Upload
dangque
View
212
Download
0
Embed Size (px)
Citation preview
PUBLICATIONS MATHÉMATIQUES ET INFORMATIQUES DE RENNES
ADAM JAKUBOWSKIA Non-Central Functional Limit Theorem for QuadraticForms in Martingale Difference SequencesPublications de l’Institut de recherche mathématiques de Rennes, 1989-1990, fascicule 1« Probabilités », p. 69-73<http://www.numdam.org/item?id=PSMIR_1989-1990___1_69_0>
© Département de mathématiques et informatique, université de Rennes,1989-1990, tous droits réservés.
L’accès aux archives de la série « Publications mathématiques et informa-tiques de Rennes » implique l’accord avec les conditions générales d’utili-sation (http://www.numdam.org/legal.php). Toute utilisation commerciale ouimpression systématique est constitutive d’une infraction pénale. Toute copieou impression de ce fichier doit contenir la présente mention de copyright.
Article numérisé dans le cadre du programmeNumérisation de documents anciens mathématiques
http://www.numdam.org/
69
A Non-Central Functional Limit Theorem for Quadratic Forms
in Martingale Difference Sequences
Adam Jakubowski* Uniwersytet Mikolaja Kopernika, Instytut Matematyki
Torun, Poland.
Abstract
For quadratic forms with nulls on the diagonal the partial-sum process is both a martingale and a stochastic integral. Using corresponding tools we derive a result on the convergence to the double Wiener-It o integral.
Let { A n = ( a J j ) } n € N be a sequence of infinite matrices with nulls on the diagonals:
< t = 0, ¿ = 1 ,2 , . . . , n G N . (1)
Let { X j } j € N be a martingale difference sequence with respect to some filtration {^/}jei\iu{o} such that
E{X2
j\T^) = \, i = 1 ,2 , . . . . (2)
Then for each n 6 N the process
Snlk= £ ¿ = 1 ,2 , . . . , (3)
l<t\j<* j < k \ i < j )
is both a square integrable martingale and a stochastic integral with respect to {Fj}-
Using the martingale structure, a Functional Central Limit Theorem for suitably scaled
{£n,A:}fceN w & s proved in [JaMe90]. In the present paper we state a non-central functional
limit theorem based on limit theorems for stochastic integrals given in [JMP89]. The limit
is identified with a double Wiener-Ito stochastic integral (see e.g. [Ito51] or [Maj81]). In
some sense it is not surprising: such limits arise, for example, in limit theory for quadratic
forms in stationary gaussian sequences exhibiting long-range dependence (see [Ros79], also
[SuHo86]).
* Supported by CPBP 01.02
1
70
For each n € N, let A n be the symmetrization of A 7 1, i.e. the matrix with entries
(alj + ali)/2' F o r A " ' d e f i n e i t s representation in Z,2([0, T ] 2 ) by the formula
AS(ti,t;) = af n t t ] f [ n v j for T > u,t> > 0. (4)
Finally, let
Xn(t) = £ X, (5) l<j<[nt]
y-(t) = n" 1 j f a&X.-X,-. (6) l<ij<[nt]
Theorem Suppose A n 's satisfy (1) and {Xj} is a martingale difference sequence such
that (2) holds.
If f\n converges in L 2 ([0, T ] 2 ) to some function A, then
Yn V F ' ( 7 )
where
Y(t) = J jA(u,v)I[0tt?(u,v)dWudWv (8)
is the classical Wiener-Ito integral P R O O F . For each e > 0 there are continuous functions I = 1,2, . . . , ra and
numbers a x , . . . , am such that
*(M) = JLa' 9
satisfies
| | A - * | | 2 < c . (9)
Hence for n > n x , ||A" — $| | 2 < 2s and by continuity of $ ||A" - $„ | | 2 < 3e for n > n 2 , where $n(s,t) = $([ns]/n, [n*]/n).
