arXiv:2011.00849v1 [math.FA] 2 Nov 2020
Kippenhahn curves of some tridiagonal matrices
⋆
Natália Bebianoa , Joáo da Providénciab , Ilya M. Spitkovskyc ,
Kenya Vazquezc
a
Bepartamento de Matemática, Universidade da Coimbra, Portugal
b
Departamento de Fı́sica, Universidade da Coimbra, Portugal
c
Division of Science and Mathematics, New York University Abu Dhabi (NYUAD),
Saadiyat Island, P.O. Box 129188 Abu Dhabi, United Arab Emirates
Abstract
Tridiagonal matrices with constant main diagonal and reciprocal pairs of offdiagonal entries are considered. Conditions for such matrices with sizes up
to 6-by-6 to have elliptical numerical ranges are obtained.
1. Introduction
Let Mn stand for the algebra of all n-by-n matrices A with the entries
aij ∈ C, i, j = 1, . . . n. We will identify A ∈ Mn with a linear operator acting
on Cn , the latter being equipped with the standard scalar product h·, ·i and
the associated norm kxk := hx, xi1/2 . The numerical range of A is defined as
W (A) = {hAx, xi : kxk = 1},
(1.1)
see e.g. [10, Chapter 1] or more recent [6, Chapter 6] for the basic properties
of W (A), in particular its convexity (the Toeplitz-Hausdorff theorem) and
invariance under unitary similarities.
The results are partially based on the Capstone project of [KV] under the supervision
of [IMS]. The latter was also supported in part by Faculty Research funding from the Division of Science and Mathematics, New York University Abu Dhabi. The work of the first
author [NB] was supported by the Centre for Mathematics of the University of Coimbra
— UIDB/00324/2020, funded by the Portuguese Government through FCT/MCTES.
Email addresses: bebiano@mat.uc.pt (Natália Bebiano), providencia@uc.pt (Joáo
da Providéncia), ims2@nyu.edu, ilya@math.wm.edu, imspitkovsky@gmail.com (Ilya
M. Spitkovsky), kvn222@nyu.edu (Kenya Vazquez)
⋆
Preprint submitted to arXiv
November 3, 2020
For our purposes it is important that W (A) is the convex hull of a certain
curve C(A) associated with the matrix A, described as follows. Using the
standard notation
Re A =
A + A∗
,
2
Im A =
A − A∗
,
2i
let λ1 (θ), . . . , λn (θ) be the spectrum of Re(eiθ A), counting the multiplicities.
Then the tangent lines of C(A) with the slope cot θ are {e−iθ (λk (θ) + it) : t ∈
R}, k = 1, . . . , n. We will thus call the characteristic polynomial of Re(eiθ A),
PA (λ, θ) = det Re(eiθ A) − λI ,
(1.2)
the NR generating polynomial of A. The connection between C(A) and W (A)
was established by Kippenhahn [11] (see also the English translation [12])
and following [6, Chapter 13] we will call C(A) the Kippenhahn curve of the
matrix A.
In algebraic-geometrical terms, C(A) is the dual of a curve of degree n,
which makes it a curve of class n. Class two curves are the same as curves
of degree two, thus implying that for A ∈ M2 the Kippenhahn curve is an
ellipse, in the case of normal A degenerating into the doubleton of its foci.
A matrix A ∈ Mn is tridiagonal if aij = 0 whenever |i − j| > 1. We will
be making use of the well known (and easy to prove) recursive relation for
the determinants ∆n of such matrices,
∆n = ann ∆n−1 − an−1,n an,n−1 ∆n−2 .
(1.3)
In this paper, we are interested mostly in tridiagonal matrices with an
additional property that their main diagonal is constant: ajj := a, j =
1, . . . , n. Let us standardize the notation as follows:
a b1
0
...
0
c1 a
b2
...
0
.. .
..
(1.4)
A(n; a, b, c) = ... . . . . . .
.
.
0 0 cn−2
a bn−1
0 . . . 0 cn−1
a
Since W (A − λI) = W (A) − λ for any A ∈ Mn and λ ∈ C, it suffices to concentrate our attention on matrices A(n; 0, b, c) with the zero main diagonal.
For n > 3, we will restrict our attention further to what we will call reciprocal
2
matrices. Namely, in addition to A being of the form A(n; 0, b, c) we will also
suppose that bj cj = 1 for all j = 1, . . . , n − 1, writing this symbolically as
A = A(n; 0, b, b−1 ).
(1.5)
Section 2 contains some properties of the NR generating polynomials and
curves of tridiagonal, in particular reciprocal, matrices. A necessary condition for ellipticity of the numerical range is also established in this section,
while its concrete implementations (along with the proof of sufficiency) for
4-by-4 and 5-by-5 reciprocal matrices are provided in Section 3. The last two
sections are devoted to 6-by-6 reciprocal matrices. In Section 4 criteria for
the Kippenhahn curve of such matrices to contain an elliptical component
are derived. By way of example it is also shown there that this component
may be the exterior one thus guaranteeing the ellipticity of W (A) without
all the components of C(A) being ellipses. Section 5 in turn is devoted to
the case when C(A) consists of three ellipses and contains an example of a
non-Toeplitz reciprocal matrix with this property.
