Journal ArticleDOI

# Addendum to `Asymptotic inference for a linear stochastic differential equation with time delay'

01 Aug 2001-Bernoulli (Bernoulli Society for Mathematical Statistics and Probability)-Vol. 7, Iss: 4, pp 629-632

AbstractIn Gushchin and Kiichler (1999) we considered the stochastic differential equation dX(t) = aX(t)dt + bX(t - 1)dt + dW(t), t > 0, (1) with initial condition X(t) = Xo(t), t E [-1, 0], where (Xo(t), t E [-1, 0]) is a given continuous stochastic process independent of a standard Wiener process (W(t), t > 0). The parameter 9 = (a, b)* is unknown and supposed to be estimated based on (X(t), t < T). (The sign * denotes matrix or vector transposition.) The solutions (X(t), t E [-1, T]) of the differential equation (1) generate a family (Pt, 9 E Rl2) of distributions on C([-1, T]). We studied local asymptotic properties of this family and corresponding properties of the maximum likelihood estimators of 9 for T -- oc.

### Summary

• (The sign denotes matrix or vector transposition.).
• The authors studied local asymptotic properties of this family and corresponding properties of the maximum likelihood estimators of W for T !1.

Did you find this useful? Give us your feedback

Content maybe subject to copyright    Report

Addendum to `Asymptotic inference for a
linear stochastic differential equation with
time delay'
ALEXANDER A. GUSHCHIN
1
and UWE KU
È
CHLER
2
1
Steklov Mathematical Institute, Gubkina 8, 117966 Moscow GSP-1, Russia.
E-mail: gushchin@mi.ras.ru
2
Institut fu
È
r Mathematik, Humboldt-Universita
È
t zu Berlin, Unter den Linden 6, D-10099 Berlin,
Germany. E-mail: kuechler@mathematik.hu-berlin.de
We strengthen the convergence result proving the local asymptotic mixed normality property in one of
the 11 cases considered in our previous paper.
Keywords: local asymptotic mixed normality; maximum likelihood estimator; stochastic differential
equations; time delay
In Gushchin and Ku
È
chler (1999) we considered the stochastic differential equation
dX (t) aX(t)dt bX (t ÿ 1) dt dW (t), t > 0, (1)
with initial condition X (t) X
0
(t), t 2 [ÿ1, 0], where (X
0
(t), t 2 [ÿ1, 0]) is a given
continuous stochastic process independent of a standard Wiener process (W (t), t > 0). The
parameter W (a, b)
is unknown and supposed to be estimated based on (X (t), t < T).
(The sign
denotes matrix or vector transposition.) The solutions (X (t), t 2 [ÿ1, T]) of the
differential equation (1) generate a family (P
W
T
, W 2 R
2
) of distributions on C([ÿ1, T ]). We
studied local asymptotic properties of this family and corresponding properties of the
maximum likelihood estimators of W for T !1.
In particular, for a ®xed W
0
(a, b)
2 R
2
the limit behaviour of
dP
W
T
dP
W
0
T
(X ), W 2 R
2
,
was studied. It turned out that this behaviour depends strongly on the position of W
0
, and one
has to consider 11 different cases (see Gushchin and Ku
È
chler 1999, Table 1, p. 1064). The
family (P
W
T
) is locally asymptotically normal in one of these cases, it is locally asymptotically
mixed normal (LAMN) in another three cases, and it is only locally asymptotically quadratic
in ®ve of the remaining cases. As for the last cases, called P1 and P2, we showed that the
family (P
W
T
) is still locally asymptotically quadratic but has an additional property that we
called periodic local asymptotic mixed normality (PLAMN).
Assume that we are in case P2, i.e. a , 1 and b , î=sin î, where î is the unique number
in (0, ð) with î a tan î (î ð=2ifa 0), or a > 1 and b , ÿe
aÿ1
. We refer to
Bernoulli 7(4), 2001, 629±632
1350±7265 # 2001 ISI/BS

