BULL. AUSTRAL. MATH. SOC.

I5A24
VOL. 10 ( 1 9 7 4 ) , 377-401.

Matrix quadratic equations

W.A. Coppel

Matrix quadratic equations have found the most diverse applications.

The present article gives a connected account of their theory, and
contains some new results and new proofs of known results.

1. Introduction
An autonomous linear hamiltonian system of differential equations has
the form

(1) x' = J hx ,
n
where x €R and

0 _J A B*
J= , H=
J o B C

are constant 2n x 2n real matrices which are respectively skew-symmetric,

symmetric.1 It is well-known, and easily established, that there is a
close connection between the solutions of (l) and the solutions of the
n *• n matrix Riccati equation

(2) R[W] E W' + A + WB + B*W + WCW = 0 .

The constant solutions of (2) are just the solutions of the matrix
quadratic equation

(3) Q[W] = A + WB + B*W + WCW = 0 .

Largely as a result of this connection, the matrix quadratic equation (3)

has found applications in many different fields; for example, optimal

Received k February 1971*-

The conjugate transpose of a matrix iB denoted by an asterisk.
377

https://doi.org/10.1017/S0004972700041071 Published online by Cambridge University Press

 378 W.A. Coppel

control, s t a b i l i t y theory, filtering theory, network theory and

differential games. In these applications i t is the symmetric solutions of
(3) which are of interest. The object of the present article is to give a
connected account of the theory of the equation (3). References to
previous work are made at various places in the text.

The following identities are easily verified. If W and W are

solutions of (3) then their difference D= W - W satisfies

(1+) D[B+CW) + [B*+W C)D = 0 ,

(5) D(B+CW ) + [B*+W C)D + DCD = 0 .

Conversely, if W is a solution of (3) and if D satisfies (h) or (5),

then W is also a solution of (3).

The coefficient matrix

B C
(6) M =
-A -B*

has the property

(7) JM = -M*J .

It follows that the characteristic polynomial

(8) cp(s) = det(sJ-M)

is an even function of s . If W is a solution of (3) then

(9) (p(s) =

since

I 0 T
0 B+CW
M
W I w I -B*-WC

Finally we recall some standard definitions. A matrix is said to be

stable if a l l i t s eigenvalues have negative real part. The inequality
W > W between two symmetric matrices W, W means that W - W is

non-negative definite. If B is an n x n matrix and C an n x m

matrix the ordered pair (S, C) is said to be controllable if the n * rnn

https://doi.org/10.1017/S0004972700041071 Published online by Cambridge University Press

 Matrix quadratic equations 379

block matrix [C BC .. . B C) has rank n . For further discussion of

this concept see, for example, Kalman et al [3]. If the pair (B, C) is
controllable then for any m x n matrix W the pair (B+CW, C) is also
controllable. If C = C* £ 0 then (B, C) is controllable if and only if

(10) ft H - I e~tBCe~tB*dt > 0

J
0

for some, and hence every, T > 0 .

LEMMA 1. If (B, C) is controllable and C £ 0 then B = B +

is stable.

Proof. We have

and hence

Bfi + fiB" = e Ce + C .

Let X be an eigenvalue of B* and 5 a corresponding eigenvector. Then

£ 0 .

Since ft > 0 it follows that RX £ 0 . Moreover, if equality holds then

C? = 0 . From the definition of B this implies B*X, = A? . Hence

Since C ^ 0 , this is a contradiction.

-r. I e-
tB
CX,e-Xtdt = 0 .

Hence B is stable.

A somewhat different proof of Lemma 1 is given by Lukes [7].

2. Extreme solutions
Throughout this section it will be assumed that C £ 0 and that
(B, C) is controllable. Moreover by a solution of (3) we will always mean
a syrmetrio solution, and we will assume that (3) has at least one solution

https://doi.org/10.1017/S0004972700041071 Published online by Cambridge University Press

 380 W.A. Coppel

W .
THEOREM 1. The matrix quadratic equation (3) has maximal and minimal
solutions W and W ; that is, every solution W satisfies

All eigenvalues of B + CW+ have non-positive real part and all eigen-
values of B + CW have non-negative real part.

We give two proofs. The f i r s t uses the relation with the Riccati
equation (2). Let W be a solution of (3) and put
B = B + CW ,

rt
Uit) = f eTBCeTB*dT
>0

Since C < 0 and (B, C) is controllable, the symmetric matrix U(t) is

a strictly decreasing function of t . In particular, U(t) > 0 for

t < 0 . As t-*-- 0 0 , U~ (t) decreases and is bounded below by 0 . Thus

lim U it) exists and is non-negative.

