Академический Документы
Профессиональный Документы
Культура Документы
Bruce C. Berndt
Fritz Gesztesy
Editors
L i b r a r y of Congress Cataloging-in-Publication D a t a
Continued fractions : from analytic number theory to constructive approximation : a volume in
honor of L. J. Lange : continued fractions, from analytic number theory t o constructive approxi-
mation, May 20-23, 1998, University of Missouri / Bruce C. Berndt, Fritz Gesztesy, editors.
p. cm. - (Contemporary mathematics, ISSN 0271-4132 ; v. 236)
Includes bibliographical references.
ISBN 0-8218-1200-9 (alk. paper)
1. Continued fractions-Congresses. I. Lange, L. J. (Leo Jerome), 1928- . 11. Berndt,
Bruce C., 1939- . 111. Gesztesy, Fritz, 1953- . IV. Series: Contemporary mathematics (Ameri-
can Mathematical Society) ; v. 236.
QA295.C663 1999
515l.243-dc21 99-30750
CIP
Leo Jerome Lange (Jerry) was the first of five children born to Leo and Clara
Lange on August 29, 1928, in New Rockford, North Dakota. For most of his
boyhood years, Jerry was raised on a farm about eleven miles from New Rockford.
This farm was homesteaded by his grandparents, and it was the birthplace of his
father. There he learned the value of hard work and experienced some of the
adversity brought on by the aftermath of the Great Depression and the dust bowl
era of the thirties.
Jerry's first eight years in school were spent in a one-room country school house.
He was fortunate to have several inspiring teachers in this school and recalls reading
most of the books (including the encyclopedias) in its limited library.
Jerry spent his four high school years, 1942-1946, at Sacred Heart Academy
(SHA) (now Schanley High School) in Fargo, North Dakota. To continue to attend
this premier high school (about 175 miles from the family home) he lived and
worked in a hospital in Fargo during his last three years at the school. To this day
he much appreciates the education he received from excellent dedicated teachers
at SHA. He has an especially fond memory of a nun who had the uncanny ability
of getting teen-age boys enthused about presenting proofs of propositions on the
blackboard.
Upon graduation from high school a t the rank of Salutatorian of his class,
Jerry now faced the problems of how t o finance his desired college education and
what to do about the fact that he was about to be caught up in the post World
War I1 military draft. The possible educational benefits of the W W I1 GI Bill
influenced him to enlist in the US Army on September 30, 1946 for an eighteen
month term. He served as a medic for more than a year in Berlin, Germany in the
3rd. Battalion, 16th Infantry Regiment, of the 1st. Division, and obtained the final
rank of Corporal.
Through the benefits of the GI Bill, by working in the summer, and with some
extra "travel money" from his parents, Jerry was able to attend Regis College (now
Regis University) in Denver, Colorado. He started his training there in the fall of
1948 and eventually decided to major in Mathematics and minor in Physics and
Philosophy. He is very appreciative of the challenges that were presented t o him and
the attitudes that were fostered in him by his professors at Regis. Jerry received
the Most Outstanding Senior Award from Regis, and graduated with a BS degree
in the spring of 1952.
Jerry had been encouraged by his mathematics teachers at Regis to pursue an
advanced degree in mathematics, and in August of 1952 he went to the University
of Colorado-Boulder (CU) to seek admission as a graduate student in mathematics.
Burton W. Jones, who was then Chair of the Mathematics Department a t CU,
xii
...
L. .I. L A N G E O N HIS 70TH BIRTHDAY Xlll
accepted him immediately. Jerry began his teaching career in the winter semester
of 1953 when he received a Teaching Assistantship in Mathematics.
In June of 1955, he married Geraldine Ryan who was teaching music in the
Denver public school system. They lived in Denver and Jerry commuted to Boulder
daily to continue his study and teaching at CU. By the end of the summer of 1955
he had completed his Masters Thesis on "Some Problems in Interpolation" and
all of the requirements for the MA degree. His Masters Thesis advisor was Kurt
A. Hirsch of the University of London, and the second reader for the work was S.
Chowla.
In 1956 Geraldine and Jerry had started a family, so she chose to relinquish her
teaching position in Denver. Jerry accepted a classified position as a mathematician
at the National Bureau of Standards, Boulder Laboratories (NBS) in July of 1956.
His position at NBS was located in the Low Frequency and Very Low Frequency
Section of the Radio Communication and Systems Division. For the next two
years he continued to take graduate courses at CU while working full-time at NBS,
making up the time he spent in class by working on weekends. Jerry did most of
the analyses and prepared most of the technical materials that were used by the
US delegation to the International Technical Discussions on Detection of Nuclear
Tests held in Geneva, Switzerland in the summer of 1958. For his work at NBS,
Jerry received two merit awards from the United States Department of Commerce,
one in 1959 and one in 1960.
In the fall of 1958, through his Ph.D. advisor Wolfgang J . Thron, Jerry was
awarded a two-year AFOSR grant for graduate research. This allowed him t o
convert his position at NBS to half-time, so he could spend his mornings at CU
attending classes and doing research for his doctoral thesis. Jerry was the first
of Thron's many Ph.D. students, and he was awarded the Ph.D. degree from CU
in the summer of 1960. His doctoral thesis consists of three parts and is entitled
"Divergence, Convergence, and Speed of Convergence of Continued Fractions 1 +
K(a,/l)". The Convergence part appears in his joint 1960 publication with Thron
in Mathematische Zeitschrifl, and t,he other two parts have played significant roles
in some of Jerry's later publications. Upon receiving t'he Ph.D., he was offered a
large increase in civil service rank if he would continue his work at NBS. However,
he chose to resign at NBS in August, 1960 so that he could accept an academic
position at the University of Missouri-Columbia (MU).
Jerry joined the faculty of the Department of Mathematics at MU on September
1, 1960. He has spent his entire academic career at this institution. He served as
Chair of the Department during the years 1988-91 and during the summers of 1966,
1968, and 1969. He also served in the position of Associate Chair during the years
1968-69, 1977-78, and 1978-79. His teaching record shows that he taught in the
neighborhood of one hundred and ninety classes at MU, most of which were at the
calculus level and beyond. He was a pioneer in the use of the computer as a teaching
aid in calculus, and for three decades he was the principal creator, teacher, and
promoter of complex analysis courses in the Department. Jerry gave unselfishly of
his time t o numerous important Departmental and Campus personnel, academic,
and policy committees. He is especially proud of his role in the Departmental
Planning Committee that had much to do with charting the path of the Department
in the last decade.
xiv I. .J L A N G E O N HIS 70TH BIRTHDAY
Jerry's principal field of research has been the analytic theory of continued
fractions. He is an internationally recognized expert on the convergence of con-
tinued fractions and their applications to function representation in the complex
domain. In each of his papers on continued fractions he has solved one or more,
usually long-standing, problems in the field. Space limitations prohibit a discussion
of Jerry's research in detail here. but the majority of his results are fundamental,
have wide-reaching significance, and are cited often by other researchers. His Uni-
form Twin Liina~onTheorem for continued fractions K ( a , / l ) . which he proved in
his 1966 paper in the Illznozs Journal of Mathematzcs, has turned out t o be one of
the most significant results in convergence region theory. Recently. in the Campznas
Proceedzngs. he settled a twenty-eight-year-old conjecture of Jones and Thron by
proving that almost all twin convergence regions for K ( a , / l ) generated by either
disk-disk, disk-halfplane, or disk-complement of disk mappings are embeddable in
those of his L i m a ~ o nTheorem. In two papers in the early 1980's, Jerry developed
the theory of 6-fractions. His &fraction work solved a number of problems dealing
with the association of continued fractions and power series for analytic functions.
Others have used his 6-fractions to approximate solutions of Riccati differential
equations, and Jerry applied them to zero location and stability problems in one
of his 1986 papers. I11 his 1986 work with Kalton, many outstanding problems in
the area of equimodular limit-periodic continued fractions were solved. More than
ten years later. some of this work was used to disprove an assertion of Ramanujan
and now appears in Berndt's Ramanujan's Notebooks, Part V. The Worpitzky and
Transcendental strips in his 1994 paper on strip convergence regions for continued
fractions amount to the first known best convergence regions that are distortions
of the famous parabolic regions and have inspired others to seek generalizations of
them. In his 1995 paper in Constructzve Approxzmatzon, Jerry completely settles
the Oval Theorem for continued fractions and proves an earlier result of others
false by actually showing that these ovals are embeddable in the parabolas of the
Uniform Parabola Theorem. His Uniform Twin Parabola Theorem of 1994 in the
Journal of Mathematzcal Analyszs and Applzcatzons generalizes work of Thron fifty
years earlier in several important ways. Uniformity and speed of convergence are
obtained, and only the even elements need to be bounded. His work on Van Vleck
fractions K ( l / b , , ) in this publication should prove to be quite valuable in future
research on complex continued fractions of this type. Jerry has participated by in-
vitation in many research workshops and conferences in his field, including Boulder
and Redstone. Colorado. Pitlochry and Aviemore, Scotland, Loen and Trondheim,
Norway. and Campinas. Brazil.
Geraldine and Jerry have four children (two sons and two daughters) and nine
grandchildren. Jerry retired on September 1 , 1998, with the status of Professor
Emeritus after thirty-eight years of service at MU. He is looking forward t o enjoying
mathematics, his family, and his other interests for a long time to come.
We wish the very best to Jerry in his future personal and professional endeavors.
Bruce C. Berndt
Fritz Gesztesy
Publications of L. J. Lange
[I] Ionospheric reflection coefficients at VLF from sferics measurements, Geofisica Pura Applicata
38 (1957) 147-153 (with A. G. Jean and J. R. Wait).
[2] Some characteristics of VLF propagation using atmospheric waveforms, Recent Advances in
Atmospheric Electrzcity, pp. 609-617, Permagon Press, London-New York-Paris-Los Angeles,
1959 (with W. L. Taylor).
(31 A two-parameter family of best twin convergence regions for continued fractions, Math. Z. 73
(1960) 295-311 (with W. J. Thron).
[4] On a family of twin convergence regions for continued fractions, Illinois J. Math. 10 (1966),
97-108.
[5] A simple irrationality proof for nth roots of positive integers, Math. Mag. 42 (1969) 242-244.
[6] &fraction expansions of analytic functions, Analytic Theory of Continued Fractions (Loen,
1981), pp. 152-175, Lecture Notes in Math. 932,Springer, Berlin-New York, 1982.
[7] &fraction expansions of analytic functions, SIAM J. Math. Anal. 14 (1983) 323-368.
[8] Equimodular limit periodic continued fractions, Analytic Theory of Contznued Fractions, I I
(Pitlochry/Aviemore, 1985), pp. 159-219, Lecture Notes In Math. 1199,Springer, Berlin-New
York, 1986 (with N. J. Kalton).
[9] Continued fraction applications to zero location, Analytic Theory of Continued Fractions, I1
(Pitlochry/Aviemore, 1985), pp. 220-262, Lecture Notes In Math. 1199,Springer, Berlin-New
York, 1986.
[lo] Continued fraction representations for functions related to the gamma function, Continued
Fractions and Orthogonal Functions (Loen, 1992), pp. 233-279, Lecture Notes in Pure and
Appl. Math. 154,Dekker, New York, 1994.
[ll] Strip convergence regions for continued fractions, Contznued Fractions and Orthogonal Func-
tions (Loen, 1992), pp. 211-231, Lecture Notes in Pure and Appl. Math. 154,Dekker, New
York, 1994.
[12] A uniform twin parabola convergence theorem for continued fractions, J. Math. Anal. Appl.
188 (1994) 985-998.
[13] Convergence region inclusion theorems for continued fractions K(a,/l), Constr. Approx. 11
(1995) 321-329.
1141 Uniformity and speed of convergence of complex continued fractions K(a,,/l), Orthogonal
Functions, Moment Theory, and Continued Fractions (Campinas, 1996), pp. 275-326, Lec-
ture Notes in Pure and Appl. Math. 199,Dekker, New York, 1998.
[15] Convergence regions with bounded convex complements for continued fractions K(l/bn), J.
Comp. Appl. Math. 105 (1999),pp. 355-366.
[16] A generalization of Van Vleck's theorem and more on complex continued fractions, Continued
Fractions: From Analytic Number Theory to Constructive Approximation, Contemporary
Mathematics 236 (1999),pp. 179-192.
[17] An elegant continued fraction for T , Amer. Math. Monthly 106 (1999), pp. 456-458.
A List of Participants
xvi
A LIST OF PARTICIPANI'S xvii
Richard Askey
A B S ~ I I A ~Recurrence
T. relations are very important when one studies orthog-
onal polynomials. Examples are given. including some where the recurrence
relations relate t o continued fractions.
1. Introduction
Gabor Szego wrote the following in the preface of his great book "Orthogonal
Polynomials" [34. p. v]. .
"The origins of the subject [orthogonal polynomials] are t o be
fourld in the investigation of a certain type of continued fractions.
bearing the name of Stieltjes, . . . Despite the close relationship
between continued fractions and the problem of moments. and
notwithstanding recent important advances in this latter subject.
continued fractions have been gradually abandoned as a starting
point for the theory of orthogonal polynomials."
\Vhen Szego wrote his book sixty years ago. it was not clear what role was
left for continued fractions. There are two dominant themes in Szego's book: the
classical orthogonal polynomials of Jacobi. Laguerre and Hermite, and the general
theory of polynomials orthogonal on a bounded interval, say [-I. 11. whose weight
function is positive in the interior of this interval and does not vanish too rapidly
as the variable approaches the endpoints of this interval. The classical orthogonal
polynomials had been studied for about 150 years when Szego wrote [MI. with the
work of Legendre and Laplace on spherical harmonics being one of the starting
places. Gauss [14] developed Gaussian quadrature from continued fractions. How-
ever. Jacobi's version [18] of this using orthogonality was much easier to motivate.
and t o extend t o other measures.
Szego's theory of orthogonal polynomials on an interval. and the corresponding
theory on the unit circle. were developed independently from continued fractions.
In the last sixty years there have been many new results obtained on orthogonal
polynomials. In one of these areas, continued fractions play a major role. There
are also very interesting combiriatorial connections with continued fractions and
orthogonal polynornials. See Viennot [39].
which satisfies
which satisfies
(2.2a) LE ( x ) L ~ , ( x ) ~ ( ~ e=- ~0,d xm # n, a > -1,
or
d
H,, (J) = 2n H,, 1 ( L ).
- -
dr
The usual notation will be used for hypergeometric series:
with
satisfies
(2.10) x
r=O
2
This is mentioned by Szego [34, Section 2.81. At the end of Chapter 2. Szego
[34]mentions an extension of this result found by Markoff in his thesis [ 2 1 ] . The
measure is qT, r = 0 , 1 , . . . , N.
In a later paper [38].Tchebychef extended the polynomials defined in ( 2 . 9 ) to
a more general set:
when rn, n = 0 , 1 , . . . . N .
These poll rloinials are called Hahn polynomials. Tcl1ebycht.f'~paper contain-
ing these polynomials was not noticed until long aftcr Hahn rediscovered these
polynomials, and found an extension of them.
Hahn was led t o these polynornials and an extension of them which contains
the polynomial found by Markoff by consideration of the following problem. Hahn
1151 had proven the theorem mentioned earlier about {p,, ( x ) ) and {p:,+, ( x ) ) being
orthogonal only for Jacobi, Laguerre and Hermite polynomials. He looked for a
similar theorem when the derivative was replaced by the operator
Hahn [ 1 6 ] showed that the general operator in (2.13) could be obtained from
its special case when h = 0 . For this case,
Here q is fixed, usually real with 1 < q < 1. See Section 3 for a definition of
the series in (2.15).
The polynoinials Hahn found have other properties in common. They satisf?
second order q-difference equations in x which are analogous to the second order
differential cquations satisfied by Jacobi. Laguerrr and Hermite polynornials. The
4 RICHARD ASKEY
Hahn only worked out the orthogonality for one of his classes of polynomials.
Ortliogonaiity relations have been worked out for all of the polynomials he found,
and for a larger class which arise from a more general operator than (2.13). In
the next section some of the different methods used to show orthogonality are
summarized.
where
and the series terminates because one value of the n,'s is y " . See 1131 for more on
hasic hypergeometric series.
For these balanced 4p:3's.t,here are two cases where thcre is ~rt~hogonality wit,h
respect t,o a positive measure. In one case the measure has an absolutely corltirll~ous
part. arid may have a finite nunlber of mass points. In the other case. whicli extends
the polyiiomials of Chebychev mentioned above, t,he orthogonality is purely discrete.
on a finite sct of points.
The second case gives the polynomials
P,, ( x ) ~ , (, ,x ) M x ) = hllbl,,,,,
and
Iteration gives
with
(3.18)
(3.20)
Pl, ( x ) =
satisfies
. h
CONTINUED FRACTIONS AND ORTHOGONAL POLYNOMIALS
'pk(x)p,(x)doc(x) = 0, k = 0 , 1 , .. . ,n - 1
and nz
k=O
Szego [[34],Section 2.21 states this in a more general form. He also wrote that
(3.21) was "not suitable in general for derivation of properties of the polynomials
in question." This is true, but only if the words "in general" are interpreted cor-
rectly. I used t o say that this representation was not useful. One of my Ph.D.
students, James Wilson, proved me wrong. He showed that for the weight function
for either R,(X(x)) or V,(x), it is possible to choose pk(x) SO that the moments
can be computed explicitly, and then showed how to take linear combinations of
the rows to explicitly evaluate the determinants. Wilson used recurence relations
to obtain explicit formulas to evaluate these determinants, and even a more general
determinant. The most general one he evaluated gave a new set of biorthogonal
rational functions. See [40].
The integral of the measure for p, (x) is a generalization of the beta integral.
It is
1 'h(x, l ) h ( x ,ql/')h(s, - l ) h ( x , -ql/')dz
h(x, a ) h ( z , b)h(x,c)h(x,d ) d =
(3.22)
po (x) a constant.
8 RICHARD ASKEY
and & ( x ) is the nleaslire used in the orthogonality relation of p,, (x). hlarkoff's
theorem requires that n and b are finite. and that is all that will be needed in the
examples below. There is a variant due to Stieltjes which allows b t o be infinite
when other conditions are added.
hlarkoff proved that
RO-
lim - (-x)= p(x)
"-"S,,( S )
when s is in the complex plaric cut on the interval n < x < b. The function E'(x)
is dcfined hj-
Smgo's proof of (4.6) uses the theorcrn of Stieltjes that Gaussian quadrature
couvergcs t o the integral of the fiinction when the functioii is continuous. See [36]
for detail\.
The first non-classicnl example I know was done by Sherman. and involves
associated Legtndrc fiinctions.
Legcndrr pol? rloriiials 5atisfy the three term recurrence relation
The associated Legelidre polvnornials arise whrn all the n's in the coefficients
in this recurrence relation are shifted by the same arriount. Sherinan [33] treated
the most important case, wlien tllc shift is by oiie. Thc general case is as follows.
P _ 1 ( x ; c )= 0 , f i ( x : c ) = 1.
IVhrn c = 1 . Sherman showed that
Shcrinan did rriort, grneral work. trcating some associatrd Jacohi polynomial
when the shift is bj- 1. Sccl Nevai [25] for gc~nr~ral
results with integer shifts.
CONTINUED FRACTIONS AND ORTHOGONAL POLYNOMIALS 9
The general case was treated twice. Once by Pollaczek [28] as part of a more
general set of orthogonal polynomials wliich will be inentiorled next. Tlw otlwr
time was by Barrucand and Dickerison [8]. The form of tlw weight fun( tion given
in these two cases is different. Barrucand and Dickimon have
as the weight function. The function PC(.r) and Q , (.r) are Legendrc~fiinction5 of
the first and second kind. T h r polynornials when c = 112 and c = 312 arose In
finding polynomials orthogonal with respect t o the ineasures
( 1 - y y l L ( 1 - k.2 .L -1IL
J 1
work done by Revs [30].
111
In a remarkablr seric.s of short notes [27]. [28],and others. and a monograph
[29]. Felix Pollaczck described a vorj interestmg 5ct of orthogonal polynomials
which he discovered. One set satisfies the following recurrence relation.
P - ] ( r ) = O . P,,(r) = 1.
Inltlally he found these pol\ nomi& u hell = f did r = 0 G a h r %ego
[35] appreciated the import ante of t he111 and cxtcndrd-~o1l;iczrk'srcwdt to iiic lude
A. Then Pollaczek added the c. Thtl assoclatrd Legciltlre pol\ nolnldls arise mlicn
a = b = O a r l d X = L.
Il'hen n> 5~ +
lbl. > + >
c > 0 nncl ( 0 or whcn n 2 lbl. 2X c 1. c > -1. the
orthogonality rtllatioll is 11 it11 respwt t o an dbsolutclj continuous rricasurc given for
l < J <lby
wit 11
(4.14)
Thc case c = O is a littlc easicr t o handle. and contains the lnost iinportant
c~xainple.The weight function satisfic.~
r ]
\I'hcn the integral in (3.15) is finitc. u 3 ( s )# O for -1 < .x. < 1. ant1 u , ( r ) is
smoot,h enough. Szcg6 has shown that the polynonliitls orthogonal with rvspect to
t,his weight function look a lot like .lacobi polynoinials. SIT [34]. Tlirse polylmrnials
whrn taken to be orthonormal grow at most like a poxwr of 11 whcn .r. = 1. I'ollaczck
polyrlomials havc more rapid growth when either n # 0 or 0 # 0. Thesc werc the
first polynoiriials t o b r fouiid explicitly m-hose weight function is not in the Szegii
class.
Szegij gave the problcm of finding out mar(' about Polla,czek's polynomials to
a Ph.D. student. A. Novikoff. Novikoff's thrsis [26] was not p~lblished.but Szegii
included a sumniary of the mait1 results in an appeidix t,o t,lie second edit ioii of
[34]. Essentiall>-nothing was tlo~iowith t,liesc, polynomials for many y ~ a r s .
Thc first rcxlated work was dolic, by Srlliah [32]. this was not plhlislled. and
the connect,ion with I'ollaczek's work was not rcdizcd. St,lliali was cxt.ending ~vork
10 RICHARD ASKEY
of Karlin and 1LlcGregor. [19], who had come across a limiting case of Pollaczek's
polynomials but with different conditions on the parameters. The Karlin and Mc-
Gregor work dealt with a set of orthogonal polynomials whose weight function is
supported on a discrete set of points with 0 as the only limit point. See [19]. Selliah
found the Stieltjes transform of the measure of orthogonality, and from this found
all of the discrete mass points. However, his lack of knowledge of Pollaczek's work
led him t o write: "It seems very unlikely that this [the Stieltjes transform] will in
general give us on inversion, a simple function for the distribution."
Askey and Isinail knew of Pollaczek's work, but not that of Selliah. when they
studied the special case of (4.12) when b = c = 0. Here is a sample of the results in
.141. A different normalization was used t o make it easier t o consider the limiting
>
G 1( 2 ) = 0. G o ( x )= 1.
There is orthogonality with respect t o a positive measure when a b > 0 and +
a>O,orwhena<-landa+b+l<O.
L$'hen a > 1. b > 0 , the orthogonality relation is contained in the ones Pollaczek
[27] and Szego [35] found. When 0 < a < 1, b > 0, in addition t o an absolutely
continuous part of the measure used for orthogonality, there are infinitely many
discrete mass points outside the interval [-d, dl, d = 2 & / ( a +
1 ) . which have d
and -d as limit points.
To understand the power of Markoff's theorem combined with other 19~''cen-
tury parts of analysis, try to evaluate the integral and sum in the following orthog-
onality relation. I do not know how t o do this directly, or know if the integral and
sum can be evaluated separately.
where
b
r ( % +i-
b(l - a)
2a
cot 8 ) 12,
and
CONTINUED FRACTIONS AND ORTHOGONAL POLYNOMIALS 11
After this paper was written, a number of other examples were worked out.
The parameter b in (4.12) was reintroduced.
A different type of generalization of Legendre polynomials had been found.
first by L. J. Rogers [31],who was unaware he had orthogonal polynomials, but
who found many other remarkable formulas for his polynomials. His polynomials,
which are now called continuous q-ultraspherical polynomials. satisfy the recurrence
relation
(4.20)
2 4 1 - Pq'"C,,(x:P I 9) = (1 - qn+')C,+1(x:P I 9 ) + (1- P q
2 71-1
) G - l ( x ; P I 91,
C-l(.GP 19) = 0, C o ( x ; P 19) = 1.
Those polynomials as orthogonal polynomials were rediscovered by Feldheim
[12] and Lanzewizky [20]. They were also rediscovered by Allaway [3]. Allaway
was more careful in giving all of the cases when orthogonal polynomials arise, and
found a new set of polynomials which are a limiting case of those of Rogers. There
are two limiting cases, one of which is found as follows.
Let W A be a kt" root of unity, say W L = exp(2nilk). Set P = s " ' + ' u J ~ , q = SUL
in (4.20). Divide by 1 - sw;" and let s -+ 1. The resulting polynomials B;(x; k)
satisfy
>
B $ ( x ; k ) = 1. BF(x;k) = 22 if k 2.
See [2] for the orthogonality of these polynomials.
Polynomials whose recurrence relation coefficients are constant except on an
arithmetic progression, are frequently called sieved orthogonal polynomials.
Both q-extensions and sieved extensions of the results in Theorem 4.1 are con-
tained in the following papers, [I],[9]. Also see [17] for related results.
Another paper which builds on work of Pollaczek is [7]. Here, associated
Laguerre and Hermite polynomials are treated as limits of other polynomials of
Pollaczek, his extensions to associated polynomials of polynomials first found by
Meixner. For associated Hermite polynomials defined by
2xHn (x; c) = H,,+I (2; c ) + 2(n + c)Hn-, (x: c),
(4.22)
H-, (2: c) = 0, Ho(x: c) = 1,
t,he orthogonality is
with
Various Pad4 type approximations havc been found using these poly~iomials.
and int,eresting Stieltjes transforms follow from t hcse results. See [7]. [41], [42] for
examples.
Fillall\: some itnpressive contirmcd fractioris stated bj- Ranianujari and first
proved and extended bj- IVatson werc reconsiderrd by Alasson [23], [24]. He was
able to extend t,he most general t,o an infinite c~nt~ilrnedfraction. while Ramanujan
and II;;rtson old\- gave their rcsult,s for ternlinating continued fract,ions. hlassori
also found a coinpanion continued fract,ion which had been rnissed. If one likes
beautiful fornullas, and t,hcse continutd fractions are beautiful. then it is useful to
have a firm knowledge of continued fractions.
References
[l] N. Al-Salaln and hI. Isrnail. O n szct~rdorthogonal p o l y n o m 7 d s . V I I I . S7evcd cassoc7ated Pol-
lnczek polynom7als. J . Approx. Theory 68 (1992). 306 321.
[2] \.V. Al-Salaln. M'. R . Allaway and R. Askcy, S z c ~ ~ eultmphcrzcal d pol:ynomzuls, Trans. A ~ n c r .
hlatlr. Soc. 284 (1981). 39 55.
[3] W. R . Allaway. T h e zdentzfieatzon of u class of orthogonal polynom7al.s. Pl1.D. Thesis. Urli-
vcrsity of Alberta. Canada. 197".
[4] R. Askey and hl. Ismail. Recurrence ,relatzon.s, c o n t m u e d frnr.tzons a n d orthogonal polynomz-
als, hlelnoirs Anic'r. Math. Soc. 300, Proviticr~ce.R I . 199.1.
[5] R . Askey arid J . IVilson, A set of orth,oyonnl polynomials t h a t g m e r a l z z e t h e R o c a h c o e f l c l e n t s
07;. 6-3 syrnbols. SIAhI J . hlath. Anal. 10 (1979). 1008 1016.
DEPAWRIESTOF ~ I X ~ I I E ~ I A TUNI\.SRSITY
ICS. OF \VIS?ONSIN ~ I . \ I ) I S O N1.I ~ 1 1 1 ~ 0 \VI
4 . 53706
E-mazl address: askeymmath. wisc . edu
Contemporary M a t h e m a t ~ c s
Volurrw 236. IWS
Abstract
1. Introduction
Between the years 1911 and 1919 Ramanujan submitted a total of 58 problems,
several with multiple parts, to the Journal of the Indian Mathematical Society.
For t,he first five. the spelling R a m a n u j a m was used. Several of the problems arc
elementary and can be &tacked with a background of only high school mathematics.
Ebr others, significant amounts of hard analysis are necessary to effect solutioiis.
and a few problems have riot been completely solved. Every problerri is either
interesting or curious in some way. All 58 problems can be found in Rarnarnljan's
Collected Papers [172, pp. 322-3341, As is customary in problems sections of
journals, edit,ors norrrlally prefer to publish solut,ions other than those given by
t,he proposer. However, if no one other t,llan the proposer has solved the problem,
or if the proposer's so1ut)ion is ~art~icularly elegant,. then the proposer's solution is
published. This was likely the practice followed by the editors of the Journal of t h e
Indian Mathernatical Societ?y, but naturally thp editors of Rarnanujan's Collected
Papers chose different criteria; only those printed solutions by Ramanujan were
reproduced in his Collected Papers.
The publication of the Collected Papers in 1927 brought Ramanujan's problems
to a wider matherriatical audience. Several problems have become quite farnous
and have attracted the attention of many mathematicians. Some problems have
spawned a p1et)hora of papers, nlany cont,ainirig generalizations or analogues. It
1991 Mathematics Subject Classzjication. Primary 11-03, 40-03; Secondary llD25, llD99,
40A15.
@ 199'3 A m e r i c a n Mrrthematlcal Society
15
16 BRUCE C . BERNDT, YOUN-SEO CHOI, AND SOON-YI KANG
2. Solutions of Equations
where X I , x ~. .,. , x,, and yl. y 2 , . . . , y, are 2n unknowns. ingeniously solved by Ra-
manujan in his third published paper [ 1 6 4 ] ,[172, pp. 18-19]. Implicit assump-
tions were made in Ramanujan's solution, and thus it should be emphasized that
(2.1) is not always solvable. For a sketch of Ramanujan's solution. see a paper by
Berndt and S. Bhargava [ 2 5 ] . Another derivation of the general solution for (2.1)
was found by hI. T. Naraniengar [148].The more general system (2.1) is also found
on page 338 in Rainanujan's second notebook [ 1 7 1 ] . We quote the discussion of
(2.1) from Berndt's book 123, p. 301.
"It is easy to see that the system (2.1) is equivalent to the single equation
.
must have a solution ul , uz, . . . u,,+l such that the n-ic form
recorded by Ramanujan on page 338 in his second notebook and solved by Berndt in
[23, pp. 27-29]. The solutions comprise 25 values for both x and y. As pointed out
in [23, pp. 28-29]. Ramariujaa's published solution. which is reproduced in [172,
pp. 322-3231. contains some mistakes. \Ve very briefly describe the solutions. Let
s=o+fi+? and ,y=a/?+py+ya,
QUESTION
507 (JIMS 5, P. 240; 6, PP. 74-77). Solve completely
I11 parts (a)-(c), the signs under the outward-most radical sign have period
three; they are, respectively, -, +, -; -, +,
-; -, +. +.
Rarnamijan's published solution in the .Journal of the Indian Mathematical
Society is correct,. However, there are four sign errors in the solution published in his
Collected Papers [172, pp. 327-3291, which we now relate. The equations in (2.2)
easily imply that .r ~at~isfiesa polynomial equation of degree 8. This polynomial
factors over ( ) ( d m )int,o a quadratic polynomial c2 T - u and t,wo cubic -
polynoinials, which are given in the Collected Papers near the end of Rarnanujan's
solution on page 328 and just before the verifications of the three exairlples (a)-
(c). Each of t,hese two cubic: polynomials has two sign errors. The polynomials are
corrr:ctly given in Ramanujan's original solut,ion and in Berndt,'~book [23, pp. 10,
11, eqs. (4.2), (4.3)].
The formulation of Question 507 indicates t,hat the three examples can be
deduced from the solutiorls of (2.2). However, in his published solution, Ramanujan
did not do this but established each ident,it,y ad hoc. In his so1ut)ionto Question 507.
M. B. Rao [176] derived each of the examples, (a)-(c), from t,he general solut,ion of
(2.2).
T H E PROBLEMS SUBMITTED BY RAMANUJAN 19
However. we must be careful in t,aking these square roots, for there are three differ-
ent square roots t o be taken and t,here are two choices for the sign of the square root
in each case. The 2" different sequences of nested radicals correspond t o the eight
root,s of thc octic polynomial arising from (2.2). Of course, one must determine the
values of a for which the eight infinite sequences convergc. In [23, pp. 14-16].
it was shown that the infinite sequences converge for a > 2. which is not tlle best
possible result. For example, it was indicated in [23, p. 151 t,hat the sequence
wit,ll signs +, -,+ of period 3, likely converges at least for a > 1.9408. In our
discussion of Quest,iorl 289. we will return to the question of the convergence of
infinite sequences of nested radicals. .
For a part,icular value of a , one can numerically check which infinite sequence
of nested radicals corresponds to a given root. In general, for the t,wo infinite
sequences of nested radicals arising from the two roots of the quadratic polynomial.
the identification is easy. However. for the remaining six roots, the problem is more
difficult. On pages 305-306 in his second not,ebook. Rarnanujan made thcse general
identifications [23, p. 17, Entry 51. In [23, p. 181, asymptotic expansions of the
six roots and the six sequences of nested radicals. as a + x, were established in
order to obtain the desired rnatchings.
There is also a very brief discussion of the system (2.2) in I.'. Cajori's book [53,
pp. 196-1971,
722 (JIILIS 8 ,
QUESTION P. 240). Solve completely
then
and that, 2f
20 BRUCE C. BERNDT, YOUN-SEO CHOI, AND SOON-YI KANG
then
Question 722 is obviously an analogue of Question 507. The first complete solu-
tion to this problem was given by XI. B. Rao [176] in 1925. He also solved Question
507. examined the analogue for five equations of the same type as Questions 507
and 722, and presented several examples. The second solution was given in 1929 by
G. N . Watson [210].who also derived Ramanujan's two examples. A considerably
shorter solution was found by A. Salam [I831 in 1943. Question 722 can also be
found in Ramanujan's third notebook [171, vol. 2, p. 3671, and a fourth solution
is given in Berndt's book [23, pp. 42-47, Entry 32, Corollary].
666 (,JIILIS 7 , P . 120; 8 ,
QUESTION P. 3 1 ) . Solve i n posztive rational numbers
N (JIILIS 3 ,
Q U E S T I O289 P. 90: 4 , P. 2 2 6 ) . find the value of
(ii)
The values of (i) and (ii) are 3 and 4, respectively. Rarnanujan's solutions in
volume 4 [172, p. 3231 are not completely rigorous. A note by T. Vijajaraghavan
at the end of Appendix 1 in Rarnanujan's Collected Papers [172, p. 3481 justifies
Rarnanujan's formal procedure. This note was considerably amplified in a letter
from Vijayaraghavan t o B. hf. \$'ilson on January 4. 1928 [36, pp. 275-2781. If
<
a , > 0 , l J < x, then Vijayaraghavan proved that a sufficient condition for the
convergence of the sequence
THE PROBLEMS SUBMITTED BY RAMANUJAN
is that
- loga,,
lirn
rr-x
-
2"
< X.
Vijayaraghavan's criterion for the convergence of infinite ~iestedradicals is a prob-
lern in P61~-aand Szegij's book [156, Prob. 162, pp. 37, 2141. A . Herschfeld [96]
also proved Vijayaragliavan's criterion. In a 1at)erquestion s u b i n i t t d t,o the Jour-
nal of the Indian Mathenmtzcul Society, Vijayaragllavan [204] claimed a stronger
theorem for t,lle convergence of t,,, which he said was best possible. However, the
problenl is flawed, and no correction or solution was evident,ly publishrd. However.
Problem 163 in P61ya and Szego's book [156, pp. 37, 2141 is likely thc criterion
that Vijayaraghavan had in mind. A sufficient condition for c:onvergencc wherl thtl
sequence {a,) is complex has been established by G. Schuskr arid \V. J. Tllron
[189]. Convergence criteria for certain iiifiriit,e nested radicals of pth roots liaw
been given by J . LI. Borwein and G. dc Barra [42].
Part (i) appeared as a problem in the \Villiam Lowell Putiianl coinpt'tition in
1966 [141]. The values of (i) and (ii) appear as examples for a gcncral tlicwrcm of
Rarrianujan on ncsted radicals in Section 4 of Chapter 12 in his second notebook
[21, p. 108, Entry 41. Further references for Qurstioii 289 and relatcd work arc
found in Section 4, and addit,ional r,xamples of infinite ~ i c s t t dradicals appear in
Entries 5 and 6 in that same chapter [21, pp. 109-1121.
Part (i) appears on the last page of Rarnanujan's second notebook [171, p.
