arXiv:submit/1087246 [math.

NA] 11 Oct 2014

On the distance from a weakly normal matrix polynomial to

matrix polynomials with a prescribed multiple eigenvalue
E. Kokabifar∗, G.B. Loghmani∗ and P.J. Psarrakos†

October 11, 2014

Abstract
Consider an n × n matrix polynomial P (λ). An upper bound for a spectral norm
distance from P (λ) to the set of n×n matrix polynomials that have a given scalar µ ∈ C
as a multiple eigenvalue was recently obtained by Papathanasiou and Psarrakos (2008).
This paper concerns a refinement of this result for the case of weakly normal matrix
polynomials. A modification method is implemented and its efficiency is verified by
an illustrative example.

Keywords: Matrix polynomial, Eigenvalue, Normality, Perturbation, Singular value.

AMS Classification: 15A18, 65F35.

1 Introduction
Let A be an n × n complex matrix and µ be a complex number, and denote by Mµ the
set of n × n complex matrices that have µ ∈ C as a multiple eigenvalue. Malyshev [10]
obtained the following formula for the spectral norm distance from A to Mµ :
 
A − µI γIn
min kA − Bk2 = max s2n−1 ,
B∈Mµ γ≥0 0 A − µI

where k · k2 denotes the spectral matrix norm (i.e., that norm subordinate to the euclidean
vector norm) and s1 (·) ≥ s2 (·) ≥ s3 (·) ≥ · · · are the singular values of the corresponding
matrix in nonincreasing order. Malyshev’s work can be considered as a theoretical solution
to Wilkinson’s problem, that is, the calculation of the distance from a matrix A ∈ Cn×n
∗
Department of Mathematics, Faculty of Science, Yazd University, Yazd, Iran ([email protected],
[email protected]).
†
Department of Mathematics, National Technical University of Athens, Zografou Campus, 15780
Athens, Greece ([email protected]).

1
that has all its eigenvalues simple to the n × n matrices with multiple eigenvalues. Wilkin-
son introduced this distance in [17], and some bounds for it were computed by Ruhe [15],
Wilkinson [18–21] and Demmel [1].
However, in the non-generic case where A is a normal matrix, Malyshev’s formula is
not directly applicable. In 2004, Ikramov and Nazari [4] showed this point and obtained
an extension of Malyshev’s method for normal matrices. Moreover, Malyshev’s results
were extended by Lippert [9] and Gracia [3]; in particular, they computed a spectral norm
distance from A to the set of matrices that have two prescribed eigenvalues and studied
a nearest matrix with the two desired eigenvalues. Nazari and Rajabi [12] refined the
method obtained by Lippert and Gracia for the case of normal matrices.
In 2008, Papathanasiou and Psarrakos [14] introduced and studied a spectral norm
distance from a n × n matrix polynomial P (λ) to the set of n × n matrix polynomials
that have a scalar µ ∈ C as a multiple eigenvalue. In particular, generalizing Malyshev’s
methodology, they computed lower and upper bounds for this distance, constructing an
associated perturbation of P (λ) for the upper bound. Motivated by the above, in this note,
we study the case of weakly normal matrix polynomials. In the next section, we give some
definitions and present briefly some of the results of [13,14]. We also give an example of a
normal matrix polynomial where the method described in [14] for the computation of the
upper bound is not directly applicable. In Section 3, we prove that the methodology of [14]
for the computation of the upper bound is indeed not directly applicable to weakly normal
matrix polynomials, and in Section 4, we obtain a modified procedure for weakly normal
matrix polynomials to improve the method. The same numerical example is considered
to illustrate the validity of the proposed technique.

2 Preliminaries
For A0 , A1 , . . . , Am ∈ Cn×n , with det(Am ) 6= 0, and a complex variable λ, we define the
matrix polynomial

P (λ) = Am λm + Am−1 λm−1 + · · · + A1 λ + A0 . (1)

The study of matrix polynomials, especially with regard to their spectral analysis, has
received a great deal of attention and has been used in several applications [2, 6, 7, 11, 16].
Standard references for the theory of matrix polynomials are [2,11]. Here, some definitions
of matrix polynomials are briefly reviewed.
If for a scalar λ0 ∈ C and some nonzero vector x0 ∈ Cn , it holds that P (λ0 )x0 = 0,
then the scalar λ0 is called an eigenvalue of P (λ) and the vector x0 is known as a (right)
eigenvector of P (λ) corresponding to λ0 . The spectrum of P (λ), denoted by σ(P ), is
the set of all eigenvalues of P (λ). Since the leading matrix-coefficient Am is nonsingular,

2
the spectrum σ(P ) contains at most mn distinct finite elements. The multiplicity of an
eigenvalue λ0 ∈ σ(P ) as a root of the scalar polynomial det P (λ) is said to be the algebraic
multiplicity of λ0 , and the dimension of the null space of the (constant) matrix P (λ0 ) is
known as the geometric multiplicity of λ0 . Algebraic multiplicity of an eigenvalue is always
greater than or equal to its geometric multiplicity. An eigenvalue is called semisimple if
its algebraic and geometric multiplicities are equal; otherwise, it is known as defective.

