Академический Документы
Профессиональный Документы
Культура Документы
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 227
Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 233
Introduction
The state of many physical, chemical and biological systems can be described
by a single time-dependent variable x(t) which satisfies the ordinary differ-
ential equation
ẋ(t) = f (λ, x(t)) . (1)
This equation depends on parameters λ = (λ1 , . . . λm ) which characterize the
properties of the environment and are usually called external parameters. For
example, the equation
ẋ = αx − x3 (2)
can be used to describe the growth of a biological population. It contains the
parameter α ∈ R which takes into account the properties of the environment.
If there is an existence and uniqueness theorem for (1), then we can de-
fine an evolution operator St in R by the formula St x0 = x(t; x0 ), where
x(t; x0 ) is the solution to (1) with x(0; x0 ) = x0 . The uniqueness theorem for
(1) and the fact that R is a totally ordered set imply that one-dimensional
equations generate monotone (or order-preserving) dynamical systems, i.e.
St x1 ≥ St x2 provided x1 ≥ x2 . This property drastically simplifies the dy-
namics. For example, for equation (2) we have either one or three equilibrium
points depending on the parameter α and every solution is attracted by an
equilibrium in a monotone way. Indeed, it is easy to see that any solution to
(2) with initial data x0 has the form
x0 eαt
x(t) = 1/2
.
(1 + x20 · α−1 · (e2αt − 1))
Therefore we have that (i) if α < 0, then x(t) → √ 0 as t → +∞ for any initial
data x0 ; (ii) if α > 0 and x0 > 0,√then x(t) → α as t → +∞ and (iii) if
α > 0 and x0 < 0, then x(t) → − α as t → +∞. Thus for the case α < 0
we have a unique globally asymptotically stable equilibrium and in the case
α > 0 we have two stable equilibria√ and√ one unstable. In the latter case the
global attractor is the interval [− α, α] (by definition, the global attractor
is a strictly invariant set which uniformly attracts every bounded set). Thus
equilibria and their stability properties completely determine the long-time
dynamics of the system.
produces solutions which are monotone with respect to initial data if and only
if a12 ≥ 0 and a21 ≥ 0. Nevertheless monotone multi-dimensional ordinary
differential equations cover important classes of mathematical models arising
in modern natural science (see discussion in Smith [102]). The mathematical
theory of deterministic monotone (order-preserving) systems is presently well-
developed due to the efforts of many authors (see, e.g., Krasnoselskii [69],
Hirsch [52, 53, 54] and also Smith [102] and the references therein). A well-
posed autonomous system of ordinary differential equations
ẋi = fi (x1 , . . . , xd ), i = 1, . . . , d ,
fi (x1 , . . . , xd ) ≤ fi (y1 , . . . , yd )
for all (x1 , . . . , xd ) and (y1 , . . . , yd ) from Rd such that xi = yi and xj ≤ yj for
j = i, where i = 1, . . . , d. For example, this relation holds for the following
system of differential equations
Chap.4]) shows that any (complicated) dynamics can occur on unstable in-
variant sets for monotone systems of sufficiently large dimension.
If the system is coupled to a fluctuating environment, then external
parameters can become stochastic quantities. In many cases these quanti-
ties can be presented as stationary random processes. We refer to Hors-
themke/Lefever [55, Chap.1] for a detailed discussion on the nature and
sources of randomness in dynamical systems.
Thus taking into account random fluctuations of the environment for the
system described by (1) leads to the equation
where λ0 corresponds to the mean state of the environment and the station-
ary process ξ(t, ω) with zero expectation on some probability space (Ω, F, P)
describes environmental fluctuations around this mean state. For example,
equation (2) turns into
As above we can show that the process x(t; ω, x0 ) which solves (3) with initial
data x0 has the form
y = (f1 ◦ f0 ◦ f0 ◦ f1 ◦ f1 ◦ f0 ◦ f1 )(x).
where Ck is one of the sets ∅, {0}, {1}, {0, 1} and P0 (∅) = 0, P0 ({0}) =
P0 ({1}) = 1/2, P0 ({0, 1}) = 1. Here {i1 , . . . , im } is an arbitrary m-tuple of
integers. For every n ∈ Z we denote by θn the left shift operator in Ω, i.e.
πn = π1 ◦ πn−1 , n ∈ N, π0 = id ,
where π1 (ω, x) = (θ1 ω, fω0 (x)). This mapping πn can be written in the form
for all n, m ∈ Z+ and ω ∈ Ω. The pair (θn , ϕ(n, ω)) is called a random
dynamical system with discrete time. The mapping θn models the evolution of
some random environment and ϕ(ω, n) describes the dynamics of the system.
If X = R and f0 and f1 are nondecreasing functions, then the mappings
ϕ(n, ω) are order preserving, i.e. the relation x1 ≥ x2 implies that ϕ(n, ω)x1 ≥
ϕ(n, ω)x2 for all n ∈ Z+ and ω ∈ Ω.
It is easy to see that ϕ satisfies the cocycle property if and only if πn given
by (4) is a semigroup, i.e. πn ◦ πm = πn+m for n, m ∈ Z+ . Thus we obtain a
dynamical system in the classical sense (i.e. a semiflow of mappings from some
space into itself). We note that semiflows of a similar structure (see (4)) arise
in the theory of nonautonomous (deterministic) differential equations and
they are known as skew-product flows (see, e.g., Chicone/Latushkin [19]
and the references therein). This observation is important in the study of the
long-time behaviour of random dynamical systems.
The aim of this book is to present a recently developed approach which is
suitable for investigating a variety of qualitative aspects of order-preserving
random dynamical systems and to give the backgrounds for further develop-
ment of the theory. We try to demonstrate the effectiveness of this approach
6 Introduction
asymptotically stable in each part of the cone. Our main result here is a
random limit set trichotomy, stating that in a given part either (i) all orbits
are unbounded, (ii) all orbits are bounded but their closure reaches out to
the boundary of the part, or (iii) there exists a unique, globally attracting
equilibrium. Several examples, including Markov chains and affine systems,
are given.
In Chapters 5 and 6 we apply the results of Chapters 3 and 4 to study the
qualitative behaviour of random and stochastic perturbations of cooperative
ordinary differential equations. These applications are the main motivations
for the development of the general theory presented in Chapters 3 and 4
and we believe that random and stochastic cooperative differential equations
merit a detailed study of its own.
In Chapter 5 we consider random cooperative differential equations in
Rd+ (real noise case). We first give conditions under which these equations
generate order-preserving random dynamical systems in Rd+ and then study
monotonicity properties of these systems. We prove several theorems on the
existence of equilibria and random attractors. Systems with concavity prop-
erties are also considered. We apply general results from Chapters 3 and 4 to
study the long-time behaviour of these systems and to obtain the limit set
trichotomy theorem for random cooperative differential equations. We con-
clude Chapter 5 with a series of examples including a class of one-dimensional
explicitly solvable equations to show possible scenarios of the long-time be-
haviour in monotone systems.
Chapter 6 is devoted to stochastic cooperative differential equations
(white noise case). The hypotheses that guarantee order-preserving prop-
erties for this case lead to a special structure of the diffusion terms. In fact
we consider here some class of stochastic perturbations of deterministic au-
tonomous cooperative differential equations. We prove several assertions on
the long-time behaviour, investigate properties of systems that possess con-
cavity properties and establish a stochastic version of the limit set trichotomy
theorem. We study the long-time dynamics in one-dimensional equations with
details. We also discuss the stochastic versions of certain examples consid-
ered in the previous chapter. Although the results for the stochastic case are
similar to the random case, Chapter 6 is not at all a duplication of Chapter 5
because the methods of proof are quite different.
1. General Facts about Random Dynamical
Systems
In this chapter we recall some basic definitions and facts about random dy-
namical systems. For a more detailed discussion of the theory and applica-
tions of random dynamical systems we refer to the monograph Arnold [3].
We pay particular attention to dissipative systems and their random (pull
back) attractors. These attractors were studied by many authors (see, e.g.,
Arnold [3], Crauel/Debussche/Flandoli [35], Crauel/Flandoli [36],
Schenk-Hoppé [89], Schmalfuss [92, 93] and the references therein). The
ideas that lead to the concept of a random attractor have their roots
in the theory of deterministic dissipative systems which has been suc-
cessfully developed in the last few decades ( see, e.g., the monographs
Babin/Vishik [13], Chueshov [20], Hale [50], Temam [104] and the liter-
ature quoted therein). The proof of the existence of random attractors given
below follows almost step-by-step the corresponding deterministic argument
(see, e.g., Chueshov [20], Temam [104]).
Throughout this book we will be concerned with a probability space by
which we mean a triple (Ω, F, P), where Ω is a space, F is a σ-algebra of
sets in Ω, and P is a nonnegative σ-additive measure on F with P(Ω) = 1.
We do not assume in general that the σ-algebra is complete. Below we will
also use the symbol T for either R or Z and we will denote by T+ all non-
negative elements of T. We will denote by B(X) the Borel σ-algebra of sets
in a topological space X. By definition B(X) is the σ-algebra generated by
the collection of open subsets of X. If (X1 , F1 ) and (X2 , F2 ) are measurable
spaces, we denote by F1 × F2 the product σ-algebra of subsets in X1 × X2
which is defined as the σ-algebra generated by the cylinder sets A = A1 × A2 ,
Ai ∈ Fi . We refer to Cohn [30] for basic definitions and facts from the
measure theory.
2. (t, ω) → θt ω is measurable;
3. θt P = P for all t ∈ T, i.e. P(θt B) = P(B) for all B ∈ F and all t ∈ T.
A set B ∈ F is called θ-invariant if θt B = B for all t ∈ T. A metric dynamical
system θ is said to be ergodic under P if for any θ-invariant set B ∈ F we
have either P(B) = 0 or P(B) = 1.
Example 1.1.1 (Periodic Case). Consider the probability space (Ω, F, P),
where Ω is a circle of unit circumference, F is its σ-algebra of Borel sets and P
is the Lebesgue measure on Ω. Let {θt , t ∈ R} be the group of rotations of the
circle. It is easy to see that we obtain an ergodic MDS (Ω, F, P, {θt , t ∈ R})
with continuous time.
Example 1.1.3 (Almost Periodic Case). Let f (x) be a Bohr almost periodic
function on R. We define the hull H(f ) of the function f as the closure of
the set {f (x + t), t ∈ R} in the norm
f
= supx∈R |f (x)|. The hull H(f ) is
a compact metric space, and it has a natural commutative group structure.
Therefore it possesses a Haar measure which, if normalized to unity, makes
H(f ) into probability space. If we define transformations {θt , t ∈ T} as shifts:
(θt g)(x) = g(x + t), g ∈ H(f ), we obtain an ergodic MDS with continuous
time. For details we refer to Ellis [42] and Levitan/Zhikov [77].
and possesses the property Rn ρ(x) dx = 1. Then ρ(x) is a density of a
probability measure on Rn . By Liouville’s theorem
f (θt x)ρ(x) dx = f (x)ρ(x) dx
Rn Rn
for any bounded continuous function f (x) on Rn and therefore in this situ-
ation an MDS arises with Ω = Rn , F = B(Rn ) and P(dx) = ρ(x)dx. Here
B(Rn ) is the Borel σ-algebra of sets in Rn . Sometimes it is also possible
to construct an MDS connected with the system (1.1), when the solution
ρ to (1.2) is not integrable but the problem (1.1) possesses a first integral
(e.g., if (1.1) is a Hamiltonian system) with appropriate properties (see, e.g.,
Mañé [79] or Sinai [100] for details).
Remark 1.2.1. (i) While the metric dynamical system (modeling the random
perturbations) is assumed to have two-sided time T = R or Z, the cocycle is
only required to have one-sided time T+ = R+ or Z+ . This reflects the fact
that evolution operators are often non-invertible. However this set-up allows
us to consider ϕ(t, θs ω) for t ∈ T+ , but starting at an arbitrary (possibly
negative) time s ∈ T which will be crucial for the construction of equilibria
and attractors. In the case of continuous time (T = R) the standard definition
of a continuous RDS requires the continuity of the mappings (t, x) → ϕ(t, ω)x
for all ω ∈ Ω (see Arnold [3, Sect.1.1]). This property is usually true for
the RDS generated by finite-dimensional random and stochastic equations.
However, as we will see, many general results on the long-time behaviour
can be proved under a weaker assumption of the continuity of the mapping
x → ϕ(t, ω)x for each t ≥ 0 and ω ∈ Ω. We also note that the cocycle
property reduces to the classical semiflow property if ϕ is independent of ω.
Hence deterministic dynamical systems are particular cases of RDS.
(ii) If in Definition 1.2.1 the cocycle is defined on a θ-invariant set Ω ∗ of
full measure, then we can extend it to the whole Ω by the formula
ϕ(t, ω) if ω ∈ Ω ∗ ,
ϕ̃(t, ω) := (1.3)
id / Ω∗ .
if ω ∈
14 1. General Facts about Random Dynamical Systems
Thus we obtain the cocycle ϕ̃(t, ω) which is indistinguishable from ϕ(t, ω).
We recall that by definition the indistinguishability of ϕ(t, ω) and ϕ̃(t, ω)
means that there exists a set N ∈ F such that P(N ) = 0 and
In our case the cocycles coincide on the θ-invariant set Ω ∗ and we can set
N = Ω \ Ω ∗ . In further considerations we do not distinguish cocycles which
coincide on θ-invariant sets of full measure.
(iii) In the definition of an RDS we require some properties to be valid for
all ω ∈ Ω. However the stochastic analysis deals usually with almost all ele-
mentary events ω. Solutions to stochastic differential equations are defined al-
most surely, for example. Therefore to construct RDS connected with stochas-
tic equations we need extend the corresponding evolution operator to all
ω ∈ Ω and prove the cocycle property for this extension. This can be done for
many cases which are important from the point of view of applications. This
procedure is usually referred to as perfection. Roughly speaking the perfection
of cocycles (or other objects) can be done in the following way. First we prove
a property for some θ-invariant set Ω ∗ of full measure. After that we define the
cocycle on Ω \ Ω ∗ in an appropriate way (cf. (1.3)). Perfection theorems have
been shown in various different cases, see, e.g., Arnold/Scheutzow [10],
Scheutzow [90], Kager/Scheutzow [61], Sharpe [98] and also the dis-
cussion in Arnold [3].
Definition 1.2.3 (Affine RDS). Let X be a linear Polish space. The RDS
(θ, ϕ) is said to be affine if the cocycle ϕ is of the form
If (θ, Φ) is a linear RDS, then the cocycle property for the mapping ϕ defined
by (1.4) is equivalent to the relation
We also suppose ϕ(0, ω) = id. It is easy to see that the sequence ϕ(n, ω)x
solves the difference equation
and the mappings ϕ(n, ω) possess the cocycle property. Thus we obtain a
discrete RDS. It is a C k -RDS, if X ⊂ Rd and f (α, ·) ∈ C k (X, X). If X is a
linear Polish space and f (α, ·) are affine mappings, i.e. f (α, x) = Kα x + hα ,
where Kα are continuous linear operators in X and hα are elements from X,
then the RDS constructed above is affine. It is a linear RDS when hα = 0 for
α ∈ Ω0 .
Since all random mappings fθn ω , n ∈ Z, are independent and identically
distributed (i.i.d.), the RDS constructed above generates (see Arnold [3,
p.53]) the homogeneous Markov chain
P (x, B) := P{Φn+1 ∈ B | Φn = x}
If X is a Banach space and g(x, ξ) = g(x) + ξ, then this equation has the
form
xn+1 = g(xn ) + ξn+1 (ω), n ∈ Z+ , x0 = x . (1.8)
A kick force model corresponds to the case when the mapping g : X → X has
the form g(x) = y(T ; x), where T > 0 is a fixed number and y(t) := y(t; x)
solves the equation
Here h is a mapping from X into itself such that equation (1.9) generates a
(deterministic) continuous dynamical system. In this case
where δ(t) is a Dirac δ-function of time. Thus the kick model describes the
situation when the deterministic system (1.9) gets random kicks with some
period T and evolves freely between kicks. We note that kick models are
sufficiently popular in the study of turbulence phenomena.
1.2 Concept of RDS 17
The next examples present the simplest versions of RDS considered in Chaps.
2, 5 and 6 with details.
Example 1.2.3 (1D Random Equation). Let θ = (Ω, F, P, {θt , t ∈ R}) be a
metric dynamical system. Consider the pathwise ordinary differential equa-
tion
ẋ(t) = f (θt ω, x(t)) . (1.10)
Under some natural conditions (see Sect. 2.1 below) on the function f :
Ω × R → R this equation generates an RDS with state space R and with the
cocycle given by the formula ϕ(t, ω)x = x(t), where x(t) is the solution to
(1.10) with x(0) = x. This RDS is affine if f (ω, x) = a(ω) · x + b(ω) for some
random variables a(ω) and b(ω).
More detailed presentation of the last three examples and their generaliza-
tions can be found in Chaps.5 and 6. We also refer to Sects. 2.1 and 2.4 in
18 1. General Facts about Random Dynamical Systems
One of the goals in this book is to describe the long-time behaviour of RDS
and the limit regimes of these systems. These limit regimes typically depend
on an event ω and therefore to characterize their attractivity properties we
should at least be able to calculate the distance between (random) trajecto-
ries and (random) limit objects and treat this distance as a random variable.
It is also crucial to decide whether the limit regimes contain a random vari-
able representing the different states of the system. These circumstances lead
to a notion of a random set which is stronger than simply a collection of sets
depending on ω. We introduce this notion of a random set following to Cas-
taing/Valadier [18] and Hu/Papageorgiou [56] (see also Crauel [32]
and Arnold [3]).
Below any mapping from Ω into the collection of all subsets of X is said
to be a multifunction (or a set valued mapping) from Ω into X.
Definition 1.3.1 (Random Set). Let X be a metric space with a metric
. The multifunction ω → D(ω) = ∅ is said to be a random set if the mapping
ω → distX (x, D(ω)) is measurable for any x ∈ X, where distX (x, B) is the
distance in X between the element x and the set B ⊂ X. If D(ω) is closed for
each ω ∈ Ω then D is called a random closed set. If D(ω) are compact sets
1.3 Random Sets 19
for all ω ∈ Ω then D is called a random compact set. A random set {D(ω)}
is said to be bounded if there exist x0 ∈ X and a random variable r(ω) > 0
such that
Remark 1.3.1. (i) The property of D being a random closed set is slightly
stronger than
being F × B(X)-measurable and D(ω) being closed; the two properties are
equivalent if F is P-complete, i.e. if for any set A ∈ F with zero probability
all subsets of A also belong to F (see Castaing/Valadier [18]).
(ii) For any x ∈ X and bounded sets A and B from X we have the relation
then D(ω) = ∩n≥0 ∪k≥n Dk (ω) for every ω ∈ Ω and ω → D(ω) is a random
bounded set (D denotes the closure of D in X).
which implies that distX (y, B(ω)) = max {0, |y − a(ω)| − r(ω)}. It is also
clear that intB(ω) = {x : |x − a(ω)| < r(ω)} is a random (open) set.
20 1. General Facts about Random Dynamical Systems
where the intersection taken over all probability measures ν on (Ω, F) and F̄ν
denotes the completion of the σ-algebra F with respect to the measure ν. We
call Fu the universal σ-algebra and F̄ν the ν-completion of F for shortness.
Recall that the P-completion F̄P is the σ-algebra consisting of all subsets A
of Ω for which there are sets U and V in F such that U ⊂ A ⊂ V and
P(U ) = P(V ). The probability measure P can be extended from F to F̄P such
that F̄P is a complete σ-algebra with respect to the extended probability
measure. For details we refer to Cohn [30], for instance. We also note that
θt F̄P = F̄P for any fixed t ∈ R. This property follows from the relation
P(θt U ) = P(U ) for any U ∈ F and t ∈ R.
Proposition 1.3.3 (Projection Theorem). Let X be a Polish space and
M ⊂ Ω × X be a set which is measurable with respect to the product σ-algebra
F × B(X). Then the set
for all t ∈ I and ω ∈ Ω. The set Q is called the separability set of the
collection {Ct }. A process {v(t, ω) : t ∈ I} is said to be separable if the
collection of random sets Ct (ω) = {v(t, ω)} is separable.
It is easy to see that {Ct : t ∈ I} is a separable collection with a separability
set Q if and only if for any t ∈ I and x ∈ Ct (ω) there exist sequences {tn } ⊂ Q
and {xn } ⊂ X such that tn → t and xn → x as n → ∞ and xn ∈ Ctn (ω).
The following proposition gives examples of separable collections of ran-
dom closed sets.
22 1. General Facts about Random Dynamical Systems
Thus by (ii) for any t ∈ I there exists a sequence {tk } ⊂ Q such that tk > t
and
h(t, ω, g(ω, y)) = lim h(tk , ω, g(ω, y))
tk →t
for all y ∈ Y and ω ∈ Ω, where Ct (ω) = h(t, ω, D(ω)). This relation gives
the separability of {h(t, ω, D(ω))}. 2
Proof. It follows from (1.14) that ∪t∈I Ct (ω) = ∪t∈I∩Q Ct (ω). Therefore we
can apply Proposition 1.3.1(v). 2
Proof. The idea of the proof is borrowed from Schenk-Hoppé [89]. It is clear
that VR (ω) is closed for any ω ∈ Ω. Due to Proposition 1.3.1(i) it is sufficient
to prove that {ω : VR (ω) ∩ U = ∅} is measurable for every open set U ⊂ X.
This is equivalent to measurability of the set
{ω : VR (ω) ∩ U = ∅} ≡ {ω : U ⊂ X \ VR (ω)} .
The following notions of random tempered sets and variables play an impor-
tant role in applications of the general theory of RDS connected with ran-
dom and stochastic equations (cf. Chaps. 4 and 5). Roughly speaking, that
a random variable which describes an influence of the random environment
is tempered means that this environment evolves in non-explosive way.
Definition 1.3.3 (Tempered Random Set). A random set {D(ω)} is
said to be tempered with respect to MDS θ = (Ω, F, P, {θt , t ∈ T}) if there
exist a random variable r(ω) and an element y ∈ X such that
sup e−γ|t| |r(θt ω)| < ∞ for all ω ∈ Ω and γ > 0 . (1.16)
t∈T
1
lim log {1 + |r(θt ω)|} = 0 for all ω ∈ Ω .
|t|→∞ |t|
which is weaker than (1.16). However we prefer to use (1.16) because it allows
us to simplify some calculations in the applications below. We also note that
if θ is ergodic, the only alternative to property (1.16) is that
1
lim log {1 + |r(θt ω)|} = +∞ for almost all ω ∈ Ω ,
|t|→∞ |t|
In this section we start to develop methods for studying the qualitative be-
haviour of random dynamical systems. Our main goal is to investigate the
behaviour of expressions of the form x(t) = ϕ(t, θ−t ω)x when t → +∞. At
first sight this object looks a bit strange. However there are at least three
reasons to study the limiting structure of ϕ(t, θ−t ω)x.
The first one is connected with the question of what limiting dynamics
we want to observe. The point is that in many applications RDS are generated
1.4 Dissipative, Compact and Asymptotically Compact RDS 25
if the limit on the right hand side exists. Thus the limiting behaviour of
ϕ(t, θ−t ω)x for all ω determines the long-time behaviour of ϕ(t, ω)x with
respect to convergence in probability.
The third reason is purely mathematical. If on the set of random variables
a(ω) with values in X we define the operators Tt by the formula
(Tt a)(ω) = ϕ(t, θ−t ω)a(θ−t ω), t ∈ R+ ,
then the family {Tt , t ∈ R+ } is a one-parameter semigroup. Indeed, using
the cocycle property we have
(Ts [Tt a])(ω) = ϕ(s, θ−s ω)(Tt a)(θ−s ω) = ϕ(s, θ−s ω)ϕ(t, θ−t−s ω)a(θ−t−s ω)
for some x0 ∈ X and random variable r(ω) and for all ω ∈ Ω. If X is a linear
space and x0 = 0, then the variable r(ω) is said to be a radius of dissipativity
of the RDS (θ, ϕ) in the universe D.
The simplest examples of dissipative RDS are the following ones.
Example 1.4.1 (Discrete Dissipative RDS). Let us consider the RDS con-
structed in Example 1.2.1. Let X = R and Ω0 = {0, 1} be a two-point set.
Assume that the continuous functions f0 and f1 possess the property
Let D be the family of all tempered (with respect to θ) random closed sets
in R. Let D ∈ D and D(ω) ⊂ {x : |x| ≤ r(ω)}, where r(ω) possesses the
property (1.16) (i.e. is a tempered random variable). Then (1.19) implies that
|ϕ(n, θ−n ω)x(θ−n ω)| ≤ an r(θ−n ω) + b · (1 − a)−1 , for all x(ω) ∈ D(ω) .
over a metric dynamical system (Ω, F, P, {θn , n ∈ Z}), where ξ(ω) is a tem-
pered random variable in X. Using (1.20) and (1.21) we have
where
∞
R(ω) = b(1 − a)−1 + ak
ξ(θ−k ω)
k=0
is a tempered random variable. It is easy to see that for every δ > 0 the ball
B δ (ω) = {x :
x
≤ (1 + δ)R(ω)} is a forward invariant absorbing set for
(θ, ϕ) in the universe D of all tempered random closed sets from X.
Example 1.4.3 (Continuous Dissipative RDS). Let (θ, ϕ) be the RDS consid-
ered in Example 1.2.3 from the random ODE ẋ = f (θt ω, x). Assume addi-
tionally that the function f (ω, x) possesses the property
where α > 0 and β ≥ 0 are nonrandom constants. Then it is easy to see that
1 d
· |x(t)|2 ≤ −α|x(t)|2 + β, t>0,
2 dt
for any solution to (1.10). Therefore, since ϕ(t, ω)x = x(t), we have
β
|ϕ(t, ω)x|2 ≤ e−2αt |x|2 + · 1 − e−2αt , t>0.
α
As in Example 1.4.1 this property implies that (θ, ϕ) is dissipative in the
universe D of all tempered (with respect to θ) random closed sets from R.
Moreover the absorbing set B = {x : |x| ≤ 1 + β/α} is a forward invariant
set from D.
28 1. General Facts about Random Dynamical Systems
d
V (ϕ(t, ω)x) + (α + (θt ω)) · V (ϕ(t, ω)x) ≤ β(θt ω) (1.22)
dt
for almost all t > 0, where (ω) is a random variable such that (θt ω)
lies in L1loc (R) for every ω ∈ Ω and
t 0
1 1
lim (θτ ω) dτ = lim (θτ ω) dτ = 0 (1.23)
t→+∞ t 0 t→+∞ t −t
for all ω ∈ Ω;
(iii) there exist positive constants b1 , b2 , δ1 , δ2 and nonnegative numbers c1 and
c2 such that
Then the RDS (θ, ϕ) is dissipative in the universe D of all tempered random
closed sets in X. Moreover there exists a tempered random variable R(ω) ≥ 0
such that for any positive the set
Proof. Let D ∈ D and x(ω) ∈ D(ω) for all ω ∈ Ω. From (1.22) we have that
t
V (ϕ(t, ω)x(ω)) ≤ V (x(ω)) · exp −αt − (θτ ω) dτ
0
t
t
+ β(θs ω) · exp −α(t − s) − (θτ ω) dτ ds .
0 s
1.4 Dissipative, Compact and Asymptotically Compact RDS 29
Therefore
0
V (ϕ(t, θ−t ω)x(θ−t ω)) ≤ V (x(θ−t ω)) · exp −αt − (θτ ω) dτ
−t
(1.26)
0
0
+ β(θs ω) · exp αs − (θτ ω) dτ ds .
−t s
It follows from (1.23) that for any ε > 0 and ω ∈ Ω there exists c(ω) > 0
such that t
(θτ ω) dτ ≤ ε|t| + c(ω), t ∈ R, ω ∈ Ω . (1.27)
0
for all ε > 0 and γ > 0 such that γ + ε < α. This implies that R(ω) is a
tempered random variable. Proposition 1.3.6 and relation (1.24) imply that
B (ω) given by (1.25) is a tempered random closed set. Let
0
e(t, ω) = exp αt − (θτ ω) dτ .
t
Since e(−t, ω) · e(s, θ−t ω) = e(s − t, ω), it follows from (1.28) that
−t
R(θ−t ω) · e(−t, ω) = β(θs ω) · e(s, ω)ds .
−∞
Therefore
V (ϕ(t, θ−t ω)x(θ−t ω)) ≤ (1 + )R(ω) .
30 1. General Facts about Random Dynamical Systems
Thus B (ω) is forward invariant. It follows from (1.24) and (1.26) that
V (ϕ(t, θ−t ω)x(θ−t ω)) ≤ b2
x(θ−t ω)
δ2 + c2 · e−αt + R(ω) .
∗
where Ω ⊆ Ω is a θ-invariant set of full measure. Without loss of generality
we can suppose that Ω ∗ = Ω (see Remark 1.2.1(ii)). Therefore we can apply
Proposition 1.4.1 with α = Eα̃ and (ω) = α̃(ω) − Eα̃.
Example 1.4.4 (Binary Biochemical Model). Consider the RDS (θ, ϕ) gen-
erated in R2 by equations (1.11) over an ergodic metric dynamical system
θ. Let the hypotheses concerning g and αi listed in Example 1.2.4 hold. We
assume in addition that
If
α0
x1 · (x2 + g(x2 )) ≤ · (x21 + x22 ) + β0 , (x1 , x2 ) ∈ R2+ ,
2
where β0 ≥ 0 is a constant, then (1.29) holds with V (x) = x21 + x22 , α̃ =
2αmin (ω) − α0 and β(ω) ≡ 2β0 . Thus the RDS (θ, ϕ) is dissipative in the
universe of all tempered random closed sets from R2 .
The following concepts are useful when the phase space X is infinite-
dimensional.
Definition 1.4.3 (Compact RDS). An RDS (θ, ϕ) is said to be compact
in the universe D, if it is dissipative in D and the absorbing set B is a random
compact set.
If the phase space X of an RDS (θ, ϕ) is compact, then (θ, ϕ) is a compact
RDS. If X is a finite-dimensional space, then any dissipative RDS is compact.
1.4 Dissipative, Compact and Asymptotically Compact RDS 31
Example 1.4.5 (Kick Model). Let (θ, ϕ) be the RDS considered in Exam-
ple 1.4.2. Assume additionally that g is a compact mapping, i.e. g(B) is a
compact set for every bounded set B from X. The set
C(ω) = ϕ(1, θ−1 ω)B δ (θ−1 ω) = g(B δ (θ−1 ω)) + ξ(ω)
is an absorbing forward invariant random compact set for (θ, ϕ) in the uni-
verse D of all tempered random closed sets from X.
