You are on page 1of 13

JOURNAL OF MATHEMATICAL ANALYSIS AND APPLICATIONS 213, 393]405 Ž1997.

ARTICLE NO. AY975517

A Power Method for Computing Square Roots of


Complex Matrices
Mohammed A. Hasan

Department of Electrical Engineering, Colorado State Uni¨ ersity,


Fort Collins, Colorado 80523

Submitted by Harlan W. Stech

Received August 22, 1995

In this paper higher order convergent methods for computing square roots of
nonsingular complex matrices are derived. These methods are globally convergent
and are based on eigenvalue shifting and powering. Specifically, it is shown for
each positive integer r G 2, a convergent method of order r can be developed.
These algorithms can be used to compute square roots of general nonsingular
complex matrices such as computing square roots of matrices with negative
eigenvalues. Q 1997 Academic Press

1. INTRODUCTION

A square root of a complex matrix A g C m= m is defined to be any


matrix B g C m= m such that B 2 s A, where C is the field of complex
numbers. If all eigenvalues of an m = m matrix A are distinct, then the
matrix equation X 2 s A generally has exactly 2 m solutions. This follows
from the fact that A is diagonalizable, i.e., there exists a similarity matrix
U such that A s UDUy1 , where D s diagŽ l1 , . . . , l m . and thus B s
Uy1 'D U, where 'D s diagŽŽy1. i1 l1 , . . . , Žy1. i m l m ., and i k s 0 or 1
' '
for k s 1, 2, . . . , m. However, if A has multiple eigenvalues, the number of
solutions will be different from 2 m as shown next. Let m s 2, then without
loss of generality, we can assume that A s l 02 or A s l 12 .
2 2

0 l 0 l
cos Ž u . sin Ž u .
Assume that A s l I, then the family l
2
½ sin Ž u . ycos Ž u . 5 0 F u - 2p forms an
l 2
infinite set of square roots of A. On the other hand, if A s 1
, then
0 l2

A has only two square roots given by "0l "1r2


"l
l
provided that l / 0.
Unlike the square roots of complex numbers, square roots of complex
393
0022-247Xr97 $25.00
Copyright Q 1997 by Academic Press
All rights of reproduction in any form reserved.
394 MOHAMMED A. HASAN

matrices may not exist. For example, when l s 0 in the last matrix no
square root exists. From this observation, it is obvious that for 2 = 2
matrices, the equation B 2 s A / 0 has a solution if and only if A has a
nonzero eigenvalue.
To understand the structure of solutions of the equation B 2 s A for
b b
m s 2, let B s b11 b12 , and A s aa1121 aa1222 such that B 2 s A. Then we
21 22

have the following four equations bi1 b1 j q bi2 b 2 j s a i j for i, j s 1, 2 which


are equivalent to F s 0, where

¡b q b12 b 21 y a11
11 b 11

~
F Ž b11 , b12 , b 21 , b 22 . s
b11 b12 q b12 b 22 y a12
Ž 1.
b 21 b11 q b 22 b 21 y a21
¢ b 21 b12 q b 22 b 22 y a22 .

The Jacobian of this system can be shown to be

2 b11 b 21 b12 0
­ F Ž b11 , b12 , b 21 , b 22 . b12 b11 q b 22 0 b12
Js s .
­ Ž b11 , b12 , b 21 , b 22 . b 21 0 b11 q b 22 b 21
0 b 21 b12 2 b 22

It can be verified that < J < s 4Ž b 11 q b 22 . 2 Ž b 11 b 22 y b 12 b 21 . s


4 Trace 2 Ž B .< B <. Here the notation < J < denotes the determinant of J, and
TraceŽ B . s Ý m < <
is1 bii . Since A is nonsingular, it follows that B / 0 and
therefore J is nonsingular if and only if b11 q b 22 / 0. Now assume that A
is nonsingular and let