Definition 2.1. Let P (λ) be a matrix polynomial as in (1). If there exists a unitary
matrix U ∈ Cn×n such that U ∗ P (λ)U is a diagonal matrix polynomial, then P (λ) is said
to be weakly normal. If, in addition, all the eigenvalues of P (λ) are semisimple, then P (λ)
is called normal.

The suggested references on weakly normal and normal matrix polynomials, and their
properties are [8, 13]. Some of the results of [13] are summarized in the next proposition.

Proposition 2.2. [13] Let P (λ) = Am λm + · · · + A1 λ + A0 be a matrix polynomial as in

(1). Then P (λ) is weakly normal if and only if one of the following (equivalent) conditions
holds.

(i) For every µ ∈ C, the matrix P (µ) is normal.

(ii) A0 , A1 , . . . , Am are normal and mutually commuting (i.e., Ai Aj = Aj Ai for i 6= j).

(iii) All the linear combinations of A0 , A1 , . . . , Am are normal matrices.

(iv) There exists a unitary matrix U ∈ Cn×n such that U ∗ Aj U is diagonal for every
j = 0, 1, . . . , m.

As mentioned, Papathanasiou and Psarrakos [14] introduced a spectral norm distance

from a matrix polynomial P (λ) to the matrix polynomials that have µ as a multiple
eigenvalue, and computed lower and upper bounds for this distance. Consider (additive)
perturbations of P (λ) of the form

Q(λ) = P (λ) + ∆(λ) = (Am + ∆m )λm + · · · + (A1 + ∆1 )λ + A0 + ∆0 , (2)

where the matrices ∆0 , ∆1 , . . . , ∆m ∈ Cn×n are arbitrary. For a given parameter ǫ > 0
and a given set of nonnegative weights w = {w0 , w1 , . . . , wm } with w0 > 0, define the class
of admissible perturbed matrix polynomials

B(P, ǫ, w) = {Q(λ) as in (2) : k∆j k2 ≤ ǫ wj , j = 0, 1, . . . , m}

and the scalar polynomial w(λ) = wm λm + wm−1 λm−1 + · · · + w1 λ + w0 . Note that the
weights w0 , w1 , . . . , wm allow freedom in how perturbations are measured.

3
 For any real number γ ∈ [0, +∞), we define the 2n × 2n matrix polynomial
 
P (λ) 0
F [P (λ); γ] = ,
γP ′ (λ) P (λ)

where P ′ (λ) denotes the derivative of P (λ) with respect to λ.

Lemma 2.3. [14, Lemma 17] Let µ ∈ C and γ∗ > 0 be a point where the singular value
s2n−1 (F [P (µ); γ]) attains its maximum
  value, and denote s∗ = s2n−1 (F [P (µ); γ∗ ]) > 0.
u1 (γ∗ ) v1 (γ∗ )
Then there exists a pair , ∈ C2n (uk (γ∗ ), vk (γ∗ ) ∈ Cn , k = 1, 2) of
u2 (γ∗ ) v2 (γ∗ )
left and right singular vectors of F [P (µ); γ∗ ] corresponding to s∗ , respectively, such that

(1) u∗2 (γ∗ )P ′ (µ)v1 (γ∗ ) = 0, and

(2) the n × 2 matrices U (γ∗ ) = [u1 (γ∗ ) u2 (γ∗ )] and V (γ∗ ) = [v1 (γ∗ ) v2 (γ∗ )] satisfy
U ∗ (γ∗ )U (γ∗ ) = V ∗ (γ∗ )V (γ∗ ).

Moreover, it is remarkable that (1) implies (2) (see the proof of Lemma 17 in [14]).
′
(|µ|) µ̄
Consider the quantity φ = ww(|µ|) µ̄
|µ| , where, by convention, we set |µ| = 0 whenever
µ = 0. Let alsoV (γ∗ )† be  the
 Moore-Penrose
 pseudoinverse of V (γ∗ ). For the pair of
u1 (γ∗ ) v1 (γ∗ )
singular vectors , ∈ C2n of Lemma 2.3, define the n × n matrix
u2 (γ∗ ) v2 (γ∗ )
 
