• No results found

SHIFT OPERATOR FOR PERIODICALLY CORRELATED PROCESSES

N/A
N/A
Protected

Academic year: 2020

Share "SHIFT OPERATOR FOR PERIODICALLY CORRELATED PROCESSES"

Copied!
6
0
0

Loading.... (view fulltext now)

Full text

(1)

SHIFT OPERATOR FOR PERIODICALLY

CORRELATED PROCESSES

A.G. Miamee

1,*

and G.H. Shahkar

2

1 Department of Mathematics, Hampton University, Hampton, VA, 23668, USA

2 Faculty of Mathematical Sciences, Ferdowsi University of Mashhad, Mashhad, Islamic Republic of Iran

Abstract

The existence of shift for periodically correlated processes and its boundedness

are investigated. Spectral criteria for these non-stationary processes to have such

shifts are obtained.

Keywords: Periodically correlated; Shift; Stationary; Linearly stationary; Bounded linearly stationary

AMS (1991) Subject Classification: 60G10, 60G12, 60G25

* E-mail: [email protected]

1. Introduction

Prediction theory of stationary stochastic processes has been extensively developed and is now considered to be complete. The existence of bounded shift for stationary processes has played a major role in this development. The existence and boundedness of shift for non-stationary processes is important [1]. An interesting class of non-stationary stochastic processes is that of periodically correlated processes. This class of processes has been studied by several authors [2-14]. However, questions concerning their shift have not yet been considered. In this note, we study these questions and obtain spectral criteria for the existence of bounded shift for PC processes.

) PC (

2. Preliminaries

Let (Ω,β,Ρ) be a probability space and 2( , , )

0 Ωβ Ρ

L

denote the space of all complex-valued random variables on Ω with zero mean and finite variance. The

inner product and norm here are given by

Ω Ρ

=

= ( ) ( ) ( ) ( ) )

,

(X Y E XY X ωY ω d ω

and X = (X,X).

Any sequence

{

Xn,nZ

}

of random variables in

) , , (

2 0 Ω β Ρ

L will be called a stochastic process and its correlation function R(m,n) is defined by

) ( ) ,

(mn E XmXn

R = .

Given a stochastic process , its shift operator V is a linear transformation which sends to , for each

n

X

n

X Xn+1

Z

(2)

to impose certain restrictions on . Before we proceed further, let’s now state the formal definition of shift operator and consider an example where this operator is not well-defined.

n

X

Definition 2.1. (a) A stochastic process is said to have a shift if the linear transformation V on

which sends each to is well defined. (b) A process is said to have a bounded shift if it has a shift which can be extended to

n

X

{

X n Z

X

L( )= n: ∈

}

Xn Xn+1 n

X

{

X n Z

}

SP X

H( )= n: ∈

as a bounded operator.

Example. Let be any nonzero stochastic process and define a new stochastic process by

n

Y

n

X

⎪⎩ ⎪ ⎨ ⎧

+ =

= =

1 2 0

2

k n if

k n if Y

Xn k

The shift operator for is clearly ill-defined because it sends a zero vector, say to a nonzero vector, say . If we take the original stochastic process to be a nondeterministic, then we get an example of a stochastic process which has no shift.

n

X

1

X

2

X

n

Y

n

X

Definition 2.2. A stochastic process is called stationary if

n

X

) , ( ) ,

(m n =R m+1n+1

R

For all m,nZ.

It is well-known that any stationary stochastic pro-cess has a bounded shift and that it is a unitary operator. However, for a non-stationary process as we saw above the shift may not even exist and in order for a stochastic process to have a shift we must impose some restrictions on the process or its correlation function

. For the following lemma one can see [1].

n

X

n

X

) , (mn

R

Lemma 2.3. Let be a stochastic process with correlation function as defined above. Then

n

X

) , (mn

R

(a) In order for to have a shift it is necessary and

sufficient that for any finite sequence

{

of complex numbers

n

X

}

n

a

0 ) 1 , 1 ( 0

) ,

( = ⇒

+ + =

amanR mn amanR m n .

(b) In order for to have a bounded shift it is necessary and sufficient to have a positive number

n

X

M

such that for any finite sequence

{

of complex numbers

}

n

a

amanR(m+1,n+1)≤M amanR(m,n).

We close this section with a brief introduction to periodically correlated processes.

Definition 2.4. A stochastic process is called periodically correlated with period p if for all

n

X

Z n m, ∈ , we have

) , ( ) ,

(mn R m p n p

R = + + .

