Communications in Statistics - Theory and Methods

ISSN: 0361-0926 (Print) 1532-415X (Online) Journal homepage: https://www.tandfonline.com/loi/lsta20

Least squares estimator for Ornstein–Uhlenbeck

processes driven by small fractional Lévy noises

Qingbo Wang, Guangjun Shen & Zhenlong Gao

To cite this article: Qingbo Wang, Guangjun Shen & Zhenlong Gao (2019): Least squares
estimator for Ornstein–Uhlenbeck processes driven by small fractional Lévy noises,
Communications in Statistics - Theory and Methods, DOI: 10.1080/03610926.2019.1653923

To link to this article: https://doi.org/10.1080/03610926.2019.1653923

Published online: 19 Aug 2019.

Submit your article to this journal

View related articles

View Crossmark data

Full Terms & Conditions of access and use can be found at

https://www.tandfonline.com/action/journalInformation?journalCode=lsta20
COMMUNICATIONS IN STATISTICS—THEORY AND METHODS
https://doi.org/10.1080/03610926.2019.1653923

Least squares estimator for Ornstein–Uhlenbeck processes

vy noises
driven by small fractional Le
Qingbo Wanga, Guangjun Shena, and Zhenlong Gaob
a
Department of Mathematics, Anhui Normal University, Wuhu, China; bSchool of Statistics, Qufu Normal
University, Qufu, China

ABSTRACT ARTICLE HISTORY

In this paper, we study the problem of parameter estimation for the Received 23 May 2019
Ornstein–Uhlenbeck processes Accepted 5 August 2019


dXt ¼ hXt dt þ dYt KEYWORDS
dYt ¼ Yt dt þ edLdt Ornstein–Uhlenbeck pro-
cess; fraction L
evy process;
driven by Ornstein–Uhlenbeck processes with small fractional Le
vy least squares estimator;
noises and Yt can be observed, based on discrete high frequency consistency; asymp-
observations at regularly spaced time points ftk ¼ nk , k ¼ 1, :::, ng on totic behavior
½0, 1: We obtain the consistency as well as the asymptotic behavior
of the least squares estimator of the unknown parameter h when MATHEMATICS SUBJECT
e ! 0 and n ! 1 simultaneously. CLASSIFICATION 2000
60G18; 60G22;
65C30; 93E24

1. Introduction
The Ornstein–Uhlenbeck process has a remarkable history in physics. It is introduced to
model the velocity of the particle diffusion process, and later the statistical inference for
the Ornstein–Uhlenbeck process driven by Brownian motion is a main research direction
in probability theory and its applications. The most popular approaches are either the
maximum likelihood estimators (MLE) or the least squares estimators (LSE), and in this
case they coincide (see, for example, Kutoyants 2004; Basawa and Scott 2012; Jiang and
Dong 2015). The drift parameter estimation problem for the Ornstein–Uhlenbeck process
driven by fractional Brownian motion (for short, fractional Ornstein–Uhlenbeck process)
has been studied in the literature. The maximum likelihood estimator was first studied in
Kleptsyna and Le Breton (2002), where the almost sure convergence of the MLE was
proved. A least squares approach has been proposed in Hu and Nualart (2010), they
obtained the strong consistence and the rate of this convergence applying a central limit
theorem for multiple Wiener integrals. Some surveys for fractional Ornstein–Uhlenbeck
process could be found in Xiao, Zhang, and Xu (2011), Hu, Nualart, and Zhou (2019),
Xiao and Yu (2019a,b), Sun and Yan (2018) and the references therein.
On the other hand, the asymptotic theory of parametric estimation for diffusion proc-
esses with small noise based on discrete time observations is well developed. Genon-

CONTACT Guangjun Shen gjshen@163.com Department of Mathematics, Anhui Normal University, Wuhu
241000, China.
ß 2019 Taylor & Francis Group, LLC
2 Q. WANG ET AL.

Catalot (1990) and Laredo (1990) studied the efficient estimation of drift parameters of
small diffusions from discrete observations when  ! 0 and n ! 1. Sorensen and
Uchida (2003) and Gloter and Sorensen (2009) studied the efficient estimation for
unknown parameters in both drift and diffusion coefficient functions using a contrast
function. Thus, in the case of small diffusions, the asymptotic distributions of the esti-
mators are normal under suitable conditions on  and n. Hu and Long (2009) discussed
the long time behavior of the LSE of discretely observed Ornstein–Uhlenbeck process
driven by a-stable L
evy noises with fixed dispersion coefficient  ¼ 1 under ergodicity
condition h > 0. Long (2009), Ma (2010) studied the problem of parameter estimation
for Ornstein–Uhlenbeck processes with small L
evy noises based on discrete observations
when  ! 0 and n ! 1 simultaneously. Long, Shimizu, and Sun (2013) discussed the
statistical inference of the drift parameter for a class of small L
evy driven SDEs with
drift function bðx, hÞ to be nonlinear in both x and h. For the case of dispersion coef-
fcient function rðXÞ, Long, Ma, and Shimizu (2017) obtained consistency and rate of
convergence of the LSE of parameter under certain regularity conditions. Ma and Yang
(2014) studied small noise fluctuations of the CIR model driven by a-stable noises.
Zhao and Zhang (2019) considered the minimum distance estimate for stochastic non-
linear differential equations with small a-stable noises. Shen and Yu (2017), Shen, Li,
and Gao (2018) obtained the consistency and the asymptotic distribution of LSE and
the minimum Skorohod distance estimation for Ornstein–Uhlenbeck process driven by
small fractional L
evy noises.
Recently, Bercu, Proïa, and Savy (2014) studied the asymptotic behavior of the max-
imum likelihood estimators of the unknown parameters of Ornstein–Uhlenbeck process
driven by Ornstein–Uhlenbeck process. The motivation comes from two observations.
On the one hand, Ornstein–Uhlenbeck process driven by Ornstein–Uhlenbeck process
are clearly related to stochastic volatility models in financial mathematics (see, for
example, Barndorff-Nielsen and Veraart 2013). On the other hand, Ornstein–Uhlenbeck
process driven by Ornstein–Uhlenbeck process also can be seen as a continuous-time
version of the first-order stable autoregressive process driven by a first-order autoregres-
sive process investigated in Bercu and Proïa (2013). El Onsy, Es-Sebaiy, and G. Viens
(2017) considered the parameter estimation problem about the Ornstein–Uhlenbeck
process driven by a fractional Ornstein–Uhlenbeck process, they obtained the strong
consistency as well as the asymptotic normality of the LSE, based either on continuous
or discrete observations.
However, there has been no study on parametric inference for Ornstein–Uhlenbeck
process with small fractional L
evy driven Ornstein–Uhlenbeck process. Motivated by
the aforementioned works, as a first attempt, in this paper, we consider the generalized
Ornstein–Uhlenbeck process X ¼ ðXt , t  0Þ starting from x0 2 R is defined as the
unique strong solution to the following SDE


