Large deviations principle for stationary solutions of stochastic differential equations with multiplicative noise
Abstract
We study the large deviations principle (LDP) for stationary solutions of a class of stochastic differential equations (SDE) in infinite time intervals by the weak convergence approach, and then establish the LDP for the invariant measures of the SDE by the contraction principle. We further point out the equivalence of the rate function of the LDP for invariant measures induced by the LDP for stationary solutions and the rate function defined by quasi-potential. This fact gives another view of the quasi-potential introduced by Freidlin and Wentzell.
keywords:
Large deviations principle; Stationary solutions; Invariant measures; Quasi-potential1 Introduction
Real world phenomena are often effected by random perturbations or under the influences of noise, random dynamical systems arise to model those problems. The stationary solution, random periodic solution, random quasi-periodic solution (see, for example [22, 19, 27, 12, 10, 13, 11]) are natural extensions correspond to the fixed point, periodic solution, quasi-periodic solution from deterministic dynamical systems to random dynamical systems, which are the fundamental concepts that describe the long time behavior of random dynamical systems. In particular, there are many phenomena with small random perturbations in nature, and small perturbations essentially influence the long time behaviour of the system in general. Following Freidlin and Wentzell’s perspective in [14], we use the LDP to characterize the long time asymptotic behavior of dynamical systems as the small random perturbation converges to zero. Therefore, it is interesting to investigate the LDP for stationary solutions, random periodic solutions and random quasi-periodic solutions in infinite time intervals with small noises.
The LDP for solutions in infinite time intervals is also an interesting problem in mathematics. The LDP for Wiener process in infinite time intervals is shown by Deuschel and Stroock (cf. Schilder’s theorem in Section 1.3 of [8]), the proof depends on the properties of the Gaussian measure. Although we can get the LDP for stationary solutions of Eq. (1.1) with additive noise by using the contraction principle, it is not feasible for multiplicative noise. The exponential tightness for stationary solutions of Eq. (1.1) with multiplicative noise in the infinite intervals plays an important role in the proof of LDP if we follow the approach of Deuschle and Stroock [8], but it is hard to obtain the exponential tightness for the lack of certain accurate estimates, such as the Fernique theorem for Gaussian measure (cf. Theorem 1.3.24 in [8]). Although, the proof of exponential tightness for non-Gaussian measures in the continuous function space is an interesting problem, unfortunately we have not found a suitable way to solve it by now.
Therefore, we use another well-known method, the weak convergence approach to prove the family of stationary solutions of Eq. (1.1) satisfies the LDP as follows. Noticing that the large deviations principle is equivalent to the Laplace principle (LP) in Polish space, it is sufficient to prove the LP by the Boué-Dupuis formula (see, for example, [3, 5, 9, 26]). This method is often referred as the weak convergence approach. Using the weak convergence approach and the Boué-Dupuis formula, the Freidlin-Wentzell type LDP for the family of solutions of SDE or SPDE driven by Wiener process in finite intervals has been re-proved (see, for example, [14, 25]). However, considering the LDP in infinite intervals, we need the Boué-Dupuis formula in infinite intervals. Since the compactness of space and the integrability of bounded functions on infinite intervals are different from those on finite intervals, we have to extend the Boué-Dupuis formula from finite intervals to infinite intervals. As far as we know, the Boué-Dupuis formula in infinite intervals has been used directly without proof in [23] by Barashkov and Gubinelli, we have to present the proof of Boué-Dupuis formula and the weak convergence approach in infinite intervals for the completeness of the present paper in Appendix A. In fact, the most difficult problem in our proof is the well-posedness with respect to the skeleton Eq. (3.1) in infinite time intervals, as we need to prove the uniqueness for the backward infinite horizon integral Eq. (5.1) for the skeleton Eq. (3.1) and construct the solution of integral equation (5.1). Besides, verifying the two conditions (cf. Condition 2.1 in Section 2.2 below) of the weak convergence approach in infinite time intervals is also quite technical.
The main purpose of this paper is to establish the LDP of stationary solutions for a class of stochastic differential equations in infinite time intervals. Moreover, since the one dimensional distribution of the stationary solution always generates an invariant measure of the SDE, here we give another illuminating view of the LDP for invariant measures by the contraction principle (cf. Section 4.2.1 in [7]). In fact, the invariant measure is unable to provide the properties for the trajectories of the dynamical systems, and there exist many systems with different dynamical behaviors but with the same invariant measure (cf. Example 4 below). By comparison, the stationary solution show more accurate information about dynamical systems, and the LDP for them will give more explicit characterization for the long time asymptotic behavior of random dynamical systems with small perturbations.
As far as we know, Freidlin and Wentzell [14], Cerrai and Röckner [6], and Brzeźniak and Cerrai [4] studied the LDP for invariant measures by taking the quasi-potential as the rate function. Moreover, it is not difficult to prove that the rate function (4.3) defined in Theorem 4.2 is equivalent to the rate function defined by quasi-potential, and the rate function (4.3) gives an natural explanation of quasi-potential. Roughly speaking, Freidlin and Wentzell [14], Cerrai and Röckner [6], and Brzeźniak and Cerrai [4] studied the LDP for invariant measures by using the result of the LDP of solutions in finite intervals and let the length of time intervals tends to infinity. It is different to our method that we directly consider the asymptotic behavior of stationary solutions for random dynamical systems in infinite intervals as perturbation converges to zero, and by using the contraction principle to get the LDP for invariant measures. Besides, we believe that our methods to show the LDP for stationary solutions is also available to explore the LDP for random periodic solutions and random quasi-periodic solutions. We give two examples (cf. Example 2 and 3 below) to show the random periodic solutions of Eq. (3.27) (see, for example, Feng, Liu and Zhao [10]) and random quasi-periodic solutions (see, for example, Hopf [15]) of equation (3.31) satisfy the LDP, and we will research the LDP for random periodic solutions and random quasi-periodic solutions for others equation in future papers.
In this paper, we consider the stochastic differential equation as follows
(1.1) |
All definitions of the symbols involved can be found later in Section 2.1.
Liu and Zhao give the representation of the stationary solution for Burgers equation with large viscosity in [19]. Under our Hypothesis 2.1, similar to the proof of [21] and [19], we have proved that there exists such that for every , the stochastic equation (1.1) exists a unique stationary solution satisfies the following equation in for any
(1.2) |
where be a separable Hilbert space, which is defined in Section 2.1.
