Hankel Matrices

Hankel Matrices
Author(s): Harold Widom

Source: Transactions of the American Mathematical Society, Vol. 121, No. 1 (Jan., 1966),
pp. 1-35
Published by: American Mathematical Society
Stable URL: http://www.jstor.org/stable/1994330
Accessed: 19-08-2017 06:59 UTC
REFERENCES
Linked references are available on JSTOR for this article:
http://www.jstor.org/stable/1994330?seq=1&cid=pdf-reference#references_tab_contents
You may need to log in to JSTOR to access the linked references.
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide
range of content in a trusted digital archive. We use information technology and tools to increase productivity and
facilitate new forms of scholarship. For more information about JSTOR, please contact support@jstor.org.
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at
http://about.jstor.org/terms
American Mathematical Society is collaborating with JSTOR to digitize, preserve and extend
access to Transactions of the American Mathematical Society
This content downloaded from 14.139.38.11 on Sat, 19 Aug 2017 06:59:05 UTC
All use subject to http://about.jstor.org/terms
HANKEL MATRICES(1)
BY
HAROLD WIDOM(2)
1. Introduction. A Hankel matrix is a matrix, finite or infinite, whose j, k entr

is a function of j + k. We shall present here certain theorems whose commo
property is that they deal with spectral properties of Hankel matrices.
Nehari [3] has shown that an infinite Hankel matrix
H=(cj+k), jk=O1,
represents a bounded operator on the Hilbert space 12 if and only if there is a
bounded function 0 such that
(1) c.=- 2 c (0)e i}0dO, j= l

and that
|HJ| = inf|| bfJl
where the infimum is taken over all 4 satisfying (1). In ?2 we shall determine
analogously the smallest closed symmetric convex set containing the numerical
range of H. This will be shown to be the intersection, taken over all 0 satisfying
(1), of certain convex sets explicitly described in terms of b. Unfortunately
the determination of the numerical range itself, rather than its symmetrization,
seems much more difficult.
?3 is mainly devoted to the study of the rate of approach to zero of the eigen-
values of positive completely continuous Hankel matrices. Formulas, valid asymp-
totically for small s, are found for the number of eigenvalues greater than e.
Rather than involving the cj directly, or the functions 0b, the formulas involve
the measure on (-1,1) whose moments are the cj. With practically no more
work we are able to obtain similar formulas for integral operators on (0, cc)
whose kernels depend on the sum of the arguments.
In ?4 we consider finite Hankel matrices. It is assumed that the cj are ultimately

positive and well behaved and various theorems are proved concerning the limit-
ing behavior of the eigenvalues and eigenvectors of the matrices
HN = (Cj+k), j, k = O, 1, ...,N-1
as N - co.
Received by the editors September 24, 1964 and, in revised form, November 15, 1964.
(1) Supported in part by N.S.F. grant GP-1645.
(2) Sloan Foundation fellow.
2 HAROLD WIDOM [January
2. The numerical range. The numerical range of the bounded Hankel matrix
His the set
W(H) = {(Ha, a): a |=1}
=, ECj+kakdj: Y Iaj 12=.

{j,k=O
(Here we use bold face letters to represent vectors in 12.) W(H), like the numerical
range of any operator, is a convex set [2]. Let W14(H) denote the smallest sym-
metric convex set containing W(H). We use "co" to denote convex hull.
LEMMA 2.1.
W,(H) = co {2(Ha, a) + -(Ha, a): || al| = 1}.
Proof. Given a vector a let us write a = a' + ia" where a' and a" have real
components. An easy computation shows that
(Ha, a) = (Ha', a') + (Ha", a")

and
(Ha, a) + (Ha-, a) = (Ha', a') - (Ha", a").
Denote by WO the set which we allege W,(H) to be. The first identity abo
gether with the fact that j|a' 112 + || a" 112 =1| a 112, shows that W(H) c WO.
since WO is convex and symmetric this implies W,(H) c WO. The second id
shows that for every vector a with jj a jj = 1, the number
1 -)+1
2(Ha) + 2(Ha, a)
is of the form (z' - z") with z', z" E W(H), and so belongs to W,(H). This prov
Wo c- W(H).
The proof of the following lemma imitates the proof of a theorem of de Leeuw
and Rudin [1] which determines the extreme points of the unit sphere of H1.
(Actually, using the theorem of de Leeuw and Rudin, the Krein-Milman theorem,
and other heavy machinery, one could quickly prove a weakened version of the
lemma which would be sufficient for our purposes. It seems better though to
simply repeat the de Leeuw-Rudin argument in the present setting.) Recall that
Hp consists of those functions in Lp( - , it) whose Fourier coefficients for negative
values of the index all vanish.
LEMMA 2.2. Every function fe H1 which has L1 norm at most I is the limit
in the L1 norm, of convex combinations of squares of H2 functions each of L2
norm at most 1.
Proof. Since f is the limit in L1 of the (C, 1) means of its Fourier series we
may assume to begin with that f(O) is a trigonometric polynomial Yaie'j0. Le
1966] HANKEL MATRICES 3
F(z) denote the corresp

point C inside the unit c
G(z)= C F(z).
Then for any complex number /3 7 0,
F(e)=4 1 _ CeiO +41)G(e) 1 _-.o

If | I,=1 then we have, since |,/(eO - C)(1 - je)-= 1,
1 ei0- 2 / ej0-o \
- T ?= 1 G(e'o) = (1 + I IG(z)I.
(= real part.) Thus if /3 is chosen so that
Rfl e -' j I G(e 0)jdO = 0

1- CeiO
we shall have
1 2 e O )2 G(e'1 ) = IG(ei0)|i ? 1.
2 ?1i -
Moreover the analytic functions
1 (/z -C G(z)
2 /3B CZ
have one fewer zero inside the unit circle than does F.
ment repeatedly we find that F(z) is a convex com
analytic on j z j< 1, nowhere zero on z I<1, and
Each Gj(z) has a square root (Gi(z))112 which is
tinuous on j z 1, and (Gi(e'0))'12 are the desired
Given a convex set K containing 0, define
w(oc, K) = sup {r > 0: ra G K}
for complex numbers a of absolute value one. The closure of K is uniquely de-
termined by the function (oc, K). In the next lemma, the crucial one, we denote
by D the class of those bounded functions 0 which satisfy (1).
LEMMA 2.3. w)(xc, WT'(H)) = info e ,p | (O) + 40(-0)| .

Proof. Given a vector a E 12 let f denote the H2 function whose Fourier co-
efficients are the components of a. We have then
Rdec[(Ha,ii) + (Ha, a)] = ?([ wCj+kakaj + z Cj+kakaj]
- F (acjk + CiXJ+k)aka
- I f [co(O) + Jiq0(_0)]f(0)2dO.
Since ||fII2 = 27|a || we conclude, using Lemma 2.1, that
(o(a, Ws(H)) = sup Q aWs(H)
= 2sup {LP j [(X b(0) + ib( -0)]f(0)2dO:fe H2 X f | 2 < i}
and by Lemma 2.2 this is
2sup{f [a (O)?+ 4( -O)]f(O)dO: feH1, lif III '}.