1 −γ∗ φ
∆γ∗ = −s∗ U (γ∗ ) V (γ∗ )† .
0 1

Theorem 2.4. [14, Theorem 19] Let P (λ) be a matrix polynomial as in (1), and let
w = {w0 , w1 , . . . , wm }, with w0 > 0, be a set of nonnegative weights. Suppose that
µ ∈ C\σ(P ′ ), γ∗ > 0 is a point where the singular value s2n−1 (F [P (µ); γ]) attains its
maximumvalue,
  and s∗ = s2n−1 (F [P (µ); γ∗ ]) > 0. Then, for the pair of singular vectors
u1 (γ∗ ) v1 (γ∗ )
, ∈ C2n of Lemma 2.3, we have
u2 (γ∗ ) v2 (γ∗ )

min {ǫ ≥ 0 : ∃ Q(λ) ∈ B(P, ǫ, w) with µ as a multiple eigenvalue}

 
s∗ 1 −γ ∗ φ †

≤ V (γ∗ ) V (γ∗ ) .
w(|µ|) 0 1

Moreover, the perturbed matrix polynomial

m  j
X wj µ
Qγ∗ (λ) = P (λ) + ∆γ∗ (λ) = P (λ) + ∆γ∗ λj , (3)
w(|µ|) |µ|
j=0

4
    
s∗ V (γ∗ ) 1 −γ∗ φ V (γ∗ )† , w and has µ

lies on the boundary of the set B P, w(|µ|) 0 1
as a (multiple) defective eigenvalue.

Some numerical examples in Section 8 of [14] illustrate the effectiveness of the upper
bound of Theorem 2.4. In all these examples,
  s∗ is a simple singular value, and con-
u1 (γ∗ ) v1 (γ∗ )
sequently, the singular vectors , ∈ C2n of Lemma 2.3 are directly
u2 (γ∗ ) v2 (γ∗ )
computable (due to their essential uniqueness). Let us now consider the normal (in par-
ticular, diagonal) matrix polynomial
     
1 0 0 −3 0 0 2 0 0
P (λ) =  0 1 0  λ2 +  0 −1 0  λ +  0 0 0  (4)
0 0 1 0 0 3 0 0 2

that is borrowed from [13, Section 3]. Let also the set of weights w = {1, 1, 1} and the scalar
µ = −4. The singular value s5 (F [P (−4); γ]) attains its maximum value at γ∗ = 2.0180,
and at this point, we have s∗ = s5 (F [P (−4); 2.0180]) = s4 (F [P (−4); 2.0180]) = 12.8841;
i.e., s∗ is a multiple singular value of matrix F [P (−4); 2.0180]. A left and a right singular
vectors of F [P (−4); 2.0180] corresponding to s∗ are

0 0
   

  0.8407    0.5416 
   
 
u1 (γ∗ )  0  v1 (γ∗ )  0 
=  and = ,
u2 (γ∗ ) 
 0 
 v2 (γ∗ ) 
 0 

 0.5416   0.8407 
0 0

respectively, and they yield the perturbed matrix polynomial (see (3))
     
1 0 0 −3 0 0 2 0 0
Qγ∗ (λ) =  0 0.0664 0  λ2 +  0 −0.0664 0  λ +  0 −0.9336 0  .
0 0 1 0 0 3 0 0 2

One can see that µ = −4 is not a multiple eigenvalue of Qγ∗ (λ). Moreover, proper-
ties (1) and (2) of Lemma 2.3 do not hold since u∗2 (γ∗ )P ′ (µ)v1 (γ∗ ) = −2.6396 6= 0 and
kU ∗ (γ∗ )U (γ∗ ) − V ∗ (γ∗ )V (γ∗ )k2 = 0.4134 6= 0.
Clearly, this example verifies that the computation of appropriate singular vectors
which satisfy (1) and (2) of Lemma 2.3 is still an open problem when s∗ is a multiple
singular value. In the next section, we obtain that for weakly normal matrix polynomials,
s∗ is always a multiple singular value, and in Section 4, we solve the problem of calculation
of the desired singular vectors of Lemma 2.3.

5
3 Weakly normal matrix polynomials
In this section, by extending the analysis performed in [5], we prove that s∗ is always a
multiple singular value of F [P (µ); γ∗ ] when P (λ) is a weakly normal matrix polynomial.
Let P (λ) be a weakly normal matrix polynomial, and let µ ∈ C\σ(P ′ ). By Proposi-
tion 2.2 (iv), it follows that there exists a unitary matrix U ∈ Cn×n such that all matrices
U ∗ A0 U, U ∗ A1 U, . . . , U ∗ Am U are diagonal. Hence, U ∗ P (µ)U and U ∗ P (µ)′ U are also diag-
onal matrices; in particular,

U ∗ P (µ)U = diag{ζ1 , ζ2 , . . . , ζn } and U ∗ P (µ)′ U = diag{ξ1 , ξ2 , . . . , ξn },

where all scalars ξ1 , ξ2 , . . . , ξn ∈ C are nonzero (recall that P ′ (µ) is nonsingular) and,
without loss of generality, we assume that

|ζ1 | ≥ |ζ2 | ≥ · · · ≥ |ζn | .

