• Aucun résultat trouvé

A LINEAR FILTER FOR THE OPERATORIAL PREDICTION OF A PERIODICALLY CORRELATED PROCESS

N/A
N/A
Protected

Academic year: 2022

Partager "A LINEAR FILTER FOR THE OPERATORIAL PREDICTION OF A PERIODICALLY CORRELATED PROCESS"

Copied!
15
0
0

Texte intégral

(1)

PREDICTION OF A PERIODICALLY CORRELATED PROCESS

ILIE VALUS¸ESCU

For a periodically Γ-correlated process{fn}in anL(E)-right moduleH, a linear operatorial Wiener filter for prediction is obtained in terms of the coefficients of the maximal outer function of an attached stationary process. Also, an evaluation of the prediction error, which in this case is a periodic function, is given in terms of the same coefficients.

AMS 2000 Subject Classification: 47A20, 47N30, 60G25.

Key words: complete correlated actions, operator model, maximal outer function, linear predictor.

1. PRELIMINARIES

Let E be a separable Hilbert space, L(E) the C-algebra of all linear bounded operators onE and Ha rightL(E)-module. Anaction ofL(E) onH is a map from L(E)× Hinto H given by (A, h)→ hA. For convenience, the action of A on h will be written as Ah, but will be understood as hA, in the sense of the right L(E)-module.

A map Γ fromH × H intoL(E) given by

(1.1) (f, g)→Γ[f, g]

such that

(i) Γ[h, h]≥0, Γ[h, h] = 0 impliesh= 0;

(ii) Γ[g, h] = Γ[h, g];

(iii) Γ[ΣiAihijBjgj] = Σi,jAiΓ[hi, gj]Bj,

for finite sums of actions of L(E) onH, is called a correlation of the action of L(E) on H.

A triple{E,H, Γ}whereΓ has the above properties was called in [5] the correlated action of L(E) on H. For prediction purposes, the Hilbert spaceE will be called the parameter space, and the right L(E)-module H will be the state space. As was seen in [5], to each correlated action {E,H, Γ}, by the

REV. ROUMAINE MATH. PURES APPL.,54(2009),1, 53–67

(2)

positively definite Aronsjain reproducing kernel (1.2) hγλ1, γλ2i= Γ[h2, h1]a1, a2

E, whereλi = (ai, hi)∈ E × H, we can attach a Hilbert space K – the measuring space.

It is easy to see thatL(E,K) is a rightL(E)-module and Γ :L(E,K)× L(E,K)→ L(E)

given by Γ[V1, V2] = V1V2, is a correlation of the action AV := V A of L(E) on L(E,K). This correlated action {E,L(E,K), Γ} was called the operator model. In [6] it was proved that an abstract correlated action {E,H, Γ} can be imbedded into the operator model by considering for each h ∈ H the operatorVh fromL(E,K) given by

(1.3) Vha=γ(a,h).

The imbedding h→Vh is unique up to a unitary equivalence.

The correlated action{E,H, Γ}is said to be acomplete correlated action, if the imbedding h→Vh ofH intoL(E,K) is surjective.

A Γ-stationary process in H is a sequence {ft}t∈G in H, where G is a locally compact group or hypergroup, such that Γ[ft, fs] only depends on the differences−t, not separately onsandt. The function Γ :G→ L(E) given by

(1.4) Γ(t) = Γ[f0, ft]

is called thecorrelation function of the process {ft}t∈G.

Operator valued Γ-stationary processes in complete correlated actions, where G isZorR, were studied in, e.g., [4]–[7], and prediction problems were solved. An appropriate Wiener filter for prediction and an estimation for the prediction error operator, have been obtained.

The previous results can be extended to the matrix valued case with the aim to apply them in the study of periodically Γ-correlated processes.

LetT ≥2 be a positive integer and

(1.5) HT =H × H × · · · × H

the Cartesian product ofT copies of the rightL(E)-moduleH. An elementX of HT will be seen as a line vector (h1, . . . , hT). On HT it is possible to have the action of L(E) on the components with the same operatorA ∈ L(E), or on each component with differentAi ∈ L(E). Also, we can consider the action of L(E)T×T onHT, taking

(1.6) A(h1, . . . , hT) := (h1, . . . , hT)A

for each matrixA= (Aij)Ti,j=1fromL(E)T×T, in the sense of the right module.