3561. The proofs of (i) and dnother co~npanionresult by Berndt 111 [23, p. 391 use
a general result of Rarnanujan [23, p. 22, Entry 101 on the sunl of t h cube ~ moth
of the three roots of d L U ~ Kpolyno~nial.The proofs of (i) and (ii) b\ N. Salikara
Aiyar in volume 6 of the Journal of the I n d ~ a nMafh~rnntzcalSo( rely arc. similar
QUESTION 682 (JIhIS 7, P. 160; 10. P. 325). Show how to find t h u~ ~ b eroot.^
of surd.^ of the ,form A + m, and d ~ d u c ethat
6)
(ii)
Of course, one can establish each of the nine identities in the four preceding
problems by taklng the appropriate powcr of each side of each equality. applying the
rnultinornial thcorcin, and simplifying. However, such a proof provides no insight
whatsoever into such an equality. nor does it indicate how Ramanujan might have
discovered it. Both the left and right sides of each of the equalities are units
in sornt. algebraic number field. Although Ramanujan never used the term unzt,
and probably did not formally know what a unit was, he evidently rralized their
funcldinental properties. He then recognized that taking certain powers of units
often led to elegant identities.
Berndt, Chan. and Zharig [33] have found generalizations of the identities
above. For example. for any real number a ,
and
(a + 2) + (1 - 4a) \i/2aL-*-l
9 3
which when we set a = 5 in the former equality and a = 7 in the latter yield (i)
and (ii), respectively. of Question 525. As another example, for any real number a ,
On page 341 in his second notebook [171], Ramanujan stated two general
radical identities involving cube roots. For example,
See [23, pp. 34-36] for these two identities and several additional examples of the
sorts we have discussed here.
Many papers have been written on simplifying radicals. In particular. we men-
tion papers by R. Zippel [224] and T. J. Osler [151]. S. Landau [116]. [117] has
given Galois explanations for several types of radical identities.
QUESTION
785 ( J I M S 8. PP. 159-160; 8, P. 232): Shour that
This i s analogous to
Hence,
QUESTION
359 (JIMS 4 , P. 78; 15. PP. 114-117). If
prove that
(+sinrcosi)"'+ f + r o s z s i n z ) " ' = (sin2t~)'/",
and verzfy the result when
1203, pp. 2-91. but their solution is also quite lengthy. It would seem desirable to
have a briefer, more elegant solution, but perhaps this is not possible.
5. Number Theory
QI;ESTION
441 (JIILIS 5 , P. 39; 6, PP. 226-227). Show that
J. Browkiri and A. Schinzel [49] proved that 2" - D = y 2 has at most one solut,ion
if D $ 0 , 4 . 7 (mod 8). Furthermore: if any solut,ion exist,s. then n < 2. They also
pointed out that, in fact, in 1956, they [48] had completely solved a diophantine
equation, easily shown to be equivalent t,o (5.6). and so had given the second solution
to Rarnanlljan's problem. R. Apiiry [8]showed t,hat. if p is a n odd prirne not dividing
D > 0. then (5.7) has at most two s~lut~ioris. Ljuriggren [133] proved t,hat when
D = 7. (5.7) has no solutjioris when p is odd. More recently. J. H. E. Cohn [70]
has shown that for 46 values of D 5 100, (5.7) has no solutions. H. Hasse [94]
and F. Beukers [37] have written excellent surveys on (5.7). Beukers' thesis [37]
and his two papers [38] and [39] contain substantial new results as well. In [38].
Beukers studied (5.7) for p = 2 and proved. among other things, a co~ljectureof
Browkin and Schinzel on the number of solutions when D > 0. 111his second payer
[39],assuming that D is a negative int,eger and p is an odd prime not dividing D .
he showed that there are a t most four solutions. hloreover. he gave a farnily of
equations having exactly three solutions. 111both papers. hypergeometric f u n h o n s
play a central role. LI. H. Le has written several papers on (5.7). most of t,hern
improving results of Beukers. To describe some of his results, assume in t,he sequel
that D < 0, t,hat ( D , p ) = 1, p is prirne, and that N ( D , p ) denotes the number of
solutions of (5.7). In [124],Le considered the case when p = 2 and showed that, in
various cases, N ( D , 2) 5 2, 5 3. = 4. In [I231 and [125], Le considered the case of
an odd prime p and showed that if m a x ( D , p ) is sufficiently large (which is made
precise), then N ( D ,p) 5 3. There are many further generali~at~ions of (5.7) studied
by Le and others. but we will not discuss these here.
QI:ESTION469 (JIMS 5 , P. 159; 15, P. 97). The number 1 + n ! is a perfect
squcre for the values 4 , 5 , 7 of n,.Find other values.
In volume 15. citing Dickson's History [73, p. 6811. RI. B. Rao wrote that the
question was originally posed by H. Brocard [46] in 1876, and then again in 1885
[47]. (The problem was also nlentioned on t,he Norwegian radio program, Verdt
aa vit,e (Worth knowing).) Furthermore. Dickson [73, p. 6821 reported t,hat A .
G6rardin [83] remarked that, if further solutions of
exist. thcn r n has at least 20 digits. H. Gupta [88] found no solutions in the range
8 5 r~ 5 63 and thereby concluded that rrr has at least 45 digits. Recent calculations
[35] have shown that there are no further solutions up to TL = 10".
In 1993. ill. Overholt [I521 proved that (5.8) has only finitely many solutions
if the weak form of Szpiro's conjecture 1s true. but this remains unproved. To state
the weak form of Szpiro's conlecturc. which is a special case of the ABC conjecture.
first set
where p denotes a prime. Let a,b, and c denote positive integers, relatively prime
+
in pairs and satisfying the equality a b = c. Then the weak form of Szpiro's
28 BRUCE C. BERNDT, YOUN-SEO CHOI, AND SOON-YI KANG
(ii)
(iii) 1-
[A+ = [J&q
All three parts of Question 723 may be found on the first page of Rarnanujan's
third notebook 1171, vol. 2, p. 3611, and proofs can be found in Berndt's book
[23, pp. 76-78]. A. A . Krislmaswami Aiyangar [2] later posed a problem giving
analogues. one involving fourth roots and one involving fifth roots, of all three parts
of Question 723. In a subsequent paper [3], he established theorems generalizing
the results in his problem. a i d so further generalized Rarnanujan's Question 723.
K . J . Cheii [60] has also established extensions of (ii) and (iii). Part (iii) appeared
on thc M'illiam L o w d Putnam exam in 1948 [85].
QLESTION 770 (.JIhIS 8. P. 120). If d(n) denotrs the r~urnherof dzvzsors of
n (p.g.,d(1) = 1,d(2) = 2 . 4 3 ) = 2, d(4) = 3 . . . . ) show that
(ii)
J
liin !VF(N&) = -;
\-x 2fi
(ii) lim log M ) ~ - J ' F ( N ~ ~ /=" )0,
v-x
where n is any integer and p is any positive nmmber: show further that i n (ii) p
cannot be zero.
Question 784 is concerned with the approximation of irrational nunlbers by
rational riumbers. The proposed equalities should be compared with Hurwitz's
theorem [150, p. 304, Thm. 6.111: Given any irrational number <. there are
infinitely man; distinct rational numbers rrrln such that
The constant 1/& is best possible. To sec how Question 784 rclates t o Hurwitz's
theorem. we more closely exanline the first assertion in Question 784, which iinplies
that there exists a sequence of positive integers {lV,:) tending t o x such that
as NA --t cc.Setting A l k = [ N A a],we may write the last formula in the equivalent
form
a more general lemma. The two published solutions in the Journal of the Indian
Mathematical Society are more complicated.
In fact, Question 427 is a special case of Gauss' theory of composition of binary
quadratic forms [69, p. 2121, [51, Chap. 71, which we very briefly describe.
Suppose that Q l (x,y ) and Q2(x,g) are integral, positive definite quadratic forms
of discriminant d. Then if Q:j is a form of discriminant d,
6)
(ii)
and that
a result which is also found in both Ramanujan's first and second notebooks [23, p.
1911. The class equation for G7y is not found in Weber's book [219]. However, on
pages 263 and 300 in his second notebook [171],Ramanujan claimed that 21/4/G70
is a root of
This result can be deduced from equivalent results due to R. Russell [182] and
later by Watson [217]. See also Berndt's book [24, pp. 193, 2751. Watson [211]
furthermore pointed out that (5.13) was explicity solved by G. P. Young 12211
in a paper devoted to the general problem of explicitly finding solutions to solv-
able quintic polynomials and to working out many examples, the first of which is
(5.13). Young remarks that (5.13) was "brought under the notice of the writer by
a mathematical correspondent," whom Watson conjectured was A. G. Greenhill,
who had also studied (5.13). A few years later, A. Cayley [58] considerably simpli-
fied Young's calculations. Since the solutions of (5.14) had not been given in the
literature, Watson [211] explicitly determined them. The work of Ramanujan and
Watson is summarized in a paper by S. Chowla [65].
32 BRUCE C. BERNDT, YOUN-SEO CHOI, AND SOON-YI KANG
D. Dummit [75] and V. M. Galkin and 0. R. Kozyrev [82] have also examined
(5.13) and (5.14) and provided some insights into the Galois theory behind the two
equations. For example, the Galois groups of the Hilbert class fields over Q are
dihedral groups of order 10. Unaware of the work of Russell, Watson, and others,
Galkin and Kozyrev rederived the class equation (5.14).
629 (JIMS 7,
QUESTION P. 40; 8 , PP. 25-30). Prove that
(5.15)
l
-
2
+ x
x
n=l
e-nn2xcos {nn2JG?)
=
n=l
(ii)
In volume 8, three solutions are given. In the first, the solver erroneously
7L= 1 71= 1
Chowla [62] pointed out these mistakes and evaluated each series above in terms
of gamma functions. The second and third solutions by N. Durai Rajan and M.
Bhimasena Rao, respectively, are correct.
Equality (5.15) can be found in Section 23 of Chapter 18 in Ramanujan's second
notebook [22, p. 2091, where it is a corollary of a more general transformation
formula, Entry 23 [22, p. 2081. Part (ii) is given as Example (iv) in Section 7 of
Chapter 17 in Ramanujan's second notebook [22, p. 1041. After Ramanujan, set
In that same Section 7, Ramanujan offers the values of p(eP"), p(epnJZ), and
p(e-2") [22, pp. 103-1041, all of which are classical. In particular [22, p. 1031,
In view of (5.16), we see that (5.17) explicitly determines cp(e-"). Part (i) is found
in the second notebook, where it is an example in the aforementioned Section 23
[22, p. 2101. In Berndt's book [22, p. 2101, one can find a short proof that uses
the same generalized theta transformation formula that leads to a proof of (5.15).
Part (i)?in the form (5.17), is also recorded in Ramanujan's first notebook [171,
p. 2851. The values p(e-:'"), p(e-7k), p(ep"), and p(e-4") are also recorded in
THE PROBLEMS SUBMITTED BY RAMANUJAN 33
the first notebook and were first proved by Berndt and Chan [27];see also Berndt's
book [24, pp. 327-3281, where several "easier" values of p are also established
124, p. 325, Entry 11. Explicit values of cp yield at once explicit values for the
i:
hypergeometric function 2 ~ L ( $ , 1; k2) and for the complete elliptic integral of the
first kind K ( k ) , for certain values of the modulus k [24, p. 323, eq. (0.4)].
We shorten Ramanujan's formulation of Question 584. in part, by using the
notation
n-1
(5.18) (a; q)n := n(l- aqk), (a:q), = hrl(a;q),,. (ql < 1.
QUESTION
584 ( J I M S 6' P. 199). Examine the correctness of the following
results:
The identities (5.19) and (5.20) are called the Rogers-Ramanujan identities.
and they have a long, interesting history, which we briefly relate here. As the word-
ing of the problem intimates. Ramanujan had not proved these identities when he
submitted them. Also. as their name suggests. they were, in fact. first discovered
and proved by L. J. Rogers [180] in 1894. Ramanujan had evidently stated them
in one of his initial letters to Hardy, for Hardy [172, p. 3141 later claimed that.
"They were rediscovered nearly 20 years later by Mr Ramanujan, who communi-
cated them to me in a letter from India in February 1913." (For a discussion of
Hardy's assertion. see Berndt and Rankin's book [36, pp. 43-44].) Hardy in-
formed several mathematicians about (5.19) and (5.20). but he obviously failed to
consult with Rogers. The "unproved" identities became well-known, and P. A.
MacMahon stated them without proof and devoted an entire chapter to them in
volume 2 of his treatise Combznatory Analyszs [135]. In 1917. Ramanujan was
perusing old volumes of the Proceedzngs of the London Mathematzcal Soczety and
accidently came across Rogers' paper [180]. Shortly thereafter, Rainanujan [l7O],
[172, pp. 214-2151 found his own proof. and Rogers [181] published a second
proof. There now exist many proofs, which have been classified and discussed by
G. E. Andrews in a very informative paper [6].
The Rogers-Ramanujan identities are recorded as Entries 38(i). (ii) in Chapter
16 in Ramanujan's second notebook [22, p. 161. It is ironic that they are. in fact.
limiting cases of Entry 7 in the same chapter, as observed by R. A. Askey [22, pp.
77-78]. Entry 7 is a limiting case of ViTatson'stransformation for 8P7, and, in view
of the complexity of Entry 7. it is inconceivable that Ramanujan could have found
it without having a proof of it.
The Rogers-Ramanujan identities have interesting combinatorial interpreta-
tions. first observed by ICIacMahorl [135]. The first. (5.19), implies that: The
number of partitions of a positive integer n into distinct parts. each two differing
by at least 2. is equinumerous with the number of partitions of n into parts that
are congruent to either 1 or 4 modulo 5. The second implies that: The number of
partitions of a positive integer n into distinct parts. with each part at least 2 and
34 BRUCE C. BERNDT, YOUN-SEO CHOI, AND SOON-YI KANG
each two parts differing by at least 2. is equinumerous with the number of partitions
of n into parts that are congruent to either 2 or 3 modulo 5.
For further history and information on the Rogers-Ramanujan identities, see
Andrews' paper [6] and books [4, pp. 103-1051, [5] and Berndt's book [22, pp.
77-79].
6. Integrals
Using this notation and the notation (5.18). we abbreviate Ramanujan's Question
386.
QUESTION386 ( J I M S 4, P. 120; 7, PP. 143-144). Show that
which is stated by Ramanujan in his paper [168, eq. ( l g ) ] , [172, p. 571. The
deduction of (6.1) from (6.2) is given on the following page in that paper [168,
eq. (24)], [172, p. 581. Ramanujan evidently never had a rigorous proof of (6.2).
for he wrote [168], [172, p. 571 "My own proofs of the above results make use
of a general formula, the truth of which depends on conditions which I have not
yet investigated completely. A direct proof depending on Cauchy's theorem will
be found in Mr Hardy's note which follows this paper." (That paper is [go]. [91,
pp. 594-5971,) The special case a = 0 of (6.2), which, of course, contains (6.1),
is stated by Ramanujan in his Quarterly Reports to the University of Madras, the
focus of which is Ramanujan's "Master Theorem." To see how Ramanujan deduced
the special case a = 0 from his "Master Theorem," see Hardy's book [92, p. 1941
or Berndt's book [20, p. 3021. The solution of Question 386 by N. Durai Rajan in
volume 7 employs a partial fraction decomposition of the integrand.
R. A. Askey [9].[lo],[ll], [12] has shown that (6.2) is a q-analogue of the beta
integral. For references to extensions and further related work, see the aforemen-
tioned papers by Askey. A nice discussion of (6.2) may also be found in his book
[7, Chap. 101 with Andrews and R. Roy. The evaluation (6.2) also appears as
Entry 14 in Chapter 16 of Ramanujan's second notebook. See Berndt's book [22,
p. 291, where several references for (6.2) and kindred integrals can also be found.
show that
This is a very beautiful result which has been greatly generalized by Berndt
and R. J. Evans [34] in the following theorem.
THEOREM 1. Let g be a strictly increasing, differentinble function o n [0,cx;)
with g(0) = 1 and g ( w ) = x.For n > 0 and t 0, define >
Suppose that
converges. T h e n
471) +4 l l n ) = 241).
To deduce (ii) of Question 783, let g(t) = l + t . Observing that v ( t ) = t ( l + t ) ' l p l ,
setting u = 1 + t , and using the value p(1) = 7r2/12, we find that Theorem 1 reduces
to (ii). The proof by N. Durai Rajan and "Zero" in volume 10 is longer than the
proof by Berndt and Evans of the more general result. Question 783 can be found
on page 373 of Ramanujan's third notebook [22, pp. 326-329, Entry 411.
308 (JIMS 3,
QUESTION P. 168; 3, P. 248). Show that
7r3 T
(i) 0 cot 0 log(sin 0)dO = - - - - log2 2,
48 4
Observe that
was studied by Ramanujan in his paper [166], [172, pp. 40-431 and in his note-
books [20, pp. 265-2671, For related results in the notebooks. set [20, pp.
268-273, 285-2901 and [24, pp. 457-4611, The function
36 BRUCE C. BERNDT, YOUN-SEO CHOI, AND SOON-YI KANG
was also examined by Ramanujan in his notebooks [20, pp. 264-265, 268, 285-
2881. In fact, (ii) is a special case of Entry 16 of Chapter 9 in the second notebook,
which we can write in the form
(see [20, p. 264, penultimate line of p. 2851). Setting x = 114 in (6.4) and
using (6.3), we deduce (ii).
For (i), observe that an integration by parts gives
+ -)
,
-
L (2k + 1) (COS~
('Lk+l).rr
cos
(2k+l).rr
~ T L
- -
, ,
L ( 2 t + 1) (cosh
i=il + cos d m )
463 (JIMS 5,
QUESTION P. 120). If
im cos (nx)
e2nA - 1
dx =d n),
then
7. Series
QUESTION
260 (JIMS 3, P. 43; 3, PP. 86-87). Show, without using calculus,
that
Question 260 is the first question submitted by Ramanujan t o the Journal of the
Indian Mathematical Society. In fact, the first two questions were communicated by
P. V. Seshu Aiyar, Ramanujan's mathematics instructor at the Government College
of Kumbakonam.
The equality (7.1) is a corollary of Entry 4 of Chapter 2 in Ramanujan's second
notebook [20, pp. 28-29]. Ramanujan was not the first to pose (7.1) as a problem;
Lionnet [I311 offered (7.1) as a problem in 1879. The problem can also be found
in G. Chrystal's book [67, p. 3221.
38 BRUCE C. BERNDT, YOUN-SEO CHOI, AND SOON-YI KANG
327 (JIMS 3,
QUESTION P. 209). Show that Euler's constant, namely [the
limit ofl
11-1 71- 1
(ii) tall-'
1
= C tan-' 1
~ = I I (2n + 2k + I ) & k=o ((2k +l ) ~ 5 ) ~
'
Alehr Chand Suri showed in volume 16 that (i) and (ii) are the special cases
r = 1 and n: = 1/& of the identity
show that
(ii)
where
Although no solutions were published, (7.2) has been rediscovered several times
in the literature. Its home is in the theory of elliptic functions, as Ramanujan
himself indicated when he proved (7.2) in his paper [165, p. 3611, 1172, p. 341.
However, Ramanujan was not the first to establish (7.2). The first mathematician
known to us to have proved (7.2) is 0. Schlomilch [187], [I881 in 1877. Although
not explicitly stated, (7.2) was also established by A. Hurwitz [loll, [I021 in his
thesis in 1881. Others who discovered (7.2) include C. Krishnamachari [112], S. L.
Malurkar [137],M. B. Rao and M. V. Ayyar [177],H. F. Sandham [184],and C.-B.
Ling [130]. The discovery of (7.2) in Ramanujan's paper [I651 or notebooks [I711
motivated the proofs by Watson [207], Grosswald [87], and Berndt [19]. More
precisely, (7.2) is an example in Section 8 of Chapter 14 in Ramanujan's second
notebook. The discussion in Berndt's book 121, p. 2561 contains many further
references.
40 BRUCE C. BERNDT, YOUN-SEO CHOI, AND SOON-YI KANG
Schlijmilch and several others cited above, in fact, proved a more general for-
mula than (7.2). Let a ,P > 0 with aP = r2.Then
The special case a = P = -ir of (7.3) yields (7.2). Equality (7.3) is Corollary (i) in
Section 8 of Chapter 14 in the second notebook [21, p. 2551. There is a further
generalization. Let a , P > 0 with a p = n2,and let n be a n integer exceeding 1.
Then
where B k ,k 2 0, denotes the kth Bernoulli number. Observe, by (7.3), that (7.4) is
not valid for n = 1. The first proof of (7.4) known to us is by Rao and Ayyar [177].
Ramanujan also discovered (7.4), and it can be found as Entry 13 in Chapter 14 in
his second notebook [21, p. 2611. If we set a = /3 = -ir in (7.4), assume that n is
+
odd, and replace n by 2n 1, we find that, for each positive integer n ,
which is also in Ramanujan's notebooks [21, p. 262, Cor. (iv)]. Apparently the
first proof of (7.5) is by J . W. L. Glaisher [84] in 1889. Many proofs of (7.4) and
(7.5) can be found in the literature, and readers should consult Berndt's book [21,
pp. 261-2621 for many of these references.
Analogues of (7.3) and (7.4) wherein negative odd powers of k appear in the
summands are also very farnous. Although we do not state the primary formula
+
here, it is generally called "Ramanujan's formula for C(2n I)." It is recorded by
Ramanujan as Entry 21(i) in his second notebook [21, pp. 275-2761, and there
exist many proofs of it. See Berndt's book 121, p. 2761 for almost two dozen
references.
Most of the proofs of (7.2)-(7.5) do not involve elliptic functions or modu-
lar forms. However, Berndt [19] has shown that all of the identities discussed
above, and others as well, can be derived from one general modular transformation
formula for a large class of functions generalizing the logarithm of the Dedekind
eta-function.
Question 358 has a beautiful generalization. Let cu, ,i3 > 0 with a0 = 7r2, and
let n be a positive integer. Then
If n is even and cu = /3 = .rr in (7.7), then (7.6) arises. Equality (7.7) is Entry 14 in
Chapter 14 in Ramanujan's second notebook [21, p. 2621, and the corollary (7.6) is
recorded immediately thereafter. The first proof of (7.7) appears to be by Malurkar
[I371 in 1925. Proofs have also been given by Nanjundiah [147] and Berndt [19,
p. 1771.
+
Analogues of (7.6) and (7.7) exist for negative powers of 2k 1 in the sum-
mands. Such results were also found by Malurkar [137], Nanjundiah [147], and
Berndt [19]. These formulas, (7.6), and (7.7) can be deduced from the same gen-
eral transformation formula; see [19] for details.
QUESTION
546 (JIMS 6, P. 80; 7, PP. 107-109, 136-141). Show that
(ii)
The evaluations (i) and (ii) are Examples (v) and (vi), respectively, in Section
32 of Chapter 9 in Ramanujan's second notebook [20, p. 2891. These and several
other evaluations of this sort were derived in [20, pp. 288-2901 from the following
two related corollaries given in the same section. If 1x1 < 1, then
(9
(ii)
The two published proofs are elementary. The first, by K. R. Rama Aiyar, uses
Euler's elementary identity
x(l
k=O
n
- ak:+1)a1a2... ak = 1 - a 1 a 2 . .. anti.
8. Continued Fractions
352 ( J I M S 4,
QUESTION P. 40). Show that
THE PROBLEMS SUBMITTED BY RAMANUJAN 43
when q = e-'", -ep", respectively. As the name suggests, R(q) was first studied
by L. .I. Rogers in 1894 [180]. In his first letter t o Hardy, Ramanujan stated both
(i) and (ii) [172, p. xxvii], [36, p. 291. and in his second letter, he gave the value
of ~ ( e - ~ " / & [172,
) p. xxviii], [36, p. 571. In both letters, Ramanujan wrote
that R(eP"fi) "can be exactly found if n be any positive rational quantity" [172,
p. xxvii], [36, pp. 29, 571. (We have quoted from the first letter; the statement
in the second letter is similar but is omitted from the excerpts of Ramanujan's
letters in the Collected Papers.) In both his first notebook [171, p. 3111 and lost
notebook [173, pp. 204, 2101, Ramanujan offered several further values for R(q).
On page 210 of [173], Ramanujan planned to list fourteen values, but only three
are actually given. Since the lost notebook was written in the last year of his life,
his illness and subsequent death obviously prevented him from determining the
values he intended to compute. In several papers [159]-[163], K. G. Ramanathan
derived some of Ramanujan's values for R(q).All of the values in the first notebook
were systematically computed by Berndt and Chan in [26], while all the values,
including the eleven omitted values, were established by Berndt, Chan, and Zhang
in [29]. Moreover, in the latter paper, the authors demonstrated for the first time
the meaning and truth of Ramanujan's claim that ~ ( e - " A )"can be exactly found"
when n is a positive rational number. More precisely, they used modular equations
to derive some general formulas for R ( e - " A ) in terms of class invariants. Thus, if
the requisite class invariants are known, R(e-"fi) can be determined exactly. A
brief expository account of this work can be found in 1311. S.-Y. Kang [106] has
proved a formula in the lost notebook [I731 that likely was used by Ramanujan to
compute values of ~ ( e - " 6 ) .
QUESTION 541 (JIMS 6, P. 79; 8, PP. 17-20 ). Prove that
where(a)o=1and(a),=a(a+l)(a+2)..~(a+n-1),n~1,witha=x=1/2,
and tjhe continued fraction of Legendre [127, p. 5091
1 a 1 a+1 2 a t 2 3
a real, x > 0,
;c+i+x+ 1 + x + x +x+...'
with z = a = 112.
44 BRUCE C. BERNDT, YOUN-SEO CHOI, AND SOON-YI KANG
where x is any complex number outside (-a, 01. In turn, Entry 43 is a corollary
of Entry 42 [21, p. 1651, which gives a continued fraction of Legendre [126] for
certain generalized hypergeometric functions 1Fl (a; b; x). More precisely, if n is a
nonnegative integer, and x 6 ( - O O , ~ ]then
,
which Watson [208] found in Laplace's treatise on celestial mechanics [121, pp.
253-2563. However, the first rigorous proof is due to Jacobi [103].
1049 (JIMS 11, P. 120). Show that
QUESTION
(i)
sin(nx)dx
-
-
1 2
Jtr
3 >
n+- - -
x +x+x+... n+n+n+...
sin(irnx)dx 1 l 2 22
- - - - -
32
(ii)
n + n + n + n+..
x+-
x + +
-
x
-
x +...
The formulas above are valid for n > 0.
If we set
and
and
T H E PROBLEMS SUBMITTED BY RAMANUJAN 45
9. Other Analysis
QUESTION
261 (JIMS 3, P. 43; 3, PP. 124-125). Show that
where n is an even positive integer. Further product evaluations of the same kind are
located on pages 279 and 287 in the second notebook [23, pp. 335-337, Entries
1-41,
QUESTION
571 (JIMS
show that
(G)
46 BRUCE C. BERNDT, YOUN-SEO CHOI, AND SOON-YI KANG
This result was proved by Ramanujan in his paper [166], [1?2, pp. 40-431.
An equivalent formulation can be seen on page 286 in his second notebook [24, p.
4611.
QUESTION
294 (JIMS 3, P. 128; 4, PP. 151-152). Show that [if x is apositive
integer]
as x tends to m. In his paper [209], Watson remarked, "I shall also give reasons,
which seem t o me to be fairly convincing, for believing that k lies between &
and
- ." To the best of our knowledge, however, it appears that this claim has never
been firmly established. However, K. P. Choi [61] has shown that, for all x 2 1,
o<++&.
Question 294 has the following connection with probability [139], [140]. Sup-
< <
pose that each of the n independent random variables X k , 1 k n, has a Poisson
distribution with parameter 1. Then S,, := C;=IX I , has a Poisson distribution
with parameter n. Thus,
C. Y. Yilbrim 12201 has studied the zeros of the partlal sum in Question 294;
a certain s l m involving these zeros arises in formulas for certain mean values of
I<(; + <
zt)lL, where denotes the Riernanri zcta-function.
Results related to Question 294 can be found in D. E. Knuth's book [ l o g , pp.
112-1171, For further discussions of this problem see the commentary in Szego's
Collected Papers [200, pp. 151-1521 and Berndt's book [21, pp. 181-1841.
QUESTION
738 (JIMS 8 , P. 40). If
show that ~ ( x = ) 1 when x lzes b e t w ~ e n0 and 1 ; and that *(x) #1 when x > 1.
Fznd the lzmzt of
{9(1 + c) - d l ) } 1 6
as t + 0 through posztive values
The first part of Question 738 can be found as a special case of a corollary in
Section 13 of Chapter 3 in Rarnanujan's second not,ebook 120, p. 701. Rarnarlujari
also communicated a version of this corollary in his Quarterly Reports to t,he Uni-
versity of hladras [20, p. 306, eq. (1.14)]. In fact, (9.1) has a long history. It can
be traced back t,o papers of J. I]. Lanibert [I151 in 1758. J . I,. Lagrange 11141 in
1770, and Euler [78]. [BO] in 1783. The most common proof utilizes the Lagrange
inversion formula. Indeed, in P6lya and Szego's t,reatise [156, pp. 125, 135, 301,
316-3171, (9.1) is given as a problem to illustrate this formula. For many further
references, see Berndt's book [20, pp. 72, 3071.
The first complete solution t,o Problem 738 was given by Szego [199], [200,
pp. 143-1521. The limit queried by Ramanujan equals -2.
F. C. Auluck [13] and Auluck and Chowla [14] conjectured that ~ ( cis) c o n -
pletely rnonot,onic for x > 1, that is, (-l)"(k)(x) >
0, k = 0 , l . 2 . . . . . S. LI. Shah
and U. C. Sharma [I911 established this inequality for k = 0 . 1 , 2 , 3 , 4 . On the other
hand, Shah [190] proved that xd(x) and e . r b ( x ) are not completely monot,onic.
Finally. R.. P. Boas. Jr. [41] proved that $(.c) is not completely monot,onic on any
interval ( c , m).
QUESTION
526 (JIMS 6, P. 39). If n zs any posztzve quant7ty. show that
It was pointed out by S. S. hlacintyre [134] that Szego actually showed that
7 4 0 (JIMS 8 ,
QI-ESTION P. 40; 8 , PP. 220-221). If
where [XI denotes the greatest integer i n x . show that ~ ( xis) a continmous function
of z for all positiue values of .r, and oscdlates from +T to ,when s becomes
infinite. Also diflerentmte ~ ( - r . ) .
Questions 730 and 753 are very closely related. with Stirling's formula playing
a key role in the solutions.
QI~ESTION
7 5 4 (JIillS 8 . P. 8 0 ; 1 2 , P. 101: 1 3 , P. 1 5 1 ) . Show that
(.r+ a
(3.1-+ n -
-
+ +
+
b)!(X.r+ 2b)!(9.r a + b)!
r.)!(3.r (1 - b c)!(12.r 3b)!
=
+ g.
This result is a straightforward application of Stirling':; forrriula.
T H E PROBLEMS SUBMITTED BY RAMANUJAN 49
On page 346 in his second notebook, Ramanujan stated a much more general
result. Under certain prescribed conditions on m , n, &. B,. a k , and b,, 1 5 k 5
m, 1 1 3 1 n ,
Question 605 is easily seen to be a special case of (9.3),and is one of several examples
offered by Rainanujan on page 346. For proofs of (9.3) and the aforementioned
corollaries. see Berndt's book [23, pp. 340-3411.
10. Geometry
Robert Larnphrrc. Jtrry Lange. Frank Olver, C. A. Rrddi. Bruce Reznick. Haakon
IIkadcland. Keiineth S. Williams. arid the referee for helpful comments and refer-
Cll('CS.
References
1. A. A. Krislmas~vamiAiyangar. Partial Solutzon to Questaon 784, J . Irldian hlath. Soc. 18
(1929 X I ) , 214 217.
2. A. A. Krishnaswarni Aiyangar. Questzon 1567, J. Indian Math. Soc. 18 ( 1 9 2 9 3 0 ) , 224.
.'3. A. A. Krishuaswarni Aiyangar, Some arithmetzcal zdentztzes. J . Indian hlath. Soc. 18
(1920 30). 278 285.
4. G. E. Antlrews, T h e Theory of Partztzons, Addison Wesley, Reading, MA, 1976.
5. G. E. Xntircws, q-Serzes: Thezr Development and Applzcatzon zn Analysis, Number The-
ory. Cornbznatorzc~,Ph:yszcs, and Corr~puterAlgebra. CBhIS Regional Confercncc Series
in hSat11ematics No. 66, American hlathematical Society, Providence, 1986.
6. G. E. i2ntlrews. O n the proofs of the Rogers Ramanzyan zdentztzes, q Series and Partitions.
Springer Verlag, New York, 1989. pp. 1- 14.
7. G. E. Antlrcws, R . A. Askcy. ant1 R . Roy. Speczal Functzons, Ca~nhridgeUniversity Press
(to appear).
8. R . Apdry. Sw ,un(: iquatzon daophantzenne. C . R . Acad. Sci. Paris 251 (1060). 1451-1452.
9. R . A. Askey. R o m m u p n ' s ertens7ons of the gamma and beta functions, Amer. Math.
hIor~thly87 (1980). :346 359.
10. R . A . Askey. Bcta 7n,tegrals zn Ramanujan's papers, hzs unpublzshed work and further
c.rc~nrplcs.Rarnanrijan Revisited. Acadernic Press, Boston, 1988, pp. 561 590.
11. R . A. Askey. Beta ~ n t e g m l sand q extensions, Proceedings of the Ramanujan Centen-
nial Intcrrmtior~alC o r ~ f c r e ~ ~( 1c e5 1 8 December 1987, Armamalainagar), T h e Ramanujan
hIatllernatica1 Soriety. hladras. 1988, pp. 85 102.
12. R . A. Askey. Rnnlanmjan and hypergeometrzc serzes, Ramanujan I~lternationalSymposium
on Analysis ( N . K. Thakarc, cd.). hlacmillan India, Madras, 1989, pp. 1 - 8 3 .
13. F. C . Ai~luck. On s o ~ n ctheorems of Ramanujan, Proc. Irdian Acad. Sci., Sect. A. 11
(1040). 376 378.
14. F. C . Auluck anti S. Chowla, Some propertzes of a functzon considered by Ramanujan, J .
Indian hlath. Soc. (N. S.) 4 (1940), 169 173.
15. hS. \'erikata R a m a Ayyar A note on threc allzed problems of Ramanujan, Half-yearly J .
hIysorc Ur~iv.Sect. B Scicnce 1 (1941), 109-114.
16. hl. V. Ayyar ant1 hl. B. Ruo, Types of solutzonw of x,' y:' + + = 1 zn zntegers (I). J .
Irdian hlath. Soc. (N. S.) 4 (1940). 47 70.
17. W. N . Bailey, A qeneralzzatzon of an zntegral due to Ra,manujan, J . London Math. Soc. 5
(1930), 200 202.
18. U'. N . Bailey, A note o n an znukgral due to R a m a n ~ u n J, . London Math. Soc. 6 (1931),
216 217.
19. B. C'. Barr~tlt.Modular transformatzons and generalzzatzons of several formulae of Ra-
mnn~rjan.Rocky hlt. J . h l a t l ~ 7 . (1977), 147 189.
20. B. C. Bcrrdt,. Ra~nanujan'sNotebooh, Part I, Springer Verlag, New York, 1985,
21. B. C. Bcrr~tlt.Ru~nanujan's notebook^, Part II. Springer Vcrlag, New York, 1989.
22. B. C. Berridt. Rnman,ujan's Notebooks, Part 111, Springer Verlag, New York, 1991.
23. B. C:. B e r r d t , Ramanujan's Notebooks, Part I V , Springer Verlag, New York, 1994.
2 4 B. C . B e r r ~ d t ,R a m a n u ~ a ~ Notebooks,
n'~~ Part V . Springer Vcrlag, New York, 1998.
25. B. C. Bcrrltlt and S. Bhargava. Ranianujan for lowbro~u:~, Amer. hlath. hlonthly 100
(1993). 644 6.56.
26. B. C. Berntlt and H. H. char^. Some ualues for the Rogers Ramanujan contznued fractzon,
C:ar~a(l.J . hlath. 47 (ISST,), 897 914.
27. B. C . Berr~dtam1 H. H. Ctlari. Ra7r~unu.lan'sezpl7czf ~ialuesfor the classzcal theta fun,ctzon,
hSathcrr~atih42 (1995), 278 294.
28. B. C:. Bcrntlt. H. H. C l ~ a n ,and L. C . Zhang, Ramanujnn,'s clans anvarzants and cubic
co'rr.tznvc,dfractaon. Acta Arith. 73 (199.5), 67 85.
29. B. C:. Bern(lt, H. H. Chan, a r d L. C. Zhang, Erplzczt evaluatzons of the Rogers Ramanujan
contznued fmctzon. J . Rcirie Angew. h l a t l ~ .480 (1996); 141 159.
T H E PROBLEMS SUBMITTED BY RAMANUJAN 51
128. D. H. Lehmer, O n the dzoph,antzne equatzon .c" + + y:' 2:' = 1. J. London Math. Soc. 31
(1956), 275--280.
129. L. Leu-in, Polylogarzthms and Assoczated Functzons, North Holland, New York, 1981.
130. C. B. Ling, Generalzzatzon of certain sum,matzons due to Ramanujan, SIAM J . Math.
Anal. 9 (1978), 34 48.
131. M. Lionnet, Questzon 1294, solutzon by M. A . Lazsant, Nouv. Ann. Math. 18 (1879).
330 332.
