See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/222497305

Normal matrices: An update

Article  in  Linear Algebra and its Applications · December 1998

DOI: 10.1016/S0024-3795(98)10161-1

CITATIONS READS
58 643

2 authors:

Ludwig Elsner Khakim Ikramov

Bielefeld University Lomonosov Moscow State University
159 PUBLICATIONS   2,514 CITATIONS    469 PUBLICATIONS   1,226 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Numerical methods for large matrix eigenproblems View project

All content following this page was uploaded by Khakim Ikramov on 09 November 2017.

The user has requested enhancement of the downloaded file.

 NORMAL MATRICES: AN UPDATE

L. Elsner
Fakultat fur Mathematik
Universitat Bielefeld
Postfach 100131

33501 Bielefeld
FRG
Kh. D. Ikramov1
Faculty of Computational Mathematics
and Cybernetics
Moscow State University

119899 Moscow
Russia
Abstract. A list of seventy conditions on an n {by{ n complex matrix A , equi-
valent to its being normal, published nearly ten years ago by Grone, Johnson, Sa,
and Wolkowicz has proved to be very useful. Hoping that, in an extended form, it
will be even more helpful, we compile here another list of about twenty conditions.
They either have been overlooked by the authors of the original list or have appeared
during the last decade.

1 Research supported by Sonderforschungsbereich 343 Diskrete Strukturen in der Mathematik

2
1. Introduction.
A square complex matrix A is called normal if
(1) AA = AA
where A is the adjoint of A . If this relation is taken as condition 0, then the
seventy conditions 1{70 equivalent to (1) constitute Section II of [6].
At the end of the introduction to their list, the authors of [6] say: "Re ecting the
fact that normality arises in many ways, it is hoped that not only will it be useful
now, but its utility will grow over time as conditions are added."
Nearly a decade after [6] was published, one can say that it (called henceforth the
GJSW list) has indeed proved to be very useful. It is perhaps the right time to
increase its utility, as the authors suggest, by supplementing the original list with
additional conditions equivalent to the normality of A .
Section 2 of this paper contains nearly twenty such conditions. They either have
been overlooked by the authors of [6] or have appeared during the last decade. Since
our list is meant to be a continuation of that in [6], our rst condition is numbered
71. The outline of proofs and/or comments to most of the conditions are given in
Section 3. In nearly all cases, we point to a publication where the corresponding
condition has rst appeared. Our proofs are, up to minor details, the original proofs
in these publications. As in [6], proofs in an obvious direction are omitted.
We conclude this section by introducing the notation, which does not in all cases
coincide with that of [6]. Also, those conditions in the original GJSW list are given
explicitly which are used in the body of our paper.
Our matrices belong (almost) invariably to the set C n;n of complex nn {matrices.
It is sometimes benecial to regard C n;n as a unitary space equipped with the inner
product
(2) < A; B >= tr (B A)
where tr(
) stands for the trace of a matrix.
Cn denotes the set of all complex column vectors with n components. If necessary,
Cn is also considered as a unitary space with the usual inner product ( ; ) . The
symbol jj
jj is used for the 2{norm of a vector and, if not stated otherwise, for the
spectral norm of a matrix. For A nonsingular, the condition number
(3) cond A = jjAjj jjA?1jj
is always meant to be the spectral one. If  is a simple eigenvalue of A , x and y
corresponding unit right and left eigenvectors, then
(4) () = j(x; y)j?1
3
is called the condition number of  .
We denote by (A) and 1(A)  : : :  n (A) the spectral radius of A and
its singular values. The commutator [B; C ] = BC ? CB is sometimes called the
self{commutator of B , if C = B  . The matrices
(5) jAj = (A A)1=2; jA j = (AA)1=2
are called the right and left absolute value of A , respectively.
In addition to well{known matrix decompositions, such as the QR or the singular
value decomposition, we also use some that are encountered less frequently. These
are the Toeplitz decomposition
(6) A = H + iK; H = H  ; K = K 
with
(7) H = 21 (A + A ); K = 21i (A ? A)
and the polar decompositions
(8) A = U1 jAj = jAjU2
with unitary matrices U1 and U2 . Note that, for a nonsingular A , U1 and U2
are dened uniquely and are, in fact, the same unitary matrix. If A is singular
then there exist (innitely) many U1 ; U2 satisfying (8). However, one can choose
them to be the same matrix in this case as well [11, p. 152].
For our purposes, matrices B and C are said to be simultaneously diagonalizable,
if
(9) B = U 1 V  ; C = U 2 V 
with unitary U; V and diagonal 1; 2 . Note that, as opposed to the singular
value decomposition, 1 and 2 are not required to be nonnegative.
The symbol Ck (A) stands for the k {th compound matrix of A [14, p. 16].
Now we list some conditions from [6]:
11. There exist unitary U and diagonal D such that A = UDU  .
14. There exist an orthonormal basis of C n consisting of eigenvectors of A .
37. If A = U jAj is the (right) polar decomposition of A , then U and jAj
commute.
38. A and jAj commute.
58. The singular values 1  : : :  n of A coincide with the moduli of its
eigenvalues.
59. If the eigenvalues 1; : : : ; n of A are numbered according to
(10) j1j  j2j  : : :  jn j;
then
1 : : : k = j1 : : : k j; k = 1; : : : ; n:

