IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 52, NO.

11, NOVEMBER 2007 2151

An Eneström-Kakeya Theorem for Hermitian Polynomial
Matrices
Gunther Dirr and Harald K. Wimmer
Abstract—We extend the Eneström-Kakeya theorem and its refinement
by Hurwitz to polynomial matrices ( ) with positive semidefinite coeffi-
cients. We determine an annular region containing the zeros of det ( ).
A stability result for systems of linear difference equations is given as an
application.
Index Terms—Block companion matrix, Eneström-Kakeya theorem,
polynomial matrices, root location, system of difference equations, zeros
of polynomials.
apply our results to systems of linear difference equations of the form
Ai x(k + i) = 0.
We use the following notation. If M 2 n2n then  (M ) and (M )
are the spectrum and the spectral radius of M . For the polyomial matrix
H (z ), we define (H ) = f 2 ; det H () = 0g and (H ) =
maxfjj;  2  (H )g.
Let A; A~ 2 n2n be Hermitian. We write A  0; A~ > 0, if A is
positive semidefinite and A~ is positive definite. Accordingly, A  A~
means that A 0 A~ is positive semidefinite. If all the eigenvalues of M
are real then max (M ) and min (M ) shall denote the largest and the
smallest eigenvalue of M , respectively. Let = fz 2 ; jz j < 1g
be the open unit disc and @ = fz 2 ; jz j = 1g the unit circle
of the complex plane. The set m + shall consist of all real polynomials
p(z ) = i=0 ai z satisfying (1).
m i

According to [1] it is of interest to know when the inequalities (2)

I. INTRODUCTION in Theorem 1.2 are sharp, or equivalently when a polynomial h(z ) in
Theorem 1.1 has no roots with modulus 1. This was first studied by
The following theorem is known as the Eneström-Kakeya theorem
Hurwitz. The following result contains a modified version of Hurwitz’s
[10, p. 4], [3, p. 12], [11, p. 255].
Theorem 1.1: Let h(z ) = am z m + 1 1 1 + a1 z + a0 be a real poly-
original contribution [6].
Theorem 1.3: Let h(z ) = a0 + a1 z + 1 1 1 + am01 z m01 + am z m
nomial with
be a real polynomial with

0 < a0 = a1 = 1 1 1 = ar 01
am  1 1 1  a1  a0  0; am > 0:
< ar = ar +1 = 1 1 1 = ar 01 < 1 1 1
(1)

i) Then all the zeros of h(z ) are in the closed unit disk. < ar = ar +1 = 1 1 1 = am :
ii) The zeros of h(z ) lying on the unit circle are simple.
The Eneström-Kakeya theorem is a powerful tool to determine re- Setk = gcd(m + 1; r1 ; . . . ; rs ). We have (h) = 1 if and only if
gions in the complex plane containing zeros of polynomials or eigen- k > 1. In that case
< a0 = 1 1 1 = ak01  ak = 1 1 1 = a2k  1 1 1
values of matrices. It has been used to analyze overflow oscillations
0
of discrete-time dynamical systems [8], to investigate properties of or-
thogonal wavelets [7], to establish the stability of allpass fractional  am0k+1 = 1 1 1 = am (3)
Hilbert transformers [9], and to determine the asymptotic behavior of and
zeros of the Daubechies filter [12] or the location of zeros of the partial
sums of cos(z ) and sin(z ) [13]. For applications to a model of high
h(z ) = (1 + z + 1 1 1 + z k01 )p(z k );
energy collisions we refer to [4]. p 2 `+01 ; (p) \ @ = ; (4)
It is not difficult to show (see, e.g., [3, p. 12]) that Theorem 1.1(i) is
equivalent to the following. with ` = (m + 1)=k .
Theorem 1.2: Let h(z ) = a0 + a1 z + 1 1 1 + am z m be a real poly-
nomial with positive coefficients. Set II. HERMITIAN POLYNOMIAL MATRICES
In the following, let:

 = 0min
im01
(ai =ai+1 );  = max (ai =ai+1 ):
0im01
H (z ) = Am z m + Am01 z m01 + 1 1 1 + A0 (5)

be a polynomial matrix with Hermitian coefficients Ai 2 2 ;i =

n n
Then the roots  of h(z ) satisfy
0; 1; . . . ; m, such that

  jj  : (2)
Am  Am01  1 1 1  A0  0; Am > 0: (6)