dXt ¼ hXt dt þ dYt , X0 ¼ 0
(1.1)
dYt ¼ qYt dt þ edLdt , Y0 ¼ 1

Where Ldt is a fractional L
evy process (the precise definition is given below in
Definition 2.1) and h 6¼ 0. Assume that these processes are observed at n regularly
spaced time points fti ¼ ni , i ¼ 1, 2, :::, ng on ½0, 1, Yt can be observed, then the
 COMMUNICATIONS IN STATISTICS—THEORY AND METHODS 3

estimator of q becomes the usual full observation estimator of a small fractional L
evy
driven Ornstein–Uhlenbeck process as in Shen and Yu (2017), i.e.
Pn
ðYt  Yt ÞYt
^ n,  ¼ i¼11 Pi n 2i1 i1
q (1.2)
n i¼1 Yti1

We see that q ^ n,  is defined explicitly autonomously via the discrete data, for simplify,
we let q ^ n,  ¼ 1, the more complicated situation will be discussed in our forthcom-
ing work.
We are interested in the study of parameter estimation based on discrete observations
ðXti Þni¼1 when e ! 0 and n ! 1. The only unknown quantity in SDE (1.1) is the par-
ameter h. We denote also the true value of the parameter by h. We shall use the least
squares method to obtain an asymptotically consistent estimator ^h n, e . The originality of
the present paper lies in the fact that the driving process is fractional L
evy process
instead of Brownian motion and fractional Brownian motion. Moreover, it is
not Gaussian.
To obtain the LSE, we introduce the following contrast function
X n  ð ti1 2

Un, e ðhÞ ¼ X X ððh þ 1ÞX h X ds þ 1ÞDt 
 t i t i1 t i1 s i1 
i¼1 0

where Dti1 ¼ ti ti1 ¼ n1. Then the LSE ^h n,  is defined as ^h n,  ¼ arg minh Un,  ðhÞ,
which can be explicitly represented as
" ð ti1 #
Xn
i¼1
ðXti Xti1 n1 Xti1 n1 ÞðXti1  Xs dsÞ
^h n, e ¼ ð ti1
0
(1.3)
Xn 2
n1 i¼1 ðXti1  Xs dsÞ
0

We shall consider two problems. Firstly, we shall prove the consistency of ^h n, e , that
is,

^h n, e !
P
h, as e ! 0, n ! 1, en2d ! 0
1

P  
where the notation ! stand for convergence in probability and d 2 0, 12 .
Secondly, we can study the asymptotic behavior of ^h n, e , that is,
Ð1 Ðs
~  ~ r drÞdLds
0 ðX s X
e ð^h n, e hÞ !
1 P 0
Ch

where Ch , X ~ t are given below in Theorem 3.2.

The rest of this paper is organized as follows. In Section 2, we give a detail informa-
tion on integrals with respect to fractional L
evy process, which will be very useful to
our main proof. In Section 3, we study the consistency and asymptotic behavior of the
LSE ^h n,  .
4 Q. WANG ET AL.

2. Preliminaries
2.1. L

evy processes
In this subsection, we mainly introduce the elementary properties of L
evy processes that
will be used in following. More studies on the L
evy processes can be found in Sato
(1999) and the references therein.
Let L ¼ fLðtÞgt0 be a L
evy process in R without Brownian component. It is deter-
mined by its characteristic function in the L
evy-Khintchine form
E½exp fiuLðtÞg ¼ exp ftwðuÞg, t0
where
ð
wðuÞ ¼ icu þ ðeiux  1  iux1fjxj1g ÞðdxÞ, u2R
R

here c 2 R and  is the L
evy measure of L on R that satisfies

ð ð
ðf0gÞ ¼ 0, ðjxj2 Ù 1ÞðdxÞ < 1, jxj2 ðdxÞ < 1
R jxj > 1

This is a necessary and sufficient condition (see Sato (1999), Example 25.12) for L to
have finite mean and variance given by
ð
varðLðtÞÞ ¼ tvarðLð1ÞÞ ¼ t x2 ðdxÞ, t  0
R

Furthermore, we restrict E½Lð1Þ ¼ 0, then

ð
c¼ xðdxÞ
jxj > 1

and
ð
wðuÞ ¼ ðeiux  1  iuxÞðdxÞ, u 2 R
R

Throughout this paper we will use a two-sided L
evy process L ¼ fLðtÞgt2R constructed by
taking two independent copies fL1 ðtÞgt0 , fL2 ðtÞgt0 of a one-sided L
evy process and putting