Such a process will be briefly called a process. Let be a process with period . Then for each integer

PC

n

X PC p

τ , the function R(n,n+τ) is periodic in with period . Therefore it has Fourier expansion

n p

− =

= +

1

0

) 2 exp( ) ( )

, (

p

k k

p ikn R

n n

R τ τ π ,

where Rk(τ) are given by

− =

− +

= 1

0

) 2 exp( ) , ( 1 ) (

p

n

k p R nn in

R τ τ

π

τ

.

For convenience, we extend the definition of these

1 1

, 0 , )

( k= , ,p

Rk τ K to all integers by k Rk(τ)

) (τ

p k R +

= .

It is shown in [3] that each Rk(τ) has a spectral representation of the form

= π θ

τ τθ

π

2

0 ( )

2 1 )

( i k

k e dF

(3)

where each dFk is a complex-valued measure on

) 2 , 0

[ π . One can then see that

∫ ∫

− −

= π π θ λ θ λ

π

2

0 2

0

) (

2 ( , )

4 1 ) ,

(m n e dF

R im n

where the spectral measure dF of Xn is given by

− − =

− =

1

1

)) 2 ( ( )

, (

p

p k

k p

k B A F B

A

F I π .

Here stands for the set of all numbers of the form with . This shows that spectral measure of any process is concentrated on

line segments

a B

a

bbB dF PC 1

2p − θ − λ = 2πk/ p, k =

, contained in the square 1

, ,

1−pK p− [0,2π)×[0,2π), and

− −

= −

= π θ

π θ

2

0

)

( ( )

2 1 ) ( ) ,

(m n r m n e dF

R im n

3. Shift for PC Processes

In this section we study the questions of existence and boundedness of shift for processes. We first obtain some basic results and then prove our main result which gives several criteria for existence and boundedness of shift including a spectral criterion.

PC

Lemma 3.1. Let Xn be a PC process with period p. (a) If Xn has shift V then V is invertible.

(b) If has bounded shift V then V is boundedly invertible.

n

X

Proof. (a) Suppose has shift V and assume for some finite sequence

n

X

0

=

anXn

{ }

an of

complex numbers. This means that

0 ) , ( =

amanR m n . Applying Lemma 2.3(a) we get

0 ) , ( +1 +1 =

amanR m n . Applying Lemma 2.3(a)

more times to the latter equation, we get 2

p

0 ) 1 ,

1

( + − + − =

amanR m p n p .

Considering that Xn is PC with period p, we get

0 ) 1 , 1 ( − − =

amanR m n ,

which means

anXn1=0. So we showed that for

any sequence of complex numbers an

0 0 ⇒ 1=

=

anXn anXn− .

But this is clearly equivalent to the existence of the inverse U of V which sends each Xn to Xn−1.

(b) Suppose has a bounded shift V . and let be a finite sequence of complex numbers. Applying Lemma 2.3(b)

n

X an

1

p consecutive times we arrive at

, ) , (

) 1 ,

1 (

1

− ≤

− + − +

n m R a a M

p n p m R a a

n m p n m

which in conjunction with the fact that is periodically correlated with period implies

n

X p

a a R(m1,n1)M −1 a a R(m,n) n m p n

m .

Therefore

2 1

2

1

− ≤ p n n n

nX M a X

a ,

which implies that backward shift U sending each to has a bounded extension to . Since

clearly , we conclude that V is boundedly invertible.

n

X

1

n

X H(X)

1 −

=V U

The following remarks follows from the proof of Lemma 3.1.

Remarks 3.2. Let be any process with period .

n

X PC

p

(a) If has a shift (bounded shift) V , then it has shifts (bounded shifts) , of any order , sending

each to . In fact, it is clear that .

n

X

k

V kZ

n

(4)

(b) The shift Vp always exists and it is unitary.

(c) If has a bounded shift V , then for any positive integer ,

n

X k

k

k V

V ≤ and VkV k(p−1).

Before we proceed further, we need to introduce some terminologies. Let be Hilbert

spaces. Their direct sum , equipped

with the Euclidean inner product

q

H H

H1, 2,K, q

q

H H

H1⊕ 2⊕K⊕

=

=

q

i

i i Y X 1

) , ( )) , ((XY

becomes a Hilbert space. Here X=X1⊕X2⊕K⊕Xq

and , with . Direct

sum of q copies of H will be denoted by .

q

Y Y

Y ⊕ ⊕ ⊕

= 1 2 K

Y Xi,YiHi

q

H

Lemma 3.3. A stochastic process Xn in H =

is periodically correlated with period if

and only ifits associated process in defined by

) , , (

2

0 Ω β Ρ

L p

n

Z Hq

1

1⊕ ⊕ + −

=Xn Xn+ Xn p

n K

Z

is stationary.