The function u(t) is the solution of the linear differential

equation

(11) U' = BU + VB* + C

which vanishes at t = 0 . For any T > 0 put

WT(t) =W+ U'XU-T) .

Using (11), it is easily verified that W (t) is a solution of (2) on the

interval [0,2") . The representation

wT(t) = [w/(t-T)+J]y~1(t-27)

shows that wAt) is non-singular on a small interval [T-6, T) and

wZ^it) •*• 0 as t •*• T . If VAt) is the solution of the 'inverse'

Riccati equation

https://doi.org/10.1017/S0004972700041071 Published online by Cambridge University Press

 Matrix quadratic equations 38 1

V' - C - VB* - BV - VAV = 0

such that V~(T) = 0 it follows that V-,[t) is non-singular and

V^(t) = Wf(t) on [T-6, T) . This shows that Wf(t) can also be defined
independently of W • The existence of W simply guarantees the existence
of WT{t) over the interval [0, T) . As T •* <*> , Vy(t) converges

uniformly on compact i-intervals to the constant matrix

(12) W+ = W + lim U~X{t) .

Thus W is a constant solution of the Riccati equation (2); that is, it

is a solution of the quadratic equation (3)- Moreover W+ > W for any
solution W of (3). Since (B, C) is controllable it follows from Lemma
1 that

-if
B + CWT(O) = B - C\ f e"TBCe"TB*dx]

is a stable matrix. Hence, letting T •*• °° , all eigenvalues of B + CW+

have non-positive real part.

By replacing W by -W we see that (3) also has the minimal solution

(13) W = W + lim U'1^) ,

and all eigenvalues of B + CW have non-negative real part.

The second proof provides a practical algorithm for the determination
of the maximal and minimal solutions. It is based on the identity
(ll») W{B+CW) + (B+CW)*W - WCW = W(B+CW) + {B+CW)*W - WCW + {W-W)C{W-W) ,
valid for arbitrary symmetric matrices C, W, W , and on the following
simple
LEMMA 2. Suppose C < 0 and the linear equation
(15) WB + B*W = C
has a solution W > 0 . If B has an eigenvalue \ = \i + iv with \i > 0
d •£/ x, is a corresponding eigenvector then Cz, = 0 .

https://doi.org/10.1017/S0004972700041071 Published online by Cambridge University Press

 382 W.A. Coppel

In fact from (15) we obtain at once

Since the left side is non-positive and the right side is non-negative they
must both be zero. Since C 5 0 this implies Ct, = 0 .

We now continue with the second proof of Theorem 1. By Lemma 1 there

exists a symmetric matrix W such that B + CW is stable. We define a

sequence {W } of symmetric matrices inductively by the linear equations

(16) W , [B+CW ) + [B+CW )*W , = W CW - A ( v> 0 ) .

V+lk v* *• VJ V+l V V
Suppose W has been defined and B + CW is stable. Then (l6) has a

unique and symmetric solution W . By the identity (l^t), with W= W

V+l

and W= W ,

(17) -A = W[B+CWv) + [B+CUj*W - WvCWv + (w^W) C[W^-W) .

Adding this to (l6) we get

(18) ( " v + 1 - ^ [P+CWj * [B+CWv)*{Wv+1-W) = [Wv-W)c{Wv-W] .

Therefore

, - v - • ,/* t[B+CW)
v
W - W= - e [W -W]C[W -W)e dt

2 0.

By the identity (lU) also, with W=W and 1/ = W ,

V+l V
(19) w _,, [S+CP/ _,_,) + fs+cv _)*»/ , - w nw _,.. =
v+ll v+l; "• v+l-* v+l v+l v+l
= -A + v(W ^.-W )C[W -W ) .
v + l v ; *• v + l \>J

Adding (17), with v replaced by v + 1 , we get

(20) (w^- J(1

Suppose 5 +CT/+ had an eigenvalue X with non-negative r e a l p a r t , and

let X, be a corresponding eigenvector. Then, by Lemma 2,

https://doi.org/10.1017/S0004972700041071 Published online by Cambridge University Press

 Matrix quadratic equations 383

Therefore

Since B + CW is stable, this is a contradiction. Thus B + CW , is

stable.