0 0
b11 b12
B0 s 0 0
b 21 b 22

be a solution of the equation B 2 s A such that b11 0


q b 22
0
/ 0. Since J is
0 0 0 0 .
nonsingular at Ž b11 , b12 , b 21 , b 22 , it follows from the implicit function
theorem that B0 is the only solution. From the eigendecomposition of A
indicated before, one can see that there are at least four square roots of
A. The implicit function theorem guarantees exactly four square roots with
nonzero traces. These square roots of A which have nonzero traces are
referred to as functions of A w1x. Essentially, B is a function of A if B can
be expressed as a polynomial in A.
Now if b11 q b 22 s 0, then it follows from Ž1. that a11 q a22 s l2 ,
a12 s a21 s 0, i.e., A is diagonal of the form l2 I. In this case the equation
SQUARE ROOTS OF COMPLEX MATRICES 395

B 2 s A has a two-dimensional family of solutions given by

"'l2 y rs r
½ s .'l y rs
2 5 r, s g C.

Note that when r s s s l sinŽ u ., we get the one-parameter family


cos Ž u . sin Ž u .
½l sin Ž u . ycos Ž u . 5 0 F u - 2p which is described before.
The following result provides conditions on the eigenstructure of A
which ensure the existence of square roots which are functions of A.

PROPOSITION 1 w1x. Let A be nonsingular and its p elementary di¨ isors be


coprime, that is, each eigen¨ alue appears in only one Jordan block. Then A
has precisely 2 p square roots, each of which is a function of A.

Several computational methods of square roots of complex matrices


have been reported in the literature. In w2x, the Newton-Raphson method
was used for computing the principal square root of a complex matrix. An
accelerated algorithm for computing the positive definite square root of a
positive definite matrix was presented in w3x. A matrix continued fraction
method was presented in w4x. The matrix sign algorithm was developed in
w5x. A Schur method for computing square roots was developed in w6x. Fast
stable methods for computing square roots were also presented in w7, 8x.
It is noted in almost all of the above methods either a linear or
quadratic convergence can be obtained. In this paper, higher order conver-
gent methods of order r G 2 will be derived. The essence of these methods
is a process whereby a sequence of matrices which in the limit converges to
a square root of A is generated. This process involves creating gaps
between the magnitudes of eigenvalues of different square roots of A so
that for sufficiently high powers the eigenvalues will become decoupled.
This is similar in principle to well-known methods such as those of
Graeffe, Bernoulli, and the qd algorithm for solving polynomial equations
in that these methods are based on eigenvalue powering. For a survey of
some of these methods the reader is referred to w9, 10x and the references
therein.
Let S be a set of commutative and thus simultaneously diagonalizable
matrices. In the sequel, the notation l i Ž X . denotes the ith eigenvalue of
the square matrix X g S relative to a fixed similarity matrix which
diagonalizes the set S . The notation s Ž A. denotes the set of eigenvalues
of A. The symbol R is used to denote the set of real numbers and 5 A 5
denotes any vector norm of the matrix A.
396 MOHAMMED A. HASAN

2. DERIVATION OF THE MAIN RESULTS

In the next theorem we will generate a sequence which converges to a


square root of a square matrix.
THEOREM 2. Let A g C m= m be a nonsingular matrix. Let r be a positi¨ e
integer such that r G 2 and define A k and Bk recursi¨ ely as follow. Let
r
r
A kq 1 s Ý ž /Ary2
k
l 2l l
Bk A , Ž 2.
ls0
2l

and
r
r
Bkq 1 s Ý ž Ary2/ly1 2 lq1 l
Bk A . Ž 3.
ls0
2l q 1 k

Then there exists an a g C such that Bk is nonsingular for all sufficiently large
k. Set X k s Bky1A k , then the initial guess A 0 s aIm and B0 s Im the sequence
X k con¨ erges to a square root W of A. Moreo¨ er,
r
X kq 1 " W s By1
kq1 Bk Ž X k " W . ,
r
Ž 4.
i.e., if the sequence X k con¨ erges, it is rth order con¨ ergent to W. Addition-
ally,

lim k ª` Ay1
kq1 A k s I
r
and y1
lim k ª` Bkq1 Bkr s I.

