Perturbación Del Radio Numérico Q de Un Operador de Desplazamiento Ponderado
Perturbación Del Radio Numérico Q de Un Operador de Desplazamiento Ponderado
Perturbación Del Radio Numérico Q de Un Operador de Desplazamiento Ponderado
www.elsevier.com/locate/jmaa
a r t i c l e i n f o a b s t r a c t
Article history: We formulate the Taylor series expansion for the q-numerical radius of a weighted shift
Received 19 November 2007 operator with periodic weights near q = 0. Coefficients up to the fourth order in the
Available online 10 May 2008 expansion are found via the perturbation theory of Hermitian matrices.
Submitted by M. Putinar
© 2008 Elsevier Inc. All rights reserved.
Keywords:
Perturbation
q-numerical radius
Weighted shift operators
1. Introduction
Let A be a bounded linear operator on a complex Hilbert space H . For 0 q 1, the q-numerical radius w q ( A ) of A is
defined by
w q ( A ) = sup | z|: z ∈ W q ( A ) ,
where
W q ( A ) = A ξ, η: ξ, η ∈ H , ξ = η = 1, ξ, η = q
is the q-numerical range of A. This concept is a generalization of the classical numerical range, namely,
W ( A ) = W 1 ( A ) = Ax, x: x ∈ H , x = 1 .
It is known that the q-numerical range W q ( A ) is a bounded convex subset of C (cf. [20]), and it is related with the height
function
h( z) = h A ( z) = sup t 0: ( z, t ) ∈ W A , A ∗ A
* Corresponding author.
E-mail addresses: [email protected] (M.-T. Chien), [email protected] (H. Nakazato).
1
Supported in part by Taiwan National Science Council.
0022-247X/$ – see front matter © 2008 Elsevier Inc. All rights reserved.
doi:10.1016/j.jmaa.2008.05.015
M.-T. Chien, H. Nakazato / J. Math. Anal. Appl. 345 (2008) 954–963 955
The authors of this paper gave in [6] another approach for the q-numerical radius of weighted shift operators with periodic
weights. In that paper, a bounded operator A on the Hilbert space 2 (N) defined by
z ∈ W ( A ). It follows from this fact and Tarski–Seidenberg’s theorem that there exists a real polynomial F (q, w ) satisfying
F (q, w q ( A )) = 0 for 0 q 1 (cf. [4]), that is, the q-numerical radius w = w q ( A ) is an algebraic function in the weights
s1 , s2 , . . . , sn of A. For instance, if m = 3, s1 = s2 = 1, s3 = 0, then
1 1/ 2
wq ( A) = 27 + 18q − 13q2 + (9 + 7q) (1 − q)(9 + 7q)
8
for 1/2 q 1 and w q ( A ) = 1 for 0 q 1/2 (cf. [13]). However exact computation of the polynomial F (q, w ) is not
performed except for some special cases. In Section 5, we provide such a polynomial for a rather simple case. By the result
w q ( A ) = w q (C ), h A ( z) = h C (| z|), a numerical algorithm in [5] is available to obtain a uniform approximation of the graph
w q ( A ) for 0 q 1. We provide another method to analyze the radius w q ( A ) which gives more precise information of
w q ( A ) near q = 0. We consider the Taylor series of the function w q ( A ) in q near q = 0. The coefficients of the Taylor series
for the q-numerical radius of the weighted shift operator A are treated, and the formulae to compute the coefficients up to
the 4th order are given. By the invariance of the q-numerical range under a unitary similarity, the q-numerical radius of A
is invariant under a cyclic replacement of the weights. Moreover, the q-numerical range is transpose invariant, we have that
the q-numerical radius of A is invariant if we reverse the oder of its weights.
In quantum mechanics, perturbation formulae have been used in many instances [2,17]. In this paper, the coefficients of
the Taylor series up to the 4th order are described explicitly as real rational functions of the weights of the shift operator A.
The explicit formula is basically deduced from (2.33) in [12] and the duality between the family of q-numerical radius of A
and the function h A .