2. Preliminary results
We start with some basic observations concerning NR generating polynomials of matrices (1.4).
Proposition 1. The NR generating polynomial of A(n; 0, b, c) is an even/odd
function of λ if n is even (resp., odd).
Proof. Matrices Re(eiθ A) − λI are tridiagonal along with A = A(n; 0, b, c).
More specifically,
Re(eiθ A(n; 0, b, c)) − λI = A n; −λ, (eiθ b + e−iθ c)/2, (e−iθ b + e−iθ c)/2
and so (1.3) implies
det Re(eiθ A(n; 0, b, c)) − λI = −λ det Re(eiθ A(n − 1; 0, b, c)) − λI
2
eiθ bn−1 + e−iθ cn−1
det Re(eiθ A(n − 2; 0, b, c)) − λI .
−
4
iθ
Since det Re(e A(1; 0, b, c) = −λ and
2
det Re(eiθ A(2; 0, b, c) = λ2 − eiθ b1 + e−iθ c1 /4,
the result follows by induction.
3
Corollary 1. The Kippenhahn curve of A = A(n; 0, b, c) is central symmetric, for n odd having the origin as one of its components.
In particular, the numerical range of A(n; 0, b, c) is central symmetric
— a fact observed for the first time (to the best of our knowledge) in [4,
Theorem 1]. If n = 3, Corollary 1 implies that C(A) consists of an ellipse
centered at the origin and the origin itself. Thus, W (A) is an elliptical disk
(degenerating into a line segment if A is normal) centered at the origin. This
fact also follows from [2, Theorem 4.2].
If A is a reciprocal matrix, then
eiθ bj + e−iθ cj
2
= |bj |2 + |bj |−2 + 2 cos(2θ) = 2(Aj + τ ),
j = 1, . . . , n − 1,
(2.1)
where for notational convenience we relabeled
|bj |2 + |cj |2 := 2Aj and cos(2θ) = τ.
(2.2)
Note that Aj ≥ 1, and the extremal case
A1 = . . . = An−1 = 1
(2.3)
is easy, due to the following:
Proposition 2. An n-by-n reciprocal matrix A is normal if and only if (2.3)
holds. If this is the case, then A is in fact hermitian, and W (A) is the real
π
.
line segment with the endpoints ±2 cos n+1
Proof. If a tridiagonal matrix (1.4) is normal, the equalities |bj | = |cj | can
be obtained via a direct verification, or by applying the normality criterion
from [2, Lemma 5.1]. For a reciprocal A from |bj | = |cj | it then follows that
bj = cj , thus making A hermitian. Moreover, such A is unitarily similar to a
tridiagonal Toeplitz matrix (1.4) with a = 0 and bj = cj = 1. The description
of W (A) then follows from the well known formula
λj = 2 cos
πj
,
n+1
j = 1, . . . , n
for the eigenvalues of A (see e.g. [1, Theorem 2.4]).
In this case C(A) = σ(A).
4
On the other hand, the case
A1 = . . . = An−1 := A0 > 1
(2.4)
is covered by [2]. According to [2, Theorem 3.3], W (A) is then an elliptical
disk. Following the proof of this theorem more closely, as in [3, Section 5],
reveals that in fact the “hidden” components of C(A) are also all elliptical.
⌈n/2⌉
Theorem 3. Let (2.4) hold. Then C(A) = ∪j=1 σj E, where E is the ellipse
p
with the foci ±1 and the axes having lengths 2(A0 ± 1), while
σj = cos
jπ
,
n+1
j = 1, . . . , ⌈n/2⌉.
(2.5)
Note that the number of non-degenerate ellipses constituting C(A) is
⌊n/2⌋. Indeed, if n is odd, then σ⌈n/2⌉ = 0, and the respective component
degenerates into the point {0}.
The only difference in the statement of Theorem 3 from the material
already contained in [3] is the explicit formula (2.5) for the s-numbers σj of
the ⌈n/2⌉ × ⌊n/2⌋ matrix X with the entries
(
1 if i − j = 0, 1
xij =
0 otherwise.
This covers in particular tridiagonal Toeplitz matrices, for which the ellipticity of their numerical ranges was established much earlier in [7].
In order to move beyond cases (2.3) and (2.4), we need to further analyze
properties of NR generating polynomials (1.2). Tracking the coefficients of
PA (λ, θ) in the proof of Proposition 1 when matrices under consideration are
reciprocal yields the following, more precise, statement. For convenience of
notation, in addition to (2.2) we also introduce
ζ = λ2 ,
k = ⌊n/2⌋.
Proposition 4. Let A ∈ Mn be a reciprocal matrix. Then its NR generating
polynomial has the form
Pn (ζ, τ ) = ζ k +
k−1
X
pj (τ )ζ j ,
(2.6)
j=0
premultiplied by −λ in case n is odd. Here pj are polynomials in τ of degree
k − j and coefficients depending only on A1 , . . . An−1 .
