Econometric Theory and Methods

Answers to Starred Exercises

118

Solution to Exercise 12.16

12.16 The k k matrix X ( 1 PW )X given in expression (12.66) is positive semidenite by construction. Show this property explicitly by expressing the matrix in the form A A, where A is a matrix with k columns and at least k rows that should depend on a g g nonsingular matrix which satises the relation = 1 . Show that a positive semidenite matrix expressed in the form A A is positive denite if and only if A has full column rank. In the present case, the matrix A fails to have full column rank if and only if there exists a k --vector , dierent g from zero, such that A = 0. Since k = i=1 ki , we may write the vector . . . . as [1 . . . . g ], where i is a ki --vector for i = 1, . . . , g . Show that there exists a nonzero such that A = 0 if and only if, for at least one i, there is a nonzero i such that PW Xi i = 0, that is, if PW Xi does not have full column rank. Show that, if PW Xi has full column rank, then there exists a unique solution of the estimating equations (12.60) for the parameters i of equation i.

If we set A ( PW )X , then we see that A A = X ( PW )( PW )X = X ( PW )X = X ( 1 PW )X . In these manipulations, we make use of the fact that the orthogonal projection matrix PW is symmetric and idempotent. Since PW is gn gn, and X is gn k , it follows that A is a gn k matrix. The requirement on the number of rows is satised because gn > k . For the second part of the question, the positive semidenite matrix is positive denite if and only if A A = 0 implies that = 0. A A = A 2, and so the quadratic form is zero if and only if A that is, if and only if A = 0. If this last relation implies that = 0, by denition A has full column rank. The matrix A can be expressed explicitly as follows: A = ( PW )X 11 PW g1 PW X1 . . . .. .. . . = . . . . . . O 1g PW gg PW 11 PW X1 g1 PW Xg . . .. . . = , . . . 1g PW X1 gg PW Xg AA But = 0, then

O . . . Xg

Copyright c 2003, Russell Davidson and James G. MacKinnon

Econometric Theory and Methods

Answers to Starred Exercises

119

where ij is the ij th element of . We can postmultiply this by a k --vector . . . . that is partitioned as [1 . . . . g ]. The result is 11 PW X1 . . A = . 1g PW X1 .. . g1 PW Xg 1 . = . . . . . gg PW Xg g j 1 PW Xj j . . . . g j =1 jg PW Xj j
g j =1

If we stack the n 1 blocks of the gn 1 vector A horizontally in an n g matrix rather than vertically, we obtain [
g j =1

j 1 PW Xj j

= [ PW X1 1 = [ PW X1 1

jg PW Xj j ] 11 1g . . .. . PW Xg g ] . . . . g1 gg PW Xg g ] .

g j =1

Clearly, the vector A is zero if and only if the matrix in the last line above is zero. But is a nonsingular g g matrix, and so A is zero for arbitrary nonzero if and only if the entire matrix [PW X1 1 PW Xg g ] is zero. But that can only be the case if PW Xi i is zero for all i = 1, . . . , g . Consequently, A is zero with a nonzero if and only if there is at least one i such that PW Xi i = 0 with nonzero i . We now turn to the third part of the question. If X ( 1 PW )X is nonsingular, then the equations (12.60) have a unique solution for , and the result is trivial. The only case that needs further study is therefore the one in which the matrix A does not have full column rank. Suppose then that A is of rank r < k . Then A can be partitioned, possibly after a reordering of its columns, as A = [A1 A1 B ], where A1 is gn r with full column rank, and B is r (k r). This result simply makes explicit the fact that k r columns of A are linear combinations of the other r columns. Two points need to be established, namely, existence and uniqueness. For existence, observe that equations (12.60) can be written as A A ( PW )y = 0. (S12.22)

1 . 2 1 . . Partition as [ ], conformably with the partition of A, so that has . 2 1 2 r elements, and has k r elements. Thus A = A1 ( + B ). Now set 2 = 0. Then we can show that equations (S12.22) have a unique solution 1 for . Indeed, equation (S12.22) becomes

A1 B A1

1 A1 ( PW )y = 0.

(S12.23)

Copyright c 2003, Russell Davidson and James G. MacKinnon

Econometric Theory and Methods Note that if

Answers to Starred Exercises

120

1 ( PW )y = 0, A1 A1

(S12.24)

then (S12.23) is true, since the last k r rows are just linear combinations of the rst r rows. But A1 has full column rank of r, and so the r r matrix A1 A1 is nonsingular. Thus the equations (S12.24) have a unique solution 1 1 , and by the for the r--vector , as claimed. Denote this solution by . 1 . . k --vector [ . 0]. In order to show uniqueness, observe that any other solution to equation , is such that (S12.22), say ) = 0. A A( (S12.25)

from the same equation This follows by subtracting (S12.22) evaluated at evaluated at . If we write , then, by an argument used earlier, (S12.25) implies that A = 0, and, by the second part of this exercise, this implies that PW Xi i = 0 for all i = 1, . . . , g , where i is dened as above as the i th block of . If, for some i, PW Xi has full column rank, then it follows and have the same i th block, which that i = 0. Thus the two solutions is therefore dened uniquely, as we wished to show.

Copyright c 2003, Russell Davidson and James G. MacKinnon