L1 ðtÞ, t0
LðtÞ ¼ (2.1)
L2 ðt Þ, t < 0

2.2. Fractional L
evy processes

In this subsection, we briefly recall the definition and properties of fractional L
evy process.
As an extension of fractional Brownian motion, fractional L
evy process is of interest
in practical applications because of its stationarity of increments and long range
dependence. However, it is not Gaussian. Actually the very large utilization of the frac-
tional Brownian motion in practice (hydrology, telecommunications) are due to these
properties (long range dependence). One prefers in general fractional Brownian motion
before other processes because it is Gaussian and the calculus for it is easier. However
 COMMUNICATIONS IN STATISTICS—THEORY AND METHODS 5

in concrete situations when the Gaussianity is not plausible for the model, one can use
for example the fractional L
evy process. There exists a consistent literature that focuses
on different theoretical and applications aspects of the fractional L
evy process. For
example, Bender, Lindner, and Schicks (2012) studied the finite variation of fractional
L
evy processes, Tikanm€aki and Mishura (2011) define fractional L
evy processes using
the compact interval representation, and proved that the fractional L
evy processes pre-
sented via different integral transformations have the same finite dimensional distribu-
tions if and only if they are fractional Brownian motions. More studies on the fractional
L
evy processes can be found in Benassi, Cohen, and Istas (2002, 2004), Lacaux (2004),
Engelke and Woerner (2013), Bender, Knobloch, and Oberacker (2015), Glaser (2015)
and the references therein.
Based on the moving average integral representation of fractional Brownian motion,
the class of fractional L
evy processes is introduced by replacing the Brownian motion
by a general L
evy process with zero mean, finite variance.
Definition 2.1. (Marquardt 2006) Let L ¼ ðLðtÞÞt2R be a zero-mean two-sided L
evy pro-
cess with E½Lð1Þ2  < 1 and without a Brownian component. For fractional integration
 
parameter d 2 0, 12 , a stochastic process
ð1 h i
1 d d
Lt :¼
d
ðtsÞþ ðsÞþ LðdsÞ, t 2 R
Cðd þ 1Þ 1
is called a fractional L
evy process (fLp).
fLp has the following sample path properties.

 H€older continuity. For every b < d, there exists a continuous modification of Ld

and there exists an almost surely positive random variable He and a constant
d > 0 such that
" #
Ldtþh ðxÞLdt ðxÞ
P x2X: sup ð Þd ¼1
0 < h < H ðxÞ hb

 Stationary increments. Ld is a process with stationary increments.

 Symmetry. fLdt gt2R ¼d fLdt gt2R .
 ~ that is, for every fixed t 2 R,
Ld is locally self-similar with parameter H,

d 
Ltþex Ldt D
lim ¼ fYH~ ðxÞg, x 2 R
e!0 H~
D
here ¼ denotes equality in (all finite-dimensional) distribution(s), YH~ is a linear
fractional stable motion with expression
ð h i
1 ~ 1
H ~ 1
H
YH~ ðxÞ ¼ ðxsÞþ ðsÞþ La ðdsÞ
a a

CðdÞ R
where La is symmetric a-stable L
evy process, 1 < a < 2, H ~ ¼ d þ 1 such
a
~ < 1.
that 0 < H
The following two lemmas give a relationship between integral with respect to fractional
L
evy process and L
evy process and the second-order property of the stochastic integral
respect to fractional L
evy process.
6 Q. WANG ET AL.

Lemma 2.1. (Marquardt 2006) Let g 2 H, H is the completion of L1 ðRÞ \ L2 ðRÞ with
Ð
respect to the norm jjgjj2H ¼ E½Lð1Þ2  R ðId gÞ2 ðuÞdu, then
ð ð
gðsÞdLs ¼ ðId gÞðuÞdLðuÞ
d
(2.2)
R R
2
where the equality holds in the L sense and Id g denotes the Riemann-Liouville fractional
integrals defined by
ð
1 1
ðI gÞðxÞ ¼
d
gðtÞðtxÞd1 dt
CðdÞ x

Lemma 2.2. (Marquardt 2006) Let jf j, jgj 2 H. Then

ð ð  ð ð
Cð12dÞE Lð1Þ2
E gðsÞdLs f ðsÞdLt ¼
d d
f ðtÞgðsÞjtsj2d1 dsdt (2.3)
R R CðdÞCð1  dÞ R R

3. Consistency and asymptotic distribution of the LSE

In this section, we will consider the consistency and asymptotic distribution of the LSE ^h n, e :
It follows from (1.1) that Xt solves the following stochastic integro-differential equation
ðt
dXt ¼ ððh þ 1ÞXt h Xs ds þ 1Þdt þ edLdt (3.1)
0

By (3.1), we can obtain

ð ti ðs
Xti Xti1 ¼ ððh þ 1ÞXs h Xr dr þ 1Þds þ eðLdti Ldti1 Þ
ti1 0

It follows from (1.3), we can rewrite the ^h n, e

" ð ðs ð ti ð ti ð ti1 #
Xn ti

i¼1
ðh ðXs  Xr drÞds þ ðXs  Xti1 Þds þ e dLs ÞðXti1 
d
Xs dsÞ
ti1 0 ti1 ti1 0
^
h n, e ¼ ð ti1
Xn
n1 i¼1
ðXti1  Xs dsÞ2
" ð ti1 ð ti
0
# " ð ti1 ð ti #
Xn Xn
i¼1
ðXti1  Xs dsÞ ðXs  Xti1 Þds e i¼1 ðXti1  Xs dsÞ dLds
0 ti1 0 ti1
¼hþ Xn ð ti1 þ Xn ð ti1
n1 i¼1
ðXti1  Xs dsÞ2 n1 i¼1
ðXti1  Xs dsÞ2
" ð ti1 ð ti
0
ðs # 0
Xn
h i¼1 ðXti1  Xs dsÞ ðXs  Xti1  Xr drÞds
0 ti1 ti1
þ Xn ð ti1
n1 i¼1
ðXti1  Xs dsÞ2
0
K2 ðn, eÞ K3 ðn, eÞ K4 ðn, eÞ
:¼ h þ þ þ
K1 ðn, eÞ K1 ðn, eÞ K1 ðn, eÞ
(3.2)
 COMMUNICATIONS IN STATISTICS—THEORY AND METHODS 7

In order to prove the main results, we need the following lemmas.