5
In particular, NR generating polynomials of reciprocal matrices are invariant under the change θ 7→ −θ. Combined with the central symmetry of
C(A), this observation implies
Corollary 2. Let A be a reciprocal matrix. Then its Kippenhahn curve C(A),
and thus also the numerical range W (A), is symmetric about both coordinate
axes.
We also need the following simple technical observation, which is a reformulation of a well known fact repeatedly used in the numerical range related
literature (see, e.g., [5] or [8]). It holds for all A ∈ Mn , not just tridiagonal
matrices.
Proposition 5. The Kippenhahn curve C(A) of A ∈ Mn contains an ellipse
centered at the origin if and only if the NR generating polynomial of A is
divisible by
λ2 − (x cos(2θ) + y sin(2θ) + z)
(2.7)
p
with x, y, z ∈ R such that z > x2 + y 2 .
Combining the results stated above, we arrive at a necessary condition
for a reciprocal matrix to have an elliptical numerical range.
Theorem 6. Let A be a reciprocal matrix with its numerical range being
an elliptical disk. Then polynomial (2.6) is divisible by ζ − (xτ + z), where
√
z ± x are the lengths of the half-axes of W (A). Equivalently,
Pn (xτ + z, τ ) = 0 for all τ.
(2.8)
Proof. According to Corollary 2, W (A) has to be axes-aligned and centered
at the origin. Since its boundary E is a component of C(A), Proposition 5
implies that Pn is divisible by (2.7) with some triple x, y, z. The geometry of
E means that y = 0 and the values of z ±x are as described in the statement.
Finally, (2.8) follows from the remainder theorem.
Treating Pn (xτ + z, τ ) as a polynomial in τ , (2.8) can be rewritten as the
system of k + 1 algebraic equations in x, z, A1 , . . . , An−1 . So, the (n − 1)tuples {A1 , . . . , An−1} for which this system is consistent form an algebraic
variety defined by k − 1 polynomial equations.
6
3. Reciprocal 4-by-4 and 5-by-5 matrices
In full agreement with Proposition 4, polynomials (2.6) for n = 4 and
n = 5 are as follows:
1
1
P4 (ζ, τ ) = ζ 2 − ζ (A1 + A2 + A3 + 3τ ) + (A1 + τ )(A3 + τ )
2
4
(3.1)
and
2
P5 (ζ, τ ) = ζ − ζ
+
1
(A1 + A2 + A3 + A4 ) + 2τ
2
1
(A1 + τ )(A3 + τ ) + (A1 + τ )(A4 + τ ) + (A2 + τ )(A4 + τ ) . (3.2)
4
Using the explicit formulas (3.1),(3.2), in the next two theorems we restate
the necessary condition of Theorem 6 in a constructive way and show that it
is also sufficient. The proofs of these results have a similar outline, varying
in computational details only.
Theorem 7. A reciprocal 4-by-4 matrix A has an elliptical numerical range
if and only if
√
A2 = φA1 − φ−1 A3 or A2 = φA3 − φ−1 A1 ,
(3.3)
is the golden ratio, and at least one of the inequalities Aj ≥ 1
where φ = 5+1
2
(j = 1, 2, 3) is strict. Moreover, in this case C(A) is the union of two nested
axis-aligned ellipses centered at the origin.
Proof. Necessity. Case (2.3) is excluded due to Proposition 2. When divided
by ζ − (xτ + z), polynomial (3.1) yields the quotient ζ − (x1 τ + z1 ), where
x1 =
3
− x,
2
1
z1 = (A1 + A2 + A3 ) − z,
2
while condition (2.8) takes the form of the system:
1
3
x x − 2 + 4 = 0
z z − 21 (A1 + A2 + A3 ) + 41 A1 A3 = 0
(x − 32 )z + x z − 21 (A1 + A2 + A3 ) + 41 (A1 + A3 ) = 0.
7
Solving the first two equations for x and z, respectively:
√
p
3± 5
1
x=
, z = (A1 + A2 + A3 ± (A1 + A2 + A3 )2 − 4A1 A3 ). (3.4)
4
4
Plugging these values of x and z into the last equation of the system reveals
that it is consistent if and only if the signs in (3.4) match, and
√
p
5
2
(A1 + A2 + A3 ) − 4A1 A3 =
(A1 + 3A2 + A3 ).
(3.5)
5
Finally, solving (3.5) for A2 yields (3.3).
Sufficiency. Given (3.3) and choosing upper signs in (3.4), we arrive
at the factorization of (3.1) as (ζ − (xτ + z)) (ζ − (x1 τ + z1 )). Condition
A1 + A2 + A3 > 3 guarantees that (0 <) x < z. From here and the relation
4(xτ − z)(x1 τ − z1 ) = (A1 + τ )(A3 + τ ) we conclude that C(A) consists of a
non-degenerate ellipse E and an ellipse E1 , degenerating into the doubleton
of its foci if A1 = 1 or A3 = 1. Furthermore, E1 lies completely inside E.