Since this holds for all kG e1 it is immediate that
o(x, Ws(H)) < inf,, c | (O) + 1o(-0) l l t o
To prove the opposite inequality take any 0 e(D and consider the linear fun
tional on H1, considered as a real normed linear space, given by
f -* I {q [aOu(O) + iq(-O)]f(O)dO.
The norm of this linear functional is at most j(x, Ws(H)), and by the Ha
Banach theorem we can find an extension of it to L1 which has no larger norm.
Thus there is a function / E Loo such that
|| oo| < a)(a Ws(H)),
2 [a+ (O) + +i((- 0)]f(O)dO = f ifr(O)f(O)dO f e Hl.
This last identity shows that the Fourier coefficients of
-1(0) = f(O) - [a b(O) + ?5/(-0)]
for nonpositive values of the index all vanish. Thus f, = + ca4 belongs to
'D and
-|1o(O) + 41(-.-) | = -2 O() + ~(-O) || X _w(c, Ws(H)).
This completes the proof of the lemma.
19661 HANKEL MATRICES 5
Given a pair of complex nu

+ (ZZ,) 1/2 and major axis le
E Z = I(2(Z + Z'):
the set of real parts of quant
{4t~(o1z+4Jz'):IflI?1
( + 2il )
Thus
c)(a,EZz,) = sup3EiEz ,
In view of Lemma 2.3 this sug
over all 4) e@, of the smalles
would be true were it not for
We must somehow use only
this is the reason for the following considerations.
Given a complex valued function f the essential range of f , here denoted by
R(f), is defined as usual to be the set of complex numbers z with the property
that for each e > 0 the set
{0: If(0)-z I < ?}

has positive measure. Analogously given two functions f,g we define R(f,g)
to be the set of pairs of complex numbers (z, z') with the property that for each
s > 0 the set
{0:If(0)-z <ae and jg(0)-z'| C<}

has positive measure.
LEMMA 2.4. If f and g are bounded and F is a continuous function of two

variables then R(F(f, g)) = F(R(f, g)).
Proof. The inclusion R(F(f, g)) = F(R(f, g)) follows trivially from the defi-
nitions and the continuity of F. To prove the opposite inclusion, let A e R(F(f, g)).
Let P and Q be closed squares, say both of side s, covering the ranges of f and g
respectively. Divide P into four closed squares P1 P2 , P3 , P4 of side sf2 and Q
similarly into closed squares Q1, Q2, Q3, Q4. Then for the same pair
i,j (li_4, 1_j 4) the set
{0:f(0) Pi,g(0) E QJ, and j F(f(0), g(0)) - A

will have positive measure for all e > 0. (One need only find a pair for which
this is true for infinitely many s's of a sequence tending to zero.) Write
p(i) for this Pi and Q(l) for this Qj. If we apply this process repeatedly we ge
a decreasing sequence of closed squares P(k) and Q(k) with side s/2k such that
for each k the set
{O :f (0) E p(k) g(O) E Q(k), and I F(f(O), g(O)) - j < 8}

has positive measure for all s > 0. Let z be the common point to all the
p(k) and z' the point common to all the Q (k). Then since
{0:f (0) C p(k) g(O) E Q(k)}
has positive measure for each k we have (z, z') e R(f, g). Moreover for each
8, A is within e of nkF(P(k), Q(k)) = F(z, z'). Therefore A = F(z,z') E F(R(f, g)).
THEOREM 2.1. The closure of W,(H) is
n co u Ez:z.
C- (D (z,z') e R(0(0),-P(-O))
Proof. Let us write
E,p = co U Ez z.
(z,z') e R(4(0),4(-O))
By Lemma 2.3 it suffices to prove
w(aL,E) =)11G
co(,EJ) = 2 || a o(o) + do(-)|
If we apply the preceding lemma with
f(O) = 4(O), g(O) = k(-0), F(z, z') = a(a, Ez . ),

we obtain
R(o(a, E0,(0),(_0))) = o(ax,E+).

Since, as we have already seen, (xoe , E -2 = x + az'|, the desired identity
is established.
3. Positive Hankel matrices. In order for the infinite Hankel matrix H -(Cj+k)
to be positive, i.e. in order that every sum
N
z Cj+kakaj
j,k =O
be non-negative, it is necessary and sufficient that there exist a nondecreasing

function p on oo , oo) such that
Cj = XJ dl(x) j = 0,1,2,....
(See, for example, Theorem 1.2 of [5].) Since in order for H to represent a bound
operator on 12 it is nec
,u(x) be constant for x < -1 and for x > 1. Therefore throughout this section
we shall assume
Cj= xidp(x), j= 0,1,2,
where ,u is a nondecreasing function on [-1,1].
THEOREM 3.1. The following are equivalent:

(a) H represents a bounded operator on 12,
(b) cj =O(j- 1) as j ->o

(c) ju(l) -p(x) = O(1 -x) as x 1+ and p(x) - p(-1) = 0(1 + x) as x 1
Proof. First we prove the equivalence of (b) and (c). If we write a = y + 112
where pj is constant for x < 0 and [2 constant for x > 0, we see that we may
assume to begin with that p is constant for x < 0. Thus
cj = xI d1(j) >(I - 6)1(p(l) - j(l - 6)).
If we set j = [3-1] we can conclude that (b) implies (c). Conversely if (c)
then integration by parts gives
c = j - x [(l) -(x)] dx ? Aj xj-1(l - x) dx = A/(j + 1)

for some constant A, so (b) holds.
Hilbert's inequality
N _ N
shows that (b) implies (a). We shall show that (a) implies (c). In fact since
cJC+kak,j = I | ajxj 2 dp(x)

as long as only finitely many aj are not zero, we have if H is bounded
/| ajxj12dji(x) ? E la 12,
and this holds whenever the power series is uniformly convergent on [-1,1].
If aj = rJ (0 < r < 1) we conclude
f' du(x) < IIHII

J -1 (1- rx)2 r2
and so
8 HIAROLD WIDOM [January
p(1) - /(r) dr 1du(x) < ||H||

(1-r2)2 Jr (1- rx)2 = 1-r2
u(l1)- I(r) < 2||H|| (1-r),

and a similar inequality holds for u(- r) - -1).
The next lemma, which we state in a general Hilbert space setting, will enable
us to pass from H to a more convenient (integral) operator with similar spectral
properties.
LEMMA 3.1. LetXY and X' be Hilbert spaces and Ta continuous linear transfor-
mation from X' into X with dense range. Define the positive self-adjoint
operator A on X)" by the bilinear form
(Ax, y) = (Tx,Ty)
and define A on T(YI) by AT(Tx)= T(Ax). Then A is well defined, extends to
a bounded positive operator on Y{, and the spectral subspaces &A of A and
St of A are related by
-ff = T &A
at every point of continuity of the spectralfamily of A. In addition A is completely
continuous if and only if A is.
Proof. A is well defined since it follows from the definition of A that Tx = Tx'
implies Ax = Ax'. If x = Tx then
A| A12 = (A Tx, A Tx) = (TAx, TAX)

= (A3X, X) < 11 A112(Ax,x X) = 11 A112 11 X 112
(We have used the fact that A3 ? A jj2A in the usual partial ordering of the
bounded self-adjoint operators on V.) Therefore A extends to a bounded oper-
ator on V. If x = Tx and y = Ty then
(AxZ,y) = (TAx, Ty) = (A2x, y)

and since A is self-adjoint and T(,V') is dense in X, the extended A is self-adjoint.
Let El be the spectral family of J. This may be defined by
_R = sup {Th:ERA = TR _ AS}
where the . denote projections. At any point of continuity of A we also have
A = inf {f': EA = if _ AS}}.
Since MA is invariant under A, TWA is invariant under A. Moreover if - E T//,

say if -=Tx with x e A, then
(Ax,x) = (A2x,x) A(Ax,x) =
This shows that T/A/4 AX = ,A. Similarly T.A a A-. Now the sub-
spaces TW&A and T//AI are mutually orthogonal, since if xe A and yEfA1
then also Ax e X and so
(Tx,Ty) = (Ax,y)=O.
In addition TWIA + TWAk = T(J,A + &') = T(Y'), which is d

follows that TWA./1 and TIWI are orthogonal complements of each o
inclusion T/IAt c: W.-#, which we have already seen, gives TWA' xA This,
together with the inclusion TIA ' ,, shows Ti/IA = &A.
Finally, A is completely continuous if and only if IAI is finite dimensional
for each A > 0, and similarly for A. Since TI/4' = .4L1, it is immediate that A
is completely continuous whenever A is. The converse would also be immediate
if Twere one-one. Thus if hX is the null space of Tand A is completely continuous
then A when restricted to X1 is completely continuous. But since A vanishes
on X, A itself is completely continuous.
To see what the lemma does for us, assume that H is a bounded Hankel matrix.
Take for A' and # the spaces 12 and L2(yU) respectively. Then if a, b E 12 we have
oo 1
(2) (Ha, b) = z Cj+kakbJ =11 ( x ax)( bjxi) dlu(x).