In this note, we deal with polynomial matrices

We associate to u 2 n
; u 6= 0, the polynomial

H (z ) = A 0 + A 1 z + 1 1 1 + A m z m hu (z ) = u3 H (z )u = ai z i
with ai = u3 Ai u; i = 0; . . . ; m: (7)
where the coefficients Ai 2 n2n are positive semidefinite Hermitian
matrices. We generalize the preceding two theorems and extend work Because of (6), the coefficients of hu (z ) satisfy (1). If H ()u = 0; u 6=
of [6] and [1] on the sharpness of the Eneström-Kakeya theorem. We 0,then hu () = 0, and we can apply Theorem 1.1(i) to hu (z ) and
obtain jj  1. Thus, the following result is obvious.
Theorem 2.1: If the polynomial matrix H (z ) satisfies (6), then
Manuscript received March 28, 2007; revised June 6, 2007. Recommended (H )  1 .
For k  2 we define
by Associate Editor A. Hansson.
The authors are with the Institute of Mathematics, University of Würzburg,
Würzburg D-97074, Germany (e-mail: wimmer@mathematik.uni-wuerzburg.
de).
Digital Object Identifier 10.1109/TAC.2007.908340 dk (z ) = 1 + z + 1 1 1 + z k01 :
0018-9286/$25.00 © 2007 IEEE
2152 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 52, NO. 11, NOVEMBER 2007

Theorem 2.2: Let jj = 1. (i) Suppose Theorem 2.4: If  2  (H ) and jj = 1 then the corresponding
elementary divisors of H (z ) have degree 1. Moreover,  6= 1, and
 2 (H ) with H ()u = 0; u 6= 0 (8) k = 1 for some k 2 ; k  m + 1. If A0 > 0 then m+1 = 1.
Proof: If  2  (H ) has an elementary divisor of degree at least
and
2 then (see, e.g., [2, p. 342]) there exist nonzero vectors u; w such that
As u 6= 0 and Ai u = 0 if i < s: (9) H ()u = 0 and H 0 ()u + H ()w = 0. Suppose jj = 1. Then
u3 H () = 0, and therefore hu0 () = u3 H 0 ()u = 0. Then  2 @
Set d(z ) = z s dk (z ). Let U = (u; u2 ; . . . ; un ) be unitary. Then would be a multiple zero of hu (z ), which is a contradiction to Theorem
1.1(ii).
U 3 H (z )U It is easy to see that 1 2 = (H ). Suppose (8). Then (11) implies
d(z )p11 (z k ) d(z )p12 (z k ) 1 1 1 d(z)p1n (zk ) dk () = (k 0 1)( 0 1)01 = 0, and we obtain k = 1. If A0 > 0
d(z )p12 (z k ) 1 111 1 then s = 0 in (11). Therefore, m + 1 = k`, and m+1 = 1.
=
1 1 111 1 Furuta and Nakamura [5] extended Theorem 1.1 (i) to polynomials
Ai z i with positive definite operator coefficients Ai . The approach
d(z )p1n (z k ) 1 111 1 in [5] is different, its relies on a power inequality for the numerical
(10)
radius of an operator acting on a Hilbert space.
Note that Theorem 1.3 completely characterizes those polynomials
with suitable polynomials p1j (z k ). h(z ) 2 m + with a0 > 0, which satisfy (h) < 1. In the case of
(ii) We have H ()u = 0 if and only if u3 H () = 0. polynomial matrices H (z ), we can only give sufficient conditions for
Proof: Because of (9) we have hu (z ) = z s h ~ (0) 6= 0. It
~ (z ); h (H ) < 1 .
follows from Theorem 1.3. that there exists a k > 1 such that m + 1 = Theorem 2.5: Assume
s + k` and
0 < A0  A1  1 1 1  Ar 01 < Ar  Ar +1
u3 H (z )u = z s dk (z )p11 (z k )  1 1 1  Ar 01 < 1 1 1 < Ar  Ar +1  1 1 1  Am : (12)
with p11 2 `+01 ;  (p11 ) \ @ = ;: (11)
(i) If gcd(m + 1; r1 ; . . . ; rs ) = 1 then (H ) < 1.
In particular, the coefficients ai in (7) satisfy (ii) In particular, we have (H ) < 1 if 0 < A0 < A1 < 1 1 1 < Am .
Proof: Let hu (z ) be defined as in (7). Then (12) implies ar 01 <
0 = a0 = 1 1 1 = as01 ar ; j = 1; . . . ; s, and Theorem 1.3 yields (h) < 1. Hence, we have
 as = 1 1 1 = as+k01  1 1 1 u3 H ()u 6= 0 for all  2 @ , and therefore (H ) < 1.
We now turn to Theorem 1.2 and its generalization. Suppose A; A~ 2
 am0k+1 = 1 1 1 = am : n2n are positive semidefinite. Let A] denote the Moore-Penrose gen-
eralized inverse of A. It is obvious that there exists a number b 2
From u3 Am u = 1 1 1 = u3 Am0k+1 u and Am  1 1 1  Am0k+1 , we such that bA  A~, if and only if
obtain u3 Am = 1 1 1 = u3 Am0k+1 = (m1 ; . . . ; mn ). Similarly,
u3 Am0k = 1 1 1 = u3 Am02k+1 = (m0k;1 ; . . . ; m0k;n ); . . . ; Ker A  Ker A:
~
u3 As+k01 = 1 1 1 = u3 As = (s+k01;1 ; . . . ; s+k01;n ); u3 As01 =
(13)
1 1 1 = u3 A0 = (0; . . . ; 0). Hence It is not difficult to verify that under the hypothesis (13) we have