Subtracting (l6) from (19) with v replaced by v - 1 we get

Therefore

w -w =-I iVcK-xVe dt
V V+l JQ
> 0 .

Thus, after the first term, the sequence {w } is monotonic decreasing and

bounded below by W . It therefore converges to a limit W . Letting

V -»• °° in (16) we see that W+ is a solution of (3). Since W+ > W for

any solution W of (3) it is in fact a maximal solution. Since

B + CW+ = lim(B+C!v' ) , all eigenvalues of B + CW have non-positive real

part.

The existence of extreme solutions seems to have been first

established by Reid [//], to whom the formulae (12), (13) are also due
(Reid [/2]). The algorithm used in the second proof was introduced in a
special case by Kleinman [4] and further developed by Wonham 1161. Its
applicability in the general case is mentioned by WiI I ems [15].

Our next result relates two arbitrary solutions of (3).

THEOREM 2. Let W^ V^ be any two solutions of (3) and set

D = U2 - Wx , B 1 = B + CW±1 B2= B + CW2 .

Let W denote the nullspace of D and let2 m, p, q be the dimensions of

2
m, p , and q are the number of zero, positive, and negative eigenvalues
of D .

https://doi.org/10.1017/S0004972700041071 Published online by Cambridge University Press

 384 W.A. Coppel

the largest subspaaes invariant under D on which D= 0 , D> 0 ,

D<0 .
Then N is invariant under B and B = B on N . Let
X , ..., \ be the eigenvalues of the restriction of B to N and let
X , ..., X be the remaining eigenvalues of B . Then the eigenvalues
of B are X , ..., X , ~\n+T> •••» -^ • Moreover among the p + q
eigenvalues X ,...., X exactly p have positive real part and q
have negative real part.
Proof. We can assume p + q > 0 , since the assertion is t r i v i a l for
0=0. The difference D is a solution of the quadratic equation
(22) DB + B*D + DCD = 0 .

Moreover (5 , C) i s also controllable. If £, € N then, on postmul-

t i p l y i n g (22) by E, , we obtain DB E, = 0 . Thus N i s invariant under

B± . Since B = + CD it is obvious that = B on

By replacing D, B , C by T*DT, T~XBT, 21"1C2"1"1 , which doesn o t

alter the hypotheses, we can assume that

0 0
D =
0 D,

where D is non-singular. The nullspace W now consists of all vectors

whose last p + q coordinates vanish. Let the corresponding decomposit-

ions of B and C be

B B C c0 1
00 01 00
B = , C=
1 o B c*
11 0 1 c1 1

Then, putting G= D , we have

(23) B^G +

Moreover

https://doi.org/10.1017/S0004972700041071 Published online by Cambridge University Press

 Matrix quadratic equations 385

B B
oo oi+coA
B1+ CD =

Since B + CD = -GB* G , this proves the assertion concerning the

eigenvalues of B .

We are going to show now that B has no pure imaginary eigenvalues.

Suppose on the contrary that B* t, = ivX, , where u is real and

C, t 0 . Then from (23) we obtain

Since C £ 0 t h i s implies C r = 0 , and also C X, = 0 . Thus if we

put

then
= 0 .

Hence

t,*[c BC ... Bn~YC) = 0 .

Since t, + 0 , this contradicts the hypothesis of controllability.

Put

e )
V •Bn + zG
•

Then, by (23), B,,(e) is a solution of the linear equation

B u (e)G + + C u - 2EG2 = 0 .

By a generalisation of Ljapunov's Lemma (see Ostrowski and Schneider [JO],

Lancaster [6]) it follows that for e > 0 , s,-,(e) has no pure imaginary

eigenvalues and further that it has p eigenvalues with positive real

part, q eigenvalues with negative real part. Since B has no pure

https://doi.org/10.1017/S0004972700041071 Published online by Cambridge University Press

 386 W.A. Coppel

imaginary eigenvalues, the same holds by continuity for i t .

COROLLARIES, (i) If W is a solution of (3) such that all eigen-

values of B + CW have nongositive veal -part then W = W+ . [Take
w2 = w+ , w1 = w .)
(ii) If W is any solution 0 / ( 3 ) then the nullspace V of
W+ - W is invariant under B + CW and the restriction of B + CW to M
has only pure imaginary eigenvalues. No other eigenvalues of B + CW are
pure imaginary.

