Anda di halaman 1dari 18

Asymptotic Poisson Character of Extremes in

Non-Stationary Gaussian Models


JEAN-MARC AZAI

S
Laboratoire de Statistique et Probabilites, Universite Paul Sabatier, 118 route de Narbonne, 31062,
Toulouse Cedex 4, France
E-mail: azais@cict.fr
CE

CILE MERCADIER
Laboratoire de Statistique et Probabilites, UMR C5583, Universite Paul Sabatier, 118 route de Narbonne,
31062, Toulouse Cedex 4, France
E-mail: mercadie@cict.fr
[Received October 14, 2003; Revised May 5, 2004; Accepted July 2, 2004]
Abstract. Let X be a non-stationary Gaussian process, asymptotically centered with constant variance. Let u
be a positive real. Define R
u
(t) as the number of upcrossings of level u by the process X on the interval (0, t].
Under some conditions we prove that the sequence of point processes (R
u
)
u >0
converges weakly, after
normalization, to a standard Poisson process as u tends to infinity. In consequence of this study we obtain the
weak convergence of the normalized supremum to a Gumbel distribution.
Key words. Gumbel distribution, non-stationary, Poisson process, supremum of Gaussian process, upcrossings
AMS 2000 Subject Classifications. PrimaryV60G70, 60G15
1. Introduction
Since the heuristic consideration of S. Rice and the development of V. A. VolkonskiN
and J. A. Rozanov, one knows that high excursions of a regular stationary Gaussian
process have a limit distribution (after renormalization of time). More precisely if the
excursions beyond the level u are identified by their starting point (called upcrossings of
level u) this point process converges weakly, as u tends to infinity, to a Poisson process.
Several proofs of such kind of results have been given with increasingly more general
hypotheses or original arguments by Volkonski and Rozanov (1961), Cramer and
Leadbetter (1967), Berman (1971), Leadbetter et al. (1983) and Wschebor (1985).
However all these works rely on the stationary assumption which sometimes is not very
well adapted to some modelling situations.
The aim of the present paper is to prove that in fact the assumption of constant
variance, instead of stationarity, is sufficient to obtain the same type of convergence
(Theorem 2.3). In addition we show that the process does not need to be exactly centered:
this first result remains true for a process with an expectation that tends to zero at a
certain speed. Classically we obtain, as a corollary of our first result, that a normalization
Extremes 6, 301318, 2003
#
2005 Kluwer Academic Publishers. Manufactured in The Netherlands.
of the supremum converges weakly to a Gumbel distribution (Corollary 2.1). Related
studies were conducted by Piterbarg and Konakov (1995) for homogeneous and isotropic
Gaussian fields under the assumption of compact supported covariance.
The key of our development is to begin the study in a simple case when the variance of
the derived process is also supposed to be constant (Theorem 2.2). Next we define the
Bunit-speed^ transformation of time to come down to such context. Furthermore, we use
some improvements with respect to known tools:
a sharp evaluation of the distribution of the supremum on intervals which sizes tend to
infinity at a certain speed (Lemma 4.l),
the normal comparison Lemma in a version due to Li and Shao (2002) (Proposition 3.2).
The organization of the paper is as following. In Section 2 we state our main results.
Principal tools for their demonstrations are reminded in Section 3. Section 4 is devoted to
the proofs.
1.1. Notations and hypotheses
We consider X X
t
. t 2 R f g a Gaussian process. Denote by m its expectation function
and r its covariance function. Assume that the variance of X is constant, without loss of
generality we can take it equal to one: for all t in R. r t. t Var X
t
1. The partial
derivatives of r are given for all (i, j) in N
2
. r
ij
s. t
0
i j
0s
i
0t
j
r s. t . Throughout the paper
X will satisfy one of the two sets of hypotheses expressed below.
& The Bunit-speed^ set:
U1 8t 2 R. r
11
t. t 1.
U2 r s. t log s t j j
s t j j !1
!0.
U3 8 0. sup
s t j j
r s. t j j < 1.
U4 r is of class C
4
.
8 0. r
01
and r
04
are bounded on fs. t 2 R
2
. s and t g.
U5

log t
_
m t
t !1
!0.
U6 m
0
t
t !1
!0.
& In general case we define:
8t 2 R. St
_
t
0

r
11
s. s
_
ds.
Y Y
t
X
S
1
t
. t 2 R
_ _
the Bunit-speed^ transformation of X and r
Y
its
covariance function.
302 AZAI

S AND MERCADIER
Then the general group of hypotheses on X consists in assuming that the transformed
process Y satisfies Conditions (U2, . . . , U6) since by construction Y verifies (Ul). It
consists of:
G1 8t 2 R. r
11
t. t 0 and lim
t !1
S t 1.
G2 r s. t log S s S t j j
S s S t j j ! 1
!0.
G3 8 0. sup
S s S t j j
r s. t j j < 1.
G4 r is of class C
4
with s ! r
11
s. s of class C
3
.
8 0. r
Y
01
and r
Y
04
are bounded on fs. t 2 R
2
. s and t g.
G5

log St
_
m t
t !1
!0.
G6
m
0
t

r
11
t. t
_
t !1
!0.
is the standard Gaussian density and F its distribution function. Set
"
1 .
For a positive real number u and a process X with C
1
Ysample paths, we define the
point process of upcrossings by X of level u:
8B & R. U
u
B # t 2 B. X
t
u. X
0
t
0
_ _
.
For any value of u, we define C
u
as the asymptotic mean number of upcrossings on
any interval of size one under Bunit-speed^ parameterization:
C
u

