Proceedings of the Iowa Academy of Science
Volume 55 Annual Issue Article 39
1948
Simultaneous Reduction of Quadratic Forms
B. Vinograde
Iowa State College
Let us know how access to this document benefits you
Copyright ©1948 Iowa Academy of Science, Inc.
Follow this and additional works at: https://scholarworks.uni.edu/pias
Recommended Citation
Vinograde, B. (1948) "Simultaneous Reduction of Quadratic Forms," Proceedings of the Iowa Academy of
Science, 55(1), 291-292.
Available at: https://scholarworks.uni.edu/pias/vol55/iss1/39
This Research is brought to you for free and open access by the IAS Journals & Newsletters at UNI ScholarWorks. It
has been accepted for inclusion in Proceedings of the Iowa Academy of Science by an authorized editor of UNI
ScholarWorks. For more information, please contact [email protected].
Offensive Materials Statement: Materials located in UNI ScholarWorks come from a broad range of sources and
time periods. Some of these materials may contain offensive stereotypes, ideas, visuals, or language.
Vinograde: Simultaneous Reduction of Quadratic Forms
Simultaneous Reduction of Quadr41tic Forms
B. VINOGRADE
Let x denote the real n-tuple (x,, .. ., xn), (x x ... ) the iterated
=
n-tuple (x,, .. ., Xn, x,, . . . , x"' . . . ), A (a i j) and A' a square
matrix and its transpose. We wish to consider the effect of non-
singular transformations on a sum x(A, + ... + A.)x' of k quad-
ratic forms when the real symmetric matrices A, and A = A, + ...
A. are subjected to certain restrictions on their ranks. The results
+
attained link a theorem in statistics with a theorem in ring theory
(see Corollary 3). Extension of the results to hermitean forms fol-
lows easily.
Theorem. Let A" .. ., A. be n-rowed real symmetric matrices with
ranks n,, .. ., n •. If
1. A = +
A, + ... A, is non-singular, and
2. n, + ... + = n. n
then there exists a real non-singular matrix Q such that Q A,Q' =
diagonal (0, L., 0), for i = 1, ... , k, where L, is an n,-rowed diagonal
matrix whose entries are the non-zero characteristic roots of A., and
the first 0 is n, + ... + n,_,-rowed.
Corollary 1: Let r(n, + ... + n,_1 + 1), ••• , r(n 1 + ... + n,)
denote the characteristic roots of A,. Then the linear transformation
x = yQ gives x A,. x' equal to the sum of terms r(j)yJ• with j running
from n, + ... + n,_1 ton,+ ... + n, and i = 1, ••• , k.
Proof of Theorem and Corollary 1: For each i there exists a real
orthogonal matrix R, such that R,'A,R, = diag (0,L,,0). Hence x
Ax'= (xx ••. ) diag (A" . . . , A.) (x, . . . , x)' = (x x . . . )
diag (Ri, . . . , R.)D diag (R,', ••. , R.') ( xx . . . )' = (xx . . . )
E. D E' (x x . . . )', D = diagonal (D,, . . . , D.) and E = diag.
(Ei. ..• , E.) where E, has zeros in all columns except those corre-
sponding to the same columns in (0, L., 0) = D., and where by vir-
tue of the associative law we have replaced with zeros those columns
of R, which do not actually participate in the matrix product. Hence
x Ax'= x Eo' x', where L = diag (L" ... and Eo consists of the non-
zero columns of E,, . . . , E. juxtaposed. So we have EoL Eo' = A.
Since A and L are non-singular, it follows that IEol is not zero.
Furthermore, since E and R, coincide in the columns numbered
0
n, + ... + n,_1+ 1 to n, + ... + n,, we may write x A, x' = x
R, diag (0, L., 0) R,'x' = x E diag (0, L., 0) Eo' x' = y diag
0
(0, L., O)y'. To get the Q of the theorem, take Q = E 0 - 1 • This com-
pletes the proof.
Corollary 2: If, in the main theorem, A = A, + ... + A. is positive
definite, then there exists a Q such that QA,Q' equals diag (0, In., 0),
i = 1, ••. , k, where In, is the n,-rowed identity matrix and reposes
in the block occupied by L, in the main theorem. Hence Q- 1 Q- 1 ' = A.
Proof: Since A is positive definite, there exists a non-singular real
B such that A= BB'. Let P be the E of the main theorem. Then we
0
have PLP' =BB', L = p-1B(P-1B)', implying that Lis positive defin-
ite, hence the r(i)'s are positive. Let L'h denote the matrix L with
291
Published by UNI ScholarWorks, 1948 1
Proceedings of the Iowa Academy of Science, Vol. 55 [1948], No. 1, Art. 39
292 IOWA ACADEMY OF SCIENCE [VOL. 55
r(i)%'s replacing the r(i)'s. Then PLP' = P L'h(PL'h)' =A, or writ-
ing Q-1 = PL'h, we have QAQ' =I. Hence xAx' = yy', where x =
yQ. The structure of PL'h implies QA,Q' = diag(O,In,,0), i = 1,
•.. , k.
Corollary 8. If in the main theorem A, + ... + A. = I, then Corol-
lary 2 implies that Q is real orthogonal. Hence the characteristic
roots of the A 1's are O's and l's. (This gives Cochrane's theorem on
sums of quadratic forms.') Furthermore, A 1Ai = d 11 A., where d, 1 is
the Kronecker delta.
Proof: Corollary 2 implies that the Q constructed there will satisfy
QQ' =I. Hence A 1 is orthogonally equivalent to In., that is, the char-
acteristic roots of A, are O's and l's, i = I, ••• , k. Furthermore, the
relations In,Ini = =
d 1iln 1 and 101 QA,Q' imply QA,Q' .QAiQ' =
QA,AiQ' = d 1 iQA,Q', or A,Ai = d,iA,.
A converse of Corollary 3 is implicit in a known general ring
theorem•. Moreover, again in connection with Corollary 3, it may be
remarked that when A = I it can be proved directly that the charac-
teristic roots of the A;'s are all O's and I's. For let R, be a real ortho-
gonal matrix which reduces I - A, to R',(I - A,)R,' = I - diag
(L,,O). But the latter equals a matrix S = B, + .•• B., where Bi
= R,AiR,'. Now the rank of Sis less than or equal ton,+ ..• + n •.
Hence S = diag(O,l0 _ 0 i) and L 1 = 10 • In this way one sees that L
= I. This fact can be utilized to obtain a simple proof of Cochrane's
Theorem by substituting directly in the main theorem.
DEPARTMENT OF MATHEMATICS,
IOWA STATE COLLEGE.
'S. S. Wilks, Mathematical Statistics, p. 105.
2N. Jacobson, The Theory of Rings, AMS Colloquium Publication,
p. 28.
https://scholarworks.uni.edu/pias/vol55/iss1/39 2