4
2. Conditions.
71 [3]. jAj = jA j .
72 [3]. A is diagonalizable:
(11) A = X X ?1;  = diag (1; : : : ; n )
with X such that cond X = 1 . (Or: for any eigenvalue  of A , the
condition number () = 1 , as long as A has distinct eigenvalues).
73 [14]. A commutes with [A; A] .
74 [8]. A commutes with A A (or AA ).
75 [9]. In the Toeplitz decomposition of A ,
H 2 + K 2 = AA (or AA):
76 [16]. tr (A2A2 ) = tr [(AA)2]:
77 [16]. (Generalization of condition 76)
tr (ApAp ) = tr [(AA)p] for some positive integer p  2:
78 [16]. (Generalization of condition 77)
tr [(ApAp )q ] = tr [(AA)pq ] for some positive integers p  2; q  1:
79. etn A is normal for a sequence ftn g , converging to zero.
80 [16]. tr (eA eA ) = tr (eA +A ):
81 [1,5]. The function fx (t) = ln jjetA xjj is convex on R for any vector x 2 C n .
82 [1]. The functions fj (t) = ln j (etA ); j = 1; : : : ; n , are convex on R .
83 [7]. A is diagonalizable as in (11), and a nonzero perturbation B exists such
that for some eigenvalue 0 of the perturbed matrix A + B the closest
eigenvalue 0 of A is uniquely dened, and
(12) j0 ? 0 j = jjB jj cond X:
84 [12]. j(Au; u)j  (jAju; u) 8 u 2 C n :
85 [12]. j(Au; u)j  (jAju; u) 8 u 2 C n :
86 [12]. (Generalization of conditions 84 and 85)
(13) j(Au; u)j  (jAju; u) (jAju; u)1? 8 u 2 C n
for some real  6= 21 .
87 [13]. If A = QR is the QR{decomposition of A then Q and R are simulta-
neously diagonalizable.
88. For k = 1; 2; : : : ; n ? 1
(14) jjCk (A)jj = (Ck (A)):
89. The matrix Lyapunov operator on C n;n dened by the formula
(15) LAX = AX + XA 8 X 2 C n;n
is normal.
5
3. Proofs and comments.
Condition 71 is immediate from (1) and (5), and condition 72 from condition 11,
if one takes into account that matrices with the spectral condition number 1 are
scalar multiples of unitary matrices. The second version of condition 72 implies that
A has a full orthonormal set of eigenvectors, hence condition 14 is applicable.
For the suciency part of condition 73, we rst mention that under the unitary
similarity
A ! U  AU
the self{commutator C = [A; A] transforms in the same way as A does:
C ! U CU:
Suppose C is nonzero. Then no generality is lost in assuming C to be in block
diagonal form
(16) C = C11  : : :  Cmm
with diagonal blocks C11; : : : ; Cmm corresponding to the distinct eigenvalues
1 (C ); : : : ; m(C ) of C . Obviously, m  2 because tr C = 0 . Since A and C
commute, the former must assume the same block diagonal form as the latter:
A = A11  : : :  Amm:
Then
Cii = [Aii ; Aii]; i = 1; : : : ; m;
which implies
tr Cii = 0; i = 1; : : : ; m;
and
i (C ) = 0; i = 1; : : : ; m:
This is a contradiction to the way by which the decomposition (16) has been intro-
duced.
Condition 74 insures that A commutes with any polynomial of AA . Since jAj is
one of those polynomials, condition 38 may be employed.
Condition 75 is an immediate consequence of the identity
H 2 + K 2 = 21 (AA + AA)
which holds for any square A .
To prove the suciency of condition 76, one can use the identity
jjAA ? AAjj2F = 2ftr [(AA)2 ] ? tr (A2A2)g
6
(this is Lemma 4.3 in [16]) that is again valid for any square A . The symbol jj
jjF
stands for the Frobenius norm of a matrix.
Assume that p > 2 in condition 77 (otherwise, it coincides with condition 76). The
idea is to show that condition 77 with the exponent p implies the same condition
with the exponent p ? 1 . Then, in a nite number of steps, we arrive at condition
76 assuring the normality of A .