1
2
exp
u
2
2
_ _

2
p u .
For t > 0 and B & R. we define M
t
= sup
s 2(0, t)
X
s
and M
B
= sup
s 2B
X
s
.
We write jBj for the length of B & R.
We define on (0, +V) the two normalizing functions
a
t

2 log S t
_
and b
t

2 log S t
_

log 2

2 log S t
_ .
The standard Poisson process is, by definition, the Poisson process associated to the
Lebesgue measure.
Throughout this paper (Const) symbolizes a generic positive constant. Its value may
change from one occurrence to another.
2. Main results
We start by reminding the result within the stationary framework, as it is stated in
Leadbetter et al. (1983).
ASYMPTOTIC POISSON CHARACTER OF EXTREMES IN NON-STATIONARY GAUSSIAN MODELS 303
Theorem 2.1: Let X X
t
. t 2 R f g be a zero-mean stationary Gaussian process with
unit variance. Assume that the covariance function P satisfies P(t) log(t) Y 0 as t tends
to infinity and that the second spectral moment `
2
= P
00
(0) is finite.
Define the family of point processes R
u
by
8B & R. R
u
B U
u
C
1
u
`

1
2
2
B
_ _
.
Then as u Y +V, R
u
converges weakly to a standard Poisson process.
By comparison our results are:
Theorem 2.2: Let X X
t
. t 2 R f g be a Gaussian process with unit variance and
suppose that Assumptions (UlYU6 ) are satisfied.
Define the family of point processes R
u
by
8B & R. R
u
B U
u
C
1
u
B
_ _
.
Then as u Y +V, R
u
converges weakly on (0, +V) to a standard Poisson process.
Theorem 2.3: Let X X
t
. t 2 R f g be a Gaussian process with unit variance and
suppose that Assumptions (GlYG6) are satisfied.
Define the family of point processes R
u
by
8B & R. R
u
B U
u
S
1
C
1
u
B
_ _ _ _
.
Then as u Y +V, R
u
converges weakly on (0, +V) to a standard Poisson process.
Comments:
If when t tends to infinity we have

log t
_
m t ! t instead of (U5) and

log S t
_
m t ! t instead of (G5), R
u
converges weakly on (0, +V) as u Y +V
to an homogeneous Poisson process with intensity e
t
dx . Arguments of the
demonstration are the same as those used for the case t = 0.
For a stationary process, Conditions (U1), the second statement of (U4), (U5) and (U6)
are obviously satisfied. Moreover in this case (U3) is a direct consequence of (U2).
Thus our hypotheses are comparable to those of Leadbetter, Lindgren and Rootzen.
Remark that under (U1) and stationary hypotheses we get `
2
= r
11
(t, t) = 1. Hence we
still look at the same family (R
u
)
u > 0
.
The restriction to the positive half line is due to the bound (U5) or (G5) on m that
makes sense only for t positive. For zero-mean processes the convergence of the
family of point processes R
u
is valid on R.
Of course by space-reversibility the same kind of theorem holds true for u that
tends to jV.
Note that hypotheses (U4) and (G4) imply that the process X admits a modification
with C
1
Ysample paths, so the number of upcrossings is well defined.
304 AZAI

S AND MERCADIER
As it is well known, the asymptotic Poisson character implies that the supremum tends
weakly to a Gumbel distribution.
Corollary 2.1: Let X X
t
. t 2 R

f g be a Gaussian process with unit variance


satisfying (G1YG5). Then
Pa
t
M
t
b
t
x
t !1
!exp e
x
.
Comments:
If (G5) is replaced by

log S t
_
m t ! t as t tends to infinity, we can prove that
the preceding probability converges to exp (j exp(tjx)).
Note that the Condition (G6) is not needed. This comes from a stochastic argument of
comparison as detailed at the end of the proof.
2.1. Example of application
An example of process satisfying the assumptions of Corollary 2.1 is given by the
process that describe the behavior of the likelihood test statistic for the simple Gaussian
mixture. We consider the following testing situation:
H
0
: N 0. 1
H
1
: 1 p
0
N 0. 1 p
0
N t
0
. 1
_
where p
0
2 (0, 1) and t
0
2 [jT, T] for some T > 0 and t
0
m 0. It is proved in Dacunha-
Castelle and Gassiat (1997, 1999) that under a local alternative assumption (t
0
fixed and
p
0

n
p
for instance) the asymptotic distribution (as the number n of independent
observations tends to infinity) of the logYlikelihood ratio test statistic is related to that of
the supremum of a Gaussian process with mean and covariance
m t
c
2
e
tt
0
1

e
t
2
1
p and r s. t
e
st
1

e
s
2
1
p
e
t
2
1
p .
It is deduced in Aza s et al. (2003) that, for such kind of mixture and other general
mixtures, the logYlikelihood ratio test has a power that tends to the level when the para-
meter T tends to infinity.
3. Main tools
In this part we express classical results about Poisson convergence and normal
comparison. Let J denote the family that consists of finite unions of bounded and
disjoint intervals of R.
ASYMPTOTIC POISSON CHARACTER OF EXTREMES IN NON-STATIONARY GAUSSIAN MODELS 305
Proposition 3.1: (Kallenberg (1973)): Let N
n