As a consequence,
 ∗
U∗ 0
      
U 0 U 0 P (µ) 0 U 0
∗ F [P (µ); γ] =
0 U 0 U 0 U∗ ′
γP (µ) P (µ) 0 U
 
diag{ζ1 , ζ2 , . . . , ζn } 0
= .
γ diag{ξ1 , ξ2 , . . . , ξn } diag{ζ1 , ζ2 , . . . , ζn }

It is straightforward to verify that there is a 2n × 2n permutation matrix R such that

 
diag{ζ1 , ζ2 , . . . , ζn } 0
R RT
γ diag{ξ1 , ξ2 , . . . , ξn } diag{ζ1 , ζ2 , . . . , ζn }
     
ζ1 0 ζ2 0 ζn 0
= ⊕ ⊕ ··· ⊕ .
γ ξ1 ζ1 γ ξ2 ζ2 γ ξn ζn

The fact that singular values of a matrix are invariant under unitary similarity implies
that the 2n × 2n matrices
     
ζ1 0 ζ2 0 ζn 0
F [P (µ); γ] and ⊕ ⊕ ··· ⊕
γ ξ1 ζ1 γ ξ2 ζ2 γ ξn ζn

have the same

 singular
 values.
 Therefore,
 in what follows, we are focused on the singular
ζ1 0 ζ2 0 ζn 0
values of ⊕ ⊕ ··· ⊕ , which are the union of the
γ ξ1 ζ1 γ ξ2 ζ2 γ ξn ζn
 
ζi 0
singular values of , i = 1, 2, . . . , n.
γ ξi ζi

6
  
ζi 0
For any i = 1, 2, . . . , n, let si,1 (γ) ≥ si,2 (γ) be the singular values of , and
γ ξi ζi
consider the characteristic polynomial of matrix
∗ 
|ζi |2 + γ 2 |ξi |2 γ ξ i ζi
   
ζi 0 ζi 0
= ,
γ ξi ζi γ ξi ζi γ ξi ζ i |ζi |2

that is,

|ζi |2 + γ 2 |ξi |2 γ ξ i ζi
    
det tI − = t2 − 2 |ζi |2 + γ 2 |ξi |2 t + |ζi |4 .
γ ξi ζ i |ζi |2

|ζi |2 + γ |ξi |2 γ ξ i ζi
 
The positive square roots of the eigenvalues of matrix are the
γ ξi ζ i |ζi |2
 
ζi 0
singular values of matrix , namely,
γ ξi ζi
v s
u
γ 2 |ξi |2 γ 2 |ξi |2
u
si,1 (γ) = |ζi |2 + + γ |ξi | |ζi |2 + ,
t
2 4

and
v s
u
γ 2 |ξi |2 γ 2 |ξi |2
u
2 2
si,2 (γ) = |ζi | + − γ |ξi | |ζi | + .
t
2 4

As γ ≥ 0 increases, si,1 (γ) increases and lim si,1 (γ) = +∞, while si,2 (γ) decreases and
γ→+∞
lim si,2 (γ) = 0 (recall that |ξi | > 0, i = 1, 2, . . . , n). Also, it is apparent that
γ→+∞

si,2 (γ) ≤ |ζi | ≤ si,1 (γ), and si,1 (0) = si,2 (0) = |ζi | .

Next we consider two cases with respect to |ζn−1 | and |ζn |.

Case 1. Suppose |ζn | < |ζn−1 |. At γ = 0, it holds that sn,1 (0) = |ζn | < |ζn−1 | =
sn−1,2 (0). According to the above discussion, as the nonnegative variable γ increases from
zero, the functions

s1,1 (γ), s2,1 (γ), . . . , sn−1,1 (γ), sn,1 (γ)

increase to +∞, whereas the functions

s1,2 (γ), s2,2 (γ), . . . , sn−1,2 (γ), sn,2 (γ)

7
decrease to 0. Let (γ0 , s0 ) be the first point in R2 where the graph of the increasing function
sn,1 (γ) intersects the graph of one of the n−1 decreasing functions s1,2 (γ), s2,2 (γ), . . . , sn−1,2 (γ),
say sκ,2 (γ) (for some κ ∈ {1, 2, . . . , n − 1}). Note that by the definition of si,1 (γ) and
si,2 (γ) (i = 1, 2, . . . , n), s0 lies in the open interval (0, |ζn−1 |) and the graph of sn,1 (γ)
cannot intersect the graph of one of the increasing functions s1,1 (γ), s2,1 (γ), . . . , sn−1,1 (γ)
for γ ≤ γ0 .
Since sn,2 (γ) and sκ,2 (γ) are both decreasing functions in γ ≥ 0, it follows that (see
Fig. 1 below, where κ = n − 1 = 2)

γ∗ = γ0 and s∗ = s0 = s2n−1 (F [P (µ); γ∗ ]) = sn,1 (γ∗ ) = sκ,2 (γ∗ ) = s2n−2 (F [P (µ); γ∗ ]).

Hence, when |ζn | < |ζn−1 |, γ∗ is the minimum positive root of one of the equations

sn,1 (γ) = sn−1,2 (γ), sn,1 (γ) = sn−2,2 (γ), . . . , sn,1 (γ) = s1,2 (γ)

and s∗ is a multiple singular value of F [P (µ); γ∗ ].

