Abstract
In this paper, we study the asymptotic properties of drift parameter estimations in reflected Ornstein–Uhlenbeck process, and establish their moderate deviations in both cases with one-sided barrier and two-sided barriers. The main methods consist of regenerative process techniques and the strong Markov property, as well as moderate deviations for martingales.
We’re sorry, something doesn't seem to be working properly.
Please try refreshing the page. If that doesn't work, please contact support so we can address the problem.
Avoid common mistakes on your manuscript.
1 Introduction and Main Results
1.1 Introduction
In many situations, the stochastic processes involved are not allowed to cross a certain boundary, or are even supposed to remain within two boundaries. For instance, the reflected Ornstein–Uhlenbeck process behaves like the standard Ornstein–Uhlenbeck process in the interior of its domain. However, when it reaches the boundary, the sample path returns to the interior in a manner that the “pushing” force is minimal. This kind of process has wide range of applications in the field of queueing system, financial engineering, and mathematical biology.
Consider the following reflected Ornstein–Uhlenbeck process with one-sided barrier \(b_L\):
where \(\theta \in (0,+\infty )\) and \(\gamma \) are unknown, \(W=\{W_{t},t\in [0,\infty )\}\) is a standard Brownian motion. Here, the process \(L=\{L_t, t\ge 0\}\) is the minimal continuous increasing process with \(L_0=0\), which makes the process \(X_t\ge b_\mathrm{L}\) for all \(t\ge 0\). The process L increases only when X hits the boundary \(b_\mathrm{L}\), satisfying \(\int _0^{\infty }I_{\big \{X_t>b_{\mathrm{L}}\big \}}\mathrm{d}L_t=0\). Denote by \(P_{\theta ,\gamma ,x_0}\) the probability distribution of the solution of (1.1) on \(C({{\mathbb {R}}}_{+},{{\mathbb {R}}})\), the space of continuous functions from \(\mathbb {R}^+\) to \(\mathbb {R}\). Without specific instruction, we will suppress \(\theta , \gamma \) and denote by \(P_{x_0}\).
For \(\theta \in (0,+\infty )\), the reflected Ornstein–Uhlenbeck process (1.1) is an ergodic Markov process [36, 37], and its properties have been extensively studied. To be explicit, we can refer to [30, 32, 36] for the transition density analysis; and [8, 10, 11] for the study of first passages time, [36, 37] for the formula of stationary distribution, as well as [26, 31] for the limit theorems of the processes \(\{X_t, L_t, t\ge 0\}\).
The reflected Ornstein–Uhlenbeck process (1.1), as an extended Vasicek model, can successfully characterize mean reversion property of short interest rate. Actually, \(\theta \) indicates the mean reversion rate, whereas \(\gamma \), along with \(\theta \), determines the long run average. Then, to estimate them is a crucial step for practical applications. By Girsanov formula in Ward and Glynn [36], Bo et al. [9], the log-likelihood ratio process can be written as
where \(\mathcal {F}_T=\sigma (W_s,s\le T)\). The maximum likelihood estimators of \(\theta \) and \(\gamma \) are given by
and
In the case of \(\theta >0, \gamma \equiv 0\), Bo et al. [9] studied the strong consistency and asymptotic normality for \(\widehat{\theta }_T\), while Zang and Zhang [38] analyzed the Cramér–Rao lower bound. Moreover, Hu et al. [25] constructed another estimator via discrete observations, and considered its asymptotic normality. For more details, one can refer to Hu and Lee [24], Lee and Song [29], and the references therein. On the other hand, Zang and Zhang [39] considered asymptotic behavior of the trajectory fitting estimator for nonergodic reflected Ornstein–Uhlenbeck processes (\(\theta <0, \gamma \equiv 0\)).
Compared with huge literature in classical Ornstein–Uhlenbeck type process [3,4,5,6,7, 14,15,16,17,18,19,20,21,22,23, 27], the large and moderate deviations for estimators in reflected Ornstein–Uhlenbeck process have been in the ascendant. In this paper, our goal is to fill this gap, refining the already known results in Bo et al. [9], Zang and Zhang [38]. Here, we will analyze the reflected Ornstein–Uhlenbeck process (1.1) in view of regenerative process, and this method is quite different from the techniques in the existed work.
Generally speaking, moderate deviation fulfills the gap between the limiting distribution and large deviation. More precisely, consider the estimation \(\mathbb {P}\left( \frac{\sqrt{T}}{\lambda _T}\left( \begin{array}{ll}\widehat{\theta }_T-\theta \\ \widehat{\gamma }_T-\gamma \\ \end{array}\right) \in A\right) \), where A is a given domain of deviations and \(\lambda _{T}\) denotes the scale of deviation. When \(\lambda _{T}\equiv 1\), this is exactly the estimation of limiting distribution result. When \(\lambda _{T}\equiv \sqrt{T}\), this corresponds to the large deviation. And when \(\lambda _{T}\) between 1 and \(\sqrt{T}\), that is, \(\lambda _T\rightarrow \infty \) and \(\frac{\lambda _T}{\sqrt{T}}\rightarrow 0\) as \(T\rightarrow \infty \), this is the so-called moderate deviation.
