Singular Values of Cauchy-Toeplitz Matrices

Evgenij E. Tyrtyshnikov
Department of Numerical Mathematics
USSR Academy of Sciences
Ryleez;a 29
Moscow 119034, USSR

Submitted by Richard A. Brualdi

ABSTRACT

The behavior of singular values of matrices A,, = [l/(i - j + g)]: j= 1 with n +m

is investigated. For any real g which is not integer it is proved that the singular
values are clustered at rTT/lsin agl, which is their upper boundary. The only o(n)
singular values are those which lie outside a given e-neighborhood of the clustering
point [o(n>/n + 0 as n 4001; o(n) = 00 nz n) holds if Igl Q i. Also proved is that the
minimum singular values of A,(g) tend to zero provided that Igl > f

1. INTRODUCTION

By a Cauchy-Toeplitz matrix we mean a matrix which is both Cauchy and

Toeplitz. Any such matrix differs from

A,,(g) = (1.1)

by at most a scalar multiplier. So we do not lose generality in investigating

only A,(g). We shall study the behavior of the singular values

of real A,,(g) as n tends to infinity; so g is supposed to be real and not

LINEAR ALGEBRA AND ITS APPLICATlONS 161:99-116 (1992) 99

0 Elsevier Science Publishing Co., Inc., 1992

655 Avenue of the Americas, New York, NY 10010 0024-3795/92/$3.50
100 EVGENIJ E. TYRTYSHNIKOV

integer. This paper is a natural extension of our previous work [lo], where
the case g = i has been investigated.
In the behavior of singular values of A,($), there are two remarkable
peculiarities concerning superior singular values and inferior ones. The
superior singular values constitute a cluster. More precisely, most of the
singular values are very close to n-, and none of them is greater than r. This
phenomenon was discovered by Moler in his experiments and then was
explained by Parter [7]. A qualitative explanation of Moler’s phenomenon is
provided by the two relationships that follow:

P~,(~)G~T, j=l ,...,n, (1.2)

lim Pl”( i) + . . * + P,,(i)

= 5r. (1.3)
n+m n

One way or another, while proving either of these relationships one has to
utilize Fourier-Laurent series that generate some Toeplitz matrices allied
with A,(i). In the paper of Parter [7], (1.3) is tackled by employing the
well-known Szego theorem on the distribution of eigenvalues of Hermitian
Toeplitz forms connected with Fourier-Laurent series [3]. In [lo] we have
shown how (1.3) can be deduced without applying the Szego theorem.
The main property of the inferior singular values of A,,(i) is that

lim p,,( +) = 0. (1.4)

n-+,x

The question whether (1.4) is true was raised by Parter [8] and then
answered by the author [lo]. It is curious that numerical tests do not
unambiguously suggest equating the limit to zero, for p,, tends to zero
rather slowly [e.g., ~~,,a(+) = 0.8, p,sa,,(~) = 0.71.
One can conjecture that the abovementioned behavior of the singular
values will take place not only for g = $ but for other values of g as well.
The investigation of this conjecture is the goal of this paper.
Some numerical experiments were carried out before formulating the
theorems. For all g chosen, most of the singular values of A,,(g) appeared to
be located in a small neighborhood of the maximum one. For instance,
setting g = 0.75 and n = 32, we get

Pjn
= 4.443, l<j<n-5,
SINGULAR VALUES OF CAUCHY-TOEPLITZ MATRICES 101

TABLE 1

g Pill g Pin
0.1 10.2 0.7 3.88
0.2 5.34 0.8 5.34
0.3 3.88 0.9 10.2
0.4 3.30 1.1 10.2
0.5 3.14 2.1 10.2
0.6 3.30 10.1 10.2

and the smallest five singular values are

0.2259, 3.905, 4.340, 4.428, 4.441.

If we set g = 0.3 and n = 32, then

Pjn = 3.883, l<j<n-5,

and the smallest five singular values are

2.784, 3.507, 3.804, 3.871, 3.882.

Table 1 shows the maximum singular values which correspond to various g.