n 2N
a sequence of point processes on
R satisfying for any B 2 J,
a E N
n
B
n !1
! B j j.
b P N
n
B 0
n !1
!exp B j j .
Then as n Y +V, N
n
converges weakly to a standard Poisson process.
Comment:
It is enough to replace J by the family consisting in elements of the form of B
[
p
i 1
c
i
. d
i
called
""
J.
A central tool in our study is the discretization of the initial Gaussian process X.
Actually we need to show the independence of the maximum of the discretized process
on disjoint and bounded sets of (0, + V) (Lemma 4.3 proved in Section 4.4). It will be
useful to know how to compare in distribution two Gaussian vectors. This result, based
on early works of Plackett (1954), Slepian (1962) and Berman (1964, 1971), is presented
here under its version improved by Li and Shao (2002).
Proposition 3.2 (Normal comparison lemma): Let us consider two Gaussian vectors
U N
k
(0, ) and V N
k
(0, L). Assume that
ii
= L
ii
=1 for all i 2 {1, . . . , k} and let
,
ij
= max(j
ij
j, jL
ij
j). Then
P U
i
w
i
. i 1. . . . . k P V
i
w
i
. i 1. . . . . k j j

1
2

1 i < j k
arcsin

ij
_ _
arcsin 0
ij
_ _

exp
w
2
i
w
2
j
2 1 ,
ij
_ _
_ _
.
4. Proofs
Once Theorem 2.2 has been proved, the proof of Theorem 2.3 is direct using the Bunit-
speed^ transformation. Thus we will only prove Theorem 2.2 and work under the set of
Hypotheses (U1YU6). The domain of convergence of R
u
being (0, +V) it suffices to
check the characterization of Kallenberg over
~
JJ which contains of family of sets
B [
p
i 1
c
i
. d
i
where 0 < c
1
< d
1
< c
2
< d
2
< . . . < d
p
< +V.
Throughout the different proofs we will consider
Bsemi-dilated^ intervals
I
u
t
u
a. t
u
b with 0 < a < b and t
u
O C

1
2
u
_ _
.
Bdilated^ intervals
D
u.i
C
1
u
c
i
. C
1
u
d
i
_
for i 2 f1. . . . . pg with B [
p
i 1
c
i
. d
i
2
~
J.
D
u
C
1
u
c. C
1
u
d
_
with 0 < c < d.
306 AZAI

S AND MERCADIER
(I
u,j
)
j =1, . . . , n
a partition of D
u
into n = n
u
Bsemi-dilated^ intervals. For the sequel, it is
important to keep in mind that n O C

1
2
u
_ _
.
We will use discretization arguments in order to apply normal comparison Lemma.
Consequently, let q = q
u
a discretization step that depends on u such that (U7) qu Y 0
as u Y +V.
Define for any subset B of R
U
q
u
B # l 2 Z : q l 1 2 B. ql 2 B. X q l 1 < u < X ql f g.
4.1. Proof of Theorem 2.2
Before turning to the proof of our main result, we express intermediate results.
Lemma 4.1: Under Assumptions (U2YU5),
P M
I
u
u E U
u
I
u
o C
1
2
u
_ _
with o C
1
2
u
_ _
uniform for all intervals I
u
such that a Q > 0.
Lemma 4.2: Under Assumptions (U1, U4YU6, U7)
E U
u
I
u
U
q
u
I
u

_ _
o C
1
2
u
_ _
with o C
1
2
u
_ _
uniform for all I
u
such that 0 < e a < b e G.
Lemma 4.3: Under Assumptions (U2YU7),
P X ql u. 8 ql 2 [
n
j 1
I
u.j
_ _

n
j 1
P X ql u. 8 ql 2 I
u.j
_ _
o 1 .
P X ql u. 8 ql 2 [
p
i 1
D
u.i
_ _

p
i 1
P X ql u. 8 ql 2 D
u.i
_ _
o 1 .
P

p
i 1
M
D
u.i
u
_ _
_ _

p
i 1
P M
D
u.i
u
_ _
o 1 .
Lemma 4.4: Under Assumptions (U1YU6),
P M
D
u
u exp d c o 1 .
The proofs of these lemmas are postponed up to next sections.
Now let B [
p
i 1
c
i
. d
i
be an element of the family J
~
. Define the function
j t sup
s t
m s j. j Assumption (U5) implies that for any positive c
lim
u !1
uj C
1
u
c
_ _
0.
We start by checking Condition (a) of Proposition 3.1.
ASYMPTOTIC POISSON CHARACTER OF EXTREMES IN NON-STATIONARY GAUSSIAN MODELS 307
Using the fact that the variance is constant, it may be seen that for any positive t the
variables X
t
and X
t
0
are independent. Consequently by Rice formula the first moment of
R
u
(B) is
E R
u
B E U
u
C
1
u
B
_ _ _ _

_
C
1
u
B
E X
0
t
_ _
u m t dt.
Since E X
0
t
_ _
E G m
0
t

_ _
where G is a standard Gaussian variable and
sup
t 2C
1
u
B
m
0
t j j ! 0. it is obvious that E X
0
t
_ _

2
p
1 o 1 with o(1) uniform
with respect to t 2 C
1
u
B. After a change of variable we get
E R
u
B
_
B
exp um C
1
u
v
_ _

