Ergodic properties of some piecewise-deterministic Markov process with application to gene expression modelling
Introduction
In this paper, we study a subclass of piecewise-deterministic Markov processes (PDMPs; see e.g. [9]), defined somewhat similarly to those in [2], [3], [7], [20]. More specifically, we are concerned with a dynamical system evolving on a Polish space through random jumps, occurring at the jump times of a Poisson process. On the time interval between two consecutive jumps, the system moves deterministically according to one of semiflows , , which are randomly switched at the jump times (in contrast, e.g. in [7] only a single semiflow is considered). The post-jump location in our model is, however, not only a result of the semiflow change (like e.g. in [2], [3], [21]), but also a transformation of the current state, which takes place just before that change. This transformation is attained by a function , randomly selected among all possible ones, and further disturbed by adding a random shift (within the given -neighbourhood of zero). The random dynamical systems with such a jump mechanism can capture various less complex models developed in literature, starting from classical iterated function systems with place-dependent probabilities [25], [29], [39], through their generalisations involving an arbitrary set of transformations [23], [41], including their perturbed versions [17], [18], and ending with a class of piecewise-deterministic processes, driven by a collection of randomly switching semiflows [2], [3], [21]. Furthermore, they offer a description of phenomena arising in certain domains of natural science, especially in molecular biology (e.g. stochastic models of gene expression [17], [30]) and population dynamics [5]. In this connection, it is also worth mentioning that the PDMPs under consideration appear as solutions of certain variations on the Poisson driven stochastic differential equation (see e.g. [19], [24]), mainly developed by Lasota and Traple [28], which has fairly important applications in biomathematics, physics and engineering (cf. e.g. [10], [37]).
More formally, the evolution of the above-described dynamical system, say , can be described as follows. Assuming that is the index of the initially running semiflow, and that is the initial state of the system, we have until some random time, say , at which the process jumps to a random point in the -neighbourhood of , where is a random variable depending on . Let denote the position of the process directly after this jump, i.e. . The index of the semiflow that follows until the next jump time is given by , which depends on both the current state and the index of the previous flow. At the time the procedure restarts for and is continued inductively. As a result, we obtain a piecewise-deterministic trajectory with jump times and post-jump locations given by , as illustrated in Fig. 1.
Clearly, neither the chain nor its interpolation does not need to have the Markov property (unless only one semiflow is considered). Therefore, instead, we investigate the chain and the continuous-time process , wherein for , which are already Markovian.
The first and foremost goal of this paper is to provide a set of relatively easily verifiable conditions which ensure the exponential ergodicity of the chain in the Fortet–Mourier distance (see [25] or [11] for the equivalent Dudley metric). By this, we mean the existence of a unique stationary distribution, which attracts all initial distributions with finite first moment at an exponential rate with respect to the above-mentioned distance. As will be clarified in Section 3, the assumptions concerning the semiflows, which govern the deterministic evolution of the process, are quite naturally met by a fairly wide class of semiflows generated by certain dissipative differential equations in reflexive Banach spaces.
The proposed conditions imply, in particular, the existence of a norm-like function and constants and , for which the transition law of the chain satisfies This property, together with the assumption that the sublevel sets of are petite, is often used for verifying a Foster–Lyapunov drift condition ([31, Lemma 15.2.8]), which guarantees geometric ergodicity in the total variation norm for -irreducible and aperiodic Markov chains ([31, Theorems 16.1.2 and 15.4.1]). It should be, however, stressed explicitly that we consider the case where the process evolves over a Polish space, which is not necessarily locally compact (as it is required e.g. in [2], [3], [5], [7]). Under these settings, the conventional methods of Meyn and Tweedie [31], [32], like e.g. the above-mentioned drift criterion or another ones, based on the Harris recurrence (which are applied in the aforementioned papers), are most often inapplicable. This is mainly due to the difficulty in verifying the -irreducibility in the absence of local compactness. Typically, it requires the use of certain highly restrictive assumptions, like e.g. the strong Feller property (cf. [31, Proposition 6.1.5]), which is not the case here. Moreover, apart from showing the -irreducibility and aperiodicity, there still remains the problem with the petiteness of the sublevel sets of , which, in the case of locally-compact spaces, often follows directly from their compactness, whenever the support of has a non-empty interior (see [31, Proposition 6.2.8 and Theorem 6.2.9]). On the other hand, the methods that are predominantly used in the case of Polish spaces, which pertain to non-expansive (in the Fortet–Mourier distance) or, more general, equicontinuous Markov operators with some concentration properties, do not provide any conclusions on the rate of ergodicity (see e.g. [4], [17], [20], [27], [29], [39], [40]).