Lemma 3.1. (Liptser and Shiryaev 1978) Let c0, c1, c2 be nonnegative constant, u(t) be a
nonnegative bounded function, and v(t) be a nonnegative function, 0  t  1, such that
ðt ðt  ðs 
uðtÞ  c0 þ c1 vðsÞuðsÞds þ c2 vðsÞ uðs1 Þds1 ds
0 0 0

then

ðt 
uðtÞ  c0 exp ðc1 þ c2 Þ vðsÞds
0

~ t be the solution of the following differential equations,

Let X


dXt ¼ hXt dt þ dYt , X0 ¼ 0
(3.3)
dYt ¼ Yt dt, Y0 ¼ 1
Hence,

1
~t ¼ ðet eht Þ, h 6¼ 1
X 1h (3.4)
tet , h¼1
and
8
ðt < 1 1 1
~ s ds ¼ et  eht þ , h 6¼ 1
:1 
X h h h (3.5)
0
tet et þ 1, h¼1
 
Lemma 3.2. Suppose that d 2 0, 12 , for any t 2 ½0, 1, we have,
 ðt 
 
jXt X~ t j  e exp ðjh þ 1jt þ jhjtÞ sup  d 
 dLs 
0t1 0
 ðt ðt   ðt 
   
 Xs ds  X ~ s ds  et exp ðjh þ 1j þ jhjÞ sup  dLds 
   
0 0 0t1 0

and
 ðt ðt 
  P
jXt X t j ! 0,  Xs ds  X s ds ! 0,
~  ~
P
e ! 0
0 0

Proof. By (3.1), we have

 ðt ðs ðs ðt 
 
~ 
jXt X t j ¼  ~
ðh þ 1ÞðXs  X s Þ  h ~
Xr dr  X r dr ds þ e dLs  d
0 0 0 0
 ðt   ðt ðs ðs   ðt 
     
  ðh þ 1ÞðXs  X ~ s Þds þ  h
  X r dr  ~
X r dr ds 
 þ e sup 
 dL d
s
0 0 0 0 0t1 0
ðt ðt ðs  ðt 
 
 jh þ 1j jXs X ~ s jds þ jhj jXr X~ r jdrds þ e sup  dLds 
 
0 0 0 0t1 0
8 Q. WANG ET AL.

Hence,
 ðt 
 
~ t j  e exp ðjh þ 1jt þ jhjtÞ sup  dLds 
jXt X  
0t1 0

since Lemma 3.1. Thus,

 ðt ðt  ðt  ðt 
   
 Xs ds  X ~ s jdt  et exp ðjh þ 1j þ jhjÞ sup  dLds 
~ s ds  jXs X
   
0 0 0 0t1 0

Therefore, by Chebyshev inequality and Lemma 2.2, we have, for all r > 0,

~ t j > rÞ  r2 EjXt X

PðjXt X ~ t j2
 ðt 2
 
 r2 e2 exp ð2jh þ 1jt þ 2jhjtÞ sup E dLds 
0t1 0
ðt ðt
 Cr2 e2 exp ð2jh þ 1jt þ 2jhjtÞ sup jsvj2d1 dsdv
0t1 0 0

 Cr2 e2 exp ð2jh þ 1jt þ 2jhjtÞ ! 0, as e ! 0

Hence,
~ tj !
jXt X
P
0, e ! 0
And apply similar techniques, we have
 ðt ðt 
  P
 Xs ds  X ~ s ds ! e ! 0
  0,
0 0

This completes the proof.

 
Lemma 3.3. Let d 2 0, 12 , for any ti1  t  ti , ti ¼ ni , i ¼ 1, 2, :::, n, we have
" ð ti1  ðt #
1 1 1  
jXt Xti1 j  2 jhXti1 j þ jðh þ 1ÞXti1  h Xs dsj þ þ e sup   d
dLs 
n n 0 n ti1 tti ti1

1
 exp ðjh þ 1j þ jhjÞ
n

and as n ! 1
 ðt 
 
jXt Xti1 j  e sup  dLds , a:s:
ti1 tti ti1
 COMMUNICATIONS IN STATISTICS—THEORY AND METHODS 9

Proof. Using (3.1), we have

 ðt ðt ðs ðt 
 

jXt Xti1 j ¼  ðh þ 1ÞXs ds  h Xr drds þ ðt  ti1 Þ þ e d
dLs 
ti1 ti1 0 ti1
ðt ðt ðt ðs
 jðh þ 1Þð Xs  Xti1 ds þ Xti1 dsÞh Xr Xti1 drds
ti1 ti1 ti1 ti1
ðt ðs ð t ð ti1  ðt 
 
h Xti1 drdsh 
Xr drdsj þ ðt  ti1 Þ þ e sup d
dLs 
ti1 ti1 ti1 0 ti1 tti ti1

ðt ðt ðs
 jh þ 1j jXs Xti1 jds þ jhj jXr Xti1 jdrds
ti1 ti1 ti1
 ðt ðs   ðt ð ti1 
   

þ h  
Xti1 drds þ  ððh þ 1ÞXti1  h Xr drÞds
ti1 ti1 ti1 0
 ðt 
 

þ ðt  ti1 Þ þ e sup d
dLs 
ti1 tti ti1

ðt ðt ðs
 jh þ 1j jXs Xti1 jds þ jhj jXr Xti1 jdrds
ti1 ti1 ti1
ð ti1  ðt 
1 1 1  
þ 2 jhXti1 j þ jðh þ 1ÞXti1 h Xr drj þ þ e sup   d
dLs 
n n 0 n ti1 tti ti1

It follows from the Lemma 3.1, we have

" ð ti1  ðt #
1 1 1  
jXt Xti1 j  2 jhXti1 j þ jðh þ 1ÞXti1  h Xr drj þ þ e sup  dLs 
d
n n 0 n ti1 tti ti1

 exp ððtti1 Þðjh þ 1j þ jhjÞÞ

This completes the proof. w

The following theorem gives the consistency of the ^h n, e .

 
Theorem 3.1. Let d 2 0, 12 . As e ! 0, n ! 1, and en2d ! 0, we have
1

^h n, e !
P
h

Proof. By (3.2), we will study the asymptotic behavior of K1 ðn, eÞ, K2 ðn, eÞ, K3 ðn, eÞ and
K4 ðn, eÞ respectively.
P
Step one. As n ! 1, e ! 0, we have K1 ðn, eÞ ! Ch
10 Q. WANG ET AL.