132. W. Ljunggren, Oppgave n r 2, Norsk Mat. Tidsskr. 25 (1943), 29.
133. W. Ljunggren, O n the Diophantzne equatzon Cx" f = yy",Pacific J. hlath. 14 (1964),
585-596.
134. S. S. Macintyre, O n a problem of Raman,ujan, J . London hlath. Soc. 30 (1955), 310 314.
135. P. A. MacMahon, Combznatory Analyszs, Val. 2, Carrlbridge University Press, London,
1916; reprinted by Chelsea, New York, 1960.
136. K. Mahler, Note on Hypotheszs K of Hardy and Littlewood, J . London Math. Soc. 11
( l 9 3 6 ) , 136-138.
137. S. L. hlalurkar, O n the applzcatzon of Herr Mellzn's integrals to some series, J . Indian
Math. Soc. 16 (1925 26), 130 138.
138. J . C. W. hlarsaglia, The zncomplete gamma functzon and Ramanujan's ratzonal approxz-
matzon to e r ,J . Stat. Comput. Simul. 24 (1986), 163 168.
139. hlathematical Gazette, Solutzons and coinments o n 67.E and 67.F, Math. Gaz. 68 (1984),
57 59.
140. hlathematical Gazette, Solutzons and comments on 67.G an,d 67.H, Math. Gaz. 68 (1984),
139 141.
141. J. H. McKay, T h e Wzllzam Lowell Putnam mathematzcal competztzon, Amer. Math. Monthly
74 (1967), 771 777.
142. N. B. hIitra, Reply to Questzon 45, Arner. hlath. Monthly 32 (1925), 506 509.
143. L. J. Mordell, The diophantzne equatzon 2" = .c% 7, Ark. Mat. 4 (1962), 455 460.
144. L. J. Mordell, Dzophantzne Equatzons, Academic Press, London, 1969.
145. T . Nagell, L@smzng tzl oppgave n,r 2, Norsk Mat. Tidsskr. 30 (1948), 62 64.
146. T. Nagell, T h e daophantzne equatzon x' + 7 = 2" , Ark. Mat. 4 (196l), 185-187.
147. T . S. Nanjundiah, Certazn summatzons due to Ramanujan, and thezr generalzsatzons,
Proc. Indian Acad. Sci. Sect. A 34 (1951), 215 228.
148. M. T. Naraniengar, A set of sznzvltaneous equatzons, J . Indian Math. Soc. 5 (1913), 61 62.
149. K. Narasimha Murthy Rao, Remarks o n a serzes of Ramanujan, hlath. Student 28 (1960),
41 43.
150. I. Niven, H. S. Zuckerrnan, and H. L. Llontgornery, A n In,troduction to the Theory of
Numbers, 5th ed., John Wiley. New York, 1991.
151. T. J . Osler, Cardan polynomials and the reductzon of radzcals; preprint.
+
152. hl. Overholt. The diophantzne equatzon n! 1 = m', Bull. London Math. Soc. 25 (1993).
104.
153. R . B. Paris, O n a generalisatzon of a result of Ramanujan connected wzth the expon,ential
serzes, Proc. Edinburgh Math. Soc. 24 (1981). 179-195.
154. E. G. Phillips, Note on a problem of Ramanujan, J . London Math. Soc. 4 (1929), 310-313.
+
155. V. D. Podsypanin, T h e equatzon z" f ?/,' 2:' = 1, Tul. Gas. Pen. Inst. Ucen. Zap. Mat.
Kaf. (1968), 4 9 .
156. G. P6lya and G . Szego, Problems and Theorems zn Analysts, Vol. 1 , Springer Vcrlag;
Berlin, 1978.
157. A. P. Prudnikov, Yu. A. Brychkov, and 0. I. Marichev, Integrals and Series, Vol. 1:
Elementary Functzons, Gordon and Breach, New York, 1986.
158. F. E. Prym. Zur Theorze der Gammafunction, J . Reinc Angew. Math. 82 (1877). 165- 172.
159. K . G. Ramanathan, O n Ramanujan's contznued fractzoiz, Acta Arith. 43 (1984), 209 226.
160. K. G . Ramanathan, O n the Rogers Ramanulan con,tznued fractzon, Proc. Indian Acad.
Sci. (Math. Sci.) 93 (1984). 67 77.
161. K . G. R a ~ n a n a t h a n ,Ramanujan's contznued fractzon, Indian J . Pure Appl. Math. 16
(1985), 695 724.
162. K . G. Rarnanathan, Some applzcations of Kronecker's limzt formula, J . Indian Math. Soc.
52 (1987), 71 89.
T H E PROBLEMS SUBMITTED BY RAMANUJAN 55
197. 3. J. Sylvester, O n a remarkable dzscovery zn the theory of canonzeal forms and hyperde-
terminants, Phil. Mag. 2 (1851), 391410.
198. J. J . Sylvester, Collected Mathematical Papers, Vol. 1: Chelsea, New York, 1973.
199. G. Szegii, ~ b e reinzge won S . Ramanujan gestellte Aufgaben, J . London Math. Soc. 3
(1928), 225-~232.
200. G. Szego, Collected papers, Vol. 2, Birkhauser, Boston, 1982.
201. G. Turnwald, A note o n the Ramanujan Nagell equatzon, Number-Theoretic Analysis (Vi-
enna, 1988 89), Lecture Notes in Math. Vol. 1452, Springer Vcrlag, Berlin, 1990, pp. 206
207.
202. R. C. Vaughan, O n Waring's problem for cubes, J . Rcine Angew. Math. 365 (1986),
122 170.
203. V. R. Thiruvenkatachar and K. Venkatachaliengar, Ramanujan at Elementary Levels;
Glimpses (Unpublished Lecture Notes).
204. T. Vijayaraghavan, Questzon 1528, J. Indian Math. Soc. 17 (1927 -28): 192.
205. T . Vijayaraghavan and G. N. Watson, Solution to Questzon 784, J . Indian hlath. Soc. 19
(1931), 12-23.
206. hl. B. Villarino, Ramanujan's approszmatzons to the arc length of a n ellzpse; preprint.
207. G . N. Watson, Theorems statcd by Ramanujan 11, J . London hlath. Soc. 3 (1928), 216 225.
208. G. N. Watson, theorem,^ stated by Ramanujan ( I V ) : Theorems on approxzmate integratzon
and summatzon of serzes, J . London hlath. Sac. 3 (1928), 282 289.
209. G. N. Watson, Theorems stated by Ramanujan ( V ) : Approxzmatzons connected wzth e " ,
Proc. London Math. Soc. (2) 29 (1929), 293 308.
210. G. N. Watson, Solutzon to Questzon 722, J. Indian Math. Soc. 18 (1929-30), 113 117.
211. G. N. Watson, Solutzon to Questzon 699, J . Indian hlath. Soc. 18 (1929-30), 273-275.
212. G. N. Watson, Solutzon to Question 770, J. Indian Math. Soc. 18 (1929 30). 294298.
213. G. N. Watson, Ramanujan's note books: J . London hlath. Soc. 6 (1931), 137 153.
214. G. N. Watson, Theorems stated by Ramanujan ( X I V ) : a szngular modulus, J . London
hlath. Soc. 6 (1931), 126-132.
215. G. N. Watson, Some szngular m,oduli (I), Quart. J . Math. 3 (1932), 81--98.
216. G. N. Watson, Some szngular moduli (11), Quart. .J. Math. 3 (1932), 1 8 9 212.
217. G. N. Watson, Szngular modulz (4), Acta. Arith. 1 (1936), 284 323.
218. G. N. Watson, Unpublished lecture on solvzng quzntzc equatzons.
219. H. Weber, Lehrbuch der Algebra, drztter Band, Chelsea, New York, 1961.
220. C. Y. Yilbrim, A s u m over the zeros of partzal sums of e ' , J . Ramanujan Math. Soc. 6
(1991), 51--66.
221. G. P. Young, Solvable quintzc equations wzth commensurable coeficzents, Amer. J . Math.
10 (1888), 99 130.
222. L.-C. Zhang, Kronecker's lzmzt formula, class anvariants and modular equations (II),
Analytic Nurr~berTheory: Proceedings of a Conference in Honor of Heini Halberstarn,
Vol. 2 (B. C. Berndt, H. G. Diamond, and A. J. Hildebrand, eds.), Birkhauser, Boston,
pp. 817-838.
223. L.-C. Zhang, Kronecker's lzmzt formula, class znvarzants nnd modular equatzons ( I l l ) ,
Acta Arith. 82 (1997), 379-392.
224. R . Zippcl, Szmplzjicatzon of expresszons znvolvzng radzcals, J . Symbolic Comp. 1 (1985),
189 210.
1. Introduction.
Any squarr integrable function f ( x ) on a finite interval ( a , h ) is completely
drternliilecl by its moments
, 171.
Proof. Assume that 0 5 Ic 5 m,. Using the assumptions of t,he theorem we get
= ll h(w){o(w)j4*(w)dw
= iik(O)
and the theorem is proved. As a particular case of this theorem we get the following.
COROLLARY 1.2. Let h ( w ) be a given zntegrable function on y. Assume that
G ( w ,t ) is any integrable function on y x [c,dl such that
Proof. Assume t,hat 0 < k < 2n - 1. Using the assumptions of the corollary we get
Since L, h ( w ) G ( w t)wkdw
, is a polynomial of degree <
2 n - 1 in t we can compute
the integral with respect to t exactly by the Gaussian quadrature. Thus
2. Generating functions.
, (x),. . . be a system of orthogonal polynomials on [a,b ] , finite or
Let P o ( x ) Pl
infinite, with respect to a given weight function h ( x ) . Let
X
G ( s ,t ) = c,,PI,(x)tl'
be the generating function of this system. For any polynomial Q ( x ) of degree <k
we clearly havc
I1
since the Fourier coeficirnts b,, (Q) = h ( r ) Q ( x P,,
) ( x ) d sof Q with respect to the
system {P,,) are equal to zero for n > I;. Thus G ( x ,t ) satisfies the nlain conditions
of corollary 1.2. As an immediate consequence the corollaries given below hold.
60 B. BOJANOV AND A. SRI RANGA
where TI,( x ) are the Tchebycheff polynomials of the first kind. Then
COROLLARY 2.1. Let F ( t ) be any non-negative function o n [ - I , 11 which van-
ishes at most o n a set of measure zero. Let { t , ) ; and {a,);' be the nodes and
coefJicients of the Gaussian quadrature rule associated with the weight function
g ( t ) = ~ ( t ) / d omn [ - I , 11. T h e n the rational approximation
For X > 0, let P;')(.c) be the Gegenbauer polynomials. It is known (see [3])
that they are orthogonal with respect to the weight function { 1 - x 2 ) ' p 1 / 2 in [-I. 11
and
1
preserves 2 n moments p~( f ) = J", f ( x ) x L ( l ~ ~ ) ~ " ' / ~k: d= r0 , l . . . . ,272
- - 1. of
the functzon
where PI, are the orthonorrnal polynornials associated with the weight function h ( t )
in the interval [a,b].
T H E O K3.1.
E ~ ~6 % ~any
. n,on-negat,ive function F ( t ) the polynomial
where { t , ) ; and { a , ) ; are the nodes ond coeficzents of the Gausszan rule wzth
respect to the wezght functzon g ( t ) = F ( t ) h ( t ) ,preserves 2 n moments p k ( f ) =
J: h ( x )f ( z ) r L d r .k = 0 . 1 , . . . ,271 - 1 . of the polynolnzal
L: h ( x ) K 2( x ,t ) z L d x=
exists. then we may call it, in accordance with [3]. a sequence of orthogonal polyno-
mials in relation to the function h ( w ) on y. These polynomials satisfy the recurrence
relation
Here the curve F , which is the image of y by the mapping w + I ) , is such that for
any v E l- also P"7: E I?. The function [ ( I ) )= c ~ n s t . v - ' / ~ h ( w ( vintegrable
)), in r,
is such that
G ( w ,t ) = c,, Pn ( w ) t n .
Then
From (4.1) this last equation means that J , & ( u , t )u"[(u)dv is a polyrlonlial of
degree k in t . Now we can state the following theorem.
SOME EXAMPLES O F MOMENT PRESERVING APPROXIMATION 63
where C is the unit circle { u : u = e'" -K < 6' < K ) . T h e recurrence relation
' ( v ) = 1 , B ,( 1 1 ( u ) = ( v - 1 ) and
wzth B,(,' B:' ' ( 2 1 ) = ( v - 1)'+2v, holds. Consequently
B $ ~ ' ) ( U ) = ll" + (-I)", n > 1.
Furthermore.
Proof. From the Tchebycheff polynomials TI,( u t ) . n 2 0, wp obtain that the real
monic polynomials
P r L ( w=
) 2Pr'" ( z ) " ~ (, -, i w ) , n > 1,
satisfy the orthogonality property
where yl = { w : % ( w )= 0 and - 1 < S(w)< 1). From the recurrence relation and
+
t ) ,= ( 1 - z u t ) / ( l - 2wt t 2 ) for T, ( U I ) we obtain
the generating function ~ ( " ) ( w
satisfy
(G.X)
with B,j"")(v) = 1 and B, ( v ) = ( v - I), holds. Furthermore,
Consequently
COROLLARY 4.4. Let g ( t ) be any non-negative function on the interval [c,dl.
Let us also assume that 1 $ [c,dl. If { t , ) ; and { a , } ; are the nodes and coeficients
of the Gaussian quadrature rule associated wzth the weight function g ( t ) on [c,dl,
then the approximation
+
preserves 2n moments p k ( f ) = Jc,f ( v ) v k { ( v 1 ) 2 X 1 / v ' ) d v ,k = 0,1,. . . ,2n - 1,
of the function
where ((11) is e weight function such that the moments pn, = J, < ( v ) d vexist for
all rn - 0 , & I , f2.. . . . It is known that when the hankel determinants H,(-") and
SOME EXAMPLES O F MOMENT PRESERVING APPROXIMATION 65
HA-"+'), n >
0 , associated with the moments are different from zero then the
polynomials B,, ( v ) exist and satisfy
(5.2) B n + l ( v ) = ( v - Pn+l)Bn(v) - ~ n + ~ ~ B n - l ( v ) , 2 1,
with B o ( v ) = 1, B l ( v ) = v - PI, /31 = p ~ / p - 1 ,and for n > 1
Now we are able to prove the result (5.5) by induction. We assume that F'LNP2)(t)=
(N-2)
pN-2tk for 1 5 k 5 N - 3 . Since ~ i ~ - ' ) ( =
t )P N - ~and H N P 2 ( t )= p N - 2 B N - 2 ( t ) ,
(N-2)
the assumption leads also to F N _ , ( t ) = p N - z t N - 2 . With these results the relation
(5.6) immediatly gives F L N ) ( t )= pNt"or 1 k < <
N - 1. The initial conditions
for the induct,ion are clearly seen. This completes the proof.
As consequences of this theorem we have the following results.
66 B . BOJANOV AND A . SRI RANGA
T H E O R E5 M
. 2 . Let the functzon F ( t ) be such that g ( t ) = t V $ ' ( t ) < ( t ) zs non-
negatrve zn [c.d] and the zntegral J~~ t v i n l g ( t ) f ( t ) d t ezzsts for nl = 0 , 1 , . . . . Let
{t,);" and {a,);' be the nodes and coeficzents of the Gausszan rule wzth respect to
the wezght g ( t ) . T h e n the polynomzal
here h ( x ) is a fixed weight function on [ a ,b]. Precisely, the linear system of equations
+ + +
bo+o (x) bl yl ( x ) . . . b,, _ $I, 1 ( x ) that preserves the monlents of f , provided
-
D f 0 . We shall consider in this section the particular case when {Oo, @ I , . . . , 4,1-I }
is the algebraic system
SOME EXAMPLES O F MOMENT PRESERVING APPROXIMATION 67
where t l < t 2 < . . . < t,, are any fixed points which do not lie in [a,b]. We assume
that these points are inside the interval [c,dl. where [a,b] n [c.d] = 0. Let us uote
first the following sirnple fact.
L E M M A6.1. For any weigh,t function h ( r ) o,n [a,b] we have
Proof. Assume the contrary Then tllcre exist numbers bo, bl , . . . , b , , 1 , at least
one bJ distinct from zero, such that
This implies
rational function of the forrn ( . r ) / Q , , ( x )with P,,-1 E xi,- 1 which has the
.
same moments ,uo(f ) , pl ( f ) . . . . p i l l ( f ) as f with respect to the algebraic systern
1, x , . . . , x l . Now c011sider the question: Does there exist a rational function
Pn ( x ) / Q T( Lx ) which preserves 2n moments of f ' If so. Pi, 1 ( x ) is completely
- -
Since Q,, ( x ) ,sQ,, ( x ) ,. . . ,xri-I Q,,( 2 ) can be written as linear corrlbinatio~lsof s'.
J .
= 0 , 1 ,. . . 2n - 1, we get in particular that
h x s , r - , x h = 0, /c = 0 . 1 , . , n - 1.
which means that P,,_1 ( x ) is the polynomial of best L2-approximation of the func-
tion f ( s ) Q , (, x ) . Thus,
68 B. BOJANOV AND A. SRI RANGA
where {L,) is the system of orthonormal polynomials with respect to the weight h ( t )
in [a,b]. Thus the computation of the moment preserving rational approximation
Pn-1 ( x ) / Q , ( x )t o a given function f , if it exists, is reduced t o an non-linear problem
of determining the zeros t l ,t a , .. . , t, of Q,(x). In view of theorem 1.1 such an
approximation exists and it is completely determined if f can be presented in the
form
with some integrable non-negative function g(t) on [c,dl. Indeed, in this case one
can show as in the proof of lemma 6.1 that the functions
where {a,) are the coefficients and i t j ) are the nodes of the Gaussian quadrature
formula for the system {Q,,)?-l on [c,d ] corresponding to the weight g(t).
The question of convergence of P,- 1 ( z ) / Q ,( z ) to
can be simplified applying the following scheme. Given z @ [c,dl choose a function
Q,(t)from span {%, . . . @ z n - l } such that
a t some points yl < . . . < yn from [c,dl. In particular one can choose y, = t,,
j = 1,.. . ,n, in which case Q, is the Hermite interpolant with double nodes a t
t l , . . . , t,. Having chosen @ we rewrite the remainder as
=
d
9 ( f ) d t-
z-t
ld +
Q,(t)g(t)dt L d
@(t)g(t)dt
SOME EXAMPLES O F MOMENT PRESERVING APPROXIMATION 69
(since the Gaussian quadrature formula for the Tchebycheff system Q o , . . . , @212-
integrates exactly a),
and because interpolates 5
at t l . . . . . t,, . we get
Let us apply this scheme in the case [a.b] is replaced by a circle r of radius
r > 1 and centered at z = 0. Choose [c,d] = [-I, I ] and h(x) = 1. Then we get
where
we get
R(f;
-
- Qr,( z ) q ( z )
1' g ( t ) Q r l ( t ) ~ ( tdt)
z-t
-
-
1 '(') ' ( ' )
z-t
- -I
+ Q, ( t )4 2 )
z-t
dt
The integral of the first term in thc square brackets above is zero since ( q ( t ) - q ( z ) )/
( z - t ) is a polynomial of degree n - 1 and thus the integral can be computed exactly
by the Gaussian quadrature formula (which gives a value 0 since Q,, ( t ) vanishes at
the nodes). Therefore the next elementary estimate
yields that / R(f ; z ) 1 tendh to zero as n tends to rx, for z far from [ - I , 11.
B. BOJANOV AND A. SRI RANGA
References
[I] B. Bojanov and L. Gori, Moment Preservzng Apprommatzons, Math. Balkanzca (New Series),
t o appear.
[2] C.F. Bracciali, J.hl.V. Capela and A. Sri Ranga, Symmetric orthogonal L-polynomials zn the
complex plane, in Orthogonal Functions, Moment Theory and Continued Fractions: Theory
and Applications, (W.B. Jones and A. Sri Ranga, eds.), Lecture Notes in Pure and Applied
hlath. 199, Marcel-Dekker Inc., New York, 1998, 57-67.
[3] T.S. Chihara, An Introduction to Orthogonal Polynomzals, Mathematics and its Applications
Series, New York: Gordon and Breach, 1978.
[4] W.B. Jones, 0. Njastati and W.J. Thron, Two-point Pad6 expansions for a famzly of analytzc
functzons, J . Comput. Appl. hlath. 9 (1983), 105-123.
[5] Ur.B. Jones, W.J. Thron and H. Waadeland, A strong Stzeltjes moment problem, Trans.
Arner. Math. Soc. 261 (1980), 503-528.
[6] A. Sri Ranga, Another quadrature rule of highest algebrazc degree of preczszon, Nurncrische
Math. 68 (1994), 283-294.
[7] A. Sri Ranga, Symmetrzc orthogonal polynomials and the assoczated orthogonal L-
polynomzals, Proc. Amer. Math. Soc. 1 2 3 (1995), 3135-3141.
N T~IATHEMATICS,
D E P A K ~ X I EOF LNIVERSITYOF SOFIA,BLVD. J , u ~ ; sBOUCHER
5. 1126
SOFIA.BULGARIA
E-mazl address: borisQfmi . m i - s o f ia.bg
C. F. Bracciali
ABSTI~A~ We
T .consider strong Stieltjes distributions $ ( t ) that satisfy
t" d q ( p 2 / t )= - ( P 2 / t ) & d * ( t ) ,
for t t ( a ,b ) C (0, m). 2w integer, 0 > 0, and the orthogonal Laurent poly-
nomials related to ~ ( t )We . study some relations that appear between these
distributions when 2w is odd and 2w is cven. Further results concerning an
example arc given.
1. Introduction
Connections between the theory of moments. continued fractions, Pad6 approxi-
rnants and orthogonal polynomials are well known (see, e.g., Akhiezer [Ak] and
Chihara [Ch]).Analogous relationships can be found in the study of two-point Pad6
approximants, continued fractions that correspond to two series. strong moment
problems. and orthogonal Laurent (or L-)polynomials (see Cochran and Cooper
[ C C ] ,and Jones et al. [ J N T 2 , J T W ] ) . In this work we consider strong Stieltjes
distributions, the related continued fractions and orthogonal L-polynomials.
Let 0 < /? < b 5 co,a = p L / b and $ ( t ) be a real, bounded and nondecreasing
function on (a, b ) , with infinitely many points of increase in (a, b ) such that the
moments
h
(1.1) p = ( ni = 0, *I, *2, ... ,
all exist. With these conditions $(t) is a strong Stieltjes distribution function on
c
( a ,b) (0, oc).
We consider the sequences of mouic polynomials of degree n,B: ' ( z ) , n 2 1.
r = 0, *I, 5 2 , ... , defined by the conditions
@ 1999 A r n r r i c a n M a t h e m a t i c a l Society
71
72 C. F. BRACCIALI
The main properties of these polynomials can be found in a paper by Sri Ranga
[ S r l ] . These polynomials are related to the orthogonal Laurent polynomials studied
by Jones et al. [ J N T l , J N T 2 , J T W ] .
From the results given in [ S r l ] , it is easy to prove that the polynomials B:)(z).
>
n 0, r = 0, 11, k 2, ... , satisfy the three term recurrence relation
and
for n = 1 , 2 , ... and m = 0, +1, k 2 , ... , are positive. Hence there exist rational
P1 P2
fi+ll+_i+,+...+- Zn+r + lower other terms
Pn+r-1
z z zJ
B?) ( z ) +
- p P 1 - p P 2 z - . . . - p-n-rzn-T-l higher other terms
when expanded accordingly. These rational functions form the two-point Pad6 table
for the two series.
From the three term recurrence relation (1.3), we see that the polynomials
B?' ( z ) , n >
0, r = 0, 1 1 , 1 2 , ..., are the denominators of the nth convergents of
the M-fraction
The polynomials A ; ) ( Z ) are the numerators of the nth convergents of this continued
fraction.
RELATIONS BETWEEN CERTAIN STRONG STIELTJES DISTRIBUTIONS 73
The coefficients a:) and PC), n > 1, r = 0, f l , f 2 , ..., can be obtained from
the moments p,, r = 0, f 1 , f 2, ... using the quotient difference relations, for n 2 1,
with crj') = 0 and pir) = p r / p r l , for all values of r , (see McCabe [Mc]).
We now set Y~:)( -
- pn + F'rom the equations (2.1), we see that
while
and
and
These relations can also be derived from the algorithms given in a paper by Jones
and Magnus [ J M ] .
74 C. F. BRACCIALI
Hence, if the coefficients a:) and ~$1"are known for a fixed r then, from (2.1).
we can calculate the coefficient ~ 1 ( +' I ) as follows
p!r+l) = + &1
We can calculate the coefficients a!,'") and p;;'"), > 2, using (2.4), then
for n
(2.5)
&+I' = (T) Pn + %+I
(7)
=P
d:),,) + a:;i1
(Tj
an (T)
+ a:? ' +
The coefficients a:-'' and @c--ll', for n > 2, can be calculated similarly from
and
and
(4.2) d q e ( t ) = t + U d @ ~ ( t ) , t t ( a ,b ) ,
zf and only if
For the case 2w0 odd, then in ( 3 . 2 ) , J = jO = - w O . While for 2w' even. in
( 3 . 4 ) ,j = j e = -w". Ram the above theorem wO = we + i,
hence j" = j'. We set
J = j 0 = j e , to simplify the notation.
If we set
d+;(t) = tJ&,ho(t) and d$,"(t) = t J d + " ( t ) ,
then. from the properties (3.6) and ( 3 . 7 ) . we obtain
+
d$$(t) is an ~ " [ w " j , P, b] = S J [ 1 / 2 ,0,b]
and
d$$ ( t )is an S" [we j , + p, b] = s"[o,
0,b].
While from ( 4 . 1 ) and (4.2), we see that
,o(j)
where l k l l = -- n+llp 1, n > 1 and I!:) = 1. We need to calculate the coefficient
1P - 1
l p ) from @ J ) .
On the other hand, if we consider the relation
t
then the coefficients pi('), u ~ y !,B$)
, and a:?!, n > 1 , are related by
,o(j)
where l g ) = - n+llP + 1 , n > 1 and 1 f ) = 1.
(3
ln-l 1 +
5 . Example
T H E O R E5.1.
M For the S"w,@, b] distribution defined by (5.1), and for any
w > 112, some of the coeficients a t ) and p;), from the recurrence relation (1.3),
are given by
= p, a , f o r r > j,
- n > r+2w+ 1,
(5.2)
,&) = P,
( T I
an+, = a , f o r r 5 j , n > -r + 2,
where
and
- w, for 2w odd
j = { 1
w , for 2u, even.
It means that some of the coefficients a$) and p::) are constants and equal to
a and p respectively, and they appear in a stair shape in the a - P table. It is easy
to visualize this behaviour from some examples of the a - P table.
RELATIONS BETWEEN CERTAIN STRONG STIELTJES DISTRIBUTIONS 77
For example, for w = 112 and j = 0, some of the coefficients a:) and ,Or)
related to the distribution
I ( I
That is
d,')=P, a:,7 . ) = a , f o r r > - 1 , n>r+3,
P?) = P, a,+l = a , for r 5 -1, n -r + 2.
While for w = 3/2 and j = -1, the distribution
78 C. F. BRACCIALI
That is
d$(t) =
That is
pp = 0, a;:) = Q, forr>-2, n>r+5,
P:;) = P,
(1.1
a,+, = a , for r 5 -2, n > -r + 2.
We now prove Theorem 5.1.
PROOF.
For w = 112, in Sri Ranga [Srl] we find that
While from the equations (2.6), we see that for r < 0 and n > -r + 2 ,
)3:/ =o and =a.
Hence, the result (5.2) holds for i~ = 112.
To show the result for w = 1 , we use the relation
d q O ( t )= ( t + P)d$" ( t ) .
I We set
1
d$"(t) = dt, w" = 112
Jb-tJt-a
and
p:,-l) = p p)
rl+l a, - n 2 3.
From the equations (2.5), (2.6) and the above coefficients & ' ) and a!,'), we
can see that
PA') = P, a:,') = n , for r > -1, n > r + 3,
(r)
=P, a,,+, = a , f o r r < -1,
(T)
n > -r+2,
hence. the result (5.2) holds for w = 1.
To show the result for w = 312, we use t,he relation
(t P ) d q ( f ) .
d$"(t) = - +
t
80 C . F. BRACCIALI
We now set
and
t
d$O(t)) = t+/3d+'(t) =
0'-1' =
Z
p, a:;:) = , n > 3,
where 1 = d m .
Again using (2.5) and (2.6) we can see that
p r ) = p, a:) a, forr>-1, n > r + 4 ,
=
~ 2= )P, crril
= a , for r < -1, n > -r + 2 ,
and then conclude that the result (5.2) holds for w = 312.
Similarly we prove the result for w = 2 , 5 / 2 , 3 , 7 / 2 ,... .
From Theorem 5.1, we see that the M-fraction (1.4),for the distribution defined
>
by (5.1) with w 112, can be written as
The function
THEOREM 5.2. For the S 3 [ w ,P, b] distribution defined by (5.4), some of the
coeficients a t ) and P?), from the recurrence relation (1.3), are related by
for r odd.
The a - P table for the distribution (5.4) is given by
where
QIP
(5.8) lo = 1, ll = -
lo 1 +
2a1' + 1 , and 1,=1+
1 + ln-1 , n22.
PROPERTY
5.1. The coeficients l,, n 0, defined by (5.8) or (5.9), satisfy
PROOF. For n = 0, the result follows directly from (5.9). For n > 0, from t,he
relation (5.8), we have
After some simple manipulations with lz+l + ln+21n (L;+, - l ) , we obtain the
result. 0
We now can prove Theorem 5.2.
PROOF. From the relations (2.5), we construct the row with superscript (2) in
the cu - p table using the row with superscript (1) given by (5.10), we then obtain
P1(2) -
- P1(1) + 0 2(1),
and
0); =p p, = Qp,n > 3.
Similarly, using the relations (2.5) we obtain the result for the rows with superscript
(3),(4),(5),... .
From the relations (2.61, we construct the row with superscript (-1) in the
cu - table using the row with superscript (0), given by (5.7), we then obtain
and
(-1) - (1)
P~=P:), a,+, -.,,+,, 7223.
Similarly. using the relations (2.6) we obtain the result for the rows with superscript
(-21, (-3), (-41, ... . 0
RELATIONS BETWEEN CERTAIN STRONG STIELTJES DISTRIBUTIONS 83
In addition, for the distribution defined by (5.4): the coefficients a f i l and at:i,
satisfy the property
We can prove this result by substituting (5.7) and (5.10) in (5.12), that is, for
TL > 1)
Acknowledgement. The author is grateful to the referee for the helpful suggestions,
resulting in improvements to t,he initial version of this paper.
References
[Ak] N.I. Akhiezer, The classzcal moment problem and some related questzons i n analyszs, (trans-
lated by N. Kemmer), Oliver k Boyd, Edinburgh/London, 1965.
[BMS] C.F. Bracciali, J.H. McCabe, A. Sri Ranga, O n a symmetry zn strong dzstrzbutzons, in Con-
tinued Fractions and Geometric Function Theory, (L. Lorentzen. 0. Njastad, and F . Ronning,
eds.), J . Comput. Appl. Math., t o appear.
[Ch] T.S. Chihara, An zntroductzon to orthogonal polynomials, Gordon & Breach, New York,
1978.
[CC] L. Cochran, S. Clement Cooper, Orthogonal Laurent polynomials on the real line, in Conti-
nued Fractions and Orthogonal Functions, Theory and Applications, (S. Clement Cooper and
W.J. Thron, eds.), Lecture Notes in Pure and Appl. Math., 154, Marcel-Dekker, New York,
1994, pp. 47-100.
[CM] A.K. Common, J.H. McCabe, The symmetric strong moment problem, 3. Comput. Appl.
Math 67 (1996), 327-341.
[JM] W.B. Jones, A. Magnus, Computatzon of poles of t w o - p a n t Pad6 approximants and thew
lzmits, J . Comput. Appl. Math. 6 (1980), 105-119.
[JNTl] W.B. Jones, 0. Nj&tad, W . J . Thron, Two-point Pad6 expanszons for a famzly of analytzc
functzons, J . Comput. Appl. Math. 9 (1983), 105-123.
[JNTZ] W.B. Jones, 0.NjLstad, W . J . Thron, Orthogonal Laurent polynomials and the strong
Hamburger moment problem, J . Math. Anal. Appl. 98 (1984), 528-554.
[JTW] W.B. Jones, W.J. Thron, H. Waadeland, A strong Stzeltjes moment problem, Trans. Amer.
Math. Soc. 261, (1980), 503 528.
[Mc] J.H. McCabe, A formal extension of the Pad6 table to include two-poznt Pad6 quotzents, J .
Inst. Math. Appl. 15 (1975), 363-372.
[Srl] A. Sri Ranga, Another quadrature rule of hzghest algebrazc degree of preczsion, Numer. Math.
68 (1994), 283-294.
[Sr2] A. Sri Rangs, Companzon orthogonal polynomzals, J . Comput. Appl. Math. 75 (1996), 23
33.
[SAM] A. Sri Ranga, E.X.L. de Andrade, J.H. McCabe, Some consequences of a symmetry i n
strong dzstrzbutions: J . Math. Anal. Appl. 193 (1995), 158-168.
[SB] A. Sri Ranga, C.F. Bracciali, A contznued fractzon assoczated with a special Stieltjes functzon:
Comm. Anal. Theory of Contin. Fractions 3 (1994), 60-64.
[SM] A. Sri Ranga, J.H. McCabe, O n pairwise related strong Stieltjes dzstrzbutions, Det Kongelige
Norske Videnskaders Selskab 3 (1996), 3 12.
DEPARTAMENTO DE C I ~ N C I A
DES COMPUTA~AO
E E S T A T ~ T I CIBILCE-CNESP,
A. 1505,1-000,
SXo J o s i DO RIO PRETO,S P , BRASIL.
Current address: School of Mathematical and Computational Sciences, University of
St Andrews, North Haugh, St Andrews, KY16 9SS, Scotland, UK.
E-mazl address: cleonicet3dcce. ibilce .unesp. br
Contemporary M a t h e m a t i c s
Volume 236, 1999
1. Introduction
I
Let a be a general distribution function on [0,co). We consider quadrature
formulas of the form Ckl
,A, f (x,,) with positive weights A ,, > O and with real
knots x,, > 0. If these quadrature formulas are exact for all integrals I , ( f ) =
h,""f (x)da(x) where f is an arbitrary Laurent polynomial of the form C y = P pa,xJ
where p = p(n) and q = q(n) are nonnegative integers such that p + q = 2n- 1, then
these are called n-point Gauss type quadrature formulas. If we want to investigate
the behavior of these quadrature formulas as n + m with both p ( n ) and q(n)
tending to co, then it is natural t o consider integrals ST
f (x)da(x) where f (z) =
fkz%uch that xr="=,
f k zk and C p = l f P k z k represent entire functions. This
class generalizes in a natural way the Laurent polynomials.
In this paper we shall give rates of convergence for these n-point Gauss-type
quadrature formulas. In Section 2 we introduce the necessary notation and defini-
tions. The main results are introduced in Section 3 and in Section 4 we apply these
general results to the special case when d a ( x ) = xa exp{-(x? +
x-Y))dx, a E IR
and y E (112, co).
85
86 A. BULTHEEL, C. D~AZ-MENDOZA,
P. GONZALEZ-VERA, AND R. ORIVE
2. N o t a t i o n a n d definitions
Let a ( x ) be a distribution function on [0,m ) , i.e. a real valued, nondecreasing
function with infinitely many points of increase such that the moment function
Let {x,,);=, be the zeros of Q,, , then there exist positive weights {A,l,)~=:L=l
such
that
>
where IIk ( k 0) denotes the space of polynomials of degree k at most. For p and
q nonnegative integers, we write for subspaces of Laurent polynomials
f + (2) and f -(1/z) being entire functions. This class of functions f is clearly the
most natural generalization of the Laurent polynomials.
We now introduce two functions R, and r, which we shall use to control the
size of respectively the largest and smallest node x,, as n + m. These R, and
r, are two arbitrary functions defined on [I, m ) which satisfy the following condi-
tions: R, > 0 and r, > 0, R, is monotonically increasing and r, is monotonically
decreasing, and
(2.5) X'X
lirn R, = cx and max{x,,, : 12 j < n ) 5 R,,
(2.6) lim r,
x-X
=0 and min{x,, : 1 <j < 7%) > rTL.
Thus xjn E [ r n , R n ] ,j = 1, . . . ,n,. n E N.
In the rest of this paper we also assume that the sequences ~ ( n and
) q(n) are
both nlonotonically increasing and that they satisfy
lim p(n) = lim q(n) = m.
TZ"X n-x
lim p(n)
-- =8 E (0, I ) .
2n n i x
Hence
and
Furthermore we introduce
lldn)
- ^in
p ( 0 ) = lim -
n+m Ra
2(1-8)
where
and Q, the nth monic orthogonal polynomial for x-p(")da(x). We also need the so
called Christoffel function of degree m for the distribution da,(x) = x-(p(")+')da(x)
a s given by
IP(z)
K m ( z , a n )= sup - --
12
IIPII~'
where IIPII: = I P ( x ) ~ ~ ~ ~ , This
( x ) .allows us t o define
For more information and properties on several of the notions that were in-
troduced, we refer to [Mha96]. There one can also find in greater detail several
(polynomial) results that are related to some of the problems we discuss in this
paper for the case of Laurent polynomials.