Definition 1.4.4 (Asymptotically Compact RDS). An RDS (θ, ϕ) is
said to be asymptotically compact in the universe D, if there exists an at-
tracting random compact set {B0 (ω)}, i.e. for any D ∈ D and for any ω ∈ Ω
we have
lim dX {ϕ(t, θ−t ω)D(θ−t ω) | B0 (ω)} = 0 , (1.30)
t→+∞
1.5 Trajectories
the tail (from the moment t) of the pull back trajectories emanating from D.
0
If D(ω) = {v(ω)} is a single valued function, then ω → γv (ω) ≡ γD (ω) is
said to be the (pull back) trajectory (or orbit) emanating from v.
In the deterministic case Ω is a one-point set and ϕ(t, ω) = ϕ(t) is a semigroup
t
of continuous mappings. Therefore in this case the tail γD has the form
t 0
γD = ϕ(τ )D = ϕ(τ )(ϕ(t)D) = γϕ(t)D ,
τ ≥t τ ≥0
t
i.e. γD is a collection of the “normal” trajectories emanating from ϕ(t)D.
We note that any tail is a forward invariant multifunction. It also follows
from Proposition 1.3.1(v) that in the case of discrete time (T = Z) the closure
γDt (ω) of any tail γ t (ω) is a random closed set. For continuous time we have
D
the following proposition.
Proposition 1.5.1. For any random closed set {D(ω)} the closure γD t (ω)
t
of any tail γD (ω) of the pull back trajectories emanating from D is a random
closed set with respect to the σ-algebra Fu of universally measurable sets.
Proof. The idea of the proof is borrowed from Crauel/Flandoli [36]. The
Representation Theorem 1.3.1 gives that D(ω) = g(ω, Y ), where Y is a Polish
space, g(ω, ·) is continuous for all ω ∈ Ω and g(·, y) is measurable for all y ∈ Y .
Therefore for every x ∈ X we have
d(t, ω) := distX (x, ϕ(t, θ−t ω)D(θ−t ω)) = inf distX (x, ϕ(t, θ−t ω)g(θ−t ω, yk )) ,
k
to F .
u
2
where Q is a separability set of the process t → ϕ(t, θ−t ω)a(θ−t ω). Therefore
we can apply Proposition 1.3.1(v). 2
Hence (t, x) → (t, ϕ̃(−t, ω)) is a bijective mapping from R × X into itself and
−1
ϕ(t, θ−t ω) = ϕ̃(t, θ−t ω) = [ϕ̃(−t, ω)] , t≥0.
Therefore
x∈ ϕ(τ, θ−τ ω)D(θ−τ ω) for all t > 0.
τ ≥t
The compactness of B0 (ω) implies that for some subsequence {nk } and some
b ∈ B0 (ω) we have that bnk → b. This implies that
Thus ΓD (ω) is nonempty. It is clear from (1.30) that any element of the form
(1.33) belongs to B0 (ω). Therefore we have ΓD (ω) ⊂ B0 (ω) and, since ΓD (ω)
is closed, ΓD (ω) is a compact set.
Let us prove that ω → ΓD (ω) is invariant. Using the cocycle property we
have
ϕ(t, ω)x = lim ϕ(t, ω) ◦ ϕ(tn , θ−tn ω)yn = lim ϕ(t + tn , θ−t−tn ◦ θt ω)yn
n→∞ n→∞
for any x ∈ ΓD (ω) of the form (1.33). Due to Proposition 1.6.1 this implies
that ϕ(t, ω)x ∈ ΓD (θt ω). Thus ϕ(t, ω)ΓD (ω) ⊂ ΓD (θt ω) for all t > 0 and
ω ∈ Ω.
Assume that x ∈ ΓD (θt ω) for some t > 0 and ω ∈ Ω. Proposition 1.6.1
implies that
x = lim ϕ(tn , θ−tn ◦ θt ω)yn , (1.35)
n→∞
Let xn ∈ γD
n
(ω) be such that
1
dist(x, xn ) ≤ dist(x, γD
n
(ω)) + , n = 1, 2, . . .
n
Since γDn
(ω) → B0 (ω) as n → ∞ for all ω ∈ Ω, there exist a subsequence
nk = nk (ω) and b ∈ B0 (ω) such that xnk → b. By Proposition 1.6.1 b ∈
ΓD (ω). Therefore
Remark 1.6.1. The existence and measurability of omega-limit sets with re-
spect to the universal σ-algebra can be proved under a weaker property than
the asymptotic compactness of RDS (θ, ϕ). Assume that {D(ω)} is a random
closed set and for every ω ∈ Ω there exists a compact set BD (ω) ⊂ X such
that
lim dX {ϕ(t, θ−t ω)D(θ−t ω) | BD (ω)} = 0 ,
t→+∞
where dX {A|B} = supx∈A distX (x, B). Then, as in the proof of Proposi-
tion 1.6.2, it follows from Proposition 1.5.1 that ΓD exists and ω → ΓD (ω)
is an invariant random compact set with respect to the universal σ-algebra
Fu . If we additionally assume that the closure γDt (ω) of the tail γ t (ω) is a
D
random closed set for every t ≥ 0 (cf. Proposition 1.5.3 and Remark 1.5.1),
then ΓD is a random compact set with respect to F. We refer to Crauel [33]
for other results concerning the measurability of omega-limit sets.
The following two assertions provide us with conditions which guarantee that
{ΓD (ω)} is a random compact set with respect to the σ-algebra F.
Proposition 1.6.3. If {D(ω)} is a forward invariant random compact set
for the RDS (θ, ϕ), then the multifunction ω → ΓD (ω) is an invariant random
compact set with respect to F and ΓD (ω) ⊂ D(ω).
38 1. General Facts about Random Dynamical Systems
Proof. The compactness of γat∗ (ω) implies that Γa (ω) is a nonempty com-
pact set for all ω ∈ Ω. Therefore we can use Proposition 1.5.2, the for-
mula ΓD (ω) = ∩n∈Z+ γD
n (ω) and the argument given in the proof of Proposi-
tion 1.6.2. 2
1.7 Equilibria
A special case of omega-limit sets are random equilibria. They are the random
analog of deterministic fixed points and generate stationary stochastic orbits
(cf. Arnold [3], Arnold/Schmalfuss [11] and Schmalfuss [94]).
Definition 1.7.1. A random variable u : Ω → X is said to be an equilibrium
(or fixed point, or stationary solution) of the RDS (θ, ϕ) if it is invariant
under ϕ, i.e. if
Remark 1.7.1. The problem of the construction of equilibria for general RDS
is rather complicated. The following example demonstrates the difficulties in
the construction of equilibria. Let us consider the RDS on R+ constructed
in the Introduction (cf. also Example 1.4.1) with f0 (x) = 12 x and f1 (x) =
1 1
2 + f0 (x) = 2 (1 + x). Both functions f0 (x) and f1 (x) have a fixed point:
f0 (0) = 0 and f1 (1) = 1. To obtain an equilibrium we should look for a
solution to the equation fω0 (u(ω)) = u(θ1 ω), where ω = {ωi | i ∈ Z} is a two-
sided sequence consisting of zeros and ones and θ1 is the left one-symbol shift
operator. It is clear that an equilibrium u(ω) is not simply a random variable
which takes as its values the fixed points 0 and 1 of the mappings f0 (x) and
f1 (x). The variable u(ω) can really depend on the sequence ω = {ωi | i ∈ Z} in
a very complicated way. However we prove in Chap.3 that this RDS possesses
a unique globally asymptotically stable equilibrium in R+ with its values
inside the interval (0, 1).
We also note that the results by Ochs/Oseledets [87] and Ochs [85]
show that it is impossible to generalize topological fixed point theorems to
the case of random dynamical systems. However, as we will see in Chaps.3–6,
there are more simple approaches which allow us to construct equilibria for
monotone RDS.
G = {(ω, u, v) : u, v ∈ D(ω), ω ∈ Ω} ⊂ Ω × X × X
V (θt ω, ϕ(t, ω)u, ϕ(t, ω)v) ≤ V (ω, u, v) for all t > 0, ω ∈ Ω ; (1.39)
Proof. Assume that the RDS (θ, ϕ) has two equilibria u1 and u2 in D such
that u1 (ω) = u2 (ω) on a measurable set U ⊂ Ω with P(U ) > 0. It follows
from condition (iii) that
V (θt ω, ϕ(t, ω)u1 (ω), ϕ(t, ω)u2 (ω)) < V (ω, u1 (ω), u2 (ω)) < ∞ (1.40)
40 1. General Facts about Random Dynamical Systems
for all ω ∈ U and t > 0. Since u1 and u2 are equilibria, (1.40) is equivalent
to
V (θt ω, u1 (θt ω), u2 (θt ω)) < V (ω, u1 (ω), u2 (ω)) < ∞
for all ω ∈ U and t > 0. From (1.39) we also have
ft (ω) := V (θt ω, u1 (θt ω), u2 (θt ω)) and f (ω) := V (ω, u1 (ω), u2 (ω))
have the same probability distribution for every t > 0, but satisfy ft (ω) ≤
f (ω) for ω ∈ Ω and ft (ω) < f (ω) for ω ∈ U . This contradicts the assumption
P(U ) > 0. Thus for any fixed t > 0 we have f (θt ω) = f (ω) on a set of full
measure. Let
Ωn = {ω : f (θn ω) = f (ω)}, n ∈ Z+ .
The sets Ωn are F-measurable and P(Ωn ) = 1. Property (1.39) implies that
f (θt ω) = f (ω) for all t ∈ [0, n] and ω ∈ Ωn . Therefore f (θn−k θs ω) = f (θs ω)
for all s ∈ [0, k] and ω ∈ Ωn , where k ≤ n. Thus
Let Ω ∗ = ∩n≥1 Ωn . It is clear that P(Ω ∗ ) = 1 and f (θt ω) = f (ω) for all t ≥ 0
and ω ∈ Ω ∗ . From (1.41) we also easily have that θs Ω ∗ ⊆ Ω ∗ for all s ≥ 0.
Therefore Ω̃ = ∩s≥0 θs Ω ∗ = ∩n∈Z+ θn Ω ∗ is F-measurable θ-invariant set such
that P(Ω̃) = 1. Since Ω̃ ⊂ Ω ∗ , we have that u1 (ω) = u2 (ω) for all ω ∈ Ω̃. 2
We note that Proposition 1.7.1 is wrong without the assumption (iii). Indeed,
the identical mapping f (x) = x in R possesses the property |f (x) − f (y)| =
|x − y| and every point x ∈ R is an equilibrium for f . See also the example
of an RDS given in Remark 4.2.1 in Chap.4.
Example 1.7.2. Consider the one-dimensional random differential equation
satisfies the hypotheses of Proposition 1.7.1. The same is true for V ∗ (u, v) :=
−V (u, v) provided that g(x) is strictly increasing.
1.8 Random Attractors 41
in probability as t → ∞, i.e.
for any δ > 0. Thus any pull back attractor is a forward attractor with respect
to convergence in probability. We refer to Ochs [86] for some discussion of
the theory of attractors based on convergence in probability. We note that
an example given in Arnold [3] shows that pull back convergence (1.42)
does not imply forward convergence, i.e. the closeness of ϕ(t, ω)D(ω) and
A(θt ω) in the topology of the space X for every ω ∈ Ω. We also refer to
Scheutzow [91] for a short survey of other (non-equivalent) definitions of a
random attractor.
(ii) An attractor depends crucially on a choice of universe D. Indeed, the
deterministic dynamical system in R generated by the equation
ẋ = x − x3
bounded subsets of R and the set {−1, 0, 1} is the attractor in the universe of
all one-point subsets of R. We also note that there exists some classification
of random attractors (see, e.g., Crauel [34]) depending on the choice of
families of sets which are attracted (set attractors, point attractors, etc.).
(iii) Sometimes it is convenient to consider random attractors which do not
belong to the corresponding universe (see Crauel [33, 34], Crauel/Debus-
sche/Flandoli [35], Crauel/Flandoli [36]).
Proof. Assume that there exist two random attractors A1 (ω) and A2 (ω) in
the universe D. Since ϕ(t, ω)A1 (ω) = A1 (θt ω), we have
for all t > 0. Therefore the attraction property (1.42) implies that
Proof. We follow the line of arguments given for the deterministic case (see,
e.g., Temam [104] or Chueshov [20]).
Let A(ω) be defined by (1.43). Proposition 1.6.2 implies that A(ω) is a
nonempty invariant set and is a compact subset of B0 (ω) for all ω ∈ Ω.
1.8 Random Attractors 43
Assume now that property (1.42) is not true for some D ∈ D. Then there
exist > 0 and sequences tn → ∞ and yn ∈ D(θ−tn ω) such that
for some ω ∈ Ω. It follows from (1.30) that there exists a sequence {bn } ⊂
B0 (ω) such that
exists for some subsequence {nk }. Proposition 1.6.1 and relation (1.45) imply
that z ∈ ΓD (ω) ⊂ A(ω). Thus we have
A(ω) = ϕ(n, θ−n ω)A(θ−n ω) ⊂ ϕ(n, θ−n ω)B0 (θ−n ω) for all n ∈ Z+ .
Therefore A(ω) ⊂ ΓN∗ (ω) for any ω ∈ Ω and N ∈ Z+ . On the other hand, it
is clear from (1.43) that ΓN∗ (ω) ⊂ A(ω). Thus (1.44) is proved.
To prove that {A(ω)} is a random compact set we use Proposition 1.3.1(iv)
and relation (1.44). 2
Remark 1.8.2. (i) It is clear that if the RDS (θ, ϕ) has a random compact
attractor, then (θ, ϕ) is asymptotically compact. Thus Theorem 1.8.1 implies
that (θ, ϕ) possesses a random compact attractor in D if and only if this RDS
is asymptotically compact in D with an attracting set from D.
44 1. General Facts about Random Dynamical Systems
(ii) Under the hypotheses of Theorem 1.8.1 similarly to the deterministic case
(see, e.g., Chueshov [20, Sect.1.5.2]) we can prove that
for any absorbing set B ∈ D of the RDS (θ, ϕ), where h(A|B) is the Hausdorff
distance defined by the equality
This property means that the set AB t (ω) := ϕ(t, θ−t ω)B(θ−t ω) provides us
with an approximate image of the random attractor A(ω) for t large enough.
We also refer to Arnold/Schmalfuss [12] for the study of stability prop-
erties of random attractors for finite-dimensional RDS.
Corollary 1.8.1. Let (θ, ϕ) be a dissipative RDS in the universe D with an
absorbing set from D. Assume that the phase space X is locally compact. Then
the RDS (θ, ϕ) possesses a unique global random attractor in the universe D.
Proof. In this case any closed bounded set is compact. Therefore (θ, ϕ) is a
compact RDS and we can apply Theorem 1.8.1. 2
Corollary 1.8.2. Assume that for the RDS (θ, ϕ) the hypotheses of Proposi-
tion 1.4.1 on the dissipativity of an RDS possessing a Lyapunov type function
hold. Let the phase space X be finite-dimensional. Then the RDS (θ, ϕ) pos-
sesses a unique random attractor in the universe D of all tempered random
closed sets in X.
Proof. Since X is finite-dimensional, Proposition 1.4.1 implies that (θ, ϕ) is
a compact RDS. Thus we can apply Theorem 1.8.1. 2
Theorem 1.8.1 and Corollaries 1.8.1 and 1.8.2 imply the existence of random
attractors for the RDS considered in Examples 1.4.1, 1.4.3, 1.4.4 and 1.4.5.
Below we also need the following simple assertion concerning attractors of
equivalent RDS (cf. Keller/Schmalfuss [63] and Imkeller/Schmalfuss
[59]).
Proposition 1.8.3. Let (θ, ϕ1 ) and (θ, ϕ2 ) be two RDS over the same MDS
θ with phase spaces X1 and X2 resp. Assume that the systems (θ, ϕ1 ) and
(θ, ϕ2 ) are conjugate by a random homeomorphism T from X1 onto X2 (see
Definition 1.2.4) and there exists a compact random attractor A1 for the
RDS (θ, ϕ1 ) in the universe D1 . Then the RDS (θ, ϕ2 ) possesses a random
attractor A2 in the universe
D2 = {T (ω, D(ω))} : {D(ω)} ∈ D1 .
where D2 (ω) = T (ω, D1 (ω)) and dX {A|B} = supx∈A distX (x, B). If d2 (ω, t)
does not tend to 0 as t → ∞ for some ω, then there exist tn → ∞ and
bn ∈ D1 (θ−tn ω) such that
for some ε > 0, where xn (ω) = ϕ1 (tn , θ−tn ω)bn . Since A1 (ω) is an attractor
for (θ, ϕ1 ), there exists a sequence {an } ⊂ A1 (ω) such that
The compactness of A1 (ω) implies that xnk (ω) → a for some subsequence
{nk } and a ∈ A1 (ω). Therefore distX2 (T (ω, xnk (ω)) , T (ω, a)) → 0. This
contradicts (1.47). Thus A2 is a random attractor for (θ, ϕ2 ). 2
for any D ∈ D.
46 1. General Facts about Random Dynamical Systems
Proof. Let r(ω) be a radius of dissipativity of (θ, Φ). Then for any D ∈ D
and for any λ > 0 there exists a time tλ,D (ω) > 0 such that
Therefore
r(ω)
sup
Φ(t, θ−t ω)v
≤ , t ≥ tλ,D (ω) .
v∈D(θ−t ω) λ
Hence
r(ω)
lim sup sup
Φ(t, θ−t ω)v
≤
t→+∞ v∈D(θ−t ω) λ
for all λ > 0. This implies (1.50).
Vice versa, (1.50) implies that the deterministic ball {x :
x
≤ 1} is an
absorbing set for (θ, Φ). 2
Remark 1.9.1. Let D be a universe such that {0} ∈ D. It is easy to see that
the dissipativity of the affine RDS (θ, ϕ) implies the dissipativity of its linear
part (θ, Φ).
exists for all ω ∈ Ω and is an equilibrium for the RDS (θ, ϕ). This equilibrium
is globally asymptotically (pull back) stable in D, i.e.
for any D ∈ D. Moreover {u(ω)} ∈ D and the RDS (θ, ϕ) possesses a unique
equilibrium with this property.
1.9 Dissipative Linear and Affine RDS 47
ψ(τ, θ−τ ω) = Φ(t, θ−t ω)ψ(τ − t, θ−τ ω) + ψ(t, θ−t ω), τ >t≥0. (1.53)
Since
ϕ(t, θ−t ω)v − u(ω) = Φ(t, θ−t ω)v − Φ(t, θ−t ω)u(θ−t ω) (1.56)
and Proposition 1.9.1 we obtain (1.52). Finally, if there exists another equi-
librium v(ω) with the property {v(ω)} ∈ D, then we have
Remark 1.9.2. If in Proposition 1.9.2 the universe D contains all bounded de-
terministic sets, then any equilibrium v(ω) coincides with u(ω) almost surely.
Indeed, since (θ, Φ) is dissipative, from Proposition 1.9.1 we have that
lim P(UN
δ
)=0
t→∞
It is clear that
Hence
for every ε > 0 and all x ∈ X, where Rε (ω) > 0 is a tempered random
variable.
We refer to Arnold [3, Part II] for conditions which guarantee the existence
of the top Lyapunov exponent and for a comprehensive presentation of the
theory of Lyapunov exponents for finite-dimensional RDS.
Following the line of argument given in the proof of Proposition 1.9.2 we
can easily prove the next assertion.
Proposition 1.9.3. Let (θ, ϕ) be an affine RDS with the cocycle given by
(1.48). Assume that the linear RDS (θ, Φ) has top Lyapunov exponent λ < 0
and for every ω ∈ Ω there exists a tempered random compact set B0 (ω) such
that
lim distX (ψ(t, θ−t ω), B0 (ω)) = 0 .
t→∞
Then the limit in (1.51) exists and belongs to B0 (ω) for all ω ∈ Ω ∗ . It is
an equilibrium on Ω ∗ , i.e. the property ϕ(t, ω)u(ω) = u(θt ω) holds for all
ω ∈ Ω ∗ . Moreover this equilibrium is unique almost surely and
1.10 Connection Between Attractors and Invariant Measures 49
lim e γt
sup
ϕ(t, θ−t ω)v − u(ω)
= 0, ω ∈ Ω∗ , (1.58)
t→+∞ v∈D(θ−t ω)
Proof. As in the proof of Proposition 1.9.2 using (1.53) we find that for any
t > 0 there exist b ∈ B0 (ω) and b1 (t) ∈ B0 (θ−t ω) such that (1.55) holds.
Since B0 (ω) is a tempered, there exists a tempered random variable r(ω) > 0
such that
b1 (t)
≤ r(θ−t ω). Therefore it follows from (1.57) that
provided λ + ε < 0. Thus (1.55) implies that the limit in (1.51) exists for
ω ∈ Ω ∗ . It is clear that u(ω) ∈ B0 (ω) for all ω ∈ Ω ∗ and it is an equilibrium
on Ω ∗ . Using relation (1.56) with an arbitrary v ∈ X, we find that
To conclude this section we refer to Arnold [3, Sect.5.6] for a more detailed
study of the asymptotic properties of affine systems with general hyperbolic
linear parts in finite-dimensional spaces.
A number of interesting properties follow from the fact that the RDS (θ, ϕ)
has a random attractor. One of them is the existence of an invariant measure
of (θ, ϕ) in the sense of the theory of RDS. In this section we introduce the
corresponding notions and briefly discuss the properties of these measures.
50 1. General Facts about Random Dynamical Systems
One can prove (see Arnold [3, p.53]) that the family of sequences
for all t ≥ 0. Since θ−t = θt−1 , we have θt A = A for all t ∈ R. The ergodicity
of θ implies that we have either P(A) = 0 or µ(A) = 1. It is clear from (1.59)
that µ(C) = P(A). Thus µ is ϕ-ergodic.
The following assertion (see, e.g., Crauel/Flandoli [36], Crauel [32, 33]
and Arnold [3]) describes the relation between invariant measures and for-
ward invariant random sets.
Proposition 1.10.1. A probability measure µ on (Ω ×X, F ×B) is invariant
for (θ, ϕ) if and only if its disintegration µω possesses property ϕ(t, ω)µω =
µθt ω P-almost surely, i.e. for any f ∈ Cb (X) we have
f (ϕ(t, ω)x)µω (dx) = f (x)µθt ω (dx) P − almost surely .
X X
Moreover there exists a forward invariant random closed set {C(ω)} such that
µω (C(ω)) = 1 for almost all ω ∈ Ω.
On the other hand for any forward invariant random compact set {C(ω)}
there exists an invariant measure µ concentrated on {C(ω)}, i.e. µ{(ω, x) :
x ∈ C(ω)} = 1. In particular if the RDS (θ, ϕ) possesses a random compact
attractor {A(ω)} in the universe D which contains all bounded determinis-
tic sets, then there exists an invariant measure µ concentrated on {A(ω)}.
Moreover in the last case every invariant probability measure is concentrated
on {A(ω)}.
We refer to Scheutzow [90] for the proof of this result. We also refer to
Schenk-Hoppé [89] for additional properties of invariant measures in the
case of invertible cocycles, i.e. for RDS with time T (not T+ ).
Let us define the future F+ and the past F− σ-algebras for RDS (θ, ϕ) by
the formulas
F+ = σ{ω → ϕ(τ, θt ω) : t, τ ≥ 0}
1.10 Connection Between Attractors and Invariant Measures 53
and
F− = σ{ω → ϕ(τ, θ−t ω) : 0 ≤ τ ≤ t} ,
where σ{fα (ω) : α ∈ Λ} denotes the σ-algebra generated by the mappings
{fα }, where α ∈ Λ.
Definition 1.10.2 (Markov Measure). A probability measure µ on (Ω ×
X, F×B) is said to be a Markov measure if its disintegration µω is measurable
with respect to the past σ-algebra F− .
The following theorem (see, e.g. Crauel [31, 32], Crauel/Flandoli [36]
and Arnold [3]) shows that invariant Markov measures supported by the
random attractor for the RDS (θ, ϕ) generate stationary probability measures
in the phase space of this RDS.
Theorem 1.10.1. Assume that the RDS (θ, ϕ) possesses a random compact
attractor {A(ω)} in the universe D which contains all bounded determin-
istic sets. Then there an exists invariant Markov measure µ supported by
{A(ω)}, i.e. µ{(ω, x) : x ∈ A(ω)} = 1. Assume additionally that the pro-
cesses {ϕ(t, ω)x : x ∈ X} form a Markov family, i.e. the stochastic kernels
Pt (x, B) := P{ω : ϕ(t, ω)x ∈ B} satisfy the Chapman-Kolmogorov equation
Pt+s (x, B) = Pt (y, B)Ps (x, dy), t, s ≥ 0, B ∈ B .
X
If the σ-algebras F− and F+ are independent, then for any invariant Markov
measure µ supported by {A(ω)} the measure on (X, B) defined by the for-
mula
(B) = µω (B)P(dω), B ∈ B ,
Ω
is a stationary probability measure for the Markov semigroup associated with
the family {ϕ(t, ω)x : x ∈ X}, i.e.
(B) = Pt (x, B)(dx), B ∈ B ,
X
and
|f (ω, x)| ≤ CK (ω), |f (ω, x) − f (ω, y)| ≤ CK (ω) · |x − y| (2.2)
for any x, y ∈ K and ω ∈ Ω. Here and below | · | is the Euclidean distance in
Rd .
We emphasize that assumptions (2.1) and (2.2) are stated here for all
ω ∈ Ω. This does not spoil generality because we can apply the following
simple perfection procedure. Assume that (2.1) and (2.2) hold almost surely
and consider the sets
b
ΩN = ω : CKN (θt ω) dt < ∞ for all a < b ,
a
and
Definition 2.1.1. A function x(t, ω) = (x1 (t, ω), . . . , xd (t, ω)) is said to
be a local solution to problem (2.3) if for every ω ∈ Ω there exists t0 =
t0 (ω, x0 ) > 0 such that x(t, ω) is continuous with respect to t from the inter-
val (0, t0 (ω, x0 )) into Rd for each ω ∈ Ω and satisfies the equation
t
x(t, ω) = x0 + f (θτ ω, x(τ, ω)) dτ, 0 < t < t0 (ω, x0 ), ω ∈ Ω . (2.5)
0
for all t from the interval (0, t0 (ω, x0 )) of the existence of the solution x(t, ω).
Moreover the determinant detDx0 x(t, ω) satisfies Liouville’s equation
t
detDx0 x(t, ω) = exp tr{Dx f (θτ ω, x(τ, ω))} dτ (2.7)
0
where x(t, ω; x0 ) is the global solution to problem (2.3) for the initial data
x0 ∈ Rd . Moreover the mapping (t, x) → ϕ(t, ω, x) is continuous for all
58 2. Generation of Random Dynamical Systems
Proof. This follows from Proposition 2.1.1. We refer to Arnold [3] for de-
tails. 2
Corollary 2.1.1. Assume that f (ω, x) satisfies (2.1) and (2.2) and there
exist random variables c1 (ω) and c2 (ω) such that t → cj (θt ω) is locally inte-
grable and
x, f (ω, x) ≤ c1 (ω)|x|2 + c2 (ω) , (2.10)
where ·, · is the inner product in Rd generated by the Euclidean norm | · |.
Then the conclusions of Theorem 2.1.1 are true.
Proof. Under condition (2.10) we obviously have (cf. Remark 2.1.1) that
1 d
· |x(t, ω)|2 ≤ c1 (θt ω)|x(t, ω)|2 + c2 (θt ω)
2 dt
on the existence semi-interval [0, t0 (ω, x0 )). Consequently the Gronwall lemma
gives
t
t
|x(t, ω)|2 ≤ |x0 |2 + 2 |c2 (θτ ω)| dτ · exp 2 |c1 (θτ ω)| dτ . (2.11)
0 0
Remark 2.1.2. (i) Under the hypotheses of Corollary 2.1.1 the cocycle ϕ(t, ω)
possesses the following property which is important in the study of pull back
trajectories (cf. Proposition 1.5.2): for every x ∈ Rd and ω ∈ Ω the function
t → ϕ(t, θ−t ω)x is right continuous on R+ . To prove this we note that (2.5)
implies the relation
0
ϕ(t, θ−t ω)x − x = f (θs ω, ϕ(t + s, θ−t ω)x) ds .
−t
It follows from (2.11) that for any T > 0 there exists CT (ω) > 0 such that
By the cocycle property we have ϕ(s + t, θ−s−t ω) = ϕ(s, θ−s ω)ϕ(t, θ−s−t ω)
for any t, s ≥ 0. Therefore (2.13) implies that
lim |ϕ(t, θ−t ω)x − ϕ(s, θ−s ω)x| = 0 for any s > 0, x ∈ Rd , ω ∈ Ω .
t→s+0
for all ω ∈ Ω with a negative constant α and that the variable max{0, c2 (ω)}
is tempered. Then under the hypotheses of Corollary 2.1.1 we can apply
Proposition 1.4.1 (see also Remark 1.4.1) with the function V (x) = |x|2
to prove that the RDS generated by (2.3) is dissipative in the universe of all
tempered subsets of Rd and possesses a random attractor in this universe (see
Corollary 1.8.2). In particular (2.12) generates a dissipative RDS provided
that the assumptions of Example 1.4.4 hold.
The following result contains useful information on the top Lyapunov expo-
nent of the linear RDS (θ, Φ). It is an easy consequence of the multiplicative
ergodic theorem (see, e.g., Arnold [3, Chaps.3,4]).
Theorem 2.1.2. Assume that the matrix A(ω) satisfies
A(·)
∈ L1 (Ω, F, P)
and
A(θt ω)
∈ L1loc (R) for all ω ∈ Ω. Let Φ(t, ω) be the linear cocycle in
Rd generated by (2.16). Then there exists a θ-invariant set Ω ∗ ⊂ Ω of full
measure such that for each x ∈ Rd \ {0} the Lyapunov exponent
1
λ(ω, x) := lim log |Φ(t, ω)x| (2.17)
t→+∞ t
exists for all ω ∈ Ω ∗ . For every ω ∈ Ω ∗ the image of the function x →
λ(ω, x) is a finite set. If θ is an ergodic metric dynamical system, then λ :=
maxx∈Rd \{0} λ(ω, x) is a constant on Ω ∗ and it is the top Lyapunov exponent
in the sense of Definition 1.9.1. Moreover in this case we have
where µmin (ω) and µmax (ω) are the least and the greatest eigenvalues of the
Hermitian part of the matrix A(ω). In particular the top Lyapunov exponent
λ belongs to the interval [Eµmin , Eµmax ].
