arXiv:2104.07893v1 [math.FA] 16 Apr 2021
On Kippenhahn curves and higher-rank
numerical ranges of some matrices ⋆
Natália Bebianoa , Joáo da Providénciab , Ilya M. Spitkovskyc
a
Departamento 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
The higher rank numerical ranges of generic matrices are described in terms
of the components of their Kippenhahn curves. Cases of tridiagonal (in
particular, reciprocal) 2-periodic matrices are treated in more detail.
1. Introduction
Let Mn stand for the algebra of all n-by-n matrices 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 [7, Chapter 6] for the basic properties
of W (A), in particular its convexity and invariance under unitary similarities.
⋆
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. The third author [IMS] was supported in part by Faculty Research
funding from the Division of Science and Mathematics, New York University Abu Dhabi.
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)
Preprint submitted to arXiv
April 19, 2021
In [6], this notion was generalized as follows: the rank-k numerical range
of A is
Λk (A) = {λ ∈ C : P AP = λP for some rank-k orthogonal projection P }.
(1.2)
Of course,
W (A) = Λ1 (A) ⊇ Λ2 (A) ⊇ · · · ⊇ Λn (A).
(1.3)
For k > n/2 the set Λk (A) is empty or a singleton {λ0 }; in the latter case
λ0 is an eigenvalue of A having geometric multiplicity at least 2k − n [6,
Proposition 2.2]. In particular, Λn (A) 6= ∅ if and only if A is a scalar multiple
of the identity, and then all the sets in (1.3) coincide.
So, for k = 1 and k > n/2 the sets Λk (A) are convex. Their convexity for
intermediate values of k was established in [16]. Shortly thereafter, in [15] it
was shown that, moreover,
\
Λk (A) =
{µ ∈ C : Re(eiθ µ) ≤ λk (θ)},
(1.4)
θ∈[0,2π)
where λk (θ) stands for the k-th largest (counting the multiplicities) eigenvalue
of the matrix Re(eiθ A). As usual, for any X ∈ Mn
Re X =
X + X∗
,
2
Im X =
X − X∗
.
2i
When applied to normal matrices, (1.4) yields
Λk (N) = ∩ conv{λj1 , . . . , λjn−k+1 },
(1.5)
with the intersection taken over all (n−k +1)-tuples from the spectrum σ(N)
of a normal matrix N. This result is also from [15], confirming a conjecture
from [5].
Our next observation is that the boundary lines
ℓθ,k = {µ ∈ C : Re(eiθ µ) = λk (θ)}
(1.6)
of the half-planes in the right hand side of (1.4), when taken for all k =
1, . . . , n, form a family the envelope of which is the so called Kippenhahn
curve C(A) of the matrix A. It was shown in [13] (see also the English
translation [14]) that W (A) = conv C(A). From the discussion above it is
2
clear that, at least in principle, not only W (A) but all the rank-k numerical
ranges of A can be described in terms of C(A).
Section 2 is devoted to generic matrices, for which C(A) splits into ⌈n/2⌉
components, each solely responsible for the respective higher rank numerical
range. These results are specified further in Section 3 for the case of tridiagonal 2-periodic matrices, when explicit formulas for λk (θ) are known. Finally,
a particular case of reciprocal 2-periodic matrices is treated in Section 4.
2. Generic matrices
For n = 2, there are only two sets in the chain (1.3), both easily identifiable. If n = 3, the middle term is either a singleton or the empty set
(since 2 > 3/2). The next proposition allows to distinguish between the two
possibilities.
Proposition 1. Let A ∈ M3 . Then Λ2 (A) 6= ∅ if and only if W (A) is an
elliptical disk, possibly degenerating into a line segment.
Proof. Directly from the definition it follows that
Λ2 (A) is a singleton {λ}
λ 0 x
if and only if A is unitarily similar to 0 λ y . Applying another unitary
u v z
similarity if needed, we may without loss of generality
suppose that u = 0.
λ y
, and W (A) = W (B)
Case 1. x = 0. Then A = (λ)⊕B, where B =
v z
is either an elliptical disk or a line segment, depending on whether or not B
is normal.
Case 2.
x 6= 0. Then A is unitarily similar to the tridiagonal matrix
λ x 0
0 z v with (1, 2)- and (2, 1)-entries having distinct absolute values. Ac0 y λ
cording to [3, Lemma 8], A is unitarily irreducible. On the other hand, its
(1, 1)- and (3, 3)-entries coincide, which implies the ellipticity of W (A) [4,
Theorem 4.2].