1.2 Main Results
Denote the stationary distribution of (1.1) by ( [36])
Now, we state our main results as follows:
Theorem 1.1
Let \(\lambda _T\) be positive numbers, satisfying as \(T\rightarrow \infty \)
Then, the family \(\left\{ \frac{\sqrt{T}}{\lambda _T}\left( \begin{array}{ll}\widehat{\theta }_T-\theta \\ \widehat{\gamma }_T-\gamma \\ \end{array}\right) , T>0\right\} \) satisfies the large deviations with speed \(\lambda _T^2\) and rate function
where \(\mu _1=\int _{b_\mathrm{L}}^{\infty }x\pi (\mathrm{d}x)\), \(\mu _2=\int _{b_\mathrm{L}}^{\infty }x^2\pi (\mathrm{d}x)\) and
Explicitly, for any \(A\in \mathcal {B}(\mathbb {R}^2)\)
Then, we can obtain immediately that
Corollary 1.1
Under condition (1.5), the families
satisfy the large deviations with speed \(\lambda _T^2\) and rate functions
respectively.
In particular, for any \(x\ge 0\), we have
and
The paper is organized as follows. In Sect. 2, by using the regenerative process techniques, we first state some properties of reflected Ornstein–Uhlenbeck process (1.1), and then give exponential equivalence for the functionals \(\int _0^TX_t\mathrm{d}t, \int _0^TX_t^2\mathrm{d}t\) to their asymptotic expectations, respectively. The proof of the main result Theorem 1.1 will be postponed to Sect. 3. In Sect. 4, we extend our results to two-sided barriers case. The main methods of this paper consist of regenerative process techniques and strong Markov property, as well as the moderate deviations for martingales. Throughout this paper, \(C_0,C_1\), depending only on \(b_L, \theta , \gamma \) and the initial point \(x_0\), denote positive constants whose values can differ at different places.
2 Regenerative Process and Exponential Equivalence
To obtain the moderate deviations for \(\left( \hat{\theta }_T, \hat{\gamma }_T\right) \), the key point is to show the functionals \(\int _0^TX_tdt\) and \(\int _0^TX_t^2dt\) are exponential equivalent to their asymptotic expectations, respectively. Notice that the existing methods (Girsanov formula technique [3,4,5,6,7, 14, 15, 19]; multiple Wiener-Itô integral [21, 27]; log-Sobolev inequality method
[13, 18, 20]) maybe not work. Here, regenerative process techniques will be employed, and we benefit a lot from Banerjee and Mukherjee [2].
2.1 Regenerative Process View of Functionals
We first briefly recall the definition of regenerative process [33, 34].
Definition 2.1
The process \(X=\big \{X_t, t\ge 0\big \}\) is a regenerative process, if there exist random times \(0\le \Theta _0\le \Theta _1\le \cdots \), such that for \(k\ge 1\),
-
(1)
\(\big (X_{\Theta _k+t}, t\ge 0\big \}\) has the same distribution as \(\big (X_{\Theta _0+t}, t\ge 0\big \}\).
-
(2)
\(\big (X_{\Theta _k+t}, t\ge 0\big \}\) is independent of \(\big (X_{t}, 0\le t\le \Theta _k\big \}\).
In particular, if \(\Theta _0=0\), the process X is called a non-delayed regenerative process. Else, X is called a delayed regenerative process.
Loosely speaking, regenerative process starts anew at regeneration times \(\big \{\Theta _k, k\ge 1\big \}\), independent of the past. Moreover, the regeneration times split the process into renewal cycles that are independent and identically distributed, possibly except the first cycle.
For the reflected Ornstein–Uhlenbeck process X (1.1), let
Now, we can define regenerative times in terms of hitting times as follows:
The strong Markov property implies that X is a regenerative process with regeneration times given by \(\Big \{\Theta _k, k\ge -1\Big \}\). Then, under \(P_{x_0}\) with \(x_0\ge b_L\),
are both independent and identically distributed sequences. Moreover, we also have the following important results
and
where the sum is 0 if the upper index is strictly less than the lower index.
The tail asymptotic of regenerative time \(\Theta _0\), renewal rewards \(\int _0^{\Theta _0}X_t\mathrm{d}t\), \(\int _0^{\Theta _0}X_t^2\mathrm{d}t\), and \(\int _0^{\Theta _0\wedge T}X_t\mathrm{d}t\), \(\int _{\Theta _{N_\mathrm{T}}}^{T}X_t\mathrm{d}t\) can be analyzed as follows:
Lemma 2.1
For all \(x_0\ge b_L\) and \(\Theta _0\) defined by (2.3), there exists some positive constants \(C_0, C_1\) depending only on \(x_0, b_L, \theta \) and \(\gamma \), such that for T large enough
Moreover, there exists some \(\eta >0\) such that \(E_{x_0}e^{\eta \Theta _0}<\infty \).