Case 2. Suppose |ζn | = |ζn−1 |. Then, it follows that sn,1 (γ) = sn−1,1 (γ) and sn,2 (γ) =
sn−1,2 (γ). Moreover, one can see that at γ = 0,

sn,1 (0) = sn,2 (0) = sn−1,1 (0) = sn−1,2 (0) = |ζn | = |ζn−1 | ,

i.e.,

s2n (F [P (µ); 0]) = s2n−1 (F [P (µ); 0]) = s2n−2 (F [P (µ); 0]) = s2n−3 (F [P (µ); 0]) = |ζn | = |ζn−1 | .

Since sn,2 (γ) and sn−1,2 (γ)) are decreasing functions in γ ≥ 0, s2n−1 (F [P (µ); γ]) attains
its maximum value s∗ at γ = 0 = γ∗ , and s∗ is a multiple singular value of F [P (µ); 0]. In
this non-generic case, an upper bound and an associate perturbed matrix polynomial can
be computed by the method described in Section 6 of [14].
Hence, we have the following result.
Theorem 3.1. Let P (λ) in (1) be a weakly normal matrix polynomial, and let µ ∈
C\σ(P ′ ). If γ∗ > 0 is a point where the singular value s2n−1 (F [P (µ); γ]) attains its maxi-
mum value, then s∗ = s2n−1 (F [P (µ); γ∗ ]) > 0 is a multiple singular value of F [P (µ); γ∗ ].

4 Computing the desired singular vectors

In this section, we apply a technique proposed in [5] (see also the proof of Lemma 5
in [10]) to compute suitable singular vectors of F [P (µ); γ∗ ] corresponding to the singular
value s∗ , which satisfy (1) and (2) of Lemma 2.3. It is remarkable that the proposed
technique can be applied to general matrix polynomials and not only to weakly normal
matrix polynomials.

8
4.1 The case of multiplicity 2
First we consider the case where γ∗ > 0 and the multiplicity of the singular value s∗ > 0
is equal to 2, and we work on the example of Section 2.
Recall that for the normal matrix polynomial P (λ) in (4) and for µ = −4, the singular
value s2n−1 (F [P (µ); γ]) = s5 (F [P (−4); γ]) attains its maximum value at γ∗ = 2.0180
and s∗ = s5 (F [P (−4); 2.0180]) = s4 (F [P (−4); 2.0180]) = 12.8841 (i.e., s∗ is a double
singular value of F [P (−4); 2.0180]). Two pairs of left and a right singular vectors of
F [P (−4); 2.0180] corresponding to s∗ , which do not satisfy properties (1) and (2) of Lemma
2.3 are
0 0
   

  0.8407     0.5416 
   

u1 (γ∗ )  0  v1 (γ∗ )  0 
= , = ,
u2 (γ∗ ) 
 0 
 v2 (γ∗ ) 
 0 

 0.5416   0.8407 
0 0

and
0 0
   

 

 0 
 

  0 

û1 (γ∗ )  −0.4222  v̂1 (γ∗ )  −0.9065 
= , = .
û2 (γ∗ ) 
 0 
 v̂2 (γ∗ ) 
 0 

 0   0 
0.9065 0.4222

In particular, we have

u2 (γ∗ )∗ P ′ (−4)v1 (γ∗ ) = −2.6396 6= 0 and û2 (γ∗ )∗ P ′ (−4)v̂1 (γ∗ ) = 4.1089 6= 0.

In Figure 1, the graphs of

s2n−1 (F [P (µ); γ]) = s5 (F [P (−4); γ]) and s2n−2 (F [P (µ); γ]) = s4 (F [P (−4); γ])

are plotted for γ ∈ [0, 10], and their common point (γ∗ , s∗ ) = (2.0180, 12.8841) is marked
with “◦”. With respect to the discussion in the previous section, it is worth noting that
in this example, the graph of s2,2 (γ) (that is, sn−1,2 (γ)) is the graph of the decreasing
functions s1,2 (γ) and s2,2 (γ) that intersects first the graph of the increasing function s3,1 (γ)
(that is, sn,1 (γ)). Moreover, it is apparent that s2n−1 (F [P (µ); γ]) and s2n−2 (F [P (µ); γ])
are non-differentiable functions at γ∗ .

9
 20
s2n−1(γ)
18 s (γ)
2n−2

16

14
o
12
(γ ,s )
* *

10

4
0 1 2 3 4 5 6 7 8 9 10
γ

Fig 1: The singular values s2n−1 (F [P (µ); γ]) (solid line) and s2n−2 (F [P (µ); γ]) (dashed line).