Gushchin and Ku
È
chler (1999) for the de®nition of the random vector V
T
, the random matrix
I
T
, and the numbers v
0
. 0, î
0
. 0, A
0
and B
0
. It was shown that the following properties
hold (Gushchin and Ku
È
chler 1999, Proposition 2.6):
(i)
log
dP
W
0
e
ÿv
0
T
ì
T
dP
W
0
T
(X ) ì
V
T
ÿ
1
2
ì
I
T
ì, ì 2 R
2
; (2)
(ii) for T
n
u nÄ, where u 2 [0, Ä)ixed,Ä ð=î
0
, n > 0,
(V
T
n
, I
T
n
) !
d
(V
1
(u), I
1
(u)), n !1,
where (V
1
(u), I
1
(u))
d
(I
1=2
1
(u) Z, I
1
(u)) and Z is a random vector independent of
I
1
(u) and distributed as N (0, I
2
). The matrix I
1
(u) is given by
I
1
(u)
1
0
e
ÿ2v
0
t
U
2
0
(u ÿ t)dt
1
0
e
ÿ2v
0
t
U
0
(u ÿ t)U
2
(u ÿ t)dt
1
0
e
ÿ2v
0
t
U
0
(u ÿ t)U
2
(u ÿ t)dt
1
0
e
ÿ2v
0
t
U
2
2
(u ÿ t)dt
0
B
B
B
@
1
C
C
C
A
, (3)
where
U
i
(t) X
0
(0)ö
i
(t) b
0
ÿ1
ö
i
(t ÿ s ÿ 1)e
ÿv
0
(s1)
X
0
(s)ds
1
0
ö
i
(t ÿ s)e
ÿv
0
s
dW (s), (4)
in which
ö
i
(t) A
i
cos(î
0
t) B
i
sin(î
0
t), i 0, 2 (5)
and
A
2
B
2

e
ÿv
0
cos î
0
ÿsin î
0
sin î
0
cos î
0

A
0
B
0

: (6)
We have referred to this asymptotic behaviour as PLAMN to indicate that the cluster points
of (V
T
, I
T
) have the same structure as in the LAMN case but that (V
T
, I
T
) converges in
distribution only if T runs to in®nity through a grid in such a way that the fractional part of
T=Ä tends to a limit.
The purpose of this note is to show that in case P2 one can prove the usual LAMN
property. To this end, we look for matrices Ö
T
, T > 0, such that
Ö
T
U
0
(T ÿ t)
U
2
(T ÿ t)

does not depend on T . In view of (4) and (5), this will be the case if
Ö
T
A
0
A
2
B
0
B
2

cos(T ÿ t)
sin(T ÿ t)

630 A.A. Gushchin and U. Ku
È
chler

is independent of T . Using (6), after elementary calculations one obtains that the latter
property holds if and only if
Ö
T
Ö
0
cos(î
0
T) ÿ sin(î
0
T) cot î
0
ÿe
ÿv
0
sin(î
0
T) csc î
0
e
v
0
sin(î
0
T) csc î
0
cos(î
0
T) sin(î
0
T) cot î
0

: (7)
De®ne Ö
T
according to (7) with Ö
0
: I
2
, where I
2
is the identity matrix. Then
~
U
0
(t)
~
U
2
(t)
!
: Ö
T
U
0
(T ÿ t)
U
2
(T ÿ t)
!
X
0
(0)
A
0
cos(î
0
t) ÿ B
0
sin(î
0
t)
e
ÿv
0
(A
0
cos(î
0
(t 1)) ÿ B
0
sin(î
0
(t 1)))
!
b
0
ÿ1
A
0
cos(î
0
(t s 1)) ÿ B
0
sin(î
0
(t s 1))
e
ÿv
0
(A
0
cos(î
0
(t s 2)) ÿ B
0
sin(î
0
(t s 2)))
!
e
ÿv
0
(s1)
X
0
(s)ds
1
0
A
0
cos(î
0
(t s)) ÿ B
0
sin(î
0
(t s))
e
ÿv
0
(A
0
cos(î
0
(t s 1)) ÿ B
0
sin(î
0
(t s 1)))
!
e
ÿv
0
s
dW (s):
Extend the de®nition of I
1
(u)tou 2 R
by the same formula (3). Then
Ö
T
I
1
(T)Ö
T
1
0
e
ÿ2v
0
t
Ö
T
U
0
(T ÿ t)
U
2
(T ÿ t)

U
0
(T ÿ t)
U
2
(T ÿ t)

Ö
T
dt,
and, therefore, the matrix
~
I
1
: Ö
T
I
1
(T)Ö
T
does not depend on T and is given by
~
I
1
1
0
e
ÿ2v
0
t
~
U
2
0
(t)dt
1
0
e
ÿ2v
0
t
~
U
0
(t)
~
U
2
(t)dt
1
0
e
ÿ2v
0
t
~
U
0
(t)
~
U
2
(t)dt
1
0
e
ÿ2v
0
t
~
U
2
2
(t)dt
0
B
B
B
@
1
C
C
C
A
:
Put
~
V
T
: Ö
T
V
T
,
~
I
T
: Ö
T
I
T
Ö
T
:
Proposition 1. In case P2 the family (P
W
, W 2 R
2
) is LAMN at every W
0
:
log
dP
W
0
e
ÿv
0
T
Ö
T
ì
T
dP
W
0
T
(X ) ì
~
V
T
ÿ
1
2
ì
~
I
T
ì (8)
and
(
~
V
T
,
~
I
T
) !
d
(
~
V
1
,
~
I
1
), (9)
where (
~
V
1
,
~
I
1
)
d
(
~
I
1=2
1
Z,
~
I
1
) and Z is a vector independent of
~
I
1
and distributed as
N (0, I
2
).
Proof. Equation (8) is obvious from (2).
Addendum to `Asymptotic inference for a linear SDE with time delay' 631