Proof
Firstly, if \(b_L\le x_0\le b_L+1\), then \(\Theta _0=\tau _X(b_L+2)\). Define the following Ornstein–Uhlenbeck process
Under \(X_0=Y_0=x_0\), we have \(X_t\ge Y_t\) for \(t\ge 0\), and then \(\tau _X(b_\mathrm{L}+2)\le \tau _Y(b_\mathrm{L}+2)\), where
By Corollary 3.1 in Alili et al. ( [1]), we have for T large enough
where \(C_0, C_1\) are positive constants depending only on \(x_0, b_L, \theta \) and \(\gamma \).
On the other hand, if \(x_0>b_L+1\), we have \(\tau _X(b_L+1)=\tau _Y(b_L+1)=\alpha _1\), and \(X_t=Y_t\) on the interval \([0,\tau (b_L)]\), where Y is defined by (2.7). It holds by strong Markov property
Using Corollary 3.1 in Alili et al. [1] again, we have for T large enough
where \(C_0, C_1\) are positive constants depending only on \(x_0, b_L, \theta \) and \(\gamma \).
Finally, by using Fubini theorem and (2.6), we can choose some \(\eta >0\) such that
which concludes the proof of this lemma. \(\square \)
Lemma 2.2
For all \(x_0\ge b_L\), there exists some positive constants \(C_0, C_1\) depending only on \(x_0, b_L, \theta \) and \(\gamma \), such that for T large enough
In particular, there exists some \(\eta >0\) such that
Proof
Firstly, if \(b_L\le x_0\le b_L+1\), then \(\Theta _0=\tau _X(b_L+2)\), and
which implies by (2.9) that
On the other hand, suppose \(x_0>b_L+1\). Then, for \(t\in [0,\tau _X(b_L)]\), \(X_t=Y_t\) and \(\alpha _1=\tau _X(b_L+1)=\tau _Y(b_L+1)\), where Y is defined by (2.7). Then, it holds that
By using the strong Markov property, we obtain
Since \(\lim _{T\rightarrow \infty }\frac{1}{T}\int _0^TY_t^2\mathrm{d}t=\frac{\theta +2\gamma ^2}{2\theta ^2}\), by Lemma 2.3 in Gao and Jiang [20] and (2.9), we have for T large enough
Finally, by Hölder inequality, Lemma 2.1 and (2.13), we have
and
which complete the proof of this lemma. \(\square \)
2.2 Exponential Equivalence
In this subsection, we will show that the following functionals \(\int _0^TX_t\mathrm{d}t,\quad \int _0^TX_t^2\mathrm{d}t\) are exponentially equivalent to their asymptotic expectations, respectively.
Since the stationary distribution of (1.1) is given by [36]
using ergodic theorem (Theorem 1.16 in [28]), we have immediately that
Lemma 2.3
As \(T\rightarrow +\infty \), under \(P_{x_0}\) with \(x_0\ge b_\mathrm{L}\), for any \(\beta \in \mathbb {R}\)
where \(\mu _1=\int _{b_\mathrm{L}}^{\infty }x\pi (\mathrm{d}x)\), \(\mu _2=\int _{b_\mathrm{L}}^{\infty }x^2\pi (\mathrm{d}x)\).
Remark 2.1
By using Proposition 7.3 in Ross [33], Lemma 2.2, (2.4), (2.5) and strong Markov property, we have
Now, we can state the exponential equivalence results as follows:
Proposition 2.1
For \(\lambda _T\) defined by (1.5) and for all \(\delta >0\) and \(x_0\ge b_L\), we have
and
In particular, for any \(\beta \in \mathbb {R}\)
Proof
To prove (2.16), applying (2.5), we have by strong Markov property
From (2.11), it follows that
Now, it is sufficient to show
Firstly, we give some estimations for \(N_\mathrm{T}\). In fact, we have for any \(\delta '>0\)
Take T large enough such that \(\frac{T}{\left[ {T(1+\delta ')}/{E_{b_\mathrm{L}+2}\Theta _{0}}\right] +1}<\frac{E_{b_\mathrm{L}+2}\Theta _{0}}{1+\delta '/2}\). Then, Under \(P_{b_\mathrm{L}+2}\), by using Lemma 2.1, \(\left\{ \Theta _{i}-\Theta _{i-1}: i\ge 0\right\} \) is a sequence of independent and identical distributed variables with some finite exponential moment. Then, we have by the large deviation results
Similarly, it holds
Secondly, take T large enough such that \(\frac{T}{\left[ {T(1+\delta ')}/{E_{b_\mathrm{L}+2}\Theta _{0}}\right] }>\frac{E_{b_\mathrm{L}+2}\Theta _{0}}{1+2\delta '}\). By (2.14), we have
Now, choose \(\delta '<\frac{\delta E_{b_\mathrm{L}+2}\Theta _{0}}{8E_{b_\mathrm{L}+2}\int _{\Theta _{0}}^{\Theta _1}X_t^2\mathrm{d}t}\), and then
Notice that, under \(P_{b_\mathrm{L}+2}\), by using Lemma 2.2, \(\left\{ \int _{\Theta _{k-1}}^{\Theta _k}X_t^2\mathrm{d}t: k\ge 0\right\} \) is a sequence of independent and identical distributed variables with some finite exponential moment. Together with (2.19), (2.21) and the large deviation results,
Finally, using (2.20) and following above procedures, we also have
Therefore,
Now, we turn to proving (2.15). Indeed, by (2.4), we can write
Applying Lemma 2.1, we have
Together with (2.18) and strong Markov property, we obtain
and
Together with (2.18, 2.22, 2.23) 2.24, following the similar line in the proof of (2.16), we have for any \(\delta >0\),
and thus
which completes the proof of this proposition. \(\square \)
3 Moderate Deviations for \(\Big (\widehat{\theta }_T, \widehat{\gamma }_{T}\Big )\)
Set
For \(\widehat{\theta }_T\) and \(\widehat{\gamma }_T\) defined by (1.2) and (1.3), we have the following key martingale decomposition by straightforward calculations
with the martingale
and the remainder term
As a martingale, \(\big \{M_T, T>0\big \}\) is the main term in our moderate deviation analysis, while \(R_T\) will be negligible.