Proof. This follows directly from Arnold [3, Theorem 3.4.1] (see also
Arnold [3, Example 3.4.15]). Relation (2.18) follows from the Birkhoff–
Khinchin ergodic theorem (see, e.g., Sinai Ya. G. [100]) and the argument
given in Hartman [51, p.56], see also Arnold [3, Theorem 6.2.8]. 2
We also note that if Eµmax < 0 and b(ω) is tempered, then the affine RDS
(θ, ϕ) generated by (2.14) over an ergodic θ is dissipative in the universe of
all tempered subsets of Rd (see Remark 2.1.2(ii)) and both Propositions 1.9.2
and 1.9.3 can be applied here.
Example 2.1.2 (1D Affine RDE). Consider the one-dimensional RDE
over an ergodic metric dynamical system θ, where α(ω) and β(ω) are random
variables such that t → α(θt ω) and t → β(θt ω) are locally integrable. This
equation generates an affine RDS in
t
R. The cocycle ϕ has the form (2.15)
with Φ(t, ω)x = x exp 0 α(θτ ω)dτ . If α ∈ L1 (Ω, F, P), then the Birkhoff–
Khinchin ergodic theorem implies that the (top) Lyapunov exponent for (θ, Φ)
is λ = Eα (see Remark 1.4.1). The RDS (θ, ϕ) is dissipative in the universe
D of all tempered subsets of R provided that Eα < 0 and β(ω) is a tempered
random variable. In this case
0
0
ψ(t, θ−t ω) = β(θs ω) exp α(θτ ω)dτ ds
−t s
in representation (1.48) and therefore (see Propositions 1.9.2 and 1.9.3 and
also Remark 1.9.2) the RDS (θ, ϕ) possesses a unique exponentially stable
equilibrium
0
0
u(ω) = β(θs ω) exp α(θτ ω)dτ ds .
−∞ s
In the case Eα > 0 the RDS (θ, ϕ) is not dissipative in D (see Remark 1.9.1).
Nevertheless a simple calculation shows that
∞
s
v(ω) = − β(θs ω) exp − α(θτ ω)dτ ds
0 0
ẋ = α + β(θt ω) · x − x3
ẋ = β(θt ω) · x(1 − x) .
Example 2.2.2 (Binary Biochemical Model). Let (θ, ϕ) be the RDS consid-
ered in Example 2.1.1. If g(0) ≥ 0, then R2+ = {x = (x1 , x2 ) : xi ≥ 0} is a
forward invariant set for (θ, ϕ). This property is important because x1 and
x2 represent concentrations of macro-molecules.
Theorem 2.2.2. Assume that (2.1) and (2.2) hold. Let O ⊆ Rd be a deter-
ministic forward invariant open set for the RDE (2.3) and D be a closed set
in Rd such that (i) D ∩ O = ∅, (ii) D has an outer normal at every point of
the set ∂D ∩ O and (iii) relation (2.19) holds for any x ∈ ∂D ∩ O. Then the
set D ∩ O is a deterministic forward invariant set for the RDE (2.3).
In the proof of Theorem 2.2.2 we rely on some ideas presented in Bony [17].
We start with the following deterministic lemma.
Lemma 2.2.1. Assume that f (t) is a continuous function on the segment
[a, b] such that
1
lim inf (f (t + h) − f (t)) ≥ −m(t)
h→0, h<0 |h|
for almost all t ∈ (a, b), where m(t) ∈ L1 (a, b). Then
t2
f (t2 ) − f (t1 ) ≤ m(τ ) dτ for all a ≤ t1 < t2 ≤ b . (2.20)
t1
for all t ∈ B, where B is a measurable set of full measure in (a, b). To obtain
(2.20) we should prove that g(t) is a nonincreasing function on [a, b]. It is
sufficient to prove that the function Φ(t) = g(t) − γt is nonincreasing for any
γ > 0. From (2.21) we have
1
lim inf (Φ(t + h) − Φ(t)) ≥ γ > 0, t∈B. (2.22)
h→0, h<0 |h|
This implies that for every t ∈ B there exists h(t) > 0 such that
Let t1 < t2 be points from B. Consider the covering of the segment [t1 , t2 ]
by intervals (t − min{h(t2 ), h(t)}, t), where t ∈ B. It is clear that there exists
a finite subcovering. Moreover we can choose the points τ1 < τ2 < . . . < τN
from B ∩ (t1 , t2 ) such that
and
τk ∈ (τk+1 − h(τk+1 ), τk+1 ), k = 1, . . . N − 1 .
Therefore from (2.23) we have
Proof of Theorem 2.2.2. Let x(t) be a local solution to (2.3) for some fixed
ω with initial data from D ∩ O. Assume that this solution may leave the set
D ∩ O. Since O is forward invariant, there exist a point x∗ ∈ ∂D ∩ O and a
semiinterval (t0 , t1 ] such that x(t0 ) = x∗ , x(t) ∈ Br (x∗ ) and x(t) ∈ D ∩ O
for t ∈ (t0 , t1 ]. Here Br (x∗ ) is an open ball with center x∗ and with radius r
chosen such that B2r (x∗ ) ⊂ O.
Let hn < 0 and hn → 0. Assume that t, t + hn ∈ (t0 , t1 ] and denote
x = x(t) and xn = x(t + hn ) for short. Let δ(t) = dist(x(t), D ∩ B2r (x∗ )).
Since xn → x, it is clear that we can suppose that there exists a sequence
{yn } ⊂ ∂D ∩ B2r (x∗ ) which converges to some element y ∈ ∂D ∩ B2r (x∗ ) such
that δ(t + hn ) = |xn − yn | and δ(t) = |x − y|. Therefore we obtain the relation
|xn − yn |2 − |x − yn |2
δ(t + hn ) − δ(t) ≥ |xn − yn | − |x − yn | = .
|xn − yn | + |x − yn |
Thus we have
|wn |2 − |wn + vn |2 −2wn , vn − |vn |2
δ(t + hn ) − δ(t) ≥ = ,
|wn | + |wn + vn | |wn | + |wn + vn |
where wn = xn − yn and vn = x − xn . We have that wn → x − y = 0 and
vn · h−1
n → −f (θt ω, x) for almost all t ∈ (t0 , t1 ). Therefore
1 w vn
n
lim inf {δ(t + hn ) − δ(t)} ≥ − lim ,
n→∞ |hn | n→∞ |wn | |hn |
x−y
=− , f (θt ω, x)
|x − y|
It follows from (2.2) that there exists a constant C(ω, t) > 0 such that
t1
C(ω, t)dt < ∞, and |f (θt ω, x) − f (θt ω, y)| ≤ C(ω, t)|x − y| .
t0
2.3 The Itô and Stratonovich Stochastic Integrals 65
Therefore
1
lim inf {δ(t + h) − δ(t)} ≥ −C(ω, t) · δ(t).
h→0, h<0 |h|
Using Lemma 2.2.1 we obtain
t
δ(t) − δ(s) ≤ C(ω, τ )δ(τ ) dτ for all t0 ≤ s < t ≤ t1 .
s
i.e. for any local solution x(t; x0 ) to problem (2.24) the property x0 ∈ D im-
plies that x(t; x0 ) ∈ D for all t ∈ (0, t(x0 )), where (0, t(ω, x0 )) is the maximal
interval of the existence x(t; x0 ).
In this section we recall several standard definitions and facts from stochastic
analysis and give a short description of stochastic integration. We need this
to construct RDS generated by Itô and Stratonovich stochastic differential
equations in the next section. For details concerning stochastic integration
66 2. Generation of Random Dynamical Systems
We note that the set of all Ft -adapted continuous random processes on [a, b]
are dense in L2 [a, b], i.e. for any f ∈ L2 [a, b] there exists a sequence {fn }
Ft -adapted continuous processes such that
b
lim |f (t, ω) − fn (t, ω)|2 dt = 0 almost surely .
n→∞ a
For any f ∈ L2 [a, b] we can uniquely define the Itô stochastic integral
m
t t
Iat (f ) = f (τ, ω), dWτ (ω) = fi (τ, ω)dWτi (ω), t ∈ [a, b] ,
a i=1 a
n
Iat (f ) = (P)- lim f (t ∧ tk ), Wt∧tk+1 − Wt∧tk (2.26)
|∆|→0
k=1
The Stratonovich integral can be defined for more general mappings G (see,
e.g., Kunita [74]). However this will not be necessary in our subsequent
considerations.
We have the following relation between the Stratonovich and Itô integrals
(see, e.g., Kunita [74]). Let G(x) be a C 2 -mapping from Rd into Rm and let
X(t, ω) = (X1 (t, ω), . . . , Xd (t, ω)) be continuous semimartingale with values
in Rd . Then
2.3 The Itô and Stratonovich Stochastic Integrals 69
t t
G(X(τ, ω)), ◦dWτ (ω) = G(X(τ, ω)), dWτ (ω)
a a
1
m d t
∂Gi
+ (X(τ, ω))d W i , Xj τ ,
2 i=1 j=1 a ∂xj
1
m d t
∂Gi
+ (X(τ, ω))hji (τ, ω)dτ .
2 i=1 j=1 a ∂xj
d
t
∂G
G(X(t, ω)) − G(X(a, ω)) = (X(τ ))dXi (τ )
i=1 a ∂xi
d
1 t ∂2G
m
+ (X(τ, ω))hil (τ, ω)hjl (τ, ω)dτ .
2 i,j=1 a
∂xi xj
l=1
where hij and ci (ω) are the same as above, then we have
d
t
∂G
G(X(t, ω)) − G(X(a, ω)) = (X(τ )) ◦ dXi (τ ) .
i=1 a ∂xi
t
similarly for the Stratonovich integral a
g(τ ) ◦ dXi (τ ).
|D f (x) − Dα f (x)|
α
f
k,δ =
f
k,0 + sup , 0<δ≤1.
x=y |x − y|δ
|α|=k
! ∂ |α|
Here α = (α1 , . . . , αd ), |α| = αi and Dα = α α
∂x1 1 ...∂xd d
. For a closed set
D ⊂ Rd we denote by Cbk,δ (D) the space of restrictions to D of functions from
Cbk,δ (Rd ).
Now we consider the following system of Itô SDEs
m
dxi = hi (x1 , . . . , xd )dt + σij (x1 , . . . , xd )dWtj , i = 1, . . . , d . (2.28)
j=1
We assume that hi (x) and σij (x) are functions from Cb0,1 (Rd ).
Definition 2.4.2 (Solutions to Itô SDE). A random process x(t, ω) on
R+ with values in Rd is said to be a solution to the system of Itô SDEs (2.28)
with initial data x∗ (ω) = (x∗1 (ω), . . . , x∗d (ω)) if it is an Ft -adapted continuous
process on R+ satisfying the integral equation
t
xi (t, ω) = x∗i (ω) + hi (x1 (τ, ω), . . . , xd (τ, ω))dτ
0
(2.29)
m
t
+ σij ((x1 (τ, ω), . . . , xd (τ, ω))dWτj (ω)
j=1 0
Example 2.4.1 (Binary Biochemical Model). Let (Wt1 , Wt2 ) be a Wiener pro-
cess in R2 . Consider the system of Itô ordinary differential equations
We assume that fi (x) ∈ Cb1,δ (Rd ) and σij (x) ∈ Cb2,δ (Rd ).
Definition 2.4.3 (Solutions to Stratonovich SDE). A continuous se-
mimartingale x(t, ω) on R+ with values in Rd is said to be a solution to the
system (2.31) with initial data x∗ (ω) = (x∗1 (ω), . . . , x∗d (ω)) if it satisfies the
integral equation
t
∗
xi (t, ω) = xi (ω) + fi (x1 (τ, ω), . . . , xd (τ, ω))dτ
0
(2.32)
m
t
+ σij ((x1 (τ, ω), . . . , xd (τ, ω)) ◦ dWτj (ω)
j=1 0
1
d m
∂σkj (x)
ck (x) ≡ · σij (x) · ∈ Cb1,δ (Rd ), k = 1, . . . , d . (2.33)
2 i=1 j=1 ∂xi
Then the system of Stratonovich SDEs (2.31) has a unique solution for every
F0 -measurable initial data x∗ (ω) ∈ L2 (Ω, F, P). Further this solution satisfies
Itô’s SDEs (2.28) with hi (x) = fi (x) + ci (x), where ci (x) is given by (2.33),
i = 1, . . . , d.
Conversely, if x(t, ω) is a solution to Itô SDEs (2.28) with hi (x) ∈
Cb1,δ (Rd ) and σij (x) as above, then x(t, ω) solves the system of Stratonovich
SDEs (2.31) with fi (x) = hi (x) − ci (x), i = 1, . . . , d.
2.4 RDS Generated by Stochastic Differential Equations 73
Example 2.4.2 (Binary Biochemical Model). Let (θ, ϕ) be the RDS generated
by (2.30). If g(x) ∈ Cb1,δ (R), then x(t) = ϕ(t, ω, x) solves the Stratonovich
equations
σ12
dx1 = g(x2 ) − α1 + x1 dt + σ1 x1 ◦ dWt1 ,
2
σ2
dx2 = x1 − α2 + 2 x2 dt + σ2 x2 ◦ dWt2 .
2
(2.34)
m
t
+ Dx σj (ϕ(τ, ω, x))Dx ϕ(τ, ω, x) ◦ dWτj
j=1 0
for all t > 0. Here f = (f1 . . . , fd ) and σj = (σ1j . . . , σdj ) are mappings from
Rd into itself. Moreover for every t > 0 the determinant detDx ϕ(t, ω, x)
satisfies Liouville’s equation
t
detDx ϕ(t, ω, x) = exp tr{Dx f (ϕ(τ, ω, x))}dτ
0
m
t
+ tr{Dx σj (ϕ(τ, ω, x))} ◦ dWτj . (2.35)
j=1 0
are independent (in particular, the past and future σ-algebras are indepen-
dent) and the processes {ϕ(t, ω)x : x ∈ Rd } form a Markov family.
Example 2.4.3 (Binary Biochemical Model). Consider the following Strato-
novich version of equations (2.30)
σ2
dx2 = x1 − α2 − 2 x2 dt + σ2 x2 dWt2 .
2
Remark 2.4.1. It is possible (see Arnold [3, Chap.2]) to prove in the two
cases described in Theorems 2.4.1 and 2.4.3 that the cocycle ϕ(t, ω) can
be extended to a cocycle ϕ̃ with two-sided time. In particular this implies
(see Arnold [3, Theorem 1.1.6]) that the function (t, x) → ϕ(t, θ−t ω)x =
[ϕ̃(−t, ω)]−1 x is continuous for every ω ∈ Ω. Therefore the mapping t →
ϕ(t, θ−t ω)v(θ−t ω) is continuous for a dense (with respect to convergence in
probability) set of random variables v(ω) (cf. Remark 1.5.1).
Theorem 2.4.3 can be applied to the affine (linear nonhomogeneous) Stratono-
vich SDE
m
dx(t) = (A0 x(t) + b0 ) dt + (Aj x(t) + bj ) ◦ Wtj , (2.37)
j=1
where ψ(t, ω) = ϕ(t, ω)0 and Φ(t, ω) is the linear cocycle in Rd generated by
the linear SDE
m
dx(t) = A0 x(t)dt + Aj x(t) ◦ Wtj . (2.39)
j=1
As in the random case (cf. Theorem 2.1.2) from the multiplicative ergodic
theorem (see Arnold [3, Chaps.3,4]) we can derive the following assertion
on the top Lyapunov exponent of the linear RDS (θ, Φ).
2.4 RDS Generated by Stochastic Differential Equations 75
It is clear that the number λ is the (top) Lyapunov exponent for (θ, Φ). If
λ < 0, then the RDS (θ, ϕ) is dissipative in the universe of all tempered
subsets of R. By Proposition 1.9.2 and Remark 1.9.2 the RDS (θ, ϕ) has a
unique equilibrium u(ω). Relations (1.51) and (2.42) imply that
0
u(ω) = β exp {−λτ − σWτ (ω)} dτ .
−∞
dx = λxdt + σdWt ,
Using the relation Wτ (θt ω) = Wt+τ (ω) − Wt (ω), it is easy to see that
d j,ε ε
dt Wt (ω) = ηj (θt ω), where
ε
ηjε (ω) = − φ̇ε (τ )Wτj (ω) dτ .
0
dxi m
= fi (x1 , . . . , xd ) + σij (x1 , . . . , xd ) · ηjε (θt ω), i = 1, . . . , d . (2.44)
dt j=1
If fi (x) ∈ Cb1,δ (Rd ) and σij (x) ∈ Cb2,δ (Rd ), then Corollary 2.1.1 implies that
the random differential equations (2.44) generates an RDS in Rd . In particular
equations (2.44) are uniquely solved on Rd for any initial data x∗ ∈ Rd .
We have the following Wong-Zakaı̈ type approximation theorem (see, e.g.,
Ikeda/Watanabe [57] and Kunita [74]).
Theorem 2.5.1. Let the hypotheses of Theorem 2.4.2 hold. Let a function
x(t, ω; x∗ ) be the solution to the Stratonovich SDE (2.31) with initial data
x∗ ∈ Rd and xε (t, ω; x∗ ) be the solution to the RDE (2.44) with the same
initial data. Then we have
lim E sup sup |x(t; x∗ ) − xε (t; x∗ )|2 =0 (2.45)
ε→0 [0,T ] |x∗ |≤R
Remark 2.5.1. Assume that the hypotheses of Theorem 2.4.2 hold. Let ϕ(t, ω)
and ϕε (t, ω) be the cocycles of the RDS generated by (2.31) and (2.44) re-
spectively. From (2.45) we have
m
lim E dτ sup sup |ϕ(t, θτ ω)x∗ − ϕε (t, θτ ω)x∗ |2 =0
ε→0 −m t∈[0,l] |x∗ |≤r
for all m, l, r ∈ N. This implies that there exists a sequence εn → 0 such that
the set Ω ∗ of all ω ∈ Ω satisfying
β
lim dτ sup sup |ϕ(t, θτ ω)x∗ − ϕεn (t, θτ ω)x∗ |2 =0 (2.46)
n→0 α [0,T ] |x∗ |≤R
for all α < β, T > 0 and R > 0, has full measure, i.e. P(Ω ∗ ) = 1. Moreover
the set Ω ∗ is θ-invariant, i.e. θt Ω ∗ = Ω ∗ for all t ∈ R.
This remark along with Theorems 2.2.1 and 2.5.1 allows us to obtain the fol-
lowing result concerning deterministic invariant domains for RDS generated
by Stratonovich SDE (2.31).
78 2. Generation of Random Dynamical Systems
Corollary 2.5.1. Assume that the hypotheses of Theorem 2.4.2 hold. Let D
be a closed set in Rd such that (i) D has an outer normal at every point of
the boundary ∂D and (ii) for any x ∈ ∂D we have
for almost all t from the interval [0, β]. If x∗ ∈ D, Theorem 2.2.1 implies that
ϕεn(k) (t, θ−t ω)x∗ ∈ D. Therefore it follows from (2.49) that ϕ(t, θ−t ω)x∗ ∈ D
for almost all t from [0, β]. Since the function t → ϕ(t, θ−t ω)x is continuous for
every ω ∈ Ω and x ∈ Rd (see Remark 2.4.1), we have that ϕ(t, θ−t ω)x∗ ∈ D
for all t ∈ [0, β], x∗ ∈ D and ω ∈ Ω ∗ with arbitrary β > 0. Now the invariance
of Ω ∗ implies (2.48). 2
Proof. Let Ω ∗ be the set given by Corollary 2.5.1. If we redefine the cocycle
ϕ of the RDS generated by (2.31) (see Theorem 2.4.3) by the formula
ϕ(t, ω) if ω ∈ Ω ∗ ,
ϕ̃(t, ω) := (2.50)
id if ω ∈
/Ω,
Example 2.5.1 (1D Stochastic Equation). Let f (x) ∈ Cb1,1 (R), σ(x) ∈ Cb2,1 (R)
and σ(x) · σ (x) ∈ Cb1,1 (R). If f (0) ≥ 0 and σ(0) = 0, then the equation
generates an RDS in R+ .
for some µ > 0. For existence and properties of solutions to (2.51) we refer
to Ikeda/Watanabe [57] or McKean [82], for instance. The stationary
solution {zk (t, ω)} can be written in the form
t
zk (t, ω) = e−µ(t−τ ) dWtk (ω) almost surely .
−∞
Proof. The existence of the variable z(ω) with properties (i)-(iii) follows from
Chueshov/Scheutzow [23, Proposition 3.1]. To obtain (iv) we note that
the ergodic theorem for stationary processes (see, e.g., Gihman/Skoro-
hod [48, p.140]) implies that (2.52) holds almost surely. Let Ω ∗ be the set
of all ω ∈ Ω such that (2.52) holds. It is clear that Ω ∗ is a θ-invariant set of
full measure. Therefore we can redefine z(ω) by zero outside of Ω ∗ . 2
∂uj (z, x)
= σji (u(z, x)), u(0, x) = x ∈ Rd , i = 1, . . . , m, j = 1, . . . , d ,
∂zi
(2.54)
and let z(ω) be the random variable given by Lemma 2.5.1. Then u(z(ω), x)
is a tempered random variable in Rd for every x ∈ Rd and the mapping
x → T (ω, x) ≡ u(z(ω), x) is a diffeomorphism of Rd for each ω ∈ Ω. Further,
if (θ, ϕ) is the RDS generated by (2.31), then
Example 2.5.3. Let us assume that m = d and σij (x) = δij · σi (xi ), i.e. we
consider the following system of Stratonovich SDEs
Simple calculation shows that condition (2.53) is satisfied and the equations
in (2.54) have the form
∂uj ∂ui
= 0, if i = j and = σi (ui ); ui (0, x) = xi . (2.56)
∂zi ∂zi
where
fi H1−1 (z1 (ω) + H1 (y1 )), . . . , Hd−1 (zd (ω) + Hd (yd ))
gi (ω, y) = + µzi (ω)
σi (Hi−1 (zi (ω) + Hi (yi )))
for i = 1, . . . , d.
Equations (2.56) can be also solved in the case when σi (u) = σi · u. In
this case we have ui (z, x) = xi exp{σi zi } for i = 1, . . . , d. In particular, this
observation means that the SDE (2.36) which arises in a stochastic binary
biochemical model is conjugate with the RDE
ẏ1 = e1 (θt ω)−1 g(e2 (θt ω)y2 ) − (α1 − µσ1 z1 (θt ω))y1 ,
Other examples of the conjugacy of SDE and RDE can be found in Kel-
ler/Schmalfuss [63], Imkeller/Schmalfuss [59] and in Chap.6 below.
3. Order-Preserving Random Dynamical
Systems
[a, b] = {x ∈ V : a ≤ x ≤ b} .
sublinear RDS. In Sect.4.1 we prove that these RDS are nonexpansive with
respect to p.
Let u ∈ V+ \ {0} and Cu be the part of the cone which contains u. Then it
easy to prove the following relations between the part metric and the u-norm
and
Rd+ = {x = (x1 , . . . , xd ) ∈ Rd : xi ≥ 0, i = 1, . . . , d}
every set
In particular this formula shows that the part metric is not strictly convex,
i.e. for some points a and b from intRd+ the set
Remark 3.1.1. Let V+ be a normal cone. Then it is easy to see that the norm
·
∗ defined by (3.5) is monotone, i.e. the relation 0 ≤ x ≤ y implies that
x
∗ ≤
y
∗ . For any monotone norm
·
we have the inequality
x − y
≤ 2ep(x,y) − e−p(x,y) − 1 · min {
x
,
y
} , x, y ∈ V+ \ {0} ,
where p(x, y) is the part metric and we suppose p(x, y) = ∞ if x and y belong
to different parts (the proof see in Krause/Nussbaum [72]).
In general the monotonicity and boundedness of a sequence does not imply its
convergence. Counterexamples can be easily constructed in the space C[0, 1]
of continuous functions on the interval [0, 1] ⊂ R with the cone of nonnegative
functions. Therefore the following concept is useful in applications.
Definition 3.1.6 (Regular Cone). Let V be a real Banach space. A cone
V+ is said to be regular if every monotone sequence
x1 ≤ x2 ≤ . . . ≤ xn ≤ . . .
of real numbers with the property |ai | < ∞, 1 ≤ p < ∞. One can prove
(see, e.g., Krasnoselskii/Lifshits/Sobolev [71]) that every regular cone
is normal.
3.1 Partially Ordered Banach Spaces 87
for any w and w∗ from V , where C1 and C2 are the operator norms of Φ−1
and Φ. 2
{ω : U ∩ {x : 0 x ≤ b(ω)} = ∅} ∈ F
Proof. Let D(ω) be bounded. Then there exists a random variable r(ω) ∈ R+
such that
Since V+ is solid, there exists u ∈ intV+ such that the interval [−u, u] contains
the unit ball of V . Therefore we have (3.9) with v(ω) = r(ω) · u. If {D(ω)} is
tempered, then r(ω) is a tempered random variable. Therefore v(ω) = r(ω)·u
is a tempered element in V+ .
90 3. Order-Preserving Random Dynamical Systems
Assume that (3.9) is valid with v(ω) ≥ 0. Then for any x ∈ D(ω) we have
0 ≤ x + v(ω) ≤ 2v(ω). Therefore the normality of the cone implies that
Thus we have (3.10) with r(ω) = (1 + 2c)
v(ω)
. Here above c is the constant
from Definition 3.1.5 and r(ω) is tempered if v(ω) is a tempered element in
V+ . 2
are random variables in V . If we additionally assume that a(ω) and b(ω) are
tempered random variables in V , then the same property is true for u(ω) and
v(ω).
Proof. The first part follows easily from Corollary 3.1.1(ii). For the second
part, due to (3.6) and (3.7) it is sufficient to prove that w+ (ω) = w(ω) ∨ 0 is
tempered for any tempered variable w(ω) ∈ V . Using (3.8) with w∗ = 0 we
have
w+ (ω)
≤ C1 C2
w(ω)
. Therefore the temperedness of w(ω) implies
the temperedness of w+ (ω). 2
Theorem 3.2.1. Assume that the cone V+ is a solid normal minihedral cone.
Let {D(ω)} be a random compact set in V . Then sup D(ω) and inf D(ω) are
random variables in V . If we assume additionally that D(ω) is tempered, then
sup D(ω) and inf D(ω) are tempered random elements in V .
Let
wN (ω) = sup{v1 (ω), . . . , vN (ω)}, N = 1, 2, . . . .
Lemma 3.2.1 implies that wN (ω) is a random variable in V for every N =
1, 2, . . .. It is also clear that
where B is the unit ball of V . Since V+ is solid, there exists an interval [−u, u]
which contains the ball B. It follows from (3.13) that for any m ∈ N there
exists nm ≤ Nk such that
vm (ω) ∈ vnm (ω) + 2−k · B ⊂ [vnm (ω) − 2−k u, vnm (ω) + 2−k u]
and therefore
Consequently
Therefore the sequence {wNk } is a Cauchy sequence. This implies the ex-
istence of the limit in (3.12). It is clear that w(ω) = sup D(ω). The tem-
peredness of sup D(ω) and inf D(ω) for tempered D(ω) follows from Propo-
sition 3.2.2. 2
and
λ−1 v(ω) ≤ u(ω) ≤ λv(ω), ω∈Ω.
In particular u(ω) ∈ Cv . Since {um (ω)} is a Cauchy sequence in Cv , for any
ε > 0 there exists Nε such that
Below we also need the following property of the part metric p(u, v).
3.3 Definition of Order-Preserving RDS 93
Proposition 3.2.4. Let u(ω) and v(ω) be random variables in V+ such that
p(u(ω), v(ω)) exists for every ω ∈ Ω. Then the function ω → p(u(ω), v(ω)) is
a random variable.
Proof. From Definition 3.1.4 we have
Ac : = {ω : p(u(ω), w(ω) < c} = {ω : e−c u(ω) < v(ω) < ec u(ω)}
the linear operator Φ(t, ω) maps V+ into itself for any t ≥ 0 and ω ∈ Ω. We
obtain particular cases of affine order-preserving RDS if in Examples 3.3.1 –
3.3.4 we additionally assume that f (α, x), f (ω, x), g(x, h(x) and σ(x) are lin-
ear functions with respect to x. We study properties of affine order-preserving
RDS in detail in Sect. 4.6.
ϕ(t, ω)u(ω) ≥ u(θt ω) for all t≥0 and all ω∈Ω; (3.18)
(ii) a super-equilibrium if
ϕ(t, ω)u(ω) ≤ u(θt ω) for all t≥0 and all ω∈Ω. (3.19)
It is clear that any equilibrium (see Definition 1.7.1) is both a sub- and
super-equilibrium. Below we will refer to sub- and super-equilibria as semi-
equilibria.
We note that inequality (3.18) is equivalent to
with λ ∈ (0, 1). This mapping is a contraction along the line l = {(s, −s) :
s ∈ R} and it is order-preserving with respect to R2+ with equilibria a =
(−1, −1) and b = (1, 1). A simple calculation shows that any element (α, −α)
with α ∈ (λ, 1] belongs to [a, b] and it does not belong to Tλ [a, b].
u(θt+h ω) − u(θt ω)
D+ u(θt ω) := lim sup ≤ f (θt ω, u(θt ω)), t ∈ R, ω ∈ Ω .
h→+0 h
Thus the sub-equilibrium u(ω) is a random variable such that the stationary
process u(t) := u(θt ω) solves (in some sense) the differential inequality
Similarly, if u(ω) is a super-equilibrium for (θ, ϕ), then the process u(t) =
u(θt ω) solves the inequality
Moreover using the comparison principle (cf. Theorem 5.3.1 below) one can
prove that a random variable u(ω) is a semi-equilibrium for (θ, ϕ) if and only
if the process u(t) = u(θt ω) solves one of these differential inequalities. In
particular a number c is a sub-equilibrium (resp. super-equilibrium) if and
only if f (ω, c) ≥ 0 (resp. f (ω, c) ≤ 0) for all ω ∈ Ω. Similar results remain
true for order-preserving RDS generated by systems of random or stochastic
differential equations.