Let r be the rank of A , then, among its singular values 1  2  : : :  n , only
the rst r are nonzero. Letting H = jAj , we see that 1; : : : ; n are the eigenva-
lues of H . Denote the corresponding orthonormal eigenvectors by v1 ; : : : ; vn , and
introduce
(17) S (p) = tr [(AA)p] ? tr (ApAp )
and t
X
D(t; s) = (i2 ? t2) (i2s ? jjAsUvi jj2):
i=1
Here U is the unitary factor in the polar decomposition of A .
It follows from [4] that
(18) S (p)  0 for any positive integer p
and
D(t + 1; s)  D(t; s) for positive integers t; s(t  n):
Hence
D(t + 1; s)  D(t; s)  D(t ? 1; s)  : : :  D(1; s) = 0;
and D(t; s) is always nonnegative. Now one can prove the inequality (this is essen-
tially Theorem 4.4 in [16])
S (p)  r2 S (p ? 1) + D(r; p ? 1):
We conclude that S (p) = 0 forces S (p ? 1) = 0 .
Turning to condition 78, assume that q > 1 (the case q = 1 gives us condition
77). Denote by tri (H ) the sum of the i largest eigenvalues of an Hermitian matrix
H . Then an inequality stronger than (17) { (18) is proved in [4], namely
(19) tri (Ap Ap)  tri [(AA)p] for any positive integer p and i = 1; 2; : : : ; n:
Inequalities (19) can be interpreted as the weak majorization relations between the
vectors
(12(Ap); 22 (Ap); : : : ; n2 (Ap))
and
(12p; 22p; : : : ; n2p):
7
Then, using the usual majorization techniques (see, for example, [15, p. 115]), one
can easily show that
(20) tr[(ApAp )q ]  tr[(AA)pq ]
for any matrix A and positive integers p; q . Moreover, equality in (20) implies
that
(21) tr(Ap Ap) = tr[(AA)p]
(this is Theorem 4.2 in [16]). In other words, condition 78 with q > 1 implies
condition 77.
For condition 79, the suciency is obvious from the relation
A = tlim (etn A ? I )=tn:
n !0
The proof of condition 80 relies on the following product exponential formula which
is valid for any n  n matrices X and Y :
X Y
(22) lim (e m e m )m = eX +Y :
m!1
This formula is proved in [2]. As a consequence of (22),
X
!1 tr[(e
mlim
m e mY )m] = tr(eX +Y ):
Using repeatedly (20) with p = 2; q = 1; 2; 3; : : : and A replaced by eA=2q , one
can show that, for any square matrix A , the sequence
h  q q 2 qi
tr eA =2 eA=2
is monotonically increasing to tr(eA +A ) (Theorem 4.6 in [16]).
Now condition 80 implies that
 h  q q 2 qi
tr(eA eA ) = tr eA =2 eA=2
for all positive integers q . This
q
is nothing else than equality (21) with p = 2q and
A replaced by Bq = eA=2 . According to condition 77, Bq must be normal. It
remains to apply condition 79.
For condition 81, the necessity part consists of checking, for any t1; t2 2 R ; t1 < t2 ,
the inequality
t + t2   fx (t1 ) + fx (t2 )
(23) fx 1
2 2
8
that amounts to
t +t
(24) jje 1 2 2 A xjj2  jjet1 A xjj jjet2A xjj:
Without loss of generality, one can assume the normal matrix A in (24) to be
diagonal: A = diag(1; : : : ; n ) . Letting
x = (x1; : : : ; xn)T ;
we then rewrite (24) as
n  t +t
X 2 n 
X n 
2 1=2 X 2 1=2
 1 2 2 i  t1 i  t2 i
(25) e xi   e xi  e xi  :
i=1 i=1 i=1
Since t +t
je 1 2 2 i xi j2 = je(t1 +t2 )i x2i j = jet1 i xi j jet2 i xi j;
relation (25) is a consequence of the Cauchy{Schwarz inequality.
Thus, we proved (23).
If condition 81 holds, then the function
f (t) = log jjetA jj = sup fx (t)
jjxjj=1
is convex on R . Letting t1 = 0; t2 = 2t (t > 0) in
t + t2   f (t1) + f (t2) ;
f 1
2 2
we have
jjetA jj  jje2tAjj1=2:
On the other hand,
jje2tAjj = jjetA
etA jj  jjetAjj2
hence
jjetA jj = jje2tAjj1=2:
By induction, we obtain
jjetA jj = jje2k tA jj2?k ; k = 1; 2; : : :
Then the spectral radius formula implies
(26) jjetAjj = (etA ):

