DEVIATION INEQUALITIES AND MODERATE DEVIATIONS FOR PROCESS WITH LINEAR DRIFT

advertisement
Elect. Comm. in Probab. 14 (2009), 210–220
ELECTRONIC
COMMUNICATIONS
in PROBABILITY
DEVIATION INEQUALITIES AND MODERATE DEVIATIONS FOR
ESTIMATORS OF PARAMETERS IN AN ORNSTEIN-UHLENBECK
PROCESS WITH LINEAR DRIFT
FUQING GAO1
School of Mathematics and Statistics, Wuhan University, Wuhan 430072, P.R.China
email: fqgao@whu.edu.cn
HUI JIANG
School of Science, Nanjing University of Aeronautics and Astronautics, Nanjing 210016, P.R.China
email: huijiang@nuaa.edu.cn
Submitted December 29, 2008, accepted in final form April 21, 2009
AMS 2000 Subject classification: 60F12, 62F12, 62N02
Keywords: Deviation inequality, logarithmic Sobolev inequality, moderate deviations, OrnsteinUhlenbeck process
Abstract
Some deviation inequalities and moderate deviation principles for the maximum likelihood estimators of parameters in an Ornstein-Uhlenbeck process with linear drift are established by the
logarithmic Sobolev inequality and the exponential martingale method.
1 Introduction and main results
1.1 Introduction
We consider the following Ornstein-Uhlenbeck process
d X t = (−θ X t + γ)d t + dWt ,
X0 = x
(1.1)
where W is a standard Brownian motion and θ , γ are unknown parameters with θ ∈ (0, +∞). We
denote by Pθ ,γ,x the distribution of the solution of (1.1).
It is known that the maximum likelihood estimators (MLE) of the parameters θ and γ are (cf.
1
RESEARCH SUPPORTED BY THE NATIONAL NATURAL SCIENCE FOUNDATION OF CHINA (10871153)
210
Deviation inequalities and MDP for estimators in OU model
211
[15])
RT
X t d X t + (X T − x) 0 X t d t
θ̂ T =
R T
2
RT
T 0 X t2 d t − 0 X t d t
RT
WT µ̂ T − 0 X t dWt
,
=θ +
T σ̂2T
−T
RT
0
RT
RT
X t d t 0 X t d X t + (X T − x) 0 X t2 d t
γ̂ T =
2
R T
RT
T 0 X t2 d t − 0 X t d t
RT
µ̂ T (WT µ̂ T − 0 X t dWt )
WT
=γ +
+
,
T
T σ̂2T
−
RT
0
where
µ̂ T =
1
T
(1.2)
T
Z
X t d t,
0
σ̂2T
=
1
T
(1.3)
T
Z
0
X t2 d t − µ̂2T .
(1.4)
It is known that θ̂ T and γ̂ T are consistent estimators of θ and γ and have asymptotic normality
(cf. [15]).
RT
For γ ≡ 0 case, Florens-Landais and Pham([9]) calculated the Laplace functional of ( 0 X t d X t ,
RT
X t2 d t) by Girsanov’s formula and obtained large deviations for θ̂ T by Gärtner-Ellis theorem.
0
Bercu and Rouault ([1]) presented a sharp large deviation for θ̂ T . Lezaud ([14]) obtained the
deviation inequality of quadratic functional of the classical OU processes. We refer to [8] and
[11] for the moderate deviations of some non-linear functionals of moving average processes and
diffusion processes. In this paper we use the logarithmic Sobolev inequality (LSI) to study the
deviation inequalities and the moderate deviations of θ̂ T and γ̂ T for γ 6= 0 case.
1.2 Main results
Throughout this paper, let λ T , T ≥ 1 be a positive sequence satisfying
λ T → ∞,
λT
p → 0.
T
(1.5)
Theorem 1.1. There exist finite positive constants C0 , C1 , C2 and C3 such that for all r > 0 and all
T ≥ 1,
€
Š
¦
©
Pθ ,γ,x |θ̂ T − θ | ≥ r ≤C0 exp −C1 r T Eθ ,γ,x (σ̂2T ) min 1, C2 r
¦
©
+ C0 exp −C3 T Eθ ,γ,x (σ̂2T )
and
¦
©
Pθ ,γ,x |γ̂ T − γ| ≥ r ≤C0 exp −C1 r T Eθ ,γ,x (σ̂2T ) min 1, C2 r
¦
©
+ C0 exp −C3 T Eθ ,γ,x (σ̂2T ) .
Remark 1.1. In this theorem and the remainder of the paper, all the constants involved depend on
θ , γ and the initial point x.
212
Electronic Communications in Probability
q
Theorem 1.2. (1).
Pθ ,γ,x
T
λT
speed λ T and rate function I1 (u) =
(θ̂ T − θ ) ∈ · , T ≥ 1 satisfies the large deviation principle with
u2
,
4θ
that is, for any closed set F in R,
!
r
1
u2
T
lim sup
log Pθ ,γ,x
(θ̂ T − θ ) ∈ F ≤ − inf
u∈F 4θ
λT
n→∞ λ T
and open set G in R,
lim inf
n→∞
q
(2).
Pθ ,γ,x
r
1
λT
log Pθ ,γ,x
T
λT
λT
rate function I2 (u) =
θu
,
2(θ +2γ2 )
lim sup
n→∞
1
λT
!
(θ̂ T − θ ) ∈ G
≥ − inf
u∈G
u2
4θ
.
(γ̂ T − γ) ∈ · , T ≥ 1
2
T
satisfies the large deviation principle with speed λ T and
that is, for any closed set F in R,
r
log Pθ ,γ,x
!
T
λT
(γ̂ T − γ) ∈ F
≤ − inf
u∈F
θ u2
2(θ + 2γ2 )
and open set G in R,
lim inf
n→∞
1
λT
r
log Pθ ,γ,x
T
λT
!
(γ̂ T − γ) ∈ G
≥ − inf
u∈G
θ u2
2(θ + 2γ2 )
.
In γ = 0 case, the deviation inequalities of quadratic functionals of the classical OU process are
obtained in [14]. For the large deviations and the moderate deviations of θ̂ T , we refer to [1],
[9] and [11]. The proofs of Theorem 1.1 and Theorem 1.2 are based on the LSI with respect to
L 2 -norm in the Wiener space and Herbst’s argument (cf. [10], [12]).
2 Deviation inequalities
In this section, we give some deviation inequalities for the estimators θ̂ T and γ̂ T by the logarithmic
Sobolev inequality and the exponential martingale method. For deviation bounds for additive
functionals of Markov processes, we refer to [3] and [18].
2.1 Moments
It is known that the solution of equation (1.1) has the following expression:
Z t