This table permits us to put forward the following hypotheses: (1) pl,( g)
depends on the fractional part of g only; (2) the minimum value of p In( g ) on
the interval 0 < g < 1 is equal to 7 and attained at g = i; (3) the function
pl,,(g - k) is even. It may not be quite right, as a matter of principle, that we
started by trying to guess the function; all the same, having tested the
function f(g) = a/ lsin z-gl, we were happy to see how well it fits the table.
Soon after that we derived a mathematically rigorous derivation of that very
function.
There are two prime results obtained in this paper. For the first, we
prove that for any g

77

Pjn(g) G~ j=l ,...,n, (I.51

lsin rg I ’

P:,(g)+ +p:,hd 73
(1.6)
lim *..

Ez-

n+m n sin” rg ’
102 EVGENIJ E. TYRTYSHNIKOV

For the second, we establish that for (gl > i

lim p,,(g) = 0. (1.7)

n--tm

In Section 2 some auxiliary statements are given. In Section 3 we prove

(1.5) and (1.6). In Section 4 the properties of inferior singular values are
studied. We prove that if Igl > i, then p,,, = O(r~‘/~-l~l). For ]g] = i we
have the two-sided estimate

Cl C2
-=Gp =G-, Cl,CZ > 0,
Inn “I In n

which is a more general form of the result brought forth in [5]. For Ig( < i
we have managed to prove only that

c
- G P,,,) c > 0.
In n

Note that we have gained an approach to the quantitative description of

the singular-value distribution when ]g] < $, and, in particular, of the Moler
phenomenon. That is, from Theorem 4.2 it follows that there are O(ln” n>
singular values which lie outside a given e-neighborhood of the maximum
singular value.
In this paper we have solved some problems which were called open in
[lo]. Section 5 discusses a new list of open issues.

2. INVERSE MATRICES

THEOREM 2.1 [lo]. Let A be a regular Cauchy matrix of order n:

(2.1)

Then its inverse is of the form

A,’ = U,AT,V,, (2.2)

SINGULAR VALUES OF CAUCHY-TOEPLITZ MATRICES 103

where

U,,=diag(u, ,..., u,,), V,,=diag(~,,...,u,). (2.3)

The components u,, . . . , u, and vl,. . , v, are defined as follows:

ui= ,fpYJ ll$Yryi)> i = 1,. ,n, (2.4)

i l#i

lfpj-Yl)lij(x.j-x,,:
Di=
i’ l#j
j=l,. .,n. (2.5)

Moreocer .

II’> (2.6)

AT,[C, ... O,,lT = [ 1 .. 11’. (2.7)

Some aspects of this theorem can he found in [2, 4, 6, 111.

COROLLARY 1. For any A,(g) of the form (1.1) we have

Ui=g~~(l-~)~(l+~), i=l,..., n, (2.8)

vj=gFi(l+f)‘~(l-;), i = l,...,n. (2.9)

We have (2.8) and (2.9) by taking

xi=i+g, Yj = j (2.10)

and substituting these values into (2.4) and (2.5).

COROLLARY 2. For any A ,,( g ) we have

ui = v ntl--i, i = l,...,n (2.11)

104 EVCENIJ E. TYRTYSHNIKOV

That is, in matrix language,

u,, = JnVn
Jr,> (2.12)

where

(2.13)

and consequently,

A ; ’ = J,,V,, J, AC’,Vn. (2.14)

COROLLARY 3. For any A,,(g) we have

"j(-8)=-vn+l-j(g)> j=l ,...,n. (2.15)

3. SUPERIOR SINGULAR VALUES

THEOREM 3.1. Far all n and all g real and not integer the inequality

(3.1)

is fulfilled.

Proof. Consider the function

(3.2)

(i is the imaginary unit) and its Fourier expansion

f(x) = fj Ckeikx (3.3)

k=-m

where, as it is easy to check,

(3.4)
SINGULAR VALUES OF CAUCHY-TOEPLITZ MATRICES 105

Build up the Toeplitz matrix (nonsymmetric)

‘,I = Ici-jl~,j=l (3.5)

Owing to (3.4) we see that

C,, = D,A,( g) D,, (3.6)

where

D,=diag(-l,l,-l,l,..., (-1)“). (3.7)

Thus, C, has the same singular values as those of A,(g) we are interested in.
So, in particular, we have

IIC,,lln = II&( g) j12. (3.8)