It was shown in the proof of Proposition 2.6 that
I
T
ÿ I
1
(T ) !
P
0:
Thus
~
I
T
!
P
~
I
1
because the matrices Ö
T
are uniformly bounded. Now the joint convergence (9) follows from
the stable limit theorem for martingales as in the proofs of Propositions 2.2±2.4 in Gushchin
and Ku
È
chler (1999). h
Remark. Contrary to the case just considered, in case P1 one cannot avoid periodic behaviour
of the likelihood and prove the LAMN property by choosing an appropriate matrix norming.
For a general statement of this type for delay processes, see Putschke (2000).
Acknowledgements
The authors express their thanks to anonymous referees, whose comments led to a more
detailed explanation of how to construct the matrices Ö
T
.
References
Gushchin, A.A. and Ku
È
chler, U. (1999) Asymptotic inference for a linear stochastic differential
equation with time delay. Bernoulli, 5, 1059±1098.
Putschke, U. (2000) Af®ne stochastische Funktionaldifferentialgleichungen und lokal asymptotische
Eigenschaften ihrer ParameterschaÈtzungen. Doctoral dissertation, Humboldt-UniversitaÈt zu Berlin.
Received June 2000 and revised March 2001
632 A.A. Gushchin and U. Ku
È
chler
##### Citations
More filters

Journal ArticleDOI
Abstract: We obtain exact large deviation rates for the log-likelihood ratio in testing models with observed Ornstein–Uhlenbeck processes and get explicit rates of decrease for the error probabilities of Neyman–Pearson, Bayes, and minimax tests. Moreover, we give expressions for the rates of decrease for the error probabilities of Neyman–Pearson tests in models with observed processes solving affine stochastic delay differential equations.

12 citations

### Cites background from "Addendum to `Asymptotic inference f..."

• ...Gushchin and Küchler [12] - [14] derived local asymptotic properties of the likelihood process in (two-parameter) models connected with a special case of affine stochastic delay differential equations....

[...]

04 Jun 2018
Abstract: In the thesis a statistical model of linear stochastic differential equation with time delay is considered. The aim of the investigation is to prove local asymptotic properties of the likelihood function.

6 citations

### Cites methods from "Addendum to `Asymptotic inference f..."

• ...Küchler and Kutoyants [27] investigated in 2000 the equation dX(t) = bX(t− θ) dt+ dW (t), where b is a known negative parameter and the delay parameter θ ∈ R is unknown....

[...]

• ...Further, by Corollary 4.12 in Gushchin and Küchler [16], (Y T (1), [Y T , Y T ](1)) D−→ (Y(1), [Y ,Y ](1)) as T →∞....

[...]

• ...[28] Küchler, U. and Sørensen, M. (2010)....

[...]

• ...By the functional central limit theorem, W T D−→W as T →∞, hence∣∣∣∣ 1T 2 ∫ T 0 W (t)X(t) dt ∣∣∣∣ 6 √( 1 T 2 ∫ T 0 W (t)2 dt )( 1 T 2 ∫ T 0 X(t)2 dt ) = √(∫ 1 0 W T (t)2 dt )( 1 T 2 ∫ T 0 X(t)2 dt ) P−→ 0 as T →∞, and the claim follows from Corollary 4.12 in Gushchin and Küchler [16]....

[...]

• ...[16] Gushchin, A. A. and Küchler, U. (1999)....

[...]

Journal ArticleDOI
Abstract: We obtain an explicit form of fine large deviation theorems for the log-likelihood ratio in testing models with observed Ornstein-Uhlenbeck processes and get explicit rates of decrease for error probabilities of Neyman-Pearson, Bayes, and minimax tests. We also give expressions for the rates of decrease of error probabilities of Neyman-Pearson tests in models with observed processes solving affine stochastic delay differential equations.

1 citations

### Cites background from "Addendum to `Asymptotic inference f..."

• ...Gushchin and Küchler [12] - [14] derived local asymptotic properties of the likelihood process in (two-parameter) models connected with a special case of affine stochastic delay differential equations....

[...]

##### References
More filters

Journal ArticleDOI
Abstract: For the stochastic differential equation d X(t)=a X(t)+bX(t-1),dt+dW(t),t≥0, the local asymptotic properties of the likelihood function are studied. They depend strongly on the true value of the parameter θ =(a,b) * . Eleven different cases are possible if θ runs through ℝ 2 . Let θ T be the maximum likelihood estimator of θ based on ( X(t),t≤T) . Applications to the asymptotic behaviour of θ T as T →∞ are given.

38 citations