Proof. Let W be any square root of A, i.e., W 2 s A and show by


induction that
k
A k " BkW s Ž aI " W . r . Ž 5.
Clearly Ž5. holds for k s 0. Assume that Ž5. holds for the positive integer
k. Then
r
kq 1 r r
Ž aI " W . r s Ž A k " BkW . s Ý ž / A ry2
k
l 2l l
Bk A
ls0
2l
r
r
" Ýž A ry2
/ ly1 2 lq1 l
Bk A W
ls0
2l q 1 k

s A kq 1 " Bkq1W ,

where the last equality follows from Ž2. and Ž3.. Hence Ž5. is true for the
integer k q 1. This shows that Ž5. is true for each nonnegative integer k.
SQUARE ROOTS OF COMPLEX MATRICES 397

The nonsingularity of A implies that there exists an a g C such that


< l j Ž aI q W .< ) < l j Ž aI y W .<, for j s 1, . . . , m. From Ž5. we have Ž aI q
k k
W . r s A k q BkW and Ž aI y W . r s A k y BkW. Solving the last two
equations for A k and Bk yields
1 k k
Ak s  Ž aI q W . r q Ž aI y W . r 4 , Ž 6.
2
and
1 k k
Bk s  Ž aI q W . r y Ž aI y W . r 4 Wy1 . Ž 7.
2
Note that the order of matrix multiplication is immaterial in this case since
the choice X 0 s aI implies that the elements of the sequence  X k 4`ks1
commute with each other. It should also be noted that for sufficiently large
k both A k and Bk are invertible since < l j Ž aI q W .< ) < l j Ž aI y W .<, j s
1, . . . , m. Thus multiplying both sides of Ž5. by By1
k yields
y1
k y1 k
By1 y1
k A k q W s Bk Ž aI q W .
r
s 2W I y  Ž aI q W . ½ Ž aI y W . r 4
y1 rk y1 2rk
s 2W I q Ž aI q W .
½ Ž aI y W . 4 q  Ž aI q W . Ž aI y W . 4 q??? . 5
Since < l i Ž aI q W .< ) < l i Ž aI q W .<, for i s 1, . . . , m, it follows from the
last equation that
k
lim By1 y1
k A k q W s lim Bk Ž aI q W .
r
s 2W.
kª` kª`

Therefore, lim k ª` Bky1A k


s W.
To prove Ž4. we have from the relation A kq 1 " Bkq1W s Ž A k " BkW . r
that
r
By1 y1 y1
kq 1 A kq1 " W s Bkq1 Bk Ž Bk A k " W . ,
r

or equivalently
r
X kq 1 " W s By1
kq1 Bk Ž X k " W . .
r

Finally, the last conclusion follows from


kq 1 kq 1 y1 k k r
By1
kq 1 Bk s  Ž aI q W .
r r
qŽ aI y W . r 4  Ž aI q W . r q Ž aI y W . r 4
y1
y1 r kq 1
s ½ ž I y Ž Ž aI q W . Ž aI y W . . 5
r
y1 rk
½ ž I y Ž Ž aI q W . Ž aI y W . . / ªI
as k ª `. Q.E.D.
398 MOHAMMED A. HASAN

An analogous algorithm to that of Theorem 2 can be obtained by setting


X k s Bky1A k as shown next.
THEOREM 3. Let A be a nonsingular matrix and let r be a positi¨ e integer
such that r G 2. Let
r
r
Ck s Ý ž /X kry2 lAl , Ž 8.
ls0
2l
r
r
Dk s Ý ž X kry2 ly1Al ,
/ Ž 9.
ls0
2l q 1