Therefore there exists a version of the cocycle ϕ (see Remark 1.2.1(ii)) such
that ϕ(t, ω)a ≥ a for all ω ∈ Ω and t ≥ 0. Thus a(ω) ≡ a is a sub-equilibrium
for (θ, ϕ). The same argument gives that b(ω) ≡ b is a super-equilibrium
provided b ∈ R satisfies the conditions h(b) ≤ 0 and σ(b) = 0.
ϕ(t, ω)as (ω) = ϕ(t, ω)ϕ(s, θ−s ω)a(θ−s ω) = ϕ(t + s, θ−s ω)a(θ−s ω)
(3.25)
= ϕ(s, θt−s ω)ϕ(t, θ−s ω)a(θ−s ω) .
(as )σ (ω) = ϕ(σ, θ−σ ω)as (θ−σ ω) ≥ as (ω) for any s, σ ≥ 0 . (3.27)
(as )σ (ω) = ϕ(σ, θ−σ ω)as (θ−σ ω) = ϕ(σ, θ−σ ω)ϕ(s, θ−s−σ ω)a(θ−s−σ ω) .
Similarly, if u1 (ω) and u2 (ω) are super-equilibria and inf{u1 (ω), u2 (ω)} ∈ X
for all ω ∈ Ω, then inf{u1 (ω), u2 (ω)} is a super-equilibrium.
3.4 Sub-Equilibria and Super-Equilibria 99
Therefore ϕ(t, ω)(sup{u1 (ω), u2 (ω)}) ≥ sup{u1 (θt ω), u2 (θt ω)}. The same ar-
gument applies for super-equilibria. 2
implies that
ϕ(t, ω)a(ω) ≤ ϕ(t, ω)w(ω) ≤ ϕ(t, ω)b(ω) for all w(ω) ∈ A(ω) .
ϕ(t, θ−t ω)a(θ−t ω) ≤ w(ω) ≤ ϕ(t, θ−t ω)b(θ−t ω) for all w(ω) ∈ A(ω) .
Since a(ω) = inf A(ω) and b(ω) = sup A(ω), the last relation implies that
Remark 3.4.2. (i) If the cone V+ is normal, solid and miniedral and if for an
RDS (θ, ϕ) on X = V there exists a random element x ∈ V such that the
closure γxτ (ω) of the orbit
γxτ (ω) = ϕ(t, θ−t ω)x(θ−t ω)
t≥τ
emanating from ϕ(τ, θ−τ ω)x(θ−τ ω) is a random compact set for some τ > 0,
then there exist a sub-equilibrium b(ω) and a super-equilibrium a(ω) for (θ, ϕ)
such that a(ω) ≤ b(ω). In fact the compactness of γxτ (ω) implies (see Sect.1.6)
that the corresponding omega-limit set
Γx (ω) = ϕ(τ, θ−τ ω)x(θ−τ ω)
t>0 τ ≥t
100 3. Order-Preserving Random Dynamical Systems
It is clear that the set A = {(1, 0), (0, 1)} is invariant and sup A = (1, 1) is a
strict sub-equilibrium.
3.5 Equilibria
and
ū(ω) = lim bs (ω) = inf bs (ω) (3.30)
s→+∞ s>0
exist, where as (ω) and bs (ω) are defined as in (3.23). These limits are equi-
libria of (θ, ϕ) such that
Proof. Let us consider the discrete (T = Z) RDS (θ̂, ϕ̂), where θ̂n = θnt0
and ϕ̂(n, ω) = ϕ(nt0 , ω). It is clear that a(ω) and b(ω) are sub- and super-
equilibria of ϕ̂. Let ân (ω) and b̂n (ω) be defined for (θ̂, ϕ̂) as in (3.23). From
Proposition 3.4.1 one can see that
a(ω) ≤ ân (ω) ≤ ân (ω) ≤ b̂m (ω) ≤ b̂m (ω) ≤ b(ω) (3.32)
3.5 Equilibria 101
ân+1 (ω) = (ân )1 (ω) = ϕ̂(1, θ̂−1 ω)ân (θ̂−1 ω) = ϕ(t0 , θ−t0 ω)ant0 (θ−t0 ω) .
Consequently ân+1 (ω) ∈ ϕ(t0 , θ−t0 ω)[a, b](θ−t0 ω) for any n ∈ Z+ . Thus the
sequence {ân (ω)} is relatively compact for every ω. The monotonicity prop-
erty (3.32) implies that this sequence has a unique limit point. Indeed, let us
assume that there exist two points v(ω) and w(ω) such that
and
w(ω) = lim âmk (ω), âmk (ω) ≤ w(ω), k = 1, 2, . . .
k→∞
for some sequences {nk } and {mk }. However for any k there exists l such
that nk ≤ ml and, therefore, ânk (ω) ≤ âml (ω) ≤ w(ω). Hence
In the same way we have w(ω) ≤ v(ω). Hence v(ω) = w(ω). Thus the limit
exists. Since for any s ∈ T+ there exists n ∈ Z+ such that (n − 1)t0 < s ≤ nt0
we have ân (ω) ≤ as (ω) ≤ ân+1 (ω). Therefore (3.33) implies that the element
u(ω) possesses property (3.29).
It remains to prove that u(ω) is an equilibrium. The continuity of the
cocycle ϕ and the structure of as (ω) imply that
This relation means that u(ω) is an equilibrium. In the same way one can
prove the existence of an equilibrium ū(ω) possessing property (3.30). The
inequalities (3.32) imply (3.31). 2
Remark 3.5.1. For regular cones (see Definition 3.1.6) Theorem 3.5.1 is
valid without the assumption concerning the relative compactness of the set
φ(t0 , ω)[a, b](ω).
102 3. Order-Preserving Random Dynamical Systems
Corollary 3.5.1. Assume that the hypotheses of Theorem 3.5.1 hold. If a(ω)
is measurable with respect to the past σ-algebra F− (see the definition in
Sect.1.10), then the equilibrium u(ω) is also F− -measurable and the random
Dirac measure δu(ω) is a disintegration of the invariant Markov measure µ
on (Ω × X, F × B(X)) which has the form
Proof. Since
F− = σ{ω → ϕ(τ, θ−t ω) : 0 ≤ τ ≤ t} ,
the mapping ω → ϕ(s, θ−s ω)x is F− -measurable for any x ∈ X. If a(ω) is
F− -measurable, then a(θ−s ω) is also F− -measurable for s ≥ 0. Therefore
as (ω) = ϕ(s, θ−s ω)a(θ−s ω) possesses the same property and (3.29) implies
that the equilibrium u(ω) is F− -measurable. It is clear that the measure
defined by (3.34) is invariant for the RDS (θ, ϕ) (see Example 1.10.1). Its
disintegration µω has the form µω (B) = χB (u(ω)), where χB (x) = 1 for
x ∈ B and χB (x) = 0 if x ∈ B. Therefore ω → µω (B) is F− -measurable for
any B ∈ B. Thus µ is a Markov measure by Definition 1.10.2. 2
a ≤ u(ω) ≤ u(ω) ≤ b .
By Corollary 3.5.1 these equilibria are F− -measurable and therefore they gen-
erate invariant Markov measures for the RDS (θ, ϕ) connected with equation
(3.16).
Example 3.5.2 (Equilibria for Binary Biochemical Model). Consider the sit-
uation described in Example 3.4.2. Theorem 3.5.1 and Corollary 3.5.1 im-
ply that the RDS (θ, ϕ) generated in R2+ by equations (2.12) possesses F− -
measurable equilibria u(ω) and u(ω) such that
Proof. The main idea of the proof is due to Hans Crauel (see Arnold [3,
Sect.1.8], where the second assertion of this proposition is proved for one-
dimensional RDS with continuous two-sided time). For the sake of simplicity
we consider the case X = R only. The proofs for other cases are similar.
We start with some preliminary observations. Let ν be a probability mea-
sure on R and
the values α(ω) and β(ω) are random variables. Since ϕ is order-preserving,
we have
(−∞, x] ⊆ ϕ(t, ω)−1 (−∞, ϕ(t, ω)x] .
Thus by the invariance of µω , we obtain the relation
µω {(−∞, x]} ≤ µω {ϕ(t, ω)−1 (−∞, ϕ(t, ω)x]} = µθt ω {(−∞, ϕ(t, ω)x]} .
Similarly,
µω {[x, ∞)} ≤ µθt ω {[ϕ(t, ω)x, ∞)} .
These properties imply that α(ω) is a sub-equilibrium and β(ω) is a super-
equilibrium for (θ, ϕ) such that α(ω) ≤ β(ω). Therefore we can apply The-
orem 3.5.1 to conclude that there exist at least one equilibrium u(ω) ∈
[α(ω), β(ω)] with properties (3.35).
The random semiinterval I u (ω) := (−∞, u(ω)] is a forward invariant
random closed set (see Remark 3.4.1), i.e. I u (ω) ⊂ ϕ(t, ω)−1 I u (θt ω). If for
some ω ∈ Ω there exists x ∈ ϕ(t, ω)−1 I u (θt ω) such that x ∈ I u (ω), then
ϕ(t, ω)x ≤ u(θt ω) and x > u(ω) which is impossible if ϕ is strictly order-
preserving. Thus I u (ω) = ϕ(t, ω)−1 I u (θt ω). This relation implies that the
set
M− = {(ω, x) : x ≤ u(ω)} ∈ F × B(R)
satisfies πt−1 M− = M− , where πt is the skew-product semiflow corresponding
to (θ, ϕ) (see (1.6)). We also have
µ(M− ) = µω {I u (ω)}P(dω) ≥ 1/2 .
Ω
The following proposition shows that the pull back omega-limit set (see Def-
inition 1.6.1) emanating from a semi-equilibrium consists of a single equilib-
rium.
Proposition 3.5.2. Assume c is either a sub- or a super-equilibrium and
t (ω)
for any ω ∈ Ω there exists t0 = t0 (ω) such that the closure γc0 (ω) of the
tail of the orbit γc0 (ω) emanating from c,
3.6 Properties of Invariant Sets of Order-Preserving RDS 105
γct0 (ω) (ω) = ϕ(t, θ−t ω)c(θ−t ω) ,
t≥t0 (ω)
Proof. Using Proposition 3.4.1 we obviously have that for any ω ∈ Ω the
sequence cn (ω) = ϕ(n, θ−n ω)c(θ−n ω) is monotone and relatively compact for
each ω. Therefore we can repeat the argument given in the proof of Theo-
rem 3.5.1. 2
In this section we prove a theorem on the structure of the random pull back
attractor (see Definition 1.8.1) of an order-preserving RDS. We obtain this
result as a corollary of the following general assertion which can also be useful
to prove the existence of equilibria. We consider an order-preserving RDS
(θ, ϕ) on a subset X of a real separable Banach space V with a closed convex
106 3. Order-Preserving Random Dynamical Systems
then there exists an equilibrium u(ω) ∈ A(ω) such that u(ω) = sup A(ω)
and
(ii) if there exists a random variable v(ω) satisfying (3.36) and such that
v(ω) ≤ A(ω) for all ω ∈ Ω , then there exists an equilibrium u(ω) ∈ A(ω)
such that u(ω) = inf A(ω) and
Proof. We prove assertion (i) only. Since A(ω) ≤ v(ω), the invariance prop-
erty of A(ω) implies that
The compactness of A(ω) and property (3.36) imply that for any ω ∈ Ω there
exist a sequence tn = tn (ω) → ∞ and an element u(ω) ∈ A(ω) such that
From (3.39) we have that a(ω) ≤ u(ω) for any a(ω) ∈ A(ω), i.e. u(ω) is the
least upper bound for A(ω). Let us prove (3.37). Assume that this relation
is not true for some ω ∈ Ω. Then there exists a sequence τk → ∞ such that
with some positive δ. As above the compactness of A(ω) and property (3.36)
allow us to extract a subsequence {τkm } and to find an element u∗ ∈ A(ω)
such that
implies that u(ω) is a random variable in X. Since u(ω) = sup A(ω), the
invariance of A(ω) implies
Therefore
The main corollary of Theorem 3.6.1 is the following result concerning the
structure of the global attractor for an order-preserving RDS.
Theorem 3.6.2. Assume that the order-preserving RDS (θ, ϕ) on X has
a random compact pull back attractor A(ω) in some universe D and that
this attractor is order-bounded in the following sense: there exists a random
interval
[b, c](ω) = {x : b(ω) ≤ x ≤ c(ω)}
such that {b(ω)}, {c(ω)} ∈ D and A(ω) ⊂ [b, c](ω). Then there exist two
equilibria u(ω) and ū(ω) in A(ω) such that u(ω) ≤ ū(ω) and the random
attractor A(ω) belongs to the interval [u, ū](ω), i.e.
These equilibria u(ω) and ū(ω) are globally asymptotically stable from below
and from above respectively, i.e.
and
lim ϕ(t, θ−t ω)v(θ−t ω) = ū(ω) (3.43)
t→+∞
for any w(ω) ≤ u(ω) and for any v(ω) ≥ ū(ω) such that {w(ω)} and {v(ω)}
belong to D.
Proof. The application of Theorem 3.6.1 gives us the existence of the equi-
libria u(ω) and ū(ω). To prove (3.42) we note that
for any w(ω) ≤ u(ω) such that {w(ω)} ∈ D. Now the compactness of A(ω)
and an argument similar to that used in the proof of Theorem 3.6.1 give
(3.42). The same argument can be applied to prove (3.43). 2
108 3. Order-Preserving Random Dynamical Systems
This remark allow us to derive from Theorem 3.6.2 the following corollary.
Corollary 3.6.1. Let the order-preserving RDS (θ, ϕ) be asymptotically com-
pact in some universe D. Assume that D contains an absorbing interval for
this RDS. If (θ, ϕ) has a unique equilibrium point u(ω) in D, then {u(ω)} is
a random attractor for this RDS in D.
In the connection with Remark 3.6.1 and Corollary 3.6.1 it is convenient to
introduce the concept of an absorbing semi-equilibrium.
Definition 3.6.1. A super-equilibrium u(ω) is said to be absorbing in the
universe D if for any B ∈ D there exists tB (ω) > 0 such that
Proof. From Proposition 3.4.1 we have that w(ω) ≤ ϕ(t, θ−t ω)w(θ−t ω) for all
t > 0. Therefore (3.44) implies that w(ω) ≤ u(ω) and the interval [w(ω), u(ω)]
is absorbing by Definition 3.6.1. It is forward invariant by Remark 3.4.1. 2
All the above considerations give ample proof of the crucial role played by sub-
and super-equilibria in the study of qualitative properties of order-preserving
RDS. One of the methods of proving their existence relies on the comparison
principle. Let V be a Banach space with a cone V+ and let X1 and X2 be
subsets of V . Let (θ, ϕ1 ) and (θ, ϕ2 ) be two RDS on X1 and X2 over the same
metric dynamical system θ and take Y ⊂ X1 ∩ X2 . The system (θ, ϕ2 ) is said
to dominate (θ, ϕ1 ) from above on Y (or (θ, ϕ1 ) dominates (θ, ϕ2 ) from below
on Y ) if
Clearly (3.46) implies that any super-equilibrium v(ω) for (θ, ϕ2 ) such that
v(ω) ∈ Y for all ω ∈ Ω is a super-equilibrium for (θ, ϕ1 ) and any sub-
equilibrium w(ω) for (θ, ϕ1 ) with the property w(ω) ∈ Y for all ω ∈ Ω is a
sub-equilibrium for (θ, ϕ2 ).
In many applications, e.g. for the construction of random attractors (see
Chaps.5 and 6 below), a nonlinear RDS can be shown to be dominated by
an affine RDS (see Definition 1.2.3), whose equilibrium then becomes a sub-
or super-equilibrium of the corresponding nonlinear RDS.
As an example of an application of the comparison principle we prove the
following assertion.
110 3. Order-Preserving Random Dynamical Systems
Since ϕ(t, ω)0 ≥ 0 and ϕ(t, ω)x ≤ ϕaff (t, ω)x for all x ∈ V+ , we have
and 0
0
u2 (ω) = u1 (θs ω) exp − α2 (θτ ω)dτ ds .
−∞ s
A simple calculation shows that u(ω) 0 is an equilibrium for (θ, ϕaff ). Thus
u(ω) is a super-equilibrium for (θ, ϕ). Proposition 3.7.1 can be applied here
with the universe D consisting of all tempered sets from R2 .
4. Sublinear Random Dynamical Systems
X = V+ = {x ∈ V : x ≥ 0} ,
i.e. we consider random dynamical systems (θ, ϕ) which possess the positivity
property: ϕ(t, ω)V+ ⊂ V+ for all t and ω. For order-preserving systems this
property is equivalent to the relation ϕ(t, ω)0 ≥ 0. Our main result in this
chapter is a random limit set trichotomy which describes all possible types
of long-time behaviour in sublinear random systems. This result is a clear
manifestation of the general experience that monotonicity, and even more
so sublinearity, drastically simplifies the possible long-term behaviour of a
dynamical system.
λϕ(t, ω, x) < ϕ(t, ω, λx) for all t>0 and ω∈Ω, (4.2)
Example 4.1.1 (Binary Biochemical Model). Let (θ, ϕ) be the RDS in R2+
generated by the equations
over a metric dynamical system θ. We assume that the function g(x) and the
random variables α1 (ω) and α2 (ω) satisfy the assumptions of Examples 2.1.1
and 3.3.5. In Sect.5.7 we prove that (θ, ϕ) is a strictly sublinear RDS if g(x)
is a sublinear mapping from R+ into R, i.e. if λg(x) ≤ g(λx) for all x ≥ 0 and
0 < λ < 1. This system is strongly sublinear if we assume additionally that
g (x) > 0 for x > 0. Similar result remains true for the stochastic case (cf.
Example 2.4.3 and Sect.6.8). We note that sublinear functions g(x) appear
in the Griffith (g(x) = x · (1 + x)−1 for x > 0) and in the Othmer-Tyson
(g(x) = (1 + x) · (k + x)−1 , x > 0, k > 1) biochemical models. We refer to
Selgrade [96] for a discussion and for the references.
for all t > 0 and ω ∈ Ω. The RDS is said to be strictly concave if in addition
we have strict inequality in (4.5) for all 0 x y and it is strongly concave
if
Proof. Since
d 1 1
· ϕ(t, ω, λx) = − 2 {ϕ(t, ω, λx) − Dx ϕ(t, ω, λx)λx} ,
dλ λ λ
we have
ν
1 1 1
ϕ(t, ω, νx) − · ϕ(t, ω, λx) = − {ϕ(t, ω, µx) − Dx ϕ(t, ω, µx)µx} dµ ,
ν λ λ µ2
for any x and z from V+ \ {0}. Therefore for any λ2 > λ1 > 0 we have
λ1 λ1
ϕ(t, ω, x + λ1 z) − ϕ(t, ω, x + λ2 z) − 1 − ϕ(t, ω, x)
λ2 λ2
λ2
λ
1
= λ1 2
[Dx ϕ(t, ω, x + µz)z − Dx ϕ(t, ω, x + λz)z] dµ dλ. (4.8)
λ1 λ 0
Consequently the (strict, strong) concavity in the differential form (4.7) im-
plies (strict, strong) concavity in the sense of Definition 4.1.2. It is clear from
(4.5) that
from R+ into itself shows that strict (strong) sublinearity does not imply
strict (strong) inequality in (4.6).
Below we also use the following concept of concavity for a C 1 -smooth RDS
which was introduced by Smith [101] in the deterministic case.
Definition 4.1.3 (S-Concave RDS). A C 1 -smooth order-preserving RDS
(θ, ϕ) on X = V+ is said to be s-concave if for any 0 x y and z ∈ intV+
we have
Dx ϕ(t, ω, y)z < Dx ϕ(t, ω, x)z for all t > 0, and ω ∈ Ω . (4.10)
It is clear from Proposition 4.1.1 that any s-concave RDS is strictly concave.
is called a part of V+ and every part C is a metric space with respect to the
part (Birkhoff) metric defined by
p(ϕ(t, ω)x, ϕ(t, ω)y) ≤ p(x, y) for all t≥0 and ω∈Ω.
p(ϕ(t, ω)x, ϕ(t, ω)y) < p(x, y) for all t > 0, ω∈Ω, (4.13)
Proof. (i) It follows from (4.1) and (4.4) that if for x, y ∈ V+ and some α ≥ 1
α−1 x ≤ y ≤ αx
then also
α−1 ϕ(t, ω)x ≤ ϕ(t, ω)y ≤ αϕ(t, ω)x for all t ≥ 0 and ω ∈ Ω
and hence by (4.12) we have p(ϕ(t, ω)x, ϕ(t, ω)y) ≤ p(x, y) for all t ≥ 0 and
ω ∈ Ω, proving (i).
(ii) Assume that x, y ∈ intV+ and there is no λ > 0 such that y = λx. In
this case p(x, y) > 0 and
e−p(x,y) ϕ(t, ω)x ϕ(t, ω)y ep(x,y) ϕ(t, ω)x for t > 0 and ω ∈ Ω .
It is clear that for every t > 0 and ω ∈ Ω there exists µ := µ(t, ω, x, y) > 0
such that
Therefore
p(ϕ(t, ω)x, ϕ(t, ω)y) ≤ p(x, y) − µ < p(x, y) .
118 4. Sublinear Random Dynamical Systems
Thus we obtain (4.13) for these x and y. If y = λx for some λ > 1, then
p(x, y) = log λ and
ϕ(t, ω)x ≤ ϕ(t, ω)y λϕ(t, ω)x for t > 0 and ω ∈ Ω . (4.14)
As above this implies (4.13). The case y = λx with 0 < λ < 1 is similar. Thus
a strongly sublinear RDS possesses property (4.13). It is also clear from (4.14)
that ϕ(t, ω)intV+ ⊂ intV+ .
If (4.13) holds for some order-preserving RDS, then for any x ∈ intV+
and 0 < λ < 1 we have
1
p(ϕ(t, ω)[λx], ϕ(t, ω)x) ≤ log −µ
λ
with some positive µ. Hence eµ λϕ(t, ω)x ≤ ϕ(t, ω)[λx]. This property and
the invariance of intV+ imply (4.3). 2
Remark 4.2.1. (i) Theorem 4.2.1 is wrong without the assumption of strong
sublinearity. Consider for example ẋ = a(θt ω)x on X = R+ , where a(θt ω) =
db(θt ω)/dt is the derivative of a stationary process t → b(θt ω) with absolutely
continuous trajectories. Then the sublinear (but not strongly or strictly sub-
linear) solution is
ϕ(t, ω)x = xe−b(ω) eb(θt ω) ,
meaning that ϕ is a coboundary, i.e. is a cocycle which is cohomologous to
the trivial cocycle ψ(t, ω) ≡ 1 (Arnold [3, Chap.5]), and any x(ω) = ceb(ω) ,
4.2 Equilibria and Semi-Equilibria for Sublinear RDS 119
The idea of the proof is the same as for Theorem 4.2.1 and Proposition 1.7.1.
We next make a statement about the global asymptotic stability of an
equilibrium u in its own part Cu .
Theorem 4.2.2. Let (θ, ϕ) be a strongly sublinear order-preserving RDS in
V+ . Assume that it has an equilibrium u : Ω → intV+ . Suppose that there
exists a constant µ0 ≥ 1 such that for all µ > µ0 the orbits emanating from
µu and µ−1 u are relatively compact in V+ . Then u is globally asymptotically
stable in Cu , i.e. there exists a θ-invariant set Ω ∗ ∈ F of full measure such
that for any w ∈ Cu
By Theorem 3.5.1
Therefore
ϕ(t, θ−t ω)µ−1 u(θ−t ω) ≤ ϕ(t, θ−t ω)w(θ−t ω) ≤ ϕ(t, θ−t ω)µu(θ−t ω) .
λϕ(t, θ−t ω)a(θ−t ω) ≤ ϕ(t, θ−t ω)λa(θ−t ω) for all t > 0 and ω ∈ Ω ,
hence
λu1 (ω) ≤ uλ (ω) ≤ u1 (ω) for all 0<λ≤1.
This means that uλ ∈ Cu1 for all λ ∈ (0, 1], thus by Proposition 4.2.1,
uλ (ω) = u1 (ω) on a θ-invariant set Ω ∗ of full measure. As in the proof
of Theorem 4.2.2 we can choose Ω ∗ independent of λ. By Theorem 3.5.1,
u1 (ω) ≥ a(ω).
For all w satisfying (4.21) with u(ω) := u1 (ω) clearly
ϕ(t, θ−t ω)α−1 a(ω) ≤ ϕ(t, θ−t ω)w(ω) ≤ ϕ(t, θ−t ω)u1 (ω) = u1 (ω) .
As for the first statement of Theorem 4.2.3, the regularity of the cone V+ and
the compactness of the trajectory γa imply (4.20) for all w(ω) satisfying
Proof. From (4.7) we have Dx ϕ(t, ω, z)z ≤ Dx ϕ(t, ω, 0)z for any z > 0.
Therefore from 1
ϕ(t, ω)x = Dx ϕ(t, ω, sx)x ds
0
we get that
Therefore (4.23) follows from Definition 1.9.1 of the top Lyapunov exponent.
2
The following assertion shows that we can choose the set Ω ∗ in (4.24) to be
θ-invariant.
Proposition 4.3.1. If u(ω) ≥ 0 is an almost equilibrium of an RDS (θ, ϕ),
then there exists a θ-invariant set Ω ∗∗ ∈ F̄P of full measure such that (4.24)
holds.
Proof. Let
Since Ω ∗ ⊆ Ω̃ and P(Ω ∗ ) = 1, we have that θs Ω̃ ∈ F̄P and P̄(θs Ω̃) = 1 for
every fixed s ∈ R. Here P̄ is the extension of P on F̄P . Using the cocycle
property we get
for all t, s ≥ 0 and ω ∈ Ω̃. Hence θs Ω̃ ⊂ Ω̃ for all s ≥ 0. Let Ω ∗∗ := ∩n≥0 θn Ω̃.
It is clear that θt Ω ∗∗ ⊂ Ω ∗∗ for t ≥ 0, P̄(Ω ∗∗ ) = 1 and (4.24) holds for
ω ∈ Ω ∗∗ . Let k − 1 ≤ t < k for k ∈ N. Then
θ−t Ω ∗∗ ⊂ θn−t Ω̃ = θn−k θk−t Ω̃ ⊂ θm θk−t Ω̃
n≥0 n≥0 m≥0
Remark 4.3.1. If (θ, ϕ) is an RDS with discrete time, then in the proof of
Proposition 4.3.1 we have Ω̃ ∈ F and therefore Ω ∗∗ ∈ F. Under this condition
it is possible (cf. Remark 1.2.1(ii)) to find a version ϕ̃ of the cocycle ϕ such
that u(ω) is an equilibrium for (θ, ϕ̃). We also refer to Scheutzow [90],
where the perfection problem of crudely invariant elements is discussed for
invertible cocycles.
For the main result of this section we need the following definitions.
Definition 4.3.2. Let U ∈ F. The orbit γa (ω) = ∪t≥0 ϕ(t, θ−t ω)a(θ−t ω) of
the RDS (θ, ϕ) in X = V+ emanating from a is said to be bounded on U if
there exists a random variable C on U such that
for any random variable v(ω) possessing the property αa(ω) ≤ v(ω) ≤ λu(ω)
for all ω ∈ Ω ∗ and for some nonrandom positive α and λ.
Proof. From Proposition 3.4.1 we get that {as (ω) := ϕ(s, θ−s ω)a(θ−s ω), s >
0} is a monotone family of sub-equilibria. Since the cone V+ is normal, there
exists an equivalent norm
·
∗ on V such that s →
as (ω)
∗ is a monotone
nondecreasing function for every ω ∈ Ω (see Remark 3.1.1) and therefore the
limit lims→∞
as (ω)
∗ exists (finite or infinite). Thus if (i) is not true, then
124 4. Sublinear Random Dynamical Systems
there exists v(ω) ∈ V+ such that v(ω) ≥ αa(ω) for some 0 < α < 1 and
ϕ(t, θ−t ω)v(θ−t ω)
∗ → ∞ for ω ∈ U , where U ∈ F and P(U ) > 0. Therefore
for any ω ∈ U there exists a sequence {tn (ω)} such that tn (ω) → ∞ as
n → ∞ and
Since (θ, ϕ) is sublinear, we have supn
atn (ω) (ω)
∗ < ∞ for ω ∈ U . This
implies that
sup
as (ω)
∗ < ∞, ω ∈ U ,
s≥0
because for any ω ∈ U and s > 0 there exists tn (ω) such that
as (ω)
∗ ≤
atn (ω) (ω)
∗ . Consider the set
Proof. Proposition 3.4.1 implies that aε (ω) := ϕ(ε, θ−ε ω)0 0 is a sub-
equilibrium for every ε > 0. It is also clear that ϕ(t, θ−t ω)x ≥ aε (ω) for all
x ∈ V+ , ω ∈ Ω and t ≥ ε. Thus we can apply Theorem 4.3.1. 2
Remark 4.3.2. We note that the uniqueness results stated in Theorem 4.2.1
and Propositions 1.7.1 and 4.2.1 remain true for almost equilibria because
the proof of Proposition 1.7.1 invokes only monotonicity arguments for scalar
measurable functions and properties of probability distributions.
In this section we prove the limit set trichotomy theorem which describes the
only three possible asymptotic scenarios for sublinear systems. We do not
assume the existence of a strongly positive sub-equilibrium here.
In the deterministic discrete time case a limit set trichotomy was discov-
ered (and so named) by Krause/Ranft [73] and generalized by Krause/
Nussbaum [72].
Below we say that a multifunction {F (ω)} belongs to the part Cv gener-
ated by a random variable v(ω) (see Definition 3.2.1) if there exists a non-
random number λ > 1 such that F (ω) ⊂ [λ−1 v(ω), λv(ω)] for all ω ∈ Ω.
Theorem 4.4.1 (Limit Set Trichotomy). Let V be a separable Banach
space with a normal solid minihedral cone V+ . Assume that (θ, ϕ) is a strongly
sublinear conditionally compact order-preserving RDS in V+ . Let v : Ω →
126 4. Sublinear Random Dynamical Systems
Then Cv is a forward invariant set, i.e. (4.27) holds for any a ∈ Cv , and
precisely one of the following three cases applies:
(i) for all b ∈ Cv , the orbit γb emanating from b is unbounded;
(ii) for all b ∈ Cv , the orbit γb emanating from b is bounded, but the
closure of γb does not belong to Cv ;
(iii) there exists a unique almost equilibrium u ∈ Cv measurable with
respect to the universal σ-algebra Fu , and for all b ∈ Cv the orbit emanating
from b converges to u, i.e.
lim ϕ(t, θ−t ω)b(θ−t ω) = u(ω) for almost all ω∈Ω. (4.28)
t→+∞
(ii) Assume that there exists a ∈ Cv for which (4.27) holds. Then for
any b ∈ Cv the orbit emanating from b does not leave Cv , i.e. Cv is forward
invariant.