2. Perturbation formulae
In this section, we apply the perturbation method introduced by Kato [12], and give some preliminary formulae for
later computations. Let B be an n × n Hermitian matrix. Suppose that {ξ1 , ξ2 , . . . , ξn } is an orthonormal eigenbasis of B
corresponding to the eigenvalues b1 , b2 , . . . , bn related by B ξi = b i ξi . Let T be an n × n Hermitian matrix, we consider a
perturbation B (k) = B + kT , where k is a small real perturbation parameter. By applying Taylor series to the eigenvalue
λi ( B (k)) of B (k) close to bi , we find that
(1) (2) (3) (4) (5)
λi B (k) = bi + c i k + c i k2 + c i k3 + c i k4 + c i k5 + · · · . (2)
We treat the eigenvalue λi ( B (k)) for an index i such that b j = b i for j ∈ {1, 2, . . . , n}\{i }. This condition is called non-
degenerated in quantum mechanics (cf. [17]).
Consider the 1-dimensional orthogonal projection P (k) satisfying
B (k) P (k) = λi B (k) P (k).
where P ξ = ξ, ξi ξi for ξ ∈ Cn . We consider the resolvent family of the matrix B, that is,
R (ζ ) = ( B − ζ I )−1 .
With respect to the orthonormal basis {ξ1 , ξ2 , . . . , ξn }, the matrix representation of the resolvent R (ζ ) is
diag 1/(b1 − ζ ), 1/(b2 − ζ ), . . . , 1/(bn − ζ ) .
1
Sξ j = ξj
b j − bi
for small |ζ − b i |.
By [12], we have
∞
B (k) − b i I P (k) = B̃ (m) km , (4)
m=1
1
B̃ (m) = R (ζ ) T R (ζ )(ζ − b i ) dζ, (5)
2π i
Γ
where Γ is a positively oriented circle enclosing b i but no other eigenvalues of B. We substitute (3) into (4) and (5), then
take the trace on (4). By using the facts that T P = P T , P 2 = P , S P = 0, we compare the coefficients of (2), and obtain the
equalities
(1)
ci = tr( P T P ) = tr( T P ) = T ξi , ξi ,
(2)
ci = −tr( P T S T P ) = −tr( T S T P ) = − T S T ξi , ξi = − S T ξi , T ξi ,
(3)
c i = tr( P T S T S T P ) − tr P T S 2 T P T P = T S T ξi , S T ξi − T P T ξi , S 2 T ξi ,
(4)
c i = tr − T S T S T S T P + T S 2 T S T P T P + tr T S T S 2 T P T P + T S 2 T P T S T P − T S 3 T P T P T P
= − S T S T ξi , T S T ξi + T S T P T ξi , S 2 T ξi + S 2 T P T ξi , T S T ξi + P T S T ξi , T S 2 T ξi − P T P T ξi , T S 3 T ξi .
By the transformation S, the coefficients become
(1)
ci = T ξi , ξi
(2)
| T ξi , ξ j |2
ci = ,
bi − b j
1 j n, j =i
(3)
( T ξ j , ξ j − T ξi , ξi )| T ξi , ξ j |2
( T ξ j , ξi T ξi , ξ p T ξ p , ξ j )
ci = +2 . (6)
(bi − b j )2 (bi − b j )(bi − b p )
1 j n, j =i 1 j < p n, j , p =i
(4)
The expression of c i is more complicated, and is given by
(4) (4) (4) (4) (4)
ci = c i ,1 + c i ,2 + c i ,3 + c i ,4 ,
where
(4)
| T ξi , ξ j |2 (( T ξ j , ξ j − T ξi , ξi )2 − | T ξi , ξ j |2 )
c i ,1 = ,
(bi − b j )3
1 j n, j =i
(4)
| T ξi , ξ j |2 | T ξ p , ξ j |2 − | T ξ p , ξi |2
c i ,2 = ,
(bi − b j )2 bi − b p
1 j = p n, j , p =i
M.-T. Chien, H. Nakazato / J. Math. Anal. Appl. 345 (2008) 954–963 957
(4) − T ξi , ξi + T ξk , ξk − T ξi , ξi + T ξ j , ξ j
c i ,3 = 2 + T ξ j , ξi T ξi , ξk T ξk , ξ j ,
(bi − b j )(bi − bk )2 (bi − a j )2 (bi − bk )
1 j <kn, j ,k=i
(4)
( T ξ p , ξi T ξ j , ξ p T ξi , ξq T ξq , ξ j )
c i ,4 = 2 . (7)
(bi − b j )(bi − b p )(bi − bq )
1 j n, j =i 1 p <qn, p ,q=i , j
We consider the m × m weighted cyclic matrix C with weights {s1 , s2 , . . . , sm } defined by (1), where we assume that
s1 , s2 , . . . , sm are nonnegative distinct numbers. Consider the m × m perturbation Hermitian matrix
⎛ ⎞
s21 k s 1 /2 0 ... 0 k sm /2
⎜ k s 1 /2 s22 k s 2 /2 ... 0 0 ⎟
2 2 ∗
⎜
⎜ s23 0 ⎟
⎟
B (k) = diag s1 , s2 , . . . , sn + k C + C /2 ⎜ 0
2 k s 2 /2 ... 0 ⎟.
