On A Matrix Decomposition and Its Applications

Download as pdf or txt
Download as pdf or txt
You are on page 1of 9

On a matrix decomposition and its applications

Fuzhen Zhang
Farquhar College of Arts and Sciences
Nova Southeastern University
Fort Lauderdale, FL 33314, USA
zhang@nova.edu

Abstract
We show the uniqueness and construction (of the Z matrix in Theorem 1,
to be exact) of a matrix decomposition and give an affirmative answer to a
question proposed in [J. Math. Anal. Appl. 407 (2013) 436-442].
AMS Classification: 15A45, 15A60, 47A30
Keywords: accretive-dissipative matrix, Cartesian decomposition, majorization, matrix de-
composition, numerical range, sectoral decomposition, unitarily invariant norm

1 Introduction
Several recent papers [2, 3, 4, 5, 10] are devoted to the study of matrices with
numerical range in a sector of the complex plane. In particular, this includes the
study of accretive-dissipative matrices and positive definite matrices as special
cases. A matrix decomposition plays a fundamental role in these works. The
aim of this paper is twofold: show the uniqueness along with other properties
of the key matrix in the decomposition and give an affirmative answer to a
question raised in [12].
As usual, the set of n × n complex matrices is denoted by Mn . For A ∈ Mn ,
the singular values and eigenvalues of A are denoted by σi (A) and λi (A), respec-
tively, i = 1, . . . , n. The singular values are always arranged in nonincreasing
order: σ1 (A) ≥ · · · ≥ σn (A). If A is Hermitian, then all eigenvalues of A are
real and ordered as λ1 (A) ≥ · · · ≥ λn (A). Note that σj (A) = λj (|A|), where |A|
is the modulus of A, i.e., |A| = (A∗ A)1/2 with A∗ for the conjugate transpose
of A. We denote σ(A) = (σ1 (A), . . . , σn (A)) and λ(A) = (λ1 (A), . . . , λn (A)).
For a square complex matrix A, recall the Cartesian (or Toeplitz) decompo-
sition (see, e.g., [1, p. 6] and [7, p. 7]) A = <A + i=A, where
1 1
<A = (A + A∗ ), =A = (A − A∗ ).
2 2i

1
There are many interesting properties for such a decomposition. For in-
stance, <(R∗ AR) = R∗ (<A)R for any A ∈ Mn and any n × m matrix R. A
celebrated result due to Fan and Hoffman (see, e.g., [1, p. 73]) sates that

λj (<A) ≤ σj (A), j = 1, . . . , n. (1)

For A ∈ Mn , the numerical range of A is the set in the complex plane

W (A) = {x∗ Ax | x ∈ Cn , kxk = 1}.

For α ∈ [0, π2 ), let Sα be the sector in the complex plane given by

Sα = {z ∈ C | <z > 0, |=z| ≤ <z tan α} = {reiθ | r > 0, |θ| ≤ α}.

Apparently, if the numerical range W (A) is contained in a sector Sα for


some α ∈ [0, π2 ), then A is nonsingular and <A is positive definite. Moreover,
W (A) ⊆ Sα implies W (R∗ AR) ⊆ Sα for any nonzero n × m matrix R.
If W (A) is contained in the first quadrant of the complex plane, then <A
and =A are positive semidefinite. We call such a matrix A accretive-dissipative.
Note that if A is accretive-dissipative and nonsingular, then W (A) ⊆ eiπ/4 Sπ/4 ,
i.e., W (e−iπ/4 A) ⊆ Sπ/4 . With continuity argument, we assume that the
accretive-dissipative matrices to be considered in this paper are nonsingular.
We write A ≥ 0 if A is positive semidefinite and A > 0 if A is positive
definite. For two Hermitian matrices A and B of the same size, we denote
A ≥ B if A − B ≥ 0. Note that A ≥ B implies λj (A) ≥ λj (B) for all j.
In Section 2, we provide a detailed analysis of the so-called sectoral de-
composition and show some important properties. In section 3, we use the
decomposition and majorization as a tool to obtain some norm inequalities; a
question raised in [12] is answered.