λ−1 ϕ(t0 , ω)a(ω) ≤ ϕ(t0 , ω)b(ω) ≤ λϕ(t0 , ω)a(ω) for all ω ∈ Ω . (4.32)
Proof. Let b(ω) ∈ γb (ω). Then there exists a sequence {tn (ω)} such that
λ−1 ϕ(tn , θ−tn ω)a(θ−tn ω) ≤ ϕ(tn , θ−tn ω)b(θ−tn ω) ≤ λϕ(tn , θ−tn ω)a(θ−tn ω) .
(4.33)
Since γa (·) ⊂ Cv , there exists µ > 1 such that
(cf. Remark 1.6.1). Since Γa ⊂ Cv , there exists a number α > 1 such that
implying
Proposition 3.2.4 implies that ft (ω) := p(w(θt ω), w(θt ω)) is an Fu -measurable
variable for any fixed t ∈ R+ . Let us prove that ft has the same distribution
for each t ∈ R+ . Let Uct = {ω : ft (ω) ≤ c}. Since Uct ∈ Fu , there exists
Ũct ∈ F such that Ũct ⊆ Uct and P(Ũct ) = P̄(Uct ), where P̄ is the extension of P
to Fu . It is clear that θ−t Ũc0 ⊂ Uct . Therefore
In a similar way the relation θt Ũct ⊂ Uc0 implies P̄(Uc0 ) ≥ P̄(Uct ). Thus all
variables ft have the same distribution.
Suppose now that w(ω) = w(ω) is not true on a set of positive probability,
i.e. there exist a measurable set U ⊂ Ω with P(U ) > 0 such that
p(ϕ(t, ω)w(ω), ϕ(t, ω)w(ω)) < p(w(ω), w(ω)) for ω ∈ U and t > 0 ,
hence
p(w(θt ω), w(θt ω)) < p(w(ω), w(ω)) for ω ∈ U and t > 0 .
However, both sides of the last inequality have the same distribution, leading,
as in the proof of Proposition 1.7.1, to a contradiction of the assumption
P(U ) > 0. Thus w(ω) = w(ω) almost surely, and (4.36) implies that u(ω) ≡
w(ω) is an almost equilibrium. Moreover Γa = {u(ω)} almost surely. It finally
follows from (4.35) that u ∈ Cv . Proposition 4.2.1 and Remark 4.3.2 imply
that this equilibrium is unique almost surely in Cv . In particular Γb = {u(ω)}
almost surely for any b ∈ Cv which gives the asymptotic stability (4.28). This
completes the proof of Theorem 4.4.1. 2
Remark 4.4.1. (i) It is clear from the proof that if the cases (i) and (ii) of
Theorem 4.4.1 do not apply and if there exists an element a ∈ Cv such that
γa (ω) is a random compact set with respect to F, then case (iii) holds with
the equilibrium measurable with respect to F.
(ii) For a discrete RDS (T = Z) the equilibrium given by Theorem 4.4.1 in
case (iii) is measurable with respect to F because the closure of any trajectory
is F-measurable (see Sect. 1.5).
(iii) Theorem 4.4.1 is wrong without the assumption of strong sublinearity,
see Remark 4.2.1(i).
By slightly strengthening hypothesis (4.27) we can also prove another version
of the trichotomy theorem.
Corollary 4.4.1. Assume that the assumptions of Theorem 4.4.1 hold and
property (4.27) is valid in a strengthened form: there exists an a ∈ Cv such
130 4. Sublinear Random Dynamical Systems
that the orbit emanating from a does not leave Cv and for any T > 0 there
exists λT > 1 such that
λ−1
T v(ω) ≤ ϕ(t, θ−t ω)a(θ−t ω) ≤ λT v(ω) for all t ∈ [0, T ] . (4.39)
Then property (4.39) holds for any a ∈ Cv , and precisely one of the following
three cases applies:
(i) for all b ∈ Cv , the orbit γb emanating from b is unbounded;
(ii) for all b ∈ Cv , the orbit γb emanating from b is bounded, but
lim sup sup p(ϕ(t, θ−t ω)b(θ−t ω), v(ω)) = 0 ; (4.40)
t→∞ ω∈Ω
Proof. Theorem 4.4.1 is applicable here. We need only prove (4.40) in case
(ii). If (4.40) is not true, then (4.39) implies that
λ−1
∞ v(ω) ≤ ϕ(t, θ−t ω)b(θ−t ω) ≤ λ∞ v(ω) for all t ≥ 0, ω ∈ Ω
γb (ω) ⊂ [λ−1
∞ v(ω), λ∞ v(ω)] ⊂ Cv
for any b(ω) with the property λ−1 u(ω) ≤ b(ω) ≤ λu(ω) for all ω ∈ Ω ∗ and
for some λ > 1.
4.4 Limit Set Trichotomy for Sublinear RDS 131
Proof. If (i) is not true, then there exist x0 > 0 and a set U ∈ F such that
P(U ) > 0 and supt∈R+ ϕ(t, θ−t ω)x0 < ∞ for ω ∈ U . Let
Since
Ũ = {ω : sup ϕ(t, θ−t ω)x0 < ∞} = {ω : ϕ(t, θ−t ω)x0 < N } ,
t∈Q∩R+
N ∈N t∈Q∩R+
the set Ũ is measurable. Thus as in the proof of Theorem 4.3.1 we can obtain
that there exists a θ-invariant set Ω ∗ of full measure such that
Therefore by Remark 1.6.1 and Proposition 1.6.4 the omega-limit set Γx0 (ω)
emanating from x0 is a nonempty invariant compact random set measurable
with respect to the σ-algebra F. Since sup B ∈ B for any compact set B ⊂ R+ ,
Lemma 3.4.1 and Remark 3.4.2(ii) imply that u(ω) := sup Γx0 (ω) ≥ 0 is an
F-measurable equilibrium on Ω ∗ . By Lemma 3.5.1 we have either u(ω) = 0
or u(ω) > 0 almost surely.
If u(ω) = 0 almost surely, then ϕ(t, θ−t ω)x → 0 almost surely for all
0 ≤ x ≤ x0 . The sublinearity implies that
x x
ϕ(t, ω)x = ϕ(t, ω) · x0 ≤ · ϕ(t, ω)x0 for all x ≥ x0 . (4.43)
x0 x0
The following two simple examples of discrete systems show that all three
cases of the limit set trichotomy can actually occur. The corresponding ex-
amples of RDS with continuous time are discussed in Chaps.5 and 6.
We start with the deterministic case.
Example 4.4.1. Let us consider the scalar function fα (x) = αx + 1+xx
on R+ .
It is easy to see that for every α ∈ R+ the mapping x → fα (x) generates a
strongly sublinear dynamical system in R+ . The hypotheses of Theorem 4.4.1
hold for this system with v = 1. If α ≥ 1, then fαn (x) → ∞ for any x > 0, i.e.
132 4. Sublinear Random Dynamical Systems
We note that assumptions (i) to (iii) are fulfilled, for instance, for the function
N
f (ω, x) = Ak (ω)xαk + b(ω) , (4.45)
k=1
sup p(ϕ(n, θ−n ω)w(θ−n ω), u(ω)) ≤ αn sup p(w(ω), u(ω)) (4.46)
ω∈Ω ω∈Ω
Corollary 4.5.1. Assume that the entries of the matrices Ak (ω) are bounded
from above by a nonrandom constant and α = maxj,k αkj < 1. Let b(ω) =
b0 (ω) · v, where v ∈ intRd+ and b0 (ω) > 0 is a scalar random variable such
that
0 < β0 ≤ β1 b0 (ω) ≤ b0 (θ1 ω) ≤ β2 b0 (ω), ω ∈ Ω ,
for some nonrandom βi . Then the part Cb generated by b is invariant for the
RDS (θ, ϕ) generated by (4.44) with f given by (4.45) and the conclusions of
Proposition 4.44 hold.
Proof. A simple calculation shows that b(ω) ≤ f (ω, b(ω)) ≤ Cb(ω) for some
constant C > 0. This implies that the part Cb is invariant and therefore we
can apply Proposition 4.5.1. 2
Remark 4.5.1. (i) In the situation of Corollary 4.5.1 the equilibrium u(ω) is
globally stable not only in Cb . It is easy to see that
provided that p(w(ω), u(ω)) is finite for each ω ∈ Ω. Therefore for each ω ∈ Ω
we have that
d ∂f (ω, x)
i
xj < α(ω)fi (ω, x), i = 1, . . . , d, x ∈ intRd+ , ω ∈ Ω ,
j=1
∂x j
where p is the part metric (see Krause/Nussbaum [72, Theorem 4.1]). Thus
under the condition α(ω) ≤ α0 < 1 we can obtain the same results as for the
mapping (4.45).
4.5 Random Mappings 135
The following assertion deals with another class of mappings and is an appli-
cation of the limit set trichotomy theorem.
Proposition 4.5.2. Assume that the measurable mapping f : Ω × Rd+ → Rd+
possesses properties (i) and (ii) and also
(iii∗ ) for each ω ∈ Ω the function f (ω, ·) is strongly sublinear, i.e.
Then there exists a unique strongly positive equilibrium u(ω) for the RDS
generated by (4.44). This equilibrium is uniformly separated from 0 and from
∞, i.e. there exist positive constants α and β such that αe ≤ u(ω) ≤ βe for
all ω ∈ Ω, where e = (1, . . . , 1). Moreover the equilibrium u(ω) is globally
asymptotically stable in intRd+ , i.e. for every x ∈ intRd+ we have
on the space all probability measures on (intRd+ , B(intRd+ )). The function
da,b (·, ·) is a distance if we restrict ourselves to measures with support in
[a, b].
Theorem 4.5.1. Assume that the hypotheses of Proposition 4.5.2 hold and
that the random mappings f (θn ω, ·), n ∈ Z, are independent and identically
distributed (i.i.d.). Let u(ω) be the equilibrium given by Proposition 4.5.2 for
the RDS (θ, ϕ) generated by (4.44). Then
136 4. Sublinear Random Dynamical Systems
is a homogeneous Markov chain with state space Rd+ and transition prob-
ability
for any interval [a, b] ⊂ intRd+ , where p(n) (x, A) = P{ω : ϕ(n, ω)x ∈ A}
for A ∈ B(Rd+ );
(iv) (P ∗n λ)(A) → ν(A) as n → ∞ for all A ∈ B(Rd+ ) and for any probability
measure λ on (Rd+ , B(Rd+ )) with compact support in intRd+ .
If the equilibrium u(ω) possesses the property
Proof. Items (i) and (ii) follow from the general assertion proved by Arnold
[3, Theorem 2.1.4] (see also the discussion in Sect.1.10). The support of ν is
a compact set in intRd+ because u(ω) ∈ [αe, βe] for all ω ∈ Ω.
(iii) Any compact K belongs to the interval [xm , xm ] with m large enough.
Therefore the relation
implies
P{ω : ϕ(n, ω)xm ≤ b} ≤ p(n) (x, [0, b]) ≤ P{ω : ϕ(n, ω)xm ≤ b}
for every x ∈ K ⊂ [xm , xm ]. From (4.50) we have
P{ω : ϕ(n, ω)z ≤ b} = P{ω : ϕ(n, θ−n ω)z ≤ b} → P{ω : u(ω) ≤ b}
as n → ∞ for any z ∈ intRd+ . Hence
In this section we consider affine and linear order-preserving RDS which leave
a cone to be invariant. The results given below show that the order-preserving
property provides us with an alternative approach to the study of affine RDS
and makes it possible to obtain additional information in contrast with the
more general affine RDS studied in Sect.1.9.
Let V be a real Banach space with closed convex cone V+ . Recall (see
Definition 1.2.3) that RDS (θ, ϕ) in V is affine if the cocycle ϕ is of the form
The affine RDS (θ, ϕ) is said to be positive (with respect to the cone V+ ) if
ϕ(t, ω)V+ ⊂ V+ for all t > 0 and ω ∈ Ω. If ψ(t, ω) ≡ 0 then the affine RDS is
said to be linear. The simplest properties of positive affine RDS are collected
in the following assertion.
Proposition 4.6.1. The affine RDS (θ, ϕ) with the cocycle ϕ of the form
(4.57) is positive with respect to the cone V+ if and only if Φ(t, ω) is positive,
i.e. maps V+ to itself, and ψ : T+ × Ω → V+ . Any positive affine RDS is a
sublinear order-preserving system. It is strongly sublinear if ψ(t, ω) 0 for
t > 0. Furthermore
Proof. If Φ(t, ω) is positive and ψ(t, ω) ≥ 0, then the RDS (θ, ϕ) is obviously
positive and order-preserving. On the other hand, if (θ, ϕ) is a positive RDS,
then ψ(t, ω) = ϕ(t, ω)0 ≥ 0. Since
1 1
Φ(t, ω)x + · ψ(t, ω) = · ϕ(t, ω)[λx] ≥ 0
λ λ
for any x ≥ 0 and λ > 0, letting λ → +∞ we obtain Φ(t, ω)x ≥ 0 for
x ≥ 0. Since w = 0 is a sub-equilibrium, Proposition 3.4.1 implies that
at (ω) = ψ(t, θ−t ω) = ϕ(t, θ−t ω)0 is also a sub-equilibrium for any t ≥ 0.
From (4.57) we have
over a metric dynamical system θ, where α(ω) and β(ω) are random vari-
ables such that t → α(θt ω) and t → β(θt ω) are locally integrable. Equation
an affine RDS
in R. The cocycle ϕ has the form (4.57) with
(4.60) generates
t
Φ(t, ω)x = x exp 0 α(θτ ω)dτ and
t
t
ψ(t, ω) = β(θs ω) exp α(θτ ω)dτ ds.
0 s
Theorem 4.6.1. Let (θ, ϕ) be a positive affine RDS with the cocycle ϕ rep-
resented in the form (4.57). Assume that there exists t0 = t0 (ω) > 0 such
that {ψ(t, θ−t ω) : t ≥ t0 } is a relatively compact set for each ω ∈ Ω. Then
for any random variable w possessing the property 0 ≤ w(ω) ≤ λu(ω) for all
ω ∈ Ω ∗ with some nonrandom constant λ > 0.
ψ(t + s, θ−t−s ω) = Φ(t, θ−t ω)ψ(s, θ−s θ−t ω) + ψ(t, θ−t ω), t, s ≥ 0 .
to prove that the orbit emanating from λu is relatively compact for any λ. 2
Example 4.6.2 (1D Positive Affine RDE). Consider the RDS (θ, ϕ) described
in Example 4.6.1. We additionally assume that θ is ergodic, α ∈ L1 (Ω, F, P),
and β(ω) ≥ 0 is a tempered random variable. If Eα < 0, then
0
0
ψ(t, θ−t ω) = β(θs ω) exp α(θτ ω)dτ ds ≤ u(ω)
−t s
Proposition 4.6.2. Assume that a system (θ, ϕ) on the solid normal cone
V+ is dominated from above by a positive affine RDS (θ, ϕaff ). Suppose that
the RDS (θ, ϕaff ) satisfies the hypotheses of Theorem 4.6.1 with ψ(t, ω) 0
for t > 0 and ω ∈ Ω. Let u(ω) be the strongly positive equilibrium for (θ, ϕaff ).
Then for any µ > 1 the random variable vµ (ω) = µu(ω) is an absorbing super-
equilibrium for (θ, ϕ) in the universe D consisting of all random closed sets
{B(ω)} such that B(ω) ⊂ [0, αu(ω)] for some α > 0. Moreover if V is a
finite-dimensional space, the RDS (θ, ϕ) possesses a random attractor in the
universe D and the conclusions of Theorem 3.6.2 hold.
Proof. Since ϕ(t, ω)x ≤ ϕaff (t, ω)x for all x ∈ V+ , we have
Thus [0, µu(ω)] is an absorbing set for (θ, ϕ) and therefore (θ, ϕ) is dissipa-
tive. If V is finite-dimensional, then Corollary 1.8.1 implies that a random
attractor exists in the universe D and we can apply Theorem 3.6.2. 2
The following assertion characterizes the linear part of positive affine RDS
with a strongly positive equilibrium.
Corollary 4.6.1. Let (θ, ϕ) be a positive affine RDS in V+ . Assume that
ψ(t, ω) 0 for t > 0 and that there exists a strongly positive equilibrium
u(ω). Then for any random element w(ω) such that 0 ≤ w(ω) ≤ αu(ω) for
all ω ∈ Ω with some nonrandom constant α > 0 we have
Proof. It is clear from (4.57) that (θ, Φ) is dominated from above by (θ, ϕ).
Therefore Proposition 4.6.2 implies that µu(ω) is an absorbing super-equilib-
rium for (θ, Φ) in the universe D consisting of all random closed sets {B(ω)}
such that B(ω) ⊂ [0, αu(ω)] for some α > 0. Hence (4.64) follows from
Proposition 1.9.1. 2
5. Cooperative Random Differential Equations
(R2) there exist random variables C1 and C2 such that t → Cj (θt ω) is locally
integrable and
where ·, · is the standard inner product in Rd and |x|2 = x, x;
(R3) f (ω, ·) is weakly positive, i.e.
where
Γi = x = (x1 , . . . , xd ) ∈ Rd+ : xi = 0 .
We note that condition (R3) is satisfied if and only if
Proposition 5.1.1. Assume that conditions (R1), (R2) and (R3) hold.
Then for any initial data x0 ∈ Rd+ at the moment t = 0 problem (5.1) has
a unique global solution x(t, ω) ≡ x(t, ω; x0 ) (see Definition 2.1.1) such that
x(t, ω) ∈ Rd+ for all t ≥ 0 and ω ∈ Ω. This solution is continuously dif-
ferentiable with respect to the initial data x0 and relations (2.6) and (2.7)
concerning the evolution of the Jacobian and its determinant hold.
Proof. We first extend the function f (ω, x) from Rd+ to Rd such that the
extended function f˜(ω, x) belongs to C 1 (Rd ) for all ω ∈ Ω and possesses
properties (2.1) and (2.2), i.e. for any compact set K ⊂ Rd there exists a
random variable CK (ω) ≥ 0 such that
a+1
CK (θt ω) dt < ∞ for all a ∈ R, ω ∈ Ω , (5.3)
a
and
|f˜(ω, x)| ≤ CK (ω), |f˜(ω, x) − f˜(ω, y)| ≤ CK (ω) · |x − y|
for any x, y ∈ K and ω ∈ Ω. It is clear that this extension exists. Now we
can apply Proposition 2.1.1 to prove that the problem
The following assertion shows that equation (5.1) generates an RDS in Rd+ .
Proposition 5.2.1. Assume that conditions (R1)–(R3) hold. Then equation
(5.1) generates a C 1 RDS (θ, ϕ) in Rd+ with the cocycle ϕ(t, ω) defined by the
formula ϕ(t, ω)x = x(t), where x(t) is an absolutely continuous solution to
the equation t
x(t) = x + f (θτ ω, x(τ )) dτ . (5.4)
0
ϕ(t, ω)(Rd+ \ {0}) ⊂ Rd+ \ {0} for all t>0 and ω∈Ω (5.5)
and
ϕ(t, ω)intRd+ ⊂ intRd+ for all t>0 and ω∈Ω. (5.6)
∗
If we additionally assume that (R3 ) holds, then (θ, ϕ) is strongly positive,
i.e.
ϕ(t, ω)(Rd+ \ {0}) ⊂ intRd+ for all t>0 and ω∈Ω. (5.7)
Proof. It follows from Proposition 5.1.1 that (5.1) generates a global C 1 RDS
(θ, ϕ) in Rd+ with properties (2.8) and (2.9).
To prove (5.5) we assume that for some fixed ω ∈ Ω, x ∈ Rd+ and t0 =
t0 (ω, x) > 0 we have x(t0 ) = ϕ(t0 , ω)x = 0. Since f (θτ ω, 0) ≥ 0 for all τ ∈ R,
equation (5.4) implies that
t0
0 ≤ x(t) ≤ − (f (θτ ω, x(τ )) − f (θτ ω, 0)) dτ, 0 ≤ t ≤ t0 .
t
Since supτ ∈[0,t0 ] |x(τ, ω)| ≤ r(ω) with some r(ω) > 0, property (R1) gives
146 5. Cooperative Random Differential Equations
where
Therefore we have
t0
|x(t)| ≤ C(τ, ω) · |x(τ )| dτ, 0 ≤ t ≤ t0 .
t
for 0 ≤ t ≤ t0 , where
which is impossible under condition (R3∗ ) provided that x(0) = x > 0. Thus
ϕ(t, ω)x 0 for all x > 0, i.e. we have (5.7). 2
∂fi (ω, x)
≥0 when i = j, x = (x1 , . . . , xd ) ∈ Rd+ . (5.10)
∂xj
As in the deterministic case (see, e.g., Hirsch [52], Krasnoselskii [68, 69],
Smith [102] and the references therein) we need the concept of irreducibility.
Recall the following definition.
Definition 5.2.1. A matrix A = {aij }di,j=1 is called irreducible if for every
nonempty, proper subset I of the set N = {1, 2, . . . , d}, there is an i ∈ I and
j ∈ N \ I such that aij = 0.
One can show that a matrix A is irreducible if and only if no nonzero proper
subspace spanned by a subset of the standard basis in Rd is mapped by A
into itself.
Theorem 5.2.1. Let (R1)–(R4) hold. Then equation (5.1) generates a stric-
tly order-preserving RDS (θ, ϕ) in Rd+ and
If the matrix
d
∂fi (ω, x)
Dx f (ω, x) ≡ (5.12)
∂xj i,j=1
ϕ(t, ω)x ϕ(t, ω)y for all 0x<y and ω∈Ω, (5.13)
The proof of this theorem follows the line of argument for the deterministic
case (see, e.g., Hirsch [52] or Krasnoselskii [68]) and relies on the following
assertion.
Lemma 5.2.1. Let (R1)–(R4) be valid. Let ϕ(t, ω) be the cocycle generated
by (5.1). Then for any x ∈ Rd+ the linear mapping ψx (t, ω) ≡ Dx ϕ(t, ω, x)
possesses the properties
ψx (t, ω)(Rd+ \ {0}) ⊂ Rd+ \ {0} for all t>0 and ω∈Ω; (5.15)
If additionally the matrix Dx f (θt ω, ϕ(t, ω, x)) is irreducible for all t ≥ 0 and
ω ∈ Ω, then ψx (t, ω) possesses the property
ψx (t, ω)(Rd+ \ {0}) ⊂ intRd+ for all t>0 and ω∈Ω. (5.17)
where x(t) = ϕ(t, ω)x. Assumption (R4) implies (R4∗ ) and therefore the
right-hand side of the equation (5.18) is weakly positive (see (R3)). Conse-
quently (5.14) follows from Proposition 2.2.1. Relation (5.15) can be proved
in the same way as (5.5).
To prove (5.16) let us assume that for some ω there exist t0 > 0,
z 0 and i ∈ {1, . . . , d} such that we have yi (t0 ) = 0 for the solution
y(t) = (y1 (t), . . . , yd (t)) to problem (5.18) with y0 = z. Since ψx (t, ω) is a
linear order-preserving operator, equation (5.14) implies that yi (t0 ) = 0 for
a solution to problem (5.18) with arbitrary initial data y0 ∈ Rd . This implies
that Detψx (t0 , ω) = 0 which is impossible because of Liouville’s equation
(2.9).
To obtain the last assertion of the lemma we apply the same method as in
the proof of property (5.7). Assume that for some ω ∈ Ω there exist a solution
y(t) = (y1 (t), . . . , yd (t)) ≥ 0 to (5.18) with nonzero initial data and a moment
t0 > 0 such that yi (t0 ) = 0 for i ∈ I and yi (t0 ) > 0 when i ∈ I, where I
is a proper subset of {1, . . . , d}. We note that the relation yi (t0 ) = 0 for all
5.2 Generation of RDS 149
Since {aij (t, ω)} is irreducible, there exists a pair {k, l} such that k ∈ I, l ∈ I
and akl (t, ω) > 0. These k and l can depend on t and ω. It follows from (5.18)
that
t0
s
yi (t) + aij (s, ω)yj (s) · exp − aii (τ, ω)dτ ds = 0
j=i t t
Since y(t) is continuous, we have yj (t) > 0 for j ∈ I and for all t ∈ [t0 − δ, t0 ]
with some δ = δ(ω) > 0. Therefore the irreducibility of {aij (t, ω)} implies
I (t, s, ω) > 0 for all s ∈ [t, t0 ] with t ∈ [t0 − δ, t0 ]. Thus from (5.19) we
that F!
have i∈I yi (t) < 0 for t ∈ [t0 − δ, t0 ) which is impossible. 2
Proof of Theorem 5.2.1. We make use of the equation
1
ϕ(t, ω, y) = ϕ(t, ω, x) + Dx ϕ(t, ω, sy + (1 − s)x)ds(y − x) (5.20)
0
Remark 5.2.1. Let (θ, ϕ) be the RDS in Rd+ generated by (5.1). Since t →
ϕ(t, θ−t ω)x is a right continuous function (see Remark 2.1.2(i)), it follows
from Proposition 1.5.2 that the closure γx (ω) of any pull back orbit γx (ω)
emanating from x ∈ Rd+ is a random closed set with respect to the σ-algebra
F.
150 5. Cooperative Random Differential Equations
implies that
ϕ(t, ω)x ≤ y(t, ω; x) for all t ∈ [0, t0 (ω, x)), ω ∈ Ω, x ∈ Rd+ . (5.23)
If
f (ω, x) ≥ g(ω, x) for all x ∈ Rd+ , ω ∈ Ω , (5.24)
then
ϕ(t, ω)x ≥ y(t, ω; x) for all t ∈ [0, t0 (ω, x)), ω ∈ Ω, x ∈ Rd+ . (5.25)
Proof. Assume (5.22). Then the function z(t) = y(t, ω; x) − ϕ(t, ω)x is a local
solution to problem
where
h(t, ω, z) = g(θt ω, ϕ(t, ω)x + z) − f (θt ω, ϕ(t, ω)x) .
5.3 Random Comparison Principle 151
for every z = (z1 , . . . , zd ) ∈ Rd+ with zi = 0 such that z ≤ ϕ(t, ω)x. Therefore
as above we can conclude that z ∗ (t) ≥ 0. Thus we obtain (5.25). 2
Proof. It is necessary to prove (ii) and (iv) only. Assume that (5.22) with
a strict inequality is valid. Suppose that for some ω ∈ Ω, x ∈ intRd+ and
t0 > 0 we have that ϕ(t0 , ω)x = ψ(t0 , ω)x. Denote x(t) = ϕ(t, ω)x and
y(t) = ψ(t, ω)x. Then from (5.22) and from assertion (i) of the corollary we
have t0
0 ≤ y(t) − x(t) ≤ − (g(θτ ω, y(τ )) − g(θτ ω, x(τ ))) dτ
t
for all t ∈ [0, t0 ]. This equality allows us to conclude that y(t) = x(t) for all
t ∈ [0, t0 ] (cf. the argument given in the proof of Proposition 5.2.1). Therefore
from (5.1) and (5.21) we have the equality
t0
(g(θτ ω, x(τ )) − f (θτ ω, x(τ ))) dτ = 0, t ∈ [0, t0 ] .
t
By Theorem 5.2.1 x(τ ) ∈ intRd+ and therefore the last equality contradicts
to the strict inequality in (5.22) for x ∈ intRd+ .
The proof of (iv) is similar. 2
then
(ii) If
then
(b) property (5.24) with strict inequality for every x ∈ intRd+ implies
(5.29).
Proof. Let x(t) = ϕ(t, ω)x and y(t) = ψ(t, ω)x with x ∈ intRd+ .
(i) We obviously have
t
y(t) − x(t) = y(s) − x(s) + [g(θτ ω, y(τ )) − f (θτ ω, x(τ ))] dτ .
s
Corollary 5.3.1 and Theorem 5.2.1 imply that y(t) > x(t) 0 for all t > 0.
Therefore it follows from (5.26) that
t
yi (t) − xi (t) > yi (s) − xi (s) + [fi (θτ ω, y(τ )) − fi (θτ ω, x(τ ))] dτ (5.30)
s
where 1
∂fi
aij (τ ) = (θτ ω, x(τ ) + λ(y(τ ) − x(τ ))) dλ .
0 ∂xj
Cooperativity condition (R4∗ ) and the relation yj (τ ) − xj (τ ) ≥ 0 imply that
t
yi (t) − xi (t) ≥ yi (s) − xi (s) + aii (τ )(yi (τ ) − xi (τ ))dτ
s
for all 0 ≤ s ≤ t0 . This implies that yi (s) = xi (s) for all s ∈ [0, t0 ]. Therefore
we can apply (R4) in (5.30) and obtain the inequality
t
xi (t) − yi (t) > xi (s) − yi (s) + [fi (θτ ω, x(τ )) − fi (θτ ω, y(τ ))] dτ
s
where I is a proper subset of {1, . . . , d}. We note that the relation yi (t0 ) =
xi (t0 ) for all i = 1, . . . , d is impossible because of Corollary 5.3.1(ii). An
argument similar to given above makes it possible to obtain (5.31) for this
case. Hence as above we can prove that yi (s) = xi (s) for all s ∈ [0, t0 ] and
i ∈ I. Therefore (5.31) implies that
t0
Fi (τ )dτ ≤ 0 for all s ∈ [0, t0 ] , (5.34)
i∈I s
where
Fi (τ ) = aij (τ )(yj (τ ) − xj (τ )) ≥ 0, i∈I.
j∈I
Since yj (t) − xj (t) > 0 for j ∈ I and for all t ∈ [t0 − δ,!t0 ] with some
δ = δ(ω) > 0, from the irreducibility condition we get that i∈I Fi (t) > 0
for t ∈ [t0 − δ, t0 ]. This contradicts to (5.34).
To prove (iii-b) we use arguments similar to ones given in the proofs of
(ii) and (iii-a). 2
As an application of the random comparison principle we prove the following
assertion.
