R. S A B R E (Dijon)
SPECTRAL DENSITY ESTIMATION FOR STATIONARY STABLE RANDOM FIELDS
Abstract. We consider a stationary symmetric stable bidimensional pro- cess with discrete time, having the spectral representation (1.1). We consider a general case where the spectral measure is assumed to be the sum of an absolutely continuous measure, a discrete measure of finite order and a finite number of absolutely continuous measures on several lines. We estimate the density of the absolutely continuous measure and the density on the lines.
1. Introduction. A complex random variable X = X
1+ iX
2is sym- metric α-stable (S.α.S) if its characteristic function is of the form
exp n
− R
S2
|t
1x
1+ t
2x
2|
αdΓ
X1,X2(x
1, x
2) o
,
where t = t
1+ it
2and Γ
X1,X2is a symmetric measure on the unit sphere S
2of R
2.
A stochastic process {X
t, t ∈ T } is (S.α.S) if all linear combinations, a
1X
t1+ . . . + a
nX
tn, are (S.α.S). When X = X
1+ iX
2and Y = Y
1+ iY
2are jointly (S.α.S) and 1 < α ≤ 2, the covariation of X with Y is defined in [3] by
[X, Y ]
α= R
S4
(x
1+ ix
2)(y
1+ iy
2)
hα−1idΓ
X1,X2,Y1,Y2(x
1, x
2, y
1, y
2), where by convention, for b > 0 and Z a complex number, Z
hbi= |Z|
b−1Z
∗, Z
∗being the complex conjugate of Z. The quantity kXk
α= [X, X]
1/ααis a norm [11] on the linear space of (S.α.S) random variables. If (ξ
t)
t∈Ris a complex (S.α.S) process with independent increments, then the measure µ
1991 Mathematics Subject Classification: 62G07, 60G35.
Key words and phrases: periodogram, Jackson kernel, double kernel method, (S.α.S) process.
[107]
defined by µ(]s, t]) = kξ
t− ξ
sk
ααis a Lebesgue–Stieltjes measure [3], called the spectral measure of ξ. When µ is absolutely continuous, its density is called the spectral density of ξ.
For stationary complex symmetric α-stable (S.α.S) processes having the spectral representation
X(t) =
∞
R
−∞
e
itλdξ(λ),
where ξ is a (S.α.S) process with independent isotropic increments, the spectral density function φ is estimated in [11].
Let us consider a bidimensional (random field), complex (S.α.S) process in discrete time, having the spectral representation
(1.1) X(n, m) =
π
R
−π π
R
−π
e
i(λ1n+λ2m)dξ(λ
1, λ
2),
where ξ is a (S.α.S) process with independent isotropic increments; that means ξ is an additive complex function defined on the Borel subsets of [−π, π]
2such that:
• for any integer k and any Borel sets B
1, . . . , B
k, (ξ(B
1), . . . , ξ(B
k)) is (S.α.S);
• for any integer k and any disjoint Borel sets B
1, . . . , B
k, ξ(B
1), . . . . . . , ξ(B
k) are complex (S.α.S) independent random variables;
• for all Borel sets B, the distribution of the random variable e
iθξ(B) is independent of θ.
For more details about the spectral representation see [8] and [1].
We define the spectral measure by [ξ(B), ξ(B)]
α= R
B
dµ(λ
1, λ
2) for any Borel subset B of [−π, π]
2, where [X, Y ]
αdenotes the covariation of X with Y [3] when X and Y are jointly (S.α.S).
To our knowledge the case we deal with is more general than those considered in [11], [6]: we suppose that the spectral measure of the process is the sum of an absolutely continuous measure, a discrete measure of finite order and a finite number of absolutely continuous measures on several lines:
dµ(λ
1, λ
2) = φ(λ
1, λ
2)dλ
1dλ
2+
q
X
j=1
a
0jδ
(w1j,w2j)(1.2)
+
q0
X
i=1
φ
i(u
1)δ
u2=aiu1+bi,
where φ and φ
ifor i = 1, . . . , q
0are nonnegative continuous functions, a
0j∈ R
+, a
i, b
i∈ R and w
1j, w
2j∈ [−π, π], j = 1, . . . , q, i = 1, . . . , q
0.