RATEOFCONVERGENCEFORQUADRATURE 89
3. Rates of convergence
In this section we give the main results about the convergence o f the quadrature
formulas I , ( f ) , n = 1 , 2 , . . . as defined in Section 2. T o understand better the
technical aspects, we start with two lemmas which give the relation between the
parameters p, T and a . They are basically contained in [Lub83].
LEMMA3.1. Let f +(a) = x,& f;zj be an entire function.
a ) Suppose that for some q f > 0 and c+ > 0, R , < (c+x)'Jffor suficiently
large x . Then
i ) p ( f + ) < l / q + =+ a R ( f + ;s ) = 0 , V s > 0
4
+ a R ( f t ; s ) < (c+s e r ( f i ) / q + ) ,Vs > 0
+
ii) p ( f +) = l/q+
b ) Suppose that for some q: > 0 and cf > 0, R , > (c;x)qf for suficiently
large x . Then
i ) p ( f + ) > l / q ; + a ~ ( f +s ); = cm,V s > 0
ii) p ( f + ) = l/q: + o R ( f i ; s ) > ( c f s e r(ft)/q:) 4:
,Vs > O
REMARK3.1. I f R , = ( K X ) ~ 'then
, equality holds in a)ii) and in b)ii),so that
a R (f + ; S ) can be estimated exactly in terms o f T ( f +).
T h e following lemma gives the analog for entire functions in l / z .
LEMMA3.2. Let f - ( z ) = C,"=lf ~ z - jbe an entire function.
a ) Suppose that for some q- > 0 and c- > 0 , l l r , 5 (c-x)q- for suficiently
large x . Then
i) p(f-) < l / q + a , ( f ; s ) = O , V s > O
b ) Suppose that for some q, > 0 and c l > 0 , l l r , > (c,x)q; for suficiently
large x . Then
i ) p ( f - ) > l / q , + a , ( f P ; s ) = m, V s > 0
-
PROOF.
We have to prove that
Cx
3=1
3
We shall only prove the second equality since the first one is proved in a completely
similar way. By the dominated convergence theorem [Rud74, Theorem 1.341, it
is sufficient t o prove that CEl 1 f J - 5 - 3 1 is a Lebesgue-Stieltjes integrable function
with respect t o d a ( x ) on [O, m). Since f - ( z ) = C,:, fJ-zPJ is holomorphic in
?t \ (0) where 6 = C U {m} is the extended complex plane, we can write
Thus we have t o check that the series C,:, f J - I O ( x p J )is absolutely convergent.
Now by c) in Lemma 3.3, one has
l If;
l f ; l I I a ( ~ - ~ )= l l I u * ( J j ( x - 95 lf;lr;2(J)
We are now in a position to state the first result concerning the estimate of the
rate of convergence of the sequence I , ( f ) , n = 1 , 2 , . . . of Gauss-type quadrature
formulas.
T H E O R E3.5.
M Let f ( 2 ) be a function satisfying (2.4) and suppose that ~ ( 8 )
and I / $ * ( $ ) as given b y (2.12) and (2.13) are both finite. Assume also that
a r i ( f T ;&)+(O) < 1 and a , ( f p :$ ) / $ * ( H ) < 1. Then
where
En(f) = En(f+)-En,(f-)
C
f:En(x3) +
E n ( f )= f[En(~-~). C
3=dn)+l 3=p(7~)+1
We study the convergence of both series in the right-hand side separately. First
we consider the series x3Eq(n)+lf +E,(xj). From the definition of u ( x ) in (2.7),
Lemma 3.3, and the monotonicity of R,, it follows that
sup 1 f j t l l " l I a ( s ~ )- I T L ( d ) I 1 /=
~ sup 1 ~ ~ ~ I I ~ ( ~ ) ( Z IJ 1 ), ( x J ) l l ' ~
-
3>dn) 3>dn)
SUP I~;~/~(R:(~)+R;)~/J
3>4(71)
SUP 1
< 21/q(7L)
-
3>~(7~)
Then
liln
n-x
<
sup ~ ; I ~ / J I E ~ ( Z J ) ~ "lim
-I
1 c - x
sup 1f;i1~'~~~!~~
J>~(TL) 3>q(7~)
( 31
= lim sup ( f,+ll/%- -
11-X
3>dn) Z(l-0) R-L ( 1 - 8 )
We now proceed as in the proof of the previous lemma. By hypothesis, we know that
a R ( f + ;&)+(0) < 1. Take p > 1 such that 77 = g(0) := p o ~ ( f +&)$(0)
; <
1, then there exists an no = n"(p) E N SO that for any n > no by (3.1)
and consequently
I,
n-x
(2j=q(n)+l
5 77l-O
As the left-hand side in the last inequality does not depend on the parameter p, we
can replace p by 1 so that
On the other hand, by assuming that a,( f -; $)/+*(8) < 1, and proceeding as
above, it can be shown that
On the other hand, from the error expansion for In( f ) (see [BDMGV098])
E , ( x ~ )) 0 , 'v'k E Z.
Hence
(3.4) 1. l
E n ( f ) 2 fpjn)+l~ , ( ~ - ( p ( n ) +) l+) f z n ) + l ~ , ( ~ q ( n ) +
Let us denote b y L n ( & x ) the Hermite interpolant from A-,(n),q(n) for
the function 4,at the nodes XI,, . . . ,x,, of the quadrature formula. B y
[BDMGV097b],it follows that
E, ( x k )= I , ( x k - L, ( x k ;x ) ) .
Therefore,
n i x
iIa(f) - 1 , ( f ) 1 ~ ' ~ ~ < m a ~ { q +rl-)l~ (q+, q- as above)
n-x
''>
IIa(f) - ~ n ( f ) l ~ / ~nla~{77:,77F),
where the parameter a E IR is not relevant for our results. The important parameter
is y E (112, m ) . Let {A,) be a sequence of nonnegative integers such that 0 <
A, 5 2n and satisfying (2.9). Let us denote by h,,(x,w,) the mth orthonormal
d z by h,,(x, w,) the corresponding mth
polynomial with respect to x p X n w 7 ( ~ ) and
monic orthogonal polynomial, i.e.,
+
where h,,,(x, w,) = K,, (wy)xn' lower degree terms.
For our further development, and to make the paper more self-contained, we
recall several results contained in [LLMF95],[BHMF96],concerning the zeros of
h,,,(x, w,) which are needed t o estimate the rate of convergence.
THEOREM
4.1. i) Let { ~ ~ , , ) 3 nbe, ~ the zeros of h,,,,(z,w,). Then. there
exzst posztzve constants R,, T, and K > 1 such that z,, E [Tn/K,KR,,]
where
1 11-7
+
o(n117) and + o(nl'~)
[X
rn = ?B(-,)I
Lw,
with B ( y ) = 6 T(-r+l) and r ( z ) being the Euler Gamma function.
ii) If 0 5 8 < 1 , then
where //PIIz= I P ~ , ( X ) ~ ~ X ~ ~ ~ ~ W ~ ( ~ ) ~ Z .
Let @a,,
(z) be the conformal transformation mapping the exterior of the unit
circle onto \ A,, , preserving the point at infinity:
with T,,(z) = (z - R, ) (z - T,, ). Moreover ga,, (z; m) and gat,(z; 0) are the Green
functions of A,, with pole at infinity and a t zero respectively. Note ga,, (z; m ) =
1% P A , ,(z)l.
For a given Bore1 measure ,LL we let V,(z) = log Iz-tlp1dp(t) be its logarithmic
potential.
With this notation. we can formulate:
4.2. Let d(z, A,,) denote the dzstance from z to A,, . Then
THEOREM
4 z ,A n ) K f 1 ( z , ~<? )I ~ A , , (z)jexp{w,, - 2V,,, (z)}
96 A. BULTHEEL, C. D~AZ-MENDOZA,P. GONZALEZ-VERA, AND R . ORIVE
+ + +
2vp,,( 2 ) x - J x - ~ An log x
{> =Wn; x E supp(pn)
wn; x E itjs u p p ( p n )
with w,, the corresponding equilibrium constant.
Furthermore, it holds V z E C that
en(^; f-,) + 2Vp,, ( 2 ) + [ 2 n - AnIgA, ( 2 ;m ) + An [log lzl + gA, ( z ;0 ) ] = W n ,
where 8,(z; f,) zs the solutzon of the Dzrzchlet problem for the functzon f,(x) =
xY +
x P Y o n An whzch zs gzven by
n-x
1 20
(4.4) .
and r , = ( K ? y- B
- ~ ()7 )
Thus
$(O) = lim -
n-x RL2(1-8)
b y (2.10). O n t h e other hand
rv*( n )
$*(el = lim,L+x
-=
r L(1-8,
' ~ n-x
( ~ )
v*(n)20
l / =
, 1.
with
PROOF. By Corollary 3.9, we only need lower bounds for p(B) and for ~ ( 8 )
which were defined in (2.14) and (2.17) respectively. We have
p(f3) =
- 'Yn
lirn -
l/s(n)
-- with
n+.x. R ~ ( ' ~ 1
7, = 1cX2 A
h:, (x)x-P(")w, ( s ) ~ x .
Thus y, = [~,,(w,)]-~ with K,,(w,) given by (4.2). Now, by (4.4) and Part ii) in
Theorem 3.5, it can be easily deduced that
1
~ ( 0=) 4(eK)lly '
Here K,(z, w,) denotes the Christoffel function associated with x-(p(,)+l) wy (5),
while K, (z, w,) represents the one corresponding t o x-p(")w, (x). For our purposes,
we can use K, (z, w,) instead of K, (z, w,).
Now recall that A, = [?, R,], hence d(0, A,) = F,. Thus, by Theorem 4.2, it
follows
/aAn (0)1 2 n + 1 - p ( n ) e ~ ( n l )i m - o (log ~ z ~ +(uz :,~ )e) ~ n ( ~ . f , )
Kn(0,wy) 5
iTF,
Therefore
It is known that
b -a
(z; m ) - log lzl) = - log Cap [a,b] = - log -
lirn {g[a.61
Iz/-= 4 '
(log Cap is the logarithmic capacity) and since g ~ (z; , 0) = gA,l (1/z; m), (we de-
note A;' = {z : l / z E A,})
lirn {ga,,
z-0
(l/z; m ) + log Izl} = - log Cap A;'
which gives
-
R, - f,
(4.6) 2-0
lim {gA, (z; 0) + log lzl} = - log -.
4R, F,
By the definition of O,(z; f,), we have
where
~ ( y :=
1 qy+i/2)
J;; r'(7 + 1) =
) lim. ~ , ( y )= -
n-cx
1
;l d-dx.
X,
x(1 - 2)
Therefore
lim & ( O ; f,) -
-
-W?)- = -B(?)l_H'
B
n-x RZ l i n ~ , ~(RnFnlY
-~
or equivalently
(4.7) n-x
lim B,(O, fY)TI; = -B(?).
Thus, by (4.5) and ( 4 . 7 ) , we find after some calculations that
-
- r&g
28
1
lim 2- 4 K e l / r
TL-x &(o, uY)l/p(n) '
k=O k=O
be given, with f l ( z ) and f 2 ( l / z ) entire functions. Then consider integrands f ( z ) =
) , that
f l ( ~ ) f 2 ( ~SO
X T
3 =0 j=1
where
Thus
and
X
:
f = "' x
/=I
(@)'
(21)!(21 + k))!
Now by ( 4 . 4 ) , we can write
2(1 - 0 )
R, = ( K ' x ) " ~ , with K' = K Y -------
?W?) . >I
RATEOFCONVERGENCEFORQUADRATURE
and
while
G IIn(f)- ~ ~ ( f ) / <
n i x
l /rnax
~"
2(1 - 8)
When considering the so-called balanced situation where p(n) = n for all n (see
[JNT83]),then obviously 8 = 112 and we then have
where the last equality holds because of the relation between K , K' and K"
References
[BDMGV097a] A. Bultheel, C. Diaz-Mendoza, P. Gonzblez-Vera, and R. Orive, Quadrature on
the half lzne and two-poznt Pad6 approximants to Stieltjes functzons. Part II:
Convergence, J . Comput. Appl. Math. 77 (1997), 53-76.
IBDMGVO97bI A. Bultheel, C. Diaz-Mendoza, P. Gonzalez-Vera, and R. Orive, Quadrature o n
the half line and two-point Pad6 approximants to Stzeltjes functzons. Part III:
T h e unbounded case, J . Comput. Appl. Math. 87 (1997), 95-117.
A. Bultheel, C. Diaz-Mendoza, P. Gonzilez-Vera, and R. Orive, O n the conver-
gence of certazn Gauss-type quadrature fonnulas for unbounded intervals, Math.
Comp. (1998), Accepted.
M. Bello-Hernandez and A. Martinez-Finkelshtein, Zero asymptotics of Laurent
orthogonal polynomials, J . Approx. Theory 85 (1996), 324 342.
W.B. Jones, 0 . Njbstad, and W . J . Thron, Two-poznt Pad6 expanszons for a
family of analytic functzons, J . Comput. Appl. Math. 9 (1983), 1 0 5 1 2 4 .
G. L6pez-Lagomasino and A. Martinez-Finkelshtein, Rate of convergence of two-
point Pad6 approxzmants and logarithmic asymptotics of Laurent-type orthogonal
polynomzals, Constr. Approx. 11 (1995), 255-286.
D. Lubinsky, Geometric convergence of Lagrangzan interpolatzon and numerzcal
zntegration rules over unbounded contours and zntervals, J . Approx. Theory 39
(1983), 338--360.
H.N. Mhaskar, Wezghted polynomial approxzmatzon, World Scientific, 1996.
W . Rudin, Real and complex analysis, 2nd ed., McGraw-Hill, New York, 1974.
DEPARTMENTOF COMPUTERSCIENCE,K.U.LEUVEN,LEUVEN,BELGIUM
E-mail address: Adhemar. BultheelOcs .kuleuven. ac .be
URL:http://www.cs.kuleuven.ac.be/-ade/
1. Introduction
The main motivation for this paper is the basic observation that was made by
Fischer and Prestin [FP97]in the case of polynomials on the real line.
This observation is that a reproducing kernel in a reproducing kernel Hilbert
space has the property that it is a localized function. If 'H is a Hilbert space of
functions defined on some set X c @, then it is called a reproducing kernel Hilbert
space if there is a kernel function k , E 'H (w E X is a parameter) such that
(f,k,) = f(w)for all f E 31,wE X. It has the remarkable property that it gives
the solution of infpEX{llf 11 : f (w)= 1,w E X) namely f = k w / k w(w). This can
be interpreted as follows: without the constraint f (w)= 1, the solution would be
f = 0. The constraint forces f to take the value 1 a t w, but the optimization
problem implies that in X \ {w),the function f is "as small as possible". If the
space 'H consists of "smooth" functions like polynomials or rationals, then f has
typically a "bump" in z = w and much smaller oscillations away from z = w.
Fischer and Prestin exploited this fact t o construct polynomial wavelets. Con-
sider the space of polynomials on an interval (possibly infinite) [a,b] & R equiped
b
with the inner product (p, q) = Ja p(x)q(x)dp(x) where p is a finite positive mea-
sure. The subspaces C, of polynomials of degree at most n form a nested sequence
Lo C C1 C . . . C C, c . . . of subspaces. If f, is the projection of some f E L2(p,EX)
onto C,, it is clear that for increasing n we shall get better and better approxima-
tions for f , i.e., more and more details of f will become visible in the projections
f,. One may say that the C, represent different scales or resolution levels of the
space LZ(P,EX).
More generally, in multiresolution analysis, one decomposes a function space in
several resolution levels and the idea is to represent the functions from that function
space by a low resolution approximation and adding t o it the successive details that
lift it t o resolution levels of increasing detail.
Wavelet analysis couples the multiresolution idea with a special choice of bases
for the different resolution spaces and for the wavelet spaces that represent the
difference between successive resolution spaces. Thus if V, are the resolution spaces:
Vo c V1 c . . . , then the wavelet spaces are given by the differences: W, = VS+leV,.
By their definition, the functions in W, will automatically be band-limited in
their detail information. This means that the basis functions for these W, (which
are called wavelet functions) will have Fourier transforms that mainly live on a
relatively small interval i.e., they are small outside that interval. Moreover, for
true wavelet functions, one wants the functions themselves to live mainly on a small
interval as well. Then a local change in the function f will only have considerable
effect on the components along the basis functions that have their main part in
the neighborhood of the change and it will much less affect the other components.
Conversely, changing one component will only have a local effect on the function.
The kernel functions which we described above do have this wavelet-like be-
havior and the problem is thus whether it is possible t o give a basis of reproducing
kernels that span the successive resolution spaces and wavelet spaces. The answer
to this question is affirmative and is completely analyzed in [ F P 9 7 ] for L2(p,R)
where the subspaces are spanned by polynomials.
In this paper we give two modifications of this idea. First, we consider complex
function spaces where orthogonality is with respect to a measure on the complex
unit circle and in the second place we generalize the space of polynomials t o the
space of rational functions whose poles are in a certain prescribed set. To be more
precise, C, will be the space of functions of the form p , / ~ , where p, is a polynomial
of degree at most n and .rr,(z) = n:=l(l - h,z). We assume that all the cyk are
inside the open unit disk. A parallel theory could have been developed where all
the a k are on the unit circle, but we shall not do this here. A similar discussion
for rational wavelets on the real line with all the poles on the real line was given in
[BGV98].
The results presented in this paper rely on two pilars: the theory of wavelets,
which has to be generalized at different places, very much like it was done in [ F P 9 7 ]
and the theory of orthogonal rational functions on the unit circle.
The plan of the paper js as follows. First we introduce the rational function
spaces we shall work with and give the basic ideas of multiresolution analysis. In
WAVELETS BY ORTHOGONAL RATIONAL KERNELS 103
Thus
or equivalently
and
or more compactly: f2n = p2np2n = pnpn+qn@,. In the next step one transforms
pn into pn12and qn/2 etc. The synthesis or reconstruction algorithm goes in the
opposite direction and generates p2nfrom pn and q,. Thus it is important that
we have an efficient algorithm t o do the transformations back and forth between
P2n ++ n ).
{ ~ 7 9n
Such a transformation should preferably be unitary for the ease of inversion
and for the numerical stability. Thus the basis {pan$ : k = 0 , . . . ,272) as well as
WAVELETS BY ORTHOGONAL RATIONAL KERNELS 105
4. Reproducing kernels
The following properties concerning reproducing kernels are well known (see
for example [AroFiO, Mes62, Don741)
THEOREM 4.1. W e consider an m-dimensional subspace K of Cn . Assume that
its reproducing kernel is denoted as k(z, w). Then
1. For any orthonormal basis {el, e2, . . . , e m ) of K , the reproducing kernel is
given by
m
k(z, w) = ek(z)ek(w).
k=l
2. For any set of distinct points {wl, w2,. . . , w,) among the points of analyt-
icity for K,
is positive semi-definite.
3. The orthogonal projection o f f E 13, onto K is given by
P K =~ ( f ( z ) ,k(z, w)) .
4. For any point w where it makes sense,
inf (11 f 1 ' : f (w) = I) = [k(w,w)]-'
f€K
and the minimizer is f (2) = k(z, w)/k(w, w)
Also the following lemma is easy to verify.
LEMMA4.2. Let {40,. . .4,) be the orthonormal basis for C, as introduced
above, then for any set of distinct points x = { x o , x l , .. . ,x,) which are points
of analyticity for functions i n C,, the matrix
(4.1)
is regular.
@n(x)
~ O ( X O. ). .
4n(50) . . .
40(xn)
4n(xn) I
PROOF.This is obvious because the 4k form a Haar system in the polynomial
as well as in the rational case. See Davis [Dav63, Chap. 11, Sect. 2.41. 0
FIGURE1. The real (left) and imaginary part (right) of the func-
tions kn(z, w) = C : = o ( z / ~ ) k (top) and ln(z,w) = C ~ , + ~ ( Z / W ) ~
(bottom) for z = eie and w = 1, plotted as functions of 8. Note
that the real parts are even functions, while the imaginary parts
are odd functions.
sin $ ( 8 - w)
, 2 = eie, = eiw. ,
k=-n
Hence these kernels are complex exponentials modulated by the Dirichlet kernel,
which is localized around 6 = w. This can be verified in Figure 1. For the spaces
Kn = Can 8 C,, the reproducing kernels are given by l,(z, w) = kzn(z,w) -
kn(z, W ) = ~ ~ , + ,+k(z)+k(w). In our example this is just (zE)"+' C;:
The plots for this trivial example are given in the bottom rows of Figure 1. It
WAVELETS BY ORTHOGONAL RATIONAL KERNELS 107
is immediately observed that these oscillate more but otherwise have properties
similar t o the kernels for C,.
Although we have plotted the kernel I, for the space K, in the previous example,
it is not as trivial as for the k, t o conclude that there are numbers {yo,.. . ,yn-1)
such that {l,(z, Y k ) ) ; ~ i will form a basis for K,. Indeed, the system {+k)gn+l
will not be a Haar system in general. Before we investigate this any further, we
shall first recall several properties of the ORF basis {dk) in the next section.
For any function f we set f,(z) = f ( l / F ) and if f, E L, (it should be clear from
the context what space L, is involved) we define f; (z) = B,(z) f. (z).
Obviously L, = span{Bo, B1,. . . ,B,). If we write 4, = anoBo+anlB1 +. .+
.
a,,B,, then the orthonormal 4, are uniquely defined if we impose the normaliza-
tion n, = +;(a,) > 0. (Note that n, is the a,, from the previous expansion.)
We shall use the notation 4, to mean this uniquely defined orthonormal rational
function.
The ORF satisfy a recurrence relation that generalizes the Szego recurrence
relation for orthogonal polynomials on the unit circle. In fact the polynomials form
a special case obtained by setting all a k = 0.
THEOREM 5.1. For the orthonormal basis functions 41, = ~ k B k+ . . . , K k > 0
in C,, a recursion of the following form exists
where O 5 k I n - 1.
108 ADHEMAR BULTHEEL AND PABLO GONZALEZ-VERA
The constant en > 0 defines the modulus and i t is given as the positive square
root of
where
where the $k are the ORF and the $k are the so called associated functions or
functions of the second kind.
The successive approximants are interpolating functions for the Riesz-Herglotz
transform of the measure:
with h(z) analytic in Izl > 1 while h(z) is analytic in lzl < 1.
Several variants of this continued fraction can be obtained for example by
applying several equivalence transforms or contractions. Alternatively, they may be
WAVELETS BY ORTHOGONAL RATIONAL KERNELS 109
derived from the recurrence relations obtained by eliminating some terms from the
coupled recursions (5.1-5.2) given in the previous theorem. For example, one could
eliminate c$i-l from (5.1) which would give a relation between c$,, 4: and q5n-l.
We shall not mention all of them. Several possibilities are given in [BGVHN98b].
The most prominent one of the alternatives is some variant of a three-term
recurrence relation, which can be written in the form
will give f k = c$k.The other solution is given by functions of the second kind.
In the polynomial case, the continued fraction corresponding to this recurrence
is related to a n M-fraction and the corresponding recurrence for the q5: is related t o
T-fractions. The convergents of these continued fractions are related t o two-point
Pad6 approximants. In the general case these become multipoint Pad4 approxi-
mants and the continued fractions are sometimes called MP-fractions. Many more
of these continued fractions related to ORF are described in [BGVHNgga].
As we have seen, it is not exactly the ORF basis that we want t o use, but
the basis of reproducing kernels. Of course, once we know the ORF basis, we can
compute the reproducing kernel basis, but it would be much more interesting if
we had a recurrence directly for the reproducing kernels. Such a recurrence does
indeed exist. In fact it is even simpler to formulate.
THEOREM
5.2. Let kn(z, w) be the reproducing kernels for the spaces C,. Then
and
where
However, the recurrence for the kernels only holds for z, w @ T. For w E T, it
is clear that pn(w) E T, so that 1 - l p , ( ~ ) =
( ~0, and the recurrence breaks down.
In the sections to follow, we do need the kernels for values of w E T,so that the
recurrence of the previous theorem is useless.
Like in the polynomial case we have a certain equivalence between the existence
of a recurrence relation and the orthogonality (Favard's theorem) [BGVHN92]and
on the other hand between the recurrence relation and the existence of a Christoffel-
Darboux type formula (see [BreSO] for the real line). In our case the Christoffel-
Darboux relation takes the following form.
THEOREM 5.3 (Christoffel-Darboux formula). Let {+k)c=o with E Ck\Ck-1
be an orthonormal basis for Cn, then the reproducing kernel kn(z, w) for C, satisfies
For this paper, the relation with quadrature formulas will be important. The
zeros of the orthogonal rational functions (ORF) are known t o be in the open unit
disk Dl so they are not suitable for the construction of the quadrature formulas
representing integrals over the unit circle T. We should rather have knots which
are located on T. Such knots are provided by the zeros of para-orthogonal functions.
The para-orthogonal functions are defined as
Qn(z1rn) = &(z) +rn+:(z)
for rn E T, n = 0 , 1 , . . . and the zeros of these are simple and on T . We have the
following rational Szego quadrature formula.
THEOREM 5.4 (Rational Szego quadrature). The zeros of the para-orthogonal
rational function Qn+l(z,r n f l ) are all simple and on T . Let us denote them by
-'
{<n,k)c=o. Moreover, defining Xnk = [k, (Cnk, Cnk)] , with k,(z, w) the reproducing
kernel for C,, then equality holds i n
n,
( f i g )= ~ n k( ~ vf
f nk)~*(tnk)l 1 E Cn.
k=O
Conversely, i f the above equality holds, then the Erik should be the zeros of the para-
orthogonal function Q,+l(z, rn+1) for some rn+lE T and Xnk = [k,(Cnk, tTLk)]-l.
The n + 1 zeros {<nk);=O can also be characterized as follows.
THEOREM 5.5. Let kn(z, w) be the reproducing kernel for C,. Define Cno = w
with w arbitrary o n T and {&k);=l the n zeros of k,(z, w). Then the numbers
{Jnk);=O are the zeros of the para-orthogonal function Q,+l (z, T,+~) with
kn(Enj,Inz) = 6zjkn(Eni:E n % ) ) i : j = 0 : . . . : n.
In terms of the matrix (4.1) this reads
@ @ =A or equivalently @,An@: = I
with An = diag(Ano,.. . , Ann) a diagonal matrix with An, = l/k,(<,,. t,,).ViTriting
@,An@: = I explicitly gives
n
A n k i Z ( ~ n k ) ' $ j ( < n k )623, % : 3-- 0 , . . .
k=O
Because also (4,:4,) = 6,, for 2, 3 = 0 , . . . . n, this means that the above quadrature
formula is exact for the inner product of all basis functions in C,. hence for the
inner prcduct of any two functions in 13,. By Theorem 5.4. this means that it can
only be the rational Szego quadrature formula. Thus the theorem follows. 0
E X A M P L E6.1. Let us consider again the trivial example where all a k = 0 and
where p is the normalized Lebesgue measure on T.It is clear that if we set T,+I =
-1: then Q,+I(z, T ~ + ' )= zn+'-1. Thus its zeros are given by {<,,k = e-"k"/(n+l) }
for k = 0 , . . . ; n. Supposing that f , E 13, is given by f n ( z ) = C 2 = o a n k q 5 k ( z )=
C y = o ~ n j k . n ( zEn?):
r thus
112 ADHEMAR BULTHEEL AND PABLO GONZALEZ-VERA
it follows t h a t the {ank) are given by the Discrete Fourier Transform (DFT) of
{p,,). In other words, the matrix @,(En) becomes a Fourier matrix.
The main question in this respect is: Can we find n numbers {rl,j),"zi
such t h a t
the functions {$,j(z) = l , ( ~ , ~ , ~ ) ) , " form
~ i a basis for K, and if possible, can it
be made orthogonal?
We have t o investigate the matrix
I 1
+ l o ... 4n+l(~,-1)
(7.1) * n ( ~ )=
(YO) . . . Q2n(~n-1)
42,
Using the following lemma, it is possible t o prove that there always exists a set of
points {1J3),"zi
on T which make this matrix regular.
Q ~ +QQFQ,
~ =o
so t h a t rank(QFQ1) = rank(QFQ3). Because Q1 is regular and rank(Q2) = n ,
we find that rank(QFQ3) = n. Now if Q4 were singular, then there would exist
a nonzero vector c E @ I X n such that c Q = ~ 0, hence also cQFQ3 = 0. In other
words, r a n k ( Q r Q 3 ) < n , which is a contradiction. 0
This makes it possible t o prove the following theorem which ensures the exis-
tence of a wavelet reproducing kernel basis (WRK).
THEOREM 7.2. Consider the zeros Ezn = { < z , , ~: k = 0, 1, . . . ,212) of the para-
orthogonal function Q ~ , + l ( z ,r 2 , + l ) for some Tzn+l € T. If we select y , = {yk :
k = 0 , . . . , n - 1) to be any n out of the 2 n + 1 zeros i n J2,, then the matrix Q,(y,)
as defined i n (7.1) will be regular.
PROOF. Since the ordering of the zeros is completely arbitrary, we can
always assume t h a t we select the yk to be the last elements of E2, = :J =
0 , . . . , 2 n ) . Row consider the matrix Q2, = @ 2 n ( < 2 n ) of (4.1) where the evaluation
is in the points of tan, then +,(y,) appears as the n x n right lower part of the
matrix @an. By Lemma 4.2. it follows that taking the block of the first n + 1
WAVELETS BY ORTHOGONAL RATIONAL KERNELS 113
columns in QZn, any selection of k different rows from it will result in a matrix of
+
rank min{k, n 1). Thus the conditions of the previous lemma are satisfied and
the theorem follows. 0
This theorem settles the question of the existence of a WRK basis of the form
{ l n ( z l Y k ) } ~ , ~It. is however not clear how to choose these { Y k ) ~t o~ make
~ the
WRK basis orthogonal, if it is possible at all. The same problem was encountered
for the polynomial case on an interval in [ F P 9 7 ] where an orthogonal WRK basis
could only be constructed for a Chebyshev weight of the second kind.
EXAMPLE 7.1. For some particular cases, it is however possible t o construct
such an orthogonal WRK basis. We give a trivial example. Consider again the
case where all cur, = 0 and where orthogonality is with respect to the normalized
Lebesgue measure of T. In that case
Note that 13, = lIn and K, = C2, @ Cn = Z ~ + ' I I , - ~ Furthermore, for some
distinct qnk, k = 0 , . . . , n - 1, all on T define the functions g n k ( z ) = l,(z, qnk),
k = 0 , . . . , n - 1. In analogy with Theorem 6.1, for these to be an orthogonal basis
of K,, we need the existence of numbers Xnk, k = 0 , . . . , n - 1 such that
The quadrature formula will be exact if and only if the nodes qnk are the zeros of
+
the para-orthogonal polynomial Q n ( z ,T ) = zn T , T E T and the weights are given
by
1 1 1
Thus in this case, one can take for example the nth roots of unity for qnk and the
basis $nk will be an orthogonal WRK basis. Note that here also the matrix Q,(qn)
will be the matrix of a DFT.
We can push this example a bit further in two directions. First by considering
orthogonal polynomials with respect to a rational modification of the Lebesgue
measure. A rational modification of the Lebesgue measure is defined as a measure
p satisfying for some a3 E ID, j = 1, . . . , m
For the polynomial h(z) of degree m we have used h*(z) = zmh, (z) = nyzl
((za,).
where CnPl is the space c.f rational functions associated with the points {G, =
C Y , + ~ + , ) ~ ~Thus,
: . if f , g E K,, then there exist E such that
and therefore
FIGURE 2. The real and imaginary part and the nlodulus of the
functions ks(z, w) (top) and 18(z,w)(bottom) for z = ele and w =
1, plotted as functions of 0.The measure is the normalized Lebesgue
for k = 1,. . . , 8
measure and the C Y ~are chosen as a k = 0 . 9 e ~ ~ " ' ~
and ak+s = 0.9e2(2k+1)a/s for k = 1 , .. . ,8. These a k ' s are plotted
in the figure on the right.
and define
(f,g),,, = f( ~ k ) m .
116 ADHEMAR BULTHEEL AND PABLO GONZALEZ-VERA
it is directly seen that the FLIF {Lnk)t=Oform an orthonormal basis for L, with
respect t o this inner product. Note that the reproducing kernels for L, with respect
to this inner product are given by x y = o Lnj (z)Lnj(xk) = Lnk (z). Thus the L,I, are
reproducing kernels and by the general theory we may say that among all functions
in C, that take the value 1 at xk, they will have minimal norm IILnkllm,,.
If we set p n k ( z ) = k , ( ~ ,x k ) and g n k ( z ) = L n k ( i ) ,then because of the repro-
ducing property of cp,k
In other words, cpnk and Pnk are biorthogonal bases for L,.
Note that we can characterize gnk in another way. We may write gnk =
xy=o cjk)q5, where dk)= [ c c ) . . . cik)] is defined by dk)an= ek where ex =
10.. . 0 1 0 . . .O] is the kth unit vector and a, = @,(x,) is the matrix (4.1). Thus
dk)is row k in the inverse of a,, so that we have
f (i)= C f (~k)?il'n/c(z).
k=O
The 4,k are orthonormal with respect to the discrete inner product
(8.2)
WAVELETS BY ORTHOGONAL RATIONAL KERNELS
This relation allows us to compute p, and q, from p2, and conversely to reconstruct
Pin from Pn and 9,.
&) and p,,, = ( f i n ,
For example, using the relations pnT = (fin, we findG,,)
that
2n
and
In the special case where the v,,k are orthogonal, i.e., when the x,,k are the
zeros [n,k of Qn+l(z, r n + ~ then
), g,,, = (F~,T/(F,,,(E,,~) and we get for the first of
the analysis forrnulas
In general, the are not orthogonal, but if we choose { y , , ) ~ to~ ~be the last n
zeros {[2n,72+~+r)::0', then by definition &,(y,,) = 6,, , r: s = 0, . . . , n - 1. Thus,
the second analysis formula reduces t o
For the reconstruction formulas, it follows from pz,,, = (f, + g,, $in,,)that
Again, if the p2n,r are orthogonal, i.e.. wlien the X n n , k are tlie zeros of
Q ~ , + I ( z ,71n+l),then &n.r = p 2 n , r / ~ 2 n , r ( & n , rand
) , the reconstruction forniula
simplifies t o
10. M u l t i r e s o l u t i o n analysis
Thus far we have only worked with finite dimensional spaces C,. With our
notation V-1 = Lo and V, = C, where n = 2,, we have the nesting
One may wonder what space is ultimately covered wlien s + co.If K is a subspace
of the Hilbert space 7-1, then for the {Vk) to form a MRA of K, one should have
that
Thus for our rational spaces, we should verify in which spaces the system
{Bk)r=o is complete. There are several possibilities.
means t h a t the Blaschke product B ( z ) = n T = l &((z diverges t o zero. Thus the ak:
should not approach the boundary T too fast. H2 represents the Hardy space of
functions in L2 which are analytic in ID.
In classical multiresolution analysis (MRA) of L 2 , several other conditions have
t o be satisfied. Some of these were recently generalized t o what is called a second
generation LIRA by Swelderis [SwegB]. We shall show however that with a gener-
alization of the definition of the shift operator as it was done for the polynomial
case in [ F P 9 7 ] , it is possible t o stay close t o tlie definition of a classical MRA.
First of all we use the term Fourier transform with respect t o the O R F basis.
Thus, if f is in cur function space, then we set 3 ( f ) = { f/)r=C=o
ftwith = (f:4k).
The scaling property of a classical NRA can be interpreted in t,his general
setting by rioting that f E V, ++ s u p p F ( f ) = ( 0 . 1 , . . . , n = 2S), whereas moving
from V, to V,+l practically doubles the support of the Fourier transform.
For the shift invariance property, one should show that the basis functions of
the resolution and the wavelet spaces are generated by shifts of one father function
arid one mother function. To obtain this property, we need an appropriate notion
of a shift. To define such a shift, we fix some TO as a reference point on T and define
a shift over T as the effect of the shift operator ST via
WAVELETS BY ORTHOGONAL RATIONAL KERNELS
In other words
--
( s f f k k = 0, 1 , .. .
0
This definition can be justified as follows. Let us write F ( 0 ) for f (eZe). Then F ( 0 ) is
a 2 ~ p e r i o d i cfunction. In classical MRA for 27r periodic functions, the orthogonal
basis is the Fourier basis 4k( t ) = eZkt, a shift F(B - T ) has the effect that the kth
Fourier coefficient of F is multiplied with 4 k ( ~ ) / O k (=0 )e-'liT . Translating this to
the functions f defined on 'IT with orthogonal basis 4 k ( z ) = z k , then a shift has the
effect that the kth Fourier coefficient is multiplied with 4 k ( < ) / @ k ( = l ) < - l i where
< = e z T .So this is just a special case of our definition.
It is now clear that if we define
n
k=n+l
Thus we have one father and one mother function per resolution level and the
other basis functions are generalized shifts of these. Consequently, these spaces are
shift-invariant, i.e., f E V , + S, f E V,.