In fact, by (3.2) we have

K1 ðn, eÞ
ð ti1 !2
1X n
¼ Xti1  Xs ds
n i¼1 0
" ð ti1 ! ð ti1 ð ti1 !#2
1X n
~ ti1  ~ s ds þ ðXti1 X ~ ti1 Þ ~ s ds
¼ X X Xs ds X
n i¼1 0 0 0
ð ti1
1X n
~ ~ s dsÞÞ2 þ 1
X n
~ ti1 Þ2
¼ ððX ti1  X ðXti1 X
n i¼1 0 n
ð ð !i¼1
2 " ð ti1 ! #
1X n t i1 t i1
~ s ds þ 2 X n
~ ti1  ~ s ds ðXti1 X ~ ti1 Þ
þ Xs ds X X X
n i¼1 0 n i¼1
" ð ti1
0
! ð ð ti1 !#
0

2X n
~ ti1  ~ s ds
ti1
~ s ds
 X X Xs ds X
n i¼1
" 0
ð ti1
0
ð ti1
0
!#
2X n
~ ti1 Þ ~ s ds
 ðXti1 X Xs ds X
n i¼1 0 0
:¼ K11 ðn, eÞ þ K21 ðn, eÞ þ K31 ðn, eÞ þ K41 ðn, eÞK51 ðn, eÞK61 ðn, eÞ
a:s: Ð 1 Ð
It is easy to see that as n ! 1, K11 ðn, eÞ ! 0 ðX ~ s s X ~ 2
0 r drÞ ds, and from (3.4) and
(3.5) we get
a:s:
K11 ðn, eÞ ! Ch
where
8
>
> 1 1
< 3 ðeh 3Þðeh 1Þ þ 2 , h 6¼ 1
Ch ¼ 12h h (3.6)
>
>
: ðe3Þðe1Þ þ 1, h¼1
2
Combining Lemma 2.2 with Lemma 3.2, we have, for any r > 0,
 X 
1 n 
PðjK21 ðn, eÞj > dÞ  d1 E ðXti1 X~ ti1 Þ2 

n i¼1
 ðt 2
 
1 ~ 2 1 2
 d Ej sup ðXt X t Þ j  d e sup E dLs  d
0t1 0t1 0
ðt ðt
 Cd1 e2 sup jsvj2d1 dsdv  Cd1 e2 ! 0, n ! 1, e ! 0
0t1 0 0

 X ð ti1 ð ti1 !2 
 1 n 
PðjK31 ðn, eÞj > dÞ  d1 E Xs ds ~ s ds 
X 
n i¼1 0 0
 ðt ðt !2 
1 
 
 d E sup Xs ds X ~ s ds   d1 E sup jXt X
~ t j2

0t1 0 0 0t1
1 2
 Cd e ! 0 , n ! 1, e ! 0
 COMMUNICATIONS IN STATISTICS—THEORY AND METHODS 11

 " ð ti1 #2

2X n 

2 ~ ti1  ~ ti1 Þ 
~ s dsÞðXti1  X
PðjK41 ðn, eÞj > dÞ  d E ðX X
 n i¼1 0 
 ð ti1 2

 2Xn 
2 ~
 d E sup ðXt  X t Þ ~
ðX ti1  X s dsÞ  Cd2 Ej sup ðXt X
~ ~ t Þj2
0t1 n i¼1 0 0t1

 Cd2 e2 ! 0, n ! 1, e ! 0

 " ð ti1 ð ti1 ð ti1 !#2

2X n 
2  ~ ti1  ~ s dsÞ ~ s ds 
PðjK51 ðn, eÞj > dÞ  d E ðX X Xs ds  X
 n i¼1 0 0 0 
 ðt ðt !
 2
 Cd2 E sup Xs ds X ~ s ds   Cd2 e2 ! 0, n ! 1,
 e ! 0
0t1 0 0

 " ð ti1 ð ti1 !#

2X n 

1 ~ ti1 Þ ~ s ds 
PðjK61 ðn, eÞj > dÞ  d E ðXti1  X Xs ds  X
 n i¼1 0 0 
 2 !2 3
1X n ð ti1 ð ti1 
 4ðX X ~ ~ 5
 d1 E ti1 ti1 Þ2
þ X s ds X s ds
 n i¼1 0 0 

 Cd1 e2 ! 0, n ! 1, e ! 0

Hence, as n ! 1, e ! 0
P
K1 ðn, eÞ ! Ch (3.7)

P
Step two. As n ! 1, e ! 0, we have K2 ðn, eÞ ! 0

In fact,
n  ð ti1  ð ti 
X 
jK2 ðn, eÞj   
Xti1  Xs ds jXs  Xti1 jds
i¼1 0 ti1
" ð ti #
X n
 jXti1  X ~ ti1 j jXs  Xti1 jds
i¼1 ti1
n  ð ti1
X ð ti1  ð ti 
 (3.8)
þ  Xs ds  ~ s ds
X  jXs  Xti1 jds
 
i¼1 0 0 ti1
n 
X ð ti1  ð ti 

þ  ~ ~ 
X ti1  X s ds jXs  Xti1 jds
i¼1 0 ti1

:¼ K12 ðn, eÞ þ K22 ðn, eÞ þ K32 ðn, eÞ

12 Q. WANG ET AL.

Using the Lemma 3.3 and t 2 ½0, 1, we have

n  
1X ~ ti1 j sup jXt  Xti1 j
K12 ðn, eÞ  jXti1  X
n i¼1 ti1 tti
" " ð ti1 ##
en X
C n
~ ti1 j sup 1 1 1
 jXti1  X 2
jhXti1 j þ jðh þ 1ÞXti1  h Xs dsj þ
n i¼1 ti1 tti n n 0 n
  ð 
1 CX n 
Xt  X
 t


þ e en  i1
~ t i1
j sup  dLds 
n i¼1 ti1 tti ti1
" " ðt ##
C
~ t j sup 1 1 1
 en sup jXt  X 2
jhXt j þ jðh þ 1ÞXt  h Xs dsj þ
0t1 0t1 n n 0 n
  ðt 
 
C
þ ee n sup jXt  X ~ t j sup  dLds 

0t1 ti1 tti ti1
(3.9)
Obviously, as e ! 0, n ! 1, by Lemma 3.2, we know for any t 2 ½0, 1,
~ tj !
jXt X
P
0, then we have
P
K12 ðn, eÞ ! 0, as n ! 1, e ! 0
P
Similar to the discussion K12 ðn, eÞ, we can obtain K22 ðn, eÞ ! 0, n ! 1, e ! 0:
Moreover,
" ð ti1 #
1X n
~ ti1  ~ s dsj sup jXt  Xti1 j
K2 ðn, eÞ 
3
jX X
n i¼1 0 ti1 tti