Now let us take arbitrary nonzero vectors

ZZ[Z” ... Z”_J, w=[Wc, ..* w,-llT, (3.9)

and note that the dot product (C,Z, w) can be written as follows:

f(x) dx (3.10)

Hence,

(3.11)

At the same time it is easy to verify that

IIC,lln = max (cnZ’w) (3.12)

= 1 llzllnll4
/1211* .
llwll2 = 1

Together the relationships (3.8), (3.11), (3.12) lead to the desired (3.1). w
106 EVCENIJ E. TYRTYSHNIKOV

THEOREM 3.2. The singular values of A,,(g) satisfy the relationship

lim &(d+ ... +d,(d =- 2 (3.13)

n+m n sin” r
h0

Proof. As mentioned above, the singular vahles of A,,(g) coincide with

those of C, of the form (3.5). Recollecting the theory of Fourier series and
making use of (3.2)-(3.41, we find

(3.14)

Let ej denote column j of the identity matrix. Then

j=l
2 IlC,ejlli = pf,( g) + . ’ . +d,(d. (3.15)

Next, let E > 0 be fixed and m be such that

- In

k=-m
c
(k+g)”
l +f kzrn
l
(k+g)”
< E. (3.16)

Then for m<j<n-m weget

n-j
d
IlC,,ejIIi =

k=_j+l
C ’
(k+g)” 2zr-E3 (3.17)

and therefore

(3.18)

Together with (3.1) and (3.15) this leads to

7Tz d,,(g)+ ... -cPLW 2

~- E Q lim <- (3.19)
sin” rg n-m n sin’rg ’
SINGULAR VALUES OF CAUCHY-TOEPLITZ MATRICES 107

where the expression in the middle is assumed to equal any limit point of the
sequence being treated. Since E is arbitrary, all the limit points are the same.
So the sequence of the averages of the squared singular values is convergent
to r”/sin” rrg. The proof is over. n

COHOLLAKY. Let F be fixed, 0 < E < rr / lsin rg 1, and let y, denote the
number of singular values pjn which satisfy the inequality

7T

Pjn <
___ - E. (3.20)
lsin 55-g]

Then

Y, = o(n) (3.21)

(o(n)/n + 0 as n -4.

REMARK. Theorem 3.2 is in fact a consequence of the recent develop-

ment of standard Toeplitz theory suggested by Parter [7] and then accom-
plished by Avram [l]. If f E L,(O, 1) and F is a real-valued continuous
function on the interval [0, Ml, M = sup11 f(x>l, we have

(3.22)

This is Avram’s generalization [l] of Parter’s theorem [7]. Though (3.22) is a

good deal more general than (3.13), the latter is well suited for our purposes
and allows a proof with rather simple tools.
We shall show in the next section that (3.21) becomes y, = O(ln” n) if
I&$

4. INFERIOR SINGULAR VALUES

The theory of gamma functions asserts that

g
lf!i -1
If
1
= r(;; g> + O(j”_‘). (4.1)
 108 EVGENIJ E. TYRTYSHNIKOV

Hence, (2.9) can be rewritten as

(n+l-j)y
uj = g +O((n+l-j)-‘-’
Ul- g)

(4.2)

Since

g sin rg
=-
(4.3)
r(i+g)ry- g) T ’

we have

Gj=y( ,,+;_j)x[l+o(f+ lj)]. (4.4)

This formula leads to all the constructions that will follow.

TIWOREM 4.1. For spectral norms of the inverses of Cauchy-Toeplitz

matrices we huve

c,nlgl+‘/” if Jgl> i,
lIA,‘(d II22
( c2 In n $ lgl=i,
(4.5)

where cl, c2 are positive constants independent of n.