The bases used also have a Riesz property, i.e. there are constants A # 0 and
B such that
All~nll I l l f I IBllpnll
where the norms are 2-norms and f E V , is given by f = pncpn. It follows from
(4.2) and P a r s e d ' s equality that
I f llL = I l ~ n @ , " ~ n=
li~
lI~n@,"ll~
when f E W, with f = q n @ ,
This generalized shift operator implies that the functions in the ORK basis
look like (cyclic) shifts of each other. The same holds for the functions in the \VRK
basis. We give some examples.
EXAMPLE 10.1. The cyclic shift is illustrated in Figure 3. Which is again our
trivial example where cuk = 0 for all Ic and the measure is the normalized Lebesgue
measure. If one introduces poles which are close to the unit circle 'IT, then this
will have an influence on the form of the shifted ORK and WRI< functions. For
example in Figure 4, a l = 0.9 while all other cxk = 0. This gives a pole near z = 1.
FIGURE 3. The real and imaginary part of the functions kn(z, w) =
C z = o ( z / w ) k (top) and In(.) w) = ~ : E ~ + , ( z / w (bottom)
)~ for
n = 16. All crk = 0. In Figure A, we took w = 1, and in Figure B,
we took w = e x p ( i 4 ~ / 1 7 ) .
i.e., near 0 = 0. This pole forces the peak near H = 0 to be more pronounced, but
it also implies oscillations of the ORK and WRK functions in the neighborhood of
that pole (here near 0 = 0). These oscillations disappear for w = 1 in kn(z,w) and
l,(z, w), but they do show when w moves away from the pole. This is almost not
visible in Figure 4. However this oscillation effect increases when the pole is given
more weight by making it a multiple pole. This implies a visual distortion of the
simple shift-like property as in Figure 5. where we took all cuk = 0.9, k >
1. The
peak a t 0 = 0 is now more pronounced than in Figure 4, and the shifted versions
are heavily perturbed and do not look much like being shifted now.
WAVELETS BY ORTHOGONAL RATIONAL KERNELS 121
11. S y m m e t r y
We have observed in several of the examples that the real part of the scaling and
wavelet functions were symmetric while the imaginary parts showed an antisymmet-
ric property. We can explain this as follows. Suppose that the measure p is real and
symmetric in the sense t h a t p(S) = p ( 3 ) where S = {e" : 0 5 do 5 d 5 dl I T)
is an arc of the upper half of the unit circle and 3 is the corresponding arc on the
lower half circle: = {z : 2 E S ) . Moreover assume that the poles l / Z k are chosen
symmetrically, that is the poles used in C, are either real or they appear in complex
conjugate pairs. In that case the symmetry that was observed will take place.
THEOREM 11.1. Under the above conditions about symmetry of the measure
and of the poles, the kernels k p ( z , 1) and therefore also the kernels 12.(z, 1) =
lczs+1 ( z , 1 ) - kzS( z , 1) have a symmetric real part and an antisymmetric imaginary
part.
12. C o m p u t a t i o n s
We have given a recurrence for the reproducing kernels in Section 5. It was
however not applicable in its forward form when w E T . There is however a
backward algorithm, which is in fact the generalized Schur algorithm of Nevanlinna-
Pick. One starts by computing the Riesz-Herglotz transform R of the measure p and
from this one computes a Schur function (bounded by 1 in modulus and analytic
in D)
where z is the variable and w a parameter such that r ( w , w) = 0 . Given the points
{ak), one can compute the pk and yk of the recurrence by the following algorithm.
Set ro = r, and compute for k = 1 , 2 , . . . the transforms rk = 7 k ( I ' k - l ) where
Ik =Ilrc o T L k 0 7is ~ ~ by
defined
The 71;'s and pk's are the parameters that are needed in the recurrence relation of
the reproducing kernels. So, given the values of R at the points c u k , it is possible
t o generate the parameters pk and 3k. However, if some (or all) of the points a k
coincide, then one has t o provide the values of derivatives of R in these points and
t o cover the general case, it is rather difficult t o write a tractable algorithm.
A similar argument holds for a Schur-type algorithm t o compute the the pa-
rameters Xk in the recurrence relations for the O R F as given in Theorem 5.1.
In this text, we were mainly interested in experiments to generate plots for
wavelets and see what the influence was of the the introduction of the poles c y k ,
we have chosen t o use the forward recursion for the ORF for some chosen a k ' s and
Xk's and from there compute the reproducing kernels. It is then however not easy
t o find what the underlying measure is. If the XI, are all in T, then the existence
of a positive measure is guaranteed by the Favard theorem [ B G V H N 9 2 ] . There
are several theorems available describing the convergence of the kernels inside or
outside the unit disk, but there are however practically no theorems t o describe the
convergence of the kernels on 'IT.
We do have thc following direct result which is provided by MAt6-Nevai-Totik
[ M N T 9 1 , Theorem 11.
THEOREM 12.1. Let all a k = 0, so that we are i n the polynomial case. Let 4k
be the orthonormal polynomials with respect t o a finite positive Bore1 measure d p
supported on 'IT. Define k,(t, t ) = l4k(t)I2,t E T . T h e n if (Szegii's condition)
J -77
i t holds that
n
lim -= pl(t).
n+m k,(t, t )
for almost every t E 'IT
WAVELETS BY ORTHOGONAL RATIONAL KERNELS 123
We do not know of a proof which generalizes this t o the rational case. We hope
t o provide such a proof in a future paper.
The previous theorem supposes however that logp' is does not vanish on an
interval of positive Lebesgue measure. For examples where Szego's condition does
not hold, we did observe in our experiments that it does not give the convergence
hoped for. Also, convergence is relatively slow, which means that a large n is needed
before convergence occurs, so that one may run into numerical difficulties before
the computation is complete.
There is however an alternative which is not completely proved either but it
turns out t o be numerically performing better. We can motivate it as follows.
First recall that if Szego's condition is satisfied then one may define the spectral
factor
u ( z ) = e x p { & ~t -~zl o g w ( t ) d ~ } t. = e Z e
for z E ID and we normalize it such that a ( 0 ) > 0. Moreover it has a radial limit t o
1 ~p l ( t ) a.e. (w.r.t. Lebesgue measure) on T.
T such that l ~ ( t ) =
For the convergence of the kernels inside the disk we have for exanlple (see
[ B G V H N 9 8 b ] or [BGVHNSO]).
THEOREM12.2. With K,(t) = k,(t, 0)/Jkn0
we ha^
where
Moreover, if Szego's condition holds, and the Blaschke product diverges, then Kn(a)
converges locally uniformly to l / a ( z ) in D.
Thus we may hope that plotting IKn(t)l-"or t E T and n sufficiently large
will give an idea about the underlying weight.
EXAMPLE12.1. The previous idea is used in Figure 6. We took all cyk = 0 and
chose / A l l = /Xzl = 0.4 and all other Xk = 0. We have plotted ks(t, 1) and 18(t.1)
and their modulus. The rightmost pictures show the unit circle with the position of
the a k = 0 and below it. we have plotted the approximation for the weight. In the
figure B below, we plotted k8(t, 1) and 18(t,1) multiplied by the weight (pictures at
the bottom).
Both of the methods given above give bad convergence when Szego's condition
is not satisfied. However. it can be experimentally verified that in those cases we
do have convergence of 1 / [ 2 ~ l S , 1 ~with
] (C, 1) CesBro means
FIGURE6. The real and imaginary part and the modulus of the
functions ks(z, 1) (top) and ls(z, 1) (bottom) where all a,+= 0 and
IX1/ = IX21 = 0.4 and all other Xk = 0. This corresponds t o a
weight which is plotted in the lower right corner. In Figure B, we
plotted the same functions (real and imaginary parts and modulus)
multiplied by the weight.
comparing with the Figure 6, one can notice the influence of the poles. The weight
function is practically unchanged, but the wavelets are much more locallized.
13. Conclusion
We have given a rational generalization of the polynomial wavelets of Fischer
and Prestin [FP97] and translated it t o the orthogonal rational functions on the
WAVELETS BY ORTHOGONAL RATIONAL KERNELS 125
unit circle. The continued fraction-like recurrences and quadrature formulas related
t o these orthogonal rational functions play a central role.
This results in complex wavelets which can be influenced by the choice of poles
near the unit circle.
When one wants to apply this to real 2~-periodicfunctions. t h e wavelets be-
ing complex is a drawback. Our theory can be generalized t o orthogonal rational
functions where the points cur, are chosen in a special way like in [BGVHN94,
126 ADHEMAR BULTHEEL AND PABLO GONZALEZ-VERA
References
N.I. Achieser, Theory of approximatron. Frederick Ungar Publ. Co., New York.
1956.
N. Aronszajn, Theory of reproducing kernels. Trans. Amer. hfath. Soc. 68 (1950).
337-404.
A. Bultheel and P. Gonzalez-Vera, Ratzonal wavelet.5 on the real line, Kumer. Funct.
Anal. Optim. (1998). Submitted.
A. Bultheel, P. Gonzalez-Vera, E. Hendriksen, and 0 . N j k t a d . A Szego theory for
rational functzons. Technical Report TW131, Department of Computer Science.
K.U. Leuven, May 1990.
A. Bultheel, P. Gonzdez-Vera. E. Hendriksen, and 0 . Nj&tad, A Favard theorem
for orthogonal rational functzons on the ~ ~ n circle,
zt Numer. Algorithms 3 (1992).
81-89.
A. Bultheel, P. Gonzalez-Vera, E. Hendriksen, and 0 . Njgstad, Quadrature formulas
on the unzt circle based on ratzonal functions, J . Comput. Appl. Math. 50 (1994).
159-170.
[BGVHN98a] A. Bultheel. P. GonzBlez-Vera. E. Hendriksen, and 0. NjLstad, Continued frac-
tions and orthogonal ratzonal functzons. Orthogonal Functions, hloment Theory
and Continued Fractions: Theory and Applications (W.B. Jones arid A.S. Ranga.
eds.), Marcel Dekker, 1998, To appear.
[BGVHN98b] A. Bultheel. P. Gonzalez-Vera. E . Hendriksen, and 0. Njkstad. Orthogonal rational
functions, Cambridge University Press, 1998, (To appear).
C. Brezinski, A dzrect proof of the Chrzstoffel-Darhoux identzty and its equzvalence
to the recuvence relatzonshzp. J . Comput. Appl. hlath. 32 (1990). 17-25.
P . J . Davis. Interpolatzon and approzamation, Blaisdell. New York. 1963, Reprint:
Dover, 1975.
M.M. Djrbashian. A survey on the theory of orthogonal systems and some open
problems, Orthogonal polynomials: Theory and practice (Boston) (P. Nevai, ed.),
Series C: Mathematical and Physical Sciences. vol. 294, NATO-ASI. Kluwer Aca-
demic Publishers, 1990, pp. 135-146.
W.F. Donoghue Jr., Monotone matrix functzons and analytic contznuatzon,
Springer, Berlin, 1974.
B. Fischer and J . Prestin, Wavelets based on orthogonal polynomzals; hlath. Cornp.
66 (1997), 1593-1618.
P. Gonzblez-Vera. J.C. Santos-Leon, and 0 . N j k t a d . Some results about numerical
quadrature o n the unit circle, Adv. Comput. Math. 5 (1996), 297-328.
H. Meschkowski, Hilbertsche Raume mit Kernfunktzon, Springer, Berlin. 1962.
A. hlat8, P. Nevai. and V. Totik, Szego's extremum problem on the unit czrcle. Ann.
of Math. 134 (1991); 433-453.
W. Sweldens. The lifting scheme: A constructzon of second generatzon wavelets,
SIAhl J . hlath. Anal. 29 (1998). 511-548.
G. Szego, Orthogonal polynomzals, 3rd ed., Amer. hfath. Soc. Colloq. Publ.. vol. 33,
Amer. Math. Soc., Providence. Rhode Island, 1967. First edition 1939.
D E P A R T ~ ~ EOF
N TC O ~ I P U T ESCIENCE.
R K.U.LEUVEN.
LEITVEN.
BELGIUM
E-mail address: Adhemar. Bultheel0cs.kuleuven. ac .be
URL:http://www.cs.kuleuven.ac.be/~ade/
ABSTRACT.In this paper, we discuss recent progress made in the explicit eval-
uations of the Rogers-Rarnanujan continued fraction via modular equations,
singular rnoduli and class invariants. We also introduce empirical processes
which will generate neur evaluations of Rogers-Ramanujan continued fraction.
1. I n t r o d u c t i o n
For Iql < 1. let
denote the famous Rogers-Ramanujan continued fraction, and let S(q) = -R(-q).
In his first letter to G.H. Hardy 13, p. 291. S. Rarnanujan asserted that
Hardy was greatly impressed by the beauty of (1.1) and (1.2). 1% quote here
Hardy's comments on these identities 17, p. 91:
" T h e s e formulas defeated m e completely. I had n e v e r seen anything
in t h e least like t h e m before. A single look a t t h e m i s enough t o
s h o w t h a t t h e y could only be w r i t t e n d o w n by a m a t h e m a t i c i a n of t h e
highest class. T h e y m u s t be t r u e because, zf t h e y were n o t t r u e , n o
o n e ulould have had t h e i m a g i n a t i o n t o i n v e n t them."
The first proofs of (1.1) and (1.2) did not appear unt,il 1929 in a paper by G.N.
Watson 1141. Using the quintuple product identity, Watson first est,ablished an im-
portant relation (known to Ramanujan) between the Rogers-Ramanujan continued
f r a ~ t ~ i oand
n infinite products, namely,
He t,llen expressed the infinite products on the right hand side in terms of Weier-
strass elliptic functions and deduced (1.1) and (1.2) using the periodicity of the
elliptic functions. We remark here, as indicated in a paper by K.G. Ranlanathan
[8], that Watson's final arguement can be simplified if we invoke the well-known
transformation formula for the Dedekind q-function
where S
where the last equality is obtained by substituting T = 5i into (1.4). Solving (1.5),
we complete the proof of (1.1). The identity (1.2) can be obtained in a similar
way by using the transfo-mation formula for q (F).
Hence, with the knowledge
of (1.3), the evaluations of the Rogers-Ramanujan continued fraction at q = ec2"
and -e-" are reduced to substituting T = 5i into transformation formulas of the
Dedekind eta functions ~ ( rand ) TI(+), respectively.
Identities (1.1) and (1.2) are not the only two evaluations Ramanujan gave for
the Rogers-Ramanujan continued fraction. In fact, other special values of R ( q ) and
S ( q ) are scattered in his Notebooks [lo] and Lost Notebook [ll].We give a list
of these continued fractions, dividing them into 3 classes. Besides the continued
fractions listed here, Ramanujan indicates that he could calculate further values of
the Rogers-Ramanujan continued fraction but he does not record them [ll,p. 2101.
Class 1B
n is a complete square
Set 2c,, = 1 + -
an + brl fi and R(ep'"") = - c,.
a,, - bn
We summarize Ramanujan's continued fractions in the following table
EVALUATIONS O F T H E ROGERS-RAMANUJAN CONTINUED FRACTION 129
Class 2
Values of ~ ( e - " f i ) where n - 3 (mod 4)
Class 3A
Values of s ( e p " m ) and ~ ( e - " G ) +
where 5 n and n is squarefree
n = 1 (mod 4)
n -
Class 3B
3 (mod 4)
In the following sections, we will give brief surveys of the continued fractions
listed in these classes. We will then indicate certain empirical processes which
will enable us to discover new values of the Rogers-Ramanujan continued fraction
belonging to these classes.
H E N G HUAT C H A N A N D VICTOR T A N
2 . Class 1
In 1984, Rarrlanat,han [B] gave another proof of t,his evaluation and found the
cornpanion identity
The first cont,inued fraction is identity (1.1) while t>herest of the continued
fractions listed in Table 1 appear on page 311 of Ramanujan's first notebook. The
value for R ( e -'"1 was first proved by R a n ~ m a t h a nusing the Rarnanujan- Weber
class %nval-iunts
Y
and
x
A=I
The proofs of all the continued fractions in Table 1 are now available in a paper
b) Berndt and Chan [4]. The identities associated with R(epJT),R(e-"7) and
R ( e p l " ) were proved using the following modular relation found in Ramanujan's
Yotebook [l,pp. 212-2131:
where
n+b
2c = ---h- 1, n = 60'" and h = 2 + h- A.
a-h
We now introducr a new empirical process which will enable us t o ..guess" all
the continued fractions except R(ep") and R(e-lG") in this class. This process is
illustrat~dwith the identity for R ( r p G " ) . Sct
(2.1)
and
It can be shown that (see [ 5 ] )the value of r:)can be written in t,he form
3. Class 2
The continued fractions listed in Class 2 are recorded on page 36 of the Lost
Notebook. In view of the relation [8]
it suffices t o evaluate just one of the two continued fractions (1.7) ancl (1.8). The
evaluation of s ( ~ c was ~ ~first
) proved by Rarnanathan [9] using the valuc
132 HENG HUAT CHAN AND VICTOR TAN
which he found via Kronecker's limit formula. A second proof of this result was
given by Chan [6] using one of Ramanujan's modular equations [I,Chapter 25,
Entry 621
where
We now introduce the empirical process associated with this class. Let y, be
defined by (2.1). We compute the numerical value of ys and assume that
&+I
Solving the above quadratic equation yields y~ = -, which leads, by (2.2), to
2
the value of s(ep"&).
The continued fractions we discussed so far were known to Ramanujan. We
now show how this empirical process enables us to discover new simple values for
the Rogers-Ramanujan continued fraction which are not found in his notebooks
and Lost Notebook.
If we compute y l l and 71, numerically using the product representation (2.1),
we may assume that
and
and
(3.5)
These values are new but we do not have proofs for them yet since our compu-
tations rely heavily on the assumptions that (3.2) and (3.3) hold.
In the remainder of this section, we will briefly sketch how one can prove (3.4)
and (3.5) using modular equations. This method of proof is new and can be applied
to give a third proof of (1.7).
To prove (3.4), we first quote a modular equation discovered by L.J. Rogers
[121.
EVALUATIONS O F T H E ROGERS-RAMANUJAN CONTINUED FRACTION 133
T H E O R E3.1.
M Let u = R(q) and v = R ( q H). Then
4. Class 3
On page 210 of his Lost Notebook. Rarnarlujan started a table recording t,lic
values of ~(e-"/"') and ~ ( e - " m )for odd integers between I and 15. Only
three continued fractions are actually listed in the table. This table has now been
co~npletedby Herndt, Chan and Zhang [5] using formulas relating s ( c - " ~ ) t o
G I , and known values for G,,. We now describe an empirical process which will
enable us t,o complete Rarnanl~jan'stable (except n = 15 which belong t o Class 2)
without using values of GI,.
Let
(4.1)
Solving the above for d7 and substituting into (4.2). we obtain (1.11).
Next, if we assume that
then
bg = (4 - 6)
(26 5&). +
From t,his arid (4.2). we deduce that
where
We now state some conjectures which we wish to prove in order to provide rigorous
proofs of identities such as (3.2) and (3.3).
3 (iriocl 3)
then th,e real u n ~ ?t: ccLn
CONJECTURE
of Z[&%].
Z[,,"zG].
If n -
5.3. Let n be a positive integer not divistble by 5. I f 7, = 1
(mod 4 ) , then A,, and 6,,. defined i n (4.1), are real units i n the ring class field
3 (mod 4), then 6,,is a real unit in the ring class field of
If Conjectures 5.1 and 5.2 are true. then we would have rigorous proofs of (3.2)
and (3.3) without using modular equations. It also shows wlq we ~ h o o s e1 1 = 11
and n = 19, as these are probably the only two riumbers satisfying the collditions
in these coi1,jectures.
If Corljcctures 5.3arid 5.4are true, then we would have proofs of the identities
listcd in Class 3. illoreover. we will be able to cornputt, many inore contiiiued
fractions. 1% conclude by showing how we compute ~ ~ ' ( e ~1Vemdetermine ~ ) .
the following identities :
HENG HUAT CHAN AND VICTOR TAN
References
B. C. Berndt , Ramanujan's notebooks, part 111, Springer-Verlag, New York, 1991.
, Ramanujan's notebooks, part V , Springer-Vcrlag, Ncw York, 1998.
B. C. Berndt and R.A. Rankin, Ramanujan: Letters and Commentary, Amer. Math. Soc.,
Providence, London Math. Soc., London, 1995.
B.C. Berndt and H.H. Chan, Some values for the Rogers-Ramanujan continued fraction,
Canad. J . Math. 47 (1995), 897-914.
B.C. Berndt, H.H. Chan, and L.C. Zhang, Explicit evaluations of the Rogers-Ramanujan
contznued fraction, J . Reine Angew. Math. 480 (1996), 141-159.
H.H. Chan, O n Ramanujan's cubic contznued fraction, Acta Arith. 73 (1996), 343-355.
G.H. Hardx Ramanujan, Chelsea, New York, 1978.
K.G. Ramanathan, O n Ramanujan's continued fraction, Acta Arith. 43 (1984), 209 226.
, O n some theorems stated by Ramanujan, Number Theory and Related Topics, T a t a
Inst. Fund. Res. Stud. Math. 12, Oxford University Press, Bombay (1989), 151-160.
S. Ramanujan, Notebooks (2 volumes), Tata Institute of Fundamental Research, Bombay,
1957.
, T h e Lost Notebook and Other Unpublzshed Papers, Narosa, New Delhi, 1988.
L.J. Rogers, O n a type of modular relatzon, Proc. London Math. Soc. 19 (1920), 387 397.
G.N. Watson, Theorems stated by Ramanujan ( I X ) : T w o contznued fractzons, J . London
Math. Soc. 4 (1929), 231-237.
, Theorems stated by Ramanujan ( V I I ) : Theorems on contznued fractzons, J . London
Math. Soc. 4 (1929), 39-48.
Contcrnporary hlathematics
Volume 236. 1999
Dov Chelst
1991 Mathematics Subject Classificatzon. Primary 82B23. 82D10; Secondary 82B05, 82B26,
30B70.
This work was partially supported by AASERT Grant No. AS0281.
I would like to thank the organizers of this conference for allowing me to communicate my
ideas t o members of the continued fractions community.
I would also like t o thank Prof. Eugene Speer for discussing the details of this paper's main
argument with me and for scrutinizing my earlier drafts
138 DOV CHELST
I 1I
the analysis to one rather than three dimensions. In one dimension, a number of
models have exact solutions.
Rigorous proofs for the non-existence of phase transitions are another rare
commodity in statistical mechanics. These theorems, when they exist, focus on
showing that some fundamental relation describing a given system is analytic. In
one-dimension, the most famous non-existence proof certainly follows this line of
reasoning. In 1950, Van Hove[5] showed that one-dimensional systems with only
finite-neighbor interactions exhibit no phase transitions.
Plasmas, which consist of charged particles, interact via a long-range Coulomb
interaction and are not covered by Van Hove's theorem. Nevertheless, in 1962,
Lenard and S.F. Edwards[4] showed, by analyzing the eigenvalues of a differential
equation of Mathieu type, that the one-dimensional two-component plasma can
exhibit no phase transitions.' This was proven in a way that avoided the continued
fraction which featured so prominently in Lenard's original thermodynamic solution
of the system.
We can readily generalize the original thermodynamic argument of [9]to include
modified two-component p l a ~ m a s . One
~ simply notes that Leriard and Prager's
original argument uses a Laplace transform technique reminiscent of the solutions
of Tonks and Takahashi for systems with only hard-cores or with only nearest-
neighbor potentials respectively.3 Then, one replaces Lenard's Laplace transform
of the identity with a more complicated Laplace transform. However, it is not clear
how to generalize Lenard's second argument, that proved the non-existence of phase
transitions, in so broad a fashion. Luckily, this is not necessary.
-
a a
w a
w
a
w
-
w
a
- ..
1 a 1
w
1 1
w
a - 1
0
Modified two-component plasmas contain an equal number of positively and
negatively charged particles with unit charges f l and --1 respectively. Each particle
resides on the positive half line, and is specified by a pair of coordinates (a,. x,)
describing its charge and position. Pairs of particles interact via a one-dimensional
Coulomb potential
'In fact, he proved it for more general multi-component charged systems satisfying certain
conditions.
he entire argument can be found in [3].
3See [13, pp.48-531. A translation of Takahashi's original article can be found in [ l o , p.251.
140 DOV CHELST
4 ~ e Percus[8]
e for a discussion of this ensemble.
5 ~ p , r ( ~corresponds
) to a portion of the isobaric-isothermal partition function and is thus a
Laplace transform of some portion i N ( L )> 0 of the canonical partition function.
CONTINUED FRACTIONS AND MODIFIED TCP'S 141
O(s)= ( voPP(s)
vsame
( s ) ) , and E = --1--(1 - O ( s ) ) .
VOPP(S)
When $ does not depend upon the neighboring particles' charges, this descrip-
tion simplifies drastically. +,,, and $,, all become equal, while O r 1 and
E = 0. Hence,
2
a1 = -2,
9(P)
a, = -z,
bo = 0, and
6, = q ( P + n2).
The fact that the description of these continued fractions hinges upon the
reciprocal of a Laplace transform, coupled with the desire to have the continued
fraction description remain valid for all positive P , motivates us to stipulate the
following condition for $.
CONDITION 1. The Laplace transform C ( e P * ) ( P )determined by any nearest
neighbor potential $ describing a modified two-component plasma will converge for
all positive P and hence in the complex right half plane Wo= { z : R ( z ) > 0).
This condition is not specific to modified plasmas; it is implicit in the standard
thermodynamic description of pure nearest neighbor systems solved by Takahashi[lO,
p.25). Thus, Condition 1 is imposed, implicitly or explicitly, on many systems con-
taining nearest-neighbor potentials.
Of course, we must now convince a reader that the reciprocals of Laplace trans-
forms, the q's, are well-defined for appropriate values of P. In other words, we need
t o show that the Laplace transforms are non-zero. Only then can we show that the
continued fractions given by (6) and ( 8 ) converge. We will explore the properties
of Laplace transforms which make each 77 well-defined in Section 4.3.2.
142 DOV CHELST
4. The Analyticity of p ( P )
4.1. The Main Result. Now, we will state our main theorem and prove it
assuming Lemma 3 that we will prove in Section 4.4.
THEOREM 2. Let a nearest-neighbor pote,ntial $ satisfy Condition 1. When $
is charge-depe,nde,nt,let i t satisfy Condition 8. Let Q(P,z ) be a function that can be
described both as a continued fraction (6) and as a power series (4) for fixed P > 0
i n z about 0 with positive coeficients. Then the implicit function p ( P ) deJned via
(2) and (3) is analytic for P > 0.
4.2. Proof of Theorem 2. To prove Theorem 2, we must traverse a number
of steps.
PROOF. 1. First, we state the lemma:
LEMMA3. If Q(P,z) satisfies the hypotheses of Theorem 2, then Q is
a separately meromorphic function of P and z for all (P,z ) E 0,x C
for some appropriate open subset 0,of @ containing R+.
While we leave the proof of Lemma 3 t o Section 4.4, we would like t o
describe its conclusion and give an idea of its proof here. Specifically, we
show that if we fix one variable, and examine a neighborhood about any
value of the other variable, there is an appropriate kth remainder of the
continued fraction, K(a,+k/b,+k), which is analytic in that neighborhood.
Thus, the full continued fraction is a rational function of the first k partial
numerators and denominators and the remainder term. Since these are all
analytic functions of the free variable, Q must be a meromorphic function
of the free variable.
This argument utilizes two theorems from the arlalytic theory of contin-
ued fractions quoted in Section 4.4 as Theorems 6 and 7. At the same time,
it relies heavily upon the asymptotic behavior of Laplace transforms de-
scribed in Section 4.3. It is not suprising that for a charge-dependent system:
we need an extra condition, given in Section 4.4 as Condition 8, t o apply
these continued fraction theorems. A charge-independent system does not
require this additional condition; yet, when $ assumes the proper form, it
automatically satisfies Condition 8.
2. An extension of Hartogs' Theorem to meromorphic functions due t o W.
Rothstein[l2] states that a separately rneromorphic function, e.g. Q, in an
open complex domain, e.g. 0,x @, is actually jointly meromorphic. Thus,
Q is jointly analytic in a neigborhood of any solution t o (2).
3. Moreover, we note that the power series description (4) of Q implies t h a t
3 is non-zero about any solution t o (2) provided that both coordinates
are positive and z lies in the radius of convergence of Q. Together with the
meromorphicity of Q, this assures us that for P > 0, Q'S first pole away from
zero occurs on the positive real axis. It also guarantees t h a t , for positive P,
a unique positive solution z ( P ) exists within this radius of convergence
and that this solution has the smallest modulus of all solutions t o (2). Thus,
this solution is actually z* ( P ) .
4. We may now employ the complex analytic implicit function theorem[G,
1.B.61 t o prove that z f ( P ) is an analytic function of P . Let us briefly review
t h e implicit function theorem.
CONTINUED FRACTIONS AND MODIFIED TCP'S 143
La (t)e-"dt 5 r 1 K
f (t)e-ltdt.
Now, simply choose w > 0 so that w K 5 3. For all real values Is21 5 w and
0 5 t 5 K , cos sat 2 51
+
and for complex s = sl is2 in the complex strip S, = { s :
W s ) > x , lQ(s)l < w),
The proof is complete when we note that since S, contains no zeros of C ( f ) , the
Laplace transform's zeros must all remain at least a distance w > 0 away from R+
in W,. 0
Before we continue, we simply construct an open set using the conclusion of
the previous lemma. Given a suitable nonnegative function f and a positive r < 3,
define the "zero-less" open set 0, = U,>, S,. Thus, because L ( f ) has no zeros and
is analytic in U,, we are assured that 77 is well-defined and analytic there as well.
Finally, we note that this lemma completes a pair of inequalites. For any
S E U,,
Lemma[2, 6.41 covers a number of possible limiting behaviors. While this does not
exhaust all possible functions, nor does it include all physically meaningful func-
tions6, we will restrict ourselves to those potentials covered by Watson's Lemma.
An industrious reader will see that those generalizations of Watson's Lemma that
are also found in [2] can be treated similarly.
Watson's Lemma states that if a function f behaves as Ata to leading order as
>
t -+ O f , for some constants A and a 0 and if L ( f ) converges in some IHI,, then
the Laplace transform C ( f ) ( s ) decays as R(s) -+ m, to leading order, like -AT(a+l)
.
+
(I? is the Gamma function which for positive integers n satisfies I'(n 1) = n! and
using Euler's integral can be written as r ( x ) = Somtx-le-tdt.)This is certainly true
for real s and the statement for complex s follows from (9).
In statistical mechanics, we frequently encounter potentials $ that contain hard
core exclusions. In terms of f = e-P@, a hard core specifies that f ( t ) = 0 for all
t < b for some hard core diameter b. If, in addition, as t --, b+, f behaves as
A(t - b)O for some constants A and a 2 0, its Laplace transform L ( f ) ( s ) decays,
to leading order, as ~ l ? ( a + l ) e - ~. ~ This
R(s)a+l
( ' ) follows from Watson's Lemma and an
elementary feature of Laplace transforms that can be found in any undergraduate
text on the subject. The reciprocal of C ( f ) in either of these two situations, b = 0
or b # 0,diverges to infinity in an appropriate manner.
We note that while Watson's Lemma also allows for -1 < a < 0, a physical ar-
gument disallows these exponential values. After all, since f = e ~ ~our @ exponent
,
a depends linearly upon the inverse temperature /3. Thus, while for some values
of /3, 0 > a = a1/3would be greater than -1, there would certainly be values of P
for which this would be false and thus the Laplace transform would not even exist.
This contradicts Condition I that we imposed upon our nearest-neighbor potential
$ in Section 3.
This section on Laplace transforms provides all that we need in this paper.
It furnishes analyticity criteria. It gives a complex open set in which all 7's are
well-defined, i.e. just the finite intersection of open sets for each separate q. It also
provides the asymptotic information regarding our 7's that allows us to apply our
continued fractions theorems.
4.4. The Meromorphicity of Q.
PROOF OF LEMMA3. As stated in the proof of Theorem 2, our aim is t o show
that some remainder of the continued fraction Q is separately analytic in 0, x @;
i.e. analytic in a single variable while the other remains fixed. For this purpose, we
employ two theorems found in chapter 4 of the book by Thron and Jones[7, Thms.
4.35 and 4.541. Note that ,f refers to the nth convergent of the continued fraction
whose value we can obtain by substituting 0 for a n + l .
THEOREM6 (Pringsheim). If / bnl > ianl + 1 for all n , then
i. K (e)converges, and
ii. lfnl 5 1 for all n.
K
6 ~ r ey - p from a potential $(y) = 5
146 DOV CHELST
THEOREM7. If:
i. a,(x), b,(x) are analytic in some domain D, and
ii. SIC1, C2 E @, s.t. for all x E D and n E Z+,
f,(x) @ {El, C2, m)
Then K (-1 is analytic in D .
We first note that all the partial numerators and denominators described in (6)
and (8) are analytic functions of P and z . In 0,, both qOppand qsam, are nonzero
and hence their ratio is an analytic function. Thus, O and Z are both analytic
functions of P. Analyticity in z is even more obvious.
Now, consider O ( P ) and E ( P ) defined in (7). By the asymptotic behavior of
the two q's, according t o Watson's Lemma,
D S ~ v g 2 ~ ' - 2 0 ( r a me(buPI'-2bBBIIIII
t
/ E ( ~1 )
)SO 7
prove problematic.
In certain simple instances, e.g., when 1C, -
the ,8 which figures in the definition of the transformed function itself e c P could
~
5. Conclusion
We have accomplished our goal of utilizing the theory of continued fraction t o
prove a fact regarding modified two-component plasmas. Of course, some questions
remain unresolved. Did we use this theory effectively? The theorems employed
provided sufficient conditions preventing the existence of phase transitions in these
systems. But, how necessary is the additional condition imposed upon plasmas with
charge-dependent nearest neighbor interactions. What happens if bopp > 2bSam,?
Continued fractions may yet hold the answer t o this question. Continued frac-
tions continue t o persist in many models related t o the two-component plasma.
They appear explicitly in the discussion of one-dimensional ion-dipole systems in
[15].Moreover, they are implicit in every article that calls upon the reader t o calcu-
late an eigenvalue of a Mathieu-type differential (differential-delay) equation with
periodic boundary values7 For these reasons, we believe that continued fractions
will continue t o provide valuable information regarding this family of statistical
models.
References
R.J. Baxter. One-dimensional gases with hard-core repulsion. T h e Phyics of Fluids. 8(4):687-
692, 1965.
Carl hl. Bender and Steven A . Orszag. Advanced Mathematical Methods for Sczentists and
Engineers. McGraw-Hill, 1978.
Dov Chelst. Modified One-Dzmensional Two-Component Plasmas and Generalizations of
Schwarr's Lemma. PhD thesis, Rutgers University, 1999.
S.F. Edwards and A . Lenard. Exact statistical mechanics o f a one-dimensional system w i t h
coulomb forces. ii. t h e method o f functional integration. Journal of Allathematical Physics,
3(4):778-790. 1962.
Harry L. Frisch and Joel L . Lebowitz. editors. T h e Equilzbmum Theory of Classzcal Fluzds.
volume 15 o f Prontiers zn Physics, chapter 1-7. W . A . Benjamin, Kew York. N Y . 1964.
Robert C. Gunning and Hugo Rossi. Analytic Functions of Several Complex Varzables.
Prentice-Hall, Englewood Cliffs, N J , 3 edition, 1965.
W i l l i a m B . Jones and W . J . Thron. Contznued Fractzons: Analytzc Theory and Applzcatzons.
volume 11 o f Encyclopedia of Mathematics and its Applzcatzons. Addison-Wesley. Reading,
M A , 1980.
J.L. Lebowitz, editor. Simple Models of Equilabrzum and Nonequilibrzum Phenomena. chap-
ter 1. Elsevier, Amsterdam, 1987.
A. Lenard. Exact statistical mechanics o f a one-dimensional system w i t h coulomb forces.
Journal of Mathematzcal Physzcs, 2(5):682-693. 1961.
Daniel Mattis. The Many-Body Problem: A n Encyclopedia of Exactly Solved Models i n One
Dzmension. World Scientific. River Edge, NJ. 1993.
S. Prager. T h e one-dimensional plasma. Advwnces i n Chemical Physzcs. IV:201+. 1961.
Wolfgang Rothstein. Ein neuer Beweis des Hartogsschen Hauptsatzes und seine Ausdehnung
a u f meromorphe Funktionen. Mathematische Zeztschrift, 53:84-95, 1950.
Colin J . Thompson. Classical Equzlzbrium Statzstzcal Mechanzcs. Oxford University Press.
London. 1988.
[14] Fernando Vericat and Lesser Blum. Statistical mechanics of ion-dipole mixtures: An exactly
solvable model in one dimension. Journal of Chemical Physics, 82(3):1492-5, 1985.
[I51 Fernando Vericat and Lesser Blum. Many-body functions of nonprimitive electrolytes in one
dimension. Journal of Statistical Physics, 61(5/6):1161-1185, 1990.
Dcdzcated to L. .Jerry Lange and Wolfgang Thron on the occaszon of thezr bzrthdays.