The first major conclusion as follows.
Theorem 1.
We could get the following result for the family of invariant measures of Eq. (1.1) satisfies the LDP by using the contraction principle .
Theorem 2.
Noticing that under Hypothesis 2.1, we actually only deal with the case where Eq. (1.1) has only one asymptotically stable fixed point for . And even the stochastic Burgers equation and -dimensional Navier-Stokes on tours are all satisfies Hypothesis 2.1 (cf. Example 1, 5 below), but there are also many types of stochastic partial differential equations that do not satisfy Hypothesis 2.1, such as the stochastic wave equations. Therefore, there are many more complex and interesting cases that we might follow, such as the stochastic wave equations, random dynamical systems with random periodic solutions, random quasi-periodic solutions. In particular, the cases that the deterministic dynamical systems for have multiple equilibrium points (see, for example, [20]) is also an attractive problem. Furthermore, the existence of invariant measure does not imply the existence of stationary solution, unless we consider in an extended probability space (see, for example, Arnold [1]). The asymptotic behavior for the stationary solutions in the extended probability space is also an interesting problem.
This paper is structured as follows. In Section 2, we present the basic notations and some definitions for the stationary solution, LDP and quasi-potential. We give Hypothesis 2.1 which the operators , and satisfy throughout the paper and the definition for stationary solution in Section 2.1, the preliminary knowledge for LDP in Section 2.2 and the definition of quasi-potential in Section 2.3.
In Section 3, we prove the family of stationary solutions of Eq. (1.1) satisfies LDP. We use the weak convergence approach to prove the LDP for stationary solution of Eq. (1.1), and verify that the Burger’s equation satisfies the Hypothesis 2.1 in Example 1. Moreover, we state two examples to show the LDP for random periodic solutions and random quasi-periodic solutions in Example 2 and Example 3.
In Section 4, we prove the LDP for invariant measures of Eq. (1.1) in Theorem 4.2 and prove the rate function defined in Theorem 4.2 is equivalent to the rate function (2.10) defined by quasi-potential in Lemma 4.1. Moreover, we give two examples to show that the rate function (4.3) defined in Theorem 4.2 of the invariant measures is consistent with quasi-potential (see, for example, [4, 6, 14]), and the LDP of the stationary solutions can provide more dynamic information.
In Section 5, we consider the well-posedness of the skeleton equation in Section 5.1, and prove there exists a such that for every , Eq. (1.1) exists a unique stationary solution in Section 5.2.
In Appendix A, we prove the Boué-Dupuis formula and the weak convergence approach in infinite intervals.
2 Preliminaries
2.1 Notations
Let be a reflexive Banach space, and be the dual space of . Let be a separable Hilbert space with inner product , and its dual space are consistent by the Riesz isomorphism. Let continuously and densely, then is called a Gelfand triple.
Let be the space of all bounded linear operators from to with the operator norm . Let be the space of all Hilbert-Schmidt operators from to with Hilbert-Schmidt norm , where is an orthogonal basis of . Let be a trace class operator, is denoted by the space of linear operators such that is a Hilbert-Schmidt operator from to , with the norm .
Let . Then is a Hilbert space with the inner product
Clearly, the imbedding of in is Hilbert-Schmidt for is a trace class operator.
Let be a -Wiener process on with respect to a probability space , where is Borel -field of , is the Wiener measure on . Let be the -completion of . Define
where are the null sets of (see, for example, [19]).
We consider the stochastic differential equation
(2.1) |
where , , be progressively measurable and satisfies the following Hypothesis.
Hypothesis 2.1.
(i). Assume there exist constants such that
and
(ii). Let be the associated semigroup on H corresponding to , for any , and , there exists constant and such that the semi-group satisfies
(iii). The function will not be zero operator and there exist positive constants such that
and
Remark 2.1.
(i). Since imbedes to , there exists a constant such that
(ii). Let and . It follows from
is a trace class operator and combining Hypothesis 2.1 (iii), we could obtain that
and
(iii). Let be the solution of the equation with initial value , where be the deririaive of u with respect to time , then we have the estimates
then , thus
For any , it implies that
(2.2) |
Definition 2.2.
(Mild solution). For any , , an -valued predictable process , , is called a mild solution of Eq. (2.1) with initial value if
for each . In particular, the appearing integrals are well defined.
Theorem 2.3.
Proof.
To define the stationary solution of random dynamical system, let be a complete probability space, be a group of -preserving ergodic transformations on .
Definition 2.4.
(cf. [1])
A crude cocycle on is a -measurable random field with the following properties:
(i). for fixed , and all , -a.s. (where the
exceptional set can depend on ).
(ii). for all , .
Definition 2.5.
An measurable random variable is said to be a stationary solution for the crude cocycle if it satisfies for any
(2.3) |
Remark 2.2.
We denote by be the standard -preserving ergodic Wiener shift on , , . Noticing that the stationary solution defined in Definition 2.5 means that for any , Eq. (2.3) holds almost surely, which is slightly different from the definition in [19, 22] needs Eq. (2.3) holds for all for all with respect to the perfect cocycle (cf. [1, 19, 22]). We prove the unique solution of the pullback integral Eq. (5.10) be the stationary solution (in the sense of Definition 2.5), and the distribution of the solution for (5.10) generates an invariant measure. Our results about the LDP are independent of the perfect cocycle, since the LDP describes the properties of the distribution of stochastic processes, but not the pathwise properties. Even the well-posedness of the pullback infinite horizon integral equation (5.10) is enough to show the LDP and depicts the long time asymptotic behavior of random dynamical systems with small perturbations.,
2.2 Preliminary knowledge for LDP
We recall some standard definitions and results of the large deviations theory (cf. [3, 9, 7]). Let be a Polish space, be a family of -valued random variables defined on a probability space . The large deviation theory concerns the exponential decay of the probability measures of rare events. The rate of such exponential decay is expressed by the rate function.
Definition 2.7.
(Rate function). A function mapping to is called a rate function if is lower semi-continuous. A rate function is called a good rate function if for each , the level set is compact.
Definition 2.8.
(Large deviation principle). The sequence is said to satisfies the LDP with rate function if for each Borel subset of ,
where and are the interior and the closure of in , respectively.
Since we can not prove the exponential tightness for stationary solutions of Eq. (2.1) with multiplicative noise in infinite intervals, which extremely depends on the properties of the Gaussian measure. We choose the weak convergence method to study the LDP of stationary solution for Eq. (2.1). Some fundamental concepts and results about the weak convergence method are stated as follows.