Lemma 3.1
For \(\lambda _T\) defined by (1.5) and \(M_T\) defined by (3.3), the families
satisfy the large deviations with speed \(\lambda _T^2\) and rate function
respectively, where \(\mu _1=\int _{b_\mathrm{L}}^{\infty }x\pi (\mathrm{d}x)\), \(\mu _2=\int _{b_\mathrm{L}}^{\infty }x^2\pi (\mathrm{d}x)\) and
Proof
Note that \(\big \{M_\mathrm{T}, T>0\big \}\) and \(\Big \{\int _0^T\left( \mu _1-X_t\right) \mathrm{d}W_t, T>0\Big \}\) are martingales with predictable quadratic variations
and \(\Big <\int _0^\cdot \left( \mu _1-X_t\right) \mathrm{d}W_t\Big >_T=\int _0^T\left( \mu _1-X_t\right) ^2\mathrm{d}t\). By Proposition 2.1, we can get for any \(\delta >0\),
Therefore, Proposition 1 in Dembo ( [12]) yields the conclusion of this lemma. \(\square \)
Lemma 3.2
For the remainder term \(R_\mathrm{T}\) defined by (3.4), we have the following results.
-
(1)
For any \(\delta >0\),
$$\begin{aligned} \lim _{T\rightarrow \infty }\frac{1}{\lambda _T^2}\log P_{x_0}\Big (\Big |\widehat{\sigma }_T^2-\big (\mu _2-\mu _1^2\big )\Big |\ge \delta \Big )=-\infty . \end{aligned}$$(3.5) -
(2)
For any \(\delta >0\),
$$\begin{aligned} \lim _{T\rightarrow \infty }\frac{1}{\lambda _T^2}\log P_{x_0}\Big (\frac{1}{\lambda _T}\Big |R_T\Big |\ge \delta \Big )=-\infty . \end{aligned}$$(3.6)
Proof
-
(1)
Since \(\widehat{\sigma }^2_T=\frac{1}{T}\int _0^TX_t^2dt-\frac{1}{T^2}\left( \int _0^TX_tdt\right) ^2\), we have
$$\begin{aligned}&P_{x_0}\Big (\Big |\widehat{\sigma }_T^2-\big (\mu _2-\mu _1^2\big )\Big |\ge \delta \Big )\\&\quad \le P_{x_0}\Big (\frac{1}{T}\Big |\int _0^TX_t^2\mathrm{d}t-\mu _2T\Big |\ge \delta /2\Big ) \\&\qquad +P_{x_0}\Big (\frac{1}{T}\Big |\int _0^TX_t\mathrm{d}t+\mu _1T\Big |\cdot \frac{1}{T}\Big |\int _0^TX_tdt-\mu _1T\Big |\ge \delta /2\Big )\\&\quad \le P_{x_0}\Big (\frac{1}{T}\Big |\int _0^TX_t^2\mathrm{d}t-\mu _2T\Big |\ge \delta /2\Big ) +P_{x_0}\Big (\frac{1}{T}\Big |\int _0^TX_t\mathrm{d}t-\mu _1T\Big |\ge |\mu _1|+1\Big )\\&\qquad +P_{x_0}\Big (\frac{1}{T}\Big |\int _0^TX_t\mathrm{d}t+\mu _1T\Big |\cdot \frac{1}{T}\Big |\int _0^TX_t\mathrm{d}t-\mu _1T\Big |\\&\quad \ge \delta /2, \frac{1}{T}\Big |\int _0^TX_tdt-\mu _1T\Big |<|\mu _1|+1\Big )\\&\quad \le P_{x_0}\Big (\frac{1}{T}\Big |\int _0^TX_t^2\mathrm{d}t-\mu _2T\Big |\ge \delta /2\Big ) +P_{x_0}\Big (\frac{1}{T}\Big |\int _0^TX_t\mathrm{d}t-\mu _1T\Big |\ge |\mu _1|+1\Big )\\&\qquad +P_{x_0}\Big (\frac{1}{T}\Big |\int _0^TX_t\mathrm{d}t-\mu _1T\Big |\ge \frac{\delta }{2(3|\mu _1|+1)}\Big ). \end{aligned}$$Now, we can complete the proof of (3.5) by using Proposition 2.1.