1
2
m
2
C
1
u
v
_ _
_ _
dv 1 o 1 .
Uniformly with respect to v 2 B and according to (U5)
exp um C
1
u
v
_ _

1
2
m
2
C
1
u
v
_ _
_ _
exp uj C
1
u
c
1
_ _ _ _
Const .
For every v 2 B, lim
u !1
exp um C
1
u
v
_ _

1
2
m
2
C
1
u
v
_ _ _ _
1.
Thus by application of Lebesgue Theorem, one obtains
E R
u
B B j j 1 o 1 . 1
Now we check Assumption (b) of Proposition 3.1.
P R
u
B 0 P

p
i 1
R
u
c
i
. d
i
0 f g
_ _
P

p
i 1
U
u
D
u.i
_ _
0
_ _
_ _
.
This last probability is related to the supremum of the process since
0 P

p
i 1
U
u
D
u.i
0
_ _
_ _
P

p
i 1
fM
D
u.i
ug
_ _
p u m C
1
u
c
1
_ _ _ _
o 1 .
Thus we deduce from Lemma 4.3 that
P R
u
B 0 P

p
i 1
M
D
u.i
u
_ _
_ _
o 1

p
i 1
P M
D
u.i
u
_ _
o 1 .
The desired result is obtained since Lemma 4.4 leads to
P R
u
B 0

p
i 1
exp d
i
c
i
o 1 exp B j j o 1 .

4.2. Proof of Lemma 4.1
The upper bound comes from the equality
M
I
u
u f g X C

1
2
u
a
_ _
u
_ _
[ X C

1
2
u
a
_ _
u. U
u
I
u
! 1
_ _
a.s. 2
308 AZAI

S AND MERCADIER
which leads, by Markov inequality, to
P M
I
u
u
"
u m C

1
2
u
a
_ _ _ _
E U
u
I
u
.
The first stage follows from the equivalence
"
u m C

1
2
u
a
_ _ _ _
o C
1
2
u
_ _
.
Set v
2
E U
u
I
u
U
u
I
u
1 . Using (2) we deduce, as it is done in Aza s et al.
(1999), that
P M
I
u
u P X C

1
2
u
a
_ _
u
_ _
PU
u
I
u
!1 P X C

1
2
u
a
_ _
u. U
u
I
u
! 1
_ _
! P X C

1
2
u
a
_ _
u
_ _
E U
u
I
u

v
2
2
P X C

1
2
u
a
_ _
u. U
u
I
u
! 1
_ _
! E U
u
I
u

v
2
2
o C
1
2
u
_ _
.
It is then sufficient to analyze the factorial moment of order two of the random variable
U
u
(I
u
),
v
2

__
I
2
u
E X
0
s
X
0
t
,X
s
X
t
u
_ _
p
x
s
. x
t
u. u ds dt.
Set A
u
s. t E X
0
s
X
0
t
,X
s
X
t
u
_ _
p
x
s
. x
t
u. u . Observe that by (U2), there exists
T
0
> 0 such that
R
T
0
sup
s t j j !T
0
r s. t j j <
1
3
.
For a given small enough c > 0, we partition I
u
2
, into the three following subsets
E
u
s. t 2 I
2
u
. s t j j < c
_ _
F
u
s. t 2 I
2
u
. c s t j j < T
0
_ _
G
u
s. t 2 I
2
u
. s t j j ! T
0
_ _
.
We begin by studying the integral on F
u
. We know by (U4) that r
01
is uniformly
bounded on F
u
and by (U3) that sup
(s,t) 2F
u
jr (s,t)j< 1. We find that
E X
0
s
X
0
t
,X
s
X
t
u
_ _

1
4
E X
0
s
X
0
t
_ _
2
,X
s
X
t
u
_ _

1
4
E X
0
s
X
0
t
,X
s
X
t
u
_ _
2
Var X
0
s
X
0
t
,X
s
X
t
u
_ _
_ _

1
4
r
01
s. t r
10
s. t
1 r s. t
_ _
2
u
2
Var X
0
s
X
0
t
_ _
_ _

r
01
s. t r
10
s. t
1 r s. t
_ _
2
u
2
4
1
Const u
2
ASYMPTOTIC POISSON CHARACTER OF EXTREMES IN NON-STATIONARY GAUSSIAN MODELS 309
where the constant depends only on c. It allows us to upper bound the integral by
__
F
u
A
u
s. t ds dt
__
F
u
Const u
2
p
x
s
. x
t
u. u ds dt.
Since R
c
sup
ts j j!c
r s. t j j < 1 because of (U3), we get
p
x
s
. x
t
u. u Const exp
u
2
1 R
c
_ _
.
and we can conclude that
__
F
u
A
u
s. t ds dt Const u
2
exp
u
2
1 R
c

u
2
4
_ _
o C
1
2
u
_ _
.
With the same arguments as for F
u
, the conditional expectation on G
u
is upper
bounded by
E X
0
s
X
0
t
,X
s
X
t
u
_ _
Const u
2
where now the constant depends only on the process X. Hence
__
G
u
A
u
s. t ds dt Const u
2
exp
u
2
1 R
T
0

