04.01.2013 Views

Springer - Read

Springer - Read

Springer - Read

SHOW MORE
SHOW LESS

Create successful ePaper yourself

Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.

66 Chapter 2 Stationary Processes<br />

where Xn (Xn,...,X1) ′ and<br />

⎡<br />

⎢<br />

⎣<br />

1 φ φ 2<br />

··· φ n−1<br />

φ 1 φ ··· φ n−2<br />

. . . . .<br />

φ n−1 φ n−2 φ n−3<br />

⎤⎡<br />

⎤<br />

a1<br />

⎥⎢<br />

⎥<br />

⎥⎢<br />

a2 ⎥<br />

⎥⎢<br />

⎥<br />

⎥⎢<br />

⎥<br />

⎦⎣<br />

. ⎦<br />

··· 1<br />

<br />

⎡<br />

φ<br />

⎢ φ<br />

⎢<br />

⎣<br />

2<br />

.<br />

φ n<br />

⎤<br />

⎥<br />

⎥.<br />

⎦<br />

(2.5.11)<br />

A solution of (2.5.11) is clearly<br />

an (φ, 0,...,0) ′ ,<br />

and hence the best linear predictor of Xn+1 in terms of {X1,...,Xn} is<br />

PnXn+1 a ′<br />

n Xn φXn,<br />

with mean squared error<br />

an<br />

E(Xn+1 − PnXn+1) 2 γ(0) − a ′<br />

n γn(1) <br />

σ 2<br />

1 − φ 2 − φγ(1) σ 2 .<br />

A simpler approach to this problem is to guess, by inspection of the equation defining<br />

Xn+1, that the best predictor is φXn. Then to verify this conjecture, it suffices to check<br />

(2.5.10) for each of the predictor variables 1,Xn,...,X1. The prediction error of the<br />

predictor φXn is clearly Xn+1 − φXn Zn+1. But E(Zn+1Y) 0 for Y 1 and for<br />

Y Xj,j 1,...,n. Hence, by (2.5.10), φXn is the required best linear predictor<br />

in terms of 1,X1,...,Xn.<br />

Prediction of Second-Order Random Variables<br />

Suppose now that Y and Wn, ..., W1 are any random variables with finite second<br />

moments and that the means µ EY, µi EWi and covariances Cov(Y, Y ),<br />

Cov(Y, Wi), and Cov(Wi,Wj) are all known. It is convenient to introduce the random<br />

vector W (Wn,...,W1) ′ , the corresponding vector of means µW (µn,...,µ1) ′ ,<br />

the vector of covariances<br />

γ Cov(Y, W) (Cov(Y, Wn), Cov(Y, Wn−1),...,Cov(Y, W1)) ′ ,<br />

and the covariance matrix<br />

Ɣ Cov(W, W) Cov(Wn+1−i,Wn+1−j) n<br />

i,j1 .<br />

Then by the same arguments used in the calculation of PnXn+h, the best linear predictor<br />

of Y in terms of {1,Wn,...,W1} is found to be<br />

P(Y|W) µY + a ′ (W − µW ), (2.5.12)<br />

where a (a1,...,an) ′ is any solution of<br />

Ɣa γ. (2.5.13)

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!