Using Jackson polynomial kernels and the double kernel method, we estimate the function φ(λ
1, λ
2) for every (λ
1, λ
2) in ]−π, π[
2. Under appro- priate conditions on the function φ, we obtain rates of convergence. Let us also indicate that our methods allow us to estimate the positive num- bers a
0j(j = 1, . . . , q) and the functions φ
i(i = 1, . . . , q
0). For brevity, we consider here only the estimation of φ; for details, see [13].
2. The periodogram and its statistical properties. Given N M observations of the process X : (X(n, m))
0≤n≤N −10≤m≤M −1
, where N and M satisfy:
• N − 1 = 2k(n − 1) with n ∈ N, k ∈ N ∪ {1/2}; if k = 1/2 then n = 2n
1− 1, n
1∈ N.
• M − 1 = 2k(m − 1) with m ∈ N, k ∈ N ∪ {1/2}; if k = 1/2 then m = 2m
1− 1, m
1∈ N.
We consider the function H
(N )(λ) =
k(n−1)
X
m0=−k(n−1)
h
k(m
0/n) cos(λm
0),
where h
kis defined in [7] such that H
(N )(λ) = 1
q
k,nsin
nλ2sin
λ2 2kwith q
k,n= 1 2π
π
R
−π
sin
nλ2sin
λ2 2kdλ.
The Jackson polynomial kernel is |H
N(λ)|
α= |A
NH
(N )(λ)|
α, where A
N= B
−1/αα,Nwith B
α,N= R
π−π
|H
(N )(λ)|
αdλ. We define H
M(λ) in the same way.
The following lemma, proved in [6], will be frequently used in the sequel.
Lemma 2.1 [6]. Let B
α,N0=
π
R
−π
sin
nλ2sin
λ22kα
dλ and J
N,α=
π
R
−π
|u|
γ|H
N(λ)|
αdλ,
where γ ∈ ]0, 2]. Then
B
α,N0
≥ 2π 2 π
2kαn
2kα−1if 0 < α < 2,
≤ 4πkα
2kα − 1 n
2kα−1if 1
2k < α < 2,
and
J
N,α≤
π
γ+2kα2
2kα(γ − 2kα + 1) · 1
n
2kα−1if 1
2k < α < γ + 1 2k , 2kαπ
γ+2kα2
2kα(γ + 1)(2kα − γ − 1) · 1
n
γif γ + 1
2k < α < 2.
The proof of the following lemma is the same as in the one-dimensional case in [3].
Lemma 2.2. If ξ is a (S.α.S ) process with independent and isotropic increments, then
E exp
i Re h R
π−π π
R
−π
f (u
1, u
2) dξ(u
1, u
2) i
= exp
−C
απ
R
−π
|f (u
1, u
2)|
αdµ(u
1, u
2)
for every f ∈ L
α(µ), where C
α= (2πα)
−1R
π−π
|cos θ|
αdθ.
Now we denote by A the set of points where there are no atoms:
A =
(λ
1, λ
2) ∈ ]−π, π[
2: λ
16= w
1j, λ
26= w
2jand λ
2− b
i− a
iλ
12π 6∈ Z for 1 ≤ i ≤ q
0and 1 ≤ j ≤ q
. In the sequel we choose large k such that kα > 1 and we consider the following periodogram:
d
N,M(λ
1, λ
2) = A
NA
MRe
k(n−1)X
n0=−k(n−1)
k(m−1)
X
m0=−k(m−1)
h
kn
0n
h
km
0m
×e
−i(λ1n0+λ2m0)X(n
0+ k(n − 1), m
0+ k(m − 1))
. Following the study realized in the one-dimensional case by Masry and Cam- banis in [11], one can show easily that, for (λ
1, λ
2) ∈ A,
(2.1) lim
N →∞
E{exp[ird
N,M(λ
1, λ
2)]} = exp[−C
α|r|
αφ(λ
1, λ
2)].