u3 H (z ) = (z m + 1 1 1 + z m0k+1 )(m1 ; . . . ; mn ) minfb 2 ; bA  A~g = max (AA

~ ] ):
+ (z m0k + 1 1 1 + z m02k+1 )(m0k;1 ; . . . ; m0k;n )
+ 1 1 1 + (z s+k01 + 1 1 1 + z s ) If A~ > 0 then (13) implies A > 0. In that case
2 (s+k01;1 ; . . . ; s+k01;n )
~ 01 ):
maxfa 2 ; aA  A~g = min (AA
= d(z )(z m0k+10s (m1 ; . . . ; mn )
+ z m02k+10s (m0k;1 ; . . . ; m0k;n ) + 1 1 1 Theorem 2.6: Let H (z ) = A0 +A1 z +1 1 1+Am z m be a polynomial
+ (s+k01;1 ; . . . ; s+k01;n )) matrix with Ai  0; i = 0; . . . ; m 0 1, and Am > 0. Suppose
= d(z ) (z k )`01 m1 + (z k )`02 m0k;1 + 1 1 1
Ker Am01  1 1 1  KerA0 : (14)
+s+k01;1 ; . . . ; (z k )`01 mn + 1 1 1 + s+k01;n
= d(z )(p11 (z k ); . . . ; p1n (z k )); Define

and we obtain (10) since all Ai are Hermitian.

 = maxfmax (Ai A]i+1 ); i = 0; . . . ; m 0 1g
(ii) Suppose H ()u = 0. Then hu () = 0, and, therefore, jj =
1 and (11) imply d() = 0. Hence, (10) shows that u3 H () = 0. and either  = 0 if det A0 = 0, or
Conversely, we can use the same line of argument if u3 H () = 0.
Corollary 2.3: Let H (z ) = (h (z )). If jj = 1 and  2  (hii )  = min min (Ai A0i+1
1 ); i = 0; . . . ; m 0 1 (15)
for some i, then  2  (H ).
Proof: Let ei be the ith unit vector of n . If  2 @ and if A0 > 0. Then
ei3 H ()ei = hii () = 0 then we have H ()ei = 0, and, therefore,
 2  (H ) .   j j   for all  2  (H ): (16)
IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 52, NO. 11, NOVEMBER 2007 2153

Proof: Suppose Am01 6= 0. Then  > 0. We have Ai+1   Ai , and  (AjW )  @ . Let the initial value v0 in (19) be decomposed
and, thus, Ai+1  i+1  Ai  i ; i = 0; . . . ; m 0 1. Moreover, Am  m > correspondingly as v0 = y0 + w0 ; y0 2 Y; w0 2 W . If (y (t)) and
0. Therefore we can apply Theorem 2.1 to the polynomial matrix (w(t)) are the solutions of

H (z ) = A0 + A1 z + 1 1 1 + Am  m z m
and we obtain jj   for all  2  (H ). If Am01 = 0 then (14) y(t + 1) = Ay(t); y(0) = y0
implies H (z ) = Am z m . Thus  (H ) = f0g and  = 0. and w(t + 1) = Aw(t); w(0) = w0
Suppose A0 > 0. Then (14) implies Ai > 0 for all i. Hence  in (15)
is well defined. The argument concerning H (z ) can be adapted for
z m H (=z ). It yields the lower bound in (16). In the case det A0 = 0
we have 0 2  (H ), and  = 0 is the trivial bound lower bound in
(16).
We remark that in the case of det A0 = 0 the lower bound  = 0
is rather conservative. A more detailed investigation of the role of the
characteristic root 0 2  (H ) would yield a sharper estimate for the
nonzero elements of  (H ).
respectively, then we have v (t) = y (t) + w(t). From (AjY ) < 1
follows limt!1 y (t) = 0. If  2  (H ) \ @ is a zero of multiplicity
r of det H (z ) then Theorem 2.4 implies AjKer (A0I ) = Ir and
k = 1 for some k  2, depending on . Let p bepthe least common
multiple of all those k . If dim W = q then AjW = Iq . Therefore,
w(t + p) = Ap w(k) = w(t) for all t  0. If A0 > 0 then pj(m + 1)
by Theorem 2.4.