Definition 2.9.
(Laplace principle). The sequence is said to satisfy the Laplace principle with a rate function if for each bounded continuous real-valued function defined on , we have
(2.6) |
We consider the stationary solution of Eq. (2.1) satisfies LDP in the Polish space with the norm
(2.7) |
According to Theorems 1.2.1 and 1.2.3 in Dupuis and
Ellis [9], if is a Polish space and is a good rate function, then the LDP and Laplace
principle are equivalent, which is the basis for the weak convergence approach. Moreover, we need the Boué-Dupuis formula in infinite intervals to prove the weak convergence approach in infinite intervals.
Although the Boué-Dupuis formula in infinite intervals has been used in [23] by Barashkov and Gubinelli, for the completeness of the article, we still give the proof of Boué-Dupuis formula and the weak convergence approach in infinite intervals in Appendix A, and illustrate the results in the following.
Let
and
It is similar to prove the set endowed with the weak topology is a Polish space (cf. [17], Theorem III.1’). In this paper, except for special instructions, the topology of is always weak topology. We also define
We provide the sufficient condition for the Laplace principle (equivalently, the LDP), which is similar to Assumption 4.3 in [5], in the follwoing.
Condition 2.1.
There exists a measurable map such that the following hold:
(i). Let for some . If converges to in distribution as -valued random elements, then converges to in distribution as .
(ii). For every , the set is a compact subset of .
Theorem 2.10.
If satisfies Condition 2.1, then the family satisfies the Laplace principle in with good rate function
(2.8) |
where the infimum over an empty set is taken as .
2.3 Definition of the quasi-potential
It is well known that [14] and [6] have proved the family of invariant measures for stochastic equations satisfies LDP with the quasi-potential as rate function. We give the definition of quasi-potential below, and compare the rate function of the LDP for invariant measures induced by the LDP for stationary solutions and the rate function defined by quasi-potential in Section 4.
For any and , we denote by any solution belonging to of the control equation
(2.9) |
And we define the action functionals by
where is the solution of Eq. (2.9) in the intervals corresponding to the control , and
Moreover, we denote
In particular, when and , we set
3 Large deviation principle for stationary solution
3.1 The proof of LDP for Large deviation principle for stationary solution
We verify that the stationary solution of Eq. (2.1) satisfies Condition 2.1, which is a sufficient condition to satisfies LDP in this section.
To define in Condition 2.1, we consider the following control equation
(3.1) |
We illustrate the following result for the skeleton equation (3.1), and Theorem 3.1 will be proved in Section 5.1.
Theorem 3.1.
For some finite and , under Hypothesis 2.1, there exists a unique solution of the backward infinite horizon integral equation
(3.2) |
and satisfies
(3.3) |
Therefore, we could define the measurable map by
where be the unique solution of Eq. (3.2) with control term and satisfies (3.3).
We consider the LDP for the family of stationary solutions of Eq. (2.1) in space , which norm is defined by (2.7). According to Theorem 2.6, we have proved that there exists such that for any , there exists a unique stationary solution of Eq. (2.1) and satisfies the integral equation
(3.4) |
Then we could define by . We define by
with the help of the Girsanov transform, be the unique stationary solution of the control equation
(3.5) |
Before we prove the stationary solution of Eq. (2.1) satisfies Condition 2.1, we first give the following Lemma, which will be used in the proof of Lemma 3.2.
Lemma 3.1.
Proof.
By using the Itô formula, Hypothesis 2.1 (i) , Remark 2.1 (ii) and Young inequality, we get
After choosing small enough, it follows from Remark 2.1 (i) that
thus for any , we can obtain that
It follows from the Burkholder-Davis-Gundy inequality (cf. Proposition 3.26 in [16]), Young inequality and Remark 2.1 (ii) that
thus for every , we could have
(3.6) | ||||
For every , it follows from (3.6) and , there exists such that
∎
The following lemmas will verify the stationary solution of Eq. (2.1) satisfies Condition 2.1 in space , which is a sufficient condition to satisfies the LDP in space .
Lemma 3.2.
Let , for some . If converges to as -valued random element in distribution, then in distribution as .
Proof.
Since converges to in probability could deduce converges to in distribution, it is sufficient to prove converges to in probability. For convenience, let . Since for any , there exists such that , then it follows from the definition of the norm (2.7) in space ,
Thus it is sufficient to prove that as in probability for every .
By using the Itô formula, Hypothesis 2.1 (i) and Remark 2.1 (ii), we can get that
It follows from the Young inequality, Remark 2.1 (i) and (ii), after choosing small enough, we can obtain that
Define
is a continuous random process for and are continuous (cf. Theorem 5.16 or Theorem 2.2 in [19]).
Then for any fixed , , there exists a constant , which only depend on such that
Let , then we have
By using the Burkholder-Davis-Gundy inequality, Young inequality and Remark 2.1 (ii), there exists a constant such that
Then using the Gronwall inequality, we could obtain that
(3.8) | ||||
It follows from is a Hilbert-Schmidt operator on , Remark 2.1 (ii), and converges to as -value random variable in distribution,
(3.9) |
It follows from Lemma 5.2, Lemma 3.1, and the Chebyshev inequality, for any fixed , it is easy to obtain that there exists , which is defined by Lemma 3.1, and a constant such that
it implies that . Thus combining inequalitys (3.1), (3.8) and (3.9), let and , we get that
(3.10) |
in probability. Moreover, for any fixed , it follows from
and (3.10) to obtain that in probability as . ∎
Lemma 3.3.
For any fixed finite positive number , the set is a compact subset of .
Proof.
Let be a sequence in , for convenience let where corresponds to the solution of Eq. (3.1) with in place of . By weak compactness of , there exists a subsequence of which converges to a limit weakly in . The subsequence is indexed by for ease of notation. Let be the solution of Eq. (3.1) responding to and . Then we obtain
it follows from Hypothesis 2.1 (i) and Remark 2.1 (i), (ii) that
For any and , by using the Young inequality and Remark 2.1 (i), we have
where is a constant only depend on .