We modify this periodogram taking the power p, with 0 < p < α/2, and multiplying by a normalization constant:
I b
N,M(λ
1, λ
2) = C
p,α|d
N,M(λ
1, λ
2)|
p.
The normalization constant C
p,αis given by C
p,α= D
pF
p,αC
αp/α, where
D
p=
∞
R
−∞
1 − cos u
|u|
1+pdu and F
p,α=
∞
R
−∞
1 − e
−|u|α|u|
1+pdu, 0 < p < α 2 . We show that
E b I
N,M(λ
1, λ
2) = [ψ
N,M(λ
1, λ
2)]
p/α, (2.2)
Var[ b I
N,M(λ
1, λ
2)] = V
α,p[ψ
N,M(λ
1, λ
2)]
2p/α, (2.3)
where V
α,p= C
p,α2C
2p,α−1− 1 and
ψ
N,M(λ
1, λ
2) = I
N,M(λ
1, λ
2) + J
N,M(λ
1, λ
2) + K
N,M(λ
1, λ
2) with
I
N,M(λ
1, λ
2) =
π
R
−π π
R
−π
|H
N(λ
1− u
1)|
α|H
M(λ
2− u
2)|
αφ(u
1, u
2) du
1du
2,
J
N,M(λ
1, λ
2) =
q
X
j=1
a
0j|H
N(λ
1− w
1j)|
α|H
M(λ
2− w
2j)|
α,
K
N,M(λ
1, λ
2) =
q0
X
i=1 π
R
−π
|H
N(λ
1− v
1)|
α|H
M(λ
2− a
iv
1− b
i)|
αφ
i(v
1) dv
1. As in [11] one can show that for (λ
1, λ
2) ∈ A, the function ψ
N,M(λ
1, λ
2) converges to φ(λ
1, λ
2). Therefore b I
N,M(λ
1, λ
2) is an asymptotically unbiased estimator of [φ(λ
1, λ
2)]
p/αbut it is not consistent because the variance is proportional to [φ(λ
1, λ
2)]
2p/α.
3. The smoothed periodogram. In this section, using two spectral windows, we smooth the periodogram b I
N,Mand we obtain consistent esti- mators of [φ(λ
1, λ
2)]
p/αat the points (λ
1, λ
2) where there are no atoms. Let W be a nonnegative, even, continuous function vanishing for |λ| > 1, with R
1−1
W (λ) dλ = 1. The spectral windows W
N, W
Mare defined by W
N(λ) = M
NW (M
Nλ) and W
M(λ) = L
MW (L
Mλ) where M
Nand L
Msatisfy
N →∞
lim M
N= ∞, lim
N →∞
M
NN = 0,
M →∞
lim L
M= ∞, lim
M →∞
L
MM = 0.
We consider the following estimator:
f
N,M(λ
1, λ
2) =
π
R
−π π
R
−π
W
N(λ
1− u
1)W
M(λ
2− u
2) b I
N,M(u
1, u
2) du
1du
2. As in [6], for giving the best rate of convergence of this estimator we intro- duce on φ two hypotheses (h
1) and (h
2) called regularity hypotheses:
(h
1) φ(λ
1− u
1, λ
2− u
2) = φ(λ
1, λ
2) + R
1(λ
1, λ
2, u
1, u
2)
with |R
1(λ
1, λ
2, u
1, u
2)| ≤ C
1k(u
1, u
2)k
γ, where 0 < γ ≤ 1, (h
2) φ(λ
1− u
1, λ
2− u
2) = φ(λ
1, λ
2) + ∂φ
∂x (λ
1, λ
2)u
1+ ∂φ
∂y (λ
1, λ
2)u
2+ R
2(λ
1, λ
2, u
1, u
2)
with |R
2(λ
1, λ
2, u
1, u
2)| ≤ C
2k(u
1, u
2)k
γ, where 1 ≤ γ < 2, C
1and C
2being nonnegative constants.
Theorem 3.1. Let λ
1, λ
2∈ A. Then:
(i) f
N,M(λ
1, λ
2) is an asymptotically unbiased estimator of the quantity [φ(λ
1, λ
2)]
p/α.
