Académique Documents
Professionnel Documents
Culture Documents
www.elsevier.com/locate/laa
Abstract
In this article we determine the eigenvalues of sequences of tridiagonal matrices that
contain a Toeplitz matrix in the upper left block. 1999 Published by Elsevier Science
Inc. All rights reserved.
1. Introduction
Although Hermitian matrices are known to have real eigenvalues only, the
evaluation of these eigenvalues remains as misty as ever. For tridiagonal ma-
trices there are several known methods describing their eigenvalues such as
Gershgorin's theorem [5], Sturm sequences for Hermitian tridiagonal matrices
[1,4], etc. The eigenvalues of a tridiagonal Toeplitz matrix can be completely
determined [11]. Attempts have been made to resolve the eigenvalue problem for
matrices which are like tridiagonal Toeplitz matrices but not entirely Toeplitz
(see [2,3,12,13]). This paper falls in the same general direction of investigation.
*
Corresponding author. Partially supported by an Oakland University Research Fellowship 1995.
E-mail addresses: tsui@oakland.edu (S.-K. Tsui), kulkarni@oakland.edu (D. Kulkarni),
schmidt@oakland.edu (D. Schmidt)
0024-3795/99/$ - see front matter 1999 Published by Elsevier Science Inc. All rights reserved.
PII: S 0 0 2 4 - 3 7 9 5 ( 9 9 ) 0 0 1 1 4 - 7
64 D. Kulkarni et al. / Linear Algebra and its Applications 297 (1999) 6380
We study tridiagonal matrices which contain a Toeplitz matrix in the upper left
block. We call them pseudo-Toeplitz to make a distinction from all the matrices
studied before in [2,3,12], etc. The major feature of our treatment is the con-
nection between the characteristic polynomials of these tridiagonal pseudo-
Toeplitz matrices and the Chebyshev polynomials of the second kind, whereby
we can locate the eigenvalues that fall in the intervals determined by the roots of
some Chebyshev polynomials of the second kind. In other words, we use these
intervals derived from roots of some Chebyshev polynomial as a reference to
determine the eigenvalues of the original pseudo-Toplitz matrix. In fact, we are
able to determine the location of all eigenvalues of some tridiagonal pseudo-
Toeplitz matrices which have either all entries real with a nonnegative product
from each o-diagonal pair or the entries on the main diagonal purely imagi-
nary with a negative product from each o-diagonal pair (see Corollary 3.4). In
Section 2, we lay down a basic tool for nding the eigenvalues of tridiagonal
Toeplitz matrices, which is markedly dierent from the traditional approach
used in [2,11,12]. In Section 3, we give a detailed account of the number of ei-
genvalues in each such interval whose end points are consecutive roots of a pair
of Chebyshev polynomials related to the given tridiagonal pseudo-Toeplitz
matrix. We also show that, for a sequence of (real) tridiagonal matrices with a
positive product from each pair of o-diagonal entries, the eigenvalues of two
consecutive matrices in the sequence interlace (see Proposition 3.1). Further-
more, we discuss a lower bound for the number of real eigenvalues for tridi-
agonal pseudo-Toeplitz matrices of a xed dimension (see Theorem 3.6). In
Section 4, we demonstrate examples of tridiagonal pseudo-Toeplitz matrices for
which we can completely determine their real eigenvalues graphically.
These techniques have been also applied in innite dimensional program-
ming [10] and in numerical solutions of heat equations [3]. Standard references
for the Chebyshev polynomials are [6,8,9].