1. Introduction
In 594 of his book on continued fractions [14],H.S. \.Val1 lists continued frac-
tion expansion formulas for many special functions. His list contains 24 continued
fractions. In our generation it was a good exercise for beginners to go through this
list and try to prove them all especially after seeing the Gauss continued fraction
in 389. Among Iliall's list. five continued fractions are listed undcr group C , titled
"Integrals involving the Jacobi elliptic functions." They do stand out as being very
different from the rest. Thcy are
where
149
150 M O U R A D E. H. ISMAIL A N D DAVlD R. MASSON
where
(1.6) n,, = (2n + 2)'(1+ k2), b,, = 2n(271 + 1)'(2n + 2 ) k L ,
where
(1.8) a,, = 2(2n + 1)'(2 - k'), b,, = (2n - 1)(2n)"2n + 1)k4.
IVall [14] attributes the continued fractions in (1.1). (1.3) and (1.4) t o Stic1tjr.s
[13],and credits Rogers [12] with finding (1.5) and (1.7). A more extensive list of
contliiued fraction expansions is in the verj7rccerit book on the sub jcct by L o r e n t ~ e n
and Waadeland [lo]. Note that the fartor z on the left-hand siclr of (1.5) was left
out in Wall [14. (94.20)] and [lo. (2.6.4)].
Obwrvr that the continued fractions (1.3) and (1.4) are equivalent in vicw of
the relationship [3. 98.1531. [4. p. 2491
(1.9) C I ~ ( IILl k. ) = d n ( u / k >k).
One can also show that (1.1) and (1.7) are equivalent using Landen's transformation
[12]. [15]. Thus, of these five continued fractions. only three arc, independent.
n a t ~ i ~ l(1.1).
y (1.3) and (1.5).
In Sections 2 arid 3 we will generalize these and derive an additional coritinucd
fraction given by
where
(1.11) a,, = (271 + 2)' + (271 + l)'kL, b,, = (2rtj2(2n + 1)'k'
Equation (1.10) can also be obtainrd from (1.3) in the followirig rnanrier. On
the lcft-hmd side of (1.3) 011e can iritegratc bj. parts to get
x
(1.12) J cnju, k) sn(ii. k ) du =~k ~ ~
0
Now substit~ltethe right-hand side of (1.3) into the right-hand side of (1.12) and
use t 11c idcnt ity
which can be derived from Lemmas I a d I1 in Rogers [12]. This identity should
also be cornpared with the identity (3.16) of Section 3 of this work.
SOME CONTINUED FRACTIONS RELATED T O ELLIPTIC FUNCTIONS 151
Alt,llough the result of t,his calculation is indeed (1.10). we prefere the more
fundamental derivation of Section 3 , which uses solutions to a differericc equation
arid yields both (1.3) and (1.10).
hlotivated by the nlornent problerns considered in [I]and 121. Isinail and Valent
[ 5 ] .[6]st,ridied more general nlonient problenls which led t o new orthogonal polyno-
mials. The generalization was essentially to replace the lemniscate furictions in [I]
and 121 by ellipt,ic int,egrals. This led Ismail and Valent 151. [6] t o study orthogonal
polyrionlial having generating fimct,ions of t,he form
where g ( t ) is given by
It turns out that thc only choices for rr and 0 arc tr = 0 , - 1 / 2 and 1) = 0 , 1 / 2 .
hence there are essentially four possible sets of orthogonal polynoniials. Tlicse
orthogonal polynomials are connected with the denonlinator approxirrlants of the
four independent conti~iuedfractions ( 1 . 1 ) , (1.3). (1.5) and (1.10). The case (1.14)
with a = / j = 1 1 2 . corresponds to (1.1) while cu = O = 0 correspolids to ( 1 . 5 ) .
We note that cu = -112% 9 = 0 or (r = 0, P = -112 corresponds t o the even
denominator approxirnants of (1.3). To see this consider the recurrence relation
associated with ( 1 . 3 ) ,namely
(1.17) X,,+l ( z ) - z X , , ( 2 ) - n L x , ,- 1 ( z ) = 0, n odd.
(1.18) + (z) - zX,, ( z ) - ~ 1 ' k ~ ~( z, ), 0. 11 C V C I ~ .
.
with birth and death rates A,,, p,,,, respectively; and s = -2'. The case n = 2171 + 1
is associated with
152 MOURAD E. H. ISMAIL AND DAVID R. MASSON
and corresponds t o (2.25) in [6].since it also has the corrcct initial conditions. On
the othcr hand, the case n = 2777 + 2 appcars t o be associated with
His principal technique was t o usc integrat,iori by part,s t o find solutioiis of difference
equations.
Having t,hr, integrations in (1.24) and (1.25) e ~ t ~ e n d i nt o
g infinity is. however.
misleading. It is 1nlic11 better t o consider bot,h end points of integration as zeros
of t,he elliptic furlctions in the integrand. The bonus for doing this is that wc are
then able to obtain both a dominant and minirnal solution t,o the corresponding
t,hree-t,erm recurrence. Thus. in Sect,ion 2. we examine the difference eq~iatioiifor
the funct ions
2 li
(1.26) C:,(2;k.) =
I
0
sn" ( u .k : ) e z "du.
/
2 /I-
where
Alilnc's interesting work [ll]discusses Rogers's method and records related contiii-
ucd fractions. This is then applied t o study the representation of integers as sunls
of squares.
2. A Difference Equation
In this section we consider a difference equations and its ~olutions.This differ-
ence equation has two explicit special cases corrt,spoilding t o thc coritinucd fractions
(1.1) and (1.5).
TVe follow the procedure that Rogers uwd for u,,(z:k) [12] and ohtdin a gen-
eralization of (1.1) dnd (1.5).
Let U,, ( 2 ; k) be as in (1.26). Integrating by parts twice, wc have. for 11 > 1, n
not necessarily an integer.
where in the last line we have used Example 4. p. 516. of [15]. \li,thus have. for
n > 1. a solution t o the three-tcrrn recurrence rclatioii
(2.1)
n(n + l ) k LY,,+2 (z: k) - (z' + n' (1 + k L ) )Y,,(z: k) + n ( r ~
- 1) Y,rF2(2; k) = 0.
Another solution t o this recurrence relation is U,,(-z; k). These two solutions
are, however. linearly dependent. To show the linear dependence. we make usc of
the properties [15]
+
Thus with n = r n 0.In = 0, k1.1t2, . . . the linear clepcwleilcc is established
If wc restrict ourselves to Rc z > 0 and rr a n integer. we have anothcr linearlv
dependent solutio~igiven h~ u,, (z. k) of (1.21). To see this we split the infinite range
of intcgratiori into intervals of lengtli 2h7 arid use (2 2) t o obtain
'I\
Note that. strictly speaking. U,,(2:k ) , and en"'U,, (z;k) are solutions which are
linearly indept:ndmt but in a trivial may. The factor e n T Lcan be attached t,o any
solution of (2.1) to geiierate linearly independent solutions. since, (2.1) is basically
a sccond-order d i f i r t w ; e equat,ion in the square of the shift operator.
Tlic solutions that we have are bounded solutions t o the recurrence relat,ion
(2.1). The general i d u r e of solutions can be seen from t,he large n as>-rript,oticsof
(2.1) through the Birkhoff-Trjit,zinsky theory [16]. For n large (2.1) approximates
t,he coiistant coeficient tlifftrence equation
(2.6) X.*X,,+~ - (1 + k') X,, + X r r - =~ 0.
This liab lirlearl>-indepe~ldentsolutions
The Birklioff-Trjit z i n s k ~theory implies that the recursion relation (2.1) will have
linearlv independent solutions which. for large n. are asymptotically proportional
to the wlutioiis (2.7). With 0 < k' < I wn.c must then have one solution of (2.1)
lvith constant asvmptotics and a linearly independent solution wl~icllbecomes un-
bounded as 11 4 30. Since 0 5 m ( u , k ) 5 1 for LL 111 the iriter~al[O, 2 K ] , we have
11,)(2: k ) bourided as a function of n and corisequentl\ ~t must he a bounded solution
t o thc recurreilce relation (2.1) ancl nsymptotic t o a coristant multiple of x:," in
(2.7).
I11 ortlcr to obtain ari unhoulldccl solutioii to (2.1). we now consider n an intcgcr
and
11, '
(2 8 ) ( z :) = /
-ih'
[ s) , n > 0.
7r
(2.9) 1' = ( 1 / 2 , 1 / : 1 1 - k ) O < kL < 1.
2
Tlie integral in (2.8) is in the coniplex u-plane along a contour from u = -7K' t o
I L = ?I{'. which nrc poles of sn(u. k ) , and avoids the pomt rr = 0, wht.re sn(71,k)
v;liilsllci. 111 (2.8) t l i ~point r l = 0 is to the right (left) of the contour w h m we use
I;T;y (Cr;,- ) ) .
Ti-e again intcgrdtt. b> parts twict, t o obtain from (2.8)
valid for rc 2 2. \\.'ith re-liormnlization this becomes the sanw recurrence as (2.1).
~ , iiit(1g~r71. n > 1. t he rcm~rsion(2.1) has mother linearly indcpcrldent
H e i ~ ( for
solut ion giwn by
The contour in (2.12) is now a small circle. t,raversed counterclockwiscl enclosing tlic,
polr singularity at u = 0. An application of Caucliy's tlleortm sliows that (2.12) is
a polynomial in z of degree n 1 with even powers of z for r l odd and odd powers of
-
z for r1 e w n (see Corollary 2.3). If we refer t o (1.2). (2.15) in the work of Isillail a i d
Valerlt [6] arid compare t,heir recurrences with our (2.10) we can see that spcciwl
cases of the polynoniials in the generating furictions (1.14) and ( 1.15) are given hy
2.1. T h e r e r ~ ~ r w n crelnflon
THEOREM r
(2.16) X , , + I(z; k. cr) - [z' + n,,(k. cr)] X,,( z ; k , a ) + h,,(k, a ) X , , l ( z :k. a ) = 0.
u~rth2n +CL > 1. 0 < k' < 1. and
(2.17)
o,,(k..cu) = (277 + a ) ' ( l + k'), b l , ( k .0 ) = (2n + tr)(2r1 + a - 1)'(2n + rr - 2)kL.
has n mznirnnl solution
PROOFTo obtain the recurqiori (2.16) with solutiori (2.18) wc use (2.1) and
the solution (1.26) with n replaced 1?\. 271 + a and wt. renormalize. The ~liirlirnal
and dominant nature of the solutions comes from thc fact that ULl,+(,(z:k ) and
- i-t)
ULr, +(,( z : k ) have nonzero constaiit large rl as) inptot ic s ( s w the discussion follow~lg
(2.7)). Similarly, (2.19) and (2.20) conic from (2.10) drld a renorinalizatio~iof (2.11)
a i d (2.12). To sce that the right-hand side of (2.20) is a polynomial of degree. 11 in
z' u.1, perform the iiidicated ~ntegrationusing C a ~ c h y ' stheorein This is done in
Corollary 2.3 below. This tomp1etc.s the proof. 0
Based on the rcsults in Tlieorcm 2.1, we now oht ail1 a continued h a r t ion ex-
pansion which g e n c d i z e s ( I .1) and (1.5).
156 MOURAD E. H. ISMAIL AND DAVID R. MASSON
Un( z :k ) 1 b , ( k ,a ) b ~ ( k0, )
(2.21)
-
- ...
IL(N -) U p ( :k ) zL + +
a ~ ( k0. ) - z L a l ( k ,0 ) - z L + a 2 ( k , a ) -
,valid for a > 1 and 0 < k 2 < 1. I n partic,ulrrr, ,we have the special cases
-
- b l ( k 1) .
(1 +exp(-2Kz)) z L t n o ( k , 1)- z' + nl ( k . 1 ) - zL + a2(k. 1)-
, ,
The result is (2.21). In (2.21) we put n = 2 and use the explicit expression
U o ( z :k ) = z p l ( l - e x p ( - 2 K z ) ) to obtain ( 2 . 2 2 ) . In (2.21) one can take the limit
as tr + 1+ to obtain (2.23). Instead we integrate Ul ( z ;k ) by parts twice to obtain
the extra recurrenu,
IVe now give more explicit expressions for the solutions P,, ,, ( z L :k ) , a = 1.2
in (2.20). This will justify our claim that these are polynomials of degree n in z L
and the formulas (2.13) and ( 2 . 1 4 ) .
COROLLARY
2.3. W e h a w the explicit expresstons
SOME CONTINUED FRACTIONS RELATED T O ELLIPTIC FUNCTIONS 157
and
PROOF. IYe apply Cauchy's theore~nt o the expression ( 2 20) aftcr rearranging
the integrand, expanding ( u / sn(u, k))'"+"e-"' in powers of v and finding the co-
'
efficient of uL"+" - in t his expansion. Thc fact that sn( u , k ) is an odd function of
u causes the odd derivatives of ( u / sn(u, k ) ) z " t " to vaniqh a t i~ = 0. The re5ult i5
(2.25)-(2.28) and the proof of the corollary is coruplete. 0
The continued fraction ( 2 . 2 1 ) and the special cases ( 2 . 2 2 ) ancl (2.23) are positive
definite J-fractions with variable x = 2'. AS such they have a special representation
as a Stieltjes transform of a probability measure. Here the measure is discrete so
that. for N > 1 ancl O < k L < 1.
with x,,, ( a ,k ) real and R,,,( a , k) > 0. For the special cases n = 1.2. the poles
(a:, k ) and residues R,,, ( N , k ) turn out t o b e explicit. The polynomlal5 PI,,, ( z ' ; k )
s,,,
are orthogonal with respect to tlle5e discrete measures.
wc realize that
is obtained from
Using these values in (2.35) we establish (2.31). Tliis completes the proof of Tlieo-
rrm 2.4. 0
All of thc above results can be carried over t o the case whcn 1 < kL < x if we
use the replacement [4, p. 2491
(2.38) sn(u, k) -t k 1 sn(nk, k-I).
Sl'e will not repeat the calculatio~lsfor this case. Instcad. we give the malor
formula changes. First we cxterid the definitiori of thc U,,'s a?
where
Then t h r recurrence relation (2.16) with 2n + (I > 1. but with 1 < kL < x,st ill
has a minimal solution
(2.42) X;,')(Z;k, t r ) = r ( 2 n + c u ) k L ' 1 ~ 2 r(, z+;ik.).
r
SOME CONTINUED FRACTIONS RELATED T O ELLIPTIC FUNCTIONS 159
Furthermore (2.21) is unchanged, but t,he left sides of (2.22) and (2.23) arc replaced
2L/k
sn2(uk,k-')e-lLz du.
and
2L/L
,f sn(uk, k l ) e " ' du
and
(rn - 1 1 2 ) ~ k " ~ 2n2k (rn- 112)pm-1/2
(2.48) s,,(l.k ) =
L
, R,,,(l,k)=-
L2 1-p2rll-l '
wherc the contour of integration in (3.2) avoids the point 11 = 0. \Vc take the origin
t o the right (left) of the contour for the + (-) superscript. A linear cornbination
of the solutions in (3.2) yields the solution
k-r1
y j L ) ( z :k) = --
2x1 . f
d n ( u , k ) [sn(u,k)] " - ' p - " '
This is seen by cornparmg the recurrence^ (3.1) with tho5e in [6. (2.24). (2.31)].
rcnornlali~ing.arid applling Cauchy's theorem t o thc integral in (3.3). The latter
stcp i5 don(, below in Corollary 3.3. Notc that Vr,(z:k ) is a boul~dedsolutiori t o
(3.1). whilt~(3.2) and (3.3) a r r uribourldtd 5olutions, as r1 + x.
\Ye now rephrase these result5 in terms of solutions t o a threr>-term recurrence
relatlon which is in standard ~nonicform.
THEOREM
3.1. The recclrrrnce relaf7on
(3.5) X,,+~(~;k,tw)-[z'+~1,,(k,n)]X,,(z:k,n)+b,,(k.n)X,,-~(z:k,n)=0
uihere
(3.6)
tr,, ( k , rr) = (2n + 1 + a)' + ( 2 n + n)'k2. b,, ( k , n ) = (2n +n - 1)'(211 + n)'k2,
w t h 2n + (I > 1 . 0 < I;' < 1 , has a rnzn~rnalsolut~on
(3.7) ~ , ( , l ) ( k,
z ; tr) = r ( 2 n + + I ) ~ ' ~ ~ V(2:~ ~k ), + ~ ,
11
which 6s of drqrer n in z 2
(3.10)
V ,( 2 ;k ) 1 bl ( L0 ) b~(k. n )
...
-
cu(a
-
- 1)V,, L ( ~ k;)
-
zL + c ~ , , ( k . .n + +
) z 2 n l ( k . ( 1 ) - z L a ~ ( k0 2, ) -
SOME CONTINUED FRACTIONS RELATED T O ELLIPTIC FUNCTIONS 161
which holds for cu > 1 and 0 < k2 < 1. For a = 0: 1 there are th,e additional special
case.5
2 11.
J cn(u, k)eP"' du
-
1 bl (k. 0) b2(k. 0) ...
(3.11) I'
z(1 + exp(-2Kz)) z' + ao(k. 0)- z' + a l ( k . 0)- z' + a2(k.0)-
To derive (3.11) wc integrate In?- parts twice to calculate the extra recurrence relation
and combine this this with the case a = 2 of (3.10). Similarly we have the additional
recurrence relation
II
(1.10) results. To rstablish (1.3) urc>use (3.15) with n = 1 on the left-hand side of
. sn(rr, k ) e "'drr and
(3.11) and we re-exprcss the rlght-lland side of (3.11) using the identity [12. Lenliria
I1 1141
Just as in thc case treated in Scction 2, we now give explicit exprcwions for the
orthogonal polynomials wliich are the drnoniinator approximants of thc continuc~l
fractions in (3.11) and (3.12).
162 MOURAD E. H. ISMAIL AND DAVID R. MASSON
COROLLARY
3.3. W e have
(3.17)
whew
(3.18)
E7 ,I =
(2n)! d2~~-21
( 2 ~ ) ! ( 2 n - 2 j-
71
) ! d u ~ ~[ ' -(sn(u,k.)
~ )~ 2 7 ' 1
d n ( u k)] 1 71 =O
,
and
(3.19)
with
(3.20)
Our next theorem describes the Stieltjes transform associated with the contin-
ued fractions in (3.11) and (3.12).
THEOREM 3.4. The continued fractions i n (3.11) and (3.12) have the Stieltjes
trnn.sforrn representations
PROOF. We expand the lcft-hand sides of (3.21) and (3.22) in terms of poles
and residues using the zeros of 1 ZIZ exp(-2Kz) and the values
(3.24)
i
II
cn(u, k) sn(u, k) \in(n~.iru/K)du=
2 d mq"'
kL h 1 + qL11' '
----
SOME CONTINUED FRACTIONS RELATED T O ELLIPTIC FUNCTIONS 163
for n,= 1 , 2 . . . . . Thesc values are equivalent to thr, Fourier series expansions [ 1 5 ]
4. Limiting Cases
The results of $2 and Q3 can he extended to the limiting case k = 1. Except
for Theorems 2.4 and 3.4, this extension easill- follows from the fact that
(4.1) lim sn(u, k ) = tan11 u , lirn cn(u, k ) = lim dn(u, k ) = sech 11.
A-1- k- 1 A--1-
important formula
Similarly
(4.6)
ask + 1.
IVe thus arrive at the following k + 1 limits for (1.1), (1.5). and their Stieltjes
trallsforin representations
The discrc,tc measures in Theorem 3.4 are dealt with in the same mariner so
that thc k -. 1 limits of (1.3) arid (1.10) can be written as
Note that the Stieltjes transform rcpresentatioiis of the continued fractions in (4.7)-
(4.10) are valid for the full convergence region Rez # 0 while the Laplace transform
representations are valid only for Rez > 0.
These four cascs correspond to special cases of continuous dual Hahn polyno-
mials. The coinparison with Section 1.3 of Koekoek and Swarttouw j8, p. 271 shows
that (4.7)-(4.10) are associated with continuous dual Hahn parameters
SOME CONTINUED FRACTIONS RELATED T O ELLIPTIC FUNCTIONS 165
Using (4.11) and [S. (1.3.I ) ] we can now statc the <FL
forinulas for the d m o ~ n -
inator polynomial approxiinants of the continued fractions in (4.7)-(4.10). They
are
and
respectively.
References
[I] C . Berg and G. Valent. The Ncvnnlzn,na parnmeterzzatzon for some mdeterrnznate Stzelt~rs
m,ornent prob1em.s assoczated wzth bzrth and death processes, hIethods and Applications of
Analysis 1 (1994), 169 209.
[2j Y. Chen ant1 hI. E. H. Isrnail. So~riemdetermznate n ~ o m e n tproblem,^ and Freud-lzke umghls,
Constructive Approxi~rlation14 (1998), t o appcar.
[3] I. S. Gradshteyn and I. hI. Ryzhik, "Tables of Integrals. Series. and Products," Acatlernic
Press. Orlando, 1980.
[4] H. Hancock, Lectures on the Theory of Elliptic Functior~s,\'olume 1: Analysis, Wiley. New
York, 1910.
[Fi] h1. E. H. Ismail and G. Valent, On n f a m ~ l yyf orthogonml p o l y n o m ~ u lrelated
~ to cllzpt~c
fun,ctzons. Illinois J. Math., t o appear.
[6] hl. E. H. Issnail and G . Valent. Some orthogon,al polynminmls related to cllzptzc func.t7ons. t o
appcar.
[7] W. B. .Jones and W. J. Thron. "Continued Fractions: Al~alyticT l ~ c o r yanti Applications,"
Addiso~~-Wesley, Rearling. h l h , 1980.
[XI R. Kockoek and R . Swarttouw, The Askcy-scheme of hypcrgcornctric orthogonal polynomials
ant1 its 9-ar~alogues,Reports of the Faculty of Technical hlathcrnatics ;md Infor~nationNo.
94-05. Delft University of Technology, Delft. 1994.
[9] .J. S. Lo~norltautl J. Brillhart. Certain Elliptic Integrals anti Their Orthogonal Polynomials.
t o appear.
[lo] L. Lorentxcr~and H. Waadclar~tl,Tent inucd Fractior~sWith Applications." North Hollaml,
Arnstcrtlam, 1992.
[ I l l S . C . hIilnc. S u m s of s p a r e s . .Iacobz ell~ptzcfunctzons an,d continued fmctzons. and Schur
functzons, t o appear.
[I21 L. J. Rogers. On the representatzon of certain asymptot~cserzes as convergent contmued
fractzons. Proc. London Math. Soc. (21, Vol. 4 (1907). 72 89.
[13] T. J . Stieltjcs. Sur la r6duction en fraction continue d'une skric prkc6tlent suivant les pouis-
sarlccs descerdants d'une variable, Ann. Fac. Sci. Toulousc. Vol. 3 (1889), H. 1 17; Oeuvres.
\ b l . 2 , 184 200.
[14] H. S. Wall, Analytic Theory of Continued Fractions. Van Nostrand, prince tor^. 1948.
[15] E. T . Whittaker and G. N . \t;atson. "A Course of hIoderr~Analysis." Fourth Edition. Carrl-
bridge University Press, 1962.
[lfi] .T. Wimp. "Cornputation Wit21 Recurrerice Rolatior~s."Pitman. Boston. 1!1X4.
166 MOURAD E. H. ISMAIL AND DAVID R. MASSON
D L P A R T ~ I E NOF
T ~ I A T H E ~ I ~ Z T ~I ( N' S I ERSITY
\ OF SOLJI H FLORIDA TAMPA
E'L 33620-5700
E - m a d address ~smailQmath. usf . edu
1. Introduction
Many special functions of importance in mathrulatical and phi-sical sciences
can be expressed in terms of Stieltjes transforms
(1.2) ci = 1
x
t"dv(l). 1 = 0.1.2 . . . . .
For corwenience we call such functions S-functions and note that all S-fu~lctions
are analytic in the cut plane
1991 Mathematics Subject Classificatzon. Primary 30B70, 30E05, 30E10, 40A15, 41A20
Research supported in part by the U S National Science Foundation under Grant No DhIS
9701028
168 WILLIAM B. JONES AND GUOXIANG SHEN
= x (Thron's criterion [ 1 4 ] )
rr=l
and
for z E ST:
where
(1.10b) c ( t ) := ~ ' ( t ) ,for
ASYMPTOTICS O F STIELTJES CONTINUED FRACTION COEFFICIENTS 169
IVe use properties of the nonnegative weight function 7 ~ ( t )in (1.10) t o deduce as-
ymptotic properties of the sequence { b , , } ~ =. , Ehllowing the approach recently em-
ployed bv Jones and Van Assche in [9],we make use of the Freud corljecture/tlleore~n
proved in 1988 by Lubinsk>, hIllaskdr and Saff [ l l ] . For later reference that theo-
rem is stated in Lemma 2.3 and is referred t o hereafter as the FLMS-theorrrn. The
main result given by Jones and Van Assche [9] involves the Bin& function J ( z )
defined by
. for z E S,,
By using the FLilIS-theorem and properties of the weight function u,(t), Jones and
Van Assche proved that the coefficients bd of the S-fraction representation of G J ( z )
satisfy the asymptotic condition
b1 1
i n = - (corijectured by Cizek and Vrscay in 1982 [3])
n-*x n2 16
The present paper extends the work of [9] by deriving the as? rnptotic behavior
of S-fraction coefficients b,, for a large class of S-functions which subsumes (1.11)
as a special case. Our main result (Theorem 2.1) is stated and proved in Section
2. Section 3 is used to describe applications of Theorem 2.1 t o functions a,, i ( z ) .
closely related to Whit taker functions ti',, ,,(z) and modified Bessel funct ions h;,( 2 ) .
Before proceeding to Section 2, we sulnrnari~e(Theorem 1.1) some kilown rc-
slilts on convergent S-fractions. which rclate asvnlptotic propertics of thr, S-fraction
coefficients b,, (as n 4 x)wlth asymptotic properties of the nth approxiinnnt g,, (z)
(as z -+ x and as n 4 m ) . These resiilts can bo found in [8. p. 1951 with a srnall
nlodification for Theorem 1.1 (C) .
THEOREM 1.1. Let G(z) be a n S-fraction (1.5) ,which c o n v e ~ y e slocally ,!mi-
forrnly on S, to a n S-function (1.10) (or more generally (1.1)). Let b,, denote the
n t h coeflcaent and g,, (z) the n t h approximant of G ( z ) . Let z E S,. Then:
( A ) Suppose lirn,,,, b,, = 0 and let S he a subset of @. If th,ere exist constants
p > 0 an.d p > 0 such, that
lb,, 1 5 inf lzl nlin
, d7.=2
, ,. zf 0 < 7 < 2 , where 6 ( 2 ) := ( 1 - i) L/i;T.
2. Main Theorem
Before stating the main result of this paper it is convenient t o introclucc a
family of fiinctio~is,dcnotecl by Q , and defined as follows: A function Q : R
a mernber of Q if Q ( z ) satisfies the following conditions:
R is -
(2.la) Q ( - s ) = Q ( x ) for all r ER (i.e., Q is even):
and
zQ1I(x)
linl ---- exists.
-+.; Q1(.r)
THEOREM 2.1. Let positive constants a , 6 a n d c be given. Let Q ( x ) be a nzernber
of t h e family Q that .satisjies t h e additional conrtitions:
(2.2) Q ) i s c o n t i r ~ u o u sfor all s >O
a n d either (Cast: a )
(2.3~) Q / ( J )= CS"- +~ ( ~ r l - l -) h as x--t+m
o r ( C n s e b)
(2.3b) rv >1 a,nd Q 1 ( s )= c s " - ' + o(l), as x 4 +x.
Let u:(.r) a n d ?j(t) he defined by
(2.4~~) 1 4 ) := , for - 3(j < .r < x
(1 n rl
T h e n : ( A ) T h e mornenis
X
(B) Let G ( z ) denote the S-function defined by the S - t m n s f o r m (1.10): let L(z)
A
denote the formal Laurent series (1.4); and let G(z) denote the corresponding S-
fraction (1.5) with coeficients b,, . T h e n
>
(C) Let N satisfy N 1. T h e n the S-fraction G(z) converges locally uniformly
o n S, to G(z) and the asymptotic behavior of the n t h approziw~antsg,(z) of G(z)
is described by (1.7) of Theorem 1.1, since b,, = 0 ( n 2 / " ) as n --t X . and y:=
2/a
satisfies 0 < y 5 2.
Our proof of Theorem 2.1 makes use of two important lemmas whose proofs
can be found in references cited. They are stated below to make the present paper
self-contained. The first of these results (Lemma 2.2) has been proven in [9],but it
was not stated explicitly there. See also [2. Theorem 9.1 of Chap. 11 for a partial
proof.
A Let ~ ( t be) a non-negative weight function on (0, no) whose mo-
L E ~ I M 2.2.
ments c,, i n (2.5) exist. Let e ( z ) , L(z) and G ( z ) d e n o t ~the corresponding S-
fraction (1.5), formal Laurent series (1.4) and Stieltjes transform (1.10). respec-
tively. Let w(x) denote the even weight functzon defined b y
lim
a,,
- = -
1
,I +" q,, 2 '
172 WILLIAM B. JONES AND GUOXIANG SHEN
Therefore, for cach integer n > 0, there exists a number B, > B t f such that
(2.16) eQ(') >
- x"+', for x L:> B,,;
hence
F'rom (2.la) and (2.17) one can con( lude that the moments p,, in (2.5) (with respect
to w ( x ) ) exist. From this aiid (2.4b) it is readily shown that the lrlomerlts c,, in
(2.5) with respect to ~ ( texist ) and 5atisfy equations (2.6). By a si~nilarargument
one can show that (A) holds if (2.3h) is satisfied. This concludes our proof of (A).
(B): Canr 0 . Suppose (2.3a) holds. A similar argument holds if we assume
(2.3b) instead of ( 2 . h ) . \Ye begin In\- showing that
whcrc q,, denotes the positive root of (2.12). Our proof is by contradiction. Hence
we nssumc tlidt (2.18) is false. It follows that tlirre cxlsts a corivergent subseyuen~e
)EO
{qIlA of {(I,,):=:=(, and we write
(2.19)
-
q := lirn q,,,
h x
< w.
Since Q1(x) is continuous for x > 0 and ( 2 . 1 ~holds,
) there cxists a constant AI > 0
such that
ds 7l
It is casilv seen that -
. B\ applying Lehesgue's dominated comer-
gencc thcorcrn [6] t o (2.19) we obtam
yields
The desired result (2.7) follows from (2.10). (2.13) and (2.23). This completes the
proof of (B).
(C) is an immediate consequence of (2.7). Thron's criterion (1.8) arid Theorem
1.1 (c').
is called the Whzttaker functzon of the first k m d . Here 1 F1is the confluent liyper-
geometric function
where a and b are complex parameters with b 6 [O, - 1. -2, -3, . . . ] and where
( a ) ( ):= 1 and ( a ) A := a ( a + 1) . . . ((L + k - I ) , for k = 1 , 2 . 3 . . . . . The branch of
~ ' ' ~in4 (3.2b) is chosen so that the branch cut is on the negative real axis arid
s t l + f > 0 for -(-. > 0. (See. e.g.. [I], [5], [li'].for more on Slrhittaker functions.)
Wh?ttcrker',sfunctton of the second kznd It', , , ( z ) is defined by
(see. e.g.. [ 4 ] )one can show that a, ~ ( z is) closely related t o incomplete ganima
functions r ( a , z ) and hence t o the error function. Modified Bessel functzons of the
second kznd K , , ( z ) can be expressed by
1 7l
(3.10b) Y ; , ( ~ ) = - - [ W ~ , K , , ( ~ ~ ) + L J ~ , K , ,for
( - I ZI a) ]r g. z l < - .
7l 2
where J , , ( z ) and y , ( z ) denote Bessel functions of the first and second kinds. re-
spectively. In particular, with z = x > 0 we have
Hence one can compute both J , , ( x ) and Y;, ( x ) from K , ( z x ) , for x > 0 , by (3.11).
Some known properties of a, j ( z ) . that are subsequently used. are stated by the
following lemma. See [5, Sections 9.12. 10.5 and 12.i3] and [4] for proofs. For
converiience we introduce a notation for the following set:
(3.12)
1 I ) or ( - ; < $ a n d 1
< a and - - < 0 5 -
2 2
LEMMA3.1. Let ( a ,P ) E A and let
Then: ( A )
v ( t )> 0, for 0 < t < cx,
and the moments ck with respect to u ( t ) exist and are given b y
ASYMPTOTICS O F STIELTJES CONTINUED FRACTION COEFFICIENTS 175
where
(D) T h e m i q u e S-fraction
By (3.15) arid [5. Theorem 11.3fI there exist constants E , € 1 , E L and ~ , j with
, 0 <
E < E , < E L < E 5 < T such that )
a;: (z) has asymptotic series expansions
(3.22)
(3.24) lim
I--+ x
[1 + m]
.rQ" (1.)
=2 and lim
1-+x
x2Q1"(x)
Q1(x)
= 0.
From (3.23a) and (3.24) we conclude that thcre exist constants N > 0 and B > 0
such that
BJ- a11 elementary cornputation one can show (from (3.21b) and (3.6)) that
< +
a i d . I T t k = ,-I, then xQ1(.x.) 6lcr 4. Therefore there exist constants Al > 0
and > 0 s ~ that ~ h( 2 . 1 ~ holds.
) Condition (2.3a) holds with c = cr = S = 2. It
is rcadil\ seen that the reniaining conditions for Q in Theorem 2.1 hold. Hence by
(2.7) ne arrivc at (3.17). which proves (A). (B) is an immediate consequence of (A)
ant1 Theorem 1.1 (C). 0
An application of Theorem 3.2 to modified Bessel functions K,,(z) in (3.9)
\-ields the following:
COROLLARY
3.3. Let 11 and z satzsfy
Thrv there exists n unique sequence of positave coeficients {a,, (v)):=~ such that
ASYMPTOTICS OF STIELTJES CONTINUED FRACTION COEFFICIENTS 177
where
where
n 2'
rr-r
where
(3.31 b) a , , ( u ) = b , , ( u ) / 2 . for 11=1,2.3.
Substitution of b,, ( u ) = 2a,, ( v ) in (3.30b) leads to the assertion (3.28)
EXAMPLE 3.4. Gargantini and He~irici[4]have computed thc m ~ f f i c i m t sn,, ( 0 ) .
2 5 n 5 41. in the S-fraction representation
A few of the computed values a,,(O)/n and 0.25 -a,, ( O ) / n are given below. rountlctl
off to 5 decimal places:
Table 1.
References
. . I.A. Stegun, Handbook of Mathematzcal Funct7oiis uvith F o ~ ~ r n r ~ l n s .
Graphs and Mathematzcal Tables. National Bureau of Stantlards. Appl. hlatli. Ser. 55. 1J.S.
Govt. Printing Office, Washington, D.C. (1964).
[2] Chihara, T.S.. A n Introductzon t o Orthogonal Polynomzals, h l a t h ~ ~ i ~ a tai ir d: , Ith Applirations
Ser.. Gordon and Breach, Ncw York (1978).
[3] Cizek, J . and E.R. Vrscay, Asymptotic estimation of tlic coefficielits of the col~thuetifraction
representing the Binet function, C.R. LIatli. Rep. Acad. Sci. Carlala. \hl. I\'. KO. 1 iAugust
1982), 201- 206.
178 WILLIAM B. JONES AND GUOXIANG SHEN
[ I ] (:;~rgaritil~i.I. ant1 Henrici, P.. X C:ontinuctl fraction algoritlirri for the corriputatiori of higher
tr;~lisc.c:r~tI~l~tal fuiictioris in rorr~plcxplane. LIatl~.Corrip. 21 (1967), 18 29.
[.i]Ht.rlriri. P., Applictl and Corriputational Cornplcx Analysis, \'ol. 2, Spccial Functions. Integral
Trm~sforlns.Asyrriptotics and Coritinuctl Fractions, Jolm VtTileyant1 Sons, New York (1977).
[ ( I ] Hvn-itt E. and Strorriherg K.. Real and Ahstract Analysis, Spririgcr Verlag, Kev; York. Inr.
( lM5).
[7] .Tonc:s. TI-illialri B. and 1V.J. Thron, Continueti Fractions: Analytic Thcory and Applica-
t i o ~ i s .El~cyclopotliaof hlatllcrnatics arid Its Applicatior~s,11, A(ldison TVcsley Publishing
C'i).. Reatling. h1.A [l!lXO): (listributcd now by Carrlbritlgc University Press. Ncw York.
[XI .JOII(Y. IVilliarri B. a r d TV..J. Throri. Contimicd Fractions in numerical analysis. Applioit Nu-
111i,1icaihIat11. 4 (1988). 1 U 290.
[:I] . l i ~ i ~IVilliatri
c~. B. anti TV. Van Assche. Asymptotic bellavior of the coliti~iuciifraction co-
c+ficic,nts of a class Sticltjcs transforms inclutling the Birict function, Orthogonal F~nlctions.
hlouic~ntTlici~ryand Continued Fractions: Tllcory a1x1 Applications (1V.B. .Tones ant1 A. Sri
R a ~ l g a .~ t l s . )Lcrt111.c
. Notes in Pure anti Applied hlatl~crnatics.hIarcel Delilicr. Iric. (1988),
257 271.
