ON THE MATRIX EQUATION X'X=A

by JOHN E. MAXFIELD and HENRYK MINC

V
(Received 26th July 1961)

1. Introduction
If A' is a matrix with non-negative entries then X'X is positive semi-definite
with non-negative entries. Conversely, if A is positive semi-definite then
there exist matrices Y, not necessarily with non-negative entries, such that
Y'Y = A. In the present paper we investigate whether, given a positive semi-
definite matrix A with non-negative entries, the equation X'X = A has a
solution X with non-negative entries. An equivalent statement of the problem
is: Can a positive semi-definite matrix with non-negative entries be expressed
as a sum of rank 1 positive semi-definite matrices with non-negative entries ?
We answer the question in the affirmative for « ^ 4 and quote the following
example due to M. Hall (1) to show that the answer is in the negative for

The matrix A = / { 0 0 4 4
0 1 * 0 4
0 i 1 4 0
4 0 4 1 0
4 4 0 0 1
is positive semi-definite, yet there exists no matrix X with non-negative entries
such that X'X = A.
There is a rather surprising geometrical consequence of this answer: In an
n-dimensional space ( « ^ 5) not every pencil of n vectors such that the angle
between any two vectors does not exceed n/2 can be placed in the non-negative
orthant of coordinates by a rotation or by a rotation followed by a reflection.
We are indebted to Marvin Marcus and Morris Newman for drawing our
attention to the problem.
2. Results
Lemma 1. Let A = (<zfj-) be a positive semi-definite matrix, a y ^ 0 . Then
for any i, j , k at most one of a,jakk-aikaJk, a^a^-a^a^, a^au-a^a* can be
negative.
Proof. S u p p o s e t h a t ajjakk<aikaJk a n d aika}i<axiajk. Then y y y
<aikajkaijajk, i.e. aiiakk<a]k w h i c h is i m p o s s i b l e s i n c e a ^ a ^ — a2^ is a p r i n c i p a l
m i n o r o f t h e p o s i t i v e semi-definite m a t r i x A .
Lemma 2. Let Abe a A-square positive semi-definite matrix, atj 2:0. Then
there exists a permutation matrix P such that all the entries in the leading 3-square
E.M.S.—I

https://doi.org/10.1017/S0013091500014681 Published online by Cambridge University Press

 126 J. E. MAXFIELD AND H. MINC
principal submatrix of C2(PAP'), the second compound matrix of PAP', are
non-negative unless aijakk<aikajk, aipmm<aimaJm, akmau<aikaim, akmajj<ajkaJm
for {i,j, k, m) = (1, 2, 3, 4) or (1, 3, 2, 4) or (1, 4, 2, 3).
Proof. The diagonal entries in C2(PAP') being principal minors of a
positive semi-definite matrix are non-negative. The remaining entries are
equal to aijakk-aikaJk, aimakk-aikakm, aJmakk-aJkakm. Hence the upper 3-square
principal submatrix of C2(PAP') has non-negative entries if and only if a row
in
231 | | 241 | |341
132 | 1 142 | |342
123 | 1 143| |243
124 1 1134 1 1234
where | ijk | = a^a^—a^a^, has non-negative entries. Suppose that all rows
have at least one negative entry. Let | ijk | <0. Then, by Lemma 1, \jki | ^ 0
and | ikj\"^O. If there is a negative entry in the yth row it must be either
| imj | or | kmj\.
(i) If | imj| <0, then \jmi | ^ 0 and | ijm | ^ 0 . Thus if the rth row contains
a negative entry it must be | kmi |. Hence | ikm | ^ 0 and all the entries in the
mth row are non-negative. For if | jkm | <0 we have:

k <aJmakm,
<aikaim,
a
jflim <aifijm
and therefore a^ajja^a^ < afka%. But anakk-afk and anamm- a)m are principal
minors of positive semi-definite matrix A and thus are non-negative. Contra-
diction.
(ii) If | kmj | <0 it is easy to see that all entries in the Jth or in the mth row
are non-negative unless | kmi | < 0 and | ijm | <0, as stated in the lemma.
The last proviso in Lemma 2 cannot be omitted. For example,

is positive semi-definite but there exists no permutation matrix P for which the
leading 3-square principal submatrix of C2(PAP') has non-negative entries.
Theorem. Let A = (a,v) be a positive semi-definite n-square matrix («^4)
with non-negative entries. Then there exists an n-square matrix X with non-
negative entries such that X'X = A.
Proof. The theorem is trivial for n = 1. We prove it for n = 2, 3 and 4.
We can assume therefore, without loss of generality, that no main diagonal
element of A is 0, for if an n-square positive semi-definite matrix has a diagonal
element 0 then every element in the same row and column must vanish and

https://doi.org/10.1017/S0013091500014681 Published online by Cambridge University Press

 ON THE MATRIX EQUATION X'X = A 127
we are essentially dealing with an (n — l)-square matrix.