2 A matrix decomposition with a sector


We begin with discussions on a matrix decomposition which we refer to as
the sectoral decomposition. The existence of the matrix decomposition with
numerical range contained in a sector has appeared in [2, Lemma 2.1]. A similar
observation was made by London [13] three decades ago (or even earlier by
A. Ostrowski and O. Taussky) to prove a number of existing results by the
factorization. This decomposition theorem, though simple as it looks, has been
heavily used in recent papers [2, 3, 4, 5, 10]. In light of its importance and for
completeness and convenience, we restate it here; we then show the uniqueness
and give a way of constructing the key matrix Z in the decomposition.

Theorem 1 (Sectoral decomposition) Let A be an n × n complex matrix


such that W (A) ⊆ Sα for some α ∈ [0, π2 ). Then there exist an invertible matrix
X and a unitary diagonal matrix Z = diag(eiθ1 , . . . , eiθn ) with all |θj | ≤ α such
that A = XZX ∗ . Moreover, such a matrix Z is unique up to permutation.

2
Proof. Existence. Write A = M + iN , where M = <A and N = IA are
Hermitian. Since W (A) ⊆ Sα , A is invertible and M is positive definite. By [7,
Theorem 7.6.4] or [16, Theorem 7.6], M and N are simultaneously *-congruent
and diagonalizable, that is, P ∗ M P and P ∗ N P are diagonal for some invertible
matrix P . It follows that we can write A = QDQ∗ for some diagonal matrix D
and invertible matrix Q. Since W (A) ⊆ Sα , we have W (D) ⊆ Sα . Thus we can
iθ1 iθn
√ 1 e ,√. . . , dn e ), where djiθ> 0 andiθ|θj | ≤ α, j = 1, . . . , n.∗ Set
write D = diag(d
X = Q diag( d1 , . . . , dn ) and Z = diag(e 1 , . . . , e n ). Then A = XZX , as
desired.
Uniqueness. Suppose that A = XZ1 X ∗ = Y Z2 Y ∗ are two decompositions
of A, where X and Y are nonsingular, Z1 and Z2 are unitary and diagonal.
We may assume Y = I (otherwise replace X with Y −1 X). We show that Z1
and Z2 have the same main diagonal entries (regardless of order). For this, we
show that β ∈ C is a diagonal entry of Z1 with multiplicity k if and only
if β is a diagonal entry of Z2 with the same multiplicity. Without loss of
generality, we may assume β = 1 (or multiply both sides by β̄ and continue
the discussion on X(β̄Z1 )X ∗ = β̄Z2 ). Let Z1 = C1 + iS1 and Z2 = C2 + iS2
be the Cartisian decompositions of Z1 and Z2 , respectively. Then C1 and C2
are positive definite. Since β = 1 is a diagonal entry of Z1 with multiplicity
k, 1 appears on the diagonal of C1 k times, so S1 has k zeros on its diagonal.
Thus rank(XS1 X ∗ ) = n − k. As XS1 X ∗ = S2 , we have rank(S2 ) = n − k.
This implies that C2 contains k 1’s on its diagonal. We conclude that Z2 is
permutation similar to Z1 .
Note that cos α is decreasing in α on [0, π2 ), the following are immediate.

Corollary 1 Let A be an n × n complex matrix such that W (A) ⊆ Sα for some


α ∈ [0, π2 ) and let A = XZX ∗ be a sectoral decomposition of A, where X is
invertible and Z is unitary and diagonal. Then

(i). I ≤ sec α (<Z).


(ii). RR∗ ≤ sec α R(<Z)R∗ for any matrix R.


(iii). σj2 (R) ≤ sec α λj R(<Z)R∗ ≤ sec α σj RZR∗ for any R and j.
 


(iv). σj2 (X) ≤ sec α λj (<A ≤ sec α σ j (A) for all j = 1, . . . , n.