u
2
2
_ _
o C
1
2
u
_ _
.
On E
u
we set
j
s
E X
0
s
,X
s
X
t
u
_ _

r
10
s. t
1 r s. t
u
j
t
E X
0
t
,X
s
X
t
u
_ _

r
01
s. t
1 r s. t
u
o
2
s
Var X
0
s
,X
s
X
t
u
_ _
1
r
2
10
s. t
1 r
2
s. t
o
2
t
Var X
0
t
,X
s
X
t
u
_ _
1
r
2
01
s. t
1 r
2
s. t
.
A
u
(s, t) being symmetric in (s,t) we may assume that s < t. For some
1
and
2
2 (s, t),
Taylor expansions of r and r
01
about the point (s, s) give
r s. t 1
t s
2
2

t s
4
4!
r
04
s.
1

310 AZAI

S AND MERCADIER
and
r
01
s. t t s
t s
3
3!
r
04
s.
2
.
Using (U4), j
t
u
t s
2
O t s
3
_ _ _ _
with a uniform remainder. Thus for c
sufficiently small we have
j
t

t s
4
u and j
s
!
t s
4
u
for example. It follows that j
t
e 0 and j
s
Q 0. After some calculations we obtain from
(U4)
o
2
t
t s
2
r
04
s.
2

3

1
4

r
04
s.
1

4!
_ _
O t s
4
_ _
with also a uniform remainder. Thus we can find a constant C
1
depending only on c such
that:
8 s. t 2 E
u
. o
2
t
. o
2
s
C
1
t s
2
.
Furthermore under the conditional distribution we have from Cauchy-Schwarz,
inequality:
E X
0
s
X
0
t
_ _

E X
02
s
_ _
_
E X
02
t
_ _
_
.
In addition, if Z N(j,o
2
) then
E Z
2
E Z
2
j
2
o
2
for j 0
EZ
2

o
2
2
exp
j
2
2o
2
_ _
for j < 0.
Combining the latter results, we get in our case
E X
0
s
X
0
t
,X
s
X
t
u
_ _
j
2
s
o
2
s
_ _1
2
o
t

2
p exp
j
2
t
4o
2
t
_ _
Const t s
2
u exp C
2
u
2
_ _
.
where C
2
is a positive constant. Moreover, for c sufficiently small
1
1 r
2
s. t

Const
t s
2
.
whence
p
x
s
. x
t
u. u
Const
t s
exp
u
2
1 r s. t
_ _
.
A
u
s. t Const t s u exp u
2
1
2
C
2
_ _ _ _
.
and
__
E
u
A
u
s. t ds dt Const u exp
_
u
2
_
1
4
C
2
__
o C
1
2
u
_ _
.
ASYMPTOTIC POISSON CHARACTER OF EXTREMES IN NON-STATIONARY GAUSSIAN MODELS 311
The proof is now complete by fixing c in order to have all relations concerning the upper
bound on E
u
, and then applying the bounds for the integrals over the two other domains.
Note that our result is uniform for all intervals I
u
, such that a Q > 0.
4.3. Proof of Lemma 4.2
The proof is similar to that of Leadbetter et al. (1983, Lemmas 7.3.1. and 8.2.2).
With the same arguments as those used in (1), we can prove that
E U
u
I
u
I
u
j jC
u
1 o 1 3
with o(l) uniform for all intervals I
u
, such that 0 < e a < b e G. It is thus enough to
study E U
q
u
I
u

_ _
. Now it is sufficient to prove that for all family of intervals J
u
, with end
points a
u
, b
u
being a multiple of q
u
and such that a
u
$ C

1
2
u
a and b
u
$ C

1
2
u
b we have
E U
q
u
J
u

_ _
J
u
j jC
u
o C
1
2
u
_ _
.
This is a direct consequence of the fact that
P
k.q
u PXkq u < u < X k 1 q u qC
u
1 o 1
uniformly for k such that C

1
2
u
< kq u < C

1
2
u
. This is proved in the next paragraph.
Set d
k
= k q(u), the considered event can be written as
X d
k
< u < X d
k 1
f g Y
1.k
u

<
q
2
Y
2.k
_ _
where Y
1. k

X d
k
X d
k 1

2
resp. Y
2.k

X d
k 1
X d
k

q
_ _
is a Gaussian variable with
mean
m
1.k

m d
k
m d
k 1

2
and variance
o
2
1.k

1 r d
k
. d
k 1

2
resp. with mean
m
2.k

m d
k 1
m d
k

q
and variance
o
2
2.k

2 1 r d
k
. d
k 1

q
2
. Note that Y
1,k
and Y
2,k
are independent.
q
1
C
1
u
P
q.k
u C
1
u
P Y
1.k u

<
q
2
Y
2.k
_ _
q
u. k
_
1
0
dy
_
1
1
dx y exp f y. x. u. k
with
f y. x. u. k y
m
2.k