Proof. If C € l / then W+E, = V% = W % , s i n c e

0 £ £*(W+-W)E, £ Z*[W+-W_)Z = 0 .

Since V is invariant under B + CW it is equally invariant under

B + CW . By Theorem 1 the restriction of B + CW+ to V has only pure

imaginary eigenvalues. By Theorem 2 , B + CW has no other pure imaginary

eigenvalue. If B + CW had an additional pure imaginary eigenvalue then,

by Theorem 2 again, it would be associated with the nullspace of W+ - W

and therefore be an eigenvalue of B + CW+ .

The second corollary has been stated without proof by Wi I I ems [75].

It will now be shown that all solutions of (3) can be expressed in

terms of the extreme solutions. Put

A = V + -V , B+ = B + CW+ , B = B + CW .

By (h) we have

(2U) B*h = -AB_ .

The nullspace V of A is invariant under B+ and, by what we have just

proved, there is a unique subspace V invariant under B+ such that the

whole space V=R satisfies

(25) V = VQ i V+ .

https://doi.org/10.1017/S0004972700041071 Published online by Cambridge University Press

 M a t r i x q u a d r a t i c equations 387

All eigenvalues of the restriction of B+ to V are pure imaginary and

all eigenvalues of the restriction of B+ to V+ have negative real part.

LEMMA 3 . Let I/., be a subspace invariant under B+ . If

V n V = 0 then V c 1/+ .

Proof. Suppose £ / 0 € 1/ . Then £ = £_ + £ , where £- € 1/ and

£+ f V+ . Since 1/ n V = 0 we must have £+ ^ 0 . Let p(s) toe the
polynomial of least positive degree such that p\PAK. - 0 . Then p(s)
has only roots with negative r e a l part and

Since VQ i s invariant under B+ and V n V = 0 i t follows that

P( B +K 0 =
° -
Therefore £ = 0 , and V c V+ .

THEOREM 3 . Let 1/ be an arbitrary subspace of V+ invariant under

B+ . If V is the subspace of all vectors £ such that A£ is
orthogonal to V then

Let P be the corresponding projection of 1/ onto V . Then

(26) W = W+P + W (I-P)

is a solution of the quadratic equation (3). Moreover all solutions of (3)

are obtained in this way, and the correspondence between \) and W is'
one-to-one.
Proof. Evidently V c 1/ . If £ € V n V then £*A£ = 0 . Since

4iO it follows that A£ = 0 . Since V n V = 0 this implies £ =0.

Thus V n V= 0 . We have also

dimAl/ = n - dimf ,

dirnl/^ = n -

https://doi.org/10.1017/S0004972700041071 Published online by Cambridge University Press

 388 W.A. Coppel

and hence

(U^) > n - di

Therefore

diml/ = diml/ + dim(Al/n/') > n - diml/

I t follows t h a t

If 5 1 € Vx and £ 2 € l/g then, by (2U),

= o,

since 1/ is invariant under B + . Thus V is invariant under B

Hence the projection P satisfies

(27) PB+P = B+P

and

(I-P)BJI-P) = B_(J-P) ;
that is,

(28) PB P = PB .

Since AC is orthogonal to V we have also

P*A(J-P) = 0 ;

that is,

P*A = P*AP .

Since the r i g h t side i s symmetric t h i s gives

(29) P*A = AP .

By (27) and (28),

+CA)(J-P)
= B (I-P) + (J-P)CA(J-P)

Therefore, by (2U),

https://doi.org/10.1017/S0004972700041071 Published online by Cambridge University Press

 Matrix quadratic equations 389

A(I-P)B + + B*A(J-P) = A(I-P)CA(I-P) .

If we define W by (26) then

D = w+ - W = A(I-P) .

By (29), D is symmetric, and hence W is also. Moreover W is a

solution of the quadratic equation (3) if 0 is a solution of the
quadratic equation

DB+ + B*D - DCD = 0 .

But we have just shown that this is the case.

Furthermore

WE, = W+E, for £ € V± , W$ = VJ, for £ € l/g .

Therefore ^ and 1/ are invariant under B + CW and all eigenvalues of

the restriction of B + CW to V , V have respectively negative, non-

negative real parts. This proves that V is uniquely determined by W .

Conversely, let W t>e any solution of (3). Then we have a unique

direct decomposition

where V and V are invariant under B + CW and all eigenvalues of the

restriction of B + CW to V, V have respectively negative, non-

negative real parts. It follows from Theorem 2 that

for 5 € ^ , »5 = " J for £ € I/,, .

