poincar´e
Anal. Non lin´eaire 17 (2000), pp. 307–369.
[2] L. Arnold, Random Dynamical Systems, Springer-verlag, Berlin, 1998.
[3] L. Arnold and P. Boxler, Additive noi s e turns a hyperbolic fixed point in to a station-
ary solution, in Lyapunov Exponents, L. Arnold, H. Crauel and J.P. Eckmann, eds.,
Springer-verlag, Berlin, 1991, pp. 159–164.
[4] V. ohm and J. Wenzelburger, On the perfo rmance of efficient portfolios, J. Econ.
Dynam. Contro l 29 (2005), pp. 721–740.
[5] W. Brock and C. Hommes, Heterogeneous beliefs and routes to chaos in a simple asset
pricing model, J. Econ. Dynam. Control 22 (1998), pp. 1235–1274.
[6] S. Dahlke, Invariant manifolds for products of random diffeomorphisms, J. Dynam.
Differential Equations 9 (1997) , pp. 157–210.
[7] C.A. Futia, Invariant distributions and the limiting beha v i or of markovian econom i c
models, Econometrica 50 (1982), pp. 377–408.
[8] M. Hillebrand and J. Wenzelburger, On the dynamics of a s set p rices and po rtfoli o s in
a m ultiperiod CAPM, Chaos Solitons Fractals 29 (2006), pp. 578–594 .
[9] C. Hommes, H. Huang, D. Wang, A robust rational route to randomness in a simple
asset pricing model, J. Econ. Dynam. Control 29 (2005), pp. 1043–1072.
[10] Y. Kifer, Random Perturbations of Dynamical Systems, Birkh¨auser, Boston Basel,
1988.
[11] M. Martelli, Introduction to Discrete Dynamical Systems and Chaos, Hohn /wiley &
Sons,Inc., 1999.
[12] H. Tong, Non-Linear Time Se rie s , Oxford University Press, New Yo rk, 1990.
[13] L.S. Young, Random perturbations of matrix cocycles, Ergodic Theory Dynam. Sys-
tems 6 (1986), pp. 311–31 9.
19
[14] H. Zmarrou and A.J. Homburg, Bifurcations of stationary measures of random dif-
feomorph i s ms, Ergodic Theory Dynam. Systems 27 (2007), pp. 1651–1692.
[15] M. Zhu, D. Wang and M. Guo, Stochastic equilib ria of an asset pricing model with
heterog eneous beliefs and random dividends, Preprint: Research Report No.64 (2008),
School of Mathematical Sciences, Peking University.
20
... See for example, Arnold and Boxler (1991) and Arnold (1998). Therefore we first developed a new mathematical tool in Zhu et al. (2008), in which we gave weaker conditions to guarantee the stability of Lyapunov exponents under small random perturbations by using some ideas of Young (1986). In fact, Young (1986) studied a more general case. ...
... In order to deal with the special cases related to the HHW model, we provided some conditions that are weaker than those given by Young (1986). On the other hand, Zhu et al. (2008) also provided the conditions to guarantee that the ratio of the random norm and the standard Euclidean norm has deterministic bounds, which plays a crucial role in obtaining the existence, stability and convergence of stationary solutions of nonlinear random difference equations in the sense of almost sure. Based on the idea of Arnold and Boxler (1991), Zhu et al. (2008) proved the existence of stationary solutions of nonlinear random difference equations under easy-touse conditions. ...
... On the other hand, Zhu et al. (2008) also provided the conditions to guarantee that the ratio of the random norm and the standard Euclidean norm has deterministic bounds, which plays a crucial role in obtaining the existence, stability and convergence of stationary solutions of nonlinear random difference equations in the sense of almost sure. Based on the idea of Arnold and Boxler (1991), Zhu et al. (2008) proved the existence of stationary solutions of nonlinear random difference equations under easy-touse conditions. In addition, Zhu et al. (2008) further studied the stability of the stationary solutions and the relationship between the stationary solutions and the corresponding fixed points of the deterministic system in the sense of almost sure. ...
Article
We investigate dynamical properties of a heterogeneous agent model with random dividends and further study the relationship between dynamical properties of the random model and those of the corresponding deterministic skeleton, which is obtained by setting the random dividends as their constant mean value. Based on our recent mathematical results, we prove the existence and stability of random fixed points as the perturbation intensity of random dividends is sufficiently small. Furthermore, we prove that the random fixed points converge almost surely to the corresponding fixed points of the deterministic skeleton as the perturbation intensity tends to zero. Moreover, simulations suggest similar behaviors in the case of more complicated attractors. Therefore, the corresponding deterministic skeleton is a good approximation of the random model with sufficiently small random perturbations of dividends. Given that dividends in real markets are generally very low, it is reasonable and significant to some extent to study the effects of heterogeneous agents' behaviors on price fluctuations by the corresponding deterministic skeleton of the random model.
... In Zhu et al. (2011b), the authors investigate the relationship between the deterministic skeleton of a heterogeneous agent model and its stochastic counterpart, namely when noise is added to the dividend process. To do so they rely on results obtained in Zhu et al. (2011a). The textbook Freidlin and Wentzell (2012) details an analysis of the effect of random noise added to dynamical systems in continuous time. ...