(ii) Choosing k so large that γ + 1 < 2kα, we have E[f
N,M(λ
1, λ
2)] − [φ(λ
1, λ
2)]
p/α=
O
1
M
Nγ+ 1 (L
M)
γif φ satisfies (h
1), O
1 M
N+ 1 L
Mif φ satisfies (h
2).
P r o o f. Using (2.2), we have E[f
N,M(λ
1, λ
2)] =
MN(λ1+π)
R
MN(λ1−π)
LM(λ2+π)
R
LM(λ2−π)
W (u)W (v)
× [ψ
N,M(λ
1− u/M
N, λ
2− v/L
M)]
p/αdu dv.
Since W is vanishing for |λ| > 1 and p/α < 1, for N and M large enough we obtain
|E[f
N,M(λ
1, λ
2)] − [φ(λ
1, λ
2)]
p/α|
≤
1
R
−1 1
R
−1
W (u)W (v)
× |ψ
N,M(λ
1− u/M
N, λ
2− v/L
M) − φ(λ
1, λ
2)|
p/αdu dv.
Using the facts that H
Nand H
Mare 2π-periodic, |H
N|
αand |H
M|
αare two kernels, and φ is uniformly continuous on [−π, π]
2, we find that
(3.1) lim
N →∞
M →∞
I
N,M(λ
1− u/M
N, λ
2− v/L
M) − φ(λ
1, λ
2) = 0.
On the other hand,
J
N,M(λ
1− u/M
N, λ
2− v/L
M)
≤
q
X
j=1
a
0jB
α,N0B
α,M0× 1
sin
12
(λ
1− u/M
N− w
1j)
2kα
sin
12
(λ
2− v/L
M− w
2j)
2kα
. Since λ
16= w
1jand λ
26= w
2jfor j ∈ {1, . . . , q}, using Lemma 2.1 we obtain (3.2) J
N,M(λ
1− u/M
N, λ
2− v/L
M) = O
1
n
2kα−1m
2kα−1.
Considering all possible cases, and partitioning the integrals, we show easily (see [13]) that
(3.3) K
N,M(λ
1− u/M
N, λ
2− v/L
M)
= O
1
m
2kα−1+ 1
n
2kα−1+ 1
n
2kα−1m
2kα−1. Using the inequality
|x
r− y
r| ≤ r
2 (x
r−1+ y
r−1)|x − y|, x, y ∈ R
+∗, r ∈ [0, 1] ∪ [2, ∞[, and the equalities (3.2), (3.3) we can show that
[ψ
N,M(λ
1− u/M
N, λ
2− v/L
M)]
p/α−1+ [φ(λ
1, λ
2)]
p/α−1is bounded by a constant, for N and M large enough. We get
|E[f
N,M(λ
1, λ
2)] − [φ(λ
1, λ
2)]
p/α| ≤ const
1
R
−1 1
R
−1
W (u)W (v)
×|ψ
N,M(λ
1− u/M
N, λ
2− v/L
M) − φ(λ
1, λ
2)| du dv.
1) If φ satisfies the hypothesis (h
1) then we have
|ψ
N,M(λ
1− u/M
N, λ
2− v/L
M) − φ(λ
1, λ
2)|
≤ C
1λ1−u/MN+π
R
λ1−u/MN−π
λ2−v/LM+π
R
λ2−v/LM−π
|H
N(t)H
M(t
0)|
α× (|u/M
N+ t| + |v/L
M+ t
0|)
γdt dt
0+ J
N,M(λ
1− u/M
N, λ
2− v/L
M) + K
N,M(λ
1− u/M
N, λ
2− v/L
M).