Thus if P is the corresponding projection of 1/ onto V then V

satisfies (26). Moreover V is invariant under B+ and V n 1/ = 0 .

Therefore, by Lemma 3, V c 1/+ . Since V + - V = A(J-P) is symmetric,

(29) holds. Therefore, since P is a projection,

P*A = P*AP ;

that is,

P*A(J-P) = 0 .

https://doi.org/10.1017/S0004972700041071 Published online by Cambridge University Press

 390 W.A. Coppel

Thus Al/ is orthogonal to V . Since V = V + V it follows from the

first part of the proof that \1 is the subspace of all E, such that A£

is orthogonal to 1/ • This completes the proof.

In particular, if A > 0 then V = 0 , V is an arbitrary

invariant subspace of S + , and V- = A 1/7 . Theorem 3 was proved for

this case by WiIlems [75].

In the statement and proof of the following result we use the notation
of Theorem 3.

THEOREM 4. Let W, W be solutions 0/(3) corresponding to the

/invariant subspaces V , V of B . Then W > W if and only if

Proof. Suppose first that V c 1/ . Then, by Theorem 3, V c V .

On Vx we have WE, = W+E, = WE, . On l>2 we have WE, = W E, = WE, .

Therefore V + V is contained in the nullspace N of W - W and is

invariant under B + CW . The remaining eigenvalues of B + CW have

negative real part, whereas the remaining eigenvalues of B + CW have non-
negative real part. Hence N = V + V . It follows from Theorem 2 that

W > W .

Conversely, suppose W > W . The nullspace W of their difference is

invariant under B + CW , and hence has a unique direct decomposition

N = N+ + N ,

where W ,N are invariant under B + CW and all eigenvalues of the

restriction of B + CW to N+, N have respectively negative, non-

negative real part. ?y the proof of Lemma 3 we have N + c 1/ , N + c V .

By Theorem 2 the eigenvalues of B + CW not belonging to N have positive

real part. Therefore dimN+ = dim!/ . Thus N+ = V and I' £ V .

It follows at once that the set of aM. solutions of (3) is a complete

lattice. This is rather remarkable, since the set of all symmetric n x n

https://doi.org/10.1017/S0004972700041071 Published online by Cambridge University Press

 Matrix quadratic equations 39 1

matrices is not a lattice if n > 1 .

3. Existence of solutions
In this section we drop the hypotheses and convention imposed at the
beginning of Section 2. If W is a solution of (3), not assumed
symmetric, such that B + CW is stable then the characteristic polynomial

<p(s) = det (sI-B-CW)

is relatively prime to ty(-s) . Since, by ( 9 ) , ty(s) divides

<p(e) = det(sI-M) , and <p(s) is even, it follows that

(30) <p(s) = (-l)n\l>(sM-s) .

On the other hand it is clear that <p(s) admits a factorisation (30),

where the (real) polynomial ty(s) is monic and relatively prime to
ijj(-s) , if and only if M has no pure imaginary eigenvalue. Moreover it
is possible to choose ty(s) so that all its roots lie in the left half-
plane.

THEOREM 5. Suppose q>(s) has a factorisation (30)., where tyis) is

monic and relatively prime to ty{-s) . Then there exists a non-singular
matrix X such that
(3D3 X*JX =J
T 0
(32) MX = X 3

0 -T*

(33) det(sJ-T) = <K«) •

t polynomials X-i(s), Xo^s) such t h a t
A. td

(3U) 1 = X l (s)^(8) + X 2 (s)lp(-8) .

Then for any vector x €F

(35) x = X^ 2

Let (/1 denote the subspace of all vectors x such that \\/(M)X = 0 , and

let M denote the subspace of all vectors x such that $(-Af)x = 0 .

A matrix X satisfying (31) is said to be symplectic.

https://doi.org/10.1017/S0004972700041071 Published online by Cambridge University Press

 392 W.A. Coppel

From (35) we obtain a t once

V± • = i?2" .

Moreover V and V are invariant under M.