/n ak a k/n1 ak /n2 ak 1
/n ax 2x/n1 ax /n2 ax 2
sinn 1 cos1 x
Un x for jxj 6 1;
sincos1 x
3. Main results
where
2 3 2 3
a c 0 ak ck1 0
6 .. .. 7 6 .. .. 7
6b . . 7 6b . . 7
Tn a; b; c 6 .. .. 7; Bk 6 k1 .. .. 7:
4 . . c5 4 . . c1 5
0 b a 0 b1 a1
a
p k 2x: 4
bc
If k is a root of (3) but not a common root of Un1 and wk , then
Un k bk ck =bcwk1 k
:
Un1 k wk k
Un x 2xUn1 Un2 x 1
pn x 2x ; n P 1;
Un1 x Un1 pn1 x
and
0
pn1 x X
n1
2
pn0 x 2 2
2 2 2 2
; n P 1;
pn1 x p p
k1 n1 n2
. . . pnk
2 X n1
2
2 2
Un1k : 5
Un1 k1
Thus, pn0 x > 0 for all n P 1 and for all x in the domain of pn . Next we denote
for 1 6 j 6 k,
Fig. 1. y pn x.
It follows that
bj cj =bcwj1 k
gj k p
aj = bc k wj1 k bj1 cj1 =bcwj2 k
bj cj =bc
p
aj = bc k gj1 k
and
0
bj cj =bc1 gj1 k
gj0 k 2
; 26j6k
aj =bc k gj1 k
b1 c1 =bc
g10 k p : 7
a= bc k2
Proposition 3.1. The eigenvalues of Ak are distinct and interlace strictly with
eigenvalues of Ak1 for k P 2.
It follows from b1 c1 > 0 that u2 k has one root in maxa1 ; a2 ; 1 and one
root in 1; mina1 ; a2 . Thus, the assertion holds for k 2. Assume that the
assertion holds for order k 1. Let q1 < q2 < < qk1 be the eigenvalues of
Ak1 , and f1 ; < <; fk2 be the eigenvalues of Ak2 , where
q1 < f1 < q2 < fn2 < qn1 by hypothesis. Now uk2 k 1k2 kk2
Qk2
lower order terms so that uk2 k j1 fj k. It follows that
1k2j uk2 > 0 on fk2j ; fk1j ; 0 6 j 6 k 2, where f0 1; fk1 1.
Since qk1j 2 fk2j ; fk1j ; 1kj uk2 qk1j > 0, 0 6 j 6 k 2.
Expanding the determinant that generates uk k by the last row yields
uk k ak kuk1 k bk1 ck1 uk2 k; k P 2: 8
kj
Then it follows from (8) and bk1 ck1 > 0 that 1 uk qk1j < 0,
0 6 j 6 k 2. So uk has a zero in qj1 ; qj ; 2 6 j 6 k 1. It remains to show that
uk has a zero in each of 1; q1 and qk1 ; 1. Observe that 1k uk k kk
lower terms, so 1k uk qk1 < 0 < 1k uk k for k suciently larger than
qn1 . Thus, uk has a zero in qk1 ; 1. Also uk q1 < 0 < uk k for k suciently
smaller than q1 . Thus, uk has a zero in 1; q1 .
bk c k 0
/~kl 0 ~
n k /n kwkl k / kw~k1l k;
bc n1
70 D. Kulkarni et al. / Linear Algebra and its Applications 297 (1999) 6380
Proof. (i) Follows immediately from the discussion before the theorem. For (ii) it
suces to show that /~nkl has n k l real roots. We note that each common
pole of gk x and pn x gives rise to a root of /kn x in Eq. (3). We may now
assume that g~kl x and pn x have no common poles. Thus, it follows from part
(i) that the /~kl
n x must have n k l real roots. If bj cj 6 0 for all 1 6 j 6 k,
then it follows from Proposition 3.1 that /kn x has n k distinct real roots.
A similar analysis of the location of roots of /kn x can be done with regards
to intervals nj1 ; nj ; 1 6 j 6 n 1, which is in the following theorem.
Fig. 2. y pn x.