Indeed,
1p
1
z − z1 = 2z − (A1 + A2 + A3 ) =
(A1 + A2 + A3 )2 − 4A1 A3
2
2
√
√
is bigger than 25 due to (3.5) while x − x1 = 2x − 23 = 25 , implying that
xτ + z > x1 τ + z1 for all τ ∈ [−1, 1].
So, W (A) is the elliptical disk bounded by E, both E and E1 are centered
at the origin due to the central symmetry of C(A), and their major axes are
horizontal because x, x1 > 0.
Visualizing reciprocal 4-by-4 matrices as points {A1 , A2 , A3 } in R3+ we see
that those with elliptical numerical ranges form a 2-dimensional manifold M4
described by (3.3). These equations show that M4 contains the ray (2.4), as
it should according to Theorem 3. The same equations imply the following
Corollary 3. The intersection of M4 with any of the bisector planes Ai = Aj
(i 6= j) consists only of the ray (2.4).
It is worth mentioning that the ray A2 = 1, A1 = A3 > 1 corresponds to
matrices with so called bi-elliptical numerical ranges, i.e. convex hulls of two
non-concentric ellipses (see [9, Theorem 7]).
8
Theorem 8. A reciprocal 5-by-5 matrix A has an elliptical numerical range
if and only if
A1 = A4 or A1 − A4 = 2(A3 − A2 )
(3.6)
and at least one of the inequalities Aj ≥ 1 (j = 1, . . . , 4) is strict. Moreover, in this case C(A) is the union of the origin and two nested axis-aligned
ellipses centered there.
So, in contrast with the case n = 4, a reciprocal matrix A ∈ M5 can
have an elliptical numerical range while some but not all of its parameters
Aj coincide.
Proof. The factor −λ of the NR generating polynomial of A corresponds to
{0} as a component of C(A). Having duly noted that, we proceed by using
(3.2) in place of (3.1) but otherwise basically along the same lines as in the
proof of Theorem 7, with the computations surprisingly being even simpler.
Necessity. Case (2.3) is excluded due to Proposition 2. When divided by
ζ − (xτ + z), polynomial (3.2) yields the quotient ζ − (x1 τ + z1 ), where
x1 = 2 − x,
1
z1 = (A1 + A2 + A3 + A4 ) − z,
2
while condition (2.8) takes form of the system
3
x(x − 2) + 4 = 0
z z − 12 (A1 + A2 + A3 + A4 ) + 41 (A1 A3 + A1 A4 + A2 A4 ) = 0
(x − 2)z + x z − 21 (A1 + A2 + A3 + A4 ) + 14 (A1 A3 + A1 A4 + A2 A4 ) = 0.
Solving the first two equations for x and z, respectively:
1
x=1± ,
2
1
z = (A1 + A2 + A3 + A4 ± D),
4
(3.7)
where
D=
p
(A1 + A2 + A3 + A4 )2 − 4(A1 A3 + A1 A4 + A2 A4 ).
Plugging these values of x and z into the last equation of the system reveals
that it is consistent if and only if the signs in (3.7) match, and D = A2 + A3 .
The latter condition is equivalent to (3.6).
9
Sufficiency. Given (3.6) and choosing upper signs in (3.7), we arrive at
the factorization of (3.2) as
P5 (ζ, τ ) = (ζ − (xτ + z)) (ζ − (x1 τ + z1 )) .
(3.8)
Conditions D = A2 + A3 ≥ 2 and A1 + A2 + A3 + A4 > 4 guarantee that
(0 <) x < z. From here and the relation
4(xτ − z)(x1 τ − z1 ) = (A1 + τ )(A3 + τ ) + (A1 + τ )(A4 + τ ) + (A2 + τ )(A3 + τ )
we conclude that C(A) consists of {0}, a non-degenerate ellipse E and a
(possibly degenerating into a doubleton) ellipse E1 . Furthermore, E1 lies
inside E, with a non-empty intersection occurring only if A2 = A3 = 1.
Indeed,
1
1
1
z − z1 = 2z − (A1 + A2 + A3 + A4 ) = D = (A2 + A3 ) ≥ 1
2
2
2
while x − x1 = 2x − 2 = 1.
For convenience of future use, let us provide the explicit form of factorization (3.8) when (3.6) holds:
(
3 +3τ
ζ − A12+τ
ζ − A1 +A2 +A
2
P5 (ζ, τ ) =
(3.9)
A3 +A4 −A2 +τ
4 +3τ
ζ − 2A3 +A
ζ
−
2
2
if A1 − A4 equals zero or 2(A3 − A2 ), respectively.
Example 1. For the 5 × 5 reciprocal matrix (1.5) with the superdiagonal
string b = (1.5, 2, 2.5, 1.5) we have
A1 = A4 = 1.3472, A2 = 2.125, A3 = 3.205,
and according to the upper line of (3.9)
6.67722 + 3τ
1.34722 + τ
ζ−
.
P5 (ζ, τ ) = ζ −
2
2
The Kippenhahn curve of this matrix, consisting of two ellipses and the origin, is shown in Figure 1.