and set X kq 1 s Dy1


k C k . Then there exists an initial matrix X 0 s aI g C
m= m
,
`
a g C for which the sequence  X k 4ks1 con¨ erges to a square root W of A.
Moreo¨ er,
r
X kq 1 " W s Dy1
k Ž Xk " W . , Ž 10 .
i.e., if the sequence X k con¨ erges it is rth order con¨ ergent to W.
Proof. Theorem 3 follows directly from Theorem 2 by setting X k s
Bky1A k . Q.E.D.
In Theorems 2 and 3 and if r s 2, quadratically convergent algorithms
are obtained as follows.
COROLLARY 4. Let A be an m = m nonsingular complex matrix and
define the following sequence A kq 1 s A2k q Bk2 A, and Bkq1 s 2 A k Bk with
A 0 s aI and B0 s I. Then for some a g C , the sequence Bky1A k con¨ erges
quadratically to 'A . Moreo¨ er, if we set X k s By1 k A k , then the iteration
X kq 1 s X k y 12 Xy1
k
Ž X k2 y A. with X 0 s aI con¨ erges quadratically to some
'A and X kq 1 y 'A s 12 Xy1 k
Ž X k y 'A . 2 .
Remark. The iteration X kq 1 s X k y 12 Xy1 k
Ž X k2 y A. of the above
corollary is exactly the Newton iteration for solving X 2 y A s 0. Several
variants of Newton’s method and their implementations were presented in
w2x, where the matrix A is assumed diagonalizable. Note that the algorithm
of Corollary 4 and the Newton method are equivalent. However, the
derivation of the above algorithm lends itself to the development of other
methods whose convergence is of any prescribed order.
A cubically convergent algorithm can be derived by setting r s 3 in
Theorems 2 and 3 as in the next result.
COROLLARY 5. Let A be an m = m nonsingular complex matrix and
define the following sequence A kq 1 s A3k q 3 Bk2 A k A and Bkq1 s 3 A2k Bk q
Bk3 A, with A 0 s aI and B0 s I. Then for some a g C , the sequence Bky1A k
SQUARE ROOTS OF COMPLEX MATRICES 399

con¨ erges cubically to a square root of A. Moreo¨ er, if we set X k s Bky1A k


then the sequence satisfies the recursion X kq 1 s X k y 2Ž3 X k2 q A.y1 Ž X k2 y
A. which with the initial guess X 0 s aI con¨ erges cubically to 'A and
X kq 1 y 'A s 2Ž3 X k2 q A.y1 Ž X k y 'A . 3.

3. ANALYSIS OF CONVERGENCE

To analyze the convergence of the sequence  X k 4`ks1 we formulate the


iteration in Theorem 2 as a fixed point iteration as established in the next
result.
THEOREM 6. Assume there exists an a g C such that < l i Ž aI y
'A .rli Ž aI q 'A .< - 1,for i s 1, . . . , m. Then the sequence X k defined in
Theorem 3 is rth order con¨ ergent, i.e.,
r
X kq 1 y 'A s O Ž X k y 'A . ,
or equi¨ alently, there exists a constant S G 0 such that
5 X kq 1 y 'A 5 F S 5 X k y 'A 5 r .

Proof. Let Z g C m= m be any square matrix of A and define C Ž Z . s


Ý rls0 2rl Z ry2 lAl and DŽ Z . s Ý rls0 2 l q
ž / ž r
1
Z ry2 ly1Al. Then
/
r
C Ž Z . " D Ž Z . 'A s Ž Z " 'A . .
Define F Ž Z . s DŽ Z .y1 C Ž Z .. Then F: C m= m ª C m=m and
r y1 r
r ry2 ly1 r ry2 l
F Ž "'A . s ½Ýž ls0
2l q 1 / Ž "'A . Al 5 ½ Ý ž /Ž
ls0
2l
"'A . Al 5
r y1 y1 r
s 2 ry1 Ž "'A . Ž "'A .
½ 5 ½2 ry1
Ž "'A . 5 s "'A .
Therefore 'A and y 'A are fixed points for the function F. Now, the
function F can be written as
y1
F Ž Z . s .'A q D Ž Z . Ž C Ž Z . " D Ž Z . 'A .
y1 r
s .'A q D Ž Z . Ž Z " 'A . ,
hence F Ž Z . " 'A s DŽ Z .y1 Ž Z " 'A . r. Note that when Z is a scalar it
can be shown
d y1 ry1
F Ž Z . s rD Ž Z . Ž Z 2 y A. .
dZ
400 MOHAMMED A. HASAN

