- Open Access
On nonergodicity for nonparametric autoregressive models
Advances in Difference Equations volume 2013, Article number: 200 (2013)
In this paper, we introduce a class of nonlinear time series models with random time delay under random environment, sufficient conditions for nonergodicity of these models are developed. The so-called Markovnization methods are used, that is, proper supplementary variables are added to a non-Markov process, then a new Markov process can be obtained.
MSC:60J05, 60J10, 60K37.
By virtue of their superduper properties, stable (ergodic or recurrent) stochastic processes are very popular among many researchers, so there has been a large literature devoted to the stable (ergodic or recurrent) or even stationary stochastic processes. For instance, Jeantheau  and Tjøstheim  established consistency of the estimator they proposed under stationarity and ergodicity conditions (see also [3–5]). Fernandes and Grammig  established conditions for the existence of higher-order moments, strict stationarity, geometric ergodicity and β-mixing property with exponential decay. This owes a great deal to the beautiful properties of stable processes, such as an ergodic Markov chain has an invariant probability measure which is finite, a recurrent stochastic process re-visits an arbitrary point in its image an infinite number of times. Just because of this, many researchers often like to target ergodicity or recurrence as their assumptions in their papers or books.
However, in this colorful world, lots and lots of phenomena exhibit instability behavior, for example, David  argued that an important lesson from economic history was that economies exhibited nonergodic behavior along many dimensions. Margolin and Barkai  indicated that time series of many systems exhibited intermittency, that is to say, at random times the system will switch from state on (or up) to state off (or down) and vice versa. One method to characterize such time series is using time average correlation functions to exhibit a nonergodic behavior.
Hence more and more researchers become increasingly interested in these instable processes. Recently, some problems of nonergodic stochastic processes have been studied by many authors. Basawa and Koul , Basawa and Brockwell , Basawa and Scott  and Feigin  studied asymptotic inference problems for parameters of nonergodic stochastic processes. Budhiraja and Ocone  proved an asymptotic stability result for discrete time systems in which the signal was allowed to be nonergodic. Durlauf  considered nonergodic economic growth. Goodman and Massey  generalized Jackson’s theorem so that the large-time behavior can be described for any nonergodic N-node Jackson network system. Griffeath  developed limit theorems for nonergodic set-valued Markov processes. Jacod  constructed the estimators for drift and diffusion coefficients of a multidimensional diffusion process and obtained consistent results without any kind of ergodicity or even recurrence assumption on the diffusion process.
Ergodicity criteria with drift functions for Markov processes have been studied by many authors. For instance, see Cline  and Tweedie [19–21] and the references therein. As for nonergodicity criteria for Markov processes, the readers are referred to [22–24]. Sheng et al.  also developed some sufficient conditions for nonergodicity of some time series models.
However, the processes considered by many researchers do not reflect the factors of the interference in a system and the system itself influenced by sudden environmental change. On the other hand, the time delay in the models studied is usually a fixed constant. In this paper, we popularize general nonparametric autoregressive models through introducing random environment and at the same we turn to a random time delay instead of a fixed time delay.
The remainder of the paper is organized as follows. Section 2 introduces the nonparametric autoregressive model with random time delay under random environment. Section 3 develops some useful lemmas and gives some sufficient conditions for nonergodicity of the proposed model as our main results. All the proofs are collected in Section 4.
2 The nonparametric AR model with random time delay under random environment
In this section, we first give some notations which will be used throughout the paper. In what follows, we always have a probability space , a finite set (r is a positive integer number), with ℋ as the σ-algebra generated by all sets of E. We also let be an m-dimensional real space and be the σ-algebra generated by all Boreal subsets of .
The nonparametric autoregressive model with random time delay under random environment is defined as follows:
where is an irreducible and aperiodic Markov chain defined on , taking values in E; , is a Borel measurable mapping; , where denotes the indicator function of a single element set ; are sequences of i.i.d. random variables defined on , taking values in . In what follows, the density function of , , is denoted as .
In this paper, we aim to obtain some sufficient conditions for nonergodicity of the proposed new model. Since the nonparametric AR model with random time delay under random environment defined by (2.1) itself does not have the Markovian property, we consider the following sequence:
For simplicity, let
From (2.1) and under mild conditions (see Lemma 1 below), it is easy to show that the sequence is a Markov chain on with the following transition probability:
where and ; denotes the indicator function of A; is the transition function of the Markov chain .
Let be the t-step transition probability of , then by the property of conditional probability and inductive approach we have: when ,
3 Main results
This section gives the main results of the new model described in Section 2. We need the following conditions.
Assumption 1 are mutually independent, and , is independent of . Moreover, for each i, is a constant independent of t and .
Assumption 2 is independent of the initial random variable .
Assumption 3 , is independent of .
Assumption 1 assures the stationarity of , . Assumption 2 and Assumption 3 guarantee the Markov property of . These are the basic conditions we know that guarantee the following lemmas can be used properly throughout the paper.