The following result gives a way of constructing the unique matrix Z.

Theorem 2 Let A be an n×n complex matrix with the Cartesian decomposition


A = M + N i, where M is positive definite and N is Hermitian. Then the
matrix Z in Theorem 1 is determined by the eigenvalues of M −1 N . Let µj be
the eigenvalues of M −1 N and let 1 + iµj = |1 + iµj |eiγj , |γj | < π2 , j = 1, . . . , n.
Then Z = diag(eiγ1 , . . . , eiγn ). Let γ(A) = maxj |γj |. Then W (A) ⊆ Sγ(A) .

Proof. Since M > 0, there is an invertible matrix P such that P ∗ M P = I


and P ∗ N P = D is diagonal (see, e.g., [16, p. 213]). Recall that when X and

3
Y are both n × n matrices, XY and Y X have the same eigenvalues. We have
λj (P ∗ N P ) = λj (P P ∗ N ) = λj (M −1 N ). It follows that P ∗ AP = I +Di and D is
the diagonal matrix of the eigenvalues µj of M −1 N . Let 1 + iµj = |1 + iµj |eiγj ,
|γj | < π2 , j = 1, . . . , n. Then Z = diag(eiγ1 , . . . , eiγn ). With γ(A) = maxj |γj |,
we see that W (Z), W (I + Di), and W (A) are all contained in Sγ(A) .

Corollary 2 Let A be an n × n complex matrix such that W (A) ⊆ Sα for some


α ∈ [0, π2 ). Then there exist a normal matrix Λ such that A = (<A)1/2 Λ(<A)1/2 .
Moreover, kΛk2 ≤ sec α for the spectral norm k · k2 on Mn .

Proof. Let A = M + N i with M = <A and N = =A. Take Λ = I +


M −1/2 N M −1/2 i. Then Λ is normal and W (Λ) ⊆ Sα . For any unit vector z,
z ∗ Λz is a point in the xy-plane with x-coordinate x = 1. It follows that the
numerical radius of Λ, i.e., w(Λ) = max{|z ∗ Λz| | z ∈ Cn , z ∗ z = 1}, is no more
than sec α (as the hypotonus of the right triangle with the adjacent leg of length
1). Since Λ is normal, all the singular values of Λ are no more than sec α. In
particular, for the spectral norm kΛk2 , we have kΛk2 ≤ sec α.
Let θa and θb be respectively the largest and smallest values of the θj ’s in
Theorems 1 and 2. For the Z in the decomposition, W (Z) is the region formed
by the portion of the unit circle from eiθa to eiθb and the line segment from eiθa
to eiθb . For the Λ in the corollary, W (Λ) is the vertical line segment x = 1 from
the point (1, tan θa ) to the point (1, tan θb ). All these figures are contained in
Sθc , where θc = max{|θa |, |θb |}, which is nothing but the γ in Theorem 2.
Below is an addition-closure property for the numerical ranges in a sector.
Proposition. Let A, B ∈ Mn . If W (A), W (B) ⊆ Sα for some α ∈ [0, π2 ), then

W (A + B) ⊆ Sα .

Proof. Consider the Cartesian decompositions of A and B,

A = R1 + iS1 , B = R2 + iS2 .

Since W (A) and W (B) are contained in Sα , we have R1 + R2 > 0. Note that
for a, b, c, d > 0, (a + b)/(c + d) ≤ max{a/c, b/d}. We compute, for any x 6= 0,
|x∗ (S1 + S2 )x| |x∗ S1 x| + |x∗ S2 x|

x∗ (R1 + R2 )x x∗ (R1 + R2 )x
x |S1 |x + x∗ |S2 |x


x∗ R1 x + x∗ R2 x
 ∗
x |S1 |x x∗ |S2 |x

≤ max ,
x∗ R1 x x∗ R2 x
≤ tan α.

This says |x∗ =(A + B)x| ≤ x∗ <(A + B)x tan α. Thus, W (A + B) ⊆ Sα .