γ ‹ −θ t γ
e
+ + e−θ t
eθ s dWs .
Xt = x −
θ
θ
0
From this expression, it is easily seen that for any t ≥ 0,

γ ‹ −θ t γ
e
+ ,
µ t :=Eθ ,γ,x (X t ) = x −
θ
θ
1
−2θ t
2
(1 − e
)
σ t :=Varθ ,γ,x (X t ) =
2θ
(2.1)
(2.2)
(2.3)
Deviation inequalities and MDP for estimators in OU model
213
and for any 0 ≤ s ≤ t,
1
Covθ ,γ,x (X s , X t ) =
(1 − e−2θ s )e−θ (t−s) .
2θ
(2.4)
Therefore
Eθ ,γ,x (µ̂ T ) =
1
T
!
T
Z
Eθ ,γ,x
Varθ ,γ,x µ̂ T
1 
θT
x−
γ‹
(1 − e−θ T ) +
θ
!2 
Z t

e−θ t
eθ s dWs d t 
0

=
Xtdt
γ
θ
,
Z T

= 2 Eθ ,γ,x 
T
0
0
1
2 −θ T
1 −2θ T
= 2 2 T−
(e
− 1) + (e
− 1)
2θ
θ
θ T
1
and so for all T ≥ 1,
Varθ ,γ,x µ̂ T ≤
1
2θ 3 T
(2.5)
(2.6)
(2θ + 1)
(2.7)
and
Eθ ,γ,x (σ̂2T ) =

γ ‹2
−2θ T
(1
−
e
)
−1
+
2θ
x
−
2θ
θ
4θ 2 T
‹2

γ
1
(1 − e−θ T )
− 2 2 (1 − e−θ T )2 x −
θ
θ T 1 −2θ T
1
2 −θ T
− 2 2 T−
(e
− 1) + (e
− 1)
2θ
θ
θ T
1
1
+
which implies
Eθ ,γ,x (σ̂2T ) −