It is easy to see thatHT is anL(E)T×T-right module and the action ofL(E) on HT is a particular case of the action ofL(E)T×T onHT for the case of diagonal matrices with the same operator, or different operators on the diagonal.

(3)

Having the action of L(E)T×T on HT, various correlations of this ac- tion can be constructed. For our goal we are interested in two operatorial correlations on HT, namely,

(1.7) Γ1[X, Y] =

T−1

X

k=0

Γ[xk, yk], where X= (x0, x1, . . . xT−1),Y = (y0, y1, . . . , yT−1) , and (1.8) ΓT[X, Y] = Γ[xi, yj]

i,j∈{0,1,...,T−1}.

Using (1.6), it is easy to verify properties (i)–(iii) of a correlation of the action of L(E), respectively ofL(E)T×T, for Γ1 and ΓT.

So, starting with the correlated action {E,H,Γ} of L(E) on H, we ob- tain the correlated actions {E,HT1} and {E,HTT} of L(E), respectively L(E)T×T, on HT. Remark that the correlation Γ1 is the trace of the matrix given by the correlation ΓT.

In [11], a unique (up to a unitary equivalence) imbedding of HT into L(E,K)T was considered defining the operator WX = (Vx1, . . . , VxT) forX = (x1, . . . , xn)∈ HT. So, the correlated actions{E,HT1}and{E,HTT}can be seen in appropriate operator models as{E,L(E,K)T1}and{E,L(E,K)T, ΓT}, where forWi= (Vi1, . . . , ViT) fromL(E,K)T we have

(1.9) Γ1[W1, W2] =

T

X

k=1

Γ[V1k, V2k] =

T

X

k=1

(V1k)V2k, respectively,

(1.10) ΓT[W1, W2] = Γ[V1j, V2k]T

j,k=1= (V1j)V2kT j,k=1.

AnotherL(E)T×T-right module which will be considered in the study of periodically correlated processes will be (HT)T with an appropriate correlation of the action of L(E)T×T.

It is known [6] that for each stationary process there exists an operator valued semispectral measureF on the torusTsuch that its correlation function has the integral form

(1.11) Γ(n) =

Z

0

e−intdF(t).

Also [6] to each semispectral measure F it is attached a maximalL2-bounded outer function Θ(λ) such thatFΘ≤F.

Under a Harnack type domination of the semispectral measure of a sta- tionary process, an operatorial linear Wiener filter was obtained [7] in terms

(4)

of its maximal outer function coefficients. Namely, if there exists a positive constantc such that

(1.12) c·dt≤F ≤c−1·dt,

then the maximal function of the process is a bounded invertible function {E,E,Θ(λ)} with a bounded inverse Ω(λ). As a consequence, the predictiable part ˆfn+1 of fn+1 can be obtained as

(1.13) fˆn+1 =

X

j=0

Ajfn−j,

where the Wiener filter{Aj}for prediction is given by the operator coefficients of Θ(λ) and its inverse Ω(λ) as

(1.14) Aj =

j

X

p=0

Θp+1j−p.

2. PERIODICALLY Γ-CORRELATED PROCESSES

Aperiodically Γ-correlated process is a sequence {ft}t∈G inHwith the property that there exists a positive T such that

(2.1) Γ[ft+T, fs+T] = Γ[ft, fs]

for any t, s∈G. The smallest T which verifies (2.1) is called theperiod of the periodically Γ-correlated process {ft}t∈G.

In this paper the discrete case G =Z is only investigated. Here T is a positive integer. For T = 1 the process is stationary Γ-correlated.

Let us remark that thecorrelation function of a periodically Γ-correlated process {ft}t∈Z with period T ≥2 is the function Γ :Z×Z→ L(E) given by (2.2) Γ(t, s) = Γ[ft, fs].

This is a periodic function with the same period T.

For a periodically Γ-correlated process{ft}we also consider the covari- ance function defined as

(2.3) B(n, t) = Γ(n+t, n).

Conversely, knowing the covariance function of a periodically Γ-correlated process, the corresponding correlation function is obviously given

(2.4) Γ(s, n) =B(n, s−n).