For the above-stated reasons, instead, we take advantage of the asymptotic coupling methods introduced by Hairer [15] (cf. also [6], [36], [41]), and we apply the results of Kapica and Ślęczka [23], which are based on them, to obtain the desired property of the chain in the Fortet–Mourier distance.
Having established the exponential ergodicity of the chain , we further prove that it obeys the strong law of large numbers (SLLN) by using a theorem of Shirikyan [35]. Another key result of our study is establishing a one-to-one correspondence between the sets of invariant measures of the chain and those of the process . This is done by applying several results from the theory of semigroups of linear operators in Banach spaces [12], [13]. Finally, using a martingale method (cf. [3]), we derive the SLLN for the PDMP from the corresponding property of the chain and the above-mentioned correspondence between invariant measures. It still remains, however, an open question whether the exponential ergodicity (in the sense described above) of the discrete-time model can imply the analogous property for the associated PDMP.
As mentioned in the beginning, from the application point of view, the dynamical system under study provides a useful tool for modelling certain biological processes. For instance, in Section 5.1, we show that may be adapted as a continuous-time model of prokaryotic gene expression in the presence of transcriptional bursting (cf. [30]). It is worth stressing here that, in our framework, the existence of a unique invariant distribution is guaranteed by the restrictions imposed on the local characteristics of the process. In contrast, applying the results of [30], the invariant measure can only be obtained by solving explicitly some differential equation and proving that its solution is a strictly positive probability density function. The second example, discussed in Section 5.2, refers to the discrete-time model for an autoregulated gene, introduced by Hille et al. [17], whose non-disturbed version also appears, for instance, in the cell cycle analysis (see [26], [41]). This model constitutes a special case of the dynamical system and indicates the importance of considering a non-locally compact space as the state space in the abstract framework.
The paper is organised as follows. In Section 1 we introduce basic notation and fundamental concepts on Markov operators (discussed more widely e.g. in [25], [31], [34]). Section 2 provides a detailed description of the model and the principal assumptions employed in the studies. Section 3 is intended to point out a general class of differential equations which generate semiflows consistent with our framework. All the main results are formulated in Section 4, which is divided into two parts: Section 4.1, devoted to the discrete-time model, and Section 4.2, pertaining to its continuous-time interpolation. In Section 5 we provide two examples of applications of our abstract framework in the gene expression analysis. Finally, the detailed proofs of all the results in the paper are carried out in Section 6. Additionally, in the Appendix, we give a rough sketch of the proof of [23, Theorem 2.1], which serves as an essential tool for the analysis contained in Section 6.1.1.
Section snippets
Preliminaries
Let us begin with introducing a piece of notation. Given a metric space , endowed with the Borel -field , we define
the space of all bounded, Borel measurable, real valued functions defined on , endowed with the norm: , ;
the subspace of consisting of continuous functions;
the subspace of consisting of Lipschitz continuous functions;
the space of all finite, countably additive functions (signed measures) on ;
the
Structure and assumptions of the model
Let be a separable Banach space, and let be a closed subset of . Further, fix , and let be an arbitrary probability measure supported on the ball , i.e. such that (in the case where , we simply take ). Moreover, assume that we are given a finite set , endowed with the discrete metric and a topological measure space with a -finite Borel measure . For simplicity, in the rest of the paper, we will
Reasonableness of the assumptions
It is essential to stress that condition (A2) is reachable by a quite wide class of semiflows acting on reflexive Banach spaces (in particular, Hilbert spaces). As will be clarified below, such semiflows can be generated by certain differential equations involving dissipative operators. Furthermore, in many cases, condition (A1) can be then easily derived from the conjunction of (A2) and (A3). To justify this claim, we first repeat some relevant definitions and results (without proofs) from [22]
The Markov chain given by the post-jump locations
In this part of the paper, we provide a criterion on the existence of a unique invariant probability measure for the operator , corresponding to the chain , which is exponentially attracting in the Fortet–Mourier distance. Having established this, we further obtain, in a relatively simple way, the SLLN for the discrete-time model.