-
(2)
For any \(L>0\),
$$\begin{aligned}&P_{x_0}\left( \frac{1}{\lambda _T\sqrt{T}}\Big |W_\mathrm{T}\big (\widehat{\mu }_T-\mu _1\big ) +\big (1-(\mu _2-\mu _1^2)^{-1}\widehat{\sigma }_T^2\big )\int _0^T\left( \mu _1-X_t\right) \mathrm{d}W_t\Big |\ge \delta \right) \\&\quad \le P_{x_0}\left( \frac{1}{\lambda _T\sqrt{T}}\Big |W_\mathrm{T}\big (\widehat{\mu }_T-\mu _1\big )\Big |\ge \delta /2\right) \\&\qquad +P_{x_0}\left( \frac{1}{\lambda _T\sqrt{T}}\Big |\big (1-(\mu _2-\mu _1^2)^{-1}\widehat{\sigma }_T^2\big )\int _0^T\left( \mu _1-X_t\right) \mathrm{d}W_t\Big | \ge \delta /2\right) \\&\quad \le P_{x_0}\left( \frac{1}{\lambda _T\sqrt{T}}\Big |W_\mathrm{T}\Big |\ge L\right) +P_{x_0}\left( \Big |\widehat{\mu }_T-\mu _1\Big |\ge \frac{\delta }{2L}\right) \\&\qquad +P_{x_0}\left( \frac{1}{\lambda _T\sqrt{T}}\Big |\int _0^T\left( \mu _1-X_t\right) \mathrm{d}W_t\Big | \ge L\right) \\&\qquad +P_{x_0}\left( \Big |1-(\mu _2-\mu _1^2)^{-1}\widehat{\sigma }_T^2\Big |\ge \frac{\delta }{2L}\right) . \end{aligned}$$Applying Proposition 2.1, Lemma 3.1, (3.5) and classical moderate deviations for the Brownian motion, we can obtain that
$$\begin{aligned}&\lim _{T\rightarrow \infty }\frac{1}{\lambda _T^2}\log P_{x_0}\left( \frac{1}{\lambda _T\sqrt{T}}\Big |W_\mathrm{T}\big (\widehat{\mu }_T-\mu _1\big ) \right. \\&\qquad \left. +\big (1-(\mu _2-\mu _1^2)^{-1}\widehat{\sigma }_T^2\big )\int _0^T\left( \mu _1-X_t\right) \mathrm{d}W_t\Big |\ge \delta \right) \\&\quad \le -\frac{L^2}{2}\Big (1\vee \big (\mu _2-\mu _1^2\big )^{-1}\Big ), \end{aligned}$$which implies immediately by letting \(L\rightarrow \infty \) that
$$\begin{aligned} \begin{aligned}&\lim _{T\rightarrow \infty }\frac{1}{\lambda _T^2}\log P_{x_0}\left( \frac{1}{\lambda _T\sqrt{T}}\Big |W_T\big (\widehat{\mu }_T-\mu _1\big ) \right. \\&\qquad \left. +\big (1-(\mu _2-\mu _1^2)^{-1}\widehat{\sigma }_T^2\big )\int _0^T\left( \mu _1-X_t\right) \mathrm{d}W_t\Big |\ge \delta \right) \\&\quad =-\infty . \end{aligned} \end{aligned}$$(3.7)
Similarly, we can also have that
Then, together with (3.5), (3.7) and (3.8), we can complete the proof of (3.6). \(\square \)
Proof of Theorem 1.1
By (3.2) and Lemma 3.2, \(\left\{ \frac{\sqrt{T}}{\lambda _T}\left( \begin{array}{ll}\widehat{\theta }_T-\theta \\ \widehat{\gamma }_T-\gamma \\ \end{array}\right) , T>0\right\} \) is exponential equivalent to \(\Big \{\frac{1}{\lambda _T\sqrt{T}}M_\mathrm{T}, T>0\Big \}\) with speed \(\lambda _T^2\). Theorem 1.1 follows from Lemma 3.1. \(\square \)
4 The Case of Two-Sided Barriers
In this section, we focus on the drift parameter estimations for the reflected Ornstein–Uhlenbeck process with two-sided barriers \(b_L\) and \(b_U\) ( \(b_U>b_L\)):
where \(\theta \in (0,+\infty )\) and \(\gamma \) are unknown, the processes \(L=\{L_t, t\ge 0\}\) and \(U=\{U_t, t\ge 0\}\) are the minimal continuous increasing processes with \(L_0=U_0=0\), which make the process \(X_t\in [b_L, b_U]\) for all \(t\ge 0\) and satisfy
The stationary distribution of (4.1) is given by
4.1 Maximum Likelihood Estimators of \(\theta \) and \(\gamma \)
By Girsanov formula in Ward and Glynn [36], Bo et al. [9], the log-likelihood ratio process can be written as
where \(\mathcal {F}_T=\sigma (W_s,s\le T)\). Therefore, the maximum likelihood estimators of \(\theta \) and \(\gamma \) are given by
and
Similar to the one-sided barrier case in Sect. 3, we have the following key martingale decomposition
where
and
4.2 Regenerative Process View of \(\int _0^TX_t\mathrm{d}t\) and \(\int _0^TX_t^2\mathrm{d}t\)
To analyze the deviation properties of \(\int _0^TX_tdt\) and \(\int _0^TX_t^2dt\), we will also employ regenerative process techniques. Let \(\tau _X(x)=\inf \Big \{t\ge 0: X_t=x\Big \}\). We can define regenerative times in terms of hitting times, which are slightly different from the one-sided barrier case.