(5)

TheL(E)-valued functionB(n, t) is a periodic function of the first argu- mentn, with the same periodT, and has a Fourier representation of the form

(2.5) B(n, t) =

T−1

X

k=0

Bk(t) exp(2πikn/T),

where Bk(t) are L(E)-valued coefficients fork= 0,1, . . . , T −1. For conveni- ence, Bk(t) can be extended toZon account of the equationBk(t) =Bk+T(t).

A simple computation on (2.5) shows that

(2.6) Bl(t) = 1

T

T−1

X

n=0

B(n, t) exp(−2πinl/T).

The correlation function and the covariance function are positively defi- nite functions. In [11], a generalization of Gladyshev’s theorem [1] to the complete correlated actions case was proved. It can be stated as follows

Theorem 1 (Gladyshev). A functionB(n, t) defined by (2.5) is the co- variance function of some periodically Γ-correlated process in H with period T, if and only if the T ×T matrix valued function

(2.7) B(t) = Bjk(t)T−1

j,k=0

is the operator correlation function of some stationary ΓT-correlated process from HT, where

(2.8) Bjk(t) =Bk−j(t) exp(2πijt/T) and the ΓT-correlation on HT is given by (1.8).

As in the stationary case, prediction spaces can be attached to a perio- dically Γ-correlated process. So, thepast and present of{ft}is the submodule of H given by

(2.9) Hfn=

X

k

Akfk; Ak∈ L(E); k≤n

,

the remote past is

(2.10) Hf−∞=\

n

Hnf

while the space generated by the process is

(2.11) Hf=

X

k

Akfk; Ak∈ L(E)

.

(6)

By means of the imbedding h → Vh of H into L(E,K), the past and present at moment tbecomes a subspace in the measuring spaceK, namely,

(2.12) Kft = _

k≤t

VfkE,

or

(2.13) Kft = _

h∈Hft

VhE,

the remote past

(2.14) K−∞f =\

n

Kfn,

and the space generated by the process

(2.15) Kf=

_

−∞

VftE.

Similar submodules inHT can be attached to a process{Xn} ⊂ HT. We namely have

(2.16) HnX =

X

k

AkXk; Ak∈ L(E)T×T, k≤n

,

(2.17) H−∞X =\

n

HnX

and

(2.18) HX =

X

k

AkXk; Ak ∈ L(E)T×T

.

Using the imbedding X → WX of HT intoL(E,K)T, the corresponding sub- spaces from KT will be

(2.19) KnX = _

k≤n

WXkE,

(2.20) K−∞X =\

n

KnX

and

(2.21) KX =

_

−∞

WXnE.

(7)

3. A LINEAR WIENER FILTER FOR PREDICTION As we have seen, imbedding the right L(E)-module H (the state space) into L(E,K), where K is the measuring space, the study of some processes from H can be done in the operator model L(E,K), which is a right L(E)- module. Also L(E,K)T is a right L(E)T×T-module or, in particular, a right L(E)-module. We have considered ΓT as a correlation of the action ofL(E)T×T on L(E,K)T, and Γ1 as a correlation of the action of L(E) on L(E,K)T. The construction can be continued to

L(E,K)TT

, which becomes a right L(E)T×T-module. In what follows the elements from

L(E,K)TT

will be writ- ten in capital boldface characters, to avoid the confusion with the elements from L(E,K)T which are only designed by capital letters.

An elementZ= (W1, . . . , WT) is a line vector with Wk∈ L(E,K)T, and Wk= (Vk1, Vk2, . . . , VkT) withVkj ∈ L(E,K).

A correlation of the action ofL(E)T×T on

L(E,K)TT

is defined as (3.1) ΓT[Z1,Z2] = Γ1[W1j, W2k]T

j,k=1,

whereW1j = (V1j1, . . . , V1jT) andW2k= (V2k1, . . . , V2kT) are elements ofL(E,K)T. Let{fn}be an arbitrary process inHand Ethe operator of multiplying by e−2πi/T. On account of the construction of the measuring space K and the action of L(E) on H(respectively L(E,K)) as rightL(E)-module, to each {fn}n∈Z fromH we can attachT sequences inHT of the form

(3.2) Xnk= Eknfn, Ek(n+1)fn+1, . . . , Ek(n+T−1)fn+T−1 , where k∈ {0,1, . . . , T−1}.