" ð ti1 # " ðt #

1XC
n
~t  ~ s dsj sup 1 jhXt j þ 1 jðh þ 1ÞXt  h 1
e n jX X Xs dsj þ
n i¼1 i1 0 0t1 n
2 n 0 n
n  ð ti1  ð ti 
1X C
~ ~

 d

þ ee n jX ti1  X s dsj sup  dLs 
n i¼1 0 ti1 tti ti1

 ð1 ðt  " ðt #
C
~t  X ~ r drjds sup 1 1 1
 Ce n jX 2
jhXt j þ jðh þ 1ÞXt  h Xs dsj þ
0 0 0t1 n n 0 n
n  ð ti1  ð ti 
1X C
~ ~

 d

þ ee n jX ti1  X s dsj dLs 
n i¼1 0 ti1

(3.10)
For the first term, as n ! 1,
 ð1 ðt  " ðt #
C
~t  X ~ r drjds sup 1 1 1 a:s:
Ce n jX 2
jhXt j þ jðh þ 1ÞXt  h Xs dsj þ ! 0
0 0 0t1 n n 0 n
 COMMUNICATIONS IN STATISTICS—THEORY AND METHODS 13

For second term, by Lemma 2.2, we have

 ð ti  !
n  ð ti1
1X
C
~ ti1  ~ s dsj




P ee n jX X  dLs  > d
d
n i¼1 0 ti1
 ð ti1  ð ti 
Xn   
1 Cn 1
 d ee E X ~ ti1  ~ s ds
X  dLs 
d
n i¼1 0 ti1
" ð  ð 2 #
Xn ti1
1
ti 
1 Cn 1 ~ s dsjE2 
 d ee ~ ti1 
jX X  dLds 
n i¼1 0 ti1

X  ð 
n ti1 
 d1 Cee n nd2 X ~ ti1  ~ s ds
C 3
X 
i¼1 0
1 C
d12
 d Cee n n ! 0, as n ! 1, e ! 0

Hence,
P
jK2 ðn, eÞj ! 0, n ! 1, e ! 0

P
Step Three. As n ! 1, e ! 0 and en2d ! 0, we have K3 ðn, eÞ ! 0.
1

In fact,

Xn  ð ti1  ð ti 
 
jK3 ðn, eÞj  e  Xt  X ds  dL d
 i1 s  s
i¼1 0 ti1

n   ð ti 
X  
e ~ 
jXti1  X ti1 j d
dLs 
i¼1 ti1

Xn  ð ti1 ð ti1  ð ti 

 
þe  X ds  ~
X ds  dLd
 s s  s
i¼1 0 0 ti1

n  ð ti1  ð ti 
X  
þe  ~ ~  d
X ti1  X s ds dLs 
i¼1 0 ti1

:¼ K13 ðn, eÞ þ K23 ðn, eÞ þ K33 ðn, eÞ

We find that
n  ð ti
X


~ tj
K13 ðn, eÞ  e sup jXt X  dL d
 s
0t1 i¼1 ti1
14 Q. WANG ET AL.

Combining Lemma 3.2 and following truth for any d > 0,

Xn  ð ti 
 Xn  ð ti



P e   d 1
dLs  > d  d e E  d
dLs 
i¼1 ti1 i¼1 ti1
X  ð ti 2
1 
n
1
d e E 2  d
dLs 
i¼1 ti1

 Cd1 en 2d
1
! 0
P
as n ! 1, e ! 0 and en2d ! 0, we can get K13 ðn, eÞ ! 0
1

P
Similar to the discussion K13 ðn, eÞ, we obtain K23 ðn, eÞ ! 0. Moreover, for any given
d > 0,
 Xn  ð ti1   ð ti 
  
P jK3 ðn, eÞj > d  d e 1  ~ ~   d
X ti1  X s dsE
3
dLs 
i¼1 0 ti1
X
n ð ti1
 Cd1 en2d
1
~ ti1 
jX ~ s dsj
X
i¼1 0

 Cd1 en2d ! 0
1

as n ! 1, e ! 0 and en2d ! 0. Then, we get

1

P
en2d ! 0
1
K3 ðn, eÞ ! 0, n ! 1, e ! 0,

P
Step Four. As n ! 1, e ! 0 and en2d ! 0, we have K4 ðn, eÞ ! 0
1

Indeed,
jK4 ðn, eÞj  ð ti 
X ð ti1 ðs
n  
 jhj jXti1  Xs dsj ðXs  Xti1  Xr drÞds
0 t t
n  ð ti1 ð ti i1  i1
i¼1
X
 jhj jXti1  Xs dsj jXs  Xti1 jds
0 ti1
n  ð ti1 ð ti ð s 
i¼1
X
þ jhj jXti1  Xs dsj jXu  Xti1 jduds
0 ti1 ti1
i¼1
X n  ð ti1 ð ti 
þ jhj jXti1  Xs dsj jXti1 ðs  ti1 Þjds
i¼1 0 ti1
:¼ K14 ðn, eÞ þ K4 ðn, eÞ þ
2
K4 ðn, eÞ
3

We can easily get

n 
X ð ti1 ð ti ð s 
K24 ðn, eÞ  jhj jXti1  Xs dsj jXu  Xti1 jduds
i¼1 0 ti1 ti1
n 
X ð ti1 
1
 jhj 2 jXti1  Xs dsj sup jXu  Xti1 j
n i¼1 0 ti1 uti
 COMMUNICATIONS IN STATISTICS—THEORY AND METHODS 15

and
ð ti1
jhj X
n
K34 ðn, eÞ  2 jXt  Xs dsjjXti1 j
n i¼1 i1 0

Similarly to the proof of K2 ðn, eÞ, we can know as n ! 1, e ! 0,

P P P
K14 ðn, eÞ ! 0, K24 ðn, eÞ ! 0, K34 ðn, eÞ ! 0
Hence, we have
P
K4 ðn, eÞ ! 0, n ! 1, e ! 0

Therefore, we have ^h n, e ! h as e ! 0, n ! 1, en2d ! 0. This completes

P 1

the proof. w

In this following, we will study the asymptotic behavior of h^ n, e .