j,k=O
The interchange of summation with integration is justified b

since | cj+k akbj I < oo . If we take b = a we see that
T: a - ajxi
is a continuous linear transformation from 12 to L2(AY). It follows from the
Weierstrass approximation theorem that T has dense range. Identity (2) shows
that the operator A of the lemma is just H. As for A, which we shall here call J,
we have whenever a E 12,
JTa = THa = xj E Cj+akak

I k
= J k zXyi +ka
-1 j, k
= 1Ta(~y) dlt(y).
J1 - xy (y
Now let fe L2(1t) be arbitrary. Then f is the limit, in L2(4u), of a sequence of
polynomials fn. Since J is continuous,
Jf(x) = lim Jf"(x) = lim My) dy(y)

n -oo Jnoo 1 dxy
These limits are taken with r
lim J 1- dyi(y) = J t d,ut(y)
for each x in (-1, 1), we have almost everywhere (p)
Jf(x) = j j d/l(y).
Thus J is the integral operator on L2(u) with kernel (1 - xy)-' .
THEOREM 3.2. The following are equivalent:

(a) H is completely continuous,
(b) cj = o(jf 1) as j -> ,

(c) ,u(l) -p(x) = o(1 -x) as x 1- and y(x) - p(-1) = o(1 + x) as x 1
Proof. We shall not give the proof of the equivalence of (b) and (c) ,which
is similar to the proof of the analogous part of Theorem 3.1.
Assume cj=o(j-'). Then given e > O we can write cj = cJ + c, where all
but finitely many cJ are zero and where I C" I < 8 /(j + 1) for all j. Then
H = H' + H" where H' has finite rank and, by Hilbert's inequality, || H" ?< n.
Thus H is completely continuous.
Finally assume H, and so J, is completely continuous. Let
f,(X) = (,u(1) - Y(r)) "21XZ ,r](X)
(X = characteristic function). For any fe L2(P) we have by Schwarz's inequal
fr(x)f(x)du(x) < (u(l) - p(r4))1 J du(x) J f|(X) 12du(x)
= Jr If(x)I2du(x) 0
as r -1 1. (Note that ,u is continuous at x = 1, by Theorem 3.1.) Thus fr converge

weakly to 0, and since J is completely continuous this implies Jfr converges
strongly (i.e. in the norm of L2(p)) to 0. Since
Jfr(x) = ((l1) - u(r))- 12 r1di(Y) > (P(1)- P(r))1x 2
we have
jf d~ I1 djt(x) > u1) -

I JfrkxIdAl(x) ~jI(j(l) - (r)) (1-rX)2 (1- r2)2
Thus p(1) - j(r) = o(1 - r) as r -- 1, and similarly p(- r) - p(- 1) = o(1 - r).
The remainder of this section is devoted to the determination of the rate of
convergence to zero of the eigenvalues of certain completely continuous H.
This is accomplished by a change of variable that transforms J into a modified

convolution operator, for which there are available techniques. Suppose that ,u
is absolutely continuous and that f is an eigenfunction of J corresponding to
the eigenvalue A, so that
jfi (x) 12'(x)dx < oo

(3)
ff lt_(y) dy = if(x).
If we set
F(x) = [p'(tanh x)]112 sech x f (tanh x)
then (3) is equivalent to

00
I fF(x) 12dx < oo,

-00
,0 [4u(tanh x)u'(tanh Y)]2 F(y)dy = F(x

J0 cosh (x - y) Fyd ~()
Thus the eigenvalues of J, and so of H, are the eigenvalues of the integral opera-
tor on L2(- 00, oo) with kernel
[,u'(tanhx)ji'(tanhy)]1/2sech (x - y).
We shall write An(A) (n = 0,1,2, ..) for the nth largest positive eigenvalue of
a completely continuous self-adjoint operator A; N(e; A) will denote the number
of eigenvalues of A which exceed s. In both cases an eigenvalue of multiplicity
m is counted m times. The minimax characterization of the eigenvalues is
n(A)-= inf sup (Af,f)

gl,...,gn f-L91,- -,9n (f9f)
(See, for example, [4], p. 237.) This has several well known and
which we collect together as a lemma without proof.
LEMMA 3.2. (a) If A1 < A2 then )ln(Al) < 2l(A2) and N(s;A1) < N(e;A2).
(b) If B is a self-adjoint operator satisfying || B ?| < 1 then A%(BAB) < An(A)
and N(e; BAB) _ N(e; A) .
(c) inl 1n2(A1 + A2) _ inl1(Al) + 2n2(A2) and N(e1 + e2;A1 + A2) < N(e1 ;A1
+ N(82A2)A-
What we are going to do now is find an asymptotic formula for
N(e; [V(x) V(y)]112sech(x - y))
under the assumption that V

This will be accomplished by dividing the x, y plane into three regions, and so
expressing the kernel as a sum of three kernels. All the eigenvalues of the first
kernel will be small. The eigenvalues of the second kernel will be dominated
by the eigenvalues of sech(x + y) on (0, oo) and these, we shall show are very
rapidly convergent to zero. The eigenvalues of the third kernel, the main con-
tribution, will be estimated in terms of the eigenvalues of sech(x - y) on a finite
interval.
We shall use certain results of [7] (Lemmas 1 and 3 of ?2) which can be stated
as follows. Given a positive number 6, let
nfr n7r
2(1 - 3)' 2(1 + 6)
Then
(4) lim s n (siy(x-y) 1 =x 1,

nR-40 7r(X - y)
and for some positive constant cj
(5) Yn, (n(xX , -1=<x, y <i) < 1/exp(c6n)

for sufficiently large n.
LEMMA 3.3. For any 3 > 0 there exists constants Aj and oca, such that when-
ever x > aj we have, for all e > 0,
?< (1 + ) - sech --E + A,

N(c; sech(x - y),O ? x, y ? a) >
where sech'lu is interpreted to be 0 if u > 1.
Proof. First a simple change of variable gives
N(c;sech(x-y), 0 _ x,y < a) = N(e; jsech-2(x-y) -- < x,y_).
Now
- f)c ei xsech xdx = isech
so that if f E L2( -1, 1) and F(4) - f'1 eigxf(x)dx, we have
:S f:X 2sech (x-y)f(

1 P2 0
< JY2
=2 - I F(4)12dX +
2 7t i: 1 nY2(x -Y)f(y)f(x) d
Therefore by Lemma 3.2 (a) and (5)
A,1(sech (x - y), 0 x, y < oa) < e 5 + 7 sech 2c!(1 + 6)
for sufficiently large n. If we choose

7;2
OcX >2c,(1 + 6)
then we shall have for all a > a5 and sufficiently large n
Ai1(sech(x-y), O < x, y < oa) < 1_ sech 2(1+6)
which gives the first of the desired inequalities. Similarly
2 f sech- IF(Q)I2dd > 2 sech11 f I F(4)I2dl
= zsech -ir f f (X f(y)f(x)dydx

and so, by Lemma 3.2(a) and (4) we shall have for sufficiently large n
A),(sech(x-y), 0 < x, y < a) > I + ,sech 2 (l -,5)
and this gives the other inequality.