It is known [11] that, for the matrix Bt = etA , the equality jjBtjj = (Bt) ensures
an orthogonal decomposition of the underlying space
CN = H1  H2 ;
9
with the projection of Bt on H1 being a scalar matrix. Then the argument above
can be applied to H2 . Continuing in this way, we nally obtain that Bt is normal.
Since this holds for any positive t , the normality of A follows from condition 79.
We mention that condition 81 has rst appeared in [5]. However, the proof above
is adapted from [1].
Turning to condition 82, we observe that its necessity can be established with the
help of the previous condition. In fact, let A be a normal matrix with the ortho-
normal eigenvectors v1 ; : : : ; vn . Then, for any real t , the matrix etA is normal as
well, and (see condition 58), with the appropriate numbering,
(27) j (etA ) = jj (etA )j; j = 1; : : : ; n:
Letting x = vj in the denition of fx (t) , we have
fvj (t) = ln jjetA vj jj = ln jj (etA )j = ln j (etA ) = fj (t):
Hence, fj (t) must be convex on R (j = 1; : : : ; n) .
Since
jjB jj = 1 (B )
for any matrix B , the proof of suciency can mimick that of condition 81. Another
possibility chosen in [1] is to prove equalities (27) by using a result from [17]. Then,
again, conditions 58 and 79 may be applied.
Condition 83 is obviously connected with the Bauer{Fike theorem. It is useful to
precede the proof of Condition 83 by some discussion of this theorem.
Let A; ; X be as in relation (11), and B an arbitrary perturbation of A . Ac-
cording to the Bauer{Fike theorem, for any eigenvalue  of the perturbed matrix
A + B there exists an eigenvalue  of A such that
(28) j ? j  cond X
jjB jj:
For A normal, X can be chosen unitary, and (28) turns into
(29) j ? j  jjB jj:
Now, again for a normal A , it is always possible to perturb A along its eigendi-
rections in such a way that (29) becomes an equality for at least some  's, and for
at least one of those  's (say 0 ) the corresponding nearest (possibly, multiple)
0 is dened uniquely.
Condition 83 found by O. Hald in [7] essentially says that the property above is
characteristic for normal matrices. Now we turn to the proof of the suciency of
this condition.
10
Let x be an eigenvector of A + B corresponding to 0 :
(30) (A + B )x = 0 x:
Using (11), one can rewrite (30) as
(31) ( + X ?1BX ) T ?1 x = 0 T ?1 x:
By a proper normalization of x , we can achieve that v = T ?1 x is a unit vector.
Now (31) assumes the form
(32) ( ? 0 I ) v = ?T ?1 BT v:
Applying (12), we deduce from (32)
cond X
jjB jj = min
j
jj ? 0j  jj( ? 0 I )vjj  cond X
jjB jj:
Hence
jj( ? 0 I )vjj = min
j
jj ? 0j:
Since the closest  (say, 0 ) is unique, v must be an eigenvector of  ? 0 I :
( ? 0 I )v = (0 ? 0 )v
which, by (32), is equivalent to
(33) Bx = (0 ? 0 )x:
Finally, using (12) and (33), we have
(34) cond X
jjB jj
jjxjj = j0 ? 0j
jjxjj  jjB jj
jjxjj:
Recall that cond X  1 for any nonsingular matrix X . Therefore, (34) implies
cond X = 1;
and the normality of A follows from condition 72.
We mention a nice corollary of condition 83 also given in [7]:
if A is not normal, then the Bauer{Fike inequality is strict for small perturbati-
ons B . Indeed, if 2 cond X
jjB jj is smaller than the minimal distance between
eigenvalues of A , then, for each  , the closest  (simple or multiple) is dened
uniquely. Since A is nonnormal, equality in (28) is impossible.
Conditions 84, 85 are two particular cases of condition 86 found in a recent pu-
blication [12]. They have been singled out just for the reason that a much simpler
proof is possible for these cases compared with that for a basic condition.
11
The necessity part is the same for all three conditions. Let A = U jAj be the polar
decomposition of A . Then, according to condition 37, U and jAj commute. Since
jAj1=2 is a polynomial of jAj , the matrices U and jAj1=2 commute as well. Now
we have, for any u 2 C n ,
j(Au; u)j = j(U jAju; u)j = j(U jAj1=2 u; jAj1=2 u)j
 jjU jAj1=2 ujj jj jAj1=2 ujj = jj jAj1=2 ujj2 = (jAju; u):
Suppose condition 84 holds. Without loss of generality, A may be assumed to be
an upper triangular matrix with diagonal entries 1 ; : : : ; n . Let (t1 ; : : : ; tn ) be
the rst row of jAj . Then the relation jAj2 = AA implies
t21 + jt2 j2 + : : : + jtn j2 = j1 j2:
On the other hand, using the condition for u = (1; 0; : : : ; 0)T gives
j1 j  t1:
Hence
t1 = j1 j; t2 = : : : = tn = 0:
Continuing in this way, we conclude that the singular values of A coincide with
the moduli of its eigenvalues. By condition 58, A is normal.
For condition 85, the suciency is proved in the same way, with A as a lower
triangular matrix and with jAj replaced by jAj .
Now we outline the proof of suciency for condition 86 given in [12]. The proof is
preceded by the following
Lemma 1. Let A 2 C n;n have all eigenvalues equal to 1 in absolute value. If
(35) j(Ay; y)j  (Ay; Ay) for all unit vectors y 2 C n ;
then either  = 1
2
or A is a unitary matrix.