 
Theorem 3.2. Let d 2 0, 12 . As e ! 0, n ! 1, en2d ! 0 and ne ! 1, we have
1

Ð1 Ð
~s  s X
P 0 ðX ~ d
1 ^ 0 r drÞdLs
e ðh n, e hÞ !
Ch
where
8
< 1
~t ¼ ðet eht Þ, h 6¼ 1
X 1 
: tet , h
h¼1

and
8
>
> 1 1
< 3 ðeh 3Þðeh 1Þ þ 2 , h 6¼ 1
Ch ¼ 12h h
>
>
: ðe3Þðe1Þ þ 1, h¼1
2

Remark. The limit distribution in Theorem 3.2 depends on the driving fractional L
evy
noises, which is quite intractable. However, this is the nature of the model (1.1) since
many of the fractional L
evy process have intractable distributions. In general, an
approximated limit distribution of the estimator is available from an empirical distribu-
tion (or histogram) by using Monte Carlo samples of e1 ð^h n, e h0 Þ

Proof. By (3.2), we have

e1 K2 ðn, eÞ e1 K3 ðn, eÞ e1 K4 ðn, eÞ
e1 ð^h n, e hÞ ¼ þ þ (3.11)
K1 ðn, eÞ K1 ðn, eÞ K1 ðn, eÞ
We firstly calculate the estimators e1 K2 ðn, eÞ which come from Theorem 3.1.
By (3.8), it follows that
e1 K2 ðn, eÞ  e1 K12 ðn, eÞ þ e1 K22 ðn, eÞ þ e1 K32 ðn, eÞ
16 Q. WANG ET AL.

By (3.9), (3.10) and Lemma 3.2, it is not difficult to see that as e ! 0, n ! 1, en2d ! 0
1

P P P
and ne ! 1, e1 K12 ðn, eÞ ! 0, e1 K22 ðn, eÞ ! 0, e1 K32 ðn, eÞ ! 0 respectively. In addition,
P
using the similar way, we can obtain e1 K4 ðn, eÞ ! 0.
Finally, we deal with the term e1 K3 ðn, eÞ

n 
X ð ti1 ð ti 
1
e K3 ðn, eÞ ¼ ðXti1  Xs dsÞ dLds
i¼1 0 ti1
Xn  ð ti 
¼ ~ ti1 Þ
ðXti1  X dLds
i¼1 ti1
Xn  ð ti1 ð ti1 ð ti 
 ð Xs ds  ~ s dsÞ
X dLds
i¼1 0 0 ti1
Xn  ð ti1 ð ti 
þ ~ ti1 
ðX ~ s dsÞ
X dLds
i¼1 0 ti1

:¼ W3 ðn, eÞW3 ðn, eÞ

1 2
þ W33 ðn, eÞ

P P
It come from Lemma 3.2, we get W13 ðn, eÞ ! 0, W23 ðn, eÞ ! 0 as n ! 1, e ! 0, and

n 
X ð ti1 ð ti 
W33 ðn, eÞ ¼ ~ ti1 
ðX ~ s dsÞ
X dLds
i¼1 0 ti1
n ð ti
X ð ti1
¼ ~ ti1 
ðX ~ s dsÞdLds
X
i¼1 ti1 0
ð1 ðs
¼ ~s 
ðX ~ r drÞdLds
X
0 0

By (3.7), we have
Ð1 Ðs
~  ~ r drÞdLds
0 ðX s X
e ð^h n, e hÞ !
1 P 0
Ch

as e ! 0, n ! 1, en2d ! 0 and ne ! 1. This completes the proof.

1

Acknowledgments
The authors are very grateful to the anonymous referee and the editor for their insightful and
valuable comments, which have improved the presentation of the paper.

Funding
This research is supported by the Distinguished Young Scholars Foundation of Anhui Province
(1608085J06), the National Natural Science Foundation of China (11601260).
 COMMUNICATIONS IN STATISTICS—THEORY AND METHODS 17