γ ‹2 2
.
+
≤ 2
θ x−
2θ
θ
θ
θ T
1
1
(2.8)
Lemma 2.1. For any 0 ≤ α ≤ θ 2 /4, for all T ≥ 1,
exp
Eθ ,γ,x
!!
T
Z
X t2 d t
α
< ∞,
0
and there exist finite positive constants L1 and L2 such that for all 0 ≤ α ≤ θ 2 /4 and T ≥ 1,
exp
Eθ ,γ,x
!!
T
Z
≤ L1 e L2 αT .
X t2 d t
α
0
Proof. For any 0 ≤ α ≤ θ 2 /4, set κ =
d Pθ ,γ,x
d Pκ,γ,x
( Z
= exp −
p
θ 2 − 2α. Then by Girsanov theorem, we have
T
0
(θ − κ)X t d X t −
)
T
Z
(αX t2
0
− γ(θ − κ)X t )d t
214
Electronic Communications in Probability
and so
X t2 d t
α
exp
Eθ ,γ,x
!!
T
Z
0
d Pθ ,γ,x
=Eκ,γ,x
d Pκ,γ,x
(
( Z
)!
T
X t2 d t
exp α
0
T
Z
0
0
(
=Eκ,γ,x
exp
≤ exp
−(θ − κ)
2
(θ − κ)T
− T) + γ
( Z
Eκ,γ,x
)!
T
Z
(X T2
2
(θ − κ)X t d t
XtdXt + γ
exp (−θ + κ)
=Eκ,γ,x
)!
T
Z
0
(θ − κ)X t d t
)!
T
exp γ
0
(θ − κ)X t d t
where the last inequality is due to θ ≥ κ. Now we have to estimate Eκ,γ,x (exp{γ
Since under Pκ,γ,x ,
µ̂ T ∼ N
1
κT
(x −
γ
κ
)(1 − e
−κT
)+
γ
1
T−
,
κ κ2 T 2
1
2κ
(e
−2κT
− 1) +
2
κ
RT
0
(θ − κ)X t d t}).
(e
−κT
− 1)
,
we have
( Z
exp γ
Eκ,γ,x
0
= exp
)!
T
(θ − κ)X t d t
γ(θ − κ) 
κ
¨
· exp
x−
γ2 (θ − κ)2
γ‹
κ
(1 − e−κT ) + γT
T−
2κ2
1
2κ
(e
−2κT
‹
− 1) +
2
κ
«
(e
−κT
− 1)
.
p
Noting θ / 2 ≤ κ ≤ θ , 0 ≤ θ − κ = 2α/(θ + κ) ≤ 2α/θ and (θ − κ)2 ≤ αθ for all 0 ≤ α ≤ θ 2 /4,
we complete the proof of the lemma.
ƒ
2.2 Logarithmic Sobolev inequality
Since the LSI with respect to the Cameron-Martin metric does not produce the concentration
inequality of correct order in large time T for the functionals
1
F (X ) := p
T
T
Z
0
g(X s )ds − E
!!
T
Z
g(X s )ds
,
0
in order to get the concentration inequality of correct order for the functionals F (X ), as pointed
out by Djellout, Guillin and Wu ([7]) we should establish the LSI with respect to the L 2 -metric.
Let us introduce the logarithmic Sobolev inequality on W with respect to the gradient in L 2 ([0, T ], R)
([10]). Let µ be the Wiener measure on W = C([0, T ], R). A function f : W → R is said to be
Deviation inequalities and MDP for estimators in OU model
215
differentiable with respect to the L 2 -norm, if it can be extend to L 2 ([0, T ], R) and for any w ∈ W ,
there exists a bounded linear operator D f (w) : g → D g f (w) on L 2 ([0, T ], R) such that