It follows from the Gronwall inequality that
(3.11) | ||||
It follows from , and Lemma 5.2 that
(3.12) |
Since , there exits simple function sequences and strong convergence to and respectively, we can choose a subsequence still record it as , such that
(3.13) |
since weak converges to in , also weak converges to . By using Remark 2.1 (ii) and Hypothesis 2.1 (iii), we can obtain that
(3.14) | ||||
It follows from is a Hilbert-Schmidt operator on and weak converges to in that there exists a subsequence, still set , such that
(3.15) |
Lemma 3.2 and 3.3 have proved that the stationary solution of Eq. (2.1) satisfies Condition 2.1 in Polish Space . By using Theorem 2.10 and the equivalence of LDP and LP in Polish space, the following theorem holds.
Theorem 3.2.
Under the Hypothesis 2.1, the family satisfies the LDP in with rate function
(3.17) |
where the infimum over an empty set is taken as .
3.2 Some Examples
Next, we will give an example that satisfies the Hypothesis 2.1.
Example 1.
We consider the stochastic Burgers equation with multiplicative noise as follows,
(3.18) |
To correspond to the notation of Hypothesis 2.1, let .
Let
We will verify and satisfies Hypothesis 2.1 (i) and (ii) below.
By using integration by parts, we can obtain that
for any , then
(3.19) | ||||
According to Lemma A.1 in [19], we could have the following property.
(3.20) |
where is the minimal constant such that Sobolev’s inequality,
holds.
For linear operator , we have known that
(3.21) |
It follows from (3.19), (3.20), (3.21) and the Young inequality that
Let be the heat semigroup, through the similar calculation to Lemma 3.3 in [19], for any , , we can get
Therefore, assume that the function satisfies Hypothesis 2.1 (iii). It follows from Lemma 2.6 that there exists such that for any , the stationary solution of Burgers equation exists and is unique, which satisfies the following equation in for any ,
and
The skeleton equation of Burgers equation is
(3.22) |
It follows from Lemma 5.5 and Lemma 5.4, for any , there exits a unique solution of Eq. (3.22), which satisfies the following equation in for any ,
We define by
By using Theorem 3.2, the family satisfies the LDP in with rate function
(3.23) |
where the infimum over an empty set is taken as .
The following two examples show that we can also study the LDP for random periodic solutions and random quasi-periodic solutions. We give the definition of random periodic solution (see, for example, [13]).
Definition 3.3.
A random periodic solution of period of the random dynamical system is an -measurable map such that for almost all ,
(3.24) |
for any . It is called a random periodic solution with the minimal period if is the smallest number such that (3.24) holds.
The following lemma is the result of the LDP for -dimensional Brownian motion.
For convenience, let be the space with norm
(3.25) |
where .
Lemma 3.4.
(cf. Schilder’s theorem in section 1.3 of [8]) -dimensional Brownian motion satisfies the LDP in space with rate function
(3.26) |
where be the derivative of with respect to time .
Example 2.
Consider the one-dimension equation shown by Feng, Liu and Zhao in [10],
(3.27) |
where is one-dimension Brownian motion.
According to [10], for every , the random periodic solution is
Let be the space with another norm
(3.28) |
Let , and defined by
For any , then after simple calculation, we obtain that
by using the Grownall inequality, there exists a constant such that
(3.29) |
It follows from the integration by parts that
for any , it implies that there exists a constant only depends on such that
(3.30) |
For any , we choose such that , it follows from (3.29) and (3.30) that
then there exists such that for any , it deduces that . It imply that is continuous, by using the contraction principle, we obtain that the random periodic solutions of Eq. (3.27) satisfies the LDP in space with rate function
where be the derivative of with respect to time .
The following graphs with respect to the numerical of random periodic solutions of Eq. (3.27) with .
The following example is to show that we can also study the LDP for random quasi-periodic solutions. We give the definition of random quasi-periodic solution (see, for example, [11]).
Definition 3.4.
Let be a measurable map from to , and be a random dynamical systems from to . Let be a d-dimension vector, which is rationally independent. Then, we say has a random quasi periodic solution , if they satisfy
(i). (shift invariant of orbit) .
(ii). (quasi periodic property)
Example 3.
We consider a random Hopf’s bifurcation model for turbulence. The original deterministic model given by Hopf in [15] is
(3.31) |
where , and
We denote . Here, is regarded as the external force acting on the “velocity field” . By the Fourier transform, we know that
where is the orthonormal basis of , and with norm
So, we get
Thus, we consider the following equation
let , it is equivalent to
(3.32) |
where
Now, we consider that is perturbed by random external force . For simplicity and illustrating our ideas, we assume that , where and is a sequence independent Brownian motion, then Hopf’s model (3.31) or (3.33) with random external force is equivalent to the following stochastic equation
(3.33) | |||||
(3.34) |
where
For some more complex random Hopf’s models, please see [2].
Moreover, assume , and , is rationally independent. Let . Set , if for some fixed , then the solution of Eq. (3.31) converges to a random quasi-periodic solution with the angle variables form a manifold of the type of a -d torus . Specially, we can choose , , , let
where , , . Then for every , the solution of Eq. (3.31) converges to a random quasi-periodic solution
where .
Let be the space with another norm
(3.35) |
Let , and defined by
where . We prove the function is continuous in the following.
For any , then after simple calculation, we obtain that
(3.36) |
For any , we choose big enough such that , it follows from Eqs. (3.35) and (3.36) that
Define be the product space of times , let be
it is sufficient to prove that is a continuous function from to space with norm , for any .
For any , , satisfies and , we have
(3.37) | ||||
We first estimate , choose small enough such that then
(3.38) | ||||
Similar estimate for , choose small enough such that , we get that
(3.39) | ||||
It follows from (3.37), (3.38) and (3.39) that is a continuous function from to space , thus is a continuous function from to .
By using the contraction principle, we obtain that the random quasi-periodic solutions of Eq. (3.31) satisfies the LDP in space with rate function
4 Invariant measure, rate function and quasi-potential
4.1 LDP for invariant measure
In this section, we will use the contraction principle to prove that the LDP for the family of stationary solutions deduce the LDP for the family of invariant measures for (2.1).
It follows from Lemma 2.6 that there exists such that for , (2.1) exists a unique stationary solution . Let be the distribution of the stationary solution at time , i.e.
then is an invariant measure of Eq. (2.1) in . Moreover, Lemma 5.11 easily implies the invariant measure of Eq. (2.1) is unique in , for any . Thus is the unique invariant measure of Eq. (2.1) in .
For any , is an -measurable random variable, therefore, in this section, we only consider the LDP of stationary solutions for Eq. (2.1) in space with the norm
(4.1) |
Similar to the proof in Section 3, we could have the following theorem.
Theorem 4.1.