If
Zn(t) = n-1 ¿2 9(i/n,j/n)XiXjt (10)
then for n > n 2
£ sup \Yn(t) - Zn{t)\2 < AE\Yn{T) - Zn{T)\2 < 4||S* - $ n | | ' < 4 • 9e 2 . (11) l < t < T
Let
Z(*) = / J$(u,v)I[0,t?{u,v)dWudWv. (12)
By the isometry property for Wiener-Ito integrals
E sup \Y( t ) - Z(t)\2 < 4E\Y(T) - Z(T)\2 = 4\\A - $ | | 2 < 4e 2 . (13)
2
71
Hence it is enough to prove that
z n v z (14) on the space D ( [ 0 , T ] ) . But
Zn(t) = 2U-1 ¿2 ^{i/nJ/^XiXj l<i<j<[nt]
m f XX
1=1 \i<.<i<H v n v n
+ £ ^ ( i / n J M j / n ) ^ ) i<i<j<M VnVnJ
= X > (jf Mv) (jQ
v~ Hu) dx^j dx:+
v p a i (jf (jf~ ̂ /(w) ^ +
+jf m») (jf ^(«) <w«) ^ )
where the convergence in distribution holds by [JMP89, Theorem 2.6] •
Example 1. Let / : [0, l ] 2 -+ R 1 . Define quadratic forms
"?j = f
If gn{u,v) = f([nu],[nv]) —v L i f(u,v), then
n _ 1 £ -T f fl{u,v)dWudWv, X<ij<[nt\ V J ° J °
where / is the symmetrization of / .
Example 2. Let
Co = 0, C X = 6i, C 2 = ¿2, • • • , Q = Q + l = &1> c d + 2 = ¿2, • . •
Define a matrix A by ai,i = c\i-j\-
In this case A" does not converge, so we cannot apply directly our theorem.
3
72
Take n = N - d and define a rearranging of coordinates:
t\ i—» ei
erf h» e(d_i)Ar+i
(The general formula is of the form e^d+i *-* e^i^+k+i if 0 < k < N — 1, 1 < z < d). Under this rearranging, the distribution of Xn remains unchanged, while A n transforms
to the form D n = ( d p ^ ) , where for p = (i - 1)N + k + 1 and q = (j - 1)N + 1 + 1
{ 0 if j = z, 1 = k
bd if j = + k &y_t-| if j > z, l> k or j < z, I <k bd~\j-i\ if j > z, / < or j < z, / > A:.
It is easy to see that now D n —* D in £ 2 ( [0 , l ] 2 ) , where 0 if u = u 6d if [d • u] = [d • v] , u + v b\[d.u]-[d.v]\ if [d • u] > [d • v] and { d • u} > {d •
r>/ \ I o r
^ " [d. u) < [d • v] and {d • ix} < { d • &cH[*"M<HI if [d-u]> [d • v] and {d • u} < {d • i?}
or [d • u] < [d • v] and {d • u} > {d - v}
Here [x] is the integer part of x and {x} = x — [#]. The convergence is easily extendible from subsequence N • d to the whole N. But we
do not have the functional convergence, since
l<ij<[nt]
and
l<«,i<[n«]
have nothing common except for t = 1, where they coincide.
4
73
References
[Ito51] Itô, K. Multiple Wiener integral, J. Math. Soc. Japan 3 (1951), pp. 157-169.
[JaMé90] Jakubowski, A. & Mémin, J. A functional central limit theorem for quadratic forms in independent random variables, preprint (1990).
[JMP89] Jakubowski, A., Mémin, J. & Pages, G. Convergence en loi des suites d'integrales stochastiques sur l'espace $1 de Skorokhod Probab. Th. Rel. Fields 81 (1990), pp. 111-137.
[Maj81] Major, P. Multiple Wiener-Ito Integrals, Lecture Notes in Math., 849, Springer 1981.
[Ros79] Rosenblatt, M. Some limit theorems for partial sums of quadratic forms in stationary gaussian variables Z. Wahrsch. verw. Gebiete 49 (1979), pp. 125-132.
[SuHo86] Sun, T.C. & Ho, H.C. On central and non-central limit theorems for nonlinear functions of a stationary Gaussian process, in: Eberlein, E. & Taqqu, M., Eds, Dependence in Probability and Statistics, Birkhäuser, Boston 1986.
5