Lemma 1 Suppose that Assumptions 1-3 hold, then the sequence is a time-homogeneous Markov chain defined on with state space .
The irreducibility and aperiodicity in Lemma 2 are standard and can be found in Meyn and Tweedie  and Tong . The two concepts are very useful to derive the nonergodicity of the sequence . But before we state the results about the irreducibility and aperiodicity of the sequence , we need the following condition about the density function of , .
Assumption 4 The density function of , is strictly positive everywhere, i.e., , .
Lemma 2 Under Assumptions 1-4, the Markov chain is irreducible and aperiodic, where φ is a measure on , is a Lebesgue measure on satisfying if , , .
Remark 1 Obviously, if the Markov chain is -irreducible, for any nontrivial and σ-finite measure which is absolutely continuous with respect to , then is also -irreducible. So, we need a normal irreducibility which can define the range of the chain much more completely than some more arbitrary irreducibility measures one may construct initially. Fortunately, Sheng et al.  and Meyn and Tweedie  proved that if is a irreducible Markov chain, then there exists a maximal irreducibility measure Q. In this paper, we use those subsets whose maximal irreducibility measure is greater than zero, so here we denote .
Our main results are as follows.
Theorem 1 Suppose that Assumptions 1-4 hold and if there exist a non-negative measurable function g on , a set , and a non-negative measurable function on satisfying
For , , we have
For , we have(3.1)
For , we have(3.2)
and there exist , , , such that
Then the Markov chain is nonergodic. Moreover, whatever a probability distribution function of is, its probability distribution function will never converge to some probability distribution function.
Remark 2 Conditions (2) and (3) in the above Theorem 1 can be substituted for (4) or (5) as follows:
For , (3.1) holds, and , (3.2) holds.
For and , (3.2) holds, where when , (3.3) holds.
That is, under conditions (1) and (4) or (1) and (5), we can also show that is nonergodic, and the method of the proof is similar to that under conditions (1)-(3).
Proof of Lemma 1 , , and ∀i, , where s is an integer number satisfying , we have
where the last equation follows from the definition of the (2.1) model, Assumption 1 and the notation .
On the other hand,
Hence the sequence is a Markov chain, and its time-homogeneity follows from the stationarity of , . This completes the proof. □
Proof of Lemma 2 Suppose that and . Since is irreducible, , , such that
that is, , such that
Then from the t-step transition probability of , , we have
so is irreducible, and the aperiodicity of follows from Tong . This completes the proof. □
For , , let
In order to deal with the proofs of Theorem 1, we need the following propositions.
Lemma 3 
Suppose is a φ-irreducible Markov chain on the state space . If there exist constants , , a set A and a nonnegative measurable function which satisfies , , such that
, , ;
One of the following (i), (ii), (iii) hold:
For , , we have
furthermore, there exist , such that
For , , , we have
For , we have
Then is nonergodic.
Remark 3 Condition (1) in Lemma 3 is usually called the Kaplan condition; readers can consult Kaplan  for details. Sheng et al.  found a class of functions satisfying this condition. That is, if there exists a constant such that
then is the function wished.
Remark 4 Generally, the function , which we use frequently in Lemma 3, goes by the name of g-drift of the point x and is often expressed as . In addition, if
In fact, , so when , .
Proof of Theorem 1 By Lemma 1 and Lemma 2 we know that is a irreducible, aperiodic and time-homogeneous Markov chain with state space . So, we only need to check the conditions in Lemma 3. For simplicity, let
Step 1: To show that . For and , , by (2.2), we have
where is a δ-function, that is, if and zero otherwise.
So, by the integral transformation theorem, we can get
where the second-to-last line follows from condition (1) in the theorem.
Step 2: To show that when , and when , . In fact,
so by condition (3) we get the results wanted.
Step 3: We show that has uniformly lower bound. Note that , , we have
where the third line comes from the inequality , .
From condition (2) of this theorem, we can get when , ; when , if , , we have , if , we have
where the last inequality lies in the fact that when , .
Therefore, for all and ,
where . In other words, for all and , we have
so by Lemma 3 we know that is nonergodic.
Next we will prove whatever an initial probability distribution function of is, its probability distribution function will never converge to some probability distribution function. We will use reductio ad absurdum to prove it. Suppose that and there exists a probability distribution π such that
where is the total variation norm. As a matter of fact, due to the equivalence of norm, we can use any norm here. Since
Define , and it is well known that
and therefore we have
but this conflicts with the nonergodicity of the Markov chain . So, there is no probability distribution function π such that (4.1) holds. This completes the proof. □
School of Science, Jiangxi University of Science and Technology, No. 86, Hongqi Ave., Ganzhou, 341000, Jiangxi, P.R. China.
Jeantheau T: Strong consistence of estimators for multivariate ARCH models. Econom. Theory 1998, 14: 70-86.