Therefore if j 2 y A s 0, then F Ž j . s j and Ž d lrdZ l . F Ž j . s 0, for


l s 1, 2, . . . , r y 1. It follows that there is an R ) 0 and a neighborhood,
SR Ž j . s  Z g C m= m < 5 Z y j 5 - R4 , of j and a constant K, 0 F K - 1
such that

FŽ Z. y FŽ j . F K 5 Z y j 5

for all Z g SR Ž j . [  Z < g C m= m 5 Z y j 5 F R4 Ži.e., F is a contractive


mapping in SR Ž j ... Hence for any X 0 g SR Ž j . the generated sequence
X kq 1 s F Ž X k ., k s 1, 2, ??? has the properties X k g SR Ž j . and 5 X k y j 5
F K k 5 X 0 y z 5 for k s 1, 2, ??? . This shows that the iteration X kq1 s
F Ž X k . converges to j for any initial guess X 0 in SR Ž j . for which
X 0 A s AX 0 . To prove the last conclusion of the theorem, we have

r
Ck " D k'A s Ž X k " 'A . , Ž 11 .

where C k [ C Ž X k . s Ý rls 0 r
ž /X
2l
ry 2 l l
k A, and Dk [ DŽ Xk . s
Ý rls0 2 l q
ž r
/ 1
X kry2 ly1Al. Since DŽ'A . s Ž2'A . ry1 is nonsingular, D k is
nonsingular for all sufficiently large k. It follows that there exists a
constant S G 0 such that 5 Dy1 k
5 F S for all sufficiently large k. Therefore
5 X kq 1 y 'A 5 F S 5 X k y 'A 5 r. This proves the r th order convergence of
Xk . Q.E.D.

4. CONTINUED FRACTION EXPANSION

There are many formulas in the literature that approximate square roots
of numbers and matrices using continued fraction expansion w4x. In this
section, we express the results of Theorem 3 in continued fraction form.

THEOREM 7. Let A be a nonsingular matrix and let W be a square root of


A. Then for each a g C for which < l i Ž aI q W .< ) < l i Ž aI y W .<, i s
1, . . . , m, W has the representation

W s aI q Ž A y a2 I . 2 aI q Ž A y a2 I .
½
y1 y1
 2 aI q Ž A y a2 I .  2 aI q ??? 4 y1 4 5 . Ž 12 .
SQUARE ROOTS OF COMPLEX MATRICES 401

It can easily be verified that the r th order truncation of the continued


fraction of Ž12. yields DŽ aI .y1 C Ž aI ., where C Ž aI . s Ý rls0 2rl a ry2 lAl , and
ž /
DŽ aI . s Ý rls0 2 l q
ž r
1 /a ry2 ly1 l
A . When A is a scalar, Formula Ž12. reduces
to

A y a2
'A saq , Ž 13 .
2 a q Ž A y a2 . r Ž 2 a q Ž A y a2 . r Ž 2 a q ??? . .

which in the case a s 1 becomes that of w4x. Note that when A is positive,
Ž13. converges for any a g C with nonzero real part. The free parameter
a provides some flexibility in choosing the initial guess in that the closer a
is to 'A the more accelerated is the convergence.

Remark on the Choice of the Initial Guesses. In this remark we provide


special cases where conditions on the initial matrix X 0 can be imposed to
ensure convergence. Assume that all eigenvalues of A are not on the
negative real line. Then each eigenvalue of 'A has nonzero real part. The
initial guess A 0 s aI with a ) 0 Ž a - 0. forces the sequence X k to
converge to a square root of A with eigenvalues having positive Žnegative.
real part. However, if some of the eigenvalues of 'A are on the imaginary
axis, then the choice X 0 s Ž a q ib . I, where a ) 0 and b ) 0, will lead to
a sequence which converges to a square root of A having eigenvalues with
nonnegative real parts. These observations show in particular that if A is
positive Žor negative definite. then the methods of Theorems 2 and 3
converge for any initial guess X 0 of the form X 0 s aI, a g R Ž ia g R . ,
where i s 'y 1 . This is an improvement over the algorithm in w2x which is
applicable only to matrices which have no negative eigenvalue.