Under the Hypothesis 2.1, the family satisfies the LDP in with good rate function
(4.2) |
where the infimum over an empty set is taken as .
We could get the following Theorem by using the contraction principle (cf. Theorem 4.2.1 in [7]).
Theorem 4.2.
Proof.
It follows from Theorem 4.1 that the stationary solution family for Eq. (2.1) satisfies the LDP in with rate function .
Let by , it is obvious that is continuous, then it follows from the contraction principle that the invariant measure family of Eq. (2.1) satisfies the LDP in , with rate function
∎
4.2 Rate function and quasi-potential
It is well known that [14] and [6] have proved the family of invariant measures for stochastic equations satisfies LDP with the quasi-potential as rate function. The definition of quasi-potential is given in Section 2.3, we will prove the rate function defined in (4.3) and quasi-potential are equivalent below.
Lemma 4.1.
Proof.
Remark 4.1.
The next example illustrate that there exists two systems with different dynamical behaviors but with the same invariant measure. It implies that the LDP for the stationary solutions gives more dynamical information than the LDP of invariant measures.
Example 4.
Define two metrics and by
We consider two stochastic equations
(4.5) |
and
(4.6) |
It follows from Theorem 4.2 that the family of invariant measures of Eqs. (4.5) and (4.6) satisfies LDP with good rate function as follows
where with norm , and be the derivative of with respect to time .
By using Theorem 3.1 in [14], we can get that
It implies that the rate functions for invariant measures of two Eqs. (4.5) and (4.6) are equivalent. Furthermore, let , the two Eqs. (4.5) and (4.6) have the same invariant measure , for every . However, we have known that the determinate equations and with respect to Eqs. (4.5) and (4.6) have different asymptotic behavior.
Define the stationary solutions of Eqs. (4.5) and (4.6) by
It follows from Theorem 2.10 that the family of the stationary solutions of Eqs. (4.5) and (4.6) satisfies LDP in space with different rate function
It implies that it is meaningful to research the LDP for stationary solutions, which gives more dynamical information. The numerical simulation in the following graphs will give a more intuitive explanation.
We choose , the following three graphs are the numerical approximate of solutions for Eq. (4.5) correspond to the cases .
The following three graphs are the numerical approximate of solutions for Eq. (4.5) correspond to the cases .
Brzeźniak and Cerrai have researched the LDP for the invariant measures of the -dimensional stochastic Navier-Stokes equations on a torus in [4]. We will consider the LDP for the solutions of the pullback integral equations of the -dimensional stochastic Navier-Stokes equations in the following.
Example 5.
For convenience, we write -dimensional stochastic Navier-Stokes equations perturbed by a small additive noise in a functional form shown by Brzeźniak and Cerrai in [4], as
(4.7) |
for on a two-dimensional torus .
Let us recall that is the Stokes operator, roughly speaking, equal to the Laplace operator composed with the Leray-Helmholtz projection , the convection is equal to , and is a Wiener process.
The space is defined by
We also define the space by setting
where , are the 1st order weak derivatives of the torus.
The proof is similar to the proof of Mattingly in [21] and Appendix 5.2, that there exists such that for any , the solution of the following pullback integral equation exists and is unique.
(4.8) |
Similar to Example 1, it follows from Lemma 5.5 and Lemma 5.4, for any , that there exits a unique solution of skeleton Eq. (4.9), which satisfies the following equation in for any ,
(4.9) |
We define by
By using Theorem 3.2, the family satisfies the LDP in with rate function
where the infimum over an empty set is taken as .
For any , although we have not proved the solutions of stochastic Navier-Stokes Eq. (4.7) form a perfect cocycle, the distribution of the solution for Eq. (4.8) is also the unique invariant measure of Eq. (4.7). Moreover, it follows from Theorem 4.2 that the family of invariant measures of Eq. (4.7) satisfies LDP with rate function
(4.10) |
5 Proofs of the well-posedness of skeleton equation in infinite intervals and stationary solution
This section is divided into two subsections, in the first subsection we prove Theorem 3.1 about the well-posedness of the skeleton Eq. (3.1). In the second subsection we prove Theorem 2.6 about the existence and uniqueness of stationary solution for Eq. (2.1).
5.1 The well-posedness of the skeleton equation
Similar to Theorem 4.4 of Sritharan and Sundar [25], we could have the following lemma.
Lemma 5.1.
Before proving Theorem 3.1, we illustrate some lemmas below.
Lemma 5.2.
(Priori estimate of skeleton equation) For any , , , for some , under the Hypothesis 2.1, there exists a constant , which only depend on , , , , such that the unique mild solution of Eq. (3.1) with initial value at time satisfies
where is defined in Hypothesis 2.1 (i), and are defined in Remark 2.1 (i) and (ii) respectively.
Proof.
The following lemmas are aim to study the asymptotic stability of dynamical systems, which is the basis of the definition . For any , , , and . Let and denote the solutions of Eq. (3.1) starting at different initial value and at time , respectively. For convenience, let .
Lemma 5.3.
Proof.
Corollary 5.1.
For any , there exists a constant such that for any ,
which imply that
Proof.
We will give the following two lemmas to prove there exist a unique solution of the backward infinite horizon integral equation for Eq. (3.1). And then we define as the unique solution of the backward infinite horizon integral equation for Eq. (3.1).
Lemma 5.4.
For any , assume that for any Moreover, if satisfies the following equation in for any ,
(5.1) |
and
(5.2) |
then is unique.
Proof.
For any , it follows from Eq. (5.1) that
Therefore, for any , is the mild solution of Eq. (3.1) with initial value . We will show the uniqueness of Eq. (5.1). Assume and are two solutions of Eq. (5.1), then for any ,
It implies that and are mild solutions of Eq. (3.1) with initial value and at time , then by using Corollary 5.1 and assumption and , let , we can have
The uniqueness have been proved. ∎
Finally, we will construct the solution of (5.1). For any and fixed , for convenience we set be the mild solution of (3.1) with initial value at initial time , let
Lemma 5.5.
Proof.
It follows from Corollary 5.1 that is a Cauchy sequence in . Since the space is complete, there exists such that in . For is arbitrary, is defined for all time, and from Lemma 5.2, we have , this implies that
(5.3) |
Finally we will show satisfies Eq. (5.1).
