The following lemma will be needed in the proof of Theorem 3.1, at a point where we will apply Lemma 2.3.

*Proof:* Let ,
be vectors such that if and only
if , *i*=1, , *n*, and if and only if
, *j*=1, , *m*. Then a vector is in
if and only if there are constants so that . For this we need

so that *d* is in if and only if the vector

is in the span of the vectors

This characterization of the image of *C* under the projection and
the fact that inner product is continuous is enough to yield the result.

We now develop a criterion slightly different from that of the previous
section. We suppose that consists of *k*
negative eigenvalues (counting multiplicity), that *ker*(*A*) is finite
dimensional (this will be implied by the requirement that is trivial), and
for some
. Suppose that is spanned by
. We may assume that , ,
are orthogonal to *ker*(*A*) by projecting onto . Define *E* to be

*Proof:* Again we need to show that *A* is strongly positive on
. By orthogonalizing the matrix *E* as in
Lemma 2.2, we may assume that if and that . Take
an . Then *x*=*y*+*z*, where and . One
easily sees that .

By arguing as in Theorem 2.1, one can show that there are
constants , , so that , where *v* is orthogonal to , , . In
addition, *v* is orthogonal to *ker*(*A*), since *y* and the 's are.
Substituting,

Since , we have , so
that holds for all *j*. Therefore

with the last inequality following from the fact that *v* is orthogonal to
all non-positive eigenvectors of *A*. We must now relate the length of *v*
to that of *x*. As before, we use Lemma 2.3. Since
is trivial, there is a so that , where is independent of *x*.

For a similar inequality relating and , we
must first show that the orthogonal projection of onto
has only the trivial intersection with the span of
. Suppose that some
is also in the orthogonal projection of onto . There is
a vector so that . Then for
each *j* from 1 to *k*,

so that is the zero vector. To apply Lemma 2.3 we must
also show that the image of under orthogonal projection onto
is closed. This follows from Lemma 3.1. Therefore,
there is some (not depending on *x* or *y*) so that . We now obtain that

for all , so that *A* is strongly positive on . As
above, this is sufficient for the result to follow.

**Acknowledgments.** I would like to thank John Maddocks for
suggesting that I extend the results of [6]. I would also like to
thank Thomas Schlumprecht for a useful conversation relating to
Lemma 3.1.

Tue Aug 20 10:23:22 CDT 1996