Recall that a matrix A ∈ Mn is generic if λ1 (θ), . . . , λn (θ) are distinct for
all θ.
Normal matrices are not generic; for n = 2 the converse is also true.
Hence, there is a direct relation with the shape of the numerical range: A ∈
3
M2 is generic if and only if W (A) is a non-degenerate elliptical disc. Already
for n = 3, things get more subtle.
Proposition 2. Let A ∈ M3 . Then A is generic if and only if W (A):
(i) has an ovular shape, or
(ii) is an ellipse with no eigenvalues of A lying on its boundary.
Note that A is unitarily irreducible in case (i) while it may or may not
be unitarily reducible (though not normal) in case (ii).
Proof. If A is unitarily irreducible, according to [12, Proposition 3.2] it is
generic if and only if W (A) has no flat portions on the boundary. These
are exactly ovular and elliptical shapes, as per Kippenhahn’s classification.
Moreover, unitary irreducibility of A implies that its eigenvalues are not on
the boundary.
Normal matrices are not generic, as was mentioned earlier. In the remaining case, W (A) is the convex hull of an ellipse E and a normal eigenvalue λ
of A. The matrix is generic if λ lies in the interior of E, which falls under
(ii), and non-generic otherwise.
Comparing Propositions 1 and 2, we see that for A ∈ M3 non-empty and
empty Λ2 (A) materialize both for generic and non-generic matrices.
Example 1. Let
0 −1/2
0
0 1/2 0
, M2 = 1/2 0 2 .
0
−1/2
M1 = 2
√
0
1 0
0 1/2
2
Figure 1 refers to the matrix M1 and Figure 2 refers to the matrix M2 .
Observe that W (M1 ) is ovular, Λ2 (M1 ) = ∅, while W (M2 ) is elliptical and
Λ2 (M2 ) = {0} is the eigenvalue of M2 different from the foci ±3/2 of W (M2 ).
Returning to generic matrices of arbitrary dimension n, note that from
their definition it immediately follows that
λk (θ) = −λn−k+1 (θ + π),
k = 1, . . . , n.
(2.1)
Since λn−k+1(θ) > λk (θ) for k > ⌈n/2⌉, the half-planes corresponding to θ
and θ + π in (1.4) are disjoint. Therefore, the rank-k numerical ranges of
generic matrices A are empty for k > ⌈n/2⌉. On the other hand, directly
4
0.6
0.4
0.2
-0.6
-0.4
0.2
-0.2
0.4
0.6
-0.2
-0.4
-0.6
Figure 1: Kippenhahn curve of M1
0.4
0.2
-1.5
-1.0
0.5
-0.5
1.0
1.5
-0.2
-0.4
Figure 2: Kippenhahn curve of M2
from (1.4) we see that for generic matrices A the inclusions in (1.3) are proper
for k = 1, . . . , ⌈n/2⌉; moreover, Λk+1 (A) lies in the interior of Λk (A).
The structure of C(A) and the related description of Λk (A) for k ≤ ⌈n/2⌉
are as follows.
Theorem 3. For a generic matrix A ∈ Mn its Kippenhahn curve C(A)
consists of the closed components
γk (A) = {hAzk (θ), zk (θ)i : θ ∈ [0, 2π]},
k = 1, . . . , ⌈n/2⌉ ,
where zk (θ) is the unit eigenvector associated with the eigenvalue λk (θ) of
Re(eiθ A). Respectively, the half-planes in the representation (1.4) of Λk (A)
are bounded by the family (1.6) of the tangent lines of γk (A).
The first statement is a rewording (in different terms) of [11, Theorem
5
13], based in particular on (2.1); the second immediately follows from the
first.
For n odd and k = ⌈n/2⌉ from (1.4), (2.1) it can be seen that in fact Λk (A)
is the intersection of the tangent lines ℓθ,k to γk (A) defined by (1.6). This
yields the following test for distinguishing between Λ⌈n/2⌉ being a singleton
or the empty set.
Corollary 1. Let A ∈ Mn be generic. If n is odd, then Λ⌈n/2⌉ (A) = γ⌈n/2⌉ (A)
if γ⌈n/2⌉ (A) is a point, and Λ⌈n/2⌉ (A) = ∅ otherwise.
Both cases are illustrated by Example 1.
Corollary 1 implies that for odd n the curve γ⌈n/2⌉ (A) cannot be convex
unless it collapses to a single point. On the other hand, the outermost curve
γ1 (A) of C(A) for a generic matrix A is always convex, and thus coincides
with the boundary ∂W (A) of its numerical range. This means in particular
that ∂W (A) does not have corners or flat portions. Other components of
C(A) may exhibit cusps and swallowtails but no inflection points.