Next we prove the needed estimate for the eigenvalues of sech(x + y) on
(0,oo).
LEMMA 3.4. There are constants A,B such that for all e >0
N(s; sech(x + y), O < x, y < oo) < A log s- 1 + B.
Proof. Let K2(x, y) be the iterated kernel

00
K2(x y) = f sech(x + z) sech(z + y) dz.
This function can be extended to be analytic in the polycyli

< 7/2, so we have for O < x, y <2 t/2
00
K2(X,y) = , aj,kxjy
j,k = 0
where, for any R > 2/r,
a aj <k ARJ+k
(Here and elsewhere we use the letter "A" to denote a constant, which will vary
with different occurrences of the letter.) We shall take R = 1 for simplicity. The
kernel
E a j, kX y
j,k=O
has rank at most n, and the operator on (0,1) with kernel

00
, aj,kX Y
j,k=n
has norm at most the square root of
1/2 1/2 oo 2
/ 2 aj,kxy 2 dy
j,k=n
Therefore by Lemma 3.2(c)
2n(K2, O < x y Y ) < A2-.

Next, we can write
00 1)j+k
K2(x,y) = 2= O-j+ k + I"'
Suppose fe L2(2, oo) and
00
f(x)e-(2i+ 1)x dx=0, j =O, 1, , n-1.

1/2
Then
K2(x,y)f(y)f(x) dydx | 2 If 2 L e-(j+k+ 1)

1/2 1/2 j,k =n
< Ae -2nl
It therefore follows from the minimax principle that
il(K2 X,-<xY < oo) < Ae 2.
To put together into a single estimate for X,1(K2) the ones we have obtained,
let P be the projection operator on L2(O, oo) given by
Pf = X(1/2,oo) f,
P' the complementary projec
Pf = X(o 1/2)
and K2 the integral operator on L2(O, oc) with kernel K2(x, y). T
3.2(c)
AdK(2)-!! A2n(PK2P' + PFK2P) + AZ(PK2P) + 4(P'K2P')

We have shown
)4PK2P) ! Ae2, A(P'K2P') ? A2".
Since the eigenvalues of the square of an operator are the squares

values of the operator, we have
2J(PK2P' + P'K2P)2 < A2J(PK2P'K2P + P'K2PK2P')

< %n(PK2P'K2P) + An(P'K2PK2P')
By Lemma 3.2(b)
At,(PK2P'K2P) 4(K2P'K2)
It is an easy fact that the set of nonzero eigenvalues of the product of two opera-
tors is independent of the order of the factors, so that
A(K2P2K2)- =
Since K2 is positive (and it is for this reason t

rather than sech(x + y)) we have
K2 <? KI2 K2
which implies
P'K2P' < 1 K2 jf P'K2P'

and so by Lemma 3.2(a)
2PK2P') ? | K- 2 1|(P'K2P') ? A2-.

Similarly
A),(P'K2PK2P') : AeC 2n
and we deduce finally
A4n(K2) < A(' + 2-nf2).
Since A.(sech(x + y), 0 ? x, y < c )2 < AJK2) the desired inequality is immediate.
LEMMA 3.5. Assume V(x) is locally integrable and nonnegatve, monotontit
for large I x I, and tends to zero at + X . Let a(s) denote the measure of the set
{x: V(x) > e /i}
and assume that as e - O, v(e) -+ o and
[() loge 8-1]2/3 = 0(f ((x)) dx)

L~~\eJogej O~j (x2 - 1)1/2 /
Then as se-0
N(e; [V(x) (y)]'2sech(x - y)) - 4 ((2+o()))2) dx.
(The condition v(e) -* oo is equivalent to V having unbounded support. The

other condition on a, the ugly one, is actually very weak. It is not hard to see
for example, that if a is absolutely continuous and
lim a(g)-413 1 a'(e)I eloge'
exists, whether finitely or infinitely, then the condition holds.)
Proof. We assume first that V is bounded and monotonic in [0, oo) and in
(-oo,0], and we write V= V1 + V2 where V2= 0 in (0,oo) and V1 = 0 in
(-oo, 0). Let av(e) (i = 1,2) be the measure of the set where Vi exceeds e /li.
Let s, 3, a > 0. (Think of e and 3 as small and a as large.) Let M be the smallest
integer such that Vj(Mcc) < se i. (If V1(x) > s /7 for all x take M = 0.) We divide
the quarter plane x, y > 0 into 2M + 1 regions, M squares
Qk = {k<x< (k +), ko< y <(k+), O<?k<M,
one quarter plane
P = {x>MoM, y>Mc*}
and M pairs of strips
Sk = {k <x<(k+ 1), y>(k+ 1)o}
u {x > (k + 1), ka < y < (k + 1)a}, 0 < k < M.
Accordingly the kernel
K(')(x,y) = [V1(x) V1(y)] 1/2 sech (x - y)
is written as a sum of 2M + 1 kernels. Let us first estimate
N(Bm; Kyln(x, y), Sk)t
By the monotonicity of VI, and Lemma 3.2(b), this number is not decrease
if first [ I/(x) V,(y)]112

of quarter planes
Sk= {x <(k + 1)oa,y >(k + 1)a}}u {x >(k +l)oc,y <(k + 1)ox}.

By the translation-invariant nature of sech(x -y),
N(Qi; V(k() sech(x - y), Sk)
= N(ii V(kx)sech (x-y), {x < 0,y > 0} u {x > 0,y < 0}).
Now the second iterate of the kernel
sech (x - y) [X(- oo,O)(x)X(O ,oo)(y) + X(O,oo)(x)X(- oo,O)(Y)]

is
K2(X, Y)x(o,,)(X)x(o,00)(Y) + K2( - - Y)X(- 00,O)(X)X(- oo,o)(Y)

where K2 is, as in the proof of the preceding lemma, the second iterate of
sech(x + y) on (0, oo). Therefore using that lemma we conclude that there are
constants A and B so that for all il > 0
N(q;, K(')(x,, y) , Sk) < A log ()+ B .

Consequently
N(EK( )(XI Y)IU SO) < z NMK)XY) ~Sk)

k~=O
(6) < A A, log M k + AM

k=O
< A z log + AM2

k=O
where A now depends on 3.
We consider next
N(q; K(1)(x, y), Qk)-
It follows from the monotonicity of VI that this number is not decreased if

[V,(x) V,(y)]112 is replaced by V(ko) and not increased if it is replaced byV((k + l)ac).
After having performed one of these replacements the resulting quantity is un-
changed if Qk is replaced by 0 < x < a, 0 < y < a, and so by Lemma 3.3 we
have, as long as a > a,
< (1+3)-2 sech 3-)e + A

N(q; K(')(x, l y), Qk) (1- s2 (r(kc+)
1 ? (1-3) sech' (1+ 8)f -A.
7i2 7rV((k +1)cc)
On P we have [V(x) V(y)]11

K(l (x, y)Xp(x,y) is at most c
with kernel sech(x - y), which is
f sech x dx = .
It follows that
N(s,K(l)(x,y),p) = 0.
Now the M + 1 kernels K(')xQK,K( )Zp are mutually orthogo

is a direct sum, so for any i1 the N(q) for the sum is the sum
N(Qj)'s. Therefore
N (? K(')(xy) ,U Qk U P)
< ( sech + AM
2 k1 rrV(kcx)
and
N 3 K(l)(x y) UQk UP )
2 M-2
> (1-)- z sech- + -AM.
k=O 7tV((k +1)oc)
In the last sum we have dropped the term k = M- 1 since, by the definition of
M, V(Mcx) < s In. If we use these inequalities together with (6) and apply Lemma
3.2(c) we find
N + a> ; K(')(x y))
A 2(~~x M-1
1 - b k=O tV(lx)
- - M- (1 2 m11 _ - M2- A og,

>( sech Alog )
k= t V(kxc)
where we have used the fact that log u- < sech' u . Now
M-1 r
E sech l < sech' V + sech dx
k=O 7tV(kLx) = trV(0) rrV(xLX)
S 1f00ur(sx).
= sech- 1 V(O)+ a| (X2 - 1)1/
The last identity holds because each of the two integrals
f sech' dx, dx
idl(x) j1(x 2 - 1)1/2 d
represents the area of the set in the first quadrant where V(x) se
since M?<'1Q )+1,
N((1 ' 3+ 3>;K(1)(x,y))
< (1 + ~ A(1 + 6) 2 00 a 1(ex) d A [cT(e) 112 A 1

a ~~j2 (x 2 - 1)1/2 [t j+9o
with a similar inequality holding in the opposit
hold for the kernel
K 2(x,y)= [V2(x) V2(y)]'12sech(x - y).