⎜ .. .. .. .. .. .. ⎟
⎝ . ⎠
. . . . .
2
k sm /2 0 0 . . . k sm−1 /2 sm
Suppose that {ξ1 , ξ2 , . . . , ξm } is the standard orthonormal basis for Cm , i.e.,
and
s21 2
sm (3) (4) (5)
λ1 B (k) = s21 + + k2 + c 1 k3 + c 1 k4 + c 1 k5 + · · · .
4(s21 − s22 ) 4(s21 − sm
2
)
We introduce an m × m Hermitian matrix
⎛0 s1 0 ... 0 sm
⎞
⎜ s1 0 s2 ... 0 0 ⎟
⎜0 ⎟
T̃ = C + C ∗ = ⎜
⎜ .
s2 0 ... 0 0 ⎟,
⎝ . .. .. .. .. .. ⎟
⎠
. . . . . .
sm 0 0 . . . sm−1 0
and consider the perturbation
k
B (k) = diag s21 , s22 , . . . , sn2 + T̃ .
2
Let m = 3. By (6), it follows
(3) 1 s1 s2 s3
c1 = ,
4 (s21 − s22 )(s21 − s23 )
(3) 1 s1 s2 s3
c2 = ,
4 (s22 − s21 )(s22 − s23 )
(3) 1 s1 s2 s3
c3 = .
4 (s23 − s21 )(s23 − s22 )
(4) (4) (4)
It is obvious that c i ,3 = c i ,4 = 0, i = 1, 2, 3, and hence c i = c (i ,41) + c (i ,42) for i = 1, 2, 3. We find that
−1
T̃ ξ j , ξ1 4
3
(4) 1 s41 s43
c 1, 1 = = − + ,
16
j =2
(s21 − s2j )3 16 (s21 − s22 )3 (s21 − s23 )3
(4) 1 T̃ ξ1 , ξ2 2 T̃ ξ3 , ξ2 2 − T̃ ξ3 , ξ1 2 T̃ ξ1 , ξ3 2 T̃ ξ2 , ξ3 2 − T̃ ξ2 , ξ1 2
c 1, 2 = +
16 (s21 − s22 )2 s21 − s23 (s21 − s23 )2 s21 − s22
2 2
1 s21 s22 − s23 s23 s22 − s21 1 s1 (s2 − s23 ) s23 (s22 − s21 )
= + 2 = + .
16 (s21 − s22 )2 s21 − s23 (s1 − s23 )2 s21 − s22 16(s21 − s22 )(s21 − s23 ) s21 − s22 s21 − s23
958 M.-T. Chien, H. Nakazato / J. Math. Anal. Appl. 345 (2008) 954–963
Hence, it follows
(4) 1 s21 (s41 − s22 s23 )[s41 − (s22 + s23 )s21 + s42 − s22 s23 + s43 ]
c1 = − .
16 (s21 − s22 )3 (s21 − s23 )3
Similarly, we have
(4) 1 s22 (s42 − s21 s23 )(s42 − (s21 + s23 )s22 + s41 − s21 s23 + s43 )
c2 = − ,
16 (s22 − s21 )3 (s22 − s23 )3
(4) 1 s23 (s43 − s21 s22 )(s43 − (s21 + s22 )s23 + s41 − s21 s22 + s42 )
c3 = − .
16 (s23 − s21 )3 (s23 − s22 )3
In the case m = 4, we see that
(4) 1 T̃ ξ2 , ξ1 T̃ ξ4 , ξ1 T̃ ξ2 , ξ3 T̃ ξ4 , ξ3 s1 s2 s3 s4
c 1, 4 = = .