From the strong Markov property of reflected Ornstein–Uhlenbeck process, X is a regenerative process with regeneration times given by \(\Big \{\widetilde{\Theta }_k, k\ge 0\Big \}\). Then, under \(P_{x_0}\) with \(x_0\in [b_\mathrm{L},b_\mathrm{U}]\),
are both independent and identically distributed sequences. Moreover, we also have the following crucial formulas
and
Parallel to Lemmas 2.1 and 2.2, we have the following decay of tail probabilities.
Lemma 4.1
For all \(x_0\in [b_\mathrm{L}, b_\mathrm{U}]\), there exists some positive constants \(C_0, C_1\) depending only on \(x_0, b_\mathrm{L}, b_\mathrm{U}, \theta \) and \(\gamma \), such that for T large enough
and
In particular, there exists some \(\eta >0\) such that \(E_{x_0}e^{\eta \widetilde{\Theta }_0}<\infty \), and
Proof
Firstly, if \(b_L\le x_0\le b_\mathrm{L}+\frac{b_\mathrm{U}-b_\mathrm{L}}{4}\), then \(\widetilde{\Theta }_0=\tau _X(b_\mathrm{L}+\frac{b_\mathrm{U}-b_\mathrm{L}}{2})\). Under \(X_0=Y_0=x_0\), we have \(X_t\ge Y_t\) for \(t\le \tau _X(b_\mathrm{U})\), and then \(\tau _X(b_\mathrm{L}+\frac{b_\mathrm{U}-b_\mathrm{L}}{2})\le \tau _Y(b_\mathrm{L}+\frac{b_\mathrm{U}-b_\mathrm{L}}{2})\), where Y and \(\tau _Y\) are defined by (2.7) and (2.8). By Corollary 3.1 in Alili et al. [1], we have for T large enough
where \(C_0, C_1\) are positive constants depending only on \(x_0, b_L, b_U, \theta \) and \(\gamma \).
Secondly, if \(b_L+\frac{3(b_\mathrm{U}-b_\mathrm{L})}{4}\le x_0\le b_\mathrm{U}\), then \(\widetilde{\Theta }_0=\tau _X(b_\mathrm{L}+\frac{b_\mathrm{U}-b_\mathrm{L}}{2})\). Under \(X_0=Y_0=x_0\), we have \(X_t\le Y_t\) for \(t\le \tau _X(b_\mathrm{L})\), and then \(\tau _X(b_\mathrm{L}+\frac{b_\mathrm{U}-b_\mathrm{L}}{2})\le \tau _Y(b_\mathrm{L}+\frac{b_\mathrm{U}-b_\mathrm{L}}{2})\). By Corollary 3.1 in Alili et al. [1], we have for T large enough
Thirdly, if \(b_\mathrm{L}+\frac{b_\mathrm{U}-b_\mathrm{L}}{4}\le x_0\le b_L+\frac{3(b_\mathrm{U}-b_\mathrm{L})}{4}\), we have \(X_t=Y_t\) on the interval \([0,\tau _X(b_\mathrm{L})\wedge \tau _X(b_\mathrm{U})]\). Under \(X_0=Y_0=x_0\), we have
Consequently, it holds by strong Markov property
Using Corollary 3.1 in Alili et al. [1] again, we have for T large enough
Therefore, together with (4.10,4.11, 4.12), we can complete the proof of (4.8).
Finally, since \(\sup _{t\in [0,\infty )}|X_t|\le |b_L|\vee |b_U|\), then (4.9) can be achieved by (4.8).
\(\square \)
4.3 Exponential Equivalence and Moderate Deviations
By (4.6, 4.7) and Lemma 4.1, and using the same procedure as in the proof of Proposition 2.1, we can state the following exponential equivalence results, while the proofs are omitted.
Proposition 4.1
Let \(\lambda _T\) be defined by (1.5). For all \(\delta >0\) and \(b_{L}\le x_0\le b_U\), we have
and
In particular, for any \(\beta \in \mathbb {R}\)
Now, following the similar line as in the proof of Lemma 3.2, we have
Lemma 4.2
Let \(\lambda _T\), \(\widetilde{M}_T, \widetilde{R}_T, \widehat{\mu }_T, \widetilde{\sigma }^2_T, \widetilde{\mu }_1, \widetilde{\mu }_2\) be defined by (1.5,4.3, 4.4, 4.5).