[ I O] Li~rcwtzc.~~. Lisa arid Haakon Waadcland, Continued Fractions with Applications, Stutlics in
C'orr~putationalhlatli.. Vol. 3. North-Holland, New York. 1992.
[l I ] Lnhinsky. D.S.. H.N. LIhaskar. E.B. Saff, A proof of Fr.cutl's conjecture for cxpoiicntial
w ~ i g l ~ tCorlstr.
s, Approx. 1 (1088). 65 83.
[12] I'crroii, O . , D7e Lehm von den Kettmhriich,en. B a r d 11. Tcuhricr. Stuttgart (1957).
[1:3] Stic1t.jt.s. T . J . , Rccllcrclles sur les fractioris coritiinic, A r m Fac. Sci. Toulouse 8 (18!94). J. 1
122: !1. 1X!)5. 1 17: (<;. \'an Dijk. ctl.). Oeuvres (.'ornpl&tcs-Coll~rt~tl Papers. \'ol. 11. Springer
\i,r.lag. Urrlil~.1!)!):1. pp. 406 570: Eriglisli trans la ti or^ or1 pp. MI!) 7-15.
[I-l] .l'llron. TV..T.. 0 1 1 Parabolic rolivcrgcricc regioris for continuctl fractioris, hIath. Zcitschr. A9
(1!).;8). 1 7 3 182.
[ I 5j \.;111 .\.;s(~li(~, TIhltcr. IIsyinytotzc.s for Orthogonal Polynon~anls.Lecturc Notes in LIathernatics
12(i5. Sprii~gcr\'tlrlag. New York (1987).
[IO] \\-all. H.S.. Anol?yt?c Theoyy of Contanued Fmctzon.~.D. \'an Nostrantl Co., Inr., New York
(l!llX).
[17] IVhittakcr. E.T. arid G.N. :Vaston, A Coimse of hlodcrn Analy,s7,s, 4th cti.. University Press,
C';rrrlbritlgc (1!)50).
b - 1 1 1 n i 1~ ~ r i dirsi w~ones@euclld.colorado.edu
A Generalization of Van Vleck's Theorem and More on
Complex Continued Fractions
A B S I . H , \ C ~In this work we give allat wc think arc basic an(l useful corl\-c.r-
gcnce tl~coryrosults for ronti~iuctlfractions It7(b,,11) with complrx elvrricmts.
\%'c also give a C O I I V L ' ~ ~ I I thcorein
~C for positi\~cPcrrorl-Caratl~do(lory
contiri-
ue(l fractions. All uf the proofs preseritcd are constructive. so t r ~ ~ i ~ c a t c3r.r.or
ion
forrnulas follow as a by-product.
1. Introduction
In our opinion, one of the niost useful and basic results in tliv anal>-tic,t licory
of cmnt,imied fractions is Van Vleck's colivcrgence theorelr~for coi~timletlfract ious
of t h r form K ( l / b , , ) wit,ll the h,, lying in a wedge. 111 Section 2 wo gi\rc, a 1 1 1 7 ~
(.onstructive proof of this t heorein that cont a i m a sharp and uscfnl trllil(.itt i o l ~
t:rror forniula. Also iri Section 2 we give a const,ructivr proof for a II(W.(,o~~\-c~gc.~i(.c>
tlicwrem for contirlued fractions K ( l / b , , )whose eleiric~nts Or, lit, in tlic riglit I i a l f ' ~ ) l i ~ ~ i ( ~
R(z) > 0 that we feel will provc t o be quit,e useful for future rescarc.11i l l the, sllt)]c~t
of contiilued fractions. \Ye have added some other results in our Vaii V1~c.kf'rirct iow.
section that wc hope will he of interest to t,he reader. In Section 3 \wrcsurrwt t li('
Uniform Circle Thcorem for which we recently gave a proof in another public.atiolr
and use it to derive a twin convergence region theoreiri for colltinurd fract ic111s
I<(l/br,) that we have called t,he lJniform Twin Circle Theorrm. Both of thew
results art. generalizations of earlier results that have an intcwstiiig history. \\.cs
close Section 3 with a new uniform simple convergence region tlicortm fhr c.ont inuc,tl
fractions h 7 ( l / b , , ) in which the boundary of each region is a c,ircle cotit a i i ~ i i ~t g 1 ~ '
origin in its interior. This rcwilt is provcd with the aid of the twin circ.1~t licolx~i~i
w(, have rneiitio~lcd.In Section 4 we usc our powerful 1966 Uniform Twill Lima(.oi~
Theorcm for conti~iutdfractions K ( c i , , / l ) to derivc a new convergcncr tlic~orc~iii
for PPC-fract,ions that 11avc played an important role ill invc>stigationsoil ~liol~i(mt
throry. ortliogonal polyliomials. a d signal processing in thc last two clecatlvs. I\(,
believc it is the first rcsult of its t>-pefor these continlied fractions.
\Ve arc, now a t a point where w~think it is appropriate t o prcwnt soli~c'ilotntioi~.
drfiilitiolis. and results that will be cmployecl in this papcr. C i v m t n-o scy11~1~.(5
where the nuinrrators { A , , ) and the denominators { B , , ) satisfy the three term
recurrence relat ions
Then the coutinuetl fraction generated by the sequence t,, is the ordered pair
a,, a1 a2 a,j
K(al,/bl,)= K - = -
k=lb,, bl + b2 + +..
f,, , also denoted bl-
is called thc nth approximant of the continued fraction. Convergence of the con-
tinued fraction t o a number f in the extended complex number system & means
convergence of the sequence { f,,) to f . The A, and B,, satisfy the important
determinant formulas
12
b=l
ab. n 2 1.
often plays a major role in convergence proofs for continued fractions. It satisfies
the simple recurrence relation
VAN VLECK'S THEOREM AND MORE ON CONTINUED FRACTIONS 181
and, moreover,
then
so that the sets K,, are nested. Hence for m > n , f,,, E h',,and
Ifrrl - fn) I dzam(K,,) =: A,,.
The sequence {A,,) is nonincreasing, so if A,, + 0, then
f := lim f, = n F K n and If - f,,/< A,,.
We are now ready t o state the convergence theorem attibuted t o Van Vleck
[13]for continued fractions
that shall occupy our attention in the next section. A nonconstructivc function-
theoretic proof of this result that uses the Stieltjes-Vitali convergencc extension
theorem may be found in each of the books on continued fractions by Jones and
Thron [3,p. 891 and by Lorentzen and Waadeland [lo. p. 1241. Gragg and IVarner
[l]gave a constructive proof of an equivalent theorem in 1983.
THEOREM
1.1 (Van Vleck 1901). Let the elements of K(l/b,,) satisfy
P := n(l+
x.
I.=%
Ib,.h,, 1 cos0)
VAN VLECK'S THEOREM AND MORE ON CONTINUED FRACTIONS 183
-
PROOF.IVith V,,, TI,and K,, related as in the Introduction we choose here t o
set V,, H. where H := { z : % ( z ) 2 0 ) . Then h',, = T , , ( H ) and it is not difficult
to show. using well known properties of linear fractional transformations, that the
diameter A,, of K,, is given by
Now let
Thesc formulas are easily scen to be truc if one uscs tlie fact that
Formulas ( 2 . 5 ) iinply that Ib,, I < m if P corivergrs. since P converges if anti olily
if C Id,, / < x.I f 'I diverges. then C jb,, I = X . since t h r div~rgenceof P implies
the convcrgr.nce of K ( l / b , , ) and b y tlw Sterii-Stolz divt~rgellcr~ theorrm I i ( l / h , , )
would diverge by oscillation if (b,,( < cc. \Ye h a w now essentially taken care of
parts (B). (C). and (D). Part ( A ) follows from seriw niajorizatioli using the filial
184 L. J . LANGE
estimate below.
Upon applying this formula to the continued fraction in Corollary 2.2 we arrive at
where y is Euler's constant. This error estimate is obviously by far not as sharp as
the one given by ( 2 . 7 ) .
The following lemma can be derived from a theorem of Gragg and Warner [l,
THEOREM, p. 11931 after replacing a,, by 1, a k by and w by f in the
statement of the theorem. The first formula involving p on page 1194 of their proof
VAN VLECK'S THEOREM AND MORE ON CONTINUED FRACTIONS 185
can be corrected by replacing the function of p that occurs there by its rrciprocal.
But for the sake of completcness, we have rhosen t o givc a proof of the lemma here.
where
>
From this formula and the relation R(h7,) R(b1,).which can be verified by induc-
tion using (2.6).it is not difficult to deduce that
where
t := 2 R ( h n ) / R ( b n >
) 2.
A simple analysis of the function F ( t ) will show that it has a unique minimum for
t > 2 at the point
we obtain
\\'it11 the aid of this lemrna. we can prove the following results:
PROOF.The proof follows easily from Lemma 2.3 and basic convergence prop-
erties of products.
and
by Theorem 2.4. After replacing the PI, in the first inequalit) in Theorem 2.4 by
their estimates above we arrive at
Furthermore, zf frl and f denote the n t h approxzmant and lzmzt of (3.21, respectzvely.
>
then, for all n 0 .
From Theorem 3.1 wc can derive the following important uniform twin conver-
gence region result for continued fractions K (lib,, ):
THEOREM 3.2 (Uniform Twin Circle Theorem). Let a and p he real n,umbers
such that p > a > 0 and let cu be a n arbitrary real number. For n 2 1 let the
elements of the continued fraction K z = l ( l / b 7 , ) satisfy the conditions
Now choose
2eP2"
d=
Jn
Then the conditions ( 3 . 4 ) for b2,,_ and b2,, become
2a 2~
>
lb~rt-1 -
Jn
2a 2~
>
Jm
which are equivalent to the single conditior~
VAN VLECK'S THEOREM AND MORE ON CONTINUED FRACTIONS 189
Our assertions concerning the convergence and its speed now follow from the Uni-
form Circle Theorem applied t o K ( l / & , )after choosing the parameter c in that
theorem in the following way:
PROOF.
This theorem is really a corollary of the Uniform Twin Circle Theorem.
Let - lEl
b ~ ~ Eand b2,, E E2, where
where
\Ye shall make use of the following theorem for continued fractions K ( a , ,11) which
we proved in 1966 [7] and to which we have recently added some luster [8]:
where
L l ( a .a ) := <
{ w : Iw6 - a ( a L- ( # ) I + aI14 a ( a L- Icr" )
LL(O,0 ) + + +
:= { w : 1%0(1 3)- (1 a ) ( a 2 11 012)1 - alu'l
- >
~ ( 1+
1 (?IL - CJL)}.
The convergence of K ( n , , / l ) is ur~zforrnwith respect to the regzons L l ( t r ,a ) and
L Z ( a ar ) an,d these regions are best twin conveTyerLce regions. F ~ ~ r t h m m o ~zf. ef,,
,
>
denotes the n t h approximant of h'(c'','l) or K ( a , , / l ) , then for all n 1
VAN VLECK'S THEOREM AND MORE ON CONTINUED FRACTIONS 191
THEOREM
4.2. Let 0 and a be real numbers .suc.h that
< a < (1+ a / .
(a(
Then the continued fraction (4.2) (and hence the continued,fi.c~ction(4.1)) conveyqes
~ ~ n z f o r m if >
l yfor n 1
1 a a" + -
121
l+n-a 2
Furthermore, the t r ~ ~ n ~ a t estimates
ion (4.3) hold. with f,, und f the 71 th approxi-
mant and value of (4.21, respectively. nnd c and d given by (4.4).
PROOF.Set
> + = . C , = / d r6 , . 4, := 1,
where the principal square root is t o be chosen Using basic 1n~qli;llitiesarid
straightforward calculatioris. it i5 not difficult to vrrifj that
&7
IcLI,-~ 4 5a dnd IC~,, * 7(1 + 0i)l > m, n = 1.2,.. .
under the conditions of our theorem. Thus we call apply Theorem 4.1 to the ( . f .
K ( c f , / l ) t o reap the conclusions we have made 111 Tlleorcm 4.2 m t l drrixe nt the
end of our brief argument.
References
[I] W. B. Gragg and D. D. Warner, Two constructive results in continrictl fractions, S1Ail.l .I.
Numer. Anal., 20 (1!)X:i), 1187 1197.
[2] K. L. Hillain, Sorm convergence criteria for continued fractions, Doctoral Theszs. U ~ ~ i v c r s i t y
of Colorado Boulder, 1962.
(31 W. B. Jorics and W. J. T h r o i ~ . Contzwucd Fractzons: Annlytzc Theory and Appltrotzons.
Encylopedia of hIathernatics and its Applications . Vol. 11. Ad(liso11 h'eslcy. Reading. blA.
1980.
[4] W. B. Jones. 0. Njdstati, anti W. J . Throii, Continued fractions associatc!d with trigoriornetric
arid other strong ~norricritproblcrns, Constr. Appms. 2 (1986), 197 211.
[ii]W. B. Jones, 0. Njastad, ailti W. .l. Thron, Perron Carath/'codory continucti fractioiis. Rn-
t l m ~ a Approrarrmtzon
l n7,d Ap2lzcc~tionsan AIathmnat~csand Physzcs, Lecture Notes in hlatll..
\'ol. 1237, 188 206, Springer, Berlin New York, 1987.
192 L. J. LANGE
Xin Li
ABSTRACT.The convergence of Laurent polynomials that interpolate func-
tions on the boundary of a circular annulus is studied. The points of in-
terpolation are chosen t o be uniformly distributed on the two circles of the
boundary. It is proved that, for functions analytic on the closure of the an-
nulus, the sequence of interpolating Laurent polynomials displays a type of
maximal convergence. It is demonstrated that, in general. there is no equicon-
vergence between the interpolating Laurent polynomials and the partial sums
of the Laurent expansion of the functions analytic in the annulus. Finally, it is
shown that the norm of thc projection operator induced by the interpolating
Laurent polynomials is of the logarithmic order.
1. Introduction
Let T ( p ) = { z E @ : lzl = p ) , the circle centered at z = 0 and of radius p > 0
in the complex plane @. Throughout this paper. we will assume p2 > pl > 0. We
write A(p1. p2) for the open circular annulus { z E @ : pl < lzl < p r ) . A ( p l ,p r ) for
its closure, and d A ( p 1 , p2) for its boundary. For n7, n >
0. let C(nr. n ) denote the
set of Laurent polynomials of the form x;=_,,,
n ~z h ( a n E @). For each positive
integer p, let ui,, = eLT1'p. Define a set of points as follows:
ZA. = p ~ u i h ~k ,= 1%
2, ...%rn,
and. with a fixed a E [ 0 , 2 r / ( n+ 1)).
, 1 , 2 . ..., rn.
L,,, ,,(zx) = f ~ ( z ~k) =
This paper is a continuation of the work of Llason [4, 51, Brutman. Vcrtesi and
Xu [2], and Pan [6], where the main concern is on t,he estimation of the norni
of the operator f H L l l , r , ~ l ( fzl;. f L ) (with fJ = f l . l r ( O l ) .j = 1 . 2 ) for f anall-tic
in A ( p - I , p ) and contiriuous on A ( p p l . p ) ( p > 1 ) . It is shown that the norm is
of the logarithnlic order as n -. m. As a consequence, it is easy t o verify the
following cor~vergenccresult: W h e n f i s a n a l y t i c o n t h e c m n u l u s A ( p 1 . p n ) , t h e n
L , , . , , I ( s ;f 1 : 5 2 ) (iuith ,fJ = f j = 1 , 2 ) converges u n i f o r m l y to f o n A ( p l p, 2 )
u s n -t x. The novelty of this paper is that the general Laurent polynornials
L,,l,ll are considered and convergence by such interpolating Laurent polynornials
is established under fairly general conditions. The main results include a type of
maximal convergence of L,l,,llas nlrn 4 r ( r > 0) and rn n + CG. The paper +
is organized as follows. In Sect,ion 2 , we prove that {L,,,,,, ( z ; f l , f2)) is convergent
for z E A ( p 1 , p 2 ) : when m 3 x and n -+ m , as long as f l and f 2 are Rirrnann
integrable functions. In Sections 3, we prove the maximal convergence theorem
ment,iond above. In Section 4. we study whether t,he equiconvergence phenomenon
occurs bet,ween the interpolating Laurent polynomials and the partial sums of the
Laurent expansions of functions analytic on the annulus ( p l , p 2 ) , as Walsh proved
for int,erpolating polyno~nialsand partial slims of the Taylor expansions of functions
analytic on t h r unit disk. Finally, in Section 5, we give the as>-rnpt,oticorder of the
norm of L ,,,.,, as nl + o and n + x.
2. Convergence in A ( p I ,p 2 )
+
Xote that z v 1 L r 1 ,, ( z ) is a polynomial of degree m n that interpolates z n 7 f l ( z )
and zr" f L ( z ) at { Z ~ } Y =and~ { z , , , + ~ } ; : : , respectively. By Lagrange's interpolation
formula, wc have
where
Note that
and
Thus.
CONVERGENCE O F INTERPOLATING LAURENT POLYNOMIALS 195
which equals
since the last sum is a Rirrrlarin surn of the Riernanri integrable function
+ + + +
since z; = p ~ + ' r 7 ( " +I)'' for k = Trr 1,m 2, ..., m n 1. The factor in front of
the sum has limit -1, and, as in the estirriation of L 1 . we can verify that the sum
approaches
- x L2 =
lirn
,,I
82 ..
196 XIN LI
Therefore, by the estimates of L 1 and L z , we deduce that the limit ,&& L,,, , , ( z )
,,-x
,,,lim
-x L,,, ,, ( z ) = --
8,-x
- x L r r L, , ( z ) = f ( z ) ,
lim
r,,
r, - x
2 E A h , p2).
and
.
T h e n L,,, (2) converges t o f (z) for each z E A(61, p ~ a)s rr/m + T cmd m +n 4x
Furthermore, zf f h a s a szngular poznt o n dA(p;. p;) a n d zj r = log(pi/pl)/
log(p2/p$). t h e n p1 = pi a n d p2 = p i , a n d
Therefore
198 XIN LI
Then
So. j1= pi alld j2= Using these and (3.5) in (3.4). we get. for z E A ( P \ , P > ) .
lim sup
r,
,, 1
rr,-r
i - x
nisi
2 € T ( / l l (12)
( 2 ) 1%
l j ( z ) - L~~~ I lnax
I-"'
I 12
{r, (T)
P/1 P',
-,
zI P1 P2
I }
Assembling cverything together. when r satisfies (3.5). we have proved the following
inequality
Now, to finish the proof of the theorem. we show that the equality must hold in
(3.6) if f has a singular point on ~ A ( P ',,p;). We need a version of Bernstein's
inequality ([3,51.41 or [7, $4.61) for Laurent polynomials.
A Let 0 < R1 < RL. Then for p,,, ,, E C ( m .T I ) , we h a w
L E ~ I M3.2.
and
We now continue our proof of Theorcrn 3.1. U'e need to show that the equality
holds in (3.6). Assume, to the contrary. that, for soinc T E (0.1) arid for all
z E x(pl,P~ we) have
.
Then, in particular,
I +'
< T -.
4
lim sup I f ( z ) - L,,, I, , ,1 1
rn - x
(2)
PI
200 XIN LI
for m large enough. Let R 2 > p2. B y (3.8) in Lemma 4 and (3.9) above. we have
and so.
Similarly, from (3.9). by using (3.7) in Lemma 3.2, we can show t h a t , for
Rl E ( 7 4 ,pi ),
Therefore, by using the maximum modulus principle, we know that the sequence
{ L r r , , l r r r Lconverges
~(~)) geometrically for R1 4 lzl <
R 2 . This implies that f has an
>
analytic continuation at least t o the annulus A ( R 1 ,R 2 ) a A ( p { ,pb), which con-
tradicts the assumption that f has a singular point on d A ( p { ,p i ) . So the equality
in (3.6) must hold, which completes the proof of Theorem 3.1.
When p1 ~2 = 1>we have the followirig result, which strengthens the implied
convergence result of [2, 4, 5, B] as mentioned in Section 1.
COROLLARY3.3. Let 1 < p < R . I f f i s analytic i n A ( R 1 R, ) and zfL,,,,,-I ( z )
i s t h e Laurent polynomial that interpolates f a t { z k } f " , w i t h pl = p p l a n d pn = p,
then
liin max If ( z ) - L,,,,- 1 (z)l"iL < f
IL-x ZI=,,,~-~ - R'
4. No Equiconvergence
We continue to use the same assumptions and notations as in Theorem 3.1.
Let xr="=_, f A z k be the Laurent expansion of f on A(pi, pi). Set l,,, ,,(z) =
C;=-,,,f L z A the
, (m. n)-Laurent polynomial of f . According to [I]and [4]. the
operator that maps f to l,, ,, is a projection with minimal norm among all projec-
tions from the space of functions analytic in A(pl. p2) and continuous on X(pl, p ~ )
into L ( m ,n) with respect to the sup-norm. Since the operator mapping f to L,, ,
is also a projection, it is natural t o compare L,,, ,, with l,, , . By Theorem 3.1 and
the well-known property of Laurent expansion, we know that
Is it possible to show that L,,, "(2) I,,, ,,(z) is convergent in an annulus that is
-
bigger than A(p{,pb)? This is true for the difference of polynomials interpolating
at the roots of unity and the partial sums of the Taylor expansion of a function
analytic on the unit disk as proved by Walsh, a result usually referred t o as the
equiconvergence theorem. In this section, we show that no such results hold between
the interpolating Laurent polynomials L,,, ,, (z) and the partial sums of the Laurent
expansion 1, ,(z). Consider the function
Then f is analytic in A(1/4,4) and p', = 114 and p; = 4. Let's take the points
of interpolation on T(1/2) and T(2). that is, let zk = 4 / 2 and z,,+h = ~ u J ~h~ w , , ,
k = 1 , 2 , .... n. Then the interpolating Laurent polynomial L,, ,, 1 ( 2 ) is given by
-
and
Therefore. there is no equiconvergence between L,,,,,- 1 and l,, ,,-I in any annulus
bigger than A(1/4,4).
202 XIN LI
iVe rcrnark that it is possible to split L,, , , I ( z ) into two parts so that one part
does equiconverge with Laurent polynomial 1,) ,,( z ) in A ( 1 / 4 , 8 ) while the other part
equiconverges with 1,)- 1 - 1 ( z ) in A ( 1 / 8 , 4 ) . As the above example points out, in
general. there is 110 combined equiconvergence in any annulus containing A ( 1 / 4 , 4 ) .
arid
rl+l
C ( n+ l ) p ; lp;")(Z71+l p;+l
P;~I(Zn~ - e ~ ( f ~ + ' ) ( k
A, (Z) := I
h=l
~ I ~ ~ ' ( ~p;")(z
;+~ - - zm+~)l'
IVe now estimate X l ( 2 ) and A2 ( z ) when IzI = pl and z / = pz.
Let z = p l e t o for some 6 E [ O , ~ T ] W
. e have
and
and
2 p I ~sin I
n+l
&(z) = C
,=, ( n + l)lpl p2e'(d+cl-H)I (1
-
LA
+ (2)
0( IT)}
-
m
I sin F1
C ml sin(:
I;= 1
-
Q= x
r,,
max Xl(z) =
0€[0.2~] 7i- 7r
So, the maximum value of the above sum is attained asymptotically at 0 = 7rlnl.
Thus. we have, as m, n + cm,
2
max
14=p1
[ X l (z) + A2 (z)] = -
7'r
log m +
From these two asymptotic estimates of X I (z)+X2 ( z ) ,we have the following estimate
for t h r order of the norm of 11 Lrl, 11.
5.1. We have
THEOREM
IIL,,,.,, I/ - 2
lim - -
If n / m -+ 1, t h e n
It is known (cf. [I,41) that the norm of the Laurent projection f (z) H lm ( 2 )
+ + +
is (4/7r2) log(rn 72 1) O(1). So, according to Mason [4], L, ,, and I , ,, are
both near-minimax approximation within a relative distance of the same order as
+
m , n 4 m. Finally, we remark that (5.1) extends (at least for m n large enough)
Mason's conjecture as verified by Pan [6].which states
References
[I] Brutman, L., On the polynomial and rational projections in the complex plane, SIAM J .
Numer. Anal. 17(1980), 366-372.
[2] Brutman, L., P. Vertesi, and Y. Xu, Interpolation by polynomials in z and z - on a n Annulus,
IhlA J. Nurner. Anal, lO(1990). 235-241.
[3] Gaier, D., Lectures o n Complex Approximation, Birkhauser, Boston, 1987.
[4] hlason, J.C., Near-best L , and L1 approximations t o analytic functions on two dimensional
regions, zn Multivariate Approximat ion (Ed. Handscomb, D.C.), pp. 115-135, Academic Press,
London, 1978.
[5] hlason, J.C., Near-minimax interpolation by a polynomial in z and z p l on a circular annulus,
IhlA J. Numer. Anal., 1(1981), 359-367.
[6] Pan, K . , On Mason's conjecture concerning interpolation by polynomials in z and z p l on a n
annulus, IhlA J. Nurner. Anal., 14(1994), 599-604.
[7] Walsh, J.L., Interpolation and Approximation, Amer. Math. Soc. Coll. Pubs., 20, 1960.
DEPZRTX~ENT
OF MATHEXIAIICS.
I'NIL'ER~ITY
OF C C N I R A FLORIDA.
L ORLANDO,
FL 32816
E - m a d address xlllpegasus . cc .ucf . edu
Contemporary M n t h e r n a t i c s
Volume 236. 1999
Lisa Lorentzen
A n S r n ~ c It
~ . has been conjectured by Ruscheweyh and the author that if
01 # V C_ @ is a n open, bounded set, and E := {a E @ : a / ( l + V ) 2 V ) contains
at least two elements, then E is a convergence set for continued fractions
K ( a , , / l ) ; that is, K ( a , , / l ) converges if all a,, E E. We proved the convergence
if the poles {-h,, ) of
1. Introduction
,-.
A set V 2 C := C U {m) containing and omitting at least one element. is a
simple value set for the continued fraction
a,
(1.2) S,~(Z)
:= -E V for all 2 E V and all n E N.
l+z
1991 Mathematzcs Subject Classzficatzon. Primary 40A15; Secondary 30D05.
Key words and phrases. Convergence. Continued fractions.
@ 1989 Amcricar~ R l a t h c ~ r r l ; t t ~ c . aSor.ivty
l
205
206 LISA LORENTZEN
nl a2 a,,
s l o s 2 0 . . .os,,(z) for n = 1 , 2 , 3 , . . .
(1.3) Si,(z) := -
1 + i + . . . =+ G
of K ( n , ,11) are then contained in the nested sets
closed curve in C.
0 w is an znner poznt of the simple arc r if it is the image of a n inner point in
I C R under a continuous univalent mapping cp such that ~ ( 1=)l?.
0 The parametrization of the boundary (or parts of the boundary) of a domain
V is negatively oriented if V lies t o the right when we follow iJV with increasing
parameter. Similarly, the paranietrizat,ion is positively orzented if V lies t o t,he
left when we follow d V with increasing parameter. If it is not, clear which
component of e
\ d V we refer to, we say that the boundary curve (arc) is
negatively or posit,ively orient,ed with respect t,o the domain V.
r is an arc in @ represented by z = z ( t ) , then
A
0 If
denotes the angle of inclination and the signed curvature of r a t the point
+
z(to) m , if they exist. If z ( t ) only has one-sided derivatives of second order
a t t = t o , we define the corresponding orie-sided curvatures hi(T, i ( t o ) ) arid
n ( r , i ( t o ) ) . For convenience we also write h i ( r , i ( t o ) ) to mean eit,her onc of
them. If I' = a V of a domain V. then (1.6) is based on n e g ~ t ~ i vorientatione of
a V with respect t o V.
0 i and 2 denote t,he first and sccond order clerivat,ives of z with respect to t.
-
- -
F ( x ) = o(g(x)) as s -+ a means that limr,,{b'(z)/g(x)} = 0.
a,, b,, shall mean that a,, b,, as n + m; that is, lim,, ,,(a,, /hi,) = 1.
CONVERGENCE CRITERIA FOR CONTINUED FRACTIONS 207
rl -
liin diam S,, (V)= 0,
x
then K (a,,/ 1 ) converges (i.e. the sequence of approximants {S,, (0)) converges in
e, which means that either { S , , ( O ) ) converges to a finite number or { l / S , , ( O ) )
converges t o 0 ) . Actuallv (1.7) combined with the nesttdness (1.3) implies that
{ S , , ( z ) ) converges uniformly in V to a constant function S , ( 2 ) = f E V . This
kind of unifornlity has not been studied much until recently, but we think it deserves
attention for at least two reasons: (i) It is of interest if one uses the approxiinants
{S,, (z,, )) instead of the classical ones { S , ,( 0 ) ) . (ii) It is important when the conver-
gence theory for continued fractions is extended to more general functions { s f , ( 2 ) )
A survey of results in this area can be found in [ 1 8 ] . We say that V is a unzforrrr
ualue set for K ( a l , / l )if (1.7) holds. If (1.7) fails. we may still be able to prove that
{S,,( 0 ) ) converges by other arguments.
It is standard to define the element set E correspondmg to V as the set
(1.8) E := { a E @ : a / ( l + V) CV).
Then E is a closed set in G ( E may be empty). and V is the value set for mery
continued fraction K ( a , , / l ) from E (i.e.. all a,, E E ) . We say that V is a value set
for E . (See for instance [ Z l ] . )The elenlent set E' in (1.8) characterizes t h r class of
self-mappings (1.2) of V. Note that we allow 0 E E, even though all a,, # 0 in a
continued fraction K ( a , , / l ) .
iVe have now set the stage for rnodern convergence tlleorv for contiriued frac-
tions based on value sets. The ideas cvolved around 1940 in works by Scott and
Wall [24].and were irnmediately taken up by others. In particular m7.J. Thron has
produced a long list of convergence criteria based on different value sets. some of
these in collaboration with his earlier students Mr. B. Jones and L. J . Lange. (See
for instance [7].)
In the following we shall give three examples of such convergence criteria. The
first one is interesting because the conditions on V are so mild:
n
aEE : -- na * f ~0) with O'V := i ) n~( 1 - all).
l+d*V
illoreover, V , , is a uniform value set for E. But the convergence is also uniform
with respcct t,o K ( a , , / l ) from by (1.12). A set E C C with this propert,y is called
a m i f o r m conwelgence set for continued fractions K ( a , , / l ) .
\Ye can also say that rvery continued fraction K ( a , , / l ) from P,, with a bounded
element, seclucncr { n , , } converges. We say t,hat an element set E with this property
is a c o n v r ~ y e n c eset locc~llyfor continued fractions K ( a , / l ) . If bounded subsets of
E are u n z f o r n ~convergence set,s. we say t,hat I3 is a locally unzforrn con,vergence set.
Note tliat with the notation from Theorem 1.1. a*V,, = aV,, and P,* = dP, in
Thron's Uniform Parabola Theorem, but still we do not need any restrictions on
A
{ a , , } c E.
The third convergence rcsult is clue t o Ant,onio C6rdova Y4venes. and can be
found in his thesis [I].It gerleralizes parts of Theorem 1.3. C6rdova started with a
generalization of Tliron's halfplane V,, in (1.10). He had the following restrictions
on his set V:
Y 1: If is i~ ,Jordan doinairl in (C with V = \ (- 1 - V ) ,0 E V, - $ E d V , x E dV
and a smooth bo11ntlar~- dV\{m) with continuous curvature. That is, it admits
a para~nctrizafion-. = z ( t ) , where r;(dV, z ( t ) ) exists and is continuous. (See
(la.)
Y2: I f { E iJV \ {x), +
the11 [ - C 2 / ( l d V ) ] n dl/' = {C).
Y3: If <
E d V \ {x) and i3V is negative1~-oriented with respect to V b,y the
z = z(t), t 1 1 r ~ 1
pi".~l~letrizi~tiOll
CONVERGENCE CRITERIA FOR CONTINUED FRACTIONS 209
We observe that by Y1, i3V = 8" V. Under these conditions C6rdova proved:
THEOREM1.4. (C6rdova's Convergence Theorem 111). Let V satzsfy YlGY3,
> 0 be gzven. Then the correspondzng element set 2s gzum by
and let A f
are the corresponding element sets, and prove a similar result for this situation. It
El and EL are closed sets in C ,and that the special case where
is evident that a l s ~
= Vl brings us back t o the original set-up. Twin regions have been exterisively
studied by Thron, Jones and Lauge. (See for instance [6, 9, 10, 12, 25, 271 and
the references therein.)
210 LISA LORENTZEN
In Section 4 we study some optimality propert,ies for t,his special case. In Section
5 we look closer at our va1il.e sets. S e ~ t ~ i o n6,s 7. 8, 9 arid 10 contain the proofs of
our five classes of convergence results in Section 2. Finally. in Section 11 we apply
Corollary 2.9 and t,he generalization of C6rdova3s Convergence Theorem t o twin
versio~lsof Lange's strip regions [ll].
2. Main results
Let (L5, I/;) be givcn twin value sets with corresponding element sets ( E l . E L )
given by (1.18). IVhcn are ( E l , E L ) convergence sets locally? And when are they
locally uniform convergence sets? \Ve shall present some conditions on (V;], Vl)
which imply that ( E l ,E L ) have such properties. For coiwenicnce we write VJ,, =
MI, VLr,_+l= Vl arid := E l . := EL for all n. We shall also write
W,, := @ \ V,, throughout this paper. We shall frequently meet other notations such
as (Uo, U1). (i3*vl, i3*Vl), etc for twin sets. Also for these we lct {U,,), {d*V,,),
etc, denote the corresponding 2-periodic sequences.
Let K ( a , , / l ) be a continued fraction frorn ( E l , E l ) . The nested sets (1.4) now
take the form
- -
l()UVl and { n , , ) 2s bounded. Then (V;),V l ) are bounded twzn v d u e sets for. K ( a , , / l )
g w n b?y
A A A
THEOREM 2.3. Let K/;,and Vl be nonempty domazns zn C such that \VJ has
one and only one unbounded component each for j = 0 , 1 , and let K O and K 1 be
+
A
for j = 1.2. We shall also use the strange, but standard notation -h,, := S r y l ( m )
for the pole of S,,. Clearly. -h,, # if is bounded, but we may have that
d(-h,, . F ) + 0. This may cause problems when we try to prove that diam S,, ( F )
+ 0. The significance of d*V; and E,* lies in thr fact that if d ( - h , , , dV,,) + 0.
tlirn d(-h,, , d*V,,) + 0, and this can only happen if d(a,, . E,:) + 0 under our
conditions. (See Leniina 7.2 and Remark R7.1.) IVe get:
We recognize this as a twin version of C6rdova.s Coridition Y2, and wc call this
the Cdrdova Condztzon. The difference is that d * V = d V in C6rdova'5 case, and
that C6rdova required the existence of some a E E, depending on (, such that
+
a / ( l 3*V)n d* V = {C) for every C E d V . It is clear that if C * 1 holds. thcn tlit
critical parts of E J and V , are
Actually, d ( - h,, , a*V,,) + 0 implies that d(-h,, , d i V , , ) + 0 when lim sup lh,,I < cc
and C*1 holds. (See Lemma 8.1 and Remark R8.1.2.) It is worth noticing that we
A
e*1 is however more restrictive that C*1. (See Remark R8.1.4.) Hence we shall
call e * 1 the strong Cdrdova Condltmn. Evidently. e * 1 holds trivially if a*X/;, has
at most one point.
Our second condition depmds on a given A f > 0:
214 LISA LORENTZEN
CS12: There exist positive constants p, q and $ (depending on (h, Vl) and AI).
such that the followiilg hold for every E 8t~, {,
with / 5 Al: = 0,l:
(a): r, ,, :=aV, n B ( { , , p ) is a simple and smooth arc. and to every
< E TI, there exists a circular disk Di in @ \ V
A
Condition CT12(n) clearly holds if r3,,has bounded curvature ~ ( r , <): /r;(rl <)I,,, ,.
5 l l q . but it may also hold in cases where n ( r , ,,,<) is not defined.
To uriderstand Condition CT12(b), we first observe that if Re{(P< - < ) / < ( I +
C)) = 0. and thus the line LC:z = <+
((1 + <)ti t E R. is tangent t o dD, at C.
then Dc is invariant lirlder the parabolic linear fractional transformation
Actually, everv circular disk D < with boundary tangent t o L, at { is invariant under
this napp ping. Hence. if C = <, ~ .there exists such a Ei C V , . then the
E 8 t ~and
2 -prriodic continued fraction
c
A A
COROLLARY 2.8. Let ( v l , Vl) be t w i n valu,e sets such that the corresponding
elernrnt .sets ( E l . E 2 ) hoth contain at least t,wo elements. I f -1 $! /;, U and
C"1 nnd C * 2 hold, t h e n ( E l : E 2 ) are t w i n c o n ~ ~ e n ~ e nsets
c e locally for continued
fractions K ( a , , / l ) .
For the caw where vl
= Vl := V we have dTV .= (2 E d * V : z ' E E and
-(1 + 2)' E E). and olir r ~ s u l t sreduce to:
CONVERGENCE CRITERIA FOR CONTINUED FRACTIONS 215
COROLLARY 2.9. Let V be a value set for h ' ( a , , / l ) such that th,e following
conditions hold:
( i ) The cor.responding element set E con,tains at least two elemxnts.