Let x , ... , x be a basis for V and x , ..., x a basis for

1 TU 1 77H"1 cYl
V , and l e t X be the non-singular matrix whose fc-th column i s x,

(& = 1, . . . , 2n) . Then

T 0'
MX = X
0 5

where T is an m x m matrix, 5 is a (2n-nt) x (2n-m) matrix, and

^(T) = 0 , *K-S) = 0 . Thus the minimal polynomial of T divides \\>(s) .
Since every prime factor of the characteristic polynomial of a matrix
divides its minimal polynomial, it follows that det(sI-T) is relatively
prime to <K-s) . On the other hand, det(sJ-r) divides

(p(s) = (_i)"i|)(e)^(-s) . Therefore det(sJ-T) divides tyis) . Thus m > 0

implies m £ n . Similarly m < 2n implies m > n . Hence m = n and
(33) holds.

It follows from (3^) that ty(-T) and are non-singular. Also,

(T),
T 0
(36) X*JX = X*JMX
0 s
= -X*M*JX
-T* 0
X*JX .
0 -5"

Since *(D = 0 , *( -S) = 0 it follows that

(37) - -f°
X*JX\
0 (\\i(-T)* 0

io i 0 0
Write, in partitioned form,

https://doi.org/10.1017/S0004972700041071 Published online by Cambridge University Press

 Matrix quadratic equations 393

X*JX =

where V* = -£/ . Then (37) implies that U =0 , U = 0 . Since

X*JX is non-singular it follows that V „ is non-singular. From (36) we

now o b t a i n

Put

I 0
X = X
,-1
0 - 12

Then (31) and (32) hold. This completes the proof.

Write in partitioned for 11

Y y
X = 1 2
Z
h 2
Then (31) is equivalent to

(38)

and (32) implies

+ CZ. =
(39)

.-1
Suppose Y is non-singular, and put W = Z Y Then, by (38)1,

W is symmetric. By (39),

B + CW = Y T

https://doi.org/10.1017/S0004972700041071 Published online by Cambridge University Press

 394 W.A. Coppel

A + UB = -Y*~1T*Z*

Hence det [sI-B-CW ) = ^(e) and

eOJ = A +

= 0 .

Let W be any s o l u t i o n of (3) such t h a t i/)(e) = det (sI-B-CW) . Then

M (B+CW)

and hence

\ji(B+CW) = 0 .

Consider i n general t h e l i n e a r equation

J
(UO) = 0 .

Writing .-1 , where

^(Af) = ty(M)X.X~

0 0
1>(M)X = X
0

and

Z
*2 '"
1 X
X' = J~ X*J =

and recalling that 4>(-T) is non-singular, we see that (1*0) is equivalent

to

= Zl •

Since ^ is non-singular there is no vector E, ? 0 such that

I ? = Z ^ = 0 . It follows that (UO) is soluble only if Y is non-
singular, and it then has the unique solution W = W.

https://doi.org/10.1017/S0004972700041071 Published online by Cambridge University Press

 Matrix quadratic equations 395

Thus the equation (3) has a solution W such that

= det [sI-B-CW) if and only if Y-^ is non-singular. When such a

solution exists, it is unique and symmetric. Similarly the equation (3)

has a solution W such that (-l)nty(-s) = det[sI-B-CW ) if and only if

Y is non-singular. Moreover W is then the unique solution of the

linear equation

(-(/ I)iKA/) = 0 ,

with the same coefficient matrix as (Uo). If W and W both exist then

W - W is non-singular, since by (38)3,

Some results in this direction have been stated by Roth [ M ] and Bucy
and Joseph [/]. However Theorem 5, which underpins the method, is absent
in these references.

THEOREM 6. Suppose <p(s) has a factorisation (30), where <Ms) is

mania and relatively prime to ^(-s) . Further suppose C 5 0 , C ± 0.

Thus we can write C = -ER~ E* , where E is an n x m matrix (m 5 n)

and R = E* > 0 is an m *• m matrix. Then either

(i) for any m x n matrix: F , det(sI-E-EF) has a fixed factor

of positive degree in common with ty(-s) , or

(ii) the quadratic equation (3) has a unique, symmetric solution

W± such that det (sI-B-CW-J = *(s) .

Similarly, either

(i)' for any m x n matrix F , det(sI-B-EF) has a fixed

factor of positive degree in common with tyis) 3 or

(ii)' the quadratic equation (3) has a unique, symmetric solution

W2 such that &et[sI-B-CW2) = {-\)nty(-s) .

Both (ii) and (ii) ' hold if and only if (B, E) is controllable;
that is, if and only if (B, C) is controllable.

https://doi.org/10.1017/S0004972700041071 Published online by Cambridge University Press

 396 W.A. Coppel

Proof. Let

X =
Z
h 2
be the matrix whose existence was established in Theorem 5- We have
already seen that i f Y i s non-singular then (ii,) holds.