To prove the lemma, one may assume that A is an upper triangular matrix. Then,
inspecting (35) for y with only two components nonzero, the rst and the sth one,
one can show that
a1s = 0; s = 2; : : : ; n;
if  6= 21 . Applying an inductive argument, one shows A to be diagonal, hence
unitary.
Returning to condition 86, we rst consider the case where A is nonsingular. Let
A = UDV be the singular value decomposition of A . For any nonzero u , set
1=2 
y = jjDD1=2 UUuujj :

12
Then (13) can be rewritten in the form of (35), with A replaced by
Ae = D1=2 V UD?1=2 :
According to Lemma 1, Ae is unitary which gives V UD = DV U . >From the last
relation, the normality of A follows easily.
In the case where A is singular, the induction on n is used. For the induction
step, one may assume, without loss of generality, that
 
A = A01 0b ;

where A1 is an (n ? 1) {square matrix and b is a column (n ? 1) {vector.

Assume that b 6= 0 , then A is not normal. Take u1 2 C n?1 such that (b; u1) 6= 0
and let
 
(36) u = uu1 with u2 > 0:
2

Examining (13) when u2 ! 1 in (36) (and u1 is xed), we observe that the left{
hand side grows linearly in u2 , with (b; u1) as the leading coecient. On the other
hand, the leading term in the right{hand side behaves like u22 for a nonzero .
For inequality (13) to be valid for all u2 , we must have   21 .
Applying the same argument to A with (13) rewritten in the form
j(Au; u)j  (jAju; u)1? (jAju; u)1?(1?) 8 u 2 C n ;
we obtain   21 . Hence,  = 21 which contradicts the assumption.
Condition 87 is an immediate consequence of the following assertion that can be
found on page 426 of [10]:
Lemma 2. A necessary and sucient condition for matrices A1 and A2 to be
simultaneously diagonalizable is that both products A1 A2 and A2 A1 are normal
matrices.

We mention that, while the unitary character of Q is vital for the validity of
condition 87, the triangular form of R has no signicance at all. Therefore, similar
conditions may be stated for other decompositions of A with one of the factors
unitary.
Condition 88 is a slightly disguised version
?n
of condition 59. One only needs to recall
that the eigenvalues of Ck (A) are the k products
i1 : : : ik
13
where 1  i1 < i2 < : : : < ik  n [14, p. 24], and the similar expressions hold
for the singular values of Ck (A) .
To prove condition 89, we begin with the following observation. In the orthonormal
basis of the unitary space C n;n composed of the matrices
Eij = (ki j` )nk;`=1; i; j = 1; : : : ; n;
the operator LA is described by the matrix
MA = In
A + A
In
(see [14, p. 9]). Thus LA is normal i MA is.
Assume that
A ! R = U AU
is the unitary similarity which transforms A into its (upper triangular) Schur form.
Then MA is unitary similar to the upper triangular matrix
MR = In
R + R
In :
We conclude that A; MA , and LA are simultaneously normal (or not normal).

14
 References
1. B. Aupetit and J. Zemanek, A characterization of normal matrices by their exponentials,
Linear Algebra Appl. 132 (1990), 119{121.
2. J. Cohen, S. Friedland, T. Kato, and F.P. Kelly, Eigenvalue inequalities for products of matrix
exponentials, Linear Algebra Appl. 45 (1982), 55{95.
3. L. Elsner and M.H.C. Paardekooper, On measures of nonnormality of matrices, Linear Alge-
bra Appl. 92 (1987), 107{124.
4. K. Fan, On a theorem of Weyl concerning eigenvalues of linear transformations I, Proc. Nat.
Acad. Sci. U.S.A. 35 (1949), 652{655.
5. S. Friedland, A characterization of normal matrices, Israel J. Math. 42 (1982), 235{240.
6. R. Grone, C.R. Johnson, E.M. Sa, and H. Wolkowicz, Normal matrices, Linear Algebra Appl.
87 (1987), 213{225.
7. O.H. Hald, A converse to the Bauer{Fike theorem, Linear Algebra Appl. 9 (1974), 267{274.
8. R.E. Hartwig and K. Spindelbock, Matrices for which A and A+ commute, Linear and
Multilinear Algebra 14 (1984), 241{256.
9. S. Hochwald, Linear algebra by analogy, Amer. Math. Monthly 98 (1991), 918{926.
10. R. Horn and C. Johnson, Matrix Analysis, Cambridge U.P., 1985.
11. R. Horn and C. Johnson, Topics in Matrix Analysis, Cambridge U.P., 1991.
12. C.R. Johnson and F. Zhang, An operator inequality and matrix normality, Linear Algebra
Appl. 240 (1996), 105{110.
13. Kh.D. Ikramov, Unitary{triangular factorizations of a normal matrix, Comput. Math. Math.
Phys. 33 (1993), 407{410.
14. M. Marcus and H. Minc, A Survey of Matrix Theory and Matrix Inequalities, Dover, 1992.
15. A.W. Marshall and I. Olkin, Inequalities: Theory of Majorization and its Applications, Aca-
demic Press, 1979.
16. W. So, Equality cases in matrix exponential inequalities, SIAM J. Matrix Anal.Appl. 13
(1992), 1154{1158.
17. T. Yamamoto, On the extreme values of the roots of matrices, J. Math. Soc. Japan 19 (1967),
173{178.

15

View publication stats