-
(1)
The family \(\Big \{\frac{\widetilde{M}_T}{\sqrt{T}\lambda _T}, T>0\Big \}\) satisfy the large deviations with speed \(\lambda _T^2\) and rate function
$$\begin{aligned} \widetilde{I}(x)=\frac{1}{2}x^{\tau }\widetilde{\Sigma }^{-1}x,\quad x\in \mathbb {R}^2, \end{aligned}$$where
$$\begin{aligned} \widetilde{\Sigma }=\big (\widetilde{\mu }_2-\widetilde{\mu }_1^2\big )^{-1}\left( \begin{array}{ll}1\\ \widetilde{\mu }_1\\ \end{array} \begin{array}{ll}\widetilde{\mu }_1\\ \widetilde{\mu }_2\\ \end{array}\right) . \end{aligned}$$ -
(2)
For any \(\delta >0\),
$$\begin{aligned}&\lim _{T\rightarrow \infty }\frac{1}{\lambda _T^2}\log P_{x_0}\Big (\Big |\widehat{\sigma }_T^2-\big (\widetilde{\mu }_2-\widetilde{\mu }_1^2\big )\Big |\ge \delta \Big ) =-\infty ,\quad \\&\lim _{T\rightarrow \infty }\frac{1}{\lambda _T^2}\log P_{x_0}\Big (\frac{1}{\lambda _T}\Big |\widetilde{R}_T\Big |\ge \delta \Big )=-\infty . \end{aligned}$$
By (4.2), Lemma 4.2, and using the same way as in the proof of Theorem 1.1, we can establish the moderate deviations for \(\left( \widetilde{\theta }_T-\theta , \widetilde{\gamma }_T-\gamma \right) \).
Theorem 4.1
For \(\lambda _T\) defined by (1.5), the family \(\left\{ \frac{\sqrt{T}}{\lambda _T}\left( \begin{array}{ll}\widetilde{\theta }_T-\theta \\ \widetilde{\gamma }_T-\gamma \\ \end{array}\right) , T>0\right\} \) satisfies the large deviations with speed \(\lambda _T^2\) and rate function \(\widetilde{I}(x)\).
References
Alili, L., Patie, P., Pedersen, J.: Representations of the first hitting time density of an Ornstein-Uhlenbeck process. Stoch. Models 21, 967–980 (2005)
Banerjee, A., Mukherjee, D.: Join-the-shortest queue diffusion limit in Halfin-Whitt regime: tail asymptotics and scaling of extrema. Annals Appl. Probab. 29, 1262–1309 (2019)
Bercu, B., Rouault, A.: Sharp large deviations for the Ornstein-Uhlenbeck process. Theor. Probab. Appl. 46, 1–19 (2002)
Bercu, B., Coutin, L., Savy, N.: Sharp large deviations for the non-stationary Ornstein-Uhlenbeck process. Stoch. Process. Appl. 122, 3393–3424 (2012)
Bercu, B., Richou, A.: Large deviations for the Ornstein-Uhlenbeck process with shift. Adv. Appl. Probab. 47, 880–901 (2015)
Bercu, B., Coutin, L., Savy, N.: Sharp large deviations for the fractional Ornstein-Uhlenbeck process. Theor. Probab. Appl. 55, 575–610 (2011)
Bercu, B., Richou, A.: Large deviations for the Ornstein-Uhlenbeck process without tears. Stat. Probab. Lett. 123, 45–55 (2017)
Bo, L., Zhang, L., Wang, Y.: On the first passages times of reflected O-U processes with two-sided barriers. Queue. Syst. 54, 313–316 (2006)
Bo, L., Wang, Y., Yang, X., Zhang, G.: Maximum likelihood estimation for reflected Ornstein-Uhlenbeck processes. J. Stat. Plan. Inference 141, 588–596 (2011)
Bo, L., Wang, Y., Yang, X.: First passage times of (reflected) Ornstein-Uhlenbeck processes over random jump boundaries. J. Appl. Probab. 48, 723–732 (2011)
Bo, L., Ren, G., Wang, Y., Yang, X.: First passage times of reflected generalized Ornstein-Uhlenbeck processes. Stoch. Dyn. 13, 1–16 (2013)
Dembo, A.: Moderate deviations for martingales with bounded jumps. Electron. Commun. Probab. 1, 11–17 (1996)
Djellout, H., Guillin, A., Wu, L.: Moderate deviations of empirical periodogram and non-linear functionals of moving average processes. Ann. Inst. H. Poincaré Probab. Stat. 42, 393–416 (2006)
du Roy de Chaumaray, M.: Large deviations for the squated radial Ornstein-Uhlenbeck process. Theor. Probab. Appl. 61, 408–441 (2017)
du Roy de Chaumaray, M.: Moderate deviations for parameters estimation in a geometrically ergodic Heston process. Stat. Inference Stoch. Process. 21, 553–567 (2018)
El Onsy, B., Es-Sebaiy, K., Viens, F.G.: Parameter estimation for a partially observed Ornstein-Uhlenbeck process with long-memory noise. Stochastics 89, 431–468 (2017)
Es-Sebaiy, K., Viens, F.G.: Optimal rates for parameter estimation of stationary Gaussian processes. Stoch. Process. Appl. 129, 3018–3054 (2019)
Fan, J., Gao, F.: Deviation inequalities and moderate deviations for estimators of parameters in TAR models. Front. Math. China 6, 1067–1083 (2011)