8 (s21 − s23 )(s21 − s22 )(s21 − s24 ) 8(s21 − s22 )(s21 − s23 )(s21 − s24 )
Similarly,
(4) s1 s2 s3 s4
c 2, 4 = ,
8(s22 − s21 )(s22 − s23 )(s22 − s24 )
(4) s1 s2 s3 s4
c 3, 4 = ,
8(s23 − s21 )(s23 − s22 )(s23 − s24 )
(4) s1 s2 s3 s4
c 4, 4 = .
8(s24 − s21 )(s24 − s22 )(s24 − s23 )
Furthermore, we have
(4) −1 T̃ ξ j , ξi 4 1 s41 s42
c 2, 1 = =− + ,
16
1 j 4, j =2
(s2i − s2j )3 16 (s22 − s21 )3 (s22 − s23 )3
and
(4) 1 s41 s44
c 1, 1 = − + ,
16 (s21 − s22 )3 (s21 − s24 )3
(4) 1 s42 s43
c 3, 1 = − + ,
16 (s23 − s22 )3 (s23 − s24 )3
(4) 1 s44 s43
c 4, 1 = − + .
16 (s24 − s21 )3 (s24 − s23 )3
By using (7), we obtain
(4) 1
T̃ ξ2 , ξ j 2 T̃ ξ p , ξ j 2 − T̃ ξ p , ξ2 2
c 2, 2 =
16
1 j = p 4, j , p =2
(s22 − s2j )2 s22 − s2p
1 s21 −s22 s22 −s21 1 s21 s24
T̃ ξ4 , ξ1 2 − T̃ ξ1 , ξ2 2 0
= + 2 + + 2
16 (s2 − s1 )2 s2 − s3
2 2 2 2
(s2 − s3 )2 s2 − s1
2 2 2 16 (s2 − s1 )2 s2 − s4
2 2 2 2
(s2 − s4 )2
2
s22 − s21
1 s22 s23 0 T̃ ξ3 , ξ4 2 − T̃ ξ3 , ξ2 2
+ +
16 (s22 − s23 )2 s22 − s24 (s22 − s24 )2 s22 − s23
1 2 2
s1 s2 2 2
s1 s2 s21 s24 s22 s23
= − 2 − + + .
16 (s2 − s21 )2 (s22 − s23 ) (s22 − s23 )2 (s22 − s21 ) (s22 − s21 )2 (s22 − s24 ) (s22 − s23 )2 (s22 − s24 )
Similarly, we verify that
(4) 1 s22 s23 s22 s23 s22 s21 s23 s24
c 3, 2 = − − + + ,
16 (s23 − s22 )2 (s23 − s24 ) (s23 − s24 )2 (s23 − s22 ) (s23 − s22 )2 (s23 − s21 ) (s23 − s24 )2 (s23 − s21 )
(4) 1 s23 s24 s23 s24 s23 s21 s24 s21
c 4, 2 = − − + + ,
16 (s24 − s23 )2 (s24 − s21 ) (s24 − s21 )2 (s24 − s23 ) (s24 − s23 )2 (s24 − s22 ) (s24 − s21 )2 (s24 − s22 )
(4) 1 s24 s21 s24 s21 s24 s22 s21 s22
c 1, 2 = − − + + .
16 (s21 − s24 )2 (s21 − s22 ) (s21 − s22 )2 (s21 − s24 ) (s21 − s24 )2 (s21 − s23 ) (s21 − s22 )2 (s21 − s23 )
M.-T. Chien, H. Nakazato / J. Math. Anal. Appl. 345 (2008) 954–963 959
(4)
Let m 4. It is clear c i ,3 = 0, for all 1 i m, and the coefficients (6) are given by
(3) 1
T̃ ξ j , ξi T̃ ξ p , ξi T̃ ξ p , ξ j
ci = = 0, (9)
4
1 j < p m, j , p =i
(s2i − s2j )(s2i − s2p )
for 1 i m.