Finally if we set
K(x, y) = [V(x) V(y)] 1'2sech (x - y)

we have
N(Qi + q';K) - [N(Qi;K()) + N(q;K(2))]

? N(q'; V(O)sech (x-y), {x < O,y > O} u {x > O,y < O})
? Alog[V(O)/f'] + B,
as we have seen before. Therefore
N(( + +23>s;K(x,y))
< (I + A )(1+ 6)2 d+A[ c) + 1]2+Aoxlog !

Now set
= max [ C()2/3(log +- )- 1/3

lf ot 3 -2+ oc, then c(e)2 < Aloges1. Since
1 _ _ _ _ _)
log = = 0 2 o dx2 )
(this is an easy consequence of c(s) -0 oo) we would have for sm
and 3
N(( TIf6
3+ 26) s;K(x,Y)) < (1+23) -2 (x) dx.
os >y X2 - 1)1/2
If ax> 3-2 + ax, then
L'X = C(8)213(jog&- 1
and our other assumption on or(8) yields the same inequality. An analogous in-
equality holds in the opposite direction.
We have assumed that V was monotonic in [0, oo) and (- oo, 0]. If now V
is an arbitrary function satisfying the conditions of the lemma we can find a
function V' which agrees with V for sufficiently large x, say for |x I > fi, and
which is monotonic in [0, oo) and (- oc,O] . We decompose the x, y plane into
the square |Ix | <,the two quarter planes
{X >11,y > I3},
{x< -/1,y< -I3}

and the remainder. We now use arguments quite similar to ones already used
and find that
N 1-Us + 33)s; [V(x)V(y)]1t2sech(x -y)
< (I ''it2
+2 2J1(X2
X v6- )1)1/2
dx+Alog-
dx+lg
where r'(8) is the measure of the set where V'(x) > s/ I. We shall not go through
the details. Since '(s) a(s) + 2,B we find that for sufficiently small 3 and s
N 1 + 33)s; [V(x)V(y)]1/2sech(x-y))
< (1 + 33) i22 (x21)12 dx.
The reverse inequality holds with 3 replaced by -3 everywhere, and since 3

is arbitrarily small the assertion of the lemma follows.
We now have immediately the following result for the eigenvalues of Hankel
matrices.
THEOREM 3.3. Assume j is absolutely continuous and V(x) = ,L'(tanhx)

satisfies the conditions of Lemma 3.5. Then
N(s;H) 2 4 0 f O&+ o(s))xdx

N(s; H) - 7g2 AX2 - 1)1/2 dx.
Let us take as an example the matrix
Hh=era+0k,+-a)1a,-3 1n 0 Jdok .fa21,-wh,
where a , - 1,-2, .. and oe>0. If a > we have
(j + a),' -F(oc
and the theorem shows, after an elementary computation, that
An(H) = 1/ exp[(2oen) 1122 + o(n 12)].
If a < 0 but is not a negative integer then H is no longer positive, but the same
asymptotic formula holds since H is a matrix of finite rank plus a matrix
0 H'
where H' is of the same form as H but with positive a.

The preceding theorem gives us no information in case ji is supported in a
closed subinterval of (-1,1). In this case the cj tend to zero exponentially

j -+ oo and it is easy to conclude that the same is true of A"(H) as n -+ oo. The
following theorem gives more precise information.
THEOREM 3.4. Assume /u' is bounded, vanishes outside the interval (

where -1 < a < b < 1, and is bounded away from zero in every closed sub-
interval. Then
lim ! log K - ) K a_
where K is the complete elliptic integral of the first kind,
rn/2
K(r) = (1 - r2 sin20)-Of/2dO.
Proof. The theorem follows easily from the special case where i' = X(a,b)*
Then
in(H) = An(sech(x - y), tanh a < x, y < tanh b)
- An(fsech f(x-y), -1 < x, y < 1)

where ,B - '(tanh b - tanh a). The conclusion follows from Theorem II of [7].
One might wonder whether it is true that the asymptotic behavior of the
as j -s oo determines the asymptotic behavior of A"(H) as n -+ oo . Theorem
shows that it is not. If for example
/ll2 1/2
cj= x3dx, C3 J x3dx
O 1~~~~~~/4
then clearly cj-Cj', but the limits

1 1
urmn- log-
are different for the two matr

j, rather than the cP, which di
values.
The ideas of this section can be applied equally well to integral operators on
L2(0, oo) with kernels of the form k(s + t) where
k(s)= fe eSXdy(x)
In analogy with the discrete case one can show that the operator is bounded
and only if ,t(x) = O(x) as x -+ oo and i,(x) - p(O) = O(x) as x .-* 0; the con-
ditions for complete continuity are obtained by replacing the O's by o's. If the
operator is bounded then the map
f(s) J x'f(s)ds
is continuous from L2(0, 0) to L2(A), and if we apply Lem

= L2(0, o e'), =' L2(), and T the above map, then A is just the integral
operator on L2(0, oo) with kernel k(s + t) and A is the integral operator on L2(dp)
with kernel (x + y) -'. If , is absolutely continuous a change of variable sho
that this operator has the same eigenvalues as
[iht(e 2) p'(e 2)]'/2 sech(x -y)
on L2(-_ c, oo), and one can easily obtain analogues of Theorems 3.3 and 3.4.
4. Finite Hankel matrices. We shall consider here certain limiting properties

of the finite Hankel matrices
HN = (cJ+k), j,k =0 1,*-,N-1.

The cj will be assumed to be nor-negative throughout and the type of result
tained will depend on the behavior of the cj as j -+ Xo . There will be four c
(;) Cj = 0(j- );
(ii) cj ' L(j) /j, L slowly varying and nondecreasing;
(iii) c; L(j)jaI L slowly varying and a > -1;
(iv) log ci /log,j -s o .
In cases (i) and (iii) it will be shown that the HN, suitably normalized, converg
in the uniform operator topology to a limiting completely continuous opera
Since this implies that the spectrat family of these normalized HN converge uni
formly to the spectral family of the limiting operator [4, p. 372], the limit
behavior of the eigenvalues and eigenvectors will have been determined. In
(ii) we shall establish an asymptotic formula for the number of eigenvalues
HN which exceed a constant times L(N), and shall determine the limiting behavio
of the eigenvector corresponding to the largest eigenvalue. In case (iv) we s
determine only the asymptotic behavior of the largest eigenvalue and corre-
sponding eigenvector.
First we prove a simple lemma.
LEMMA 4.1. Let AN be operators on Hilbert space, A a completely con-

tinuous operator. Assume that Jor any uniformly bounded sequence of vectors
{XN} , XN + X weakly implies ANXN -Ax strongly. Then JJ AN- A J - 0.
Proof. If the conclusion of the theorem were false there would exist a sequence
of vectors XN with JXN II = 1 such that
11 (AN- A)xN JJ ++ 0.
By choosing a subsequence if necessary we may assume that XN converges weakly

to some vector x. Then by our main assumption ANXN -Ax strongly, and by
the complete continuity of A, AXN- Ax strongly. Thus
11 (AN-A)xN JJ 0,
which is a contradiction.
Let PN be the projection operator on 12 defined by
PN{aO, ...* aN..15aN,* } = {aO. * 5aN-1 ...
If cj = o(j-1) then the infinite Hankel matrix H is completely continuous (T