10
1.0
0.5
-2
1
-1
2
-0.5
-1.0
Figure 1: b1 = 1.5, b2 = 2, b3 = 2.5, b4 = 1.5. The curves are exactly elliptical.
Now let A be a reciprocal matrix (1.5) with b = (1.5, 2, 2, 3). Then A1 −
A4 ≈ −3.2, A3 − A2 = 0, and so (3.6) does not hold. The Kippenhahn curve
of this matrix is shown in Figure 2. Its components are non-elliptical; the
best fitting ellipses are pictured as dotted curves.
1.5
1.0
0.5
-2
1
-1
2
-0.5
-1.0
-1.5
Figure 2: b1 = 1.5, b2 = 2, b3 = 2, b4 = 3. The curves look elliptical but they are not
exactly elliptical.
11
4. Reciprocal 6-by-6 matrices with C(A) containing an ellipse
When n increases, things get more complicated. It may become impossible to state the divisibility condition from Proposition 5 in exact arithmetic.
Besides, this condition does not any longer guarantee the ellipticity of W (A).
We will illustrate these phenomena for n = 6.
Formula (2.6) then takes the form:
1
P6 (ζ, τ ) = ζ 3 − ζ 2 (A1 + A2 + A3 + A4 + A5 + 5τ )
2
1
+ ζ (A1 (A3 + A4 + A5 ) + A2 (A4 + A5 ) + A3 A5 + (3(A1 + A5 )
4
1
+ 2 (A2 + A3 + A4 )) τ + 6τ 2 − (A1 + τ ) (A3 + τ ) (A5 + τ ) . (4.1)
8
Consequently,
1
P6 (xτ +z, τ ) = τ 3 8x3 − 20x2 + 12x − 1 +τ 2 Q1 (x, z)+τ Q2 (x, z)+Q3 (x, z).
8
Here
Q1 = q11 z + q10 , Q2 = q22 z 2 + q21 z + q20 , Q3 = z 3 + q32 z 2 + q31 z + q30 (4.2)
with qij given by
3
5
q11 =3x2 − 5x + , q22 = 3x − ,
2
2
1
q10 = − (A1 + A2 + A3 + A4 + A5 ) x2
2
1
1
+ (3A1 + 2A2 + 2A3 + 2A4 + 3A5 ) x − (A1 + A3 + A5 ) ,
4
8
1
q21 = − (A1 + A2 + A3 + A4 + A5 ) x + (3A1 + 2A2 + 2A3 + 2A4 + 3A5 ) ,
4
1
q20 = (A1 A3 + A5 A3 + A1 A4 + A2 A4 + A1 A5 + A2 A5 ) x
4
1
− (A1 A3 + A5 A3 + A1 A5 ) ,
8
1
q32 = − (A1 + A2 + A3 + A4 + A5 ) ,
2
1
q31 = (A1 A3 + A5 A3 + A1 A4 + A2 A4 + A1 A5 + A2 A5 ) ,
4
1
q30 = − A1 A3 A5 .
8
12
Condition (2.8) therefore is nothing but the requirement for all three Qj from
(4.2) to vanish at the same root of the cubic equation
8x3 − 20x2 + 12x − 1 = 0.
(4.3)
In turn, this happens if and only if the resultant R1 of Q1 and Q2 , as well
as the resultant R2 of Q1 and Q3 , are equal to zero. A somewhat lengthy
but straightforward computation, with the repeated use of (4.3), reveals that
R1 , R2 can be treated as quadratic polynomials in x. Namely,
Rj (x) = rj2 x2 + rj1x + rj0 ,
j = 1, 2,
(4.4)
where
r12 = − 8(A21 + A2 A3 + A3 A4 + A25 ) − 28(A1 A2 − A1 A4 − A2 A5 + A4 A5 )
+ 4(A1 A3 + A3 A5 + A23 ) − 12(A22 + A24 − A1 A5 ) + 32A2 A4 ,
r11 =12(A22 + A24 ) + 18(A1 A2 − A2 A4 + A3 A5 ) − 24(A1 A4 + A2 A5 )
+ 6(A1 A3 − A23 + A3 A5 ),
r10 =4(A21 − A22 − A24 + A25 + A2 A3 + A3 A4 ) − (A1 A2 + A2 A4 + A4 A5 )
− 7(A1 A3 + A3 A5 ) + 6(A1 A4 − A1 A5 + A2 A5 ) + 3A23
and
r22 = − 12(A31 + A35 ) − 48(A21 A2 + A4 A25 ) − 60(A1 A22 + A24 A5 ) − 16(A32 + A34 )
+ 44(A1 A23 − A1 A2 A3 + A1 A2 A5 + A23 A5 + A1 A4 A5 − A3 A4 A5 )
− 36(A22 A3 − A22 A4 − A2 A24 + A3 A24 ) − 24(A2 A23 + A23 A4 − A1 A24 − A22 A5 )
+ 8(A21 A4 + A2 A25 ) + 20(−A1 A2 A4 + A21 A5 + A2 A4 A5 + A1 A25 )
+ 68(A1 A3 A4 + A2 A3 A5 ) − 4A33 + 40A2 A3 A4 − 84(A1 A3 A5 ),
r21 =14(A31 A35 − A1 A24 − A22 A5 ) + 56(A1 A22 + A24 A5 ) + 16(A32 + A34 )
− 12(A21 A3 + A3 A25 ) + 30(A22 A3 + A3 A24 ) − 50(A1 A23 + A23 A5 )
+ 20(A2 A23 + A23 A4 ) − 22(A21 + A22 A4 + A2 A24 + A2 A25 )
− 28(A1 A2 A4 + A21 A5 + A2 A4 A5 + A1 A25 ) − 66(A1 A3 A4 + A2 A3 A5 )
− 44(A1 A2 A5 + A1 A4 A5 ) + 48(A21 A2 + A4 A25 ) + 46(A1 A2 A3 + A3 A4 A5 )
+ 6A33 + 158A1 A3 A5 − 52A2 A3 A4 ,
13
r20 = − 2(A31 + A32 + A33 + A34 + A35 + A2 A23 + A23 A4 )
− 4(A21 A2 + A1 A22 + A1 A24 + A22 A5 + A24 A5 + A4 A25 )
+ 6(A21 A3 + A1 A2 A4 − A22 A4 − A2 A24 + A1 A2 A5 + A1 A4 A5 + A2 A4 A5 + A3 A25 )