Using twice the inequality
(3.4) |x + y|
r≤ |x|
r+ |y|
rif 0 < r ≤ 1, 2
r(|x|
r+ |y|
r) if r ≥ 1, we show that
1
const |E[f
N,M(λ
1,λ
2)][φ(λ
1, λ
2)]
p/α|
≤ C
1M
Nγ1
R
−1
W (u)|u|
γdu
+ C
1 1R
−1
W (u)
λ1−u/MN+π
R
λ1−u/MN−π
|H
N(t)|
α|t|
γdt du
+ C
1L
γM1
R
−1
W (v)|v|
γdv
+ C
1 1R
−1
W (v)
λ2−v/LM+π
R
λ2−v/LM−π
|H
M(t
0)|
α|t
0|
γdt
0dv
+
1
R
−1 1
R
−1
W (u)W (v)
× [J
N,M(λ
1− u/M
N, λ
2− v/L
M) + K
N,M(λ
1− u/M
N, λ
2− v/L
M)] du dv.
On the other hand, since H
Nis an even function, we have
λ1−u/MN+π
R
λ1−u/MN−π
|H
N(t)|
α|t|
γdt
≤
π
R
−π
|H
N(t)|
α|t|
γdt + (2π)
γ|λ1|+|u/MN|+π
R
π
|H
N(t)|
α|t|
γdt.
Using Lemma 2.1, for N large enough we have (3.5)
|λ1|+|u/MN|+π
R
π
|H
N(t)|
α|t|
γdt = O(T
N(λ
1)), where
T
N(λ
1) =
1
n
2kα−1if λ
16= 0, 1
M
Nn
2kα−1if λ
1= 0.
We gather the results (3.2), (3.3) and (3.5) to obtain
|E[f
N,M(λ
1, λ
2)] − [φ(λ
1,λ
2)]
p/α|
= O
T
N(λ
1) + T
M(λ
2) + 1
M
Nγ+ 1 L
γM+ 1
n
γ+ 1 m
γ+ 1
n
2kα−1+ 1
m
2kα−1+ 1
n
2kα−1m
2kα−1. Since 0 < γ ≤ 1 and γ + 1 < 2kα, it is clear that
|E[f
N,M(λ
1, λ
2)] − [φ(λ
1, λ
2)]
p/α| = O
1
M
Nγ+ 1 L
γM.
2) If φ satisfies the hypothesis (h
2), using the facts that H
N, H
Mare 2π-periodic kernels, and (3.5), (3.4) we get
1
const |E[f
N,M(λ
1, λ
2)] − [φ(λ
1, λ
2)]
p/α|
≤
∂φ
∂x (λ
1, λ
2)
1
R
−1
|u/M
N|W (u) du
+
∂φ
∂y (λ
1, λ
2)
T
M(λ
2)
+
∂φ
∂x (λ
1, λ
2)
T
N(λ
1) +
∂φ
∂y (λ
1, λ
2)
1
R
−1
|v/L
M|W (v) dv
+ 2
γC
2M
Nγ1
R
−1
|u|
γW (u) du + 2
γC
2L
γM1
R
−1
|v|
γW (v) dv
+ 2
γC
2λ1−u/MN+π
R
λ1−u/MN−π
|H
N(t)|
α|t|
γdt
+ 2
γC
2λ2−v/LM+π
R
λ2−v/LM−π
|H
M(t
0)|
α|t
0|
γdt
0+
1
R
−1 1
R
−1
W (u)W (v)
× [J
N,M(λ
1− u/M
N, λ
2− v/L
M)
+ K
N,M(λ
1− u/M
N, λ
2− v/L
M)] du dv.
We use again the results (3.2), (3.3) and (3.5) to obtain
|E[f
N,M(λ
1, λ
2)] − [φ(λ
1, λ
2)]
p/α|
= O
1 M
N+ 1 L
M+ 1 L
γM+ 1
M
Nγ+ T
N(λ
1) + T
M(λ
2) + 1
n
γ+ 1
m
γ+ 1
n
2kα−1+ 1
m
2kα−1+ 1
n
2kα−1m
2kα−1. Since 1 < γ ≤ 2, we have 1/k < (γ + 1)/(2k) < α. Thus, we obtain
|E[f
N,M(λ
1, λ
2)] − [φ(λ
1, λ
2)]
p/α| = O
1 M
N+ 1 L
M. Now we show the following lemma which will be used in the sequel.
Lemma 3.2. Let (λ
1,1, λ
2,1) and (λ
1,2, λ
2,2) belong to A. Write Q
N,M(λ
1,1, λ
1,2, λ
2,1, λ
2,2)
=
π
R
−π π
R
−π
|H
N(λ
1,1− v
1)H
M(λ
2,1− v
2)
× H
N(λ
1,2− v
1)H
M(λ
2,2− v
2)|
α/2dµ(v
1, v
2).