rem 3.2) and HN may be identified in an obvious way with PNHPN.
THEOREM 4.1. If cj = o(j-') then PNHPN -+ H in the uniform operator

topology.
Proof. If aN-+ a weakly then PNaN -+ a weakly (since PN + identity strongly)

and the complete continuity of H implies that HPNaN - Ha strongly. Consequently
also PNHPNaN -Ha strongly and we may apply the lemma.
If A is an eigenvalue of HN with corresponding eigenvector {ao,-, aN.
then A is an eigenvalue of the integral operator on L2(0, 1) with kernel
Nc[Nx] +[NY]
and has corresponding eigenfunction a[NX]. Moreover all eigenvalues and eigen-
functions of the kernel arise in this way and we may therefore consider the integral
operator as completely equivalent to HN. In case (iii), (NCN)- I times this operator
converges uniformly'.
THEOREM 4.2. If cj L(j)jf with L slowly varying and a > -1 then the
integral operator on L2(0,1) with kernel
c-1
CN C(NX] + [Ny]
converges in the uniform operator topology to the operator with kernel (x + y)'.
Proof. We first show that for any 3 > 0 we have
(7) L(Nx) < x-"L(N)

if Nx is sufficiently large. (Throughout the argument x and y will be confined
to the open interval (0.1).) We know that for sufficiently large u we shall have
L(u) < (1 + 3) L(eu)

and so by induction
L(u) < (1 + 6)'L(e'u).
If we set n = [logx-1], u = Nx we deduce
L(Nx) <x -L(e(lo9x- 1Nx)
and so (7) holds for large Nx.

Next we show that for some constant A,
(8) CN C[Nx]+[NyI < (x + Y) + A

for all N,x,y.
It follows from (7) that if N(x + y) is sufficiently large,
C[Nx] + [Ny] ([Nx] + [Ny]\aL([Nx] + [Ny]) a-d5

CN N / L(N)
If N(x + y) remains bounded then
C[Nx] + [NY] < A

N C -N(x
Now let fN be a sequence of functions in L2(0, 1) with
weakly. It is easy to see that for each x > 0
CN C(Nx]+[Ny] - (x + y),
uniformly in y. Therefore, for each x > 0,
J4l C(C[Nx] + [NY] d
=Jfo [ CX(N (x + y)] fN(y)dy + f (x + y)efN(y)dy
(X + y)af(y)dy.
But from (8), using Schwarz's inequality,
Jl0 C(NX]+(N(y) dy < 2 f (x + y)2(a-) dy + A.
If 3 is chosen so that a -3> -1 then the function on the right is integrable over
(0,1), and we can therefore use the Lebesgue dominated convergence theorem
to conclude that
1~~~~~~~~~~~i 1
c CNc CNX]+cNy]fN(Y) dy (
in L2(0, 1). The theorem therefore follows from Lemma 4.1.

In case (ii) we can apply an entirely analogous argument, but since in this case
(x + y) l is not the kernel of a completely continuous operator Lemma 4.1 is
not applicable and we cannot assert (in fact we do not have) uniform convergence
of the operator with kernel CNK1 CENX]+[NY to the operator with kernel (x + y)-.
Of course we do have strong convergence, and this has some consequences [4, pp.
368-369] but they are of less interest than the consequences of uniform convergence.
For example little can be concluded about the limiting behavior of the largest
eigenvalue or its corresponding eigenfunction. We therefore use a different approach.
In the following theorem the hypothesis concerning the cj is stronger, but
not very much so, than that stated in (ii).
THEOREM 4.3. Assume that L(x) is absolutely continuous in x ? 0 and that

for large x,
L (x) , O
xL(x)
Let cn= L(n)(n + 1). Then for any a in O < a < 1, N(ltaL(N);HN), the
number of eigenvalues of HN which exceed icaL(N), is asymptotic as N oo
to the area in the first quadrant bounded by
x=log N, L(eX) sech - y = a L(N).
Proof. If Lis changed on a finite interval then N(e; HN) is changed by at most
a bounded amount and the same is true of the area described. Therefore we may
assume that L(x) (x + 1) -1 is nonincreasing throughout x ? 0. If we denote
the eigenvalues of HN by 2N,j (O <j < N) then we have for each positive integer n,
N-1 = N-1 L(jn +j1)L(j1 +j2)... L(n-1 +jn)

j=- j--hjn=? (in J il + h )+Ul + X 2 + ) + (in-1 + in + 1)
If in the sum on the right any one of the j's is restricted to be zero the resultin
sum is O(L(N)n). Therefore if we use the well-known inequalities comparing
series of monotone terms with a corresponding integral we obtain
jE NJ ..(.x11(1x21..x~+~1 dxl . dxn

s =s 0 N A (Xn + )(X+ XI + 1 (X * X2 - (X,,-1 + Xn + )
(9)
+ O(L(N)n).
Set
M(tl, * **, tn- 1~ X)
L[(1 + tl)x-1] L[(1 + tn-.1)tn-.2 ... tlx -]L[(t, ... tn_I. + l)x]
(I + ti ... tn-1)(ti + l)(2 + 1) * *(tn-1 + 1)
Then if we define the new variables t- -, tn_-1, x by
Xk + 1= t ... tkx, 1?< k < n,
Xn + I =X,
the integral on the right side of (9) becomes
(10) mdtx, ...dt dtn_ 1

1/2 X J
where the inner n - 1-fold integral is taken over the region
t_ .. tk-lx<N+-, 1? k<n.
Now given 5 > 0 there is a constant X1 so that
L(yx)?(1 +y)5L(x), y 0,x XI.

Therefore if x > X1 we have
M(tls , 'tn-l,X)
[1 + t1l]- P ( + tn_ l)tn-2 - t1]6 [tl ... tn - I + 1] L n

= ~~~(1 + tl * * tn-1) (tl + 1) ... (trtl -I+ 1
=N(tj, *', tn- I)L(X)n,
say. If 6 is sufficiently small,
N .. dt, ... dt,-l < oo.
Therefore given 8 > 0 there is a region RE whose complement (in 0 ? tk < ?o

RE is bounded and bounded away from the coordinate hyperplanes and such that
f ... f dt, .. dtn-I < a
Since L is slowly varying there is an X2 so that whenever x > X2,

(tl, *-- tn- 1) c RE we have
{ <(1 + e)L(x)n,
L[(1 + tj)x -1] **L[(l + tn- 1)tn- 2 ... tlX - ]L(tl ... tn - 1 + l)x]
Consequently if X = max(X1, X2),
? dx JJ l/2L(x)n 1
h f Ix R J (I + t1 tn-1)(t; + 1) (tn
+ N+ /2 L(X) dx.
Set
(11) In = ___________ J dt, ... dtn- I
Then we find
fXN+l/2@Y.fMdtl,. dtn-l
U [(1+ )In + , fN+ /2L(x)n dx.
It is easy to see that
N -x
N dx ~IiM dt, ...dtn- 1
J ./2 X J
is bounded as N -+ oo, and so we conclude that the integral in (10) is at most
[(1U+ )In + 2E] JN L(X) dx
for large N. Similar reasoning shows that the integral is at least
[(1- eq)2In + 2,] N L() dx;
the main difference here is that R. must also be so chosen that the integral in
(11), extended over R' rather than 0 < tk < oo, is at least (1 - n)I. If we use
the easy facts
L(N)n = o t dx)
N L(x)(JNLx) d (x),
fJ X-?I dx = O(L(N)n f LXdx)
then we obtain, recalling (9),
N-1 N (X)__n_I_ L(x)

(12) N -A = L(X) dx + o(L(N)n 1f N dx)
the ~~N It ax
If in the integral (11) we set
t, = exp [2xj],
t2 = exp[2(X2 -X1
tn-i = exp [2(xn-1 - Xn-2)]
then we get
= f ... s sechx1 sech(x2- x1) ... sech(xn-2 - x1)dxl ... dXn_1,

-00 -00
which is the value at 0 of the n-fold convolution of sechx. This is the integral
over (- oo, oo) of the nth power of
00
feiyx sech x dx = - sech 2y.