− 10(A1 A2 A3 + A3 A4 A5 − A21 A4 − A2 A25 ) − (A22 A3 + A3 A24 )
+ 12(A1 A23 + A23 A5 ) + 11(A1 A3 A4 + A2 A3 A5 ) + 8(A21 A5 + A1 A25 )
+ 19A2 A3 A4 − 65A1 A3 A5 .
We thus arrive at the following
Theorem 9. A reciprocal 6-by-6 matrix A has an elliptical component in its
Kippenhahn curve C(A) if and only if both polynomials (4.4) share a common
root with equation (4.3).
Note that the requirement on the elliptical component of C(A) to be
centered at the origin is not included in the statement of Theorem 9. It
holds automatically for the following reason: if C(A) contains an ellipse E
not centered at the origin, then due to the symmetry it also contains −E
which is different from E. But them the third component of C(A) also has
to be elliptical and, since there cannot be four, this one is then centered at
the origin.
Plugging in the approximate values
x1 ≈ 0.0990311, x2 ≈ 0.777479, x3 ≈ 1.62349
for the roots of (4.3) into (4.4) thus delivers the numerical tests for C(A)
to contain an elliptical component. Due to the structure of rij , these are
systems of two homogeneous polynomial equations (one quadratic and one
cubic) in five unknowns A1 , . . . , A5 . Fixing any three of Aj (or otherwise
reducing the number of free parameters to two) and solving for the other two
yields the specific examples of matrices satisfying Theorem 9.
Example 2. Under additional constraints A1 = A5 := uA3 , A2 = A4 := vA3
the polynomials (4.4) both have the root x = x3 if and only if the pairs (u, v)
are as follows: (1, 1), (8.84369, −2.49077), (−1.80414, 2.24796), and
(1.7724359313231006, 0.6562336702811362).
(4.5)
The first solution corresponds to the situation (2.4) in which all three
components of C(A) are elliptical; the next two are irrelevant because of nonpositivity. The remaining pair (4.5) delivers a non-trivial example in which
14
one component of C(A) is an ellipse while two others are not (the latter fact
follows from Theorem 11 below). The respective C(A) is shown in Figure 3.
Only the outer component is an exact ellipse, though deviations of two others
from being elliptical are almost negligent.
2
1
-3
-2
1
-1
2
3
-1
-2
Figure 3: C(A) for A1 = A5 = 3.28117, A2 = A4 = 6.5623367, A3 = 5.
5. Reciprocal 6-by-6 matrices with C(A) consisting of three ellipses
From Theorem 9 it follows that a reciprocal matrix A ∈ M6 has C(A)
consisting of three concentric ellipses (and thus an elliptical numerical range)
if and only if R1 (xj ) = R2 (xj ) = 0 for all three roots of (4.3). This means
exactly that rjk = 0 for all j = 1, 2; k = 0, 1, 2.
We have arrived at the system of six polynomial equations in five variables
A1 , . . . , A5 , which therefore seems overdetermined. As was already observed
in Section 4, the ellipticity of two components of C(A) automatically implies
that the third one is an ellipse, centered at the origin. So, the number of equations can be easily reduced to four. An alternative approach shows, however,
that in fact reciprocal 6-by-6 matrices with C(A) consisting of three ellipses
are characterized by just three equations, and so form a two-parameter set.
15
Q
To describe this approach, observe that (4.1) factors as 3j=1 (ζ − (xj τ + zj ))
if and only if
z1 + z2 + z3 = 21 (A1 + A2 + A3 + A4 + A5 )
z1 z2 z3 = 81 A1 A3 A5
z1 z2 + z1 z3 + z2 z3 = 14 (A1 A3 + A1 A4 + A1 A5 + A2 A4 + A2 A5 + A3 A5 )
5
x1 + x2 + x3 = 2
x1 x2 x3 = 81
x1 x2 + x1 x3 + x2 x3 = 23
z1 x2 x3 + z2 x1 x3 + z3 x1 x2 = 1 (A1 + A3 + A5 )
8
1
(A1 A5 + A1 A3 + A3 A5 )
z
z
x
+
z
z
x
+
z
z
x
=
1 2 3
1 3 2
2 3 1
8
z (x + x ) + z (x + x ) + z (x + x ) = 3 (A + A ) + 1 (A + A + A ).