References
Barndorff-Nielsen, O. E., and A. E. Veraart. 2013. Stochastic volatility of volatility and variance
risk premia. Journal of Financial Econometrics 11 (1):1–46. doi:10.1093/jjfinec/nbs008.
Basawa, I. V., and D. J. Scott. 2012. Asymptotic optimal inference for non-ergodic models. Vol. 17.
New York: Springer Science & Business Media.
Benassi, A., S. Cohen, and J. Istas. 2002. Identification and properties of real harmonizable frac-
tional L
evy motions. Bernoulli 8 (1):97–115.
Benassi, A., S. Cohen, and J. Istas. 2004. On roughness indices for fractional fields. Bernoulli 10 (2):
357–73. doi:10.3150/bj/1082380223.
Bender, C., R. Knobloch, and P. Oberacker. 2015. Maximal inequalities for fractional L
evy and
related processes. Stochastic Analysis and Applications 33 (4):701–14. doi:10.1080/07362994.
2015.1036167.
Bender, C., A. Lindner, and M. Schicks. 2012. Finite variation of fractional L
evy processes.
Journal of Theoretical Probability 25 (2):594–612. doi:10.1007/s10959-010-0339-y.
Bercu, B., and F. Proïa. 2013. A sharp analysis on the asymptotic behavior of the Durbin–Watson
statistic for the first-order autoregressive process. ESAIM: Probability and Statistics 17:500–30.
doi:10.1051/ps/2012005.
Bercu, B., F. Proïa, and N. Savy. 2014. On Ornstein–Uhlenbeck driven by Ornstein–Uhlenbeck
processes. Statistics & Probability Letters 85:36–44. doi:10.1016/j.spl.2013.11.002.
El Onsy, B., K. Es-Sebaiy, and F. G. Viens. 2017. Parameter estimation for a partially observed
Ornstein–Uhlenbeck process with long-memory noise. Stochastics 89 (2):431–68. doi:10.1080/
17442508.2016.1248967.
Engelke, S., and J. H. Woerner. 2013. A unifying approach to fractional L
evy processes.
Stochastics and Dynamics 13 (2):1250017. doi:10.1142/S0219493712500177.
Genon-Catalot, V. 1990. Maximnm contrast estimation for diffusion processes from discrete
observations. Statistics 21 (1):99–116. doi:10.1080/02331889008802231.
Glaser, S. 2015. A law of large numbers for the power variation of fractional L
evy processes.
Stochastic Analysis and Applications 33 (1):1–20. doi:10.1080/07362994.2014.962045.
Gloter, A., and M. Sorensen. 2009. Estimation for stochastic differential equations with a small
diffusion coefficient. Stochastic Processes and Their Applications 119 (3):679–99. doi:10.1016/j.
spa.2008.04.004.
Hu, Y., and H. Long. 2009. Least squares estimator for Ornstein–Uhlenbeck processes driven by
a-stable motions. Stochastic Processes and Their Applications 119 (8):2465–80. doi:10.1016/j.spa.
2008.12.006.
Hu, Y., and D. Nualart. 2010. Parameter estimation for fractional Ornstein–Uhlenbeck processes.
Statistics & Probability Letters 80 (11–12):1030–38. doi:10.1016/j.spl.2010.02.018.
Hu, Y., D. Nualart, and H. Zhou. 2019. Parameter estimation for fractional Ornstein–Uhlenbeck
processes of general hurst parameter. Statistical Inference for Stochastic Processes 22 (1):111–42.
doi:10.1007/s11203-017-9168-2.
Jiang, H., and X. Dong. 2015. Parameter estimation for the non-stationary Ornstein–Uhlenbeck
process with linear drift. Statistical Papers 56 (1):257–68. doi:10.1007/s00362-014-0580-z.
Kleptsyna, M., and A. Le Breton. 2002. Statistical Inference for Stochastic Processes 5 (3):229–48.
doi:10.1023/A:1021220818545.
Kutoyants, Y. A. 2004. Statistical inference for ergodic diffusion processes. Berlin, Heidelberg:
Springer.
Lacaux, C. 2004. Real harmonizable multifractional L
evy motions. Annales de L’institut Henri
Poincare (B) Probability and Statistics 40 (3):259–77. doi:10.1016/S0246-0203(03)00064-5.
Laredo, C. F. 1990. A sufficient condition for asymptotic sufficiency of incomplete observations
of a diffusion process. The Annals of Statistics 18 (3):1158–71. doi:10.1214/aos/1176347744.
Liptser, R. S., and A. N. Shiryaev. 1978. Statistics of random processes. New York: Springer.
Long, H. 2009. Least squares estimator for discretely observed Ornstein–Uhlenbeck processes
with small L
evy noises. Statistics & Probability Letters 79 (19):2076–85. doi:10.1016/j.spl.2009.
06.018.
18 Q. WANG ET AL.

Long, H., C. Ma, and Y. Shimizu. 2017. Least squares estimators for stochastic differential equa-
tions driven by small L
evy noises. Stochastic Processes and Their Applications 127 (5):1475–95.
doi:10.1016/j.spa.2016.08.006.
Long, H., Y. Shimizu, and W. Sun. 2013. Least squares estimators for discretely observed stochas-
tic processes driven by small L
evy noises. Journal of Multivariate Analysis 116:422–39. doi:10.
1016/j.jmva.2013.01.012.
Ma, C. 2010. A note on least squares estimator for discretely observed Ornstein–Uhlenbeck proc-
esses with small L
evy noises. Statistics & Probability Letters 80 (19–20):1528–31. doi:10.1016/j.
spl.2010.06.006.
Ma, C., and X. Yang. 2014. Small noise fluctuations of the cir model driven by a-stable noises.
Statistics & Probability Letters 94:1–11. doi:10.1016/j.spl.2016.09.014.
Marquardt, T. 2006. Fractional L
evy processes with an application to long memory moving aver-
age processes. Bernoulli 12 (6):1099–1126. doi:10.3150/bj/1165269152.
Sato, K.-i. 1999. L
evy processes and infinitely divisible distributions. Cambridge: Cambridge
University Press.
Shen, G., Y. Li, and Z. Gao. 2018. Parameter estimation for Ornstein–Uhlenbeck processes driven
by fractional L
evy process. Journal of Inequalities and Applications 2018 (1):356. doi:10.1186/
s13660-018-1951-0.
Shen, G., and Q. Yu. 2017. Least squares estimator for Ornstein–Uhlenbeck processes driven by
fractional L
evy processes from discrete observations. Statistical Papers 1–19.
Sorensen, M., and M. Uchida. 2003. Small-diffusion asymptotics for discretely sampled stochastic
differential equations. Bernoulli 9 (6):1051–69. doi:10.3150/bj/1072215200.
Sun, X., and L. Yan. 2018. Central limit theorems and parameter estimation associated with a
weighted-fractional brownian motion. Journal of Statistical Planning and Inference 192:45–64.
doi:10.1016/j.jspi.2017.07.001.
Tikanm€aki, H., and Y. Mishura. 2011. Fractional L
evy processes as a result of compact interval
integral transformation. Stochastic Analysis and Applications 29 (6):1081–1101. doi:10.1080/
07362994.2011.610172.
Xiao, W., and J. Yu. 2019a. Asymptotic theory for estimating drift parameters in the fractional
vasicek model. Econometric Theory 35 (1):198–231. doi:10.1017/S0266466618000051.
Xiao, W., and J. Yu. 2019b. Asymptotic theory for rough fractional vasicek models. Economics
Letters 177:26–9. doi:10.1016/j.econlet.2019.01.020.
Xiao, W., W. Zhang, and W. Xu. 2011. Parameter estimation for fractional Ornstein–Uhlenbeck
processes at discrete observation. Applied Mathematical Modelling 35 (9):4196–207. doi:10.
1016/j.apm.2011.02.047.
Zhao, H., and C. Zhang. 2019. Minimum distance parameter estimation for sdes with small a-sta-
ble noises. Statistics & Probability Letters 145:301–11. doi:10.1016/j.spl.2018.10.009.