Using the imbeddingsh →Vh and X →WX of H intoL(E,K), respec- tivelyHT intoL(E,K)T, we obtainT sequences in L(E,K)T by taking

(3.3) Znk=WXk

n, k∈ {0,1, . . . , T −1}, and a sequence in

L(E,K)TT

, by taking

(3.4) Zn= 1

T Zn0, Zn1, . . . , ZnT−1 .

Using Gladyshev’s theorem, in what follows a linear predictor for periodi- cally Γ-correlated processes in complete correlated actions will be obtained, thus generalizing the scalar case [3]. To do this, we will first see that the process attached by (3.4) to a periodically Γ-correlated process {fn} from H is an explicit form of an attached stationary process.

(8)

Theorem 2. A process {fn} from H is periodically Γ-correlated with period T if and only if {Zn}n∈Z from

L(E,K)TT

attached by (3.4) is a sta- tionary ΓT-correlated process.

Proof. If {fn} from H is periodically Γ-correlated with period T, then for each entry of the matrix

ΓT[Zm,Zn] =1

1[Zmj , Znk]T−1 j,k=0

we have

Γ1[Zmj , Znk] = Γ1[WXj

m, WXk n] =

T−1

X

p=0

Γ[Ej(m+p)Vfm+p, Ek(n+p)Vfn+p] =

=

T−1

X

p=0

Vfm+pVfn+pE−j(m+p)+k(n+p)=

=VfmVfnE−jm+kn+

T−1

X

p=1

Vfm+pVfn+pE−j(m+p)+k(n+p)=

= Γ[Vfm, Vfn]E−jm+kn+

T−1

X

p=1

Γ[Vfm+p, Vfn+p]E−j(m+p)+k(n+p) =

= Γ[Vfm+T, Vfn+T]E−j(m+T)+k(n+T)+

T−1

X

p=1

Γ[Vfm+p, Vfn+p]E−j(m+p)+k(n+p)=

=

T

X

p=1

Γ[Vfm+p, Vfn+p]E−j(m+p)+k(n+p)=

=

T−1

X

s=0

Γ[Vfm+s+1, Vfn+s+1]E−j(m+s+1)+k(n+s+1)=

=

T−1

X

s=0

Γ[Ej(m+1+s)Vfm+1+s, Ek(n+1+s)Vfn+1+s] =

= Γ1[WXj

m+1, WXk

n+1] = Γ1[Zm+1j , Zn+1k ].

This implies that ΓT[Zm,Zn] = ΓT[Zm+1,Zn+1], i.e., the process {Zn}n∈Z is stationary ΓT-correlated in

L(E,K)TT

.

Conversely, if {Zn}n∈Z is stationary ΓT-correlated, then each entry of the ΓT-correlation matrix verifies the relation

1

1[Zmj , Znk] = 1

1[Zm+1j , Zn+1k ].

(9)

For j=k= 0 we have

Γ1[WXm0 , WXn0] = Γ1[WX0

m+1, WX0

n+1],

T−1

X

p=0

Γ[Vfm+p, Vfn+p] =

T−1

X

s=0

Γ[Vfm+1+s, Vfn+1+s], hence, by reducing the similar terms,

Γ[Vfm, Vfn] = Γ[Vfm+T, Vfn+T],

i.e., the process {fn}n∈Z is periodically Γ-correlated with periodT. We will see that for a periodically Γ-correlated process{fn}n∈Z, the sta- tionary ΓT-correlated process{Zn}n∈Zdefined by (3.4) verifies the conditions of Gladyshev’s theorem. Indeed, by (2.6), the coefficients of the ΓT-correlation matrix function (2.7) have the form

Bjk(t) = 1

1[Ztj, Z0k] = 1

1[WXj

t, WXk 0] =

= 1 T

T−1

X

p=0

Γ[Vft+p, Vfp]E−j(t+p)+kp = 1 T

T−1

X

p=0

Γ(t+p, p)E−j(t+p)+kp =

=E−jt1 T

T−1

X

p=0

B(p, t)Ep(k−j)=E−jtBk−j(t).

So, Bjk(t) =Bk−j(t) exp(2πijt/T) and condition (2.8) is verified.

According to the definition of thepast and present given in the previous section, the past and present of the process{Zn}n∈Zfrom

L(E,K)TT

will be a subspace from (KT)T given by

(3.5) KnZ= _

k≤n

ZkE, where the elements are of the form

(3.6) Z=X

k≤n

AkZkak,

while the action of Ak ∈ L(E)T×T is understood in the sense of the right L(E)T×T-module

L(E,K)TT

.