Theorem 5.3.3. Assume that f satisfies (R1)–(R4) and
where A(ω) = {aij (ω)}di,j=1 and b(ω) = (b1 (ω), . . . , bd (ω)) ≥ 0 are tempered
random variables such that t →
A(θt ω)
and t → |b(θt ω)| are locally in-
tegrable. Assume also that θ is an ergodic metric dynamical system and the
linear RDS (θ, Φ) generated by
has top Lyapunov exponent negative. Then there exist a θ-invariant set Ω ∗
of full measure and a version (θ, ϕ̃) of the RDS (θ, ϕ) generated by (5.1) such
that ϕ̃(t, ω) = ϕ(t, ω) for all ω ∈ Ω ∗ and t ≥ 0, and the RDS (θ, ϕ̃) possesses
an absorbing super-equilibrium v(ω) 0 in the universe D of all tempered
subsets of Rd+ .
5.3 Random Comparison Principle 155
Proof. The comparison principle implies that the RDS (θ, ϕc ) generated by
the equation
ẋ(t) = A(θt ω)x(t) + b(θt ω) + c · e ,
where e := (1, . . . , 1) ∈ Rd+ , dominates the system (θ, ϕ) from above for any
c ∈ R+ . Condition (5.35) and the weak positivity property (R3) of f imply
that aij (ω) ≥ 0, i = j. Therefore (θ, ϕc ) is a strictly order-preserving RDS in
Rd+ (see Theorem 5.2.1). By (2.15) the cocycle ϕc (t, ω) has the form
ϕc (t, ω)x = Φ(t, ω)x + ψc (t, ω) ,
where t
ψc (t, ω) := Φ(t − s, θs ω) (b(θs ω) + c · e) ds .
0
Property (5.11) implies that ψc (t, ω) 0 for all ω ∈ Ω, t > 0 and c > 0. By
Theorem 2.1.2 and Definition 1.9.1 we also have
0 √
|ψc (t, θ−t ω)| ≤ Rε (θs ω)e−(λ+ε)s |b(θs ω)| + c · d ds, ω ∈ Ω ∗ ,
−t
where ε > 0 is arbitrary and Ω ∗ is the θ-invariant set of full measure given
by Theorem 2.1.2. Therefore we can choose ε > 0 such that {|ψc (t, θ−t ω)| :
t ≥ 0} is bounded for every ω ∈ Ω ∗ . Consequently by Proposition 1.9.3 for
any c > 0 the system (θ, ϕc ) possesses an equilibrium wc (ω) on Ω ∗ such that
lim eγt sup |ϕ(t, θ−t ω)v − wc (ω)| = 0, ω ∈ Ω∗ ,
t→+∞ v∈D(θ−t ω)
for any tempered random closed set D ⊂ Rd+ and γ < −λ. It is also clear
from Theorem 4.6.1 that wc (ω) wc (ω) when c > c , ω ∈ Ω ∗ . Therefore, if
we redefine f by f (ω, x) = −x + e on Ω \ Ω ∗ , then
wc (ω) if ω ∈ Ω ∗ ,
vc (ω) =
(1 + c)e if ω ∈ Ω \ Ω ∗ ,
is a D-absorbing super-equilibrium for the RDS (θ, ϕ̃) with the cocycle ϕ̃
defined by the formula
ϕ(t, ω)x if ω ∈ Ω ∗ ,
ϕ̃(t, ω)x =
e−t x + (1 − e−t )e if ω ∈ Ω \ Ω ∗ .
2
Corollary 5.3.2. Under the hypotheses of Theorem 5.3.3 the RDS (θ, ϕ̃)
generated by (5.1) has a global random attractor in the universe D of all tem-
pered subsets of Rd+ and it possesses the properties stated in Theorem 3.6.2.
Proof. This follows directly from Theorem 5.3.3, Corollary 1.8.1 and Theo-
rem 3.6.2. 2
156 5. Cooperative Random Differential Equations
Proof. Let f (ω, 0) > 0. Assume that there exists t0 > 0 such that ϕ(t0 , ω)0 =
0 for some ω ∈ Ω. Then the same argument as in the proof of property (5.5)
gives us that ϕ(t, ω)0 = 0 for all t ∈ [0, t0 ]. Thus x(t) ≡ 0 is a stationary
solution to (5.1) which is impossible because f (ω, 0) > 0.
Assume now that (5.36) holds. Denote x(t) = ϕ(t, ω)0. Suppose that there
exists t0 > 0 such that xi (t0 ) = 0 for some i and ω ∈ Ω. Then
t0
−xi (t) = fi (θτ ω, 0, . . . , 0, xi (τ ), 0, . . . , 0) dτ
t
t0
+ (fi (θτ ω, x(τ )) − fi (θτ ω, 0, . . . , 0, xi (τ ), 0, . . . , 0)) dτ
t
Thus as in the proof of Proposition 5.2.1 we find that xi (t) ≡ 0 for t ∈ [0, t0 ].
t
Hence (5.37) implies t 0 fi (θτ ω, 0)dτ ≤ 0 which is impossible. 2
γct0 (ω) (ω) = ϕ(t, θ−t ω)x ,
t≥t0 (ω)
is a compact set in Rd+ . Then there exists an equilibrium u(ω) for the RDS
(θ, ϕ) generated by (5.1). This equilibrium is positive when f (ω, 0) > 0.
Proof. Since t → ϕ(t, θ−t ω)x is a right continuous function (see Remark 2.1.2)
and therefore a separable process, Remark 1.6.1 and Proposition 1.6.4 imply
that the omega-limit set Γx (ω) exists and is an invariant random compact
set. Therefore we can apply Lemma 3.4.1. Hence v(ω) := inf Γx (ω) is a super-
equilibrium such that v(ω) ≥ 0. Since 0 is a sub-equlibrium, the existence of
an equilibrium u(ω) ∈ [0, v(ω)] now follows from Theorem 3.5.1. 2
is bounded, then there exists an equilibrium u(ω) for the RDS (θ, ϕ) generated
by (5.1) in Rd+ such that 0 ≤ u(ω) ≤ sup B. In this case there also exists a sub-
equilibrium w(ω) with the properties w(ω) ≥ u(ω) and inf B ≤ w(ω) ≤ sup B.
We also have u(ω) > 0 provided that f (ω, 0) > 0 for all ω ∈ Ω and u(ω) 0
if (5.36) holds.
The proof of this theorem relies on the following lemma.
Lemma 5.4.1. Let assumptions (R1)–(R3) be valid and assume that there
exists a C 1 function W (x) from intRd+ into R+ such that for some nonrandom
R we have (5.38) provided that W (x) = R. Then the set B given by (5.39)
is a forward invariant set with respect to ϕ(t, ω), i.e. ϕ(t, ω)B ⊂ B for all
ω ∈ Ω. Here ϕ(t, ω) is the cocycle generated by (5.1).
now follows from Theorem 3.5.1. This last relation and Proposition 5.4.1
imply the positivity properties of u(ω). 2
Corollary 5.4.1. Let (R1)–(R4) hold. Assume that there exist positive num-
bers R and α such that
d
xiα−1 fi (ω, x1 , . . . , xd ) ≤ 0
i=1
!d
provided that i=1 xα i = R and x ∈ intR+ . Then there exists an equilibrium
d
1/α
u(ω) lying in the interval [0, R e] for the RDS generated by (5.1) in Rd+ .
d
Here e is the element from intR+ given by the formula e = (1, . . . , 1).
!d
Proof. We apply Theorem 5.4.1 with W (x) = i=1 xα
i . 2
Below we make use of the following simple sufficient condition for the exis-
tence of super- and sub-equilibria for problem (5.1).
Proposition 5.4.3. Let (R1)–(R3) be valid. Assume that there exists a non-
random element w ∈ Rd+ such that f (ω, x) satisfies (R4) for all x ∈ [0, w]
and
fi (ω, w) ≤ 0, for all i = 1, . . . , d and ω ∈ Ω . (5.40)
Then w(ω) ≡ w is a super-equilibrium for the RDS (θ, ϕ) generated by (5.1).
If the inequality in the formula above is reversed and (R4) holds for all x ≥ w,
then w(ω) ≡ w is a sub-equilibrium.
Proof. If w = 0, then the weak positivity (R3) and equation (5.40) imply
that f (ω, 0) = 0 and therefore w is an equilibrium.
Assume that there exists w ∈ intRd+ such that fi (ω, w) ≤ 0 for all ω ∈ Ω
and i = 1, . . . , d. The cooperativity condition implies that
where {ei } is the standard basis in Rd . Therefore from (5.41) we have that
f (ω, x), νx = αi fi (ω, x) ≤ 0 .
i∈I
Thus Theorem 2.2.2 implies that the set [0, w]∩intRd+ is invariant with respect
to (θ, ϕ). Hence [0, w] is also an invariant set and w is a super-equilibrium
(see Remark 3.4.1).
Assume now that w ∈ ∂Rd+ \ {0}. For the sake of simplicity we consider
the case w1 = 0 and wj > 0, j = 2, . . . , d (for other cases the proof is similar).
The weak positivity condition (R3) and (5.40) imply that
Theorem 5.4.2. Let assumptions (R1)–(R4) be valid and assume that there
exists a C 1 function W (x) from Rd+ into R such that
f (ω, x), ∇W (x) + (β + (ω)) · W (x) ≤ C(ω) for all ω∈Ω, (5.44)
for all ω ∈ Ω. Then the RDS (θ, ϕ) possesses a random attractor A(ω) in
the universe D of all tempered random closed subsets of Rd+ . This attractor
is bounded from above and from below and there exist maximal and minimal
equilibria ū and u such that the random interval [u, ū] contains the attractor as
well as all other possible tempered equilibria. In particular, if the equilibrium
u is unique, then A = {u}.
Remark 5.4.1. The hypotheses of Theorem 5.4.2 hold with W (x) = |x|2 , if
f (ω, x) satisfies (R1), (R3), (R4) and also the inequality
with β, (ω) and C(ω) possessing the properties listed in Theorem 5.4.2.
Lemma 5.5.1. Assume that conditions (R1)–(R4) hold and for any ω ∈ Ω
the function f (ω, ·) is a sublinear mapping from Rd+ into Rd , i.e.
for 0 < λ < 1 and for all x ∈ Rd+ and ω ∈ Ω. Then the RDS (θ, ϕ) gener-
ated by (5.1) is sublinear. Moreover (θ, ϕ) is strongly sublinear if one of the
following conditions is satisfied:
(i) λfi (ω, x) < fi (ω, λx) for all i = 1, . . . , d, 0 < λ < 1, x ∈ intRd+ and
ω ∈ Ω;
(ii) the matrix (5.12) is irreducible for all x ∈ intRd+ and ω ∈ Ω and prop-
erty (5.45) holds with strict inequality for every x ∈ intRd+ .
Proof. The function xλ (t) = λ · ϕ(t, ω)x is the solution to the problem
where fλ (ω, x) = λf (ω, λ−1 x). From (5.45) we have fλ (ω, x) ≤ f (ω, x).
Therefore the comparison principle (see Corollary 5.3.1) gives
for any random variable v(ω) possessing the property 0 ≤ v(ω) ≤ λu(ω)
for all ω ∈ Ω ∗ and for some nonrandom λ > 0.
Proof. Proposition 5.4.1 implies that ϕ(t, ω)0 0 for all t > 0 and ω ∈ Ω.
It is also clear that any finite-dimensional RDS is conditionally compact.
Therefore we can apply Corollary 4.3.1. 2
162 5. Cooperative Random Differential Equations
We can also apply the trichotomy theorem (see Sect.4.4) in our situation.
Theorem 5.5.2 (Limit Set Trichotomy). Let conditions (R1), (R3),
(R4) and either (i) or (ii) of Lemma 5.5.1 be valid. Instead of (R2) we assume
that there exist positive nonrandom constants a and b such that
lim ϕ(t, θ−t ω)b(θ−t ω) = u(ω) for almost all ω∈Ω. (5.48)
t→+∞
ẏ1 (t) = −a · |y1 (t)|1 · e and ẏ2 (t) = b · (1 + |y2 (t)|1 ) · e (5.49)
with initial data y1,2 (0) = e and t0 := sup{s : y1 (t) ≥ 0, t ∈ [0, s)}. It is easy
to see that
Thus the orbit emanating from e does not leave Ce and therefore we can
apply Theorem 4.4.1. By Remark 5.2.1 the trajectory emanating from e is a
random set with respect to F. Therefore by Remark 4.4.1(i) the equilibrium
u(ω) is F-measurable. This completes the proof. 2
Remark 5.5.1. Relation (5.51) shows that under the conditions of Theo-
rem 5.5.2 property (4.39) holds with a(ω) ≡ e. Therefore by Corollary 4.4.1
statement (ii) in Theorem 5.5.2 is valid in the form: for all b ∈ Ce , the orbit
γb emanating from b is bounded, but
lim sup sup p(ϕ(t, θ−t ω)b(θ−t ω), v(ω)) = ∞ ,
t→∞ ω∈Ω
or (b) there exists a unique equilibrium u(ω) such that 0 u(ω) ≤ α · e for
ω ∈ Ω ∗ and we have (5.52) for all v ∈ Ce .
164 5. Cooperative Random Differential Equations
Proof. If under the condition in (i) option (a) is not valid, then the orbit
emanating from w is bounded. Therefore (b) follows from the first part of
Theorem 4.2.3 (see also Remark 4.2.2).
As for assertion (ii), the second part of Theorem 4.2.3 implies the existence
of an equilibrium 0 ≤ u(ω) ≤ w(ω) ≤ α · e. Lemma 3.5.1 and (5.53) give that
either u(ω) = 0 or u(ω) 0 on a θ-invariant set of full measure. In the first
case we obtain (5.54). In the second case we obtain (b). 2
Proof. We first note that y(t) = Dx ϕ(t, ω, x)z is the solution to the problem
Hence
ϕ(t, ω, x) ≥ λ · Dx ϕ(t, ω, 0)x + λ · b(t, ω, x, λ) , (5.56)
5.5 Random Equations with Concavity Properties 165
where
λ
−1
b(t, ω, x, λ) = λ (Dx ϕ(t, ω, sx) − Dx ϕ(t, ω, 0)) x ds
0
Thus (5.56) implies that ϕ(t, ω, x) 0 for every x > 0, whence (5.53). 2
Hence λϕ(t, ω, y) ϕ(t, ω, λy) for all t > 0 and y > 0. Thus (θ, ϕ) is strongly
sublinear. Therefore we can apply the same argument as in the proof of
Proposition 5.5.1 to obtain (a)/(b) dichotomy.
Since 1
f (ω, x) = f (ω, 0) + Dx f (ω, sx)ds x ,
0
Therefore the affine RDS generated by (5.58) dominates (θ, ϕ) from above.
Hence w(ω) ∈ intRd+ is a super-equilibrium for (θ, ϕ). Thus the orbit γw
emanating from w is bounded, whence (b).
The proof of the last assertion follows from Proposition 4.2.2 (cf. Theo-
rem 2.1.2). 2
in the case 0 < a < ∞ (see Proposition 5.4.3) and by the relation
in the case [0, a] = R+ , where C1 and C2 are random variables such that
t → Cj (θt ω) is locally integrable (see Proposition 5.2.1).
If
g(x) > 0 and f (x) = g(x) · (γ1 G(x) + γ2 ) for 0<x<a, (5.61)
where γ1 and γ2 are constants and G(x) is a primitive for [g(x)]−1 on the
interval (0, a) (i.e. G (x) = [g(x)]−1 , x ∈ (0, a)), then the cocycle ϕ can be
represented in the explicit form. Indeed, if x(t) is a solution to (5.59), then
y(t) = G(x(t)) solves the linear equation
t
x exp 0 α(θτ ω)dτ
ϕ(t, ω, x) =
1/N
(5.64)
t s
1 + N xN 0 |β(θs ω)| exp N 0 α(θτ ω)dτ ds
for x > 0 and ϕ(t, ω, 0) = 0. Since β(ω) ≤ 0, the RDS (θ, ϕ) is dominated
from above by the system generated by the linear equation ẋ = α(θt ω) · x.
168 5. Cooperative Random Differential Equations
By the ergodic theorem (see (5.60)) the Lyapunov exponent for this system
is λ = Eα. Therefore if Eα < 0, then limt→∞ ϕ(t, θ−t ω, u(θ−t ω)) = 0 for any
tempered u(ω) ≥ 0. Thus A(ω) = {0} is a random attractor for (θ, ϕ) in the
universe D of all tempered subsets of R+ .
Assume that Eα > 0 and β(ω) ≤ −β0 < 0 for all ω ∈ Ω. Then using
Theorem 5.4.2 (see Remark 5.4.1) we can prove that (θ, ϕ) possesses a random
attractor A(ω) = [0, u(ω)] in the universe D. Here u(ω) ≥ 0 is an equilibrium.
Using (5.64) it is easy to find that
0
0 −1/N
u(ω) = N |β(θs ω)| exp −N α(θτ ω)dτ ds .
−∞ s
lim eγt |ϕ(t, θ−t ω, x) − u(ω)| = 0 for all x > 0 and ω ∈ Ω . (5.65)
t→∞
If for some b > 0 we have α(ω) + β(ω) · bN ≤ 0 for all ω ∈ Ω, then the equi-
librium u(ω) belongs to the interval (0, b]. Indeed, in this case by Proposition
5.4.3 b is a super-equilibrium, i.e. ϕ(t, θ−t ω, b) ≤ b. Therefore (5.65) implies
that u(ω) ≤ b.
where g ∈ C 1 (R), g(x) > 0 for x ∈ (0, 1) and g(0) = g(1) = 0. This equation
generates an RDS (θ, ϕ) in [0, 1] with the cocycle
t
ϕ(t, ω, x) = G−1 G(x) + β(θτ ω)dτ , (5.67)
0
where G(x) is a primitive for [g(x)]−1 on (0, 1) (relation (5.61) holds with
γ1 = γ2 = 0). It is clear that G(x) is an increasing function such that G(x) →
+∞ as x → 1 and G(x) → −∞ as x → 0. Therefore using (5.60) and (5.67)
we observe the following behaviour of trajectories:
for all t > 0, ω ∈ Ω and x ∈ (0, 1), i.e. all trajectories emanating from points
x ∈ (0, 1) are separated from the equilibria 0 and 1 uniformly with respect
to t.
On the other hand, if we assume that β ∈ L2 (Ω, F, P), Eβ = 0, and the
process β(θt ω) satisfies the central limit theorem, i.e. the limit
t 2
−1/2
lim E t β(θτ ·)dτ =σ>0
t→∞ 0
exists and
t a
1 ξ2
−1/2
lim P ω : t β(θτ ω)dτ < a =√ e− 2σ dξ
t→∞ 0 2πσ −∞
for any a ∈ [−∞, +∞], then it follows from (5.67) that under the conditions
g (0) > 0 and g (1) < 0 we have
√ √
1
lim P ω : e− t log t ≤ ϕ(t, ω, x) ≤ e− t/ log t =
t→∞ 2
and √ √
1
lim P ω : e− t log t ≤ 1 − ϕ(t, ω, x) ≤ e− t/ log t =
t→∞ 2
for all x ∈ (0, 1). We refer to Chueshov/Vuillermot [24] for the proof of
the last two formulas and for other facts on the long-time dynamics of the
RDS generated by (5.66). The last two relations imply that
for any δ > 0 and x ∈ (0, 1). This means that dist (ϕ(t, ω, x), {0, 1}) → 0 in
probability as t → ∞. Thus the two-point set A = {0, 1} is a weak point
attractor, i.e. an attractor with respect to convergence in probability.
170 5. Cooperative Random Differential Equations
over an ergodic metric dynamical system θ, where α(ω) and β(ω) are random
variables such that t → α(θt ω) and t → β(θt ω) are locally integrable. This
equation generates an RDS in R. Since any interval [2πk, 2π(k + 1)] is an
invariant set and equation (5.68) is invariant with respect to the change
x → x+2πk, we consider the problem on the unit circle C which is interpreted
as the interval [0, 2π] with identified end-points.
A simple calculation shows that relation (5.61) holds with γ1 = −1, γ2 = 0
and G(x) = − cot x2 . Therefore equation (5.68) generates an RDS (θ, ϕ) in C
with the cocycle
for x ∈ C \ {0}. Assume that α ∈ L1 (Ω, F, P). It follows from the considera-
tions presented in Example 2.1.2 that in both cases Eα > 0 and Eα > 0 the
RDS (θ, ϕ) on the circle C has two equilibria u0 ≡ 0 and either
0 0
u+ (ω) = 2arccot − β(θs ω)e− s
α(θτ ω)dτ
ds if Eα > 0 ,
−∞
or
∞ s
u− (ω) = 2arccot β(θs ω)e 0
α(θτ ω)dτ
ds if Eα < 0 .
0
for any random variable v(ω) from the interval [ε, 2π − ε], where ε > 0 is
arbitrary.
In the case Eα < 0 using the representation
t
u− (θt ω) u− (ω)
y(t, ω; y0 ) = − cot + y0 + cot · e− 0 α(θτ ω)dτ
2 2
5.7 Applications 171
we obtain that
ϕ(t, θ−t ω)v(θ−t ω) = 2arccot z(t, ω)
for any v(ω) ∈ C \ {0}, where
0
u− (ω) v(θ−t ω) u− (θ−t ω)
z(t, ω) = − cot − cot − cot · e− −t α(θτ ω)dτ .
2 2 2
provided that
u− (ω)
v(ω)
cot − cot ≥ δ(ω) > 0 ,
2 2
where δ(ω)−1 is a tempered random variable.
Thus we observe the exchange of stability between the equilibria u0 (ω)
and u± (ω) when the value Eα changes the sign in the RDS (θ, ϕ) generated
by (5.68) in the circle C.
5.7 Applications
for every ω ∈ Ω and for every x > 0. We assume also that t → b(θt ω) and
t → c(θt ω) are locally integrable. We use g to denote the derivative with
respect to the space variable.
172 5. Cooperative Random Differential Equations
where A is the matrix with all entries equal to zero, except ajj = −αj ,
aj,j−1 = 1 and a1d = a, B(ω) is the column whose only nonzero element is
b1 (ω) = b(ω). Since the eigenvalues λ of the matrix A satisfy the equation
&d &d
j=1 (λ + αj ) = a, it is easy to see that Reλ < 0 provided that j=1 αj > a,
which we assume to make A stable.
Assume also that b(ω) is tempered. Then it is easy to see that t →
ξ(t, θ−t ω) is bounded for all ω ∈ Ω. Therefore by Theorem 4.6.1 and Re-
mark 4.6.1
0
w(ω) := lim ξ(t, θ−t ω) = e−τ A B(θτ ω) dτ (5.75)
t→∞ −∞
this equilibrium uniformly attracts all tempered random sets with exponential
speed, i.e. there exists γ > 0 such that
for any D(ω) ∈ D, where D is the universe of all random tempered sets in
Rd+ .
By the comparison principle the random variable µw(ω) is a super-
equilibrium for the RDS (θ, ϕ) generated by (5.69) and (5.70) for any µ ≥ 1.
5.7 Applications 173
and
0
0
vj (ω) = vj−1 (θτ ω) · exp − αj (θs ω)ds dτ, j = 2, . . . , d .
−∞ τ
For every ω the equilibrium v(ω) belongs to intRd+ and v(ω) ≤ w(ω), where
w(ω) is given by (5.75). The comparison principle gives that µv(ω) is a sub-
equilibrium of (θ, ϕ) for any 0 ≤ µ ≤ 1. Therefore the random attractor of
(θ, ϕ) is contained in the interval [v(ω), w(ω)]. According to Theorem 3.6.2
this attractor lies between two equilibria u(ω) and u(ω) such that 0 v(ω) ≤
u(ω) ≤ u(ω) ≤ w(ω). Moreover, u(ω) = u(ω) almost surely provided that the
function g(ω, x) possesses the property
In this case the condition (ii) of Lemma 5.5.1 holds. This implies that the
system (θ, ϕ) is strongly sublinear and therefore we can apply Theorem 4.2.1
on the uniqueness of strongly positive equilibria.
We now consider the case when g(ω, 0) = 0 for all ω ∈ Ω. In this case
the attractor A lies between the two equilibria u(ω) ≡ 0 and u(ω) ≥ 0. We
can guarantee that u(ω) 0 if we assume, for instance, that there exist
nonrandom constants αj∗ , j = 1, . . . , d, and a function g0 (x) such that
and
'
d
g0 (x)
αj∗ < lim sup ≤ +∞ . (5.76)
j=1
x→+0 x
Indeed, let
174 5. Cooperative Random Differential Equations
'
d '
d
v (n) = (v1n , . . . , vdn ) := εn αj∗ , αj∗ , . . . , αd∗ , 1 0 ,
j=2 j=3
for every n almost surely. If (5.77) is not true, then there exists another
strongly positive equilibrium w(ω) such that ϕ(t, θ−t ω)v (n) → w(ω). This
contradicts the uniqueness of strongly positive equilibria for strongly sublin-
ear RDS. Thus the equilibrium u(ω) is stable.
If g (ω, x) is a strictly decreasing function for every ω ∈ Ω, then Propo-
sition 5.5.2 implies that the RDS (θ, ϕ) is s-concave. Therefore if for some
nonrandom g0∗
'
d
g (ω, 0) ≤ g0∗ < αj , (5.78)
j=1
the system (θ, ϕ) is dominated from above by the RDS generated by the
linear equations
ẋ1 (t) = g0∗ · xd (t) − α1 x1 (t) , (5.79)
ẋj (t) = xj−1 (t) − αj xj (t), j = 2, . . . , d . (5.80)
However, assumption (5.78) means that all the eigenvalues of problem (5.79)
and (5.80) possesses the property Reλj < 0. Thus the zero equilibrium of
(5.79) and (5.80) is exponentially stable. This implies that the random at-
tractor A(ω) of (θ, ϕ) is trivial, i.e. A(ω) = {0}.
5.7 Applications 175
fj (ω, x, p − x) ≤ 0, j = 1, . . . , d ,
It is easy to see that equations (5.81) possess a local solution for any initial
data from {x : 0 x p}. Assumption (b) and Proposition 5.4.3 imply
that the interval [0, p] is a forward invariant set. This makes it possible to
guarantee the global existence of solutions to (5.81) with initial data from
[0, p] for every ω and therefore equations (5.81) generate an RDS with state
space X = [0, p] ⊂ Rd+ . Assumption (c) implies that this RDS is strictly
order-preserving (see Theorem 5.2.1). It is clear that w(ω) ≡ p is a super-
equilibrium and v(ω) ≡ 0 is a sub-equilibrium. Therefore Theorem 3.5.1
implies the existence of an equilibrium u(ω) with the property
If f (ω, 0, p) = (f1 (ω, 0, p), . . . , fd (ω, 0, p)) > 0 then u(ω) > 0 by Proposi-
tion 5.4.1.
We note that deterministic version of the equations (5.81) first appeared in
Hirsch [53] as a generalization of a gonorrhea transmission models previously
176 5. Cooperative Random Differential Equations
d
fj (ω, x, p − x) = −αj (ω)xj + (pj − xj ) βji (ω)xi . (5.82)
i=1
Here αj (ω) and βji (ω) are random variables such that t → αj (θt ω) and
t → βji (θt ω) are locally integrable and satisfy the inequalities
d
−αj (ω)xj + βjj (ω)(pj − xj )xj ≤ fj (ω, x, p − x) ≤ −αj (ω)xj + pj βji (ω)xi
i=1
for x ∈ [0, p]. Using the Comparison Theorem 5.3.1 we find that the RDS
generated by (5.81) and (5.82) is dominated from above by a linear system
and from below by the direct product of one-dimensional RDS. These prop-
erties make it possible (see Theorem 5.5.3 and Example 5.6.1) to give some
condition on αj (ω) and βji (ω) which ensure one of the following cases:
(a) u(ω) ≡ 0 is globally stable and A = {0};
(b) there exists a strictly positive equilibrium u(ω) and A ⊂ [0, u], where
u(ω) ≡ 0 is unstable and u(ω) is stable.
where αj (ω, x) and gi (ω, x) are functions on Ω × R+ such that the conditions
on hj mentioned above are valid. We also assume that
ẋ = x · αj (θt ω, x) (5.86)
has a positive equilibrium for every j = 1, . . . , d. For this we can assume, for
example, that
αj (ω, x) ≥ αj0 > 0 for 0 ≤ x ≤ δ
and
αj (ω, x) ≤ −βj0 < 0 for x ≥ R ,
where αj0 , δ, βj0 and R are positive nonrandom constants. We note that equa-
tion (5.86) describes the evolution of the j-th population independent of
the others. Denote the positive equilibrium for (5.86) by vj (ω). Assumptions
(5.85) ensure that the RDS generated in Rd+ by
ẋj = xj · αj (θt ω, xj ), j = 1, . . . , d ,
dominates the RDS (θ, ϕ) generated by (5.83) with (5.84) from below. There-
fore v(ω) = (v1 (ω), . . . , vd (ω)) is a positive sub-equilibrium for (θ, ϕ). On the
other hand under condition (5.85) the system generated by
dominates (θ, ϕ) from above. If βj0 > (d − 1)M this RDS has a super-
equilibrium w(ω) such that w(ω) > v(ω). Therefore Theorem 3.5.1 implies the
existence of an equilibrium u(ω) such that v(ω) ≤ u(ω) ≤ w(ω). This equi-
librium attracts (from below) the collection of equilibria (v1 (ω), . . . , vd (ω))
which correspond to the isolated dynamics of each population. We hence
observe that the interaction results in a benefit for all populations.
d
ẋi = aij (θt ω) · hj (xj ), i = j, . . . , d,
j=1
where the matrix aij (ω) satisfies the cooperativity condition, i.e. aij (ω) ≥ 0
for all i = j and ω ∈ Ω and
d
aij (ω) = 0 for all ω ∈ Ω, j = 1, . . . , d ,
i=1
is a forward invariant set for the RDS (θ, ϕ) generated by (5.87). Therefore
the closure Σ β of Σβ is also a forward invariant set.
5.7 Applications 179
Let us prove that the RDS (θ, ϕ) is nonexpansive with respect to the
!d
l1 -norm defined by the formula |x|1 = i=1 |xi | for x = (x1 , . . . , xd ), i.e.
Indeed, let
x(t) = ϕ(t, ω)x, y(t) = ϕ(t, ω)y and z(t) = ϕ(t, ω)z ,
where z = x ∨ y ≡ sup{x, y}. It is clear that z(t) ≥ x(t) ∨ y(t) for all t > 0.
Therefore
d
|x(t) − y(t)|1 = (2 max{xi (t), yi (t)} − xi (t) − yi (t))
i=1
(5.90)
d
≤ (2zi (t) − xi (t) − yi (t)) .
i=1
for any
x, y ∈ R+ .
d
Thus we have (5.89)
If the matrix ∂x∂fi
j
(ω, x) is irreducible for every x 0 and ω ∈ Ω, then
by Theorem 5.2.1 the RDS (θ, ϕ) is strongly order-preserving in int Rd+ . In
this case the restriction (θ, ϕβ ) of (θ, ϕ) to Σβ is contractive for each β > 0.