| f (w + g) − f (w) − D g f (w)|
lim
kgk L2
kgk L2 →0
= 0.
If f : W → R is differentiable with respect to the L 2 -norm, then there exists a unique element
∇ f (w) = (∇ t f (w), t ∈ [0, T ]) in L 2 ([0, T ], R) such that
D g f (w) = ⟨∇ f (w), g⟩ L 2 , f or al l g ∈ L 2 ([0, T ], R).
Denote by C b1 (W /L 2 ) the space of all bounded function f on W , differentiable with respect to
the L 2 -norm, such that ∇ f is also continuous and bounded from W equipped with L 2 -norm to
L 2 ([0, T ], R). Applying Theorem 2.3 in [10] to the Ornstein-Uhlenbeck process with linear drift,
we have
!
Z T
2
2
2
|∇ t f | d t , f ∈ C b1 (W /L 2 )
(2.9)
Ent Pθ ,γ,x ( f ) ≤ 2 Eθ ,γ,x
θ
0
where the entropy of f 2 is given by
Ent Pθ ,γ,x ( f 2 ) = Eθ ,γ,x ( f 2 log f 2 ) − Eθ ,γ,x ( f 2 ) log Eθ ,γ,x ( f 2 ).
Lemma 2.2. For any |α| ≤ θ 2 /4,
(
Z T
Eθ ,γ,x
X t2 d t
exp α
0
− Eθ ,γ,x
(
!!)!
T
Z
X t2 d t
0
≤ Eθ ,γ,x
exp
(
Z
and
€
¦
Eθ ,γ,x exp αT
€
µ̂2T
Eθ ,γ,x (µ̂2T )
−
Š©Š
≤ Eθ ,γ,x
exp
4α2
4α2
θ2
X t2 d t
0
)!
T
X t2 d t
θ2
)!
T
Z
.
0
Proof. We apply Theorem 2.7 in [12] to prove the conclusions of the lemma. Take A1 =
{α f ; |α| ≤ θ 2 /4} and A2 = {αh; |α| ≤ θ 2 /4}, where
T
Z
w 2t d t,
f (w) =
0
Define
Γ1 (g1 ) =
4 g12
θ2 f
, g 1 ∈ A1 ;
h(w) =
1
T
!2
T
Z
wt d t
.
0
Γ2 (g2 ) =
4 g22
θ2 h
, g 2 ∈ A2 .
Then for any λ ∈ [−1, 1], g1 ∈ A1 and g2 ∈ A2 , λg1 ∈ A1 , λg2 ∈ A2 , Γ1 (λg1 ) = λ2 Γ1 (g1 ),
Γ2 (λg2 ) = λ2 Γ2 (g2 ) and by Lemma 2.1
Eθ ,γ,x exp{λΓ1 (g1 )} < ∞, Eθ ,γ,x exp{λΓ2 (g2 )} < ∞.
Choose a sequence of real C ∞ -functions Φn , n ≥ 1 with compact support such that limn→∞ sup|x|≤M |Φn (x)−
e x | = 0 for all M ∈ (0, ∞). For any g1 = α f ∈ A1 and g2 = αh ∈ A2 , set
Fn (w) = Φn g1 (w)/2 , H n (w) = Φn g2 (w)/2 .
216
Electronic Communications in Probability
Then for any g ∈ L 2 ([0, T ], R),
|Fn (w + g) − Fn (w) − αΦ′n g1 (w)/2 ⟨w, g⟩ L 2 |
lim
=0
kgk L 2
kgk L2 →0
and
|H n (w + g) − H n (w) − αΦ′n g2 (w)/2
lim
1
T
RT
0
wt d t
RT
g t d t|
0
kgk L 2
kgk L2 →0
= 0.
Therefore, Fn , H n ∈ C b1 (W /L 2 ), ∇Fn = αΦ′n g1 (w)/2 w, and
∇H n =
α
T
Z
T
w t d tΦ′n g2 (w)/2
0
and so by (2.9), we have
Ent Pθ ,γ,x
€
Fn2
Š
≤
θ2
!
T
Z
2
2
|αw t | d t
Eθ ,γ,x
0
€
Φ′n
g1 (w)/2
Š2
and