We note here that fractional roots (powers) of elements in Banach algebras
are studied in [9] by means of numerical range sectors.

4
3 Norm inequalities for partitioned matrices
Recall that a norm k · k on Mn is unitarily invariant if kU AV k = kAk for any
A ∈ Mn and all unitary U, V ∈ Mn . The unitarily invariant norms of matrices
are determined by nonzero singular values of the matrices via symmetric gauge
functions (see, e.g., [16, Theorems 10.37 and 10.38]). If B is a submatrix of
A ∈ Mn , then kBk is understood as the norm of the n × n augmented matrix
B with 0’s, and conventionally B has n singular values with the trailing ones
0; that is, σ(B) = (σ1 (B), . . . , σr (B), 0, . . . , 0) ∈ Rn , where r is the rank of B.
Thus σ(A) and σ(B) are both in Rn .
Let A be an n-square complex matrix partitioned in the form
 
A11 A12
A= , where A11 and A22 are square. (2)
A21 A22

In [12], the following norm inequalities are proved (in Hilbert space).
LZ1 [12, Theorem 3.3]: Let A ∈ Mn be accretive-dissipative and partitioned as
in (2). Then for any unitarily invariant norm k · k on Mn ,

max{kA12 k2 , kA21 k2 } ≤ 4kA11 k kA22 k. (3)

LZ2 [12, Theorem 3.11]: Let A ∈ Mn be accretive-dissipative and partitioned


as in (2). Then for any unitarily invariant norm k · k on Mn ,

kAk ≤ 2 (kA11 k + kA22 k). (4)

It is
√ asked in [12] as an open problem whether the factor
√ 4 in (3) and the
factor 2 in (4) can be improved. Indeed, the factor 2 in (4) is optimal.
To construct such an accretive-dissipative matrix, we can find a matrix whose
numerical range is contained in the sector Sπ/4 , then rotate it by +π/4. The
normal matrix B = 10 01 + 01 10 i = 1i 1i has eigenvalues 1 + i and 1 − i. So the
     

matrix A = eiπ/4
√ B is accretive-dissipative. A and B have the same repeated
singular value 2. Thus, for the trace norm (sum of all singular values),
√ √ √
2 2 = kAk = 2 (kA11 k + kA22 k) = 2 (1 + 1).

However, the factor 4 in (3) can be improved to 2 (see Corollary 3). In this
section, we show some more general results than (3) and (4).
We adopt the following standard notations. Let x = (x1 , . . . , xn ), y =
(y1 , . . . , yn ) ∈ Rn . We denote the componentwise product of x and y by x ◦ y.
i.e., x ◦ y = (x1 y1 , . . . , xn yn ). We write x ≤ y to mean xj ≤ yj for j = 1, . . . , n.
We say that x is weakly majorized by y, written as x ≺w y, if the partial sum
of the first k largest components of x is less than or equal to the corresponding
partial sum of y for k = 1, . . . , n. We write x ≺ y if x ≺w y and the sum of all
components of x is equal to that of y. (See, e.g., [14, p. 12] or [16, p. 326].)
It is well known (see, e.g., [14, p. 368] or [16, p. 375]) that, for A, B ∈ Mn ,
kAk ≤ kBk for all unitarily invariant norms k · k on Mn if and only if σ(A) ≺w

5
σ(B). So, to some extend, the norm inequalities are essentially the same as the
singular value majorization inequalities. The Fan-Hoffman inequalities (1) yield
immediately k<Ak ≤ kAk for any A ∈ Mn and any unitarily invariant norm
k · k on Mn . The following is a reversal. Two useful facts are: the singular value
majorization of product σ(AB) ≺w σ(A) ◦ σ(B) (see, e.g., [16, p. 363]) and its
companion norm inequality kABk2 ≤ kAA∗ k kB ∗ Bk (see, e.g., [6, p. 212]).