Indeed, if x, y ∈ Σβ for some β > 0 and x = y, then z > x and z > y. Since
(θ, ϕ) is strongly order-preserving, the last relation implies that
Consequently z(t) > x(t) ∨ y(t) for t > 0. Therefore we have strict inequality
in (5.90) and obtain (5.89) with strict inequality provided that x, y ∈ Σβ
and x = y. Thus (θ, ϕβ ) is contractive for each β > 0. Proposition 1.7.1
implies that every set Σβ can contain a unique (up to indistinguishability)
equilibrium.
Let us attempt to prove the existence of these equilibria. Since Σ β is
a compact set, the RDS (θ, ϕβ ) possesses a random attractor Aβ ⊆ Σ β .
Lemma 3.4.1 implies that
uβ (ω) := lim ϕ(t, θ−t ω)wβ (θ−t ω) = sup ϕ(t, θ−t ω)wβ (θ−t ω)
t→+∞ t>0
It follows from the last relation that the RDS (θ, ϕ) possesses infinitely many
equilibria. Since x(t) = uβ (θt ω) solves (5.87), we have that
d
d
uβi (θt ω) = uβi (ω), t > 0, ω ∈ Ω .
i=1 i=1
Thus the random systems generated by (5.91) and (5.92) are conjugate. Since
P = P −1 is an order isomorphism, the RDS (θ, ϕ) is order-preserving with
respect to the cone Km if and only if (θ, ψ) is order-preserving with respect
to the cone Rd+ . Hence after simple calculations we find that the condition
∂fi (ω, x)
(−1)mi +mj ≥ 0, i = j, x = (x1 , . . . , xd ) ∈ D, ω ∈ Ω , (5.93)
∂xj
4
d 2
|x| ≤ 2 (ai − bi0 xi )x2i ≤ −|x|2 + C
dt i=1
are forward invariant for (θ, ϕ). The restriction (θ, ϕ1 ) of (θ, ϕ) to V1 is gen-
erated by the equations
for which the cooperativity condition (R4∗ ) holds. Hence (θ, ϕ1 ) is order-
preserving in V1 with respect to the standard cone. It follows from (5.95) and
from Proposition 5.4.3 that there exists a super-equilibrium w = (w1 , w2 , 0, 0)
with wi ≥ R for (θ, ϕ1 ) in V1 .
Similarly, the restriction (θ, ϕ2 ) of (θ, ϕ) to V2 is an order-preserving RDS
with respect to the standard cone and there exists a super-equilibrium w =
(0, 0, w3 , w4 ) with wi ≥ R for (θ, ϕ2 ) in V2 . We obviously have
and
ϕ(t, ω)w = (0, 0, ϕ2 (t, ω)(w3 , w4 )) ≥m (0, 0, w3 , w4 ) = w .
Thus w is a super-equilibrium and w is a sub-equilibrium for (θ, ϕ) with
respect to the cone K(0,0,1,1) . Moreover w ≥m w and the absorbing set B
belongs to the interval {x : w ≤m x ≤m w̄}. Consequently by Theorem 3.6.2
the attractor A(ω) possesses the property
where u(ω) and u(ω) are equilibria. It is also easy to find that u(ω) ∈ V1 and
u(ω) ∈ V2 . Moreover u (resp. u) is globally asymptotically stable from above
in V1 (resp. V2 ) with respect to the standard cone.
6. Cooperative Stochastic Differential
Equations
where Wt (ω) = (Wt1 (ω), . . . , Wtm (ω)) is a Wiener process with values in Rm
and two-sided time R, m ≥ 1. Below we denote by θ the metric dynamical
system corresponding to this process (see Example 1.1.7).
In this chapter our main assumptions are follows:
(S1) every function fi : Rd+ → R belongs to the class Cb1,δ (Rd+ ) (see Def-
inition 2.4.1), i.e. fi (x) is a continuously differentiable function, with
derivatives bounded and globally δ-Hölder continuous:
∂f
i ∂fi
(x) − (y) ≤ C|x − y|δ , 0 < δ ≤ 1, i, j = 1, . . . , d ;
∂xj ∂xj
fi (x) ≤ fi (y), i = 1, . . . , d ,
∂fi (x)
≥ 0, i = j, x ∈ Rd+ .
∂xj
Proof. Let f˜(x) = (f˜1 (x), . . . , f˜d (x)) be a function from Rd into itself such
that f˜i (x) ∈ C 1,δ (Rd ) and f˜i (x) = fi (x) for all x ∈ Rd+ , i = 1, . . . , d. Let
σ̃ij (x) ∈ C 2,δ (R) be an extension of σij (x) from R+ to R such that σ̃ij · σ̃ij
1,δ
belongs to Cb (Rd+ ), i = 1, . . . , d, i = 1, . . . , m. It follows from Theorem 2.4.3
that the stochastic equations
m
dxi = f˜i (x1 , . . . , xd )dt + σ̃ij (xi ) ◦ dWtj , i = 1, . . . , d ,
j=1
generate a C 1 RDS in Rd . Property (S3) implies that the set D = Rd+ satisfies
the assumptions of Corollaries 2.5.1 and 2.5.2. Therefore from Corollary 2.5.2
there exists a unique (up to indistinguishability) continuous C 1 RDS (θ, ϕ)
in Rd+ generated by the system of Stratonovich SDEs (6.1) in the sense of
Theorem 2.4.3. 2
Theorem 5.2.1 implies that for every ε > 0 equations (6.3) generate an order-
preserving C 1 RDS (θ, ϕε ) in Rd+ . Therefore
β
l(ϕε (t, θ−t ω)y − ϕε (t, θ−t ω)x)dt ≥ 0, 0≤x≤y,
α
for all 0 ≤ α < β, where Ω ∗ is the θ-invariant set of full measure defined
in Remark 2.5.1. From this relation and from the continuity of the function
t → ϕ(t, θ−t ω)x (see Remark 2.4.1) we obtain that the inequality x ≤ y
implies ϕ(t, ω)x ≤ ϕ(t, ω)y for all t ≥ 0 and ω ∈ Ω ∗ . The invertibility of the
cocycle ϕ(t, ω) of the RDS generated by (6.1) (see [3, Theorem 2.3.32]) implies
that ϕ(t, ω)x < ϕ(t, ω)y for all 0 ≤ x < y, t ≥ 0 and ω ∈ Ω ∗ . Therefore after
redefining the cocycle ϕ by formula (2.50) (cf. Corollary 2.5.2) we obtain a
strictly order-preserving RDS.
If for some x0 > 0, t0 > 0 and ω ∈ Ω we have ϕ(t0 , ω)x0 = 0, then
ϕ(t0 , ω)y = 0 for all 0 ≤ y ≤ x0 , which is impossible because of the invert-
ibility of the cocycle ϕ(t, ω). Thus we have (6.2). 2
As for the random case (see Proposition 5.4.3) we have the following simple
condition for the existence of semi-equilibria for the RDS generated by (6.1).
Proposition 6.2.3. Let (S1)–(S3) be valid. Assume that there exists an el-
ement c = (c1 , . . . , cd ) in Rd+ such that σij (cj ) = 0 for each i = 1, . . . , d and
j = 1, . . . , m. If f (x) satisfies (S4) for all x ∈ [0, c] and
188 6. Cooperative Stochastic Differential Equations
Proof. Applying Proposition 5.4.3 to the RDS (θ, ϕε ) generated by the ap-
proximate equation (6.3) we obtain
for some µ > 0 and possesses the properties described in Lemma 2.5.1.
To present the main idea clearly we start with the simplest case of linear
diffusion terms.
Theorem 6.3.1. Assume that (S1)–(S3) hold. Let (θ, ϕ) be the RDS gener-
ated in Rd+ by (6.1). If σij (xi ) = sij · xi are linear functions, then (θ, ϕ) is
equivalent to the RDS (θ, ψ) generated in Rd+ by the RDE:
6.3 Conjugacy with Random Differential Equations 189
with
gi (ω, y1 , . . . , yd ) = esi (ω)−1 · fi (y1 · es1 (ω), . . . , yd · esd (ω)) + µyi zis (ω) . (6.6)
m
zis (ω) = sij zj (ω), i = 1, . . . , d , (6.7)
j=1
where the random variables zj (ω) are given by Lemma 2.5.1. Moreover we
have the relation
σij (xi ) = σi (xi ) · sij , σi (x) > 0, x > 0, σi (0) > 0 , (6.9)
where sij are constants. To obtain a theorem on conjugacy for this case we
need the following results.
Lemma 6.3.1. Suppose that Hi (x) is a primitive for σi (x)−1 on R+ \ {0}
and zis (ω) is defined by (6.7). Let Ti (ω, ·) : R+ → R+ be the random mapping
given by the formula
Then Ti (ω, y) ∈ C 3 (R+ \{0})∩C 1 (R+ ) for all ω ∈ Ω and the random mapping
T (ω, ·) : Rd → Rd+ defined by the relation
Ti (ω, y)
exp {−a|zis (ω)|} ≤ ≤ exp {a|zis (ω)|} , y>0, (6.11)
y
and
σi (y)
exp {−a|zis (ω)|} ≤ ≤ exp {a|zis (ω)|} , y>0, (6.12)
σi (Ti (ω, y))
and
Hi (yeaz ) ≤ z + Hi (y) ≤ Hi (y), y > 0, z < 0 .
From the monotonicity of Hi−1 we get (6.11).
Let ηi (x) = σi (Hi−1 (x)). Since dx d
ηi (x) = σi (Hi−1 (x)) · ηi (x), we obtain
Hi (y)
σi (y) ηi (Hi (y)) −1
= = exp σi (Hi (ξ))dξ .
σi (Ti (ω, y)) ηi (zis (ω) + Hi (y)) zis (ω)+Hi (y)
Lemma 6.3.2. If the functions fi and σij satisfy (S1)-(S3) and (6.9) holds,
then the functions
fi (T (ω, y))
gi (ω, y) = σi (yi ) · + µσi (yi ) · zis (ω) (6.13)
σi (Ti (ω, yi ))
Proof. This is rather simple and it relies on the properties the functions
Ti (ω, x) described in Lemma 6.3.1. We leave the details to the reader. 2
Theorem 6.3.2. Assume that the functions fi and σij satisfy (S1)-(S3) and
(6.9) holds. Let (θ, ϕ) be the RDS generated in Rd+ by (6.1). Then (θ, ϕ) is
equivalent to the RDS (θ, ψ) generated in Rd+ by the RDE (6.5) with gi (ω, y)
given by (6.13). Relation (6.8) holds with the diffeomorphism T (ω, ·) : Rd+ →
Rd+ given by the formula (6.10). Moreover we have the relations (6.2) and
Proof. Lemma 6.3.2 and Proposition 5.2.1 imply that the RDE (6.5) with
gi (ω, y) given by (6.13) generates a strictly positive RDS (θ, ψ) in Rd+ such
that
ψ(t, ω)intRd+ ⊂ intRd+ for any t ≥ 0, ω ∈ Ω . (6.15)
Let y(t, ω) = ψ(t, ω)T −1 (ω, x) with x ∈ intRd+ . Then we can apply Itô’s
formula (see Theorem 2.3.1) to xi (t, ω) := Ti (θt ω, yi (t, ω)) and find that
x(t, ω) = (x1 (t, ω), . . . , xd (t, ω)) satisfies (6.1) with initial data x(0, ω) = x for
every x ∈ intRd+ . Now using the continuous dependence of solutions to (6.1)
on initial data we obtain that T (θt ω, ψ(t, ω)T −1 (ω, x)) is also a solution to
this equation with initial data x from Rd+ . Therefore using (6.8) we can define
the perfect cocycle ϕ which satisfies in Rd+ the conclusions of Theorem 2.4.3.
2
Corollary 6.3.1. Assume that fi and σij satisfy (S1)-(S3) and (6.9) holds.
Let (θ, ϕ) be the RDS generated in Rd+ by (6.1). If f = (f1 , . . . , fd ) is strongly
positive, i.e. if
fi (x) > 0, i = 1, . . . , d , (6.16)
for all x ∈ Rd+ \ {0} of the form x = (x1 , . . . , xi−1 , 0, xi+1 , . . . , xd ), then (θ, ϕ)
is a strongly positive RDS, i.e. ϕ(t, ω)(Rd+ \ {0}) ⊂ intRd+ for any t ≥ 0 and
ω ∈ Ω.
Proof. In this case the function g given by (6.13) satisfies (R3∗ ). Therefore
we can apply Proposition 5.2.1. 2
192 6. Cooperative Stochastic Differential Equations
Analogously to the Random Comparison Theorem 5.3.1 we can prove the cor-
responding stochastic version (for the one-dimensional case see Ikeda/Wata-
nabe [57] and Karatzas/Shreve [62], for the Rd case see Ladde/Lakshmi-
kantham [75] and also Geiss/Manthey [46] and the references therein).
Here we consider the simplest case assuming (S1)–(S4) for one of the system.
We refer to Geiss/Manthey [46] for more general comparison theorems.
6.4 Stochastic Comparison Principle 193
implies that
ϕ(t, ω)x ≤ ψ(t, ω)x for all t > 0, ω ∈ Ω and x ∈ Rd+ ; (6.21)
(ii) if
fi (x) ≥ gi (x) for all x ∈ Rd+ , (6.22)
then
ϕ(t, ω)x ≥ ψ(t, ω)x for all t > 0, ω ∈ Ω and x ∈ Rd+ . (6.23)
Proof. Under the conditions listed above, equations (6.19) generate an RDS
in Rd+ (see Proposition 6.2.1). Let us consider together with (6.3) the system
of random differential equations
m
ẋi = gi (x1 , . . . , xd ) + σij (xi ) · ηεj (θt ω), i = 1, . . . , d , (6.24)
j=1
where ηεj (θt ω) is defined as in (6.3). By Proposition 5.2.1 both equations (6.3)
and (6.24) generate RDS in Rd+ . Let ϕε and ψ ε be the corresponding cocycles.
Let (6.20) hold. The random comparison principle (see Corollary 5.3.1(i))
implies that ϕε (t, ω)x ≤ ψ ε (t, ω)x for all t > 0, ω ∈ Ω and x ∈ Rd+ . Hence
β
l(ψ ε (t, θ−t ω)x − ϕε (t, θ−t ω)x)dt ≥ 0, x ≥ 0, ω ∈ Ω ,
α
Remark 6.4.1. (i) If the diffusion coefficients σij (x) satisfy (6.9), then using
Theorem 6.3.2 we can also prove assertions which are similar to Corollary
5.3.1 and Theorem 5.3.2 for RDS generated by stochastic equations.
(ii) The result of this section remains true if we interpret the equations
(6.1) in the Itô sense. The point is that by Theorem 2.4.2 the system of Itô
stochastic equations
m
dxi = fi (x1 , . . . , xd )dt + σij (xi ) · dWtj , i = 1, . . . , d . (6.25)
j=1
Now we give a result on the existence of equilibria and attractors for the
stochastic systems considered. As in the random case we note that under
assumptions (S1)–(S4) Proposition 6.2.1 implies that the element x ≡ 0 is a
sub-equilibrium for the RDS (θ, ϕ) generated by (6.1) in Rd+ .
Throughout this section we assume that the diffusion terms in (6.1) have
the following particular form:
and
d
∂V (x)
· σi (xi ) = γ · V (x) for all x ∈ intR+
d
, (6.29)
i=1
∂xi
where α > 0, β > 0 and γ ∈ R are constants. Let
0
R(ω) := β · exp{ατ − γWτ(s) (ω)} dτ ,
−∞
(s) !m
where Wt = j=1 sj Wtj . Let (θ, ϕ) be the RDS in Rd+ generated by (6.1).
Then the random set
B(ω) := x ∈ Rd+ : V (x) ≤ 2β/α + 2R(ω)
absorbs every deterministic bounded set, i.e. for any bounded set B from Rd+
there exists t0 = t0 (ω, B) > 0 such that ϕ(t, ω)B ⊂ B(θt ω) for t ≥ t0 . If we
additionally assume that
where aj , αj , bj are positive constants, then the RDS (θ, ϕ) possesses a random
attractor A(ω) in the universe D of all tempered subsets of Rd+ . This attractor
is measurable with respect to the past σ-algebra F− . If in addition (S4) holds,
then the attractor A(ω) is bounded from above and from below and there exist
maximal and minimal equilibria ū and u such that the random interval [u, ū]
contains the attractor as well as all other possible tempered equilibria. In
particular, if the equilibrium u is unique, then A = {u}.
d
Vε (t) ≤ −α + γ · ηε(s) (θt ω) Vε (t) + β .
dt
Therefore we have
t
Vε (t) ≤ V (x) exp −αt + γ ηε(s) (θτ ω)dτ
0
t
t
+β· exp −α(t − ξ) + γ ηε(s) (θτ ω)dτ dξ .
0 ξ
196 6. Cooperative Stochastic Differential Equations
Hence
0
V (ϕ (t, θ−t ω)x) ≤ V (x) exp −αt + γ
ε
ηε(s) (θτ ω)dτ
−t
0
0 (6.32)
+β· exp αξ + γ ηε(s) (θτ ω)dτ dξ .
−t ξ
d (s),ε d j,ε m
ηε(s) (θt ω) = Wt (ω) := sj Wt (ω) ,
dt j=1
dt
Since τ → Wξj (θτ ω) is a continuous function for every ξ and ω (see (2.25)),
we have that
0
m
(s)
lim ηε(s) (θτ ω)dτ = − sj Wξj (ω) = −Wξ (ω) . (6.33)
ε→0 ξ j=1
Let Ω ∗ be a θ-invariant set of full measure such that (2.46) holds. As in the
proof of Corollary 2.5.1 (cf.(2.49)) we conclude that for any ω ∈ Ω ∗ and
x ∈ Rd+ there exists εk → 0 such that
ϕεk (t, θ−t ω)x → ϕ(t, θ−t ω)x for almost all t > 0 .
(s)
V (ϕ(t, θ−t ω)x) ≤ V (x) exp −αt − γW−t (ω)
0
(6.34)
(s)
+β· exp αξ − γWξ (ω) dξ
−t
for almost all t > 0. Since t → ϕ(t, θ−t ω)x is continuous (see Remark 2.4.1),
we have inequality (6.34) for all ω ∈ Ω ∗ , t > 0 and x ∈ Rd+ which implies
that
(s)
V (ϕ(t, θ−t ω)x) ≤ V (x) exp −αt − γW−t (ω) + R(ω) . (6.35)
β
V (ϕ(t, θ−t ω)x) ≤ V (x)e−αt + 1 − e−αt , t > 0, ω ∈ Ω ∗ . (6.36)
α
Inequalities (6.35) and (6.36) imply that the random set B(ω) absorbs every
deterministic bounded set from Rd+ .
It is clear that R(ω) is a tempered random variable. Therefore under
condition (6.30) we have B ∈ D. From (6.35) and (6.36) we also have that B is
D-absorbing for the RDS (θ, ϕ) (cf. the proof of Proposition 1.4.1). Therefore
we can apply Theorem 1.8.1 on the existence of random attractors and assert
that the RDS (θ, ϕ) generated in the space Rd+ by problem (6.1) possesses
a random global D-attractor A(ω). Since R(ω) is F− -measurable, it follows
from (1.44) that A(ω) is F− -measurable. The existence of the maximal and
minimal equilibria ū and u and their properties under condition (S4) follow
from Theorem 3.6.2. 2
d
d
xκ−1
i · fi (x) ≤ −α · xκi + β, x ∈ Rd+ . (6.37)
i=1 i=1
Then the RDS (θ, ϕ) generated by (6.1) possesses a random attractor A(ω)
in the universe D of all tempered subsets of Rd+ and all the conclusions of
Theorem 6.5.1 hold.
!d
Proof. It is easy to see that the function V (x) = i=1 xκi satisfies all the
hypotheses of Theorem 6.5.1. 2
Corollary 6.5.2. Let assumptions (S1)–(S3) and (6.27) hold. Assume ad-
ditionally that
and
σi (x)
= λi + O(x−γi ) when x → ∞, i = 1, . . . , d , (6.38)
x
where λi > 0 and γi > 0 are constants, and there exist positive numbers α, β
and 0 ≤ κ < (min λi ) · (max λi )−1 such that
198 6. Cooperative Stochastic Differential Equations
Then the RDS (θ, ϕ) generated by (6.1) possesses a random attractor A(ω)
in the universe D of all tempered subsets of Rd+ and all the conclusions of
Theorem 6.5.1 hold.
Proof. Let
d
x
dξ
V (x) = Vi (xi ) with Vi (x) = exp δ ,
i=1 1 σi (ξ)
d
fi (x1 . . . , xd )
f (x), ∇V (x) = δ Vi (xi ) ·
i=1
σi (xi )
d
xi d
gi (x1 . . . , xd )
≤ −αδ · Vi (xi ) + δ Vi (xi ) · . (6.41)
i=1
σi (xi ) i=1
σi (xi )
αδ d
gi (x1 . . . , xd )
f (x), ∇V (x) ≤ − · V (x) + δ Vi (xi ) · +C . (6.42)
2λ i=1
σi (xi )
and
δ/σi (0) δ/σi (0)
C1 · xi ≤ Vi (xi ) ≤ C2 · xi if 0 < xi < 1, i = 1, . . . , d ,
6.6 One-Dimensional Stochastic Equations 199
These inequalities imply that under the condition δ ≥ max σi (0) we have
Vi (xi ) ( ) ( )
≤ C · 1 + V (x)1−λi /δ and xi ≤ C · 1 + V (x)λi /δ ,
σi (xi )
d
gi (x1 , . . . , xd ) ( ∗
)
Vi (xi ) · ≤ C · 1 + V (x)κ , (6.43)
i=1
σi (xi )
and
200 6. Cooperative Stochastic Differential Equations
Proof. We suppose that σ (0) > 0 and σ(x) > 0 if x > 0 for the definiteness.
We first assume that −α ≤ f (x) ≤ ax + b for some α > 0. Denote by
χN (z) a function from C 2 (R) with the properties (i) χN (z) = N + 1/2 for
z ≥ N + 1, (ii) χN (z) = z for z ∈ [−∞, N ] and (iii) 0 ≤ χ N (z) ≤ 1 for
all z ∈ R. Then fN (x) := χN (f (x)) ∈ Cb1,δ (R+ ). Since condition (S4) holds
automatically for the one-dimensional case, Proposition 6.2.2 implies that the
equation
dx = fN (x)dt + σ(x) ◦ dWt (6.47)
generates a strictly order-preserving C 1 RDS (θ, ϕN ) in R+ . Since fN (x) ≤
fN +1 (x) ≤ ax + b, Comparison Theorem 6.4.1 implies that
where (θ, ϕ̄) is the RDS generated by (6.44) with f (x) = ax + b. Relation
(6.48) implies that the limit
exists. By Theorem 6.3.2 the RDS (θ, ϕN ) is equivalent to the RDS (θ, ψN )
generated by the RDE
fN (T (θt ω, y))
ẏ = σ(y) · + µσ(y) · z(θt ω) ,
σ(T (θt ω, y))
f (T (θt ω, y))
ẏ = σ(y) · + µσ(y) · z(θt ω), y0 ∈ R + .
σ(T (θt ω, y))
6.6 One-Dimensional Stochastic Equations 201
Remark 6.6.1. Using Feller’s test for non-explosion (see, e.g., Karatzas/
Shreve [62, p.348]) it is also possible to give the sufficient and necessary
conditions on the functions f (x) ∈ C 1,δ (R+ ) and σ(x) ∈ Cb2,δ (R+ ) for gener-
ation of a C 1 RDS by equation (6.44) (cf. Arnold [3, p.96] and Kunita [74,
p.181-184]).
f (x) |σ(x)|
lim sup <0 and = λ + O(x−γ ), x → ∞ , (6.51)
x→∞ x x
where λ > 0 and γ > 0 are constants. Then the RDS (θ, ϕ) generated by (6.44)
possesses a random attractor A(ω) in the universe D of all tempered subsets
of R+ . This attractor is measurable with respect to the past σ-algebra F− .
Moreover A(ω) = [u(ω), ū(ω)], where ū and u are F− -measurable tempered
equilibria such that 0 ≤ u(ω) ≤ ū(ω).
We note that as in the random case (see Example 5.6.1) the cocycle ϕ can
be represented in the form
x exp {αt + σWt (ω)}
ϕ(t, ω, x) = 1/N
t
1 + βN xN 0 exp {N (ατ + σWτ (ω))} dτ
for x > 0 and ϕ(t, ω, 0) = 0. Therefore if α < 0, then A(ω) = {0}. In the case
α > 0 we have A(ω) = [0, uα,β,N (ω)], where
0 − N1
uα,β,N (ω) := βN exp {N (ατ + σWτ (ω))} dτ . (6.53)
−∞
Theorem 6.6.1. Assume that hypotheses (6.45), (6.46) and (6.51) hold. Let
A(ω) be the random attractor in the universe D of all tempered subsets of R+
for the RDS generated by (6.44).
(i) If f (0) = 0 and m((0, 1]) = ∞, then A(ω) = {0} almost surely.
(ii) If f (0) = 0 and m((0, 1]) < ∞, then A(ω) = [0, u(ω)] for some F− -
measurable equilibrium u(ω) such that u(ω) > 0 almost surely. There
are no other (up to indistinguishability) F− -measurable equilibria in the
set (0, u(ω)] and
m(B)
lim P{ω : ϕ(t, ω)x ∈ B} = P{ω : u(ω) ∈ B} ≡ (6.57)
t→∞ m(R+ )
for any ε > 0. Indeed, choose 0 < γ < ε such that ε·m((γ, ε]) > m([ε, ∞)) and
consider a function f˜ ∈ C 1,δ (R+ ) possessing the properties (a) f˜(x) ≥ f (x)
for x ∈ R+ , (b) f˜(x) = f (x) if x ≥ γ and (c) m̃((0, 1]) < ∞, where m̃(dx) is
defined by (6.55) with f˜ instead of f . It is easy to construct a such function
f˜ choosing f˜(x) = k · x with k > 0 at a vicinity of 0. Let (θ, ϕ̃) be the RDS
generated by (6.44) with f˜ instead of f . Comparison Theorem 6.4.1 implies
that ϕ(t, ω)x ≤ ϕ̃(t, ω)x for all t ∈ R, ω ∈ Ω and x > 0. We also have
s̃(0) = −∞ and s̃(+∞) = ∞, where s̃(x) is the scale function for the RDS
(θ, ϕ̃). Therefore by [78, Theorem 7, Chap.4] we have
Thus
for any random variable v(ω) and N ∈ N, it follows from (6.58) that
for every {v(ω)} ∈ D and ε > 0. If A(ω) = {0} is not true almost surely,
then by Corollary 6.6.1 there exists an equilibrium u(ω) ≥ 0 such that P{ω :
u(ω) ≥ ε} > 0 for some ε > 0. However the relation
for any x > 0 and B ∈ B(R+ ) (see, e.g., Mandl [78, Theorem 7, Chap.4]). In
particular, this implies that A(ω) = [0, ū(ω)], where ū(ω) is an F− -measurable
equilibrium such that ū(ω) > 0 almost surely. Indeed, if ū(ω) = 0 on a set of
positive measure, then by Lemma 3.5.1 u(ω) = 0 almost surely. In this case
6.6 One-Dimensional Stochastic Equations 205
A(ω) = {0} almost surely and therefore P{ω : ϕ(t, ω)x ≥ δ} → 0 for any
x > 0 and δ > 0, which is impossible because ([δ, ∞)) > 0.
The uniqueness of the stationary measure and Theorem 1.10.1 imply
that
Remark 6.6.2. Assume that f (x) satisfies a Lipschitz condition on each com-
pact subset of R+ , f (0) = 0 and |f (x)| ≤ ax+b for some positive a and b. Let
σ(x) ∈ C 2 with bounded first and second derivatives and σ(0) = 0. It was
proved by Scheutzow [91] that A(ω) = {0} is a random attractor for the
RDS generated by (6.44) if and only if m((0, 1]) = ∞ and m([1, ∞)) < ∞.
β1 xN +1 ≤ g(x) ≤ β2 xN +1 , x ≥ 0,
where β1 , β2 > 0 and N > 0. Proposition 6.6.1 and Corollary 6.6.1 are applied
here. Therefore (6.61) generates a strictly order-preserving RDS (θ, ϕ) in
R+ which has a random attractor A(ω) = [0, u(ω)] in the universe D of all
tempered subsets of R+ . Using the comparison principle it is easy to see that
where (θ, ϕα,β ) is the RDS generated by (6.52). Therefore A(ω) = {0} if α < 0
and A(ω) = [0, u(ω)] when α > 0, where the F− -measurable equilibrium u(ω)
satisfies the inequality
Here uα,β,N (ω) is given by (6.53). It follows from (6.57) that u(ω) attracts
every trajectory ϕ(t, ω)x with x > 0 with respect to convergence in distribu-
tion. However using properties of the RDS generated by (6.52) and relation
(6.62) we can prove that
206 6. Cooperative Stochastic Differential Equations
almost surely. Indeed, (6.54) and (6.62) imply that the omega-limit set Γx (ω)
emanating from x > 0 (see Definition 1.6.1) possesses the property
Thus by Lemma 3.4.1 and Remark 3.4.2(ii) u(ω) := inf Γx (ω) and u(ω) :=
sup Γx (ω) are F− -measurable equilibria such that
It is clear that u(ω), u(ω) ∈ (0, u(ω)]. Hence, applying Theorem 6.6.1(ii) we
obtain that u(ω) = u(ω) = u(ω) almost surely. Thus Γx (ω) = u(ω) almost
surely and (6.63) holds. As in Arnold [3, Theorem 9.3.3] it is also possible
to prove the convergence property (6.63) for a random variable x = x(ω) > 0
such that x(ω) and x(ω)−1 are tempered.
and
σ(x) ∈ Cb2,δ ([l, r]), σ(l) = σ(r) = 0, |σ(x)| > 0 if l < x < r . (6.65)
Here above δ ∈ (0, 1]. Under these conditions by Proposition 6.2.3 equation
(6.44) generates a strictly order-preserving RDS (θ, ϕ) in the interval [l, r].