Since s∗ = s5 (F [P (−4); 2.0180]) =  s4 (F [P (−4);

  2.0180])  = 12.8841
 is a double
 sin-

u1 (γ∗ ) û1 (γ∗ ) v1 (γ∗ ) v̂1 (γ∗ )
gular value, the pairs of unit vectors , and ,
u2 (γ∗ ) û2 (γ∗ ) v2 (γ∗ ) v̂2 (γ∗ )
form orthonormal bases of the left and right singular subspaces corresponding to s∗ , re-
spectively. So, recalling that in Lemma 2.3, assertion (1) yields assertion (2), henceforth
we are looking for a pair of unit vectors
           
ũ1 (γ∗ ) u1 (γ∗ ) û1 (γ∗ ) ṽ1 (γ∗ ) v1 (γ∗ ) v̂1 (γ∗ )
=α +β , =α +β (5)
ũ2 (γ∗ ) u2 (γ∗ ) û2 (γ∗ ) ṽ2 (γ∗ ) v2 (γ∗ ) v̂2 (γ∗ )

such that

ũ2 (γ∗ )∗ P ′ (µ)ṽ1 (γ∗ ) = 0, (6)

where the scalars α, β ∈ C satisfy |α|2 + |β|2 = 1. By substituting the unknown singular
vectors of (5) into (6), we obtain
 
  α
α β M = 0, (7)
β

where
u2 (γ∗ )∗ P ′ (µ)v1 (γ∗ ) u2 (γ∗ )∗ P ′ (µ)v̂1 (γ∗ )
 
M= . (8)
û2 (γ∗ )∗ P ′ (µ)v1 (γ∗ ) û2 (γ∗ )∗ P ′ (µ)v̂1 (γ∗ )

Lemma 4.1. The matrix M in (8) is always hermitian.

10
Proof. Recall that γ∗ and s∗ are positive. By the proof of Lemma 17 in [14], it follows
that the diagonal entries of matrix M are real.
By the definition of the pairs of singular vectors
       
u1 (γ∗ ) v1 (γ∗ ) û1 (γ∗ ) v̂1 (γ∗ )
, and ,
u2 (γ∗ ) v2 (γ∗ ) û2 (γ∗ ) v̂2 (γ∗ )

of F [P (µ); γ∗ ] corresponding to s∗ , we have

     
P (µ) 0 v 1 (γ ∗ ) u1 (γ∗ )
= s∗ ,

′

 γ∗ P (µ) P (µ)   v2 (γ∗ )  u2 (γ∗ ) 


P (µ) 0 v̂1 (γ∗ ) û1 (γ∗ )
= s∗ ,


γ∗ P ′ (µ) P (µ) v̂2 (γ∗ ) û2 (γ∗ )


or equivalently,

 P (µ)v1 (γ∗ ) = s∗ u1 (γ∗ ),
γ∗ P ′ (µ)v1 (γ∗ ) + P (µ)v2 (γ∗ ) = s∗ u2 (γ∗ ),


(9)

 P (µ)v̂1 (γ∗ ) = s∗ û1 (γ∗ ),
γ∗ P ′ (µ)v̂1 (γ∗ ) + P (µ)v̂2 (γ∗ ) = s∗ û2 (γ∗ ),


and
  
∗ ∗ P (µ) 0
= s∗ v1 (γ∗ )∗ v2 (γ∗ )∗ ,
   
 u1 (γ∗ ) u2 (γ∗ )

′

 γ∗ P (µ) P (µ) 
P (µ) 0
 û1 (γ∗ )∗ û2 (γ∗ )∗ = s∗ v̂1 (γ∗ )∗ v̂2 (γ∗ )∗ ,
    

′
γ∗ P (µ) P (µ)

or equivalently,

u1 (γ∗ )∗ P (µ) + γ∗ u2 (γ∗ )∗ P ′ (µ) = s∗ v1 (γ∗ )∗ ,



u2 (γ∗ )∗ P (µ) = s∗ v2 (γ∗ )∗ ,


(10)
û (γ )∗ P (µ) + γ∗ û2 (γ∗ )∗ P ′ (µ) = s∗ v̂1 (γ∗ )∗ ,
 1 ∗ ∗


û2 (γ∗ ) P (µ) = s∗ v̂2 (γ∗ )∗ ,

By multiplying the fourth equation in (9) by u2 (γ∗ )∗ from the left, and the second
equation of (10) by v̂2 (γ∗ ) from the right, we obtain

γ∗ u2 (γ∗ )∗ P ′ (µ)v̂1 (γ∗ ) + u2 (γ∗ )∗ P (µ)v̂2 (γ∗ ) = s∗ u2 (γ∗ )∗ û2 (γ∗ ), (11)

and

u2 (γ∗ )∗ P (µ)v̂2 (γ∗ ) = s∗ v2 (γ∗ )∗ v̂2 (γ∗ ), (12)

11
respectively. As a consequence,

γ∗ u2 (γ∗ )∗ P ′ (µ)v̂1 (γ∗ ) = s∗ (u2 (γ∗ )∗ û2 (γ∗ ) − v2 (γ∗ )∗ v̂2 (γ∗ )) . (13)

Performing similar calculations, one can verify that

γ∗ û2 (γ∗ )∗ P ′ (µ)v1 (γ∗ ) = s∗ (û2 (γ∗ )∗ u2 (γ∗ ) − v̂2 (γ∗ )∗ v2 (γ∗ )) . (14)

Clearly, equations (13) and (14) imply that the non-diagonal entries of matrix M are
complex conjugate.