In the case m 5 the equation
T̃ ξ p , ξi T̃ ξq , ξi T̃ ξ p , ξ j T̃ ξq , ξ j = 0
holds for distinct 1 i , j , p , q m, and hence
(4) 1
T̃ ξ p , ξi T̃ ξq , ξi T̃ ξ p , ξ j T̃ ξq , ξ j
c i ,4 = =0
8
1 j n, j =i 1 p <qn, p ,q=i , j
(s2i − s2j )(s2i − s2p )(s2i − sq2 )
(4) 1
T̃ ξ3 , ξ j 2 T̃ ξ p , ξ j 2 − T̃ ξ p , ξ3 2
c 3, 2 =
16
1 j = p m, j =3, p =3
(s23 − s2j )2 s23 − s2p
1
T̃ ξ3 , ξ j 2 T̃ ξ p , ξ j 2 − T̃ ξ p , ξ3 2
=
16
( j , p )=(2,4),(4,2),(4,5),(2,1)
(s23 − s2j )2 s23 − s2p
1 s22 s23 s22 s23 s23 s24 s21 s22
= − − + + . (12)
16 (s23 − s22 )2 (s23 − s24 ) (s23 − s24 )2 (s23 − s22 ) (s23 − s24 )2 (s23 − s25 ) (s23 − s22 )2 (s23 − s21 )
4. Periodic weights
Let A be a shift operator with periodic nonnegative weights {s1 , s2 , . . . , sm }, m 3. Suppose that s2 > s j for all j =
1, 3, . . . , m. It is shown in [6] that the second-order perturbation of the q-numerical radius of A is determined, namely,
(s22 − s21 )(s22 − s23 )
w q ( A ) = s2 − q2 + c 3 q3 + · · ·
2s2 (2s22 − s21 − s23 )
for sufficiently small q. In this section, we proceed to find up to the fourth-order corrections of the q-numerical radius
when the largest weight has multiplicity 1. It does not matter which weight is the largest among the weights because
of the invariance of w q ( A ) under cyclic changes of weights. In the following theorems, we specify the largest weight for
simplicity.
Theorem 4.1. Let A be the shift operator with periodic nonnegative weights {s1 , s2 , s3 , s4 , . . . , sm }, m 5, such that s3 >
max{s1 , s2 , s4 , . . . , sm }. Then the perturbation of the q-numerical radius is
(s23 − s22 )(s23 − s24 ) (4)
w = s3 − q2 + c 3 q4 + O q5 ,
2s3 (2s23 − s22 − s24 )
where
for x ∈ W ( ( A )), x 0, where ( A ) = ( A + A ∗ )/2. By [4, Theorem 2], the graph of h is the upper part of the numerical
range W ( ( A ) + i A ∗ A ), and is obtained as the envelope of the one-parameter family of straight lines
y + kx = max σ T (k) , k 0. (13)
By hypothesis s3 > s j 0 for j = 3, we may assume that max σ ( T (k)) = λ3 ( T (k)) near k = 0. Together with (8)–(12), the
straight line (13) becomes
1 s22 s23 1 s42 s43 s22 s23 s22 s23
− y − kx + s23 + + k2 + − − − −
4 s23 − s22 s23 − s24 16 (s23 − s22 )3 (s23 − s24 )3 (s23 − s22 )2 (s23 − s24 ) (s23 − s24 )2 (s23 − s22 )
s23 s24 s21 s22
+ + k4 + O k5 = 0.
(s23 − 2 2 2
s4 ) (s3 − s25 ) (s23 − 2 2 2
s2 ) (s3 − s21 )
The coefficients c 2 , c 4 of the Taylor expansion of the function y = h(x) near x = 0
h(x) = s23 + c 2 x2 + c 4 x4 + O x5 ,
where
1 s22 s23
α= + ,
4 s23 − s22 s23 − s4 2
β = c 3(4,1) + c 3(4,2)
1 s4 s4 s22 s23 s22 s23 s23 s24
= − 2 22 − 2 32 − 2 − 2 + 2
16 (s3 − s2 )3 (s3 − s4 )3 (s3 − s2 )2 (s3 − s4 ) (s3 − s4 )2 (s3 − s2 ) (s3 − s4 )2 (s23 − s25 )
2 2 2 2 2 2 2
s21 s22
+ 2 . (15)
(s3 − s22 )2 (s23 − s21 )
By solving the equation H (x, y ) = 0 in y, we obtain the Taylor expansion of h:
1 β
y = h(x) = s23 − x2 + x4 + O x5 .
4α 16α 4
Moreover the function
φ(x) = h(x) − x2
is expressed as
4α + 1 4β s23 − 16α 4 − 8α 3 − α 2
φ(x) = s3 − x2 + x4 + O x5 = s3 + α̃ x2 + β̃ x4 + O x5 .