Then
Q
N,M(λ
1,1, λ
1,2, λ
2,1, λ
2,2)
≤ O
1
n
2kα−1+ 1
m
2kα−1+ 1
n
2kα−1m
2kα−1+ sup(φ) π
2π 2
4kα1 (nm)
2kα−1<(δ, λ
1,2, λ
1,1)<(δ
0, λ
2,2, λ
2,1), where the function < is defined by
<(x, y, z) = π
sin
x22kα+ 2x
inf
sin
|y−z|+x2 kα, sin
|y−z|4 kα, with two real numbers δ, δ
0satisfying
0 < δ < inf[π − λ
1,2; π + λ
1,1; |λ
2,1− λ
1,1|/2], 0 < δ
0< inf[π − λ
2,2; π + λ
2,1; |λ
2,2− λ
2,1|/2].
P r o o f. First using the expression of dµ in (1.1) and the inequality x
α/2y
α/2≤
12(x
α+ y
α), we obtain
Q
N,M(λ
1, λ
2, x
1, x
2) ≤ B +
12J
N,M(λ
1,1, λ
2,1)
12J
N,M(λ
1,2, λ
2,2)
+
12K
N,M(λ
1,1, λ
2,1) +
12K
N,M(λ
1,2, λ
2,2),
where
B =
π
R
−π π
R
−π
|H
N(λ
1,1− v
1)H
M(λ
2,1− v
2)
× H
N(λ
1,2− v
1)H
M(λ
2,2− v
2)|
α/2φ(v
1, v
2) dv
1dv
2. Since φ is bounded on [−π, π]
2, we have
B ≤ sup(φ)
R
π−π
|H
N(λ
1,1− v
1)H
N(λ
1,2− v
1)|
α/2dv
1× R
π−π
|H
M(λ
2,2− v
2)H
M(λ
2,1− v
2)|
α/2dv
2.
We split the first integral into five integrals: over a neighbourhood of λ
1,1, a neighbourhood of λ
1,2and the remainders:
π
R
−π
|H
N(λ
1,1− v
1)H
N(λ
1,2− v
1)|
α/2dv
1= 1
B
α,N0λ1,1−δ
R
−π
+
λ1,1+δ
R
λ1,1−δ
+
λ1,2−δ
R
λ1,1+δ
+
λ1,2+δ
R
λ1,2−δ
+
π
R
λ1,2+δ
sin
n2
(λ
1,1− w)
sin
λ1,12−w· sin
n2
(λ
1,2− w) sin
λ1,22−wkα
dw.
The first, third and last integrals are respectively bounded by 1
B
α,N0· λ
1,1− δ + π
sin
δ22kα, 1
B
α,N0· λ
1,2− λ
1,1− 2δ sin
δ22kαand
1
B
α,N0· π − λ
1,2− δ sin
δ22kα. Using the inequality sin(nx) ≤ n sin x, we obtain
1 B
α,N0λ1,1+δ
R
λ1,1−δ
sin
n2
(λ
1,1− w) sin
λ1,12−wsin
n2
(λ
1,2− w) sin
λ1,22−wkα
dw
≤ 1
B
α,N0· 2δn
kαinf
sin
|λ1,2−λ21,1|+δkα, sin
|λ1,2−λ4 1,1|kα.
Similarly, the remaining integral is bounded by the same quantity. Thus
using Lemma 2.1, we have
π
R
−π
|H
N(λ
1,1− v
1)H
N(λ
1,2− v
1)|
α/2dv
1= 1 2π
π 2
2kα1 n
2kα−1×
2π
sin
δ22kα+ 4δn
kαinf
sin
λ1,2−λ21,1+δkα, sin
λ1,2−λ4 1,1kα. Just as before, we have
π
R
−π
|H
M(λ
2,1− v
2)H
M(λ
2,2− v
2)|
α/2dv
2≤ 1 π
π 2
2kα1
m
2kα−1<(δ
0, λ
2,2, λ
2,1).