Suppose Y i s singular, and l e t E, f it be any vector such that

7? = 0 . From (39)i and (38)i we obtain

Since Y E = 0 it follows that £*Z*CZ £ = 0 . Therefore, since i? > 0 ,

(Ul) = 0 .

Putting n = Z C 5 we now obtain from (39),

(1*2) Y^E, = 0 ,

-B*T) = Z

This shows that we can replace £ by TE, in the preceding argument. Thus
for any non-negative integer k ,

=o ,

^ =o .
For any m x n matrix F ,

r\*(B+EF) = -C*T*Z* +

= -E,*T*Z* .

By induction on & , using (U3) and (39)2> it is easily shown that

T)*(B+EF))k = (-l Z .

Hence for any polynomial p(s) ,

https://doi.org/10.1017/S0004972700041071 Published online by Cambridge University Press

 Matrix quadratic equations 397

n*P(-B-EF) = |

If Z p(T)5 = 0 then p(T)? = 0 , since Y p(T)E, = 0 and X is non-

singular.

Fix a particular E, + 0 in the nullspace of Y and let w(s) he

the monic polynomial of least positive degree such that U)(!F)£ = 0 . Then
d)(s) divides the minimal polynomial of T , and hence also ijj(s) .
Furthermore to(s) is the monic polynomial of least degree such that

r\*u(-B-EF) = 0 .

Thus w(-s) divides the minimal polynomial, and also the characteristic
polynomial, of B + EF . This proves (i). Since

B + CW = B + E\-R~1E*W

and ij>(s) is relatively prime to IJJ(-S) , it is clear that (i) and (£i)
cannot hold simultaneously.

3y replacing ip(s) hy (-l)"<K-s) and Y , Z by Y , Z we obtain

(i) ' and (ii) '. Suppose now that (B, E) is controllable. Then by a
well-known property of controllability (see, for example, Heymann [2]),
neither (i) nor (i) ' holds. Therefore both (ii) and (ii) ' hold.
Conversely, suppose that (B, E) is not controllable. Then there exists a
non-singular n x n matrix P such that

B B E
ll 12 l
X
PBP~ = , Pi? =
0 B
22. 0

where > 1 . Write

Then

22

and hence

https://doi.org/10.1017/S0004972700041071 Published online by Cambridge University Press

 398 W.A. Coppel

&et(sI-B-EF) = det (sT-B -E F) det (sI-B ) .

If both (ii) and (ii) ' hold then, taking F = -R~1E*W. {i = 1 , 2 ) , we see
that det[sJ-B ) divides both tp(s) and ip( —s) . Since they are
r e l a t i v e l y prime and det (sJ-Bpp) i s of positive degree t h i s i s a
contradiction.
COROLLARY. Suppose C< 0 . Then the following statements are
equivalent:
(i) (B, C) is controllable and M has no pure imaginary
eigenvalues;
(ii) the equation (3) has a solution W such that B + CW
is stable and a solution W such that -{B+CW ) is
stable.
This Corollary has been established by Molinari [S], [9] by a rather
long argument involving spectral factorisation of matrices of rational
functions. The key idea in the proof of Theorem 6, namely, the derivation
of (Ul) and (U2), is due to Kucera [5]. However Kucera treats only a
special case and uses the Jordan canonical form.
We next use Theorem 5 to discuss the applicability of an algorithm
introduced by Roberts [73].
LEMMA 4. Let A be a linear transformation of the vector space V
which has no pure imaginary eigenvalues3 and let 1/ = l/_ + l/+ be the
direct decomposition of V into subspaces invariant under A such that
the eigenvalues of the restriction of A to V , V+ have respectively
negative, positive real parts. Set
/ 2 (V
(MO AQ=A , 4v+1= ( v C ) "0) •
Then the sequence {A } is defined for all v and A •* A as v-*- 0 0 .,

where A is the linear transformation defined by

(It5) A £ = -£ if £ € V_ , At = E if £ € l/+ .

https://doi.org/10.1017/S0004972700041071 Published online by Cambridge University Press

 Matrix quadratic equations 399

Proof. Since the effect of changing A into -A is otrvious we may

suppose that A is stable. If A has the eigenvalues A,

(k = 1, , n) then

A = (A+A )/2

has the eigenvalues

= 1
' •••' n ) •

Hence A is also stable. It follows that A - I and A - I are non-

singular and

Therefore the sequence {A } is defined for all V and

Since all eigenvalues of (A+I)(A-I) have absolute value less than 1

this shows that [A +j) (d - l ) " 1 •* 0 , and hence A -*• -I .