8α s3 128α 4 s3
3
By using this expression, we compute the Taylor series of the q-numerical radius w = w q ( A ) near q = 0. We recall (cf. [6])
that
w q ( A ) = max qx + 1 − q2 φ(x): 0 x w 1 ( A ) .
The implicit function expression of w = w q ( A ) is obtained as the discriminant of the quartic polynomial
G (x) = − w + qx + 1 − q2 s3 + α̃ x2 + β̃ x4 ,
M.-T. Chien, H. Nakazato / J. Math. Anal. Appl. 345 (2008) 954–963 961
with respect to x. We introduce another parameter τ = tan(arc sin(q)/2) ∈ [0, 1]. Then we have the identities
2τ 1−τ 2
q= , 1 − q2 = .
1 + τ2 1 + τ2
The relation
w = s3 + c 2 τ 2 + c 4 τ 4 + O τ5
implies that
c2 c2 c3
w = s3 + q2 + + q4 + O q5 ,
4 8 16
since
q q3 q2 q4 q4
τ= + + O q5 , τ2 = + + O q6 , τ4 = + O q6 .
2 8 4 8 16
The discriminant (14) becomes
3 2 2
H̃ (τ , w ) = 16β̃ 2 τ2 + 1 τ 2 − 1 w 3 − 8β̃ α̃ 2 − 6β̃ s3 τ 2 − 1 τ 2 + 1 w2
2 2 2
+ τ 2 − 1 α̃ 4 − 16β̃ s3 τ 2 − 1 α̃ 2 + 36β̃ τ 2 α̃ + 48β̃ 2 s23 τ 2 − 1 w
4 2 4 2
+ s3 τ 2 − 1 α̃ 4 − τ τ 2 − 1 α̃ 3 + 36s3 β̃ τ τ 2 − 1 α̃ + β̃ 16β̃ s3 τ 2 − 1 − 27τ 4 .
2 2 3
Notice that the coefficients of the perturbations up to the fourth order are independent of the weights s j with | j − i | > 2,
where si > sk for all k = i.
Remark. The hypothesis s3 > max{s1 , s2 , s4 , . . . , sm } is essential for the proof of Theorem 4.1. For example, if m = 5 and
s1 = s3 > max{s2 , s4 , s5 }, then the expression β of (15) contains an undefined fraction s21 s22 /(s23 − s22 )2 (s23 − s21 ). Indeed, it
is shown in [6] that in this case the q-numerical radius w q ( A ) is strictly decreasing near q = 0 which is not applicable to
Theorem 4.1.
We may apply the perturbation method used in Theorem 4.1 for m = 4, and obtain the following analogous formula.
Theorem 4.2. Let A be a shift operator with periodic nonnegative weights {s1 , s2 , s3 , s4 } such that
s3 > max{s1 , s2 , s4 }.
Then the perturbation of the q-numerical radius is
(s23 − s22 )(s23 − s24 ) (4)
w = w q ( A ) = s3 − q2 + c 3 q4 + O q5 ,
2s3 (2s23 − s22 − s24 )
where
Theorem 4.3. Let A be a shift operator with periodic nonnegative weights {s1 , s2 , s3 } such that s3 > max{s1 , s2 }. Then, for sufficiently
small q, the perturbation of the q-numerical radius is
(s23 − s21 )(s23 − s22 ) s1 s2 (s23 − s21 )2 (s23 − s22 )2 γ (s23 − s21 )2 (s23 − s22 )2
w q ( A ) = s3 − q2 + q3 − q4 + O q5 ,
2s3 (2s23 − s21 − s22 ) s33 (2s23 − s21 − s22 )3 8s53 (2s23 − s21 − s22 )5
where
γ = 16s12 2 2 10 4 2 2 2 8 6 4 2 2 4 6 6
3 − 32 s1 + s2 s3 + 30s1 + 72s1 s2 + 30s2 s3 − 11s1 + 93s1 s2 + 93s1 s2 + 11s2 s3
+ s81 + 34s61 s22 + 162s41 s42 + 34s21 s62 + s82 s43 + 3s81 s22 − 75s61 s42 − 75s41 s62 + 3s21 s82 s23 + 36s61 s62 .