Equalities (3.2) and (3.3) give the result of this lemma.
Theorem 3.3. Let (λ
1, λ
2) ∈ A be such that φ(λ
1, λ
2) 6= 0. Then:
(i) Var[f
N,M(λ
1, λ
2)] converges to zero.
(ii) If φ satisfies (h
1) or (h
2), and M
N= n
cand L
M= m
c0, where c and c
0are two real numbers satisfying
inf 2k
2α
2+ 1
6α
2k
2, kα + 2 3(kα + 1)
< c, c
0< 1 2 , then
Var[f
N,M(λ
1, λ
2)] = O
1
n
2(1−2c)+ 1 m
2(1−2c0).
P r o o f. When φ(λ
1, λ
2) = 0 we do not need to smooth b I
N,M, since its variance tends to zero. We have
Var[f
N,M(λ
1, λ
2)]
=
π
R
−π π
R
−π π
R
−π π
R
−π
W
N(λ
1− u
1)W
M(λ
2− u
2)W
N(λ
1− u
01)W
M(λ
2− u
02)
× Cov[b I
N,M(u
1, u
2), b I
N,M(u
01, u
02)] du
1du
2du
01du
02. Putting
x
1= M
N(λ
1− u
1), x
01= M
N(λ
1− u
01),
x
2= L
M(λ
2− u
2), x
02= L
M(λ
2− u
02).
and using the fact that W is vanishing for |λ| > 1, for N and M large enough
we have
Var[f
N,M(λ
1, λ
2)]
=
1
R
−1 1
R
−1 1
R
−1 1
R
−1
W (x
1)W (x
2)W (x
01)W (x
02)
× Cov[b I
N,M(λ
1− x
1/M
N, λ
2− x
2/L
M),
I b
N,M(λ
1− x
01/M
N, λ
2− x
02/L
M)] dx
1dx
2dx
01dx
02. We define
L
1= {(x
1, x
01) ∈ [−1, 1]
2: |x
1− x
01| > σ
N}, L
2= {(x
2, x
02) ∈ [−1, 1]
2: |x
2− x
02| > σ
0M},
L
3= {(x
1, x
01, x
2, x
02) ∈ [−1, 1]
4: |x
1− x
01| ≤ σ
Nor |x
2− x
02| ≤ σ
M0}, where σ
N, σ
0Mare two nonnegative real sequences converging to 0. We split the last integral into the integral over L
3and the integral over L
1and L
2:
Var[f
N,M(λ
1, λ
2)] = R R R R
L3
+ R R
L1
R R
L2
=: J
1+ J
2.
Using (2.3) and the uniform (in x
1, x
2) convergence of ψ
N,M(λ
1− x
1/M
N, λ
2− x
2/L
M) to φ(λ
1, λ
2), we obtain
J
1≤ const h R R
|x2−x02|≤σ0M
W (x
2)W (x
02) dx
2dx
02+ R R
|x1−x01|≤σN
W (x
1)W (x
01) dx
1dx
01i .
Thus, J
1≤ const[sup(W )]
2[σ
N+ σ
M0]. Hence J
1tends to zero.
It remains to show that J
2tends to zero. First we define for simplicity λ
1,1= λ
1− x
1/M
N, λ
1,2= λ
1− x
01/M
N,
λ
2,1= λ
2− x
2/L
M, λ
2,2= λ
2− x
02/L
M, and
C(λ
1, λ
2)
= Cov[ b I
N,M(λ
1− x
1/M
N, λ
2− x
2/L
M), b I
N,M(λ
1− x
01/M
N, λ
2− x
02/L
M)].