1
5
2
3
4
1 2
3
2 1
3
3 1
2
4
2
The three equations of this system not containing variables zj mean exactly
that x1 , x2 , x3 are the roots of (4.3). The three equations linear in zj can be
rewritten as
1
1
1
1
z1
(A1 + A2 + A3 + A4 + A5 )
2
x2 + x3 x1 + x3 x1 + x2 z2 = 3 (A1 + A5 ) + 1 (A2 + A3 + A4 ) .
4
2
1
z3
x2 x3
x1 x3
x1 x2
(A
+
A3 + A5 )
1
8
Solving this system:
zj =
R(xj )
,
8(xj − xi )(xj − xk )
j = 1, 2, 3,
(5.1)
where {i, k} = {1, 2, 3} \ {j} and R(x) = r2 x2 + r1 x + r0 with
r2 = 4(A1 + A2 + A3 + A4 + A5 ),
r1 = 4(A2 + A3 + A4 ) − 6(A1 + A5 ),
r0 = A1 + A3 + A5 . (5.2)
Finally, the remaining three nonlinear equations in zj yield the following:
16r02 + 24 (r1 + r2 ) r0 + 12r12 + 15r22 + 28r1 r2
+ 112 (A1 A3 + A5 A3 + A1 A4 + A2 A4 + A1 A5 + A2 A5 ) = 0;
32r02 + 8 (3r1 + r2 ) r0 + 4r12 + 3r22 + 6r1 r2 + 112 (A1 A3 + A5 A3 + A1 A5 ) = 0;
1 3
64r0 + 16(10r1 + 13r2 )r02 + 8(12r12 + 27r2 r1 + 13r22 )r0 + 8r13 + r23
64
2
2
+ 12r1 r2 + 20r1 r2 + 49A1 A3 A5 = 0
16
or, plugging in the values of rj from (5.2):
− (A22 + A24 ) − 2(A21 + A25 ) + 2A3 (A1 − A2 − A4 + A5 )
+ 3(A1 A4 + A1 A5 + A2 A5 ) − 4(A1 A2 + A4 A5 ) + 5A2 A4 = 0,
and
(5.3)
A22 − A23 + A24 + 3(A1 A3 + A1 A5 + A3 A5 )
(5.4)
− 2(A21 + A25 + A2 A3 − A2 A4 + A3 A4 ) − (A1 + A5 )(A2 + A4 ) = 0,
− A31 + A32 + A33 + A34 − A35 + (A1 − A3 + A5 )(A22 + A24 )
− 2(A2 + A4 )(A21 + A23 + A25 ) + 2A2 A4 (A1 − A3 + A5 )
− 3(A2 + A4 ) (A3 (A1 + A5 ) − A2 A4 ) − 3(A1 + A5 )(A23 + A1 A5 )
− 4(A21 A3 + A3 A25 + A1 A2 A5 + A1 A4 A5 ) + 41A1 A3 A5 = 0,
(5.5)
respectively.
We have thus reached the conclusion.
Theorem 10. A reciprocal 6-by-6 matrix A has its Kippenhahn curve consisting of three concentric ellipses if and only if Aj defined by (2.2) satisfy
(5.3)–(5.5), and in addition A1 + · · · + A5 > 1.
Note that the difference of (5.3) and (5.4),
A23 − A3 (A1 + A5 ) − 2(A22 + A24 ) − 3(A1 A2 + A4 A5 )
+ 3A2 A4 + 4(A1 A4 + A2 A5 ) = 0,
(5.6)
is somewhat simpler than either of these conditions, and for computational
purposes it thus might be useful to replace (5.3) or (5.4) (but not both) with
(5.6).
The semialgebraic subset M6 of R5+ defined by the system (5.3)–(5.5)
is of a more complicated structure than the pairs of hyperplanes (3.3),(3.6)
corresponding to n = 4, 5 cases. Even though we do not have an explicit
description of M6 , a direct verification confirms that it contains the ray
A1 = . . . = A5 > 0 — as it should, in order to agree with Theorem 3. Our
next result, somewhat analogous to Corollary 3, shows that this ray is in fact
the intersection of M6 with either of the hyperplanes A2 = A4 and A1 = A5 .
Theorem 11. If (A1 , . . . , A5 ) ∈ M6 and A2 = A4 or A1 = A5 , then all Aj
coincide.
17
Proof. Equations (5.3)–(5.6) are homogeneous, so by scaling without loss of
generality we may set A3 = 1.
Case 1. A2 = A4 := s. Relabeling also for convenience of notation
A1 = x, A5 = y, we may rewrite (5.6) as
(s − 1)(x + y − (s + 1)) = 0,
and conclude from there that s = 1 or x + y = s + 1.
