06.06.2013 Views

Theory of Statistics - George Mason University

Theory of Statistics - George Mason University

Theory of Statistics - George Mason University

SHOW MORE
SHOW LESS

Create successful ePaper yourself

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

792 0 Statistical Mathematics<br />

• Z T Z is nonnegative definite, and positive definite if and only if Z is <strong>of</strong> full<br />

column rank;<br />

• Z T Z = 0 =⇒ Z = 0.<br />

The generalized inverses <strong>of</strong> Z T Z have useful properties. First, we see from<br />

the definition, for any generalized inverse, (Z T Z) − that ((Z T Z) − ) T is also a<br />

generalized inverse <strong>of</strong> Z T Z. (Note that (Z T Z) − is not necessarily symmetric.)<br />

Another useful property <strong>of</strong> a Gramian matrix is that for any matrices B<br />

and C (that are conformable for the operations indicated),<br />

BZ T Z = CZ T Z ⇐⇒ BZ T = CZ T .<br />

The implication from right to left is obvious, and we can see the left to right<br />

implication by writing<br />

(BZ T Z − CZ T Z)(B T − C T ) = (BZ T − CZ T )(BZ T − CZ T ) T ,<br />

and then observing that if the left side is null, then so is the right side, and if<br />

the right side is null, then BZ T − CZ T = 0. Similarly, we have<br />

Also,<br />

Z T ZB = Z T ZC ⇐⇒ Z T B = Z T C.<br />

Z(Z T Z) − Z T Z = Z.<br />

This means that (Z T Z) − Z T is a generalized inverse <strong>of</strong> Z<br />

An important property <strong>of</strong> Z(Z T Z) − Z T is its invariance to the choice <strong>of</strong><br />

the generalized inverse <strong>of</strong> Z T Z. Suppose G is any generalized inverse <strong>of</strong> Z T Z.<br />

Then we have<br />

ZGZ T = Z(Z T Z) − Z T ;<br />

that is, Z(Z T Z) − Z T is invariant to choice <strong>of</strong> generalized inverse.<br />

The squared norm <strong>of</strong> the residual vector obtained from any generalized<br />

inverse <strong>of</strong> Z T Z has some interesting properties. First, just by direct multiplication,<br />

we get the “Pythagorean property” <strong>of</strong> the norm <strong>of</strong> the predicted<br />

values and the residuals:<br />

X − Zβ 2 = X − Z β 2 + Z β − Zβ 2<br />

where β = (Z T Z) − Z T X for any generalized inverse. We also have<br />

and<br />

E(Z β) = Zβ,<br />

E((Z β − Zβ) T (Z β − Zβ)) = V(Z β).<br />

Because for any vector y, we have<br />

y 2 = y T y = tr(y T y),<br />

<strong>Theory</strong> <strong>of</strong> <strong>Statistics</strong> c○2000–2013 James E. Gentle

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

Saved successfully!

Ooh no, something went wrong!