As can be seen from Fig. 1, cusps (but not swallowtails) materialize already when n = 3. The emergence of swallowtails will be demonstrated in
Section 4, see Fig. 5–9.
Convexity of γ1 (A) implies that the subsequent components lie strictly
inside of it. This, however, does not preclude γj (A) with j > 1 from intersecting, as soon as there are at least two of them (i.e., when n ≥ 5 – see
Fig. 3 in Section 3 for an example corresponding to n = 5). Note that this
is happening in spite of strict inclusions in (1.3).
3. Tridiagonal 2-periodic matrices
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 ,
(3.1)
implying in particular that ∆n is invariant under transpositions ai+1,i ↔ ai,i+1
of its off-diagonal pairs.
Suppose now that these pairs are unbalanced, i.e.,
|ai+1,i | 6= |ai,i+1 | for i = 1, . . . , n − 1.
6
(3.2)
Then hermitian matrices Re(eiθ A) will be proper tridiagonal, i.e., their entries
directly above and below the main diagonal will be non-zero. According to
[3, Corollary 7], the eigenvalues of Re(eiθ A) are simple for all θ, thus implying
the genericity of A.
Example 2. Let
1
1
0
0
0
1/4 2 1/2 0
0
M3 = 0 1/4 0 3/4 0
.
0
0 1/4 −2 1
0
0
0 1/4 −1
This matrix is generic, since (3.2) holds. According to Corollary 1, Λ3 = ∅.
0.4
0.2
-2
-1
-0.2
-0.4
1
2
Figure 3: Kippenhahn curve of M3 . Notice that γ2 intersects γ3 .
We will say that a tridiagonal matrix A is 2-periodic if so are the sequence
of its diagonal entries and of its (non-ordered) off-diagonal pairs. For such
matrices we will use the notation a1 , a2 for the first two diagonal entries,
and {b1 , c1 }, {b2 , c2 } for the first two (once again, non-ordered) pairs of the
off-diagonal entries.
Along with A, for any θ the hermitian matrix Re(eiθ A) will be 2-periodic
as well, with αj (θ) =: Re(eiθ aj ) (j = 1, 2) as the period of its main diagonals. Transposing their off-diagonal pairs as needed, we may arrange for the
superdiagonal to also be 2-periodic, with
βj (θ) =: (eiθ bj + e−iθ cj )/2,
j = 1, 2
(3.3)
as the first two entries. According to (3.1), this rearrangement preserves the
characteristic polynomial of Re(eiθ A). Therefore, explicit formulas from [9]
can be used to compute λk (θ) in our setting. The respective straightforward
computation shows that
s
2
α1 + α2
α1 − α2
λk,n−k+1 =
±
+ |β1 |2 + |β2 |2 + 2 |β1 β2 | Qk (3.4)
2
2
7
for k = 1, . . . , m := ⌊n/2⌋, while λm+1 = α1 if n is odd.
kπ
Here Qk = cos m+1
if n is odd, and the k-th (in the decreasing order) root
of the m-th degree polynomial qm defined recursively via
q0 = 1, q1 (µ) = µ + |β2 /β1 | , qk+1 (µ) = µqk (µ) − qk−1 (µ) for k ≥ 1 (3.5)
if n is even.
For odd n, directly from the formula for λm+1 we obtain
Proposition 4. Let A ∈ Mn be tridiagonal and 2-periodic. If n is odd, then
γ⌈n/2⌉ (A) = {a1 }, the (1,1)-entry of A.
According to Corollary 1, for such matrices Λ⌈n/2⌉ (A) = {a1 }. Also, by
Proposition 4 a 2-periodic tridiagonal matrix A ∈ M5 cannot have intersecting γ2 and γ3 . For n = 6, however, this becomes a possibility; see Fig. 8 in
Section 4.
The parameters Qk are explicit and constant when n is odd, and implicit
(and in general depending on θ) if n is even. This makes consideration of
even-sized matrices much harder. However, in the case
b1 c2 = c1 b2
(3.6)
treated in [1], the ratio |β2 /β1 | is the same as |b2 /b1 | and thus θ-independent.
According to (3.5), Qk then do not depend on θ for even n as well. Formulas
(3.4), with some addtional nontrivial computations, provide an alternative
approach to the complete description of rank-k numerical ranges of 2-periodic
tridiagonal matrices satisfying (3.6). In agreement with [1], they all happen
to be elliptical disks.