Florens-Landais, D., Pham, H.: Large deviations in estimate of an Ornstein-Uhlenbeck model. J. Appl. Probab. 36, 60–77 (1999)
Gao, F., Jiang, H.: Deviation inequalities and moderate deviations for estimators of parameters in an Ornstein-Uhlenbeck process with linear drift. Electron. Commun. Probab. 14, 210–223 (2009)
Gao, F., Jiang, H.: Deviation inequalities for quadratic Wiener functionals and moderate deviations for parameter estimators. Sci. China Math. 60, 1181–1196 (2017)
Gao, F.: Long time asymptotics of unbounded additive functionals of Markov processes. Electron. J. Probab. 22, 94 (2017)
Guillin, A., Liptser, R.: Examples of moderate deviation principles for diffusion processes. Discret. Contin. Dyn. Syst., Ser. B 6, 77–102 (2006)
Hu, Y., Lee, C.: Drift parameter estimation for a reflected fractional Brownian motion based on its local time. J. Appl. Probab. 50, 592–597 (2013)
Hu, Y., Lee, C., Lee, M., Song, J.: Parameter estimation for reflected Ornstein-Uhlenbeck processes with discrete observations. Stat. Inference Stoch. Process. 18, 279–291 (2015)
Huang, G., Mandjes, M., Spreij, P.: Limit theorems for reflected Ornstein-Uhlenbeck processes. Stat. Neerlandica 68, 25–42 (2014)
Jiang, H., Liu, J., Wang, S.: Self-normalized asymptotic properties for the parameter estimation in fractional Ornstein-Uhlenbeck process. Stoch. Dyn. 3, 1950018 (2019)
Kutoyants, Y.A.: Statistical inference for ergodic diffusion processes. Springer, Cham (2003)
Lee, C., Song, J.: On drift parameter estimation for reflected fractional Ornstein-Uhlenbeck processes. Stochastics 88, 751–778 (2016)
Linetsky, V.: On the transition densities for reflected diffusions. Adv. Appl. Probab. 37, 435–460 (2005)
Mandjes, M., Spreij, P.: A note on the central limit theorem for the idleness process in a one-sided reflected Ornstein-Uhlenbeck model. Stat. Neerlandica 71, 225–235 (2017)
Ricciardi, L., Sacerdote, L.: On the probability densities of an Ornstein-Uhlenbeck process with a reflecting boundary. J. Appl. Probab. 24, 355–369 (1987)
Ross, S.M.: Introduction to probability models. Academic Press, Cambridge (2014)
Thorisson, H.: Coupling, stationarity, and regeneration. Springer, New York (2000)
Ward, A., Glynn, P.: A diffusion approximation for Markovian queue with reneging. Queue. Syst. 43, 103–128 (2003)
Ward, A., Glynn, P.: Properties of the reflected Ornstein-Uhlenbeck process. Queue. Syst. 44, 109–123 (2003)
Xing, X., Zhang, W., Wang, Y.: The stationary distributions of two classes of reflected Ornstein- Uhlenbeck processes. J. Appl. Probab. 46, 709–720 (2009)
Zang, Q., Zhang, L.: A general lower bound of parameter estimation for reflected Ornstein-Uhlenbeck processes. J. Appl. Probab. 53, 22–32 (2016)
Zang, Q., Zhang, L.: Asymptotic behaviour of the trajectory fitting estimator for reflected Ornstein-Uhlenbeck processes. J.Theor. Probab. 32, 183–201 (2019)
Acknowledgements
The authors would like to express great gratitude to the anonymous referee and the editor for the careful reading and constructive comments which led to an improved presentation of this paper.
Author information
Authors and Affiliations
Corresponding author
Additional information
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Hui Jiang is supported by the National Natural Science Foundation of China (Grant Nos.11771209, 11971227); Qingshan Yang is supported by the National Natural Science Foundation of China (Grant Nos.11401090, 11971097, 11971098) and the Fundamental Research Funds for the Central Universities (Grant No.2412019FZ031)
Rights and permissions
About this article
Cite this article
Jiang, H., Yang, Q. Moderate Deviations for Drift Parameter Estimations in Reflected Ornstein–Uhlenbeck Process. J Theor Probab 35, 1262–1283 (2022). https://doi.org/10.1007/s10959-021-01096-3
Received:
Revised:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s10959-021-01096-3
Keywords
- Maximum likelihood estimator
- Moderate deviation principle
- Reflected Ornstein–Uhlenbeck process
- Regenerative process