€
Ent Pθ ,γ,x H n2
Š
2
1
≤ 2 Eθ ,γ,x 
T
θ
!2
T
Z
α
wt d t
€
Φ′n

Š 2 
g2 (w)/2
.
0
Letting n → ∞ and by Lemma 2.1, we get
Ent Pθ ,γ,x (e g1 ) ≤
1
2
Eθ ,γ,x Γ1 (g1 )e g1 ,
Ent Pθ ,γ,x (e g2 ) ≤
1
2
Eθ ,γ,x Γ2 (g2 )e g2 ,
and so the conclusions of the lemma hold by Theorem 2.7 in [12] and T µ̂2T ≤
RT
0
(2.10)
X t2 d t.
ƒ
2.3 Deviation inequalities
Š
€
Since X T ∼ N µ T , σ2T , and under Pθ ,γ,x
µ̂ T ∼ N
1
θT
(x −
γ
θ
)(1 − e
−θ T
)+
γ
1
,
θ θ2T2
T−
1
2θ
(e
−2θ T
− 1) +
it is easily to get from Chebyshev inequality, for any r > 0,
Š
¦
©
€
Pθ ,γ,x X T − Eθ ,γ,x (X T ) ≥ r ≤ 2 exp −θ r 2 ,
Pθ ,γ,x
where we used (2.7).
€
¨
«
Š
θ 3 Tr 2
µ̂ T − Eθ ,γ,x (µ̂ T ) ≥ r ≤ 2 exp −
2θ + 1
2
θ
(e
−θ T
− 1)
,
(2.11)
(2.12)
Deviation inequalities and MDP for estimators in OU model
217
Lemma 2.3. There exist finite positive constants C0 , C1 , C2 such that for all r > 0 and all T ≥ 1,
!
!
Z T
Z T
2
2
Xt dt
X t d t − Eθ ,γ,x
≥ r T ≤ C0 exp −C1 r T min 1, C2 r
Pθ ,γ,x
0
0
and
Pθ ,γ,x
€
Š
µ̂2T − Eθ ,γ,x (µ̂2T ) ≥ r ≤ C0 exp −C1 r T min 1, C2 r
.
In particular, there exist finite positive constants C0 , C1 , C2 such that for all r > 0 and all T ≥ 1,
€
Š
Pθ ,γ,x |σ̂2T − Eθ ,γ,x (σ̂2T )| ≥ r ≤ C0 exp −C1 r T min 1, C2 r .
Proof. We only prove the first inequality. By Lemma 2.2 and Lemma 2.1, there exist finite positive
constants L1 and L2 such that for all T ≥ 1, for any |α| ≤ θ 2 /4,
!!)!
(
Z T
Z T
Eθ ,γ,x
exp α
0
2
≤L1 e L2 α T .
X t2 d t
X t2 d t − Eθ ,γ,x
0
Therefore, by Chebyshev inequality, for any r > 0, T ≥ 1 and |α| ≤ θ 2 /4,
!
Z T
Z T
Pθ ,γ,x
0
and
X t2 d t − Eθ ,γ,x (
T
Z
Pθ ,γ,x
0
0
0
≤ L1 e−(αr−L2 α
2
)T
!
T
Z
X t2 d t − Eθ ,γ,x (
X t2 d t) ≥ r T
X t2 d t) ≤ −r T
≤ L1 e−(αr−L2 α
2
)T
.
Now, by
sup {αr − L2 α2 } ≥
|α|≤θ 2 /4
θ2r
2r
,
min 1,
8
L2 θ 2
we obtain the first inequality of the lemma from the above estimates.
ƒ
Lemma 2.4. There exist finite positive constants C0 , C1 and C2 such that for all r > 0 and all T ≥ 1,
‚
Œ

γ‹
Pθ ,γ,x WT µ̂ T −
≥ r T ≤ C0 exp −C1 r T min 1, C2 r .
θ
Proof. Since for any r > 0 and T ≥ 1,
¨
«

γ‹
WT µ̂ T −
≥ rT
θ
¨
⊂
¦
©
WT (µ̂ T − Eθ ,γ,x (µ̂ T )) ≥ r T /2 ∪
¦
⊂ |WT | ≥
p
©
r T /2 ∪
¦

WT
Eθ ,γ,x (µ̂ T ) −
p ©
(µ̂ T − Eθ ,γ,x (µ̂ T )) ≥ r ∪
γ‹
θ
«
≥ r T /2
(
WT ≥
θ rT
Š
€
γ
2| x − θ |
)
,
218
Electronic Communications in Probability
by (2.12) and WT ∼ N (0, T ), we get
‹