- 002
Thus
In = 2 7 (sech -y dy - (rsech -y) dy

~'2 0 2] J 2/
and so (12) may be written
Ain J = (r L(x)sech jy) dyx
+ o(L(N)nif N (x) dx).
This was proved for n = 2,3, * and a somewhat simpler argument shows that
it holds also for n = 1. It follows that for any polynomial P without constant
term
N P(L(N) JAN,j) = f J (ZL(x) sech y)dy-

j =O L(N) 2
+ o(LNN)1
Since by Hilbert's inequality
11 HN || _ rmaXN.S 2(N-1)L(N')
we see that the quantit

L(N) 'L(x) sech I 7ry (1 < x < N). Given any bounded t interval, any a in
0 < a < 1, and any ? in 0 < E < a we can find polynomials P1, P2 without con-
stant term so that on this interval
P I(t) <X(ira, oo)(t) <:! P2 (t)
P2(t) Pl(t) ? e jt I + (a-e, a+e)(t/7r.

Therefore we have
w Z(xa.O)(L(1ANj) < { f X(a,oo() (LM)sech y dy
(13) + (28 + o(1))L(N)Y J L dx
+ f [sechl (a -) j) -sech-'(a+ ) k-
The left side of the inequality is the number of eigenvalues w

and the double integral on the right side, which we call shall I, represents the
area described in the statement of the theorem.
We show next that
(14) L(N)-1 f ! ) dx = 0(I).
If L'(x) = 0 for arbitrarily large x, then, by our main assumption, L is constant

for sufficiently large x and both the left side of (14) and I are asymptotically
nonzero constants times logN. Assume therefore that L' is ultimately positive
and set
p(N) = inf{x: L(x) > 2 L(N)}.
Then
2 rsN e 1 aL(N) dx
I=- sech -
7 1 L(x) x
2 sech
> - -1 2a- {N
- dx
7r + a p(N) X
Let A be so large that xL'(x) L(x)-1 is nonincreasing for x > A. Then if in

the integral fAx - L(x)dx we make the change of variable y = L(x) we
obtain
(N L(x) d L(N) L(x)
A X L(A) xL'(x)
?2IL(N) L(x)d
1/2L(N) xL'(x)
- 2f L(x) dx
p(N) X
? 2L(N)f JNdx
p(NL( ) X
Therefore jx- 'L(x)dx= O(IL(N)) and (14) is established.

For the remaining integral in (13) take e < a /2. Then by the monotonicity
of L,
(a-c) L(x) > 2
Now there is a constant A with the property that if a /2 < u < v then
sech-lu - sech lv < A(v - u)1/2
so the integrand in the last integral of (13) is at most
L(N)) /
But if (a - c) L(N) /L(x) _ 1 the integrand is zero. Therefore the integrand is at

most
A (-) < 2A (-)
and so the last term in (13) is at most
A'12 log N
with a different A. If we use this, the fact that log N = O(I), and (t4), then we
obtain from (13)
N-1
N: X(na, oo)(L(N) 2iNJ) < (1 + o(1))I.

j=o
The opposite inequality can be obtained similarly and so the theorem is proved.
We have the following corollary, which was conjectured by Wilf [8].
COROLLARY. For the Hilbert matrix
((j . k . X)-'), j* k = 0, N-
with oc 0, -1, -2, the number of eigenvalues which exceed na (O < a < 1)
is asymptotically
- log N sech-la.
Let us return to a general Hankel matrix HN in case (ii) and consider the behavior
of the norm (largest eigenvalue) and corresponding eigenvector. As was mentioned
in the discussion preceding the statement of Theorem 4.3, the integral operator on
L2(O, 1) with kernel
c-1
CN C[Nx] + [Ny]
converges strongly to the integral operator with kernel (x + y)-1, an operator

with norm 7r. It follows easily that
lim inf L(N) ' -r.

Conversely
1HN 1 <itr max (n +1)c.

L(N) n<2N-2 L(N)
If either (n + 1)cn is nondecreasing for all n > 0 or L(N) -+ oo the

approach m as N -0 oo and we shall have | | HN || -L(N). How
L(N) LOO < oo
but L is not nondecreasing for all n we may well have lim inf || HN || > rLOO.
(For example if Cn = n -1 for n > 0 then || HN c| c - X for all N.) Nevertheless
|| HN || does approach a finite limit in this case, namely the norm of the infinite
Hankel matrix H. In any case let us set
L unbounded,
limL(N) L bounded.
Then A _ T and
1I HN N- .
L(N) N~o
Let {aN,j} (0 _j< N) be the eigen

eigenvalue || HN ||; we assume all a
negative entries) and the aN,j are
1 N-1
N , aN- j
We write
fN(x) = aN,[Nx] 0 ? x < 1.

THEOREM 4.4. Let 4 be the smallest positive root of
*7r
siln 2= -A ,
so 0 < 4 _ . Then fN(x) converges to f(x) (a) uniformly in (3, 1) for each 3 > 0
and (b) weakly in the sense that f fN(x)g(x)dx f 1f(x)g(x)dxfor all continuo
g, where
f ( x _ _ _ _ _ _
f(x) 2- 2 (2 + 2) Jx 2 2 I2 2 xds.
Proof. The matrix equation

N-1
S Cj+kaN,k= || HN I|aN
is equivalent to the integral equation

1
(15) fCN IC[Nx]N+[Ny]fN(y) dy =(NCN) 1| HN j|fN(X
Now to prove the theorem it suffices to prove that every sequence {N'} has a
subsequence {N"} for which the desired convergence holds. Since the fN are
non-negative with integral one, we can find a function ,(x) nondecreasing
for 0 < x ? 1 and a subsequence {N"} such that
g(x)fN-(x) dx - g(x) dy(x)
for any g which is continuous on 0 < x ? 1. We shall henceforth write N for N".
Since for any x > 0, (x + y) 1 is continuous on O ? y ? 1 and
CN C[Nx]+[Ny] '(X + y)1

uniformly, we have
f C 1C[Nx]+[Ny]fN(y) dy)
= J'1 c~Tx+~- 2 fNxy) dy + 1 _fNCy) dy

Jo( CN X + y) t x + y
* r + du(y).
j0 x+y
Since (NCN)1 HN -|Ai it follows from this and (15) that
f(x) = lim fN(X)
exists for each x > 0 and that
(16) j -4 y dp(y) = if(x).
But since the fN are uniformly bounded on every interval (3, 1] (this follows from
(15), the uniform boundedness of CN 1CrNx]?rNy], and the fact that ||fN || = 1) ,u is
absolutely continuous on (0, 1] and ji'(x) =f(x) there. We can therefore write (16) as
p(? + ) - p(? + X f(y) dy-i f(x).
But f e L, implies that the integral on the left side is at most o(x'1) as x -+ 0, so
that if jl(O + ) - j(O) 7 0 we would have 2f(x) Ll, which is a contradiction.
Thus ,i is continuous at zero and
(17) f(y) dy = Af(x).