Proof. Let m denote a positive integer such that, in accordance with

(4.4),

1i
0 L+

_i
n+l-j1 11
<m
( 1
-+
n+l-j
1
(4.6)
_I i

Then for n sufficiently large the formulae (2.6), (2.12), and (4.4) together
SINGULAR VALUES OF CAUCHY-TOEPLITZ MATRICES 109

yield

_++ . . . + Uf,) = ;(c; + . .. +cn”)

11
1 sin'rg Ilfl-4l~I j 28

H i
1
a-7 1+0 1,
n j=4m
= i
r&+1-j j n+l-j

1 sin”rg *+l-47n j 2/:

2----p-
2n =
j=Am i n+l-j I (4.7)

It is easy to see that if 1g 1> i, then for some C > 0 we have

Thus, we get (4.5) by setting

Now let Igl = i. Then

nfl-4m
c j =rL+E4rnn+i-j

j=4m n+l-j j=4m

~(n+l~n+~~m~-n~const,nInn,
j=4m J

which immediately implies

110 EVGENIJ E. TYRTYSHNIKOV

The estimate obtained is less accurate than the one we wish to prove.
Evidently, a tighter lower bound can be derived by choosing another initial
vector, instead of [I . . . llT as in (4.7). It is noticed in [S] that the vector

w=

[
q’ . . . -1
2! n
1T (4.6)

is much more appropriate to this case. In accordance with (4.4), with lgl = i
we have

Further, by (2.14)

(4.10)

It is not difficult to verify that

(Flj_f+g[2cln2n_l+l, i=l,...,n, (4.11)

where c is some positive constant which does not depend on n. As a

consequence, using (4.4) (4.6) with lgl = i, we obtain

,,A,l(g)(,~B~~+~4mn+1-~InYn+l-i
i = 4171
i i

~ ,rcn~/21 [(n +I)/21 ln2 [(n +I)/21

2T” i=4m
i i

> &rln3n, E> 0. (4.12)

SINGULAR VALUES OF CAUCHY-TOEPLITZ MATRICES 111

Thus, via (4.9) and (4.12) with )g( = i we conclude that

and that completes the proof.

THEOREM 4.2. For 1g ( < + the relationship

(4.13)

is valid.

Proof. According to (2.14) we write

IIA;l(g)ll;.= t 5 a-4
i=l j=l (j-i+ g)’

(4.14)

where

min(n + k,n)
5r4
s,= c v~+l_j+$“+- (4.15)
min(l+ k,l)
sin4 rg

and

k=j-i. (4.16)

Consider first the case k < 0. From (4.4) we have

j(n+l-j-lkl) 2g
(j+Ikl)(n+l-j)

II
1 1 1 1
(4.17)
n+l-j-lkl+j+lkI+T+ n+l-j ’
112 EVGENIJ E. TYRTYSHNIKOV

Suppose that g > 0. Since, obviously,

j(fz+1-j-lkl)
(j+lkl)(n+l- j) “’

we find

Sk<<-lkl+O(lnn). (4.18)

If g < 0 we need a bit more algebra:

(j+lkl)(n+l-j) Ikl
1+
j(n+l-j-lkl) n+l-j-/k1

=(l+&)(l+
n+!y_i)
1
n+l-i

BecauseO<2/gl<l, weget

S, = n - Ik(+ Iklo(lnn) + (4.19)

Evidently, if (4.18) is true then (4.19) is. So (4.19) holds in both cases.
Now let k > 1. This time we have

j(n+l-j+k)
s,= 5 (j-k)(n+l-j)
j=l+k

1 1 1
-+-+
n+l-j+k+j-k j
SINGULAR VALUES OF CAUCHY-TOEPLITZ MATRICES 113

In a similar way to the above we write

j(n+l-j+k)

(j-k)(n+1-j) =(l+&)(l+n+:-j)

1 1
=l+k -+
j-k n+l-j

1
+
n+l-j

and so conclude that if g > 0 then (4.19) holds, and if g < 0 then (4.18) does,
and that implies (4.19) as well. Thus, (4.19) is fulfilled in all the cases, and
therefore,
n-1

c sk
k=-n+l (k+d

n-1

c l
‘nk=_n+l (k+g)’

k2 1
+ (k+g)2 n+l-lkl

Tr2
< ~n+O(ln2n).
sin2 5rg

In accordance with (4.14) we find

Since the squared Frobenius norm of a matrix equals the sum of its squared
singular values, the inequality obtained above can be written in the form

From Theorem 3.1, each summand of the left side is nonnegative, and that
114 EVGENIJ E. TYRTYSHNIKOV

means that we actually have the equality

!, (i--F] =O(ln”n), (4.20)

which is equivalent to (4.13). This completes the proof. W

COROLLARY 1. If 161 <i, then we have

(4.21)

Indeed, from (4.19) we find

l/A,‘(g)ll,=L< ~+O(lnzn))l”=O(lnn),
P nn

which is what we want to prove.