γ
Pθ ,γ,x WT (µ̂ T − ) ≥ r T
θ

≤2 exp −
Tr
¨
+ 2 exp −
8
«
θ 3 Tr
+ 2 exp
2θ + 1

θ 2 r2 T 
− €
.
Š
γ 2

8 x−θ

ƒ
Lemma 2.5. For each β ∈ R fixed, there exist finite positive constants C0 , C1 , C2 such that for all
r > 0 and all T ≥ 1,
!
Z T
Pθ ,γ,x
X t − β dWt ≥ r T ≤ C0 exp −C1 r T min 1, C2 r .
0
Proof. It is known that for α ∈ R,
( Z T
(α)
MT
= exp α
0
X t − β dWt −
α2
2
)
T
Z
0
Xt − β
2
dt
,
T ≥0
is F T -martingale, where F T := σ(Wt , t ≤ T ). Therefore, by Hölder inequality, we can get that for
any ε ∈ (0, 1],
)!
( Z T
X t − β dWt
Eθ ,γ,x exp α
0
(
≤
Eθ ,γ,x
exp
Eθ ,γ,x
exp
0
(1 + ε)2 α2
2ε
)!!
T
Z
2ε
(
=
(1 + ε)2 α2
Xt − β
2
Xt − β
2
)!!
0
dt
T
Z
ε
1+ε
1
1+ε
((1+ε)α)
Eθ ,γ,x M T
ε
1+ε
.
dt
In particular, take ε = 1, then by Lemma 2.1, there exists finite positive constants L1 = L1 (θ , β, γ, x)
and L2 = L2 (θ , β, γ, x) such that for all T ≥ 1, for any α2 ≤ θ 2 /16, by Cauchy-Schwartz inequality,
)!
( Z T
X t − β dWt
Eθ ,γ,x exp α
0
(
≤
exp 2α2
Eθ ,γ,x
≤
≤L1 e
(
4
X t2 d t
0
L 2 α2 T
)!! 1
T
Z
exp 4α2
Eθ ,γ,x
2
(X t − β)2 d t
0
(
)!! 1
T
Z
Eθ ,γ,x
)!! 1
T
Z
exp 4α2
4
(−2β X t + β 2 )d t
0
.
Therefore, by Chebyshev inequality, the conclusion of the lemma holds.
ƒ
Deviation inequalities and MDP for estimators in OU model
219
Proof of Theorem 1.1
We only show the first inequality. The second one is similar. By
θ̂ T − θ =
Š RT €
Š
€
γ
γ
WT µ̂ T − θ − 0 X t − θ dWt
T σ̂2T
for any r > 0 and T ≥ 1,
€
Š
Pθ ,γ,x |θ̂ T − θ | ≥ r
€
Š
≤Pθ ,γ,x σ̂2T − Eθ ,γ,x (σ̂2T ) ≥ Eθ ,γ,x (σ̂2T )/2
!
Z T

γ‹
γ‹
Xt −
−
dWt ≥ Eθ ,γ,x (σ̂2T )r T /2
+ Pθ ,γ,x WT µ̂ T −
θ
θ
0
Therefore, by Lemmas 2.3, 2.4 and 2.5, we obtain the first inequality of the theorem.
ƒ
3 Moderate deviations
In this section, we show Theorem 1.2. By (1.2) and (1.3), we have the following estimates
(θ̂ T − θ ) +
≤
T
Z
2θ
T

Xt −
0
€
Š
γ
|WT µ̂ T − θ |
θ
(3.1)
dWt
RT €
|2θ σ̂2T − 1|
+
T σ̂2T
γ‹
0
Xt −
γ
θ
Š
dWt
T σ̂2T
and for
(γ̂ T − γ) −
≤
WT
T
2γ
+
T
€
Š
γ
|µ̂ T ||WT µ̂ T − θ |
T σ̂2T
T
Z

Xt −
0
+
γ‹
θ
|2γσ̂2T − µ̂ T |
(3.2)
dWt
RT €
0
γ
θ
Xt −
Š
dWt
.
T σ̂2T
Lemma 3.1. (1). For any r > 0,
lim sup
T →∞
lim sup
T →∞
1
λT
1
λT

log Pθ ,γ,x
log Pθ ,γ,x
µ̂ T −
γ
θ
θ
WT ≥
T

γ‹
Z
lim sup
T →∞
λT
log Pθ ,γ,x
σ̂2T
−
1
2θ
Xt −
0
and
1
γ
µ̂ T −
T
Z
0

Xt −
θ
p
= −∞,
!
dWt ≥
γ‹
θ
‹
T λT r
p
T λT r
= −∞
!
dWt ≥
p
T λT r
= −∞.
220
Electronic Communications in Probability
(2). For any δ > 0,
lim sup
T →∞
1
λT
(θ̂ T − θ ) −
log Pθ ,γ,x
T
Z
2θ
T
γ‹