Theorem 4.1 Suppose that conditions (A1)–(A5) hold, and that Then the Markov operator , generated by (2.8), has a unique invariant probability measure
Applications in gene expression analysis
This section is intended as an attempt to demonstrate the generality of the abstract model. We show that it is flexible enough to capture at least two completely different dynamical systems, regarding the gene expression process. Although the first of them is spanned on a finite-dimensional space (and might possibly be analysed by using some simpler tools), we believe that the combination of these two examples shows the universality of the presented approach.
Proofs
In this part of the paper, we provide the proofs of all our main results, which have been gathered in Section 4. Before we proceed to the analysis, let us go back to the definition of , that is, the metric in . As we have stressed in Section 2, all the results work under the assumption that the constant , appearing in (2.7), is sufficiently large. The choice of depends on the parameters appearing in conditions (A1)–(A3) as follows: where is a fixed
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Acknowledgment
We acknowledge useful discussions with S.C. Hille about the biological issues.
References (41)
- et al.
Limit theorems for some Markov chains
J. Math. Anal. Appl.
(2016) - et al.
Lower bound technique in the theory of a stochastic differential equation
J. Differential Equations
(2006) - et al.
Invariant measures related withPoisson driven stochastic differential equation
Stochastic Process. Appl.
(2003) A note on a theorem of Karlin
Stat. Probab. Lett.
(2001)Exponential rate of convergence for some Markov operators
Statist. Probab. Lett.
(2013)- et al.
Essential Cell Biology
(2013) - et al.
Quantitative ergodicity for some switched dynamical systems
Electron. Commun. Probab.
(2012) - et al.
Qualitative properties of certain piecewise deterministic Markov processes
Ann. Inst. Henri Poincare Probab.
(2014) - et al.
Criterion on stability for Markov processes applied to a model with jumps
Semigroup Forum
(2013) - et al.
Probabilistic and piecewise deterministic models in biology
ESAIM Proc. Surv.
(2017)
Exponential ergodicity for Markov processes with random switching
Bernoulli
Stability and ergodicity of piecewise deterministic Markov processes
SIAM J. Control Optim.
Generation of semigroups of nonlinear transformations on general Banach spaces
Amer. J. Math.
Piecewise-deterministic Markov processes: a general class of non-diffusion stochastic models
J. R. Stat. Soc. Ser. B
On the stability of the cells size distribution
J. Math. Biol.
Convergence of Baire measures
Studia Math.
Markov processes, Vol. I
Selected Papers of E.B. Dynkin with Commentary
DNA and Biotechnology
Exponential mixing properties of stochastic PDEs through asymptotic coupling
Probab. Theory Related Fields
Cited by (19)
Exponential ergodicity in the bounded-Lipschitz distance for some piecewise-deterministic Markov processes with random switching between flows
2022, Nonlinear Analysis, Theory, Methods and ApplicationsOn absolute continuity of invariant measures associated with a piecewise-deterministic Markov process with random switching between flows
2021, Nonlinear Analysis, Theory, Methods and ApplicationsA useful version of the central limit theorem for a general class of Markov chains
2020, Journal of Mathematical Analysis and ApplicationsCitation Excerpt :Such a model has already been investigated in terms of its exponential ergodicity and the strong law of large numbers in [4,3].
The Central Limit Theorem for Markov Processes that are Exponentially Ergodic in the Bounded-Lipschitz Norm
2024, Qualitative Theory of Dynamical Systems
- 1
The author is supported by the Foundation for Polish Science (FNP), Poland .