In turn, (5.4) in our abbreviated notation amounts to
4s2 − 4s − 1 + (3 − 2s)(x + y) − 2(x2 + y 2 ) + 3xy = 0.
(5.7)
If s = 1, (5.7) simplifies further to
2(x2 + y 2) − 3xy − (x + y) + 1 = 0,
which for x, y > 0 is possible only when x = y = 1.
On the other hand, plugging x + y = s + 1 into (5.7) yields xy = s, so
that
x = s, y = 1, or x = 1, y = s.
(5.8)
Since (5.5) in our setting is nothing but
8s3 −4s2 −4s+1−(x+y)3 −4s(x+y)2 +(4s2 −6s−3)(x+y)−4(x2 +y 2 )+41xy = 0,
in view of (5.8) it amounts to 7(s − 1)3 = 0, thus implying s = 1. According
to (5.8), then also x = y = 1.
Case 2. A1 = A5 := s. Letting now A2 = x, A4 = y, rewrite (5.4) as
(x + y − (s + 1))2 − 2(s − 1)2 = 0.
So,
x+y = s+1±
√
2(s − 1).
(5.9)
On the other hand, in our abbreviated notation, (5.5) is nothing but
(x + y)3 + (2s − 1)(x + y)2 − (8s2 + 6s + 2)(x + y) − 8s3 + 33s2 − 6s + 1 = 0
or, plugging in the expression for x + y from (5.9):
√
√
√
−1−10s+25s2 −8s3 −(s± 2(−1+s))−2s(s± 2(−1+s))−8s2 (s± 2(−1+s))
√
√
√
+ 2(s ± 2(−1 + s))2 + 2s(s ± 2(−1 + s))2 + (s ± 2(−1 + s))3 = 0.
18
The left hand side of the latter equation is simply (3 ±
therefore s = 1. Equation(5.6) therefore takes the form
√
2)(s − 1)3 , and
−1 − 2(x + y)2 + x + y + 7xy = 0.
Observing that x + y = 2 due to (5.9) and s = 1, we conclude that xy = 1,
and so in fact x = y = 1.
Non-trivial elements of M6 can be constructed as follows. Fix two of the
parameters A1 , A2 , A4 , A5 (making sure to avoid the A1 = A5 or A2 = A4
situations), and solve (5.3)–(5.4) for the other two, expressing the solutions
as function of A3 . Then find A3 as a solution to (5.5).
Example 3. Set A1 = 20, A5 = 40. Then
A2 = 64.9396, A3 = 36.9387548, A4 = 28.9008
deliver a solution to (5.3)–(5.5) The respective C(A) curve is plotted in Figure 4.
3
2
1
-4
2
-2
4
-1
-2
-3
Figure 4: C(A) for A1 = 10, A2 = 32.4698, A3 = 18.4693774, A4 = 14.4004, A5 = 20.
19
References
[1] A. Böttcher and S. M. Grudsky, Spectral properties of banded Toeplitz
matrices, SIAM, Philadelphia, 2005.
[2] E. Brown and I. Spitkovsky, On matrices with elliptical numerical
ranges, Linear Multilinear Algebra 52 (2004), 177–193.
[3] K. A. Camenga, L. Deaett, P. X. Rault, T. Sendova, I. M. Spitkovsky,
and R. B. J. Yates, Singularities of base polynomials and Gau-Wu numbers, Linear Algebra Appl. 581 (2019), 112–127.
[4] M. T. Chien, On the numerical range of tridiagonal operators, Linear
Algebra Appl. 246 (1996), 203–214.
[5] M. T. Chien and K.-C. Hung, Elliptical numerical ranges of bordered
matrices, Taiwanese J. Math. 16 (2012), no. 3, 1007–1016.
[6] U. Daepp, P. Gorkin, A. Shaffer, and K. Voss, Finding ellipses, Carus
Mathematical Monographs, vol. 34, MAA Press, Providence, RI, 2018,
What Blaschke products, Poncelet’s theorem, and the numerical range
know about each other.
[7] M. Eiermann, Fields of values and iterative methods, Linear Algebra
Appl. 201 (1993), 167–197.
[8] H.-L. Gau, Elliptic numerical ranges of 4 × 4 matrices, Taiwanese J.
Math. 10 (2006), no. 1, 117–128.
[9] T. Geryba and I. M. Spitkovsky, On some 4-by-4 matrices with bielliptical numerical ranges, arXiv:2009.00272v1 [math.FA] (2020), 1–19.
[10] R. A. Horn and C. R. Johnson, Topics in matrix analysis, Cambridge
University Press, Cambridge, 1994, Corrected reprint of the 1991 original.
[11] R. Kippenhahn, Über den Wertevorrat einer Matrix, Math. Nachr. 6
(1951), 193–228.
[12]
, On the numerical range of a matrix, Linear Multilinear Algebra
56 (2008), no. 1-2, 185–225, Translated from the German by Paul F.
Zachlin and Michiel E. Hochstenbach.
20