Proof. “only if” part: If is a process with period , then for any and n in

n

X PC

p m Z, we can write

=

+ +

=

+ + +

+

=

+ + −

=

+ +

=

+ =

+ =

=

p

i

i n i m

p

i

i n i m p

n p m

p

i

i n i m n

m p

i

i n i m n

m

X X

X X X

X

X X X

X

X X

1

1

1 1

1 1

0

) , (

) , ( ) , (

) , ( ) , (

) , ( )) , ((Z Z

. )) , ((

) , (

1 1 1

0

1 1

+ + −

=

+ + + +

= =

n m p

i

i n i

m X

X

Z Z

This means that is stationary. Proof of “if” part is similar.

m

Z

Lemma 3.4. If is a process with period then the spectral measure of its associated stationary process introduced in last lemma is , where is the part of the spectral measure of supported on the main diagonal of the square mentioned in section 2.

n

X PC p

n

Z pdFo dFo dF Xn

Proof. For the proof, we refer the reader to [7].

Definition 3.5. A stochastic process in the Hilbert

space is called linearly stationary if there exists a stationary process in another Hilbert space

n

X

) , , (

2

0 Ω Ρ

=

Η L β

n

W

κ and an invertible transformation T:κ→H

such that Xn =TWn, for all . A linearly stationary process is called bounded linearly stationary if the transformation can be chosen to be bounded.

Z n

n

X

T

It is clear that linearly stationary processes are in general non-stationary. Nevertheless prediction properties of linearly stationary processes can easily be investigated. Because one can transfer a prediction problem concerning a linearly stationary process to one about its stationary counterpart , we find the solution for this stationary process and then transfer the result back to the original process For more detail, one can see [15].

n

X

n

W

n

W

n

X

In what follows, we will use the following notations and terminologies.

Let be a process with period p in

and in , be its associated stationary process introduced in Lemma 3.3, namely:

n

X PC

) , , (

2

0 Ω Ρ

=L β

H Zn H p

1

1⊕ ⊕ + −

= Xn Xn+ Xn p

n K

Z

Now let denote the orthogonal projection which maps any vector in to its first coordinate, i.e.

p p P:HH

p

(5)

1 2

1 )

(X X X X

P ⊕ ⊕Kp = ,

for any X1,X2,K,XpH . We denote by

κ

the subspace of spanned by all and Q to stand for the restriction of

p

H Zn's

P to κ.

Theorem 3.6. For any process with period , the following statements are equivalent.

PC Xn p

(a) Xn has a bounded shift V . (b) Xn is bounded linearly stationary.

(c) The operator Q:κ →H defined above is boundedly invertible.

Proof. We prove (a) ⇒ (b) ⇒ (c).

(c)

(a) ⇒ : Take a finite linear combination . We can write:

anXn

. )

1

( 2 2( 1) 2

2 1

2 2

2 1

2 1 2

2

− −

− +

+

+ + + ≤

+ +

+ =

+ +

+ =

n n p

n n p

n n n

n

p n n

n n n

n n

n

X a V

V

X a V

X a V X a

X a

X a X

a a

K K

K

Z

which shows the inverse of Q exists and is bounded.

) (b

(c) ⇒ : By Lemma 3.3, where is the stationary process associated to . Since each is clearly in

n

n Ρ

X = Z Zn

n

X Zn

κ and Q is the restriction of Ρ to

κ

then we get for all and this completes the proof.

n

n Q

X = Z nZ

) (a

(b) ⇒ : Suppose there is a stationary process and boundedly invertible operator

such that

n

W T:H(W) )

(X

H

) ( n

n T W

X = for all nZ .

Let be the well-known unitary shift of the stationary process and define by

U

n

W V:H(X)→H(X)

1

=TUT

V .

One can check that V is the shift of our process . Now since T and U are bounded is bounded.

n

X

1

=TUT V

Theorem 3.7. Let be a process with period . The following statements are equivalent.

n

X PC p

(a) Xn has a shift.

(b) Xn is linearly stationary.

(c) The operator Q:κ → H defined above is invertible.

Proof. : Suppose has a shift, say V and suppose a finite linear combination of is zero, i.e.

(c)

(a) ⇒ Xn

n

X

0

=

anXn . Applying V to both sides of this

equation p−1 times, we get

. 0 ,

,

0 1

1=

=

anXn+ K anXn+p

Thus

0

1

0

2 2

= =

∑ ∑

=

+

p

i

i n n n

n a X

a Z

which means

anZn =0. Hence Q is invertible.

(b)

(c) ⇒ : By Lemma 3.3, where is the stationary process associated to . Since each is clearly in

n

n Ρ

X = Z Zn

n

X Zn

κ and Q is the restriction of Ρ to κ then we get Xn =QZn for all and this completes the proof.