Xt −
0
θ
r
dWt ≥ δ
λT
!
= −∞
T
and
lim sup
T →∞
1
λT
(γ̂ T − γ) −
log Pθ ,γ,x
WT
−
T
2γ
T
Z
T

Xt −
0
γ‹
θ
r
dWt ≥ δ
λT
T
!
= −∞.
Proof. (1). We only give the proof of the third assertion in (1). The rest is similar. For any L > 0,
(
)
Z T
‹
p
1
γ
σ̂2T −
dWt ≥ T λ T r
Xt −
2θ
θ
0


¨
«
Z T
‹

 1
γ
1
r
Xt −
∪ p
dWt ≥ L .
⊂ σ̂2T −
≥

 Tλ
2θ
L
θ
0
T
By Lemma 2.3, and Lemma 2.5, we have
lim sup
T →∞
and
lim sup
T →∞
1
λT
‚
1
λT
σ̂2T
log Pθ ,γ,x

log Pθ ,γ,x  p
T
Z
1
T λT
−
1
2θ

0
Xt −
≥
γ‹
θ
r
Œ
= −∞
L

dWt ≥ L  ≤ −L 2 C1 C2 .
Hence,
lim sup
T →∞
1
λT
log Pθ ,γ,x
σ̂2T
−
1
2θ
T
Z

Xt −
0
γ‹
θ
!
dWt ≥
Letting L → ∞, we obtain the third conclusion.
(2). It follows from (3.1) and (3.2) that
r !
Z T
2θ
γ‹
λT
(θ̂ T − θ ) −
dWt ≥ δ
Xt −
T 0
θ
T

 
p
Z


T λT  
γ‹
2
2
| ≥ δσ̂ T
∪ |2θ σ̂ T − 1|
⊂ |WT µ̂ T −

θ
2  


≤ −L 2 C1 C2 .

T
0

|WT
µ̂ T −
γ‹
θ
p
| ≥ δEθ ,γ,x (σ̂2T )

Xt −
γ‹
θ
p
dWt ≥ δσ̂2T
©
T λT  ¦ 2
∪ |σ̂ T − Eθ ,γ,x (σ̂2T )| ≥ Eθ ,γ,x (σ̂2T )/2
4 


T
Z
∪
T λT r


⊂
p


|2θ σ̂2T − 1|
0

Xt −
γ‹
θ
p
dWt ≥ δEθ ,γ,x (σ̂2T )
T λT 
4 
T λT 
2 
Deviation inequalities and MDP for estimators in OU model
221
and
(γ̂ T − γ) −
2γ
T
Z
T
Xt −
0
r
γ‹

dWt ≥ δ
θ

⊂



|µ̂ T ||WT
µ̂ T −
γ‹
θ
| ≥ δσ̂2T
!
T


p
λT

T
Z
T λT  
∪ |2γσ̂2T − µ̂ T |
2  

Xt −
0
γ‹
θ
p
dWt ≥ δσ̂2T
T λT 
2 
©
γ ª ¦ 2
γ
∪ |σ̂ T − Eθ ,γ,x (σ̂2T )| ≥ Eθ ,γ,x (σ̂2T )/2
≥
µ̂ T −
θ
2θ


p

 3γ
T λT 
γ‹
∪
| ≥ δEθ ,γ,x (σ̂2T )
|WT µ̂ T −
 2θ
θ
4 
§
⊂


∪


2γσ̂2T −
γ
θ
+ µ̂ T −
T
Z
γ ‹
θ
γ‹

Xt −
0
θ
p
dWt ≥ δEθ ,γ,x (σ̂2T )
T λT 
.
4 
Therefore, by Lemmas 2.3 and (1), we get the conclusions.
ƒ

§
Lemma 3.2. For each β, κ ∈ R fixed,
Pθ ,γ,x
RT
pκ
T λT 0
θ u
.
κ2 (θ +2(γ−θ β)2 )
2 2
LDP with speed λ T and rate function J(u) =
‹
ª
X t − β dWt ∈ · , T ≥ 1
satisfies the
Proof. By (2.12) and Lemma 2.3, we can get for any δ > 0,
!
Z T
1
1
1
1
2
2
lim
≥ δ < 0.
log Pθ ,γ,x
+ 2 (γ − θ β)
Xt − β dt −
T →∞ T
T 0
2θ
θ
(3.3)
Therefore, Proposition 1 in [4] yields the conclusion of the lemma.
ƒ
Proof of Theorem 1.2
q
By Lemma 3.1, {Pθ ,γ,x (
T
λT
q
(θ̂ T − θ ) ∈ ·), T ≥ 1} and {Pθ ,γ,x (
T
λT
T
!
(γ̂ T − γ) ∈ ·), T ≥ 1} are expo-
nential equivalent to
(
r
T 2θ
Pθ ,γ,x
and
(
r
Pθ ,γ,x
λT T
T
WT
λT
T
respectively. Noting for γ 6= 0,
1.2 follows from Lemma 3.2.
WT
T
Z