(i) n = 2. X=

VI A

(ii) n = 3. Letg = dmg{y/all,yJa22,Ja33}andletB = (biJ) = Q xAQ~l.

Then it is easy to see that unless all b,j = 1, in which case we choose X =
at least one principal minor of order 2 of B is strictly positive. Let 1 — bfj>0.
Then l>Z>y and either b^—b^b^O or bik-bijbjk^O. In either case there
exists a permutation matrix P such that PBP' = C = (cy) where l - C i 2 > 0
and c23 — c 12 c 13 ^0. Then X = YPQ where
y=

Vic |

(iii) « = 4. Let Q = Va22. Va33» Va44> and let 5 = (btJ)

(a) First suppose that the conditions of Lemma 2 are satisfied and that
there exists a permutation matrix P such that all the entries in the leading
3-square submatrix of C2(PBP') are non-negative.
Let PBP' = C = and letZ =

where F
Then Z'Z = =

Thus we can find a matrix Z with non-negative entries such that Z'Z = C if
a 3-square matrix 7 with non-negative entries can be found such that
Y'Y = Cn — V'V. Now, Cx x - V V is precisely the leading 3-square submatrix
of C2(PBP') all of whose entries are non-negative. Hence, by part (ii), the
required Y can be be found.
(b) Suppose now that b^-b^b^, bij-bimbjm, bkm-bikbim and bkm-bJkbJm
are all negative. Without loss of generality we can assume that (/, j , k, rri)
= (1, 2, 3, 4). We first show that there exists a matrix R with non-negative
entries such that {R'^'BR'1 = C = (c,) has at least 4 entries equal to 0.

https://doi.org/10.1017/S0013091500014681 Published online by Cambridge University Press

 128 J. E. MAXFIELD AND H. MINC
Let R~l = •/• 1 -bi2 -xt -x2
0 1 0 0
0 0 1 0
o o "; o i
where xu x2 are non-negative numbers to be determined. We have c 12 = 0,
c
i3 = b13~Xi, c14 — bi4.—x2, c 23 = b23-b12b13, c 24 = b24.-bl2bl4,
c3A=f(.xu x2) = b^-b^Xi-b^Xi + x^^ Since b12-b13b23<0 and
bi2~*i4*24<0 both c 23 and c 24 are, by Lemma 1, non-negative. Choose
xu x2 so that xt^b13, x 2 g £ 1 4 andf(xu x2) = 0. This can always be done
since/(0, 0) = 6 3 4 ^ 0 while f(b13, 6 14 )<0. Then C, which is positive semi-
definite, has 0 in the (1, 2), (2, 1), (3, 4) and (4, 3) positions arid non-negative
entries elsewhere.
Let Z = (z y ) be a 4-square matrix with
z 12 = z 14 = z 21 = z 24 = z 32 = z 33 = z 41 = z43 = 0.
We show that we can determine the other z y in such a way that z o -^0 and
Z'Z = C, i.e.

42 = Zia + zfa = zli + zL = 1

and 1
Z
11Z13 = C
13> Z
31Z34 = C
14> Z
22Z23 = C
23> Z
42 Z 44 = C
24-

It will follow that all entries in ZRQ are non-negative and (ZRQ)'ZRQ = A.
A straightforward computation gives

Denote z\x by z and the minor of c y by | C y |. The equation becomes

(1) z 2 I C1X | - z ( | C | - 2 c 1 3 I C 13 |) + cj 3 I C 33 I = 0.
We prove first that the roots of (1) are real. The discriminant of (1) is
D = (| C | - 2 c 1 3 I C 13 | ) 2 - 4 c 2 3 I C n I I C 33 |.
But, by the classical theorem of Jacobi on the minors of adj(C),
| C n l | C 3 3 | - | C 1 3 | 2 = |C|(l-c!4).

Therefore D = | C | 2 - 4 c 1 3 | C | | C 13 |+4c 2 3 | C 13 | 2

-4c23|C|+4c23c24|C|-4c23|C13|2'
= I C |{| C | - 4 c 1 3 ( | C 13 l + c 1 3 -c 1 3 c 2 4 )}
= |C|(|C|+4c 1 3 c 1 4 c 2 3 c 2 4 )
which is non-negative. Therefore the roots of (1) are real. The coefficient

https://doi.org/10.1017/S0013091500014681 Published online by Cambridge University Press

 ON THE MATRIX EQUATION X'X = A 129
of z in (1) is
- | C | + 2c13|C13|= -|C|-2c13{c13(l-cl4)+c12c24c34}
which is non-positive and, since c\3 \ C33 |2sO, both roots are non-negative.
Hence there is a real non-negative root for z1,.
The corresponding equations for z22, z33 and z 44 yield similar results.

REFERENCE
(1) MARSHALL HALL, Jr., A Survey of Combinatorial Analysis, Surveys in Applied
Mathematics IV (Wiley, 1958), 35-104.

UNIVERSITY OF FLORIDA
GAINESVILLE, FLORIDA

https://doi.org/10.1017/S0013091500014681 Published online by Cambridge University Press