It is easily seen that if A has at least one pure imaginary

eigenvalue then the sequence {A } is either not defined for all v or

lim A does not exist.

Suppose now that the matrix M in (6) has no pure imaginary

eigenvalues. Then, by Theorem 5, there exists a symplectic matrix X such
that

T 0
M =X
0 -T*
vhere T is stable. Hence
_j 0
„-!
M = X
0)
0 I

Writing

https://doi.org/10.1017/S0004972700041071 Published online by Cambridge University Press

 400 W. A . Coppe1

Y
2
M = , X =
-S* z Z
i 2

we get

A =

If and J are non-singular then C is non-singular and the linear

equations

(1*6) CW = -I - B , CW = I - B

have the unique solutions W = Z JT , 1/ = Z J . Thus whenever (3)

has both stabilizing and antistabilizing solutions they can be found by

Roberts' algorithm. This result is independent of whether C 5 0 , and
consequently should be useful in the applications to differential games.

References
[/] Richard S. Bucy, Peter D. Joseph, Filtering for stochastic processes
with applications to guidance (interscience Tracts in Pure and
Applied Mathematics, 23. Interscience [John Wiley & Sons], New
York, London, Sydney, Toronto, 1968).

[2] Michael Heymann, "Comments 'On pole assignment in multi-input

controllable linear systems' ", IEEE Trans. Automatic Control 13
(1968), 7U8-7U9.

[3] R.E. Kalman, P.L. Falb, M.A. Arbib, Topics in mathematical system
theory (McGraw-Hill, New York; Toronto, Ontario; London;
1969).

[4] David L. Kleinman, "On an iterative technique for Riccati equation

computations", IEEE Trans. Automatic Control 13 (1968), llU-115.

[5] Vladimfr Kucera, "A contribution to matrix quadratic equations", IEEE

Trans. Automatic Control 17 (1972), 3UU—3I+7.

https://doi.org/10.1017/S0004972700041071 Published online by Cambridge University Press

 Matrix quadratic equations 40 1

[6] Peter Lancaster, "Explicit solutions of linear matrix equations",

SI AM Rev. 12 (1970), 5UU-566.
[7] Dan lard L. Lukes, "Stabilizability and optimal control", Funkcial.
Ekvao. 11 (1968), 39-50.
[8] B.P. Molinari, "The stabilizing solution of the algebraic Riccati
equation", SIAMJ. Control 11 (1973), 262-271.
[9] B.P. Molinari, "Equivalence relations for the algebraic Riccati
equation", SIAMJ. Control 11 (1973), 272-285.
[70] Alexander Ostrowski and Hans Schneider, "Some theorems on the inertia
of general matrices", J. Math. Anal. Appl. 4 (1962), 72-81*.
[7 7] William T. Reid, "Riccati matrix differential equations and non-
oscillation criteria for associated linear differential
systems", Pacific J. Math. 13 (1963), 665-685.
[72] Wi Iliam T. Reid, "A matrix equation related to a non-oscillation
criterion and Liapunov stability", Quart. Appl. Math. 23 (1965),
83-87.
[73] J.D. Roberts, "Linear model reduction and solution of the algebraic
Riccati equation by use of the sign function", Dep. Eng. Univ.
Cambridge, Cambridge, England, Rep. CUED/B Control/TR 13, 1971-
[74] WiIliam E. Roth, "On the matric equation X2 + AX + XB + C = 0 ",
Proc. Amer. Math. Soc. 1 (1950), 586-589.
[75] Jan C. Willems, "Least squares stationary optimal control and the
algebraic Riccati equation", IEEE Trans. Automatic Control 16
(1971), 621-631*.
[76] W.M. Wonham, "On a matrix Riccati equation of stochastic control",
SIAM J. Control 6 (1968), 681-697; "Erratum: On a matrix
Riccati equation of stochastic control", SIAM J. Control 1
(1969), 365.

Department of Mathematics,
Institute of Advanced Studies,
Australian National University,
Canberra, ACT.

https://doi.org/10.1017/S0004972700041071 Published online by Cambridge University Press