Step 1: For any and , we will show that satisfies
For any , we can find such that . Fixed , it follows from Hypothesis 2.1 (ii), we obtain that
Since in , it deduces that
It follows from Remark 2.1 (ii) and (iii),
Since in , for any , it implies that
At the same time it follows form Remark 2.1 (iii) that , converge strongly to and in respectively, hence Eq. (5.1) holds.
Step 2: We next prove that satisfies Eq. (5.1). From Eq. (5.1), it is easy to know that for any ,
moreover,
Thus combining (5.3) and Remark 2.1 (iii), we obtain that and converge to 0 as . Therefore,
is a Cauchy sequence in with respect to for any . Let , we can obtain that
Moreover, by using Remark 2.1 (iii), Thus it follows from Eq. (5.1) that satisfies Eq. (5.1). ∎
The proof of Theorem 3.1 in the following.
5.2 Stationary solution
Similar to the proofs of Mattingly [21] and Liu and Zhao [19], we prove Theorem 2.6 in this subsection. This subsection under Hypothesis 2.1, and we first give the energy estimate for Eq. (2.1) in the following lemma.
Lemma 5.6.
(Energy Estimate) For convenience, we fixed and denote be the solution of Eq. (2.1) with initial value at time , then we have
Proof.
For every , it follows from the Itô formula that
For , it follows from the Itô formula and Hypothesis 2.1 (i) that
For convenience set , then for is a local martingale there exists a sequence of stopping time , with as , such that is a martingale, so the Optional Stopping Time Theorem implies , we denote by
When , . And when , . So, is non-negative for all . Thus we can use Fatou’s Lemma to get
which imply the conclusion. ∎
Set , we prove the following lemma by using Lemma 5.6.
Lemma 5.7.
Proof.
Thus
then , set , then by induction we get that
where is constant only depends on , which imply that is just the linear combination of the moments of order less that or equal to of . ∎
Let be a sequence of real random variables. Let . Define the random variable to be the smallest positive integer such that for almost every ,
Mattingly has proved the following Bounding Lemma in [21].
Lemma 5.8.
(Bounding Lemma) Assume that
for some and , then
(i). if then a.s..
(ii). is finite for .
Let and be the solutions of Eq. (2.1) starting from different initial value respectively. Define , we will consider the asymptotically stable property of Eq. (2.1) in the following lemmas.
Lemma 5.9.
Let , for any , there exists such that for any there exists a almost finite random variable satisfies
for every .
Proof.
It follows from the Itô formula and Hypothesis 2.1 that
Set , for control we need to control . By using the Burkholder-Davis-Gundy inequality and Lemma 5.7, for any there exists a constant such that
the second inequality comes from Hölder’s inequality.
Let . By the Chebyshev’s inequality, we have
(5.7) |
By using Bounding Lemma 5.8, we can get the following Lemma.
Lemma 5.10.
Let be a random variable, measurable with respect to , such that is finite, then
(i). For any fixed , if , then for any , , and is finite almost surely.
(ii). and is finite for .
Lemma 5.11.
For any and . Let be initial condition, measurable with respect to , such that , for some . Let and be the solutions of Eq. (2.1) starting from different initial value respectively. Then there exists and a sequence almost finite random variable such that
for every , .
Proof.
Lemma 5.12.
(cf. Theorem 2 in [21]) Fix and . Let be a sequence of random variable with . Assume that is measurable with respect to and that is uniformly bounded in for some . Then there exists such that for any the following hold:
(i). With probability one, there exists a random time such that for every and all with , we have
Here is the set In addition for any .
(ii). Let be a second sequence of random variable with measurable with respect to and that is uniformly bounded in for some . Then with probability one, there exists a random time such that for every and all with , we have
And for any .
Proof.
Only need to prove (i), it follows from Lemma 5.11, there exists and a sequence almost finite random variable such that
for every , .
Set , then using Bounding Lemma 5.8 we have that is almost finite and is finite a.s. Set , then for every and
∎
Lemma 5.13.
(cf. Corollary 1 in [21]) Fix , and . There exist such that for any , with probability one, there is a positive random variable such that for all and all , if , we have
We first consider the uniqueness of the solution of the infinite horizon stochastic integral Eq. (5.10), which is a crucial technical condition to obtain the stationarity solution of Eq. (2.1).
Lemma 5.14.
For every , where is defined in Lemma 5.11. Assume that is a -measurable, -adapted process, and for any . Moreover, if satisfies the following equation in for any
(5.10) |
and
(5.11) |
then is unique.
Proof.
For any and ,
Assume that is another solution of Eq. (5.10), then
which imply that and are solutions of Eq. (2.1) with initial value and at time . It follows from Lemma 5.12 (ii) that there exists and a sequence almost surely finite stopping time such that for every , when ,
under the condition (5.11), let , we can get that for any
it implies ∎
The following Lemma illustrate that if Eq. (5.10) has unique solution , then is also the stationary solution of Eq. (2.1).
Proof.
Since is -preserving on the probability space , for any , we have, almost surly,
set , then
it follows from the uniqueness of Eq. (5.10) that for any
∎
Next, we will construct the solution of Eq. (5.10). For , , we define by
Lemma 5.16.
Proof.
It follows from Lemma 5.13 that there exists , such that for any , , is a Cauchy sequence in . And since the space is complete, there exists a such that in . Since is arbitrary, is defined for all time, and from Lemma 5.7, we have . This implies that
(5.12) |
We will through two steps to prove satisfies Eq. (5.10).
Step 1. Firstly, we will prove that for any , satisfies
(5.13) |
For any fixed , there exists such that ,
We have proved that as in for any , then according to Hypothesis 2.1 (ii), for any ,
where the sign means that the left side is less than or equal to the right side of a constant multiple.
It follows from the Burkholder-Davis-Gundy inequality and Remark 2.1 (ii) and (iii) that
Thus there exists a subsequence still set as such that
in At the same time, converge strongly to and in respectively, hence let , Eq. (5.13) holds for any .
Step 2. Finally, we will prove that satisfies Eq. (5.10).
It follows from Eq. (5.13), for any ,
Thus
It follows from the Itô equality and Remark 2.1 (ii), (iii) that
is a Cauchy sequence, we obtain that as .
The proof of Theorem 2.6 in the following.
Proof.
Proof.