As above we introduce the speed measure m(dx) on [l, r] and the scale
function s : [l, r] → R ∪ {±∞} by the formulas (see, e.g., Karatzas/Shre-
ve [62])
x
f (ξ) dx
m(A) = exp 2 2
dξ , A ∈ B([l, r]) , (6.66)
A c σ (ξ) |σ(x)|
and x
y
f (ξ) dy
s(x) = exp −2 dξ , x ∈ [l, r] ,
c c σ 2 (ξ) |σ(y)|
where c is a fixed point from (l, r).
Similar to Theorem 6.6.1 we can prove the following result.
Theorem 6.6.2. Assume that (6.64) and (6.65) hold. Let A(ω) be the ran-
dom attractor for the RDS (θ, ϕ) generated by (6.44) in the interval [l, r].
(i) If f (l) = f (r) = 0, then A(ω) = [l, r].
(ii) If f (l) = 0 and f (r) < 0, then
6.6 One-Dimensional Stochastic Equations 207
m(B)
lim P{ω : ϕ(t, ω)x ∈ B} = P{ω : u(ω) ∈ B} ≡ (6.67)
t→∞ m([l, r])
Proof. Assertion (i) follows from the property ϕ(t, ω)[l, r] = [l, r] for all t > 0
and ω ∈ Ω. To prove the other assertions of the theorem we note that f (l) > 0
(resp. f (r) < 0) implies that m((l, c]) < ∞ (resp. m([c, r)) < ∞). Therefore
we can apply the same argument as in the proof of Theorem 6.6.1. 2
Now we consider the case f (l) = f (r) = 0 in details. We are interested in the
description of the long-time behaviour of trajectories inside the attractor.
The following assertion is an almost direct consequence of the well-known
theorems on the boundary behaviour of one-dimensional diffusion processes
(see, e.g., Ikeda/Watanabe [57] or Karatzas/Shreve [62]).
Theorem 6.6.3. Assume that (6.64) and (6.65) hold. Let f (l) = f (r) = 0.
Denote by (θ, ϕ) the RDS generated by (6.44) in the interval [l, r].
(i) If m([l, r]) < ∞, then there exists an F− -measurable equilibrium u(ω)
such that l < u(ω) < r almost surely and (6.67) holds for all x ∈ (l, r)
and B ∈ B([l, r]). Moreover
P ω : lim inf ϕ(t, ω)x = l = P ω : lim sup ϕ(t, ω)x = r = 1
t→∞ t→∞
(6.68)
for any x ∈ (l, r) and the process ϕ(t, ω)x is recurrent, i.e. for any
y ∈ (l, r) we have
Proof. (i) As in the proof of Theorem 6.6.1(ii) it is easy to see that (x) :=
m((l,x]) r
m([l,r]) solves
the stationary Fokker-Plank equation on (l, r), l (x)dx = 1,
and (B) = B (x)dx, B ∈ B((l, r)), is a unique ergodic stationary measure.
Therefore using Theorem 2.3.45 Arnold [3] we can prove that the limit
r r
h(x)µω (dx) := lim h(ϕ(t, θ−t ω)x)(x)dx
l t→∞ l
r r
h(x)µθt ω (dx) = h(ϕ(t, ω)x)µθt ω (dx) for all ω ∈ Ω .
l l
we have that l < u(ω) < r almost surely and (6.67) holds for x ∈ (l, r).
Since the property m([l, r]) < ∞ implies that s(l) = −∞ and s(r) = ∞
(see, e.g., Crauel et al. [38, Lemma 2.4]), we can apply Proposition 5.5.22
Karatzas/Shreve [62] (see also Ikeda/Watanabe [57, Theorem 6.3.1])
to obtain (6.68) and (6.69).
To prove (ii) and (iii) we can repeat with a slight modification the ar-
gument given in the proof of Theorem 6.6.1(i) and apply Proposition 5.5.22
Karatzas/Shreve [62].
(iv) To prove (6.72) we consider the process
* +
y(t, ω; z) := G ϕ(t, ω) G−1 (z) , z ∈ R ,
where G(x) is a primitive for [σ(x)]−1 on the interval (l, r). This process
solves the SDE
f (G−1 (y))
dy = · dt + dWt , y(0) = z . (6.74)
σ(G−1 (y))
Theorem 6.6.3 implies the following result on a random attractor with respect
to the convergence in probability (see Ochs [86] for the theory of attractors
based on this type of convergence).
Corollary 6.6.2. Assume that (6.64) and (6.65) hold. Let f (l) = f (r) = 0
and m((l, r)) = ∞. Then the two-point set A := {l, r} is a weak point random
attractor for the RDS (θ, ϕ) generated by (6.44) in the interval [l, r]. This
means that (i) ϕ(t, ω)A = A for all t ≥ 0 and ω ∈ Ω; (ii) ϕ(t, ω)x converges
to A in probability for every x ∈ [l, r], i.e.
for any ε > 0; (iii) A is a minimal set possessing the properties (i) and (ii).
for some bounded interval [l, r] ⊂ R and δ ∈ (0, 1]. In this case the speed
measure m has the form
b
dx
m([a, b]) = , [a, b] ⊂ [l, r], if α = 0 ,
a σ(x)
and
,
a -
b
1 dξ dξ
m([a, b]) = · exp 2α − exp 2α , if α = 0 ,
2α c σ(ξ) c σ(ξ)
where c is a fixed point from (l, r). For the scale function s we have the
representation x
dξ
s(x) = , x ∈ [l, r], if α = 0 ,
c σ(ξ)
and
x
1 dξ
s(x) = · 1 − exp −2α , if α = 0 .
2α c σ(ξ)
where G(x) is a primitive for [σ(x)]−1 on the interval (l, r), and using the law
of the iterated logarithm for the one-dimensional Wiener process (see, e.g.,
Friedman [45, p.40]) it is easy to prove that in the case α = 0 the interval
[l, r] is the pull back omega-limit set for the trajectory emanating from any
point x ∈ (l, r).
If in addition we assume that σ (l) > 0 and σ (r) < 0, then relying on
(6.77) after a simple calculation (see Chueshov/Vuillermot [25]) we find
that
1
lim log |ϕ(t, ω)x − l| = ασ (l), P − a.s., x ∈ (l, r), α<0,
t→∞ t
and
1
lim log |ϕ(t, ω)x − r| = ασ (r), P − a.s., x ∈ (l, r), α>0.
t→∞ t
To interpret these results we recall (see, e.g., Khasminskii [64]) that (a) an
equilibrium u (either l or r) is said to be stable in probability if the relation
lim P ω ∈ Ω : sup |ϕ(t, ω)x − u| > ε = 0 (6.78)
x→u t>0
for every x ∈ (l, r); (c) u is unstable in probability if (6.78) does not hold.
Thus in this example, on the one hand, we have the weak point attractor
A = {l, r} which does not depend on α, on the other hand, we observe the
following bifurcation picture (with respect to the parameter α):
(i) if α < 0, l is globally asymptotically stable, whereas r is unstable in
probability;
(ii) if α = 0, both l and r are unstable in probability;
(iii) if α > 0, r is globally asymptotically stable, whereas l is unstable in
probability.
212 6. Cooperative Stochastic Differential Equations
Example 6.6.4. Consider now an Itô SDE on the interval (l, r) of the same
form as (6.75):
dx = ασ(x)dt + σ(x)dWt , (6.79)
where α ∈ R is a parameter and σ(x) satisfies (6.76). Assume that σ (l) > 0
and σ (r) < 0 hold. This equation can be written as a Stratonovich SDE of
the form
1
dx = α − σ (x) σ(x)dt + σ(x) ◦ dWt ,
2
The speed measure m and the scale function s for this equation are repre-
sented by the formulas
x
dξ dx
m(A) = σ(c) exp 2α 2
, A ∈ B([l, r]) ,
A c σ(ξ) σ (x)
and x
y
1 dξ
s(x) = exp −2α dy , x ∈ [l, r] ,
σ(c) c c σ(ξ)
where c is a fixed point from (l, r). It is easy to see that m((l, r)) = ∞ for
every α ∈ R. Therefore, as in Example 6.6.3, Corollary 6.6.2 implies that the
two-point set {l, r} is a weak point random attractor for the RDS in [l, r]
generated by (6.79) for every α ∈ R. Applying Theorem 6.6.3 we obtain the
following more precise information on the dynamics:
(i) if α ≤ 12 σ (r), then (6.70) holds;
(ii) if 12 σ (r) < α < 12 σ (l), then we have (6.73);
(iii) if α ≥ 12 σ (l), then (6.71) holds.
Thus we observe that in contrast with the RDS generated by (6.75) the long-
time behaviour of ϕ(t, ω)x is characterized by the absence of any recurrent
and oscillatory behavior for all values of α. Moreover, it is also possible to
prove (see Chueshov/Vuillermot [26]) that the bifurcation picture in this
example is following:
(i) if α ≤ 12 σ (r), l is globally asymptotically stable, whereas r is unstable
in probability;
6.6 One-Dimensional Stochastic Equations 213
(ii) if 12 σ (r) < α < 12 σ (l), both l and r are stable in probability;
(iii) If α ≥ 12 σ (l), r is globally asymptotically stable, whereas l is unstable
in probability.
Thus the exchange of stability between the equilibria l and r, when the pa-
rameter α varies from minus infinity to plus infinity, is slower (softer) in this
example in contrast with the picture that we can see in Example 6.6.3. How-
ever in both examples the weak point random attractor A = {l, r} is the
same for all α ∈ R.
The following example shows that the idea outlined in Sect.5.6 can be also
used in the stochastic case.
Example 6.6.5. Consider the Stratonovich SDE
dx = (α1 sin x + α2 (1 − cos x)) dt
(6.80)
+ σ1 sin x ◦ dWt1 + σ2 (1 − cos x) ◦ dWt2 ,
where αi and σi are parameters and Wt = (Wt1 , Wt2 ) is a Wiener process
in R2 . This equation generates an RDS in R. However, as in Example 5.6.3,
it is natural to consider equation (6.80) on the unit circle C which is inter-
preted as the interval [0, 2π] with identified end-points. Using Itô’s formula
for the Stratonovich integrals (see Theorem 2.3.1) it is easy to see that (6.80)
generates an RDS (θ, ϕ) in C with the cocycle
ϕ(t, ω)x = 2arccot (−ψ(t, ω)[− cot(x/2)]) , 0 < x < 2π ,
where ψ(t, ω) is the cocycle in R generated by the affine Stratonovich equation
dy = (−α1 y + α2 ) dt − σ1 y ◦ dWt1 + σ2 ◦ dWt2 .
In the case α2 = σ1 = 0 we obtain the SDE
dy = −α1 y dt + σ2 dWt2 .
Therefore the equation
dx = α1 sin x dt + σ2 (1 − cos x) ◦ dWt2
generates an RDS (θ, ϕ̃) with the cocycle
t
−α1 t x −α1 (t−s) 2
ϕ̃(t, ω)x = 2arccot e · cot − σ2 e dWs , 0 < x < 2π .
2 0
This formula implies that for any α1 = 0 the RDS (θ, ϕ̃) has two equilibria
u0 ≡ 0 and u(ω) ∈ (0, 2π) which stability properties can be easily derived
from the results presented in Example 2.4.4. The case α1 = 0 is covered by
Example 6.6.3 with α = 0.
Remark 6.6.3. We also note that it is possible to give examples which show
that all the cases listed in Theorem 6.6.3 are really occur. We refer to
Scheutzow [91], where the corresponding examples are presented in the
case [l, r) = R+ .
214 6. Cooperative Stochastic Differential Equations
We start with conditions that ensure that the RDS generated by the system
of stochastic cooperative differential equations (6.1) is sublinear.
Lemma 6.7.1. Assume that conditions (S1), (S3) and (S4) hold. Let σij (x)
be linear functions, i.e.
where σij are constants. If the function f (·) is a sublinear mapping from Rd+
into Rd , i.e. if
then the RDS (θ, ϕ) generated by (6.1) is sublinear. Moreover (θ, ϕ) is strongly
sublinear if one of the following conditions is satisfied:
(i) f is a strongly sublinear mapping, i.e. λf (x) f (λx) for all 0 < λ < 1
and x ∈ intR d
+ ;
Proof. The function xλ (t) = λ · ϕ(t, ω)x is the solution to the problem
m
dxλi (t) = fiλ (xλ (t))dt + σij · xλi (t) ◦ dWtj , i = 1, . . . , d ,
j=1
with initial data xλ (0) = λ · x, where f λ (x) = λf (λ−1 x). From (6.82) we
have f λ (x) ≤ f (x). Therefore the stochastic comparison principle (see The-
orem 6.4.1(i)) gives
To prove that (θ, ϕ) is strongly sublinear under the conditions either (i) or
(ii) we note that by Theorem 6.3.1 (θ, ϕ) is equivalent to the RDS (θ, ψ)
generated by (6.5) with g given by (6.6). Therefore we can use Lemma 5.5.1.
2
We note that Lemma 6.7.1 remains true if we understand (6.1) in the Itô
sense (cf. Remark 6.4.1). It is also easy to see that the RDS generated by
equation (6.52) gives an example of a strongly sublinear RDS.
Now we apply the results presented Chap.4 for sublinear systems. The
application of Corollary 4.3.1 gives the following assertion.
6.7 Stochastic Equations with Concavity Properties 215
Theorem 6.7.1. Assume that conditions (S1), (S3), (S4) and (6.81) hold
and f (0) 0. Let the assumption either (i) or (ii) of Lemma 6.7.1 be valid.
Then either
(i) for any x ∈ Rd+ we have |ϕ(t, θ−t ω)x| → ∞ almost surely as t → ∞
or
(ii) there exists a unique almost equilibrium u(ω) 0 defined on a θ-
invariant set Ω ∗ of full measure such that
for any random variable v(ω) possessing the property 0 ≤ v(ω) ≤ λu(ω) for
all ω ∈ Ω ∗ and for some nonrandom λ > 0.
Proof. The property f (0) 0 implies that the function g(ω, x) given by
formula (6.6) satisfies (5.36). Therefore it follows from Proposition 5.4.1 and
Theorem 6.3.1 that ϕ(t, ω)0 0 for all t > 0 and ω ∈ Ω. Thus (θ, ϕ) is
strongly positive. It is also clear that any finite-dimensional RDS is condi-
tionally compact. Therefore we can apply Corollary 4.3.1. 2
Now we prove a stochastic version of the trichtomy theorem (for the random
case see Theorem 5.5.2).
Theorem 6.7.2 (Limit Set Trichotomy). Let conditions (S1), (S3), (S4),
(6.81) and either (i) or (ii) of Lemma 6.7.1 hold. Assume that the coefficients
σij ≡ σj independent of i and there exist positive constants a and b such that
(σ) !m j
where ν is a positive constant and Wt = j=1 σj Wt . Denote by Cη =
Cη (ω) the collection of random variables w : Ω → R+ possessing the property
d
for some nonrandom number α ≥ 1. Let (θ, ϕ) be the RDS generated by the
equation
(σ)
dxi (t) = fi (x1 (t), . . . , xd (t))dt + xi (t) ◦ dWt , i = 1, . . . , d .
Then any orbit of (θ, ϕ) emanating from a ∈ Cη does not leave Cη , i.e.
lim ϕ(t, θ−t ω)b(θ−t ω) = u(ω) for almost all ω∈Ω. (6.86)
t→+∞
where y (1) (t) and y (2) (t) are solutions to the problems
(σ)
dy (1) (t) = −a · y (1) (t) · dt + y (1) (t) ◦ dWt , y (1) (0) = η(ω)e , (6.88)
and
(σ)
dy (2) (t) = b·(1+|y (2) (t)|1 )·e·dt+y (2) (t)◦dWt , y (2) (0) = η(ω)e . (6.89)
(1) (σ)
yi (t) = η(ω) · exp −at + Wt , i = 1, . . . , d . (6.90)
Therefore
(2) (σ)
yi (t) ≤ |y (2) (t)|1 = d · η(ω) · exp bd · t + Wt
t
(6.92)
(σ)
+ bd · exp bd · (t − τ ) + Wt − Wτ(σ) dτ .
0
6.7 Stochastic Equations with Concavity Properties 217
This implies the invariance of Cη (ω). Thus we can apply Theorem 4.4.1 and
Corollary 4.4.1. 2
In order to show that all three cases of the limit set trichotomy can actually
occur we consider the following example.
Example 6.7.1. Consider the Stratonovich stochastic differential equation on
X = R+ given by
where
x
f (x) = αx + , α∈R,
1+x
which is strongly sublinear for any α ∈ R, hence by Lemma 6.7.1 the RDS
(θ, ϕ) generated by (6.93) is strongly sublinear for any α, σ ∈ R. Theo-
rem 6.7.2 is applicable here for any α and σ.
The point x = 0 is always an equilibrium.
Consider first the case α > 0. Since f (x) ≥ αx, the comparison principle
yields
ϕ(t, ω)b(ω) ≥ b(ω)eαt+σWt (ω) ,
hence
ϕ(t, θ−t ω)b(θ−t ω) ≥ b(θ−t ω)eαt−σW−t (ω) .
Consequently, for any initial random variable b such that 1/b(ω) is tempered,
the orbit γb of ϕ emanating from b is unbounded, in fact converges to infinity
with probability one.
Now consider the case −1 < α < 0. Then f (x) < 1 + αx and Corol-
lary 6.6.1 implies that (θ, ϕ) possesses a random attractor A(ω) = [0, u(ω)],
where u(ω) ≥ 0 is an F− -measurable tempered equilibrium. A simple calcula-
tion shows that the speed measure m given by (6.55) satisfies m((0, 1]) < ∞
if α > −1. Therefore by Theorem 6.6.1 u(ω) > 0 almost surely and it follows
from Theorem 4.2.2 that u(ω) is attractive in the part Cu .
Here we can also clearly see why the initial values have to be restricted
somehow, e.g. to those in Cu . Take, for example, a random variable a(ω) > 0
218 6. Cooperative Stochastic Differential Equations
which is not tempered, i.e. for which lim supt→∞ 1t log a(θ−t ω) = +∞. Non-
tempered random variables exist on any standard probability space and for
any ergodic and aperiodic θ Arnold/Cong/Oseledets [9, Lemma 8.6],
which includes to the present situation. Then a ∈ Cu and since
there exists α < 0 such that the right-hand side tends to +∞ along some
sequence tn → ∞. Hence the orbit emanating from a does not converge to u.
Finally, if α < −1 then ϕ is dominated by the linear cocycle generated by
dx = (α + 1)xdt + σx ◦ dWt , hence
almost surely. If b is even ε-slowly varying, i.e. if b(θt ω) ≤ b(ω)eε|t| for some
ε > 0 such that ε + (α + 1) < 0, then the orbit γb is bounded, but the
closure of γb contains elements (namely 0) which do not belong to any part
Cv ⊂ intR+ .
We can combine several of the equations (6.93) to produce more compli-
cated limit behaviour.
The following assertion gives conditions under which the RDS generated by
(6.1) is s-concave (see Definition 4.1.3).
Proposition 6.7.1. Assume that assumptions (S1), (S3), (S4) and (6.81)
are met and that the matrix Dx f (x) is irreducible in intRd+ . Let the function
f (x) be s-concave, i.e.
As in the random case (see Theorem 5.5.3) we can also prove the following
assertion.
Theorem 6.7.3. Let (S1), (S3), (S4) and (6.81) hold. Assume that the
fuction f (x) is s-concave and that the matrix Dx f (x) is irreducible for all
x ∈ intRd+ . If f (0) > 0, then either (a) for all v(ω) ≥ 0, the orbit γv ema-
nating from v is unbounded; or (b) there exists a unique equilibrium u 0
such for every v(ω) possessing the property 0 ≤ v(ω) ≤ α · u(ω) with some
6.8 Applications 219
6.8 Applications
for some constants a and b. Proposition 6.2.2 implies that equations (6.98)
and (6.99) generate a strictly order-preserving RDS (θ, ϕ) in the cone Rd+ .
To construct sub- and super-equilibria for (6.98) and (6.99) we consider the
following auxiliary affine problem
where Φ(t, ω) is the cocycle generated by (6.101) and (6.102) with b = 0 and
b = (b, 0, . . . , 0) ∈ Rd+ . If a is small enough, then the cocycle Φ(t, ω) has the
negative top Lyapunov exponent, i.e. there exists λ < 0 such that
for every ε > 0, where Rε (ω) > 0 is a tempered random variable and
Ω ∗ ∈ F is a θ-invariant set of full measure. We can suppose Ω ∗ = Ω (see Re-
mark 1.2.1(iii)). By Theorem 5.6.5 (Arnold [3]) the RDS (θ, ϕaff ) possesses
a unique tempered equilibrium
∞
v(ω) ≡ Φ(τ, θ−τ ω)b dτ, ω ∈ Ω .
0
for any D ∈ D. This means that for any µ > 1 the super-equilibrium µv(ω)
is absorbing for the RDS (θ, ϕ). On the other hand the affine RDS generated
by
dy1 (t) = (g(0) − α1 y1 (t))dt + σ1 · y1 (t) ◦ dWt1 ,
dyj (t) = (yj−1 (t) − αj yj (t))dt + σj · yj (t) ◦ dWtj , j = 2, . . . , d ,
dominates (θ, ϕ) from below. This system possesses a uniformly attracting
equilibrium w(ω) such that 0 ≤ w(ω) ≤ v(ω). It is easy to find for w(ω) =
(w1 (ω), . . . , wd (ω)) the recurrence formulae
0
1
w1 (ω) = g(0) eα1 t−σ1 Wt dt (6.103)
−∞
6.8 Applications 221
and
0
j
wj (ω) = wj−1 (θt ω)eαj t−σj Wt dt, j = 2, . . . , d . (6.104)
−∞
and
lim ϕ(t, θ−t ω)v(θ−t ω) = ū(ω)
t→+∞
for any tempered w(ω) and v(ω) such that w(ω) ≤ u(ω) and v(ω) ≥ ū(ω).
Assume in addition to (6.100) that g(0) > 0, g (x) > 0 for x > 0, and
g(x) is strictly sublinear, i.e. λg(x) < g(λx) for any 0 < λ < 1 and x > 0.
Lemma 6.7.1 implies that (θ, ϕ) is a strongly sublinear RDS. In this case
the sub-equilibrium w(ω) given by (6.103) and (6.104) is strongly positive.
Therefore we can apply Theorem 4.2.1 and Corollary 3.6.1 to prove that
the random attractor A(ω) is a one-point set consisting of a unique globally
asymptotically stable equilibrium.
We consider the following stochastic version of the problem (5.81) and (5.82):
. /
d
dxj = −αj xj + (pj − xj ) βji xi dt + σj xj (pj − xj ) ◦ dWtj . (6.105)
i=1
X = [0, p] = {u ∈ Rd : 0 ≤ u ≤ p} ,
where
The system in (6.106) is decoupled, hence for each j equation (6.106) gener-
ates a strictly order-preserving RDS (θ, ψj ) in the one-dimensional interval
[0, pj ]. Theorem 6.4.1 implies that the direct product of these systems domi-
nates the system (θ, ϕ) generated by (6.105) from below.
Under the condition αj < pj βjj the speed measure (6.66) for problem
(6.106) on [0, pj ] possesses the property m((0, cj ]) < ∞ for any cj ∈ (0, pj ).
Therefore by Theorem 6.6.2(ii-b) there exists unique F− -measurable equilib-
rium vj (ω) ∈ (0, pj ) having a distribution with the density
. /
x
Nj 2hj (v)
j (x) = · exp 2
dv , 0 < x < pj ,
gj (x) cj gj (v)
where Nj is the normalizing factor and cj ∈ (0, pj ). Thus the system gener-
ated by (6.105) has a strongly positive sub-equilibrium
dxj = xj αj (xj ) + gi (xi ) · dt + σj xj ◦ dWtj , j = 1, . . . , d , (6.107)
i=j
where σj are constants and αj (x) and gi (x) are smooth functions on R+ with
properties like (S1) and such that
KI = {x = (x1 , . . . , xd ) ∈ Rd+ : xj = 0, j ∈ N \ I}
that describe the existence of each population independent of the others. Let
αj (x) be chosen such that the RDS generated by (6.109) in R+ has a positive
equilibrium for every j = 1, . . . , d. This is, for example, implied by
(see Theorem 6.6.1(ii)). Denote by vj (ω) the positive equilibrium for (6.109).
As in Subsect.5.7.3 under the condition βj > (d − 1)M we can prove the
existence of an equilibrium u(ω) = (u1 (ω), . . . , ud (ω)) for the RDS (θ, ϕ)
generated by (6.107) such that uj (ω) > vj (ω) and which attracts (from below)
the collection (v1 (ω), . . . , vd (ω)) of equilibria that correspond to the isolated
dynamics of each population. Thus as in the random case we observe an
interaction results in a benefit for all populations.
2
where (i) = exp{−δ|i|1 } for i ∈ Zd with some δ > 0. Let V+ := l,+ (Zd )
be the cone of nonnegative elements in l2 (Zd ), i.e. {xi : i ∈ Zd } ∈ V+ if
and only if xi ≥ 0 for all i ∈ Zd . The matrix {aij } and the function f define
mappings of l2 (Zd ) into itself via the formulae
(Ax)i = aij · xj , (F (x))i = f (xi ), i ∈ Zd , x ∈ l2 (Zd ) .
j∈Zd
for some µ > 0 (cf. Sect.2.5). The method used in the proof of Theorem 5.3.1
can be applied here. Therefore the RDS (θ, ψ) is order-preserving. Thus (θ, ϕ)
is also an order-preserving RDS.
Let us consider the affine system
dxi = aij · xj + αxi + γ · dt + dWti , i ∈ Zd . (6.113)
j∈Zd
It follows from Da Prato/Zabczyk [39, Theorem 3.4] again that for any γ
this equation generates an RDS (θ, ϕγ ) in the space l2 (Zd ). The comparison
principle (cf. Theorem 6.4.1) and property (6.111) give
ϕγ1 (t, ω)x ≤ ϕ(t, ω)x ≤ ϕγ2 (t, ω)x for all x ∈ l2 (Zd ) , (6.114)
for some ε > 0. Therefore for any γ ∈ R the affine RDS (θ, ψγ ) possesses a
F− -measurable equilibrium wγ (ω) ∈ l2 (Zd ). Inequality (6.114) implies that
wγ1 (ω) is a sub-equilibrium and wγ2 (ω) is a super-equilibrium for (θ, ϕ).
2
We obviously have the relation wγ1 (ω) ≤ wγ2 (ω). Since the cone l,+ (Zd )
is regular, Theorem 3.5.1 and Remark 3.5.1 imply the existence of an F− -
measurable equilibrium u(ω) in l2 (Zd ) for (θ, ϕ). This equilibrium generates a
Markov invariant measure for the stochastic equation (6.110) (see Sect.1.10).
References
19. Chicone C., Latushkin Yu. (1999) Evolution Semigroups in Dynamical Sys-
tems and Differential Equations. Amer Math Soc, Providence, Rhode Island
20. Chueshov I.D. (1999) Introduction to the Theory of Infinite-Dimensional Dis-
sipative Systems. Acta, Kharkov (in Russian)
21. Chueshov I.D. (2000) Order-Preserving Random Dynamical Systems Gen-
erated by a Class of Coupled Stochastic Semilinear Parabolic Equations. In:
Fiedler B, Gröger K., Sprekels J. (Eds.) International Conference on Differen-
tial Equations, EQUADIF 99, Berlin, Aug 1–7, 1999, vol 1. World Scientific,
Singapore, 711–716
22. Chueshov I.D. (2001) Order-Preserving Skew-Product Flows and Nonau-
tonomous Parabolic Systems. Acta Appl Math 65:185–205
23. Chueshov I.D., Scheutzow M. (2001) Inertial Manifolds and Forms for
Stochastically Perturbed Retarded Semilinear Parabolic Equations. J Dyn Diff
Equations 13:355–380
24. Chueshov I.D., Vuillermot P.-A. (1998) Long-Time Behavior of Solutions
to a Class of Quasilinear Parabolic Equations with Random Coefficients. Ann
Inst Henri Poincaré, Analyse non Linéaire 15:191–232
25. Chueshov I.D., Vuillermot P.-A. (1998) Long-Time Behavior of Solutions
to a Class of Stochastic Parabolic Equations with White Noise: Stratonovitch’s
Case. Probab Theory and Relat Fields 112:149–202
26. Chueshov I.D., Vuillermot P.-A. (2000) Long-Time Behavior of Solutions
to a Class of Stochastic Parabolic Equations with White Noise: Ito’s Case. Stoch
Anal Appl 18:581–615
27. Chung K.L., Williams R.J. (1983) Introduction to Stochastic Integration.
Birkhäuser, Boston Basel Stuttgart
28. Coddington E.A., Levinson N. (1955) Theory of Ordinary Differential Equa-
tions. Springer, New York
29. Cornfeld I.P., Fomin S.V., Sinai Ya. G. (1982) Ergodic Theory. Springer,
New York
30. Cohn D.L. (1980) Measure Theory. Birkhäuser, Boston Basel Stuttgart
31. Crauel H. (1991) Markov Measures for Random Dynamical Systems. Stochas-
tics and Stoch Reports 37:153–173
32. Crauel H. (1995) Random Probability Measures on Polish Spaces. Habilita-
tionsschrift, Universität Bremen
33. Crauel H. (1999) Global Random Attractors are Uniquely Determined by
Attracting Deterministic Compact Sets. Ann Mat Pura Appl, Ser IV 176:57–72
34. Crauel H. (2001) Random Point Attractors versus Random Set Attractors. J
London Math Society (2) 63:413–427
35. Crauel H., Debussche A., Flandoli F. (1997) Random Attractors. J Dyn
Diff Equations 9:307–341
36. Crauel H., Flandoli F. (1994) Attractors for Random Dynamical Systems.
Probab Theory Relat Fields 100:365–393
37. Crauel H., Flandoli F. (1998) Additive Noise Destroys a Pitchfork Bifurca-
tion. J Dyn Diff Equations 10:259–274
38. Crauel H., Imkeller P., Steinkamp M. (1999) Bifurcations of One-
Dimensional Stochastic Differential Equations. In: Crauel H., Gundlach M.
(Eds.) Stochastic Dynamics. Conference on Random Dynamical Systems, Bre-
men, Germany, April 28 - May 2, 1997. Springer. Berlin Heidelberg New York,
27–47
39. Da Prato G., Zabczyk J. (1995) Convergence to Equilibrium for Classical
and Quantum Spin Systems. Probab Theory Relat Fields 103:529–552
40. Da Prato G., Zabczyk J. (1996) Ergodicity for Infinite Dimensional Systems.
Cambridge University Press, Cambridge
References 229