5. COMPUTATIONAL RESULTS

To demonstrate the performance of the algorithms proposed in this


paper we consider in this section some examples to show the behavior of
some of these methods in finite-precision arithmetic. These computations
were carried out on approximately seven decimal digit accuracy. For the
purpose of comparison we will apply Theorem 3 to four examples. The
notation X k, r will denote the computed square root using k iterations and
r th order method. The error is measured in the Frobenious norm 5 X k,2 r y
A5 F .
402 MOHAMMED A. HASAN

EXAMPLE 1. Consider the complex 3 = 3 matrix

8 q 15i y1 y 3i y4 y 9i
3 3 3
y1 y 3i 5 q 9i y1 y 3i
As .
3 3 3
y4 y 9i y1 y 3i 8 q 15i
3 3 3

Applying five iterations of the algorithm of Theorem 3 with r s 2, X 0 s


Ž1 q i . I, yields

1.938066 q 1.123146i y0.2334078 q y 0.2188987i


X5, 2 s y0.2334079 q y 0.2188985i 1.565499 q 0.8928874i
y0.6059740 q y 0.4491569i y0.2334077 q y 0.2188986i
y0.6059739 q y 0.4491573i
y0.2334077 q y 0.2188989i ,
1.938066 q 1.123145i

which agrees with the exact square root to five decimal places, i.e.,
5 X5,2 2 y A 5 F s O Ž10y6 .. Comparable accuracy can also be achieved using
only four iterations with a third order method with the same initial guess
in which case we obtain 5 X 4,2 3 y A 5 F s O Ž10y6 ..
EXAMPLE 2. Consider the 4 = 4 matrix w2x

5 4 1 1
4 5 1 1
As .
1 1 4 2
1 1 2 4

The eigenvalues of this matrix are lŽ A. s  1, 2, 5, 104 and 2-norm condi-


tion number k 2 Ž A. s 10. When X 0 s I and 4 iterations are used with
r s 3, we obtain

1.988520 0.9885167 0.1852414 0.1852417


0.9885170 1.988519 0.1852422 0.1852421
X 4, 3 s ,
0.1852420 0.1852420 1.917762 0.5035480
0.1852418 0.1852418 0.5035481 1.917762

and 5 X 4,2 3 y A 5 F s O Ž10y6 .. We also note that 5 X 4,2 2 y A 5 F s O Ž10y3 .


while 5 X5,2 2 y A 5 F s O Ž10y7 ..
SQUARE ROOTS OF COMPLEX MATRICES 403

EXAMPLE 3. Consider the 4 = 4 near singular matrix

0 .07 .27 y.33


1.31 y0.36 1.21 0.4
As .
1.06 2.86 1.49 y1.34
y2.64 y1.84 y0.24 y2.01

The eigenvalues of this matrix are lŽ A. s  0.03, 3.03, y1.97 " i4 . Apply-
ing the iteration of Theorem 3 with r s 3 and X 0 s Ž1 q i . I we obtain

0.2441971 y9.0401828 E y 02 0.1997281


1.317404 1.181676 0.2569374
X 7, 3 s
1.0659751 E y 02 0.1509080 1.370772
y0.6756389 y1.982304 0.3442460
y8.5186012 E y 02
0.8455525
,
y1.248934
y0.1964209

with error 5 X 7,2 3 y A 5 F s O Ž10y4 .. However, when r s 2, similar accuracy


can be obtained with the same X 0 as

0.2437900 y9.0409003E y 02 0.1997136


1.317573 1.181780 0.2569094
X6, 2 s
1.0955708 E y 02 0.1507968 1.370820
y0.6754091 y1.982364 0.3442680
y8.5148215E y 02
0.8455746
,
y1.249036
y0.1964408

with error measured in Frobenious norm as 5 X6,2 2 y A 5 F s O Ž10y4 ..