Let be another stationary solution of Eq. (2.1). Denote and . It follows from the cocycle property and Lemma 5.11 that there exists and a sequence almost surely finite stopping time such that for every , when ,
(5.14) | ||||
As is a -preserving ergodic Wiener shift, we have, for any ,
In fact, by the inequality (5.14), we know that
and
However,
This implies that . ∎
Appendix A: Weak convergence method in infinite intervals
The Boué-Dupuis formula in infinite intervals has been used directly without proof in [23] by Barashkov and Gubinelli. Although we believe that the experts in field of LDP and Gaussian measures are familiar with the Boué-Dupuis formula in infinite intervals, we still present the proof of Boué-Dupuis formula and the weak convergence approach in infinite intervals for the convenience of readers.
The main difference between the proof in infinite intervals and finite intervals appears in the lower bounded proof of Theorem A.5, since bounded functions are integrable in finite intervals, but not in infinite intervals. However, the integrable function of infinite intervals can be approximated by simple function, and then the proof for infinite intervals can be transformed into finite intervals. Other routine proofs are shown for the completeness of the present paper.
A.3 Boué-Dupuis formula in infinite interval
Lemma A.18.
(cf. Problem 3.19 in [16])
The following conditions are equivalent for a continuous martingale .
(i). It is a uniformly integrable family of random variables.
(ii). It converges in , as .
(iii). It converges -a.s. (as ) to an integrable random variable , such that
is a martingale.
(iv). There exists an integrable random variable , such that -a.s., for every .
Let be an abstract Wiener space. Namely, is a separable Banach space, is a separable Hilbert space densely and continuously embedded in , and is the Gaussian measure over . If we identify the dual space with itself, then may be viewed as a dense linear subspace of so that whenever and , where denotes the inner product in .
We now recall some notations from [26] about the filtration in abstract Wiener space. In what follows, we fix a continuous and strictly monotonic resolution of the identity in ,
(i). For each , is an orthogonal projection.
(ii). .
(iii). For , .
(iv). For any and , .
For any , there exists a sequence such that . Thus, there exists a such that
The is also written as , called the Skorohod integral of .
After taking , we can obtain a continuous and strictly monotonic resolution of the identity in , i.e.
(i). For each , is an orthogonal projection.
(ii). .
(iii). For , .
(iv). For any and , .
If we take another transform, we can also get the filtration on .
Definition A.1.
The continuous filtation on is defined by
where is the collection of all the null sets in with respect to . We write as , and remark that .
Below, we shall consider the filtered probability space . If there is no special declaration, the expectation and the term are always taken with respect to the Wiener measure .
Definition A.2.
For every , let be the collection of all cylindrical function with the form
(A.1) |
In particular, the elements in are measurable with respect to , We write as .
We have the following simple approximation result.
Lemma A.19.
For a fixed , lat be an measurable and bounded function on with bound . There exists a sequence such that
In particular, is dense in .
Definition A.3.
An valued random variable is called adapted to if for every and . All the adapted valued random variables in is denoted by . The set of all bounded elements in is denoted by , i.e.
A is called simple if it has the following form
where . The set of all simple elements in is denoted by . We write .
Proposition A.1.
is a closed subspace of , and is dense in .
Basing on this Proposition, for any , we can define Itô’s integral such that
On the other hand, for any norm finite and real Borel measurable function on and , we may define the following integral with respect to the vector valued measure
such that
It is standard to prove the following result.
Lemma A.20.
Let be a left-continuous adapted process the and bounded by . Then for any , there exists a sequence such that
and
Proof.
First of all, we define for every
Then, by the dominated convergence theorem we have
For each and , by Lemma A.19 one can find such that
Finally, we define
By the diagonalization method, we may find the desired sequence . In fact, the condition can be ignored. ∎
Proposition A.2.
Let be a Borel measurable function on . Then there exists a such that
Proof.
Set . Then is a uniformly martingale bounded from above by and from below by . By Lemma A.18 and the representation formula of martingales, there is a such that
Now define
Then, clearly and is a martingale with square variation process . Thus, we have
The desired formula follows. ∎
We also need the following Clark-Ocone formula.
Proposition A.3.
Let be the set of all satisfying
For , we define
and
(A.2) |
Then by the Girsanov theorem, we have for any
Lemma A.21.
For , let be defined by (A.2). Then
Theorem A.4.
Let be any bounded Borel measurable function on . Then
where is defined by (A.2). Moreover, the infimum is unique attained at some .
Proposition A.4.
Let be any bounded Borel measurable function on . For any , there are two such that
Moreover
(A.3) |
where denotes the law of in under .
Reader interested in proof can refer to [26].
Theorem A.5.
Let be a bounded Borel measurable function on . Then we have
(A.4) |
Proof.
(Upper bound)Let . By Proposition A.1 we may choose a sequence of such that
So, converges in distribution to in , and converges weakly to . Noting that by Eq. (A.3)
and we have ([26] Lemma2.1(ii))
Therefore,by Theorem A.4, we get the upper bound
(Lower bound) We divided the proof into two steps.
Step 1: We first assume that with the form
By Proposition A.3, there is a such that for all
(A.5) |
By Proposition A.3, we in fact find from the proof of Proposition A.2 that
Thus, . By Lemma A.20 there exists a sequence satisfying for some
such that
By extracting a subsequence if necessary, we may further assume that
we have by the dominated convergence theorem
Here we have used the uniform integrability of . Moreover, by Eq. (A.5) and Lemma A.21
Step 2: Let be a bounded measurable function on . By Lemma A.19 we can choose a sequence such that , and a.s.. For any and , by Step 1 there exists a such that
(A.6) |
In view and , we have
So there is a subsequence such that convergence. We have
Dominated convergence theorem gives that for sufficiently large ,
Since , we thus complete the proof of the lower bound. ∎
A.4 The proof of Laplace principle for infinite intervals
Since the large deviation principe is equivalent to Laplace principle in Polish space, we only need to prove the Laplace principle for infinite intervals, i.e. Theorem 2.10.
Proof.
In order to prove the Laplace principle, we must show that Eq. (2.6) holds for all real valued, bounded and continuous functions on space .
Then for every , there exists such that
(A.7) | ||||
We will prove that
We claim that we can assume without loss of generality that for all and a.s.
(A.8) |
for some finite number .
To see this, observe that if then . Now define random variable
The processes are in , being as before the indicator function, and furthermore
This observation implies that the right side of inequality (A.7) is at most
Hence it suffices to prove with replaced by . This proves the claim.
Henceforth we will assume that (A.8) holds. Pick a subsequence along which converges in distribution to as -valued random elements.