We know that fN(x) -f(x) weakly. Uniform co
the facts that fN(x) -f(x) boundedly and poin
converges to (x + y) - uniformly for x ? 6, 0 < y < 1.
It remains to prove that f(x) is what it is claimed to be. Since Sff(x)dx = 1
can write (17) as
J
To~(x +y y-
(xi:~ x ) =( 2Lf(x)-
f +)(y)dy d-i( X)
Let this identity, which holds for 0 < x < 1, be used to define f(x) for x > 1.
Then 2f(x) - x-1 is O(x-1) for small x and O(x-2) for large x. Let 0 < Rs < 1.
If we multiply both sides of the equation by xs and integrate with respect to x
over (0, oo) we obtain
fXCO (if (x) - -)xs dx = yf(y) dy I +-dx
=-- csc st | f(y)ysdy,
the interchange of the order of integration being justified by

~'1
4D(s) = J f(x)xs dx,
(s) = (if(x) - X)X dx.
Then 1D is analytic in MPs > 0 and bounded in RAs ? 6 > 0 and P is analytic is
Ms < 1 and bounded in RAs < 1 - 6 < 1. Moreover
(A + 7 csc 7s)FD(s)
S
=1 -T(s).
The next step in this complex-variable solution of (17) is what is essentially the
Weiner-Hopf factorization of A + m csc is. We have (recall the definition of 4)
sin (c2 +s)cos2 (-s)

I + csc 7rs = -- -
sin s cos -s
2 2 2 2 2 2
____ - + _
(2 2) 2 -2 (2 +2 -2)r 2 2 s)
so
i2 2 , D(s)
(1 -)(F + 2?)
The left side is analytic on R s and is bounded as s s oj ; the right side is

analytic on qs < 2, except for a simple pole at s = 0 with residue
z r( 2) (2 +2)
and is bounded as Is | s o . Since also 4>(s) -> 0 as s

analytic continuations of each other and equal
z~ ~~ 2( )r+2)
in their respective domains of definition. Therefore 1D is found and the invers
formula for the Mellin transform gives the desired expression for f(x).
Finally we consider case (iv), when the c1 tend to infinity more rapidly than
power of j. In [6] we determined the asymptotic behavior as N -+ oo of the larg
eigenvalue (normn) and corresponding eigenfunction of the integral equat
k(x + y)f(y)dy = 2f(x)
under the assumption that k increases rapidly and regularly. Quite analogously
one can treat the case of Hankel matrices. We shall only state the result here.
Let {aN,jI (0 _ ? < N) be the eigenvector of HN+j corresponding to the large
eigenvalue jj HN+1 jj, normalized so that
N
aN, j _ 0 a2N j=
j=O
THEOREM 4.5. Assume cj is ultimately increasing and
Cj _ I =o(M) j 0-o.
cj-Cj_1 Cj. 1-Cy- 2
Then as N -+ oo we have
c2
I|HN+1II C2N
C2N - C2N-2
N [ (C2N -C2N- 2) I 1/2 ] 0

j=O = C2N
REFERENCES
1. K. de Leeuw and W. Rudin, Extreme poinats and extremum problems in HI, Pacific J
Math. 8 (1958), 467-485.
2. F. Hausdorff, Der Wertvorrat einer Bilinearform, Math. Z. 3 (1919), 314-316.
3. Z. Nehari, On bounded bilinear forms, Ann. of Math. 65 (1957), 153-162.
4. F. Riesz and B. Sz.-Nagy, Functional analysis, Ungar, New York, 1955.
5. J. A. Shohat and J. D. Tamarkin, The problenm of moments, Mathematical Surveys,
Vol. I, Amer. Math. Soc., Providence, R. I., 1943.
6. H. Widom, Rapidly increasing kernels, Proc. Amer. Math. Soc. 14 (1963), 501-506.
7. , Asymptotic behavior of the eigenvalues of certain integral equations. II, Arch.
Rational Mech. Anal. 17 (1964), 215-229.
8. H. S. Wilf, On Dirichlet series and Toeplitz forms, J. Math. Aiial. Appl. 8 (1964), 45-51.
CORNELL UNIVERSITY,
ITHACA, NEW YORK

Hankel Matrices

Încărcat de

Informații document

Descriere originală:

Drepturi de autor

Formate disponibile

Partajați acest document

Partajați sau inserați document

Opțiuni de partajare

Vi se pare util acest document?

Este necorespunzător acest conținut?

Drepturi de autor:

Formate disponibile

Hankel Matrices

Încărcat de

Drepturi de autor:

Formate disponibile

Hankel Matrices

Author(s): Harold Widom

1. Introduction. A Hankel matrix is a matrix, finite or infinite, whose j, k entr

(1) c.=- 2 c (0)e i}0dO, j= l

|HJ| = inf|| bfJl

In ?4 we consider finite Hankel matrices. It is assumed that the cj are ultimately

W(H) = {(Ha, a): a |=1}

=, ECj+kakdj: Y Iaj 12=.

W,(H) = co {2(Ha, a) + -(Ha, a): || al| = 1}.

(Ha, a) = (Ha', a') + (Ha", a")

F(z) denote the corresp

Then for any complex number /3 7 0,

F(e)=4 1 _ CeiO +41)G(e) 1 _-.o

(= real part.) Thus if /3 is chosen so that

Rfl e -' j I G(e 0)jdO = 0

w(oc, K) = sup {r > 0: ra G K}

LEMMA 2.3. w)(xc, WT'(H)) = info e ,p | (O) + 40(-0)| .

Rdec[(Ha,ii) + (Ha, a)] = ?([ wCj+kakaj + z Cj+kakaj]

Since ||fII2 = 27|a || we conclude, using Lemma 2.1, that

(o(a, Ws(H)) = sup Q aWs(H)

= 2sup {LP j [(X b(0) + ib( -0)]f(0)2dO:fe H2 X f | 2 < i}

and by Lemma 2.2 this is

2sup{f [a (O)?+ 4( -O)]f(O)dO: feH1, lif III '}.

o(x, Ws(H)) < inf,, c | (O) + 1o(-0) l l t o

|| oo| < a)(a Ws(H)),

2 [a+ (O) + +i((- 0)]f(O)dO = f ifr(O)f(O)dO f e Hl.

This last identity shows that the Fourier coefficients of

-1(0) = f(O) - [a b(O) + ?5/(-0)]

-|1o(O) + 41(-.-) | = -2 O() + ~(-O) || X _w(c, Ws(H)).

This completes the proof of the lemma.

Given a pair of complex nu

{0: If(0)-z I < ?}

{0:If(0)-z <ae and jg(0)-z'| C<}

LEMMA 2.4. If f and g are bounded and F is a continuous function of two

{0:f(0) Pi,g(0) E QJ, and j F(f(0), g(0)) - A

{O :f (0) E p(k) g(O) E Q(k), and I F(f(O), g(O)) - j < 8}

{0:f (0) C p(k) g(O) E Q(k)}

THEOREM 2.1. The closure of W,(H) is

Proof. Let us write

By Lemma 2.3 it suffices to prove

If we apply the preceding lemma with

f(O) = 4(O), g(O) = k(-0), F(z, z') = a(a, Ez . ),

R(o(a, E0,(0),(_0))) = o(ax,E+).

be non-negative, it is necessary and sufficient that there exist a nondecreasing

Cj= xidp(x), j= 0,1,2,

where ,u is a nondecreasing function on [-1,1].

THEOREM 3.1. The following are equivalent:

(b) cj =O(j- 1) as j ->o

cj = xI d1(j) >(I - 6)1(p(l) - j(l - 6)).

c = j - x [(l) -(x)] dx ? Aj xj-1(l - x) dx = A/(j + 1)

cJC+kak,j = I | ajxj 2 dp(x)

f' du(x) < IIHII

p(1) - /(r) dr 1du(x) < ||H||

u(l1)- I(r) < 2||H|| (1-r),

A| A12 = (A Tx, A Tx) = (TAx, TAX)

(AxZ,y) = (TAx, Ty) = (A2x, y)

_R = sup {Th:ERA = TR _ AS}

where the . denote projections. At any point of continuity of A we also have

A = inf {f': EA = if _ AS}}.

Since MA is invariant under A, TWA is invariant under A. Moreover if - E T//,

(Ax,x) = (A2x,x) A(Ax,x) =

In addition TWIA + TWAk = T(J,A + &') = T(Y'), which is d

(2) (Ha, b) = z Cj+kakbJ =11 ( x ax)( bjxi) dlu(x).