Xt −
0
2γ
Z
T

γ‹
θ
dWt ∈ ·
γ‹
!
)
,T ≥1
!
)
dWt ∈ · , T ≥ 1 ,
Xt −
T 0
θ
R €
Š
R 2γ T
γ
2γ T
γ
1
dWt , Theorem
+ T 0 X t − θ dWt = T 0 X t − θ + 2γ
+
222
Electronic Communications in Probability
ƒ
Acknowledgments The authors are grateful to referees for their comments and suggestions.
References
[1] B. Bercu, A. Rouault. Sharp large deviations for the Ornstein-Uhlenbeck process. Theory of
Prob. and its Appl., 46(2002), 1-19. MR1968706
[2] S. G. Bobkov, F. Götze. Exponential Integrability and Transportation Cost Related to Logarithmic Sobolev Inequalities. Journal of Functional Analysis, 163(1999), 1-28. MR1682772
[3] P. Cattiaux, A. Guillin. Deviation bounds for additive functionals of Markov process. ESAIM:
Probability and Statistics. 12(2008), 12-29. MR2367991
[4] A. Dembo. Moderate Deviations for Martingales with Bounded Jumps. Electronic Communications in Probability, 1 (1996),11-17. MR1386290
[5] A. Dembo, D. Zeitouni. Large Deviations Techniques and Applications, Springer-Verlag, 1998.
MR1619036
[6] J. D. Deuschel, D. W. Stroock. Large Deviations, New York, 1989. MR0997938
[7] H. Djellout, A. Guillin, L. M. Wu. Transportation cost-information inequalities and applications to random dynamical system and diffusions. Ann. Probab., 32(2004), 2702-2732.
MR2078555
[8] H. Djellout, A. Guillin, L. M. Wu. Moderate deviations for non-linear functionals and empirical spectral density of moving average processes. Ann. Inst. H. Poincaré Probab. Statist.,
42(2006), 393-416. MR2242954
[9] D. Florens-Landais, H. Pham, Large deviations in estimate of an Ornstein-Uhlenbeck model.
Journal of Applied Probability, 36(1999), 60-77. MR1699608
[10] M. Gourcy, L. M. Wu, Logarithmic Sobolev inequalities of diffusions for the L 2 metric.
Potential Analysis, 25(2006), 77-102. MR2238937
[11] A. Guillin, R. Liptser, Examples of moderate deviation principles for diffusion processes.
Discrete and continuous dynamical systems-series B, 6(2006), 803-828. MR2223909
[12] M. Ledoux, Concentration of Measure and Logarithmic Sobolev Inequalities. Séminaire de
probabilités XXXIII, Lecture Notes in Mathematics, 1709(1999), 120-216. MR1767995
[13] M. Ledoux, The Concentration of Measure Phenomenon. Mathematical Surveys and Monographs 89, American Mathematical Society, 2001. MR1849347
[14] P. Lezaud, Chernoff and Berry-Essen’s inequalities for Markov processes. ESAIMS: Probability and Statistics, 5(2001), 183-201. MR1875670
[15] Yury A. Kutoyants, Statistical Inference for Ergodic Diffusion Processes. Springer Series in
Statistics, London, 2004, MR2144185
Deviation inequalities and MDP for estimators in OU model
[16] B. L. S. Prakasa Rao, Statistical Inference for Diffusion Type Processes. Oxford University
Presss, New York,1999.
[17] D. Revuz, M. Yor, Continuous Martingales and Brownian Motion. Springer-Verlag, 1991.
MR1083357
[18] L. M. Wu, A deviation inequality for non-reversible Markov processes. Ann. Inst. Henri.
Poincaré, Probabilités et Statistiques. 36(2000), 435-445. MR1785390
223
Download