(Upper bound) Since is bounded . Let be arbitrary, and let be such that
Choose such that
and
By using the variational presentation, for bounded and continuous functions , we could have
Now from Condition 2.1 (i), as
converges to . Thus
Since is arbitrary, the proof is complete. The Condition 2.1 (ii) guaranteed that is a good rate function. ∎
Acknowledgments
We are very grateful to Professor Z. Dong for his help and suggestions, and also to Dr. W.L. Zhang for participating in the discussion. Y. Liu appreciates Professor H.Z. Zhao and Professor C.R. Feng for their discussion about random stationary solutions, random periodic solutions, etc. Y. Liu thanks to Mr. C. Bai for their discussion about the random Hopf model.
Y. Liu is supported by CNNSF (No. 11731009, No.11926327) and Center for Statistical Science, PKU. Z.H. Zheng is supported by CNNSF (No.12031020, No.12090014), the Key Lab. of Random Complex Structures and Data Sciences, CAS and National Center for Mathematics and Interdisplinary Sciences, CAS.
References
References
- [1] L. Arnold, Random dynamical systems, Springer Monographs in Mathematics, Springer-Verlag, Berlin, 1998.
- [2] C. Bai, Some problems of ergodic properties of Hopf Model with perturbation, Master’s thesis, Peking University, 2019.
- [3] M. Boué, P. Dupuis, A variational representation for certain functionals of Brownian motion, Ann. Probab. 26 (4) (1998) 1641-1659.
- [4] Z. Brzezniak, S. Cerrai, Large deviations principle for the invariant measures of the 2D stochastic Navier-Stokes equations on a torus, J. Funct. Anal. 273 (6) (2017) 1891-1930.
- [5] A. Budhiraja, P. Dupuis, A variational representation for positive functionals of infinite dimensional Brownian motion, Probab. Math. Statist. 20 (1) (2000) 39-61.
- [6] S. Cerrai, M. Röckner, Large deviations for invariant measures of stochastic reaction-diffusion systems with multiplicative noise and non-Lipschitz reaction term, Ann. Inst. H. Poincaré Probab. Statist. 41 (1) (2005) 69-105.
- [7] A. Dembo, O. Zeitouni, Large deviations techniques and applications, Vol. 38 of Stochastic Modelling and Applied Probability, Springer-Verlag, Berlin, 2010, corrected reprint of the second (1998) edition.
- [8] J.D. Deuschel, D.W. Stroock, Large deviation, Springer-Verlag, Universitext, New York, 1984.
- [9] P. Dupuis, R.S. Ellis, A weak convergence approach to the theory of large deviations, Wiley Series in Probability and Statistics: Probability and Statistics, John Wiley & Sons, Inc., New York, 1997, a Wiley-Interscience Publication.
- [10] C.R. Feng, Y. Liu, H.Z. Zhao, Numerical approximation of random periodic solutions of stochastic differential equations, Z. Angew. Math. Phys. 68 (5) (2017) 68-199.
- [11] C.R. Feng, B.Y. Qu, H.Z. Zhao, Random quasi-periodic paths and quasi-periodic measures of stochastic differential equations, J. Differential Equations 286 (2021) 119-163.
- [12] C.R. Feng, H.Z. Zhao, Random periodic solutions of SPDEs via integral equations and Wiener-Sobolev compact embedding, J. Funct. Anal. 262 (10) (2012) 4377-4422.
- [13] C.R. Feng, H.Z. Zhao, B. Zhou, Pathwise random periodic solutions of stochastic differential equations, J. Differential Equations 251 (1) (2011) 119-149.
- [14] M.I. Freidlin, A.D. Wentzell, Random perturbations of dynamical systems, 3rd Edition, Vol. 260 of Grundlehren der mathematischen Wissenschaften [Fundamental Principles of Mathematical Sciences], Springer, Heidelberg, 2012, translated from the 1979 Russian original by Joseph Szücs.
- [15] E. Hopf, Repeated branching through loss of stability, an example. Proceedings of the conference on differential equations (dedicated to A. Weinstein), University of Maryland Book Store, College Park, Md., 1956, 49056. Also in Selected works of Eberhard Hopf with commentaries. Edited by Cathleen S. Morawetz, James B. Serrin and Yakov G. Sinai. American Mathematical Society, Providence, RI, 2002.
- [16] I. Karatzas, S.E. Shreve, Brownian motion and stochastic calculus, 2nd Edition, Vol. 113 of Graduate Texts in Mathematics, Springer-Verlag, New York, 1991.
- [17] A.N. Kolmogorov, S.V. Fomin, Elements of the theory of functions and functional analysis. Vol. 1. Metric and normed spaces, Graylock Press, Rochester, N.Y., 1957, translated from the first Russian edition by Leo F. Boron.
- [18] W. Liu, M. Röckner, Stochastic partial differential equations: an introduction, Universitext, Springer, Cham, 2015.
- [19] Y. Liu, H. Z. Zhao, Representation of pathwise stationary solutions of stochastic burgers’ equations, Stoch. Dynam. 9 (04) (2009) 613-634.
- [20] D. Martirosyan, Large deviations for stationary measures of stochastic nonlinear wave equations with smooth white noise, Comm. Pure Appl. Math. 70 (9) (2017) 1754-1797.
- [21] J. C. Mattingly, Ergodicity of 2d navier stokes equations withrandom forcing and large viscosity, Commun. Math. Phys. 206 (2) (1999) 273-288.
- [22] S.-E.A. Mohammed, T.S. Zhang, H.Z. Zhao, The stable manifold theorem for semilinear stochastic evolution equations and stochastic partial differential equations, Mem. Amer. Math. Soc. 196 (917) (2008) 1-105.
- [23] B. Nikolay, G. Massimiliano, A variational method for , Duke Math. J. 169 (17) (2020) 3339-3415.
- [24] C. Prévôt, M. Röckner, A concise course on stochastic partial differential equations, Vol. 1905 of Lecture Notes in Mathematics, Springer, Berlin, 2007.
- [25] S. Sritharan, P. Sundar, Large deviations for the two-dimensional Navier-Stokes equations with multiplicative noise, Stochastic Process. Appl. 116 (11) (2006) 1636-1659.
- [26] X.C. Zhang, A variational representation for random functionals on abstract wiener spaces, J. Math. Kyoto. U 49 (3) (2009) 475-490.
- [27] H.Z. Zhao, Z.H. Zheng, Random periodic solutions of random dynamical systems, J. Differential Equations 246 (5) (2009) 2020-2038.