Also, for the process{Zn}from

L(E,K)TT

another “past and present” denoted by KZn can be considered in KT, namely, the linear span of the finite sums of the form

(3.7) X

k≤n

AkZkjak, 0≤j≤T −1,

(10)

or

(3.8) KZn = _

k≤n

ZkjE.

Due to the particular form of the stationary ΓT-correlated process at- tached to a periodically Γ-correlated process, the geometry of the past and present spaces can be described as follows.

Theorem 3. The past and present of {Zn}n∈Z has the structure

(3.9) KnZ = KnZT

and

(3.10) KnZ=Kfn× Kfn+1× · · · × Kn+Tf −1.

Proof. For each finite linear combination fromKnZ, in (KT)T we have X

k≤n

AkZkak = 1

√T X

k≤n

Ak(Zk0, Zk1, . . . , ZkT−1)ak =

= 1

√T X

k≤n

AijkT−1

i,j=0(Zk0, Zk1, . . . , ZkT−1)ak=

= 1

√ T

X

k≤n

T−1

X

j=0

ZkjAj0k ak,

T−1

X

j=0

ZkjAj1k ak, . . . ,

T−1

X

j=0

ZkjAj,T−1k ak

.

It follows that each component of the linear combination considered is of the form P

k≤n T−1

P

j=0

AjikZkjak, i.e., it belongs toKZn. Therefore, KnZ⊂(KZn)T. Conversely, each linear combination from (KZn)T has the formZ= (Z0, Z1, . . . , ZT−1), where Zi = P

k≤n T−1

P

j=0

AjikZkjak. It follows that Z = P

k≤n

AkZkak ∈ KnZ. Therefore, (KZn)T ⊂KnZ and (3.9) is proved.

To prove (3.10), let us note that from the definition (3.3) of a component Zmk, wherem≤n, we have

Zmka=WXk

ma= (EkmVfma, Ek(m+1)Vfm+1a, . . . , Ek(m+T−1)Vfm+T−1a) as elements from Kfn× Kfn+1× · · · × Kfn+T−1. Therefore,

KZn ⊂ Kfn× Kfn+1× · · · × Kn+Tf −1 ⊂ KT.

(11)

Conversely, a linear combination of the form Z =

T−1

P

j=0

E−j(m+k)Zmj a, where Zmj are the components of Zm has the form

Z=

T−1

X

j=0

WXj

mE−j(m+k)a=

T−1

X

j=0

Vfm+lEj(m+l−j(m+k)

aT−1 l=0 =

=

T−1

X

j=0

Vfm+lEj(l−k)aT−1 l=0 =

Vf m+la

T−1

X

j=0

Ej(l−k) T−1

l=0

= Vf m+la·T δlkT−1 l=0

for each m≤nand 0≤k≤T−1. Therefore for each 0≤k≤T−1 we have {0} × · · · × {0} × Kfn+k× {0} × · · · × {0} ⊂ KZn,

consequently

Kfn× Kfn+1× · · · × Kn+Tf −1 ⊂ KnZ. This completes the proof.

In what follows, as in [6] we suppose that the L(E)T×T-valued semi- spectral measure attached to the ΓT-correlation function of the process {Zn} satisfies a Harnack type boundedness condition (1.12). Then the invertible maximal matrix function

(3.11) Θ(λ) = Θij(λ)T−1

i,j=0

has a bounded inverse

(3.12) Ω(λ) = Ωij(λ)T−1

i,j=0

and the predictable part of Zn+1 can be obtained as

(3.13) Zˆn+1 =

X

k=0

AkZn−k,

where the Wiener filter

(3.14) Ak= AijkT−1

i,j=0

for prediction is given in terms of the coefficients of its maximal function in a similar way as in (1.14).

Theorem 4. If {fn}n∈Z is a periodically Γ-correlated process and the predictable part of the attached stationary ΓT-correlated process {Zn}n∈Z is given by (3.13) and (3.14), then the predictable part of {fn}n∈Z is

(3.15) fˆn+1=

X

k=0

Ckfn−k,

(12)

where Ck =

T−1

P

j=0

1

TAj0k Ej(n−k).