2
p
o
2.k
_ _
2

q
2
x
2
y
2
o
2
2.k
4o
2
1.k

u m
1.k
_ _
o
2.k
qxy

2
p
o
2
1.k
and
u. k
o
2.k
o
1.k
exp
u
2
2o
2
1.k
1 o
2
1.k
_ _

m
2
1.k
2o
2
1.k

um
1.k
o
2
1.k
_ _
.
312 AZAI

S AND MERCADIER
um
1.k
uj C

1
2
u

_ _
. Consequently u m
1,k
and m
1,k
converge to 0.
m
2.k

sup
x 2 C
1,2
u
.C
1,2
u

m
0
x j j. Hence m
2,k
also converges to 0.
o
1,k
and o
2,k
converge to 1. Indeed
o
2
1.k
1


r d
k
. d
k 1
1
2

r d
k
. d
k 1
r d
k
. d
k

2

q sup
x 2 d
k
.d
k 1

r
01
d
k
. x j j.
Since r
01
is bounded on C
1,2
u
. C
1,2
u

_ _
. we can conclude. Now, let us look at what
occurs for o
2
2.k
. The Taylor expansion of order four of r is
r d
k
. d
k 1
1
q
2
2

r
04
d
k.t
_ _
q
4
4
where t 2 (d
k
, d
k + 1
). Assumption (U4) leads to
o
2
2.k
1


r
04
d
k.
t
_ _
q
2
4
Const q
2
u !1
!
0.
First of all, one can deduce from preceding convergences that (u, k) converges to 1
uniformly in k. Moreover, we have the upper bound
f y. x. u. k y
m
2.k

2
p
o
2.k
_ _
2

u m
1.k
_ _
o
2.k
qxy

2
p
o
2
1.k
y
m
2.k

2
p
o
2.k
_ _
2

u m
1.k
_ _
o
2.k
qy

2
p
o
2
1.k
obtained for x 1
y
2

2
p
m
2.k
o
2.k

u m
1.k
_ _
o
2.k
q

2
p
o
2
1.k
_ _
y.
Consequently when y
2

2
p
m
2.k
o
2.k

2
p
u m
1. k
_ _
o
2. k
q
o
2
1. k
one has f y. x. u. k
y
2
2
. Thus
the integral for ( y, x) on
2

2
p
m
2.k
o
2.k

2
p
u m
1.k
_ _
o
2.k
q
o
2
1.k
. 1
_ _
1. 1 tends by
dominated convergence to
2
_
1
0
y exp y
2
_ _
dy 1.
The remaining integral is upper bounded by
2

2
p
m
2.k
o
2.k

2
p
u m
1.k
_ _
o
2.k
q
o
2
1.k
_ _
2
exp 2

2
p
m
2.k
o
2.k

u m
1.k
_ _
o
2.k
q

2
p
o
2
1.k
_ _
2
_
_
_
_
which converges to 0. Since q
1
C
1
u
P
q.k
u ! 1 the proof is complete.

4.4. Proof of Lemma 4.3


The first two statements come from same arguments. To prove the first part we denote by
L the variance matrix of the vector
_
Xql. 8ql 2 [
n
j 1
I
u. j
_
. It can be written using L
j
1
j
2
, the
ASYMPTOTIC POISSON CHARACTER OF EXTREMES IN NON-STATIONARY GAUSSIAN MODELS 313
variance matrix of X(ql) for all ql 2 I
u, j
1
and X(ql ) for all ql 2 I
u, j
2
. Let be the diagonal
matrix exit of L.
0

0
11
0
12
. . . 0
1n
0
21
0
22
. . . 0
2n
.
.
.
.
.
.
.
.
.
0
n1
0
n2
0
nn
_

_
_

0
11
0 0
0 0
22
0
.
.
.
.
.
.
0 0 0
nn
_

_
_

_
We write {d
1
, . . . , d
N
} the set of discretized points present in D
u


n
j 1
I
u. j
. One can
describe these two matrices in the following way:
- elements of L are the `
l,k
= r(d
l
,d
k
) for (l, k) 2 {1, . . . , N}
2
- those of are denoted by o
l,k
for (l, k) 2 {1, . . . , N}
2
.
By applying normal comparison Lemma, we have:
P X ql u. 8 ql 2
_
n
j 1
I
u. j
_ _

n
j 1
PX ql u. 8 ql 2 I
u. j
_

1
2

1 l <k N
arcsin `
l. k
_ _
arcsin o
l. k
_ _

exp
u m d
l

2
um d
k

2
2 1 `
l. k

_ _
_ _
.
On the diagonal of L, in other words when l and k satisfy d
l
, d
k
2 I
u, j
, we know that
`
l,k
= o
l,k
. Thus, it is sufficient to carry out this sum on the couples
K fl. k. 1 l < k N. 9 j
1
6 j
2
. d
l
2 I
u. j
1
. d
k
2 I
u. j
2
g.
The preceding inequality becomes:
P X ql u. 8 ql 2
_
n
j 1
I
u. j
_ _

n
j 1
P X ql u. 8 ql 2 I
u. j
_ _

Const

l. k 2K
r d
l
. d
k
j j exp
u m d
l

2
u m d
k

2
2 1 r d
l
. d
k
j j
_ _
Const

l. k 2K
r d
l
. d
k
j j exp
u
2
1 r d
l
. d
k
j j

u m d
l
m d
k

1 r d
l
. d
k
j j
_ _
.
4
Recall that we have defined T
0
such that R
T
0
sup
s t j j T
0
r s. t j j <
1
3
. Let us introduce
the following sets:
K
1
l. k 2 K. d
l
d
k
j j < T
0
f g
K2 l. k 2 K. T
0
< d
l
d
k
j j < C
1
u
d c ,n
_ _
K3 l. k 2 K. C
1
u
d c ,n < d
l
d
k
j j < C
1
u
d c
_ _
.
314 AZAI