We use the equality
|x|
p= D
−1p∞
R
−∞
1 − cos(xu)
|u|
1+pdu = D
p−1Re
∞
R
−∞
1 − e
ixu|u|
1+pdu, for all real x and p ∈ ]0, 2[. We have
(3.6) I b
N,M(λ
1, λ
2) = 1 F
p,αC
αp/αRe
∞
R
−∞
1 − e
iudN,M(λ1,λ2)|u|
1+pdu.
From (2.1) we obtain
(3.7) E b I
N,M(λ
1, λ
2) = 1 F
p,αC
αp/α∞
R
−∞
1 − exp{−C
α|u|
αψ
N,M(λ
1, λ
2)}
|u|
1+pdu.
By using (3.6) and (3.7) we show easily that
C(λ
1, λ
2) = F
p,α−2C
α−2p/α∞
R
−∞
∞
R
−∞
E
h Y
2k=1
cos(u
kd
N,M(λ
1,k, λ
2,k)) i
− exp n
−C
α 2X
k=1
|u
k|
αψ
N,M(λ
1,k, λ
2,k)
o du
1du
2|u
1u
2|
1+p. From the equality 2 cos x cos y = cos(x + y) + cos(x − y), we have
E h Y
2k=1
u
kd
N,M(λ
1,k, λ
2,k) i
= 1 2 exp h
−C
αR
2
X
k=1
u
kH
N(λ
1,k− v
1)H
M(λ
2,k− v
2)
α
dµ(v
1, v
2) i
+ 1 2 exp h
−C
αR
2
X
k=1
(−1)
k−1u
kH
N(λ
1,k− v
1)
×H
M(λ
2,k− v
2)
α
dµ(v
1, v
2) i . Substituting this expression in C(λ
1, λ
2) and changing the variable u
2to
−u
2in the second term, we obtain C(λ
1, λ
2) = F
p,α−2C
α−2p/α∞
R
−∞
∞
R
−∞
(e
−K− e
−K0) du
1du
2|u
1u
2|
1+p, where
K = C
α πR
−π π
R
−π
2
X
k=1
u
kH
N(λ
1,k− v
1)H
M(λ
2,k− v
2)
α
dµ(v
1, v
2),
K
0= C
α 2X
k=1
|u
k|
απ
R
−π π
R
−π
|H
N(λ
1,k− v
1)H
M(λ
2,k− v
2)|
αdµ(v
1, v
2)
= C
α 2X
k=1
|u
k|
αψ
N,M(λ
1,k, λ
2,k).
Since K, K
0> 0, we have
|e
−K− e
−K0| ≤ |K − K
0|e
|K−K0|−K0. Using the inequality
kx + y|
α− |x|
α− |y|
α| ≤ 2|xy|
α/2, x, y ∈ R and 1 ≤ α ≤ 2, we obtain
|K − K
0| ≤ 2C
α|u
1u
2|
α/2Q
N,M(λ
1,1, λ
1,2, λ
2,1, λ
2,2).
Therefore, C(λ
1, λ
2)
≤ F
p,α−2C
α−2p/α2C
αQ
N,M(λ
1,1, λ
1,2, λ
2,1, λ
2,2)
∞
R
−∞
∞
R
−∞
e
|K−K0|−K0|u
1u
2|
1+p−α/2du
1du
2. Now we have
(3.8) |K − K
0| − K
0≤ −C
α2
X
k=1
|u
k|
α[ψ
N,M(λ
1,k, λ
2,k) − Q
N,M(λ
1,1, λ
1,2, λ
2,1, λ
2,2)].
Let δ(N ), δ
0(M ) be two real numbers depending respectively on N and M such that
0 < δ(N ) < inf[|λ
1,1− λ
1,2|/2; π − λ
1,2; π + λ
1,1], 0 < δ
0(M ) < inf[|λ
2,1− λ
2,2|/2; π − λ
2,2; π + λ
2,1].
Moreover, we suppose that
(3.9) lim
N →∞
δ(N )M
Nσ
N= 0, lim
M →∞
δ
0(M )L
Mσ
M0= 0.
Using Lemma 3.2 and the inequality
(3.10) sin(x/2) ≥ x/π, 0 ≤ x ≤ π, we obtain
π
R
−π