EXAMPLE 4. Consider the 3 = 3 matrix w2x

1 1 y2
A s y1 2 1 ,
0 1 y1

which has the eigenvalues lŽ A. s  y1, 1, 24 , i.e., this matrix has a nega-
tive eigenvalue. Thus the initial matrix X 0 s aI should be chosen so that
404 MOHAMMED A. HASAN

the imaginary part of a is nonzero. Using X 0 s Ž1 q i . I, we obtain

1.0285987 y 0.1666667i 0.4714045 y 0.3333333i


X 9, 3 s y0.4142135 y 0.0000000i 1.4142135 q 0.0000000i
y0.0285954 q 1.1666666i 0.4714045 y 0.3333333i
y1.0285954 q 1.16666667i
0.4142135 q 0.0000000i ,
y0.0285954 q 1.1666666i

with error 5 X5,2 2 y A 5 F s O Ž10y1 0 .. If a method of order three is used,


then 5 X 3,2 3 y A 5 F s O Ž10y8 ..
In summary, some square roots of nonsingular complex matrices can be
computed using only initial matrices of the form X 0 s Ž a q ib . I. The case
b / 0 is required for matrices having negative eigenvalues.

6. CONCLUSION

A new set of algorithms for computing square roots of nonsingular


complex matrices was developed. Given any positive integer r G 2 we
presented a systematic way of deriving an r th order convergent square
root algorithm. The convergence and its speed are largely affected by the
ratios R i s < l i  X 0 y 'A 4<r< l i  X 0 q 'A 4< - 1, i s 1, . . . , m, which are ob-
viously dependent on the initial guess X 0 . Thus X 0 serves as a free
parameter for which a faster convergence can be achieved by choosing X 0
so that the R i ’s are closer to zero. It should be observed that the
conclusion of Theorem 3 is still valid if X 0 s aI is replaced by any other
m = m matrix provided that X 0 and A commute. For r s 2, this tech-
nique becomes the Newton method for solving the equation X 2 s A. One
shortcoming of these algorithms is that they cannot be applied to compute
all square roots of A using only initial guesses of the form X 0 s aI,
a g C . In order to obtain all square roots, a matrix sign function of all
square roots is to be computed, i.e., for each square root 'A , a matrix S
such that S 2 s I and S'A s 'A S must be generated and then use S as
an initial guess in Theorems 2 or 3. Finally, this work can be generalized to
compute the nth roots of complex matrices. This is the subject of a paper
submitted for publication w11x.

ACKNOWLEDGMENT

The author thanks the referees for their helpful remarks and suggestions which improved
the quality of this work.
SQUARE ROOTS OF COMPLEX MATRICES 405

REFERENCES

1. N. J. Higham, Computing real square roots of a real matrix, Linear Algebra Appl. 88r89
Ž1987., 405]430.
2. N. J. Higham, Newton’s method for the matrix square root, Math. Comp. 46, No. 174
Ž1986., 537]549.
3. E. D. Denman, Roots of real matrices, Linear Algebra Appl. 36 Ž1981., 133]139.
4. L. S. Shieh and N. Chahin, A computer-aided method for the factorization of matrix
polynomials, Appl. Math. Comput. 2 Ž1976., 63]94.
5. E. D. Denman and A. N. Beavers, The matrix sign function and computation of systems,
Appl. Math. Comput. 2 Ž1976., 63]94.
6. A. Bjorck and S. Hammarling, A Schur method for the square root of a matrix, Linear
Algebra Appl. 52r53 Ž1983., 127]140.
7. W. D. Hoskins and D. J. Walton, A fast method of computing the square root of a matrix,
IEEE Trans. Automat. Control AC-23, No. 3 Ž1978., 494]495.
8. W. D. Hoskins and D. J. Walton, A fast, more stable method for computing the pth root
of positive definite matrices, Linear Algebra Appl. 26 Ž1979., 139]163.
9. P. Henrici, ‘‘Applied and Computational Complex Analysis,’’ Vol. 1, Wiley, New York,
1974.
10. A. S. Householder, ‘‘The Numerical Treatment of a Single Nonlinear Equation,’’ Mc-
Graw]Hill, New York, 1970.
11. M. A. Hasan, Higher order convergent algorithms for computing nth roots of complex
matrices, submitted for publication.