Condition (3.6) holds in particular for tridiagonal Toeplitz matrices. If in
addition either the super- or the subdiagonal vanishes, then the dependence
on θ disappears in (3.4) alltogether. In other words, γk are then concentric
circles, and Λk (A) the respective circular disks. This covers the result on
shift operators from [8].
Example 3. To illustrate other possible shapes of Kippenhahn curves for 2periodic tridiagonal matrices, let M4 ∈ M7 have the zero main diagonal and
b1 = 3, b2 = 6, c1 = c2 = 2.
See the next section for more specific examples.
8
2
1
-6
-4
2
-2
4
6
-1
-2
Figure 4: Kippenhahn curve of M4
4. Reciprocal matrices
Recall the notion of reciprocal matrices introduced in [2]. These are
tridiagonal matrices with constant (without loss of generality, zero) main
diagonal and the off diagonal pairs satisfying ai+1,i ai,i+1 = 1. Reciprocal
matrices are of course proper tridiagonal. Denoting |aj+1,j |2 +|aj,j+1|2 =: 2Aj
we see that Aj ≥ 1. Condition (3.2) for such matrices takes the form Aj > 1,
j = 1, . . . , n − 1.
A 2-periodic reciprocal matrix A is completely characterized by its size
n and the values a1 := |a12 | , a2 := |a23 | (alternatively, by A1 and A2 ). For
n ≥ 4 (the only interesting setting), Im A has multiple eigenvalues if A1 or
A2 is equal to one, and so conditions A1 , A2 > 1 are not only sufficient but
also necessary for A to be generic.
p
Moreover, for reciprocal matrices (3.3) yields |βj | = (Aj + τ )/2, where
τ = cos(2θ). So, according to (3.4) λk,n−k+1 in this case are the square roots
of
p
1
ζk = (A1 + A2 + 2τ ) + (A1 + τ )(A2 + τ )Qk , j = 1, . . . , m. (4.1)
2
Observe that the right hand side of (4.1) is invariant under the substitutions θ 7→ −θ and θ 7→ θ + π. Thus, we arrive at the following
Corollary 2. Let A ∈ Mn be a 2-periodic reciprocal matrix. Then each component γ1 , . . . γm of its Kippenhahn curve C(A), and consequently its rank-k
numerical ranges Λk (A) for k = 1, . . . , m, are symmetric with respect to both
horizontal and vertical coordinate axes. Also, γm+1 = Λm+1 = {0} if n is
odd.
9
Furthermore, γk is an ellipse if and only if ζk = xτ + y with some constant
y > x > 0. If A1 = A2 := A, this happens to be the case for all k, since then
ζk = (A + τ )(1 + Qk ),
with Qk constant (note that (3.6) holds in a trivial way). So, the rank-k
numerical ranges of such matrices are elliptical disks with the boundaries
{γk }m
k=1 forming a family of nested ellipses whose axes are coincident with
the coordinate axes.
On the contrary, when A1 6= A2 we have
Theorem 5. Let A be a 2-periodic reciprocal matrix of odd size n and A1 6=
A2 . Then none of its rank-k numerical ranges has an elliptical shape if n = 1
mod 4. Otherwise, exactly one of them, namely Λ(n+1)/4 (A), is an elliptical
disk.
Proof. The first summand in the right hand side of (4.1) is of desired form.
kπ
The second term, however, is such only if Qk = 0. Since Qk = cos m+1
for
odd n, the result follows.
Observe that for generic 4-by-4 matrices γ1 and γ2 (consequently, Λ1 and
Λ2 ) are elliptical only simultaneously. Recall also that the numerical range
of a reciprocal matrix A ∈ M4 is elliptical if and only if
A2 = φA1 − φ−1 A3 or A2 = φA3 − φ−1 A1 ,
(4.2)
where φ is the golden ratio, and at least one of the inequalities Aj ≥ 1 is
strict [2, Theorem 7]. If A in addition is 2-periodic, i.e. A1 = A3 , then (4.2)
implies A2 = A1 . In other words, neither of rank-k numerical ranges of such
A is elliptical, unless A1 = A2 .
We suspect that this is the case for generic 2-periodic reciprocal matrices
A ∈ Mn for all even n > 2, not just n = 4. Formulas (4.1) should be
instrumental in proving this conjecture; the difficulty lies in the implicit
nature of Qk for even values of n.