Z n

) (a

(b) ⇒ : Let be the stationary process and be the linear transformation with

n

W

) ( ) (

:LW L X

T

n

n TW

X = and be the standard unitary shift operator of the stationary process , then

the linear transformation clearly serves the desirable shift for .

) ( ) (

:H W H W

U

n

W

1

=TUT V n

X

Next theorem gives our spectral characterization for a PC process to have a shift.

Theorem 3.8. Let be a process with period whose spectral measure is concentrated on

n

X PC p

) (⋅

(6)

line segments θ−λ=2πk/p, k=1−p,K,0,K,p−1

of

[

0,2π

) (

× 2π,0

]

with the measure on the diagonal being dF0.

(a) has a bounded shift if and only if there exists a positive number

n

X

K such that

∫ ∫

πφ θ θ ≤ 2π π φθ φ λ θ λ

0 2

0 2

0 0

2

) , ( ) ( ) ( )

( )

( dF K dF

for any trigonometric polynomial φ(θ)=

aneinθ . (b) Xn has a shift if and only if

, 0 )

(

0 ) , ( ) ( ) (

2

0 0

2 2

0 2

0

= ⇒

=

∫ ∫

π π π

θ φ

λ θ λ φ θ φ

dF dF

for any trigonometric polynomial function φ(θ)=

.

inθ

ne

a

Proof.(a) If has a bounded shift then by Theorem 3.6, the operator is boundedly invertible. This means there exists some such that

n

X

Q

0

>

M

anZnM anXn

for every finite sequence of complex numbers. Squaring both sides and rewriting it in terms of the spectral measure we get

n

a

∫ ∫

πφ θ ≤ 2π πφθ φ λ θ λ

0 2

0 2 2

0 0

2

) , ( ) ( ) ( )

( dF M dF

p

where . This shows that (1) holds

with . Now assume that inequality (1) holds. We can rewrite it as

= θ

θ

φ in ne

a

) (

p M K = 2/

2 2

) /

(l p

anZnK

anXn ,

which means the operator in part (c) of Theorem 3.6 is boundedly invertible. This in virtue of Theorem 3.6 completes the proof of part (a).

Q

Proof of part (b) is similar to the proof of part (a).

References

1. Getoor R.K. The shift operator for non-stationary stochastic processes. Dufe Math J., 75-187 (1956). 2. Gardner W.A. Introduction to Random Processes with

Applications to Signals and Systems. Macmillan, New York (1985), 2nd Ed. McGraw-Hill (1989).

3. Gladyshev E.G. Periodically correlated random sequences. Soviet Math Dokl., 385-388 (1961).

4. Hurd H.L. An investigation of periodically correlated stochastic processes. Ph.D. Dissertation, Duke University, Durham, North Carolina (1969).

5. Miamee A.G. and Salehi H. On the prediction of periodically correlated stochastic processes. In: P.R. Krishnaiah (Ed.), Multivariate Analysis-V. North Holland, Amsterdam, 167-179 (1980).

6. Pourahmadi M. and Salehi H. it On subordination and linear transformation of harmonizable and periodically correlated processes. Lecture Notes in Math. 1080, Springer-Verlag, Berlin, New York, 195-213 (1984). 7. Miamee A.G. Periodically correlated processes and their

stationary dilations. SIAM J. Appl. Math., 1194-1199 (1990).

8. Miamee A.G. Explicit formulas for the best linear predictor of periodically correlated sequences. SIAM J.

Math. Anal., 703-711 (1993).

9. Gladyshev E.G. Periodically and almost periodically correlated random processes with continuous time parameter. Theory Prob. Appl., 173-177 (1963).

10. Hurd H.L. Stationarizing properties of random shifts.

SIAM J. Appl. Math., 203-211 (1974).

11. Hurd H.L. Representation of strongly harmonizable periodically correlated processes and their covariances. J.

Multiv. Anal., 53-67 (1989).

12. Makagon A., Miamee A.G. and Salehi H. Periodically correlated processes and their spectrum. In: Miamee A.G. (Ed.), Nonstationary Stochastic Processes and Their Application. World Scientific, Singapore, 147-164 (1992).

13. Makagon A., Miamee A.G. and Salehi H. Continuous time periodically correlated processes : spectrum and predictions. Stochastic Process and Their Applications, 277-295 (1994).

14. Hurd H.L. and Kallianpur G. Periodically correlated processes and their relationship to L2[0,T) valued stationary sequence. In: Miamee A.G. (Ed.), Nonstationary Stochastic Processes and Their Application. World Scientific, Singapore, 256-284 (1992).

References

Related documents