Proof. Under the hypotheses, the computations in Section 3 show that the perturbation eigenvalue λ3 ( B (k)) in (8) close to
s3 becomes
(2) (3) (4)
λ3 B (k) = s23 + c 3 k2 + c 3 k3 + c 3 k4 + O k5 ,
where
(2) s23 s22
c3 = + ,
4(s23 − s21 ) 4(s23 − s22 )
(3) s1 s2 s3
c3 = ,
4(s23 − s21 )(s23 − s22 )
(4) 1 s23 (s43 − s21 s22 )(s43 − (s21 + s22 )s23 + s41 − s21 s22 + s42 )
c3 = − .
16 (s23 − s21 )3 (s23 − s22 )3
5. An example
Consider m = 3 and s3 = 3, s2 = 2, s1 = 1. In this case the equation of the boundary of the range W ( ( A ) + i A ∗ A ) is
given by
By using this equation, we conclude that the q-numerical radius w = w q ( A ) of A satisfies the following equation for
0 q 1 (cf. [6]).
M.-T. Chien, H. Nakazato / J. Math. Anal. Appl. 345 (2008) 954–963 963
and thus c 2 = −20/39. By successive substitutions, we can determine the coefficients of the Taylor expansion. This result
coincides with the one deduced from Theorem 4.3 up to the fourth order.
Acknowledgment
The authors are grateful to the referee for his (or her) careful reading of the manuscript and valuable suggestions.
References
[1] M. Aleksiejczyk, On the diameter of the generalized numerical range, Demonstratio Math. 30 (1997) 129–136.
[2] H. Baumgartel, Analytic Perturbation Theory for Matrices and Operators, Birkhäuser, Basel, 1985.
[3] J.W. Bunce, J.A. Deddens, A family of simple C ∗ -algebras related to weighted shift operators, J. Funct. Anal. 19 (1975) 13–24.
[4] R. Benedetti, J.J. Risler, Real Algebraic and Semi-Algebraic Sets, Hermann, Paris, 1990.
[5] M.T. Chien, H. Nakazato, Davis–Wielandt shell and q-numerical range, Linear Algebra Appl. 340 (2002) 15–31.
[6] M.T. Chien, H. Nakazato, The q-numerical radius of weighted shift operators with periodic weights, Linear Algebra Appl. 422 (2007) 198–218.
[7] M.T. Chien, H. Nakazato, The boundary of the q-numerical range of a reducible matrix, Linear Multilinear Algebra 55 (2007) 275–292.
[8] R. Curto, Quadratically hyponormal weighted shifts, Integral Equations Operator Theory 13 (1990) 49–66.
[9] C. Davies, The shell of a Hilbert space operators II, Acta Sci. Math. 31 (1970) 301–318.
[10] E. Gutkin, E.A. Jonckheere, M. Karow, Convexity of the joint numerical range: Topological and differential geometric viewpoints, Linear Algebra
Appl. 376 (2004) 143–171.
[11] I.B. Jung, S. Park, Cubically hyponormal weighted shifts and their examples, J. Math. Anal. Appl. 247 (2000) 557–569.
[12] T. Kato, A Short Introduction to Perturbation Theory for Linear Operators, Springer-Verlag, New York, 1982.
[13] C.K. Li, H. Nakazato, Some results on the q-numerical ranges, Linear Multilinear Algebra 43 (1998) 385–410.
[14] P.J. Psarrakos, Definite triples of Hermitian matrices and matrix polynomials, J. Comput. Appl. Math. 151 (2003) 39–58.
[15] P.J. Psarrakos, On the estimation of the q-numerical range of monic matrix polynomials, Electron. Trans. Numer. Anal. 17 (2004) 1–10.
[16] P.J. Psarrakos, P.M. Vlamos, The q-numerical range of matrix polynomials, Linear Multilinear Algebra 47 (2000) 1–9.
[17] L.I. Shiff, Quantum Mechanics, McGraw–Hill, New York, 1968.
[18] Q.F. Stout, The numerical range of a weighted shift, Proc. Amer. Math. Soc. 88 (1983) 495–502.
[19] T.Y. Tam, On a conjecture of Ridge, Proc. Amer. Math. Soc. 125 (1997) 3581–3592.
[20] N.K. Tsing, The constrained bilinear form and C -numerical range, Linear Algebra Appl. 56 (1984) 195–206.