Article
We extend the agent-based model of Anufriev and Bottazzi (2010) to the case with many risky assets. We show that under the procedural equilibrium, all assets with nonzero aggregate demand must have the same price returns. Heterogeneity in price returns appears when some assets face zero demand. In this case, the returns are essentially driven by the matrix of wealth-weighted products between demands and they generate a rich variety of patterns for agents’ market shares. We formalize the dynamics of deterministic skeletons in our market model and consider the associated Jacobian matrix, for which we provide a closed form expression up to a matrix inversion. We then introduce the characteristic ratio of an investment strategy, which involves the average dividend yields of the risky assets and the investors’ portfolio compositions. When equilibrium returns are uniform and when the number of assets is not smaller than the number of agents, the equilibrium with a single survivor can only be stable if the survivor has the maximal characteristic ratio. Moreover, we prove that this stability criterion holds as long as the noise in the system is sufficiently small. We confirm all our findings with a thorough empirical analysis of numerical simulations, with and without noise. All in all, our results form a theoretical rationale for portfolio strategies tilted towards firms that pay high dividend yields.
Article
Full-text available
Book
Mathematicians often face the question to which extent mathematical models describe processes of the real world. These models are derived from experimental data, hence they describe real phenomena only approximately. Thus a mathematical approach must begin with choosing properties which are not very sensitive to small changes in the model, and so may be viewed as properties of the real process. In particular, this concerns real processes which can be described by means of ordinary differential equations. By this reason different notions of stability played an important role in the qualitative theory of ordinary differential equations commonly known nowdays as the theory of dynamical systems. Since physical processes are usually affected by an enormous number of small external fluctuations whose resulting action would be natural to consider as random, the stability of dynamical systems with respect to random perturbations comes into the picture. There are differences between the study of stability properties of single trajectories, i. e., the Lyapunov stability, and the global stability of dynamical systems. The stochastic Lyapunov stability was dealt with in Hasminskii [Has]. In this book we are concerned mainly with questions of global stability in the presence of noise which can be described as recovering parameters of dynamical systems from the study of their random perturbations. The parameters which is possible to obtain in this way can be considered as stable under random perturbations, and so having physical sense. -1- Our set up is the following.
Article
Lyapunov exponents for two-dimensional matrix cocycles are shown to be stable under certain stochastic perturbations.
Article
We investigate dynamical properties of a heterogeneous agent model with random dividends and further study the relationship between dynamical properties of the random model and those of the corresponding deterministic skeleton, which is obtained by setting the random dividends as their constant mean value. Based on our recent mathematical results, we prove the existence and stability of random fixed points as the perturbation intensity of random dividends is sufficiently small. Furthermore, we prove that the random fixed points converge almost surely to the corresponding fixed points of the deterministic skeleton as the perturbation intensity tends to zero. Moreover, simulations suggest similar behaviors in the case of more complicated attractors. Therefore, the corresponding deterministic skeleton is a good approximation of the random model with sufficiently small random perturbations of dividends. Given that dividends in real markets are generally very low, it is reasonable and significant to some extent to study the effects of heterogeneous agents' behaviors on price fluctuations by the corresponding deterministic skeleton of the random model.
Chapter
The concept of random dynamical system is a comparatively recent development combining ideas and methods from the well developed areas of probability theory and dynamical systems. Due to our inaccurate knowledge of the particular physical system or due to computational or theoretical limitations (lack of sufficient computational power, inefficient algorithms or insufficiently developed mathematical or physical theory, for example), the mathematical models never correspond exactly to the phenomenon they are meant to model. Moreover when considering practical systems we cannot avoid either external noise or measurement or inaccuracy errors, so every realistic mathematical model should allow for small errors along orbits not to disturb too much the long term behavior. To be able to cope with unavoidable uncertainty about the ``correct'' parameter values, observed initial states and even the specific mathematical formulation involved, we let randomness be embedded within the model to begin with. We present the most basic classes of models in what follows, then define the general concept and present some developments and examples of applications.
Chapter
Suppose (*)Ȧ=A(v t w)x is hyperbolic, i.e. all of its Lyapunov exponents are different from zero. Then Ȧ=A(v t w)x+f(v t w,x)+b(v t w) with f(w,·) locally Lipschitz and f(w,0)=0 has a (unique) stationary solution in a neighborhood of x=0 provided f and b are ‘small’. ‘Smallness’ in being described in terms of a random norm measuring the non-uniformity of the hyperbolicity of (*).
Article
We investigate asset pricing dynamics in an adaptive evolutionary asset pricing model with fundamentalists, trend followers and a market maker. Agents can choose between a fundamentalist strategy at positive information cost or choose a trend following strategy for free. Price adjustment is proportional to the excess demand in the asset market. Agents asynchronously update their strategy according to realized net profits in the recent past. As agents become more sensitive to differences in strategy performance, the fundamental steady state becomes unstable and multiple steady states may arise. As the traders’ sensitivity to differences in fitness increases, a bifurcation route to chaos sets in due to homoclinic bifurcations of stable and unstable manifolds of the fundamental steady state.
Article
This paper investigates the dynamics in a simple present discounted value asset pricing model with heterogeneous beliefs. Agents choose from a finite set of predictors of future prices of a risky asset and revise their ‘beliefs’ in each period in a boundedly rational way, according to a ‘fitness measure’ such as past realized profits. Price fluctuations are thus driven by an evolutionary dynamics between different expectation schemes (‘rational animal spirits’). Using a mixture of local bifurcation theory and numerical methods, we investigate possible bifurcation routes to complicated asset price dynamics. In particular, we present numerical evidence of strange, chaotic attractors when the intensity of choice to switch prediction strategies is high.
Article
We present a numerical case study of the dynamics of a financial market in which heterogeneous investors described by linear mean–variance preferences and multiperiod planning horizons interact. The focus is on the induced price, portfolio, and wealth processes as well as on the transaction volume. Numerical evidence is provided that multiperiod planning horizons are a natural source of empirically observed clustered volatility and that heterogeneous planning horizons may amplify booms and busts.