Kippenhahn curves of several reciprocal matrices are pictured below. The
matrices are described by the triples {n, |a1 | , |a2 |}, or {n, A1 , A2 }. In Fig. 7,
8 and 10, the dotted curves are the best fitting ellipses to the components of
C(A) which look elliptical but in fact are not.
10
0.5
-1.5
-1.0
0.5
-0.5
1.0
1.5
-0.5
Figure 5: n = 4, a1 = 2, a2 = 21/20. The numerical range Λ1 is bounded by the exterior
component, while Λ2 is bounded by the interior component with its swallowtails removed;
Λ3 = ∅.
0.5
-2
1
-1
2
-0.5
Figure 6: n = 5, a1 = 2, a2 = 21/20. The picture is similar to Fig. 5, except that now
Λ3 = {0}.
11
1.0
0.5
-2
1
-1
2
-0.5
-1.0
Figure 7: n = 6, A1 = 1.25, A2 = 1.5. The components of C(A) are nested, with γ1 and γ2
being convex and so coinciding with the boundaries of Λ1 , Λ2 , respectively. On the other
hand, Λ3 is bounded by the “middle portion” of γ3 .
0.4
0.2
-1.5
-1.0
0.5
-0.5
1.0
1.5
-0.2
-0.4
Figure 8: n = 6, A1 = 1.05, A2 = 1.62. The component γ1 and γ2 are still convex. As
opposed to Fig. 7, γ3 is intersecting with γ2 .
0.4
0.2
-2
1
-1
2
-0.2
-0.4
Figure 9: n = 7, A1 = 1.05, A2 = 1.62. The picture is similar to Fig. 8, except that γ2 is
an exact ellipse, and there emerges γ4 = {0}.
12
0.6
0.4
0.2
-2
1
-1
2
-0.2
-0.4
-0.6
Figure 10: n = 7, A1 = 2, A2 = 1.5. The components γj are convex for j = 1, 2, 3 and
visually indistinguishable from ellipses, though only the middle one is a genuine ellipse.
References
[1] M. Adam, A. Aretaki, and I. M. Spitkovsky, Elliptical higher rank numerical range of some Toeplitz matrices, Linear Algebra Appl. 549
(2018), 256–275.
[2] N. Bebiano, J. Providéncia, I. M. Spitkovsky, and K. Vazquez,
Kippenhahn curves of some tridiagonal matrices, arXiv:2011.00849v1
[math.FA] (2020), 1–20 (to appear in Filomat).
[3] E. Brown and I. Spitkovsky, On flat portions on the boundary of the
numerical range, Linear Algebra Appl. 390 (2004), 75–109.
[4]
, On matrices with elliptical numerical ranges, Linear Multilinear
Algebra 52 (2004), 177–193.
[5] M.-D. Choi, J. A. Holbrook, D. W. Kribs, and K. Życzkowski, Higherrank numerical ranges of unitary and normal matrices, Operators and
Matrices 1 (2007), no. 3, 409–426.
[6] M.-D. Choi, D. W. Kribs, and K. Życzkowski, Higher-rank numerical
ranges and compression problems, Linear Algebra Appl. 418 (2006),
no. 2-3, 828–839.
[7] 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.
13
[8] H. Gaaya, On the higher rank numerical range of the shift operator, J.
Math. Sci. Adv. Appl. 13 (2012), no. 1, 1–19.
[9] M. J. C. Gover, The eigenproblem of a tridiagonal 2-Toeplitz matrix, vol.
197/198, 1994, Second Conference of the International Linear Algebra
Society (ILAS) (Lisbon, 1992), pp. 63–78.
[10] R. A. Horn and C. R. Johnson, Topics in matrix analysis, Cambridge
University Press, Cambridge, 1994, Corrected reprint of the 1991 original.
[11] E. A. Jonckheere, F. Ahmad, and E. Gutkin, Differential topology of
numerical range, Linear Algebra Appl. 279 (1998), no. 1-3, 227–254.
[12] D. Keeler, L. Rodman, and I. Spitkovsky, The numerical range of 3 × 3
matrices, Linear Algebra Appl. 252 (1997), 115–139.
[13] R. Kippenhahn, Über den Wertevorrat einer Matrix, Math. Nachr. 6
(1951), 193–228.
[14]
, 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.
[15] C.-K. Li and N.-S. Sze, Canonical forms, higher rank numerical ranges,
totally isotropic subspaces, and matrix equations, Proc. Amer. Math.
Soc. 136 (2008), no. 9, 3013–3023.
[16] H. J. Woerdeman, The higher rank numerical range is convex, Linear
Multilinear Algebra 56 (2008), no. 1-2, 65–67.
14