By Lemma 2.3 (1), equation (7) has always a nontrivial (i.e., nonzero) solution, and
hence, the hermitian matrix M in (8) cannot be (positive or negative) definite. In our
numerical example, M has a negative and a positive diagonal entries (namely, −2.6396
and 4.1089), and thus, it is an indefinite hermitian matrix.
To derive an explicit solution of (7), suppose that η1 , η2 ∈ C are the (real) eigenvalues
of matrix M , with η1 > 0 > η2 , and let w1 , w2 ∈ C2 be unit eigenvectors of M corre-
sponding to η1 and η2 , respectively. Then, it is straightforward to see (keeping in mind
the orthogonality of the eigenvectors) that the unit vector
  s s
α |η2 | |η1 |
= w1 + w2
β |η1 | + |η2 | |η1 | + |η2 |

satisfies
 
  α |η1 | η2 |η2 | η1
α β M = + = 0.
β |η1 | + |η2 | |η1 | + |η2 |

Finally, in order to verify the validity of this refinement, we return again to the
normal matrix polynomial P (λ) in (4), and by applying the above methodology, we obtain
α = 0.6254 and β = 0.7803. Consequently, the desired vectors in (5) are (approximately)

0 0
   

  0.6560    0.4226 
   
 
ũ1 (γ∗ )  −0.2640  ṽ1 (γ∗ )  −0.5669 
= and = .
ũ2 (γ∗ ) 
 0 
 ṽ2 (γ∗ ) 
 0 

 0.4226   0.6560 
0.5669 0.2640

In particular, it holds that

ũ∗2 (γ∗ )P ′ (−4)ṽ1 (γ∗ ) = −4.4409 · 10−16 ,

12
and for the n × 2 matrices Ũ (γ∗ ) = [ũ1 (γ∗ ) ũ2 (γ∗ )] and Ṽ (γ∗ ) = [ṽ1 (γ∗ ) ṽ2 (γ∗ )], we have

∗ ∗ −6
Ũ (γ )Ũ (γ ) − Ṽ (γ )Ṽ (γ∗ = 1.1383 · 10 .
)

∗ ∗ ∗
2

Thus, Lemma 2.3 is verified.

Moreover, using the matrices Ũ (γ∗ ) and Ṽ (γ∗ ), Theorem 2.4 yields the upper bound
0.9465 for the distance from P (λ) to the set of 3 × 3 quadratic matrix polynomials that
have µ = −4 as a multiple eigenvalue, and the perturbed matrix polynomial
     
1 0 0 −3 0 0 2 0 0
Q̃γ∗ (λ) =  0 0.0680 0.0152  λ2 + 0 −0.0680 −0.0152  λ+ 0 −0.9320 0.0152 
0 −0.1552 0.5986 0 0.1552 3.4014 0 −0.1552 1.5986

that lies on the boundary of B (P, 0.9465, w) and has spectrum

 
σ Q̃γ∗ (λ) = {1, 2, 4.1982, −0.5140, −4.0000 + i 0.0031, −4.0000 − i 0.0031} .

In addition, the lower bound 0.4031 of the distance is given by Theorem 11 in [14]. (All
computations were performed in Matlab with 16 significant digits.)

4.2 The case of multiplicity greater that 2

Suppose that γ∗ > 0, and the multiplicity of the singular value s∗ > 0 is r ≥ 3. For
weakly normal matrix polynomials this means that the graph of the increasing func-
tion sn,1 (γ) intersects the graphs of more than one of the n − 1 decreasing functions
s1,2 (γ), s2,2 (γ), . . . , sn−1,2 (γ), at the point (γ∗ , s∗ ). Let also
" # " # " #
(1) (2) (r)
u1 (γ∗ ) u1 (γ∗ ) u1 (γ∗ )
(1) , (2) , ..., (r)
u2 (γ∗ ) u2 (γ∗ ) u2 (γ∗ )

and
" # " # " #
(1) (2) (r)
v1 (γ∗ ) v1 (γ∗ ) v1 (γ∗ )
(1) , (2) , ..., (r)
v2 (γ∗ ) v2 (γ∗ ) v2 (γ∗ )

be orthonormal bases of the left and right singular subspaces of F [P (µ); γ∗ ] corresponding
to s∗ , respectively. Then, we are looking for a pair of unit vectors
r r
" #  " #
  X (j)  X (j)
ũ1 (γ∗ ) u1 (γ∗ ) ṽ1 (γ∗ ) v1 (γ∗ )
= αj (j) , = αj (j) (15)
ũ2 (γ∗ ) u2 (γ∗ ) ṽ2 (γ∗ ) v2 (γ∗ )
j=1 j=1

13
such that

ũ2 (γ∗ )∗ P ′ (µ)ṽ1 (γ∗ ) = 0, (16)

where the scalars α1 , α2 , . . . , αr ∈ C satisfy |α1 |2 + |α2 |2 + · · · + |αr |2 = 1.