III. A DIFFERENCE EQUATION REFERENCES

We now apply Theorem 2.1 and Theorem 2.4 to a linear system of [1] N. Anderson, E. B. Saff, and R. S. Varga, “On the Eneström-Kakeya
difference equations of the form theorem and its sharpness,” Linear Algebra Appl., vol. 28, pp. 5–16,
1979.
A0 x(t) + 1 1 1 + Am01 x(t + m 0 1) + Am x(t + m) = 0 [2] H. Baumgärtel, Analytic Perturbation Theory for Matrices and
x(0) = x0 ; . . . ; x(m 0 1) = xm01 : (17)
Operators, Operator Theory: Advances and Applications. Basel:
Birkhäuser-Verlag, 1985, vol. 15.
[3] P. Borwein and T. Erdèlyi, Polynomials and Polynomial Inequali-
Assuming ties. New York: Springer-Verlag, 1995.
0  A0  1 1 1  Am0  Am ; 0 < Am [4] M. Brambilla, A. Giovannini, and R. Ugoccioni, “Maps of zeros of the
1 (18) grand canonical partition function in a statistical model of high energy
collisions,” J. Phys. G: Nucl. Part. Phys., vol. 32, pp. 859–870, 2006.
we shall prove that in general each solution (x(t)) can be decomposed [5] T. Furuta and M. Nakamura, “An operator version of the Eneström-
into a norm-decreasing part (y (t)) and a periodic part (w(t)). Kakeya theorem,” Math. Japan, vol. 37, pp. 459–497, 1992.
Theorem 3.1: Suppose (18), and let (x(t)) be a solution of (17). [6] A. Hurwitz, “Über einen Satz des Herrn Kakeya,” Tôhoku Math. J., vol.
4, pp. 89–93, 1913.
Then [7] J. R. Karam, “An application of the Kakeya-Enestroem theorem, Proc.
x ( t) = y ( t) + w ( t) ; t  0 ;
Int. Conf. Image Sci.,” in Syst. Technol. CISST’04, 2004, pp. 354–358.
[8] T. Ooba, “On companion systems with state saturation nonlinearity,”
IEEE Trans. Autom. Control, vol. 50, pp. 1580–1584, 2003.
such that limt!1 y (t) = 0 and (w(t)) is periodic, i.e., there exists a [9] S.-C. Pei and P.-H. Wang, “Maximally flat allpass fractional Hilbert
p  2 such that w(t + p) = w(t) for all t  0. In particular, if A0 > 0 transformers,” in Proc. IEEE Int. Symp. Circuits Syst., 2002, vol. 5, pp.
then p = m + 1 is a period of (w(t)). V-701–V-704.
[10] V. V. Prasolov, Polynomials. New York: Springer, 2004.
Proof: Because of [11] Q. I. Rahman and G. Schmeisser, Analytic Theory of Polynomials.
A0m1=2 H (z )Am
01=2 = Iz m + 1 1 1 + A01=2 A0 A01=2 Oxford, U.K.: Oxford Univ. Press, 2002.
m m [12] J. Shen and G. Strang, “Asymptotic analysis of Daubechies polyno-
mials,” Proc. Amer. Math. Soc., vol. 124, pp. 3819–3833, 1996.
we can assume Am = I . Set [13] R. S. Varga and A. J. Carpenter, Zeros of the partial sums of cos(z) and
sin(z), I, Numerical Algorithms, vol. 25, pp. 363–375, 2000.
v(t) = (x(t)T ; x(t + 1)T ; . . . ; x(t + m 0 1)T )T
and define v0 conforming to (17). Let
0 I 0 ... 0
0 0 I ... 0
A= 1 1 1 ... 1
1 1 1 ... 1
0A 0A 0A
0 1 2 ... 0Am0 1

be the block companion matrix associated with H (z ) = Ai z i . Then

(17) is equivalent to

v(t + 1) = Av(t); v(0) = v0 : (19)

Thus,  (H ) =  (A)   . Suppose (H ) = (A) = 1. Let Y and W

be A-invariant subspaces such that n = Y 8 W and  (AjY )