Proof. Let us consider the predictable part Zˆn+1 = ( ˆZn+10 ,Zˆn+11 , . . . ,Zˆn+1T−1) =PKZ

nZn+1=PKZ

n(Zn+10 , Zn+11 , . . . , Zn+1T−1).

Considering the zero component of ˆZn+1 and using Theorem 3, we get Zˆn+10 =PKZ

nZn+10 =PKZ

n(Vfn+1, Vfn+2, . . . , Vfn+T) =

= (PKf

nVfn+1, . . . , PKf

n+T−1Vfn+T).

On the other hand, using (3.13), we have Zˆn+1 =

X

k=0

AkZn−k =

X

k=0

Aijk 1

T(Zn−k0 , Zn−k1 , . . . , Zn−kT−1) =

= 1

√ T

X

k=0

T−1

X

j=0

Zn−kj Aj0k ,

T−1

X

j=0

Zn−kj Aj1k , . . . ,

T−1

X

j=0

Zn−kj Aj,Tk −1

=

= 1

√ T

X

k=0 T−1

X

j=0

Zn−kj Aj0k , . . . , 1

√ T

X

k=0 T−1

X

j=0

Zn−kj Aj,Tk −1

=

= Zˆn+10 ,Zˆn+11 , . . . ,Zˆn+1T−1 . It follows that

n+10 = 1

√ T

X

k=0 T−1

X

j=0

Zn−kj Aj0k = 1

√ T

X

k=0 T−1

X

j=0

Aj0k Zn−kj =

=

X

k=0 T−1

X

j=0

√1

TAj0k WXj

n−k =

X

k=0 T−1

X

j=0

√1

TAj0k Ej(n−k+s)Vfn−k+s

T−1 s=0 =

=

X

k=0 T−1

X

j=0

√1

TAj0k Ej(n−k+s)Vfn−k+s T−1

s=0

.

Therefore, VP

Hf

nfn+1 =PKf

nVfn+1 =

X

k=0 T−1

X

j=0

√1

TAj0k Ej(n−k)Vfn−k, where PHf

n is the “Γ-orthogonal projection” on the submoduleHfn of H.

So, the corresponding operatorial Wiener filter for the prediction of a periodically Γ-correlated process from His given by (3.15).

(13)

Remark, that in the periodic case the prediction error (3.16) ∆(n) = Γ[fn+1−fˆn+1, fn+1−fˆn+1]

is a periodic function, not an operator as in the stationary case. Therefore, we have

(3.17) ∆(n) =

T−1

X

k=0

kexp(2πijk/T) and, conversely, the coefficients ∆k can be obtained as

(3.18) ∆k= 1

T

T−1

X

j=0

∆(j) exp(−2πijk/T).

The next theorem gives a characterization of the prediction error for a periodically Γ-correlated process {fn} in terms of the coefficients of the maximal function of the attached ΓT-correlated process {Zn}.

Theorem 5. The prediction error ∆(n) of a periodically Γ-correlated process {fn} has the form

(3.19) ∆(n) =

T−1

X

k=0

DkE−k(n+1),

where the operator coefficients Dk ∈ L(E) are the entries of the zero line of the prediction error matrix of the attached stationary process {Zn}, namely,

(3.20) Dk=

T−1

X

s=0

Θs0Θsk,

where Θij = Θij(0) from the maximal function (3.11) of the process {Zn}.

Proof.Let ∆ be the prediction error matrix of the stationary ΓT-correlated process {Zn}attached to {fn}by (3.4). Then, for each n∈Z,

∆ = ΓT[Zn+1−Zˆn+1,Zn+1−Zˆn+1] = ∆ijT−1 i,j=0, where the operators ∆ij are given by

ij = 1

1[Zn+1i −Zˆn+1i , Zn+1j −Zˆn+1j ].

It is known [6] that if Θ(λ) is the maximal function of a stationary process, then the prediction error is given by

(3.21) ∆ = Θ(0)Θ(0).