COROLLARY 2. Zf 161 <i, then we have

(4.22)

COROLLARY 3. kt 161~ i. Then using the notation introduced in Theo-

rem 3.2, we have

y, = O(ln”n). (4.23)

In fact, for all pjn satisfying (3.201, it is easy to check that

1 lsin 5rg I
------->&E;
Pjn T

hence, by (4.20)

EZyn = O(ln’n),

which obviously yields (4.23).

SINGULAR VALUES OF CAUCHY-TOEPLITZ MATRICES 115

5. DISCUSSION OF THE RESULTS

Note that our results give the most complete information about minimum
singular values of A,,(g) for Igl = $. In this case (4.5) and (4.21) bracket p,,,
as follows:

Cl c2
-<p,,,G---> c,,c, > 0.
In n In n

The left inequality still holds for all Igl < i. But we do not know how
strong it is, so the question of getting a good upper bound in this case is still
open. What is more, we have as yet no idea whether p,, tends to zero or not
if lgl<+.
If lgl > i, then from (4.5) we have

1/2-lCl
P nn < cn c > 0.

We guess, however, that a more precise estimate might exist.

In accordance with (4.23) if 1g 1< f , then the number of singular values
lying outside a given e-neighborhood of the maximum singular value is
O(ln2n). If lgl > +, we can only state that this number is o(n), so the
problem of getting a more definite result in this case is still open.

The author wishes to thank Professor Seymour Parter for encouraging this
work, and the referee for pointing out Reference (11 underlying the remark in
Section 3.

REFERENCES

1 F. Avram, On bilinear forms in Gaussian random variables and Toeplitz matrices,

Probab. Theory Related Fields 79:37-45 (1988).
2 S. M. Belotserkovskij and I. K. Lifanov, Numerical Methods in Singular Integral
Equations, (in Russian), Nauka, Moscow, 1985.
3 V. Grenander and G. Szego, Toeplitz Forms and Their Applications, Univ. of
California Press, Berkeley, 1958.
4 G. Heinig and K. Rost, Algebraic Methods for Toeplitz-like Matrices and Opera-
tors, Akademie-Verlag, Berlin, 1984.
5 P. Kolesnikov, On the spectral condition number of a special Cauchy-Toeplitz
matrix (in Russian), in Computer Architecture and Numerical Methods, Dept. of
Numerical Mathematics, USSR Academy of Sciences, Moscow, 1990, pp. 13-18.
116 EVGENIJ E. TYRTYSHNIKOV

6 Yu. M. Nechepurenko, Some properties of Cauchy matrices (in Russian), in

Computer Architecture and Numerical Methods, Dept. of Numerical Mathemat-
ics, USSR Academy of Sciences, Moscow, 1988, pp. 93-102.
7 S. V. Parter, On the distribution of the singular values of Toeplitz matrices,
Lineur Algebra Appl. 80:115-130 (1986).
8 S. V. Parter, Remarks on the solution of Toephtz systems of equations, in
Proceedings of the Second International Symposium on Numerical Analysis,
Prague, 1987, pp. 75-81.
9 S. V. Parter, Extreme eigenvalues of Toeplitz forms and applications to elliptic
difference equations, Trans. Amer. Math. Sot. 99(1):153-192 (1961).
10 E. E. Tyrtyshnikov, Cauchy-Toeplitz matrices and some applications, Lineur
Algebra Appl., to appear.
11 E. E. Tyrtyshnikov, Toeplitz Matrices, Some of Their Analogs and Applications
(in Russian), Dept. of Numerical Mathematics, USSR Academy of Sciences,
Moscow, 1989.
12 V. V. Voevodin and E. E. Tyrtyshnikov, Toeplitz matrices and their applications,
in Computing Methods in Applied Sciences and Engineering, North-Holland,
Amsterdam, 1984, pp. 75-85.

Received 11 July 1990;final manuscript accepted 28 Not-ember 1990