S AND MERCADIER
Set Z(a) the number of occurrences of (l, k) 2 K such that jd
l
j d
k
j = qa. It is easy to see
that:
Za an 1 for qa <
C
1
u
n
d c
Za
C
1
u
q
d c for every a.
Using #K
1
e (nj1) (T
0
+ T
0
j1 + . . . + 2 + 1) and 0 e jr (d
l
, d
k
)j e 1, we have:

l.k 2 K
1
r d
l
. d
k
j jexp
u
2
1 r d
l
. d
k
j j

u m d
l
m d
k

1 r d
l
. d
k
j j
_ _
Const C

1
2
u
exp
u
2
2
Const u
_ _
Const exp
u
2
4
Const u
_ _
u !1
!0.
The condition on R
T
0
ensures that:

l.k 2 K
2
r d
l
. d
k
j j exp
u
2
1 r d
l
. d
k
j j

u m d
l
m d
k

1 r d
l
. d
k
j j
_ _


C
1
u
n
d c
qa T
0
j j 1
Z a exp
u
2
1 R
T
0
Const u
_ _

C
1
u
n
C
1
u
n
n 1 exp
u
2
1 R
T
0
Const u
_ _
Const
C
2
u
n
exp
u
2
1 R
T
0
Const u
_ _
u !1
!0.
We write , t sup
s t j j t
r s. t j j, using changes of variable we get successively:

l.k 2K
3
rd
l
. d
k
j j exp
u
2
1 jrd
l
. d
k
j

u m d
l
m d
k

1 r d
l
. d
k
j j
_ _

C
2
u
q
2
_
C
1
u
d c
C
1
u
d c
n
, t exp
u
2
1 , t
2 u j C
1
u
c
_ _
_ _
dt

C
2
u
q
2
u
2
_
d c
d c
n
u
2
, tC
1
u
_ _
exp
u
2
1 , C
1
u
_ _ 2 u j C
1
u
c
_ _
_ _
dt

1
q
2
u
2
_
dc
d c
n
u
2
, tC
1
u
_ _
exp u
2
, tC
1
u
_ _
2 u j C
1
u
c
_ _ _ _
dt.
ASYMPTOTIC POISSON CHARACTER OF EXTREMES IN NON-STATIONARY GAUSSIAN MODELS 315
Since u
2
, tC
1
u
_ _
! 0 by (U2) and using (U5), Lebesgue Theorem implies that
_
d c
d c
n
u
2
, tC
1
u
_ _
exp u
2
, tC
1
u
_ _
2 u j C
1
u
c
_ _ _ _
dt
u !1
!0.
Thus we get the convergence to 0 of the sum on K
3
for a convenient choice of q and
finally the convergence to 0 of (4).
Now we prove the last statement of Lemma 4.3. As in the relation (4.7) we can show
that
0 P

p
i 1
X ql u. 8 ql 2 D
u.i
_ _
_ _
P

p
i 1
M
D
u.i
u
_ _
_ _
o C
1
2
u
_ _
. 5
Applying what precedes, we have
P

p
i 1
X ql u. 8 ql 2 D
u.i
_ _
_ _

p
i 1
P X ql u. 8 ql 2 D
u.i
_ _
o 1 6
and then using (4.5) applied to D
u,1
, . . . , D
u,p
and to each D
u,i
separately we get
the result.
4.5. Proof of Lemma 4.4
The proof is derived from a partition of D
u
into Bsemi-dilated^ intervals. According
to Lemma 4.2, we know that E U
u
I
u. j
_ _
U
q
u
I
u. j
_ _ _ _
is an o C
1
2
u
_ _
uniform for
(I
u, j
)
j = 1, . . . , n
, which allows us to write
0 P X ql u. 8ql 2
_
n
j 1
I
u. j
_ _
P M
D
u
u

n
j 1
P X c
j
_ _
u
_ _
P X c
j
_ _
u. U
u
I
u. j
_ _
! 1. U
q
u
I
u. j
_ _
0
_ _ _ _

n
j 1
"
u m c
j
_ _ _ _
E U
u
I
u. j
_ _
U
q
u
I
u. j
_ _ _ _ _ _
and to obtain
0 P X ql u. 8 ql 2
_
n
j 1
I
u. j
_ _
P M
D
u
u n o C
1
2
u
_ _
o 1 . 7
By using what precedes and the first part of Lemma 4.3, we can write
P M
D
u
u

n
j 1
P X ql u. 8 ql 2 I
u. j
_ _
o 1 .
316 AZAI

S AND MERCADIER
As in the proof of the relation (7), we have
P X ql u. 8 ql 2 I
u. j
_ _
P M
I
u. j
u
_ _
o C
1
2
u
_ _
. 8
Furthermore, according to Lemma 4.1
P M
I
u. j
u
_ _
1 I
u. j

C
u
o C
1
2
u
_ _
1
d c
n

n
with (
n
)
n
satisfying n sup
n 2N

n
! 0 as u !1. Using (8), the same kind of relation
is true after discretization
P X ql u. 8ql 2 I
u. j
_ _
1
d c
n
~
n
with again n sup
n 2N
~
n
! 0 as u !1. We obtain

n
j 1
P X ql u. 8ql 2 I
u. j
_ _
u !1
!exp d c
by upper and lower bounds.
4.6. Proof of Corollary 2.1
Step 1
Let X satisfying assumptions of Corollary 2.1 and also (G6), in other words
m
0
X
t

r
11
t. t
_
t ! 1
!0.
Denote Y Y
t
X
S
1
t
. t 2 R
_ _
the process constructed by Bunit-speed^ transforma-
tion. It satisfies assumptions of Theorem 2.2. Set M
Y
. U
Y
u
. R
Y
u
and m
Y
the obvious
notations related to the process Y.
Let x 2 R. t exp x and u
2
2 log S t x log