(14)

From (3.11), putting Θij = Θij(0), we get

(3.22) ∆ij =

T−1

X

s=0

ΘsiΘsj. On the other hand, from (3.18) we have

k= 1 T

T−1

X

j=0

∆(j)Ekj = 1 T

T−1

X

j=0

Γ[fj+1−fˆj+1, fj+1−fˆj+1]Ekj =

= 1 T

T−1

X

j=0

V

fj+1fˆj+1Vf

j+1fˆj+1Ekj = 1 TE−k

T−1

X

j=0

V

fj+1fˆj+1Vf

j+1fˆj+1Ek(j+1)=

= 1 TE−k

T−1

X

j=0

Γ[Vf

j+1fˆj+1, Ek(j+1)Vf

j+1fˆj+1] =

=E−k1

1[Z10−Zˆ10, Z1k−Zˆ1k] =E−k0k=E−k

T−1

X

s=0

Θs0Θsk. It follows from (3.17) that

∆(n) =

T−1

X

k=0

E−k

T−1

X

s=0

Θs0ΘskE−kn=

T−1

X

k=0

DkE−k(n+1),

where Dk =

T−1

P

s=0

Θs0Θsk, and the proof is complete.

Acknowledgements.This work was partly supported by Grant ANCS 2-CEx06-11- 34/2006.

REFERENCES

[1] E.G. Gladyshev,Periodically correlated random sequences. Soviet. Math. Dokl.2(1961), 385–388.

[2] P. Masani, Recent trends in multivariate prediction theory. In: Multivariate Analysis (Proc Internat. Sympos., Dayton, Ohio,1965), pp. 351–382. Academic Press, New York, 1966.

[3] A.G. Miamee, Explicit formula for the best linear predictor of periodically correlated sequences. SIAM J. Math. Anal.24(1993),3, 703–711.

[4] I. Suciu and I. Valu¸sescu,Factorization of semispectral measures. Rev. Roumaine Math.

Pures Appl.21(1976),6, 773–793.

[5] I. Suciu and I. Valu¸sescu, Essential parameters in prediction. Rev. Roumaine Math.

Pures Appl.22(1977),10, 1477–1495.

(15)

[6] I. Suciu and I. Valu¸sescu,Factorization theorems and prediction theory. Rev. Roumaine Math. Pures Appl.23(1978),9, 1393–1423.

[7] I. Suciu and I. Valu¸sescu, A linear filtering problem in complete correlated actions. J.

Multivariate Anal.9(1979),4, 559–613.

[8] B. Sz.-Nagy and C. Foias, Harmonic Analysis of Operators on Hilbert Space. Acad.

Kiado, Budapest & North Holland, Amsterdam, 1970.

[9] I. Valu¸sescu,The maximal function of a contraction. Acta Sci. Math. (Szeged)42(1980), 1–2, 183–188.

[10] I. Valu¸sescu,Operator methods in prediction theory, I & II. Stud. Cerc. Mat.33(1981), 3, 343–401 &33(1981),4, 467–492. (Romanian)

[11] I. Valu¸sescu, On nonstationary periodically correlated processes in complete correlated actions. Proc. ICTAMI-2007, Acta Univ. Apulensis Math. Inform. No. 15 (2008), 353–372.

[12] N. Wiener and P. Masani,The prediction theory of multivariate stochastic processes, I

& II. Acta Math.98(1957), 111–150 &99(1958), 93–139.

Received 2 May 2008 Romanian Academy

“Simion Stoilow” Institute of Mathematics P.O. Box 1-764

010702 Bucharest, Romania Ilie.Valusescu@imar.ro

Références

Documents relatifs

A self-administered dietary history questionnaire, especially developed for use in a large French prospective cohort study, was tested for accuracy of food intake measurement

Taraina a calculé que l'âge moyen de ses élèves est légèrement supérieur à 14 ans, or pour inscrire son groupe au Heiva dans la catégorie « Adolescents », l'âge moyen du

For the sake of comparison, we also implemented, in addition to the proposed method, another 5 techniques of multispectral demosaicking including bilinear interpolation [18],

In the impact limit an important simplification occurs, since the time of flight during the impact is taken to be zero. This allows more analysis to be done on periodic orbits. fjx,

method for solving the shell-model equations in terms of a basis which includes correlated sub-.. systems is

Cédric Duée, Marie Hénault, Thomas Lecas, Laifa Boufendi, Xavier Bourrat, Hubert Haas, Henry Pillière.. To cite

The results provided by the principal components analysis are satisfactory for the 2 classes defined by K-Means clustering, and make it possible to reduce significantly the data

The asymptotic normality of B b (s, τ ) for strictly periodic time series under ϕ-mixing assumption was shown in Hurd and Miamee (2007) (see Preposition 9.13).. , 24) (black)