In case k 1, we have
/1n x a1 a 2x/0n x b1 c1 /0n1 x
and so
/kn x ak a/nk1 x bk ck /nk2 x 2x/nk1 x: 10
By the induction hypothesis, the rst two terms on the right side of (10) are
linear combinations of Unk1 x; . . . ; Unk1 x. The last term on the right side
of (10) is of the form
X
aj 2xUj x:
nk1 6 j 6 nk1
D. Kulkarni et al. / Linear Algebra and its Applications 297 (1999) 6380 73
Proof. Suppose that /kn x has fewer than n k real roots Q in 1; 1, say
m
f1 6 f2 6 6 fm , m < n k. Consider a polynomial f x j1 x fj of
degree m on 1; 1. f x can be Pmwritten as a linear combination of
U0 x; U1 x; . . . ; Um x, i.e., f x j0 aj Uj x. Next consider the weighted
inner product
Z 1
1=2
h/kn x; f xi 1 x2 /kn xf x dx;
1
which is nonzero since /kn x and f x are of either the same sign or the op-
posite sign over each of the following intervals 1; f1 ; f1 ; f2 ; . . . ; fm ; 1. On
the other hand, it follows from Proposition 3.5 that
X
/kn x bj Uj x
nk 6 j 6 nk
and hence
* +
X X
m
/kn ; f bj Uj x; aj Uj x 0; m < n k;
nk 6 j 6 nk j0
4. Examples
If b1 c1 < 0, then there may be zero, one or two additional roots of (12) in the
interval (gi1 ; gi ).
D. Kulkarni et al. / Linear Algebra and its Applications 297 (1999) 6380 75
Proof. Let d1 and d2 be the parts of the graph of g1 x for x < e1 =2 and for
x > e1 =2, respectively. We observe that if gi1 < e1 =2 < gi , from Fig. 1, we see
that d1 meets each component of the graph y Un x=Un1 x once in the i 2
intervals on the left of gi1 ; g1 , and d2 meets each component in n i 1
intervals once on the right of gi1 ; g1 , producing
p n 1 roots of (12). This
holds, if b1 c1 > 0, then y g1 x b1 c1 = bce1 2x is decreasing on each
interval 1; e1 =2 and e1 =2; 1 as depicted in Fig. 3; or if b1 c1 6 0. Now
if b1 c1 > 0, the component of the graph of y Un x=Un1 x in the distin-
guished interval, gi1 ; gi , meets both d1 and d2 , and we get two additional
roots of (12) (see Fig. 1 along with Fig. 3). If b1 c1 < 0, the graph of y g1 x is
increasing on 1; e1 =2 and on e1 =2; 1. With bc; a and a1 xed,
b1 and c1 can be chosen so that b1 c1 < 0 and each of the three illustrations in
Fig. 4 occurs.
Fig. 3. y g1 x.
76 D. Kulkarni et al. / Linear Algebra and its Applications 297 (1999) 6380
where
a1 a a2 a b1 c 1
p e1 ; p e2 ; d
bc bc bc
and Un x denotes the nth degree Chebyshev polynomial of the second kind.
We have seen in Corollary 3.4, that if b2 c2 > 0, b1 c1 > 0 and bc > 0,
Tn2 a; b; c has n 2 real distinct eigenvalues.
With the help of the graph of
b2 c2 e1 2x
y g2 x ;
bc4x2 e1 e2 2x e1 e2 d
and
2
D e1 e2 4d:
b2 c2 e1 2x
Fig. 5. g2 x bc4x2 e1 e2 2x e1 e2 - d
.
78 D. Kulkarni et al. / Linear Algebra and its Applications 297 (1999) 6380
Theorem 4.2. If D < 0, the Eq. (14) has n real roots, at least one lying in each
interval gj1 ; gj for j 1; 2; . . . ; n.
Proof. The result follows from looking at the graphs given in Fig. 5 comparing
them with the graph of y Un x=Un1 x in Fig. 1.
Fig. 6. y g2 x; D > 0.
D. Kulkarni et al. / Linear Algebra and its Applications 297 (1999) 6380 79
Proof. (i) Fig. 6 depicts the graph of y g2 x for four cases. From Figs. 1 and
6, it can be seen that g2 x and pn x have at least one intersection in any given
nondistinguished interval gj1 ; gj , for some 1 6 j 6 n, for x gj1 , and x gj
are vertical asymptotes of pn x, and the x-axis is a horizontal asymptote of
g2 x.