Lemma 1 Let A ∈ Mn have W (A) ⊆ Sα for some α ∈ [0, π2 ). Then

σ(A) ≺w sec α λ(<A).

Equivalently, for all unitarily invariant norms k · k on Mn ,

kAk ≤ sec αk<Ak.

Proof. Let A = XZX ∗ be a sectoral decomposition of A, where X is invertible


and Z is unitary and diagonal. Then

σ(A) = σ(XZX ∗ ) ≺w σ(X) ◦ σ(Z) ◦ σ(X ∗ ) = σ 2 (X) ≤ sec α λ(<A).

The last “≤” is by Corollary 1 (iv). The norm inequality follows at once.

Theorem 3 Let A ∈ Mn be partitioned as in (2) and assume W (A) ⊆ Sα for


some α ∈ [0, π2 ). Then for any unitarily invariant norm k · k on Mn ,

max{kA12 k2 , kA21 k2 } ≤ sec2 α kA11 k kA22 k. (5)

Proof. Let A11 be p × p. By Theorem 1, let A = XZX ∗ be a sectoral


decomposition of A, where
  X is invertible and Z is unitary and diagonal. We
X1
partition X as X = , X1 ∈ Mp×n . Then <A11 = X1 (<Z)X1∗ , <A22 =
X2
X2 (<Z)X2∗ , and A12 = X1 ZX2∗ . Using Corollary 1 (ii), we have

kA12 k2 = kX1 ZX2∗ k2 ≤ kX1 X1∗ k kX2 Z ∗ ZX2∗ k


≤ sec2 αkX1 (<Z)X1∗ k kX2 (<Z)X2∗ k
= sec2 αk<A11 k k<A22 k
≤ sec2 αkA11 k kA22 k.

So (5) is true for A12 . The inequality for A21 is similarly proven.
If A is a positive definite matrix, then α = 0 and sec α = 1 in (5).

Corollary 3 Let A ∈ Mn be accretive-dissipative and partitioned as in (2).


Then for any unitarily invariant norm k · k on Mn ,

max{kA12 k2 , kA21 k2 } ≤ 2 kA11 k kA22 k. (6)

6
Proof. Set α = π/4 in the theorem. Then sec2 α = 2.
(6) is stronger than (3). Moreover, the constant factor 2 is best possible
for all accretive-dissipative matrices and unitarily invariant norms. Let B =
1 1−i ∗
0 1 . One may check that <B > 0 and <B ≥ ±=B, which yield x (<B)x ≥
|x∗ (=B)x| for all x ∈ C2 . (Note that <B 6≥ |=B|.) So W (B) ⊆ Sπ/4 and
A = eiπ/4 B is accretive-dissipative. For the trace norm, apparently, kA12 k2 =
2 = 2(1 · 1) = 2kA11 k kA22 k. This answers a question raised in [12, p. 442].
To present next theorem, we need a lemma which is interesting on its own.
h i
Lemma 2 Let H = H∗11 H∗22 be an n × n positive semidefinite matrix, where
H11 and H22 are square submatrices (possibly of different sizes). Then

λ(H) ≺ λ(H11 ) + λ(H22 ). (7)

Consequently, for all unitarily invariant norms k · k on Mn ,

kHk ≤ kH11 k + kH22 k. (8)



Proof. Note that a matrix P his positive
i semidefinite
h ∗ iif and only if P = Q Q
S∗ S S ∗ ∗
for some matrix Q. Let H = T ∗ [S, T ] = ∗ T ∗T with H11 = S S and

H22 = T T . Using the fact that matrices XY and Y X have the same nonzero
eigenvalues for any (p × q) matrix X and any (q × p) matrix Y , we arrive at
  S∗     ∗
S
λ(H) = λ [S, T ] = λ [S, T ]
T∗ T∗
= λ(SS ∗ + T T ∗ ) ≺ λ(SS ∗ ) + λ(T T ∗ )
= λ(H11 ) + λ(H22 ).