Following the arguments of the methodology described in the previous subsection, we
can compute the desired vectors in (15) that satisfy (16). In particular, we need to find a
solution of the equation
 
α1
   α2 
α1 α2 · · · αr Mr  .  = 0, (17)
 
 .. 
αr

where the r × r matrix

 (1) (1) (1) (2) (1) (r)

u2 (γ∗ )∗ P ′ (µ)v1 (γ∗ ) u2 (γ∗ )∗ P ′ (µ)v1 (γ∗ ) · · · u2 (γ∗ )∗ P ′ (µ)v1 (γ∗ )
 (2)
 u2 (γ∗ )∗ P ′ (µ)v1(1) (γ∗ ) u(2) ∗ ′ (2)
2 (γ∗ ) P (µ)v1 (γ∗ ) · · ·
(2) (r)
u2 (γ∗ )∗ P ′ (µ)v1 (γ∗ )


Mr =  .. .. .. .. 
. . . .

 
(r) (1) (r) (2) (r) (r)
u2 (γ∗ )∗ P ′ (µ)v1 (γ∗ ) u2 (γ∗ )∗ P ′ (µ)v1 (γ∗ ) · · · u2 (γ∗ )∗ P ′ (µ)v1 (γ∗ )

is hermitian and not definite. Considering a unit eigenvector wmax ∈ Cr of Mr corre-

sponding to the maximum eigenvalue ηmax > 0 of Mr and a unit eigenvector wmin ∈ Cr
corresponding to the minimum eigenvalue ηmin < 0 of Mr , it is straightforward to verify
that the unit vector
s s
|ηmin | |ηmax |
wmax + wmin
|ηmax | + |ηmin | |ηmax | + |ηmin |

satisfies (17).

References
[1] J.W. Demmel, On condition numbers and the distance to the nearest ill-posed prob-
lem, Numer. Math., 51 (1987), 251–289.

[2] I. Gohberg, P. Lancaster and L. Rodman, Matrix Polynomials, Academic Press, New
York, 1982.

[3] J.M. Gracia, Nearest matrix with two prescribed eigenvalues, Linear Algebra Appl.,
401 (2005), 277–294.

14
 [4] Kh.D. Ikramov and A.M. Nazari, Computational aspects of the use of Malyshev’s
formula, Comput. Math. Math. Phys., 44(1) (2004), 1–5.

[5] Kh.D. Ikramov and A.M. Nazari, On a remarkable implication of the Malyshev for-
mula, Dokl. Akad. Nauk., 385 (2002), 599–600.

[6] T. Kaczorek, Polynomial and Rational Matrices: Applications in Dynamical Systems

Theory, Springer-Verlag, London, 2007.

[7] P. Lancaster, Lambda-Matrices and Vibrating Systems, Dover Publications, 2002.

[8] P. Lancaster and P. Psarrakos, Normal and seminormal eigenvalues of matrixp oly-
nomials. Integral Equations Operator Theory, 41 (2001), 331–342.

[9] R.A. Lippert, Fixing two eigenvalues by a minimal perturbation, Linear Algebra
Appl., 406 (2005), 177–200.

[10] A.N. Malyshev, A formula for the 2-norm distance from a matrix to the set of matrices
with a multiple eigenvalue, Numer. Math., 83 (1999), 443–454.

[11] A.S. Markus, Introduction to the Spectral Theory of Polynomial Operator Pencils,
Amer. Math. Society, Providence, RI, Translations of Mathematical Monographs,
Vol. 71, 1988.

[12] A.M. Nazari and D. Rajabi, Computational aspect to the nearest matrix with two
prescribed eigenvalues, Linear Algebra Appl., 432 (2010), 1–4.

[13] N. Papathanasiou and P. Psarrakos, Normal matrix polynomials with nonsingular

leading coefficients, Electron. J. Linear Algebra, 17 (2008), 458–472.

[14] N. Papathanasiou and P. Psarrakos, The distance from a matrix polynomial to matrix
polynomials with a prescribed multiple eigenvalue, Linear Algebra Appl., 429 (2008),
1453–1477.

[15] A. Ruhe, Properties of a matrix with a very ill-conditioned eigenproblem, Numer.

Math., 15 (1970), 57–60.

[16] F. Tisseur and K. Meerbergen, The quadratic eigenvalue problem, SIAM Rev., 43
(2001), 235–286.

[17] J.H. Wilkinson, The Algebraic Eigenvalue Problem, Claredon Press, Oxford, 1965.

[18] J.H. Wilkinson, Note on matrices with a very ill-conditioned eigenproblem, Numer.
Math., 19 (1972), 175–178.

15
[19] J.H. Wilkinson, On neighbouring matrices with quadratic elementary divisors, Nu-
mer. Math., 44 (1984), 1–21.

[20] J.H. Wilkinson, Sensitivity of eigenvalues, Util. Math., 25 (1984), 5–76.

[21] J.H. Wilkinson, Sensitivity of eigenvalues II, Util. Math., 30 (1986), 243–286.

16