(ii) Suppose b2 c2 > 0 and b1 c1 > 0. Then g20 x < 0 for all x 6 h1 ; h2 . If
J1 6 J2 , then each Jj ; j contains two roots accounting for all
2 2 n 2 n 2 roots using (i). If r1 J2 , then J1 contains three roots
accounting for all 3 n 1 n 2 roots using (i). In either case, all roots of
(14) are accounted for, and thus all nondistinguished intervals contain exactly
one root of (14) again. Suppose b2 c2 > 0 and b1 c1 > 0 and b1 c1 < 0. The dis-
tinguished interval J1 or J2 that contains h2 contain two roots (Fig. 6). If
J1 6 J2 , this interval contains at least two roots of (14) accounting for
2 n 2 n roots, using (i). If J1 J2 , J1 contains at least one root of (14)
in J1 \ h1 ; 1 accounting for 1 n 1 n roots, by (i) again.
(iii) The number of real roots is at least n 2 by Theorem 3.6. In addition, if
b2 c2 < 0, b1 c1 < 0 and J1 6 J2 , then it can be seen from Fig. 6 that the dis-
tinguished interval that contains h1 necessarily contains two roots of (14) ac-
counting for 2 n 2 n roots. Finally, if J1 J2 , then one root of (14) is
guaranteed in J1 \ 1; h1 of (14), accounting for 1 n 1 n roots.
Theorem 4.4. If D 0, then there are at least n distinct real roots of 13.
Fig. 7. y g2 x; D 0.
80 D. Kulkarni et al. / Linear Algebra and its Applications 297 (1999) 6380
Remark 4.5. The positions of the real roots discussed in Theorems (4.2)(4.4)
can be determined completely by the graphs of pn x and g2 x.
References
[1] K.E. Atkinson, An Introduction to Numerical Analysis, Wiley, New York, 1978.
[2] R.M. Beam, R.F. Warming, The asymptotic spectra of banded Toeplitz and quasi-Toeplitz
matrices, SIAM J. Sci. Comput. 14 (4) (1993) 9711006.
[3] B. Cahlon, D.M. Kulkarni, P. Shi, Stepwise stability for the heat equation with a nonlocal
constraint, SIAM J. Numer. Anal. 32 (2) (1995) 571593.
[4] J.M. Franklin, Matrix Theory, Prentice-Hall, New Jersey, 1968.
[5] G.H. Golub, C.F. van Loan, Matrix Computations, Johns Hopkins University Press,
Baltimore, MD, 1989.
[6] U.W. Hochstrasser, Orthogonal polynomials, in: M. Abramowitz, I.A. Stegun, (Eds.),
Handbook of Mathematical Functions with Formulas, Graphs and Mathematical Tables,
1966, pp. 771802.
[7] R.A. Horn, C.A. Johnson, Matrix Analysis, Cambridge University Press, Cambridge, 1985.
[8] A. Korsak, C. Schubert, A determinant expression of Chebychev polynomials, Canad. Math.
Bull. 3 (1960) 243246.
[9] T.J. Rivlin, Chebychev Polynomials, Wiley, New York, 1974.
[10] I.E. Schochetman, R.L. Smith, S.-K. Tsui, in: A.V. Fiacco (Ed.), Solutions existence for
innite quadratic programming, the positive semi-denite case, to appear in Mathematical
Programming with Data Perturbations, Marcel Dekker, 1997.
[11] G.D. Smith, Numerical Solution of Partial Dierential Equations: Finite Dierence Methods,
3rd ed., Clarendon Press, Oxford, 1985.
[12] W.F. Trench, On the eigenvalue problem for Toeplitz band matrices, Linear Algebra Appl. 64
(1985) 199214.
[13] S.-K. Tsui, Matrices related to orthogonal polynomials, Preprint 1995.