Here we regard λ(H11 ) and λ(H22 ) as vectors in Rn (by adding 0’s).


h i
Remark: It is known [14, p. 308] that if H = H∗11 H∗22 is Hermitian, then

λ(H11 ), λ(H22 ) = λ(H11 ⊕ H22 ) ≺ λ(H).
h i
It is also known (see [15] or [11]) that if H = H 11 K
K H22 is positive semidefinite,

where K is Hermitian or skew-Hermitian, then

λ(H) ≺ λ(H11 + H22 ).

We must also point out that (8) has appeared in [6, p. 217] and a more
general result is available in [8, Theorem 2.1]. We include our proof here as it
is short and elementary, and the most elegant one in author’s opinion.

Theorem 4 Let A ∈ Mn be partitioned as in (2) and let W (A) ⊆ Sα for some


α ∈ [0, π2 ). Then for any unitarily invariant norm k · k on Mn ,

kAk ≤ sec α (kA11 k + kA22 k). (9)

7
h i
<A11 ∗
Proof. By Lemma 1 and noticing that <A = ∗ <A22 > 0, we have

kAk ≤ sec αk<Ak ≤ sec α(k<A11 k + k<A22 k).

The desired inequality follows at once since k<Xk ≤ kXk for any X.
If A is positive definite, then α = 0 and Theorem 4 reduces to (8). If A is
accretive-dissipative, then (4) is immediate by setting α = π/4 in (9).

Acknowledgement
The author is thankful to S.W. Drury and M. Lin for reference [5] which initiated
this work; he is also indebted to M. Lin for his valuable input and discussions.

References
[1] R. Bhatia, Matrix Analysis, GTM 169, Springer-Verlag, New York, 1997.
[2] S.W. Drury, Fischer determinantal inequalities and Higham’s Conjecture,
Linear Algebra Appl. (2013), http://dx.doi.org/10.1016/j.laa.2013.08.031.
[3] S.W. Drury, A Fischer type determinantal inequality, Linear Multilinear
Algebra (2013), http://dx.doi.org/10.1080/03081087.2013.832244.
[4] S.W. Drury, Principal powers of matrices with positive defintie real part,
preprint.
[5] S.W. Drury and M. Lin, Singular value inequalities for matrices with nu-
merical ranges in a sector, preprint.
[6] R.A. Horn and C.R. Johnson, Topics in Matrix Analysis, Cambridge Uni-
versity Press, New York, 1991.
[7] R.A. Horn and C.R. Johnson, Matrix Analysis, Cambridge University
Press, New York, 2nd edition, 2013.
[8] E.-Y. Lee, Extension of Rotfel’d theorem, Linear Algebra Appl. 435 (2011)
735-741.
[9] C.-K. Li, L. Rodman, and I. Spitkovsky, On numerical ranges and roots, J.
Math. Anal. Appl. (2003) 329-340.
[10] C.-K. Li and N. Sze, Determinantal and eigenvalue inequalities for matrices
with numerical ranges in a sector, J. Math. Anal. Appl. 410 (2014) 487-491.
[11] M. Lin and H. Wolkowicz, An eigenvalue majorization inequality for posi-
tive semidefinite block matrices, Linear Multilinear Algebra 60 (2012) 1365-
1368.
[12] M. Lin and D. Zhou, Norm inequalities for accretive-dissipative operator
matrices, J. Math. Anal. Appl. 407 (2013) 436-442.
[13] D. London, A note on matrices with positive definite real part, Proc. Amer.
Math. Soc. 82 (1981) 322-324.

8
[14] A.W. Marshall, I. Olkin, and B. Arnold, Inequalities: Theory of Majoriza-
tion and Its Applications, Springer, New York, 2nd edition, 2011.
[15] R. Turkmen, V. Paksoy, and F. Zhang, Some inequalities of majorization
type, Linear Algebra Appl. 437 (2012) 1305-1316.
[16] F. Zhang, Matrix Theory: Basic Results and Techniques, Springer, New
York, 2nd edition, 2011.

You might also like