Tjøstheim D: Estimation in nonlinear time series models. Stoch. Process. Appl. 1986, 21: 251-273. 10.1016/0304-4149(86)90099-2
Fan J, Yao Q Springer Series in Statistics. In Nonlinear Time Series: Nonparametric and Parametric Methods. Springer, New York; 2003.
Hafner CM, Preminger A: On asymptotic theory for multivariate GARCH models. J. Multivar. Anal. 2009, 100: 2044-2054. 10.1016/j.jmva.2009.03.011
Lintona O, Sancetta A: Consistent estimation of a general nonparametric regression function in time series. J. Econom. 2009, 152: 70-78. 10.1016/j.jeconom.2009.02.006
Fernandes M, Grammig J: A family of autoregressive conditional duration models. J. Econom. 2006, 130: 1-23. 10.1016/j.jeconom.2004.08.016
David, PA: Path-dependence: putting the past into the future of economics. Institute for Mathematical Studies in the Social Sciences. Stanford University (1988)
Margolin G, Barkai E: Nonergodicity of blinking nanocrystals and other Levy-walk processes. Phys. Rev. Lett. 2005, 4: 1-4.
Basawa IV, Koul HL: Asymptotically minimax tests of composite hypotheses for nonergodic type processes. Stoch. Process. Appl. 1983, 14: 41-54. 10.1016/0304-4149(83)90045-5
Basawa IV, Brockwell PJ: Asymptotic conditional inference for regular nonergodic models with an application to autoregressive processes. Ann. Stat. 1984, 12: 161-171. 10.1214/aos/1176346399
Basawa IV, Scott DJ Springer Lecture Notes in Statistics 17. In Asymptotic Optimal Inference for Nonergodic Models. Springer, New York; 1983.
Feigin PD: Conditional exponential families and a representation theorem for asymptotic inference. Ann. Stat. 1981, 9: 597-603. 10.1214/aos/1176345463
Budhiraja A, Ocone D: Exponential stability in discrete-time filtering for non-ergodic signals. Stoch. Process. Appl. 1999, 82: 245-257. 10.1016/S0304-4149(99)00032-0
Durlauf ST: Nonergodic economic growth. Rev. Econ. Stud. 1993, 60: 349-366. 10.2307/2298061
Goodman JB, Massey WA: The nonergodic Jackson network. J. Appl. Probab. 1984, 21: 860-869. 10.2307/3213702
Griffeath D: Limit theorems for nonergodic set-valued Markov processes. Ann. Probab. 1978, 6: 379-387. 10.1214/aop/1176995524
Jacod J: Parametric inference for discretely observed nonergodic diffusions. Bernoulli 2006, 12: 383-401. 10.3150/bj/1151525127
Cline DBH: Regular variation of order 1 nonlinear AR-ARCH models. Stoch. Process. Appl. 2008, 117: 840-861.
Tweedie RL: Sufficient conditions for regularity, recurrence and ergodicity of Markov processes. Math. Proc. Camb. Philos. Soc. 1975, 78: 125-136. 10.1017/S0305004100051562
Tweedie RL: Criteria for ergodicity, exponential ergodicity and strong ergodicity of Markov processes. J. Appl. Probab. 1981, 18: 122-130. 10.2307/3213172
Tweedie RL: Drift conditions and invariant measures for Markov chains. Stoch. Process. Appl. 2001, 92: 345-354. 10.1016/S0304-4149(00)00085-5
Choi BD, Kim B: Non-ergodicity criteria for denumerable continuous time Markov processes. Oper. Res. Lett. 2004, 32: 574-580. 10.1016/j.orl.2004.03.001
Kima B, Leeb I: Tests for nonergodicity of denumerable continuous time Markov processes. Comput. Math. Appl. 2008, 55: 1310-1321. 10.1016/j.camwa.2007.07.003
Kaplan M: A sufficient condition for nonergodicity of a Markov chain. IEEE Trans. Inf. Theory 1979, 25: 470-471. 10.1109/TIT.1979.1056059
Sheng ZH, Wang T, Liu DL: Stability Analysis of Nonlinear Time Series Models - Ergodic Theory and Applications. Science Press, Beijing; 1993.
Meyn SP, Tweedie RL: Markov Chains and Stochastic Stability. Springer, New York; 1993.
Tong H: Nonlinear Time Series: A Dynamical System Approach. Oxford University Press, Oxford; 1990.
The authors would like to thank the editor and anonymous referees for their valuable suggestions, which greatly improved our paper. This research is supported by the NSF of Jiangxi Province (No. 20132BAB211005), the SF of Jiangxi Provincial Education Department (No. GJJ12356), Key Scientific and Technological Research Project of Department of Education of Henan Province (No. 12B110006), and Foundation of Jiangxi University of Science and Technology (No. jxxj12064).
The authors declare that they have no competing interests.
All authors jointly worked on the results and they read and approved the final manuscript.