2
p _ _ _ _
.
P M
t
u P M
Y
S t
u
_ _
$ P U
Y
u
S t 0
_ _
P R
Y
u
C
u
S t 0
_ _
A simple calculation shows that C
u
S(t) = t. Moreover, from Theorem 2.2, we have the
convergence P R
Y
u
t 0
_ _
t !1
!
exp t from which we get
P M
t
u
t !1
!exp e
x
.
The result follows at once since u is written in the form x,a
t
b
t
o a
1
t
_ _
.
Step 2
We now assume that X only satisfies Assumptions (GlYG5).
Let us consider K a C
1
function with support included in [0, l] and satisfying
_
K t dt 1. We define
m
1
t sup
s t
m
Y
s . m
2
t m
1
K t and v t m
2
S t .
ASYMPTOTIC POISSON CHARACTER OF EXTREMES IN NON-STATIONARY GAUSSIAN MODELS 317
One can check the following assertions:
since m
2
is C
1
it follows that v is at least C
1
,
lim
t !1
m
0
2
t 0 which yields
v
0
t

r
11
t. t
_ m
0
2
S t
t !1
!0.
from m
2
(t) Q m
1
(t) Q m
Y
(t) we deduce that v Q m
X
,
0 e m
2
(t) e m
1
(t j 1) and so lim
t !1

log S t
_
v t 0.
The Gaussian process V with mean i and covariance function r is stochastically superior
to X. Furthermore, it verifies assumptions of the first part of this proof. Thus
limsup
t !1
P a
t
M
X
t
b
t
_ _
x
_ _
! lim
t !1
P a
t
M
V
t
b
t
_ _
x
_ _
expe
x
.
With a similar proof (using inf
s > t
m
Y
(s) instead of m
1
(t)) we can construct a Gaussian
process W stochastically inferior to X checking again assumptions of Step 1 and get the
inverse inequality.
References
Aza s, J.-M. and Wschebor, M. Book in preparation.
Aza s, J.-M., Cierco-Ayrolles, C. and Croquette, A., BBounds and asymptotic expansions for the distribution of
the maximum of a smooth stationary Gaussian process,^ ESAIM Probab. Statist. 3, 107Y129, electronic, (1999).
Aza s, J.-M., Gassiat, E

. and Mercadier, C., Asymptotic distribution and power of the likelihood ratio test for
mixtures: Bounded and unbounded cases: Technical report. Toulouse III, France: Laboratoire de Statistique
et Probabilites, http://www.lsp.ups-tlse.fr/Fp/Mercadier/lrt.pdf, (2003).
Berman, S.M., BLimit theorems for the maximum term in stationary sequences,^ Ann. Math. Statist. 35,
502Y516, (1964).
Berman, S.M., BAsymptotic independence of the numbers of high and low level crossings of stationary
Gaussian processes,^ Ann. Math. Statist. 42, 927Y945, (1971).
Berman, S.M., BExcursions above high levels for stationary Gaussian processes,^ Pacific J. Math. 36, 63Y79, (1971).
Cramer, H. and Leadbetter, M.R., Stationary and Related Stochastic Processes. Sample Function Properties
and their Applications, John Wiley & Sons Inc., New York, 1967.
Dacunha-Castelle, D. and Gassiat, E

., BTesting in locally conic models, and application to mixture models,^


ESAIM Probab. Statist., 1, 285Y317, electronic, (1997).
Dacunha-Castelle, D. and Gassiat, E

., BTesting the order of a model using locally conic parameterization:


population mixtures and stationary ARMA processes,^ Ann. Statist. 27(4), 1178Y1209, (1999).
Kallenberg, O., BCharacterization and convergence of random measures and point processes,^ Z. Wahrschein-
lichkeitstheorie und Verw. Gebiete 27, 9Y21, (1973).
Leadbetter, M.R., Lindgren, G. and Rootzen, H., Extremes and Related Properties of Random Sequences and
Processes, Springer-Verlag, New York, (1983).
Li, W.V. and Shao, Q.-M., BA normal comparison inequality and its applications,^ Probab. Theory Related
Fields 122(4), 494Y508, (2002).
Piterbarg, V.I. and Konakov, V.D., BHigh level excursions of Gaussian fields and the weakly optimal choice of
the smoothing parameter. I,^ Math. Methods Statist., 4(4), 421Y434, (1995).
Plackett, R.L., BA reduction formula for normal multivariate integrals,^ Biometrika, 41, 351Y360, (1954).
Slepian, D., BThe one-sided barrier problem for Gaussian noise,^ Bell System Tech. J., 41, 463Y501, (1962).
VolkonskiN , V.A. and Rozanov, J.A., BSome limit theorems for random functions. II,^ Teor. I Primenen. 6,
202Y215, (1961).
Wschebor, M., Surfaces aleatoires, volume 1147 of Lecture Notes in Mathematics, Springer-Verlag, Berlin,
Mesure geometrique des ensembles de niveau. [The geometric measure of level sets], 1985.
318 AZAI

S AND MERCADIER

Anda mungkin juga menyukai