(12) Condition C * l holds with V;] = Vl = V .
(iii) For every A 1 > 0 there exist positive constants p, q and [ such th,at the
following h,old for every E d i V n B ( 0 ,M ) :
(a) F, := d V n B ( < o , ~is) a simple, smooth arc, and to every ( E F p
there is a circu10,r disk DC in \ V0 with radius q and with on its <
boundary.
<
(b) For every E ,'I we have cos[arg{(Pc - ()/(I + >
<)}I (, where PC
zs the center of D i .
If either {a,,) is bounded and -1 @ V , or if V is bounded, then K ( a , , / l ) con,veyqes.
Theorem 2.7 was inspired by C6rdova's ideas. The conditions in Corollary
2.9 are milder than the ones in Cordova's Convergence Theorem, but we do not
get that V is a unzform value set for bounded subsets of E. If we only want to
prove convergence of K ( n , , / l ) . WP can do so by even simpler means in the following
situation:
THEOREXI 2.10. Let ( q l , Vl) be fwzn value sets for K ( a , , / l ) , where oo E dQl n
aV,, 0 @ afi U dVl, -1 @ d V l U dVl and sup / a , / < 1\/f < oo. If there exzst an
MI > 0 and a subsequence { n k ) of N such that mk := n k - n k - 1 1s bounded and
-
s!,:'::,(V,,,+,) c BI V,,, for all I;, where Bn are czrcular dzsks dependzng on
K ( a , , / l ) wzth radzz 5 M I , then K ( a , / l ) converges.
The proof of this result is found in Section 9. If we set all mk := m in this
theorem and require that every continued fraction K (a,, / I ) with value sets (V;,, Vl)
has this inclusion property. we get the following alternative to Theorrm 2.7:
COROLLARY 2.11. Let ('I/;],Vl) be tuizn value sets wzth x E aV;, n dVl. 0 @
dV;) u dVl and -1 @ aV;] U dV1, and let ( E l , Ez) be the correspondzng element spts
A
(See the proof of Theorem 2.7.) To guarantee that (2.9) holds, we shall impose
>
some extra conditions. Let Rf := {z E R : x O), and let R denote the class
(2.11)
z = z, (t, w,) := w, + eq,l {t + i&t2 + gJ ( t ) )
(negative orientation w.r. t. V,)
1. Condition C i 1 is consistent with the set-up in this paper. Neither the condition
that &/;,and Vl are open. nor the condition that &,n(- 1- Vl ) = 0 is particularly
restrictive. (See Remark R2.1.2, Proposition 5.4 and Remark R5.1.)
2. Condition C T 2only concerns the parts r J ,of dV,. This is sufficient since we
can prove that if d(-h,, dV,,) + 0, then d(-h,,, drV,,) + 0. (See Lemma 7.2,
Lemma 8.1 and Remark R8.1.2.)
,
Part (a) demands that these arcs r, are smooth with well defined curva-
ture. (That z,(t. w,) is twice differentiable at t = 0 follows from the fact that
<
ly, (t)l d ( t ) l t ( . Hence, the curvature exists at z = w,. This holds true for all
w, E T, ,.) The condition ui(t)ltl 5 f may be replaced by Reg,(t) -(1 - E ) >
for some 0 < E < 1, independent of w, and . <,
We do not loose any generality by assuming that the parametrization has
the form (2.11), since the angle of inclirmtion y(dV,, w,) and the curvature
~(84 w,), are assumed to exist. This can be seen as follows: If
ZJ ( s , wI) := w1 + p l Y {as
~ + bs' + z[jJi!,sL + F, ( s ) ) ;
A
a > 0. b E R.
then the substitution t := as + bs2 brings it over to the form (2.11), at least
locally.
3. Straightforward computation shows that
CONVERGENCE CRITERIA FOR CONTINUED FRACTIONS 217
- w;
(2.13) I ~ ( d V , , w , ) l5 l / q and K
nB(<,p):C E dtV,-l}.
T h e n ( E l ,E2) are twin convergence sets locally and (v),
V l )are uniform value sets
for every continued fraction K ( a , , / l ) with bounded element sequence from (El,& ) .
This theorem is proved in Section 10. The restriction to ( E l , El) makes sure
that c, E 6 - 1 when c, is close to d + ~ , - This~ . is a technical condition which may
be superfluous. Actually, we shall see in Section 3 (Theorem 3.1 in the particular
A
but what do these sets look like? Lange [ll]has pointed out that it is often easier
to describe the sets
nlentioned. we let If7,, := (C \ V,, for all n . both in this section and throughout
this paper. We shall see later (Proposition 5.4) that it is natural to assume that
1 -Vl) = 8 : that is,
Actually, it follows from Proposition 5.4 that if (&), Vl) are twin value sets for
h ' ( a , , / l ) , then so are also (Uo, U 1 ) . where
if Uo # e
8. Clearly, U1 C \ (-1 - Uo) We therefore get:
COROLLARY 3.2. Let ( - C f . - C f ) be the element sets correspondzng to the
grven value sets (&), Vl). Then C , C q n ( - n , ) w h e r e 0 , := l + f l / I ; U ( - l - V , - ~ )
for J = 1 , 2 .
PROOF. \Ve first note that 0, = 1 +e
\ UJ, where U, is given by (3.4) for
J = 0,1. Thc rcsult holds trivially if CJ = 8. Lct CJ # 8. Thc rcsult is also trivial
A
if U j = 8. since then also UJ+ 1 = 8. and thus f2, = @. If UJ # 8,then (Uo, UI ) arc
dl50 twill value wts for (-Cf , C f ) , and the result follows from Theorem 3.1.
CONVERGENCE CRITERIA FOR CONTINUED FRACTIONS 219
In the particular case where we have pi) = -1 - vl,then Theorem 3.1 shows
that
It is a remarkable fact that if we add a few conditions, then we can actually obtain
equality in (3.5). The idea is due t o Cbrdova, [ I ] .I11 the twin situation it takes the
form:
Since 6
l n r ( 0 ) = r l ( Q ) / r ( 0 )is strictly monotone, this can only happen if Q1 = QL.
which again gives that 81 = QO. 0
R E ~ I A H KR3.1.
S
1. The condition that z = r(Q)e7'; Q E I is a parametrization of r means that
r is starshaped with respect t o the origin. We do not really need that r is
srnooth t o conclude (3.7). The monotonicity of r l ( Q ) / r ( Q )can be replaced by
+
rnonotonicity of the function F ( Q ) := l n r ( 8 h ) - l n r ( 0 ) for every h > O for
all Q E I for which also 0 h E I . +
2. In accordance with Lange [ll]. we say that the arc I?: starshaped with respect
t o 0, is loyartfhmically convex if it satisfie5 the c o n d ~ t ~ o nofs Proposition 3.4.
Lye extend this t o arcs r which are starshaped with respect t o the origin. for
u-h~chF ( Q ) descr~bedabove, is strictly monotone for every fixed h > 0. It
follows then that the conclusion of Theorem 3.3 holds if q/;, is a Jordan domain
with 0 E V;], -1 @ Vo. x E aq, and logar~thrn~cally convex boundary which is
starshaped with respect t o -1 and 0. and also Vl := @ \ (-1 V o ) is a Jordan
domain.
3. If r is a smooth Jordan curve with O @ r and differentiable parametric repre-
+
sentation z = z(f) = x ( t ) zy(t). t E T i R. then
d 4t)
- arg z(t) = Irn -,
dt 4t)
and thus r is starshaped with respect t o the origin if Im(iZ) > 0 for all t ,
or I m ( i ~ <
) 0 for all t . Then we can choose Q := arg z ( t ) to be a rnonotone
mapping. Then r ( O ( t ) ):= Iz(f) is well defined. and we have
cl rl(Q) - Re(i2)
-lnr(Q)=- - --- = cot [ a r g ( i ~ ) ]
0% r($) Irn(i2)
(3.11) 3 + ( t ) := lini
i(t +at) - i(t)
, z P ( t ) : = lirn
i(t +a t ) - i(t)
-Itdo+ a t A ~ - ~ ~ - a t
CONVERGENCE CRITERIA FOR CONTINUED FRACTIONS 221
and
@ and Vl := C \ (-1 - x)
be ,Jordan donmin,s such
th,nt C1 and C2 given by (4.4) are strip regions. If (3.6) and C i 2 hold. then ( E l ,E2)
= (-CP, -C$) are twin convergence sets locally, 7uherea.s (El ~ { t r ) . E2) with a E C\
El are not twin convergence sets locnlly for continued fractions K ( n , , / l ) . Moreover.
(fi,Vl) are ,value sets for ( E l . &).
The second kind of optimality we shall st,udy is t,he following: IVe say that V
is the best limit set corresponding t o an element set E if V is a value set for f3.
and every 7u E V \ (0. m ) is the value of a convergent continued fraction from E .
(We disregard 0 and x in this definition since the elements a,, of K ( a , , / l ) are not
allowed t o t,ake these values.)
This is a rather recent notion. As far as we know: it was first used in [3].where
the idea of convergence based on { S , , ( z , , ) )was ~ t a r t ~ e dEarlier.
. the standard
approach had been t o concentrat,e on {S,,(O)), which made it natural t,o say that
V is t,he best linlue set corresponding t o E if cvery ru E V \ (0. w}is the value of a n
approximarlt S,, (0) of a continued fraction from E . If V is a Jordan domain with
0 E. V , t,hen V is the best value set, if and only if V is the best limit set.
It was proved alrcadl- in 1942 [13]that the halfplane V,, is the best value set for
P,, in (1.11) when cu = 0. and the same was proved for general a: - ~ / 2< (2 < ~ / 2 ,
in 1953 [8]. So, indeed. V , is the best limit set corresponding t o P,, for these
a's. Actually, V,, is the best limit set for the very thin element set dl',. This
remarkable result was proved bj- i'aadeland in 1989 1291. He even proved that
every 711 E \ (0, x)is t,he valuo of a 2--periodic continued fraction of parabolic
type. from dP,,. where the period begins after the first element. We shall see t,hat,
t,hc same holds true undcr more general condit,ions:
@ \ (-1 - V o ) ,and let the element sets (El, E2) corwsponding to (I/;,. Vl) halie
- (dVIp1)' \ {co) C E, for j = 1 , 2 . Finally, assume that arg <inm-eases from tr to
(2 + T as ( follows the boundary a v , , where a is a jized c o n s t m t 0 < a < T and
( ( 0 ) = < ( a+ T) = m . T h e n every .ul E Vo \ (0, m) zs the v n l ~ ~o,f
e n 2-p~riodic
continued frc~ctionK ( n , , / l ) ,from (El,E2),uihere
(4.5) a E - ( , -(1
n ~ ,= + <)', aL,,+1 = -<'for all n , for a C E d h .
This means of course that every w E V o\ (0, m ) is the value of a convergent
+
continued fraction from ( E l , E L ) . If Vl = E 2 / ( 1 v , ) , wc also have that every
E \ (0, m ) is tlw value of a convergent continued fraction from ( E L ,E l ) . Lye
say that the twin value sets (v). ~ twan 1zrn7t sets for ( E l ,EL).
Vl) are t h best Note
that we do not requirc that every continued fraction from (-C;. -Cj) converges.
Actually. Theorem 4.3 shows that if 8%)is also starshaped with respect t o -1, then
(Vo.V 1 ) are the bcst limit scts for ( d E I , d E 2 ) .
+
PROOF.Let z = C,(fl) := r(0)el' for cr < 0 < n T be a paramctri~ation
of ax,. andlet w := pe" E \ (0, x ) be arbitrarily chosen. wtlerc ti > 0 and
0 Ip < 2T. IR want to prove that
224 LISA LORENTZEN
This shows in particular that the closure of the value set,s in Corollary 4.2 have
this property if their boundaries are starsllapetl with respect to the origin, since
this "double starshapenrss" implies that arg( does not vary too much along the
boundaries.
Our third optimalitg propcrty is uniqueness of the value sets. Value sets V for
a given element set E are in gcncral not unique. This is for iristarice evident from
I'roposition 2.2. However. under the following conditions we have uniqueness:
THEOREM 4.4. Let (ql, V l )be closed best limit s d s for ( E l ,Eq) such that Vl =
A
@\ v/;,o
# 0.@. Then (v/;P.
V y ) are the unique open twin d u e sets for ( E l . E2) i n
the senst. that ~f(IJ,l. U , ) clre open tsuin v a h sets for ( E l , & ) , then U, \ (0, m)=
v,:-\ ( 0 % ~fofi,r.
) j = 0.1.
P R ~ O FLct . ( l / l l .C r I ) he open twin value sets for ( E l ,E L ) . It is easy t o see. [5,
Tlmi.4.1]. that V l land V I are subsets of the closurcs of every palr of twin value
set s for ( E l .E 2 ) . 111 particular, /;,C and C F . and thus also V y \ ( 0 , m)C
C', \ { 0 , x)for J = 0.1. A
.
- A
On the other hand. also (C\ (-1 lJI) @I \ (- 1 - G))art. opcn twin value set$
-
Finding alternative value sets for element sets ( E l ,E2) corresponding t o (\$I. V l ) .
This can not always be done (see Theorem 4.4). but we give examples in Lemma
5.1 and Proposition 5.4.
0 The part 8vj of d F j is important. For every C E aiv1 \ (0. - 1 , m ) tlierc is a
2-periodic continued fraction of parabolic type from ( E l ,EZ) converging to <.
This is the reason for the optiinality properties connected with d f l / ; ) .But when
do points in a*%/;, belong t o d i h ? This is a tricky question if the shape of the
corresponding element sets ( E l :E 2 ) is not known. Some necessary coridit,ions
are given in [ll].and some others follow from Propositions 5.5 and 5.6.
0 Some sort of contraction is needed t o obtain diam S,,(V,,)-+ 0. In Proposition
5.3 we prove that under mild conditions s,,(V,,) docs not fill V,, 1 . This snilcl
form of ~ o n t r a c t ~ i oisn not necessarily sufficient. though. In Proposition 5.7 wc
connect part of our sufficient conditions. C S 2 ( b )and C i 2 ( h ) , to the shape of
8%).
We start by proving two simple, but useful observations:
x)
L E M M A5.1. Let (qI. be twrn value sets for K ( a , , / l ) . T h e n (e\ ( - 1 -
A A
+
PROOF. Since s,, (V,,) C V,,-l. where s,, ( 2 ) := a , , / ( l 2 ) is a bijective mapping
of @ onto @, it follows that V,, C s;' ( V , , - ] ) which is equivalesit to 5;' ( l i ; , - 1 ) C
A
MIrI := @ \ V,, . Since s ; ' ( z ) = -1 - s,,(-1 - z ) , this proves that s,, (V,,)C V,, 1 is -
equivalent t o s,, ( - 1 - b/,,-l) 2 -1 - IV,, . Since both and lI'l colltain at least
one element by definition, it follows that also ( - 1 - Pi'[. -1 - It\/;,)arc twin value
sets for K ( a , / l ) . 0
LEMMA 5.2. Let (h, V l ) be twzn value sets for K ( a , , / l ) . If - / I , , E \Ir,, for nn
n = no E W, then -h,, E LIT7, for all n >
n m If -1 @ V l , then h , , E IT:, f o ~all
+
n E W . If h,, E 1 V,,+l for an n = no E W, then h,, E 1 +V,,+1 for d l n > If
+
0 E h.then h,, E 1 V,,+l for all n E W .
PROOF. IVe have -h7, = s z l ( - h , , - 1 ) and s;'(ll<,-l) C IT;, for all T I . Hence
-h,, E LITn for all r t >
no if this holds for n = rill. Since h
A A
= -1. we have
-h7, E Urn for all n if -1 E IV1. Since (@ \ (-1 - Vl). C \ (-1 - V ; , ) ) are value
sets for K ( n 7 11)
, by Lemma 5.1, and - 1 E - 1 - v/;, iff 0 E I/;,, also the rcmaini~ig
statements follow. 0
Our first result says that s , , ( x , ) # T/;,-, under mild conditions. It is ii tn-in
version of [20, Thm 3.31, where we have removed the condition that V should be
a n open set:
PROPOSITION 5 . 3 . Lrt (v,,
V l ) be bounded trum lialue sets f o r I i ' ( a , , / l ) sut lr
that both El and EL each contarns at least truo d e m e n t s . Then there ezzsfn an
E > 0 , dependzng only on (q/;,,
V , ) . such that for every n 2 2
-
PROOF. Since vI -
Then sl(Vl) = v)
dsitl s>(V;))= Vl It is also clear that n l # 0. since otherwise
L', = {O}, i.e.. a onc-point set. \vhlch is impossible. In fact. 0 @ I/;,. since V;) =
a1 /( 1 + Vl ) where Vl i5 hourldcd Hence 0 # E l . Similarly we find that 0 # Vl and
0 # E L . Let 2, E El with 21 f n l , and let Z1(z) := 31 /(1 z). Then +
and
Lrt a , /Zl =. r t '". Repeated applications of (5.3) then shows that r " P ~ ' "o' v - h
C
for all n . w1m-c V;, is bounded and bounded awa, from 0. Hence r = 1, and we
ma\ set 0 < 0 < 27r ~ I I I C CG I # n l . This means that every rotat loll r - "'"v/;, of is
tontalilet1 ln for tlus particular 8. Sirnllarl> n2/ZL = e'y for a p E (0.27r), and
c\er\- rotation ep"'-Vl of I; is contained in Vl for this particular p. \Ye have in
fact b> (5.3). (5.4) nrld (5.5) that
+ c
A
poirlt,s in K have the distance R t o 1 , but at least one of the points in K has
larger distance t o t h r origin than t o -1. This is a contradiction sirlcc ql C E ( 0 , R ) .
Hencc. S 2 ( v 1 )is a proper subset of for every pair ( a l .as) E ( E l ,E 2 ) . and thus
(5.1) holds for some E > 0 for each such pair ( a l ,a z ) . It remains to prove the
rxistencr of t,lie uniform E > 0. This follows by a standard compactness argument,
since El and E 2 arc. closed, bounded sets:
For every pair ( n 1 , n2) E ( E l ,E 2 ) we defirlc ~((11.as) := sup{& > 0 : S 2 ( Q l )C
.
\ B ( z , E ) for a z E K)). Assulrie that t,lie set I := { ~ ( a al ?, ) : ( a l ,as) E ( E l E z ) )
has a clustcr point at 0. T11t:n there exists a seqlience {E,,) from E converging to
0. Since ( E l .E 2 ) is co~npact.thew exists a subsequmce { n k ) of N such that
CONVERGENCE CRITERIA FOR CONTINUED FRACTIONS 227
which is impossible. since S; with a ; and aT; also maps Kl into a proper subset of
V;), even if one or both of a ; , a$ arc zero. Hcrlce E := inftEEe > 0. Siniilarl) it
follows that (5.1) holds for s2 o s 3 , and thus for all n. 0
IVe shall extend yet another result from [20]to the twin situation. It is a uscful
result which in some cases allows us to reduce the size of given value sets:
PROPOSITION 5.3. Let (Kl. V,)be twin value sets with -1 @ Vll n V 1 . such
that the corresponding element sets El and E2 each ~ o n t a i n ~ats least ttuo elements.
Then also (G \ (-1 - Vl): Vl \ (-1 - v ) ) ) a,re twin value sets for ( E l . E 2 ) .
PROOF. Since (-1 - Ilrlr -1 - W70) are twin value sets for ( E l . E L )by Lemma
5.1. it follows that Uo := q / ; , \ ( - 1 V 1 ) and UI := Vl\(-1-Kl) satisfy E 1 / ( l + U 1 ) C
+
Uo and E 2 / ( l Uo) C U1. IVe just have t o prove that Uo and U1 are non-empty.
As5umc that Uo = 0.Then also 11, = 0 since El/(1+U1 ) C;,.Hence Vl C -1 - V;,
and V;) -1 - Vi, and thus Vl = -1 - Kl. Since E l / ( l + Vl) C V;),wr then have
c
E l l ( 1 - 1 - h) K) = -1 - Vl. That is. sP1(Q/;,)C Vl for all n E E l . On the
other hand s(V1) C v/;,
by definition. IVe therefore have s(Vl) = V;, for all n E E l .
Similarly. s(V;/;,)= Vl for all n E E L . If m E To,then -1 = s - l ( c c ) E TI for
n E El \ (0). Hence -1 @ Vo However. we must have x E -1 - V O = V 1 ,w h ~ h
implies that -1 E Vo This is a coiltrddict ion. Hcrlcr K) is bounded, and so is L;.
But this contradicts the result in Proposition 5.3. Herice lJO # CI arid UI # 8.
REMARK R5.1. It is clear that thc twin value sets (Uo. C1) mith l', = C; \
(- 1 - % + I ) satisfy U j n (-1 Ul+l) = 0.as required 111 C'1 It is also t l w r that
-
d*U, = d*Vl and dl UJ = dl Vl, so this operation does not change these iinportdiit
parts of dV,. This is consistent with Theorem 4.4. which aavs that if dl V, = dV,,
then dV, cannot be changed at all. at least not under the coiiditions o f t hat tlieorern.
The next result shows that if wc want t o study the situation wherc 8 V J # 0.
then we ma) work with (I(), Vl) with non-empty interiors. or, in view of Remarks
R2.1.2 and R'2.1.3, with (V;,. Vl) which are unions of Jordan domains. without
loosing much generality. For convenience we introduce the class Cc of curves for a
given C E C\{0. -1). It consists of the line L C : 2 = C + C ( l + C ) t for -x< t K . <
and all circles tangent t o L < at C. That is. C, is the class of generalized circles which
are invariant under mapping by the parabolic linear fractional transformation
PROPOSITION 5.5. Let (V;,,V1) be t w m value sets with { E df V;, \ (0. -1. x).
where conwists of (1 finite union of simple. recfifiable arcs T I : j = 1 . 2 . . . . .I\-.
Let ( E l , E 2 ) be the corresponding element sets. Then. uizthouf loss of generalltg
(cutting r, i n two parts rf necessary), each r , is part of a C, E Ci. I f there emsts
.
a j E { I . . . . N } such that C",, is not a circle uiith center at th,e origin or n lzne
through the origin, then El \ (0) = {-C2}.
<
Straightforward computat,ion shows that 0 E Li only if E R. Similarly. Ci
contains a circle with ccnter at the origin only if R e < = -1. It is also easy to
see that Ci contains a t most onc such circle, and that if C', is slich a circle. the11
el := -(1 + C)2/(1 + C J )is not a circle with center at 0. Hrnce. Propositioii 5.5
shows that at, least one of the two elemtilt sets is a one-point sct , disregarding the
228 LISA LORENTZEN
point 0. If both V;, and Vl contain arcs from circles with centers # 0 (or from LC
<
when @ R).t h r n E l = +
and E L = {- ( 1 c ) ' ) . again disregarding ( 0 ) . and
E L )are (Vl.V l ) = ({C). { - I - 0 ) .
the bcst limit sets for (El,
PROOF. The esential thing here is that < E dTV;l, and thus-<' E E l and
v)
-(I+<)' E E L . So in particular we have - ~ ~ / ( l + VCl ) and - ( l + < ) L / ( l + CI v ) )
V , . Tf7c shall see that this implies the shape of b;,. Let r be a simple arc in 4 1 . If
where
-
C @ T . then SLr,,(r)C rl)for rn large enough, for some arc roin V;/;,
S L n L ( 2:=) T 0 T 0 . .. 0 ~ ( 2 ) .
PROOF. Let A 1 > 0 be chosen so large that E;; := E , n B(O. A l ) ; j = 1.2. each
contains a t least t,wo points. Let (GI.
GI) be bounded value sets for (El.&); t,hat
A
Assume that d v 1 has a cusp or a corner a t an inner point of die, with <
angle a E [0, 27r]: a # 7r, where t,lie angle is measured ill thc interior of q/;,.Since
A A
Our last result in this section gives some simpler sufficient conditions for Cow
dition C;, 2(b) or Condition C12(b). Thcse conditions will make the chrcking part
easier, and the choosing part simpler. Condition Ci2(b) requires that
,-
CT2(b):
21P, I < l / q and 2dJ - 1111
1
2 y for ,,
all u; E T I := i3Vl n B(CJ.q )
where E d'Vl n B(O. Al). OJ := - i r ; ( d ~ , .tr~,),7, := 7(3V1.u.,),
which of course makes sense only if the curvature k-(dVl,111,) exist5 for all ut1 E r,,
Here q and x are given positlve numbers A5 a cornparison. Coildition CT12(h)
requires that the circular disk D< 2 w,
with radius q > 0. tangcnt to dll', at (.
satisfies
cos[arg{(Pi - ( ) / < ( I + < ) ) I 2 [ f o r all <
E r,,, := i3V, n B ( < l , p ) .
where <j E dtV, n B(0,AI) and P<is the center of D,.
Here p and [ are some fixed positive numbers. Clearly. such disks exist if the onr-
sided curvatures ~ ' ( 8 4 ,C) exist and are absolutely bounded by 1/(2q) for every
< E I?, ,. (They may exist even if r;+ ( 8 5 , <) and/or K - ( ~ V ,C), does not exist .)
With this notation we have:
PROPOSITION
5.7. Let r be a bounded arc i n @, let d > 0 and
and let a ( < ) := arg[<(l + C)] for all < E ro T h e n the following statements are
equivalent:
A l . There exists a c > 0 such that s i n [ a ( < )- y(T,()I > c for all { E r o .
A2. There elcasts a [ > 0 such that cos > [ for all C E To ulherr
PCw the center of the desk Dc 7n CY12(b) nboue.
<
If In addztlon r;'(r, () r m t s for 011 E rO.thrn A1 and AL (LW rqu~ualentto:
230 LISA LORENTZEN
This can only hold if the tangent of dVl at CJ has angle of inclination strictly
less than t h r ray < , ( I + (,)t. uniformly with respect t o the points C, E 8 ' 4 .
This follows from t h (quivale11~r
~ in Proposition 5.7. since
CONVERGENCE CRITERIA FOR CONTINUED FRACTIONS 231
If aTv1
= dV;] and rx, E a v , , as in Corollary 2.14 or in C6rdova's Convergence
Theorem, then it follows froin the inequalities above that ~ ( d h<), + 0 as
C + x along dQ1. That is. d v l is almost a straight line when it approches m.
3. Coriditioii Ct2(b) corresponds t o C6rdova's Condition Y3 combined with his
continuity condition for r;(dV. 5). In Thron's Uniform Parabola Theorem it
corresponds t o a # *7r/2. In the latter example it is easy t o see that for
cu = &7r/2. the corresponding element set reduces t o the negative r c d axis
which definitely is not a convergence set of any kind.
4. The inequalities in Remark R5.2.2 above have a n interpretation based on the
angles
and
of their complement s, 117, := C\V,for J = 0, 1 each has one and only one unbounded
cornponerit. This fact allows 11s to assume that they actually are simply connected:
LEMMA6.1. Let ( \ / ; IVl) , be open and connected t w i n value sets for h7(a,,/1)
such tttat \I/;h,as o n e and ondy o n e unbounded component U , for. j = 0 . 1 . T h e n
A - A -
the simply conn,ected d o m a i n s (C\ U o , @ \ U1) are also twin value sets for K(tr ,,1 1 ) .
A A
PROOF.Since a ~ , , + , / ( 1 + 4)
C KIel, it follows that V , # According to c.
Riernann's hlapping Theorem. there exist anal) tic, univalent functions pJ in V, such
that p , ( V , ) = U := { z E C : 12 < 1 ) for J = 0 , l . Then fL,,-l := 900 srll-l 0 p,l
and fLrl := p1 o S J ~ o, prl
map U mto K := p , , ( K ( , )U p l ( K l ) , and
It follows then by careful use of Schwarz' Lemma. as done in [14],that FLr,(U) con-
verges to a orleepoint set. uniformly with respect t o { f , , ) . Actually. an expression
for a uniform bound (2,)
depending only on K . such that diam F,, ( U ) d?, for all <
n is found in [14]. Since S L r l= p t l o FLrlo p ~ where
. 90is fixed. this means that
diam S2,,(q,)is bounded by some uniform dl,, + 0 . By continuity we thus have
<
diam S2,,(VO) d2,,. The nestedness (2.1) implies that also diam SLIl+l(V1) d ~ , , . <
0
Our third lemma shows that even { S , ,( 0 ) ) converges under our conditions. It
is related to Theorem 1.2:
L E M M A6 . 3 . Let (q,,V l ) be hounded twin ? d u e sets for K ( a , , / l ) , where q/;,
contains at least two elements, and assurne that there exist two sequences {u,,) and
{v,) such that
(6.2) n -
i n S ,(
x
7 , ) =
-
i nx S
n
( 7 , ) = f u,, , u,, E V,, , lim inf Iu,, - t i , , I > 0.
Then { h ~ ~ has
~ all+ zts ~km7t
) pomts
~ ~ en~1 + V , + , ; J = 0 . 1 , and lim,,,, S,,(O) =
f.
We say that K ( a l ,11) converges generally t o f when lirn S,, ( u , , ) = lim S,, (I,,,) =
f for two sequences { u , , ) . {u,) with liminf d, (u,,, c',,) > 0 . This concept was
introduced in [4],where it also was proved that if this holds. then
PROOF.Since L$ and thus also Vl contain a t least two elements. we can choose a
g E -1-Vl such that g # f . Since s l l ( - 1 - V , , ) C -l-V,,+l for all n. it follows that
5';' ( g ) E - 1 - Vn+lfor all n. Moreover, by (6.4), d,(S;' ( g ) , S;' ( m ) )+ 0 since
f E and thus f # x . Hence { h z n + , ) ~ = =has, all its limit points in 1 for + F+l
j = 0 , l . Moreover, w,, := x can be used in (6.3). and thus S , , ( x ) = S n e l ( 0 ) -+ f .
0
To prove part B. we first observe that also (V;,,V l ) are twin value sets for
(El,E2) if K j C VJ
C for J = 0.1. Clearly we may choose q/;, and Vl to be
A A
Hence / Q 2 , L5
) (2M + 1 )(hi+ l)d2,, /Aill. On the other hand
and so
where 0
-
< d := d ( - 1 , & ) < Ih2,, /
A A
PROOFOF L E ~ I N 7.1. v)
A Evidently both and Vl contain at least two elements
in our 5ituation. Assume that {Sir, (zO)}has a limit point cr # 0 for some z(, E
V;) \ d * v ) . Let the subsequence { S ~ , ,(zo))
, converge t o a. {S,,(2)) is a sequence on
A
the compact Rieinann sphere for every z E @. Without loss of generality (taklng
subscquencrs) we ma! therefore assume that also SillI (z,)), S L ,(, ~ z ) ~dnd SL,~,
(ZL)
Converge. where zo, zl and 2 2 are distinct, and also zl E &. and that hL,, converges
to a limit h . (Infinite limits are acceptable.) According t o Theorem A we then have
the three possibilities:
A
A. Then S2r11
values ill C.Herice this case is ruled out also.
(c) { S L , , ~convergcs
} t o a constant function in a set D containing zo, z1 and
22. Assume first that h L r I h= x for infinitely many indiccs. Then we may without
loss of generality assume that all hLrli = X . So Sir,,( z ) = QLTa,+ a. whereas
SlrIr ( z l ) - SjrIi(zo) = d)L7,i( z l - q , ) 4 0, which of course is impossible. Hence
CONVERGENCE CRITERIA FOR CONTINUED FRACTIONS 235
# cc
h2,1A= cc for a t most a finite riurnber of indices. and we can assume that hL71A
for all k . Assume first that h # -20. Then
+
This is a contradiction. Hence h = -20. However, Q L r t h= C?(zO h L I I)'A. and thus
-11 E d*I/;/;,by the arguments in case (b). Since zo 6 d *I/;), this case is ruled out.
Herice {Si7,(zo))does not have a limit point cr # 0. In other words. Si,,(zo) -t 0.
Similarly we find that SL7,+l( z ~-), 0 for zl E Vl \ d*Vl.
The second lernma in this section is connectcd t o Lemma 6.3. but its conditions
and conclusion are weaker.
LE~~M A Let (K, Vl) be bounded twm value sets for K ( a , , / l ) , and n,sumr
7.2.
that the correspondmg element sets El and EL each contazn, a t lrnst two demrnts
+
Then {h2,,+,) has (dl zts lzmzt poznts zn 1 V,+I. J = 0 . 1
PROOF.By Proposition 5.4 we may without loss of gcnerality assume that
Kn(-l -F)
= a*vl,since V, := V,\(-1-V7+1) for J = 0 , 1 have t h ~ property
s arid
d*? = d*V,. Since El and ELeach contains at least two eleinents, it follows that so
do also 1/;, and Vl. Actually. both I/;, and Vl must contain infinitely nianv elements.
It follows froin Lemma 7.1 that lirn,,,, Sill+7 (zJ) = 0 for all z , E V, \ d*V,. This
means that if we use the represeritatiori (7.2) for S,,, then Q,, + 0 arid {A,,} has
all its limit points in I/;,since S,,(V,,) C I/;) for all n .
Let h be a limit point for { h ~ , ~and
) , let I L ~ , ,+
, h . \I'ithout loss of generalitv
(taking subsequences) we may assume that ALnA+ A. Lct r ~ , , , E -1 - V,, ul0 # .4
'
be chosen. Then Sly1( w ~ E) - 1 - V,,+l for all n since s; (- 1 V,,) C - 1 - V,,+1
-
This means that if h ~ , , ,# cx, for infinitely many indices k , then (SG~'~
( u i ~ ) ) + h L)r l A
-)
the fact that SL,,' (IDO)is contained in the bounded set -1 - V,. Hence this cannot
+
occur. Hence h E 1 K. Similarly we can prove that { h r l l + l ) has all its limit
+
points in 1 G. 0
REMARKR 7 . l . It follows from Lcinrna 7.2 that if (l4,V,) are bounded twin
value sets for K ( a , , / l ) . arid the corresponding elenlent sets contain at least two
elements each. thcn {-hZr,+,) has no limit points in y \ d * V,. So, if d(-h,, , dV,,) 4
0, then d(-h,,. a*V,,) 0.-)
236 LISA LORENTZEN
that \ d*U1 # 0.
Let first { - h L r , ) have a limit point -h # d*UO = a * h . Then -h # d f i by
Lernma 7.2. Actually, - h # K. Let further {-h2,,) converge t o - h , and let
u E & \ ~ * U OThen . SL,, ( u ) + 0 by Lemma 7.1. Since
where lu - ul <
diam l/;) < x,it follows that diam S2,,, (K) + 0. and thus that
diarn S,, (K)4 0 by the nestedness of the ST,(V,, ).
This result follows similarly if { - h L , , + I )has a limit point # d*Ul = d* V l .
That also {S,,( 0 ) ) converges t o the one and only point in lim S,, (K)follows
now by Lemma 6.3. 0
1. Lemma 8.1 does not imply that the continued fraction in question is h i t 2-
periodic. It is actually an open question whether there exist other continued
fractions K(a,,/l) than the limit 2-periodic ones which have d(-h,,, d*V,,)i. 0
under our conditions. It is clear that the 2-periodic continued fractions are the
only ones in Lemma 8.1 if dTI/;/;,is a finite set.
3. The particular case where K(a,/l) is limit %periodic is treated in [19]. There
it is shown that every such continued fraction from ( E l .E s ) converges under
mild conditions.
4. Condition C*1 holds if e*1 is satisfied. This follows by the following argumer~ts:
Let e*1hold, and let s,+l o s,+l(d*VJ)n d*V, # (2). Then there exist (,E a*%,
&+I E V J + ~ E a*%w ~ that
and < J + L h (,+I = S ~ + L ( ( , + L ) and (,= s , + ~ ( < , + ~ ) .
Since -1 - <,+I = -1 - s y +1l ( < J ) = s,+i(-1 - C,) E s,+l(V,+i) C V,, it
A
follows that <,+I E a*V,+l. It follows now from C * l that <,+h-l is the o n l ~
point in KA:= s,+~(d*V,+k) nd*V,+~-,;k = 1.2. Since = s,+h(<,+h)
if and only if -1 - (,+A = s l +(-1 ~ - <,+A-1). we therefore must have that
i J + n - 1 = -1 - <,+A, and thus also a,+h = -C1/Lih_,
- -(I + <I+#.
-
In the next lemma we use the notation (7.2) for S,,. This lemma. combined with
Lemma 5.2, shows that diam S,,(K)4 0 if and only if G,, := Q,,/d(-h,, . dV,,) + 0
under mild conditions:
LEMMA8.2. Let (I/;), Vl) be twsn - value sets for K(a,,/l) such that I/;) contams
at least two elements, and let -h,,,, E II.',,,, for a n no E N. T h r n the followzng holds
for n = no:
A . diani S,, (K)= IQ,, ( . diarn (K)sf h,, = K .
B. diam Sr,(V,)< 2lQn Ild(-hr,%8x1) zf h,# x.
C. If cx, E K. then diam s,, (K)> /Q,,I/d(-h,, , dV,,) ~f h,, # x
D. If x @ and Ih,, I 5 Al, then there mists a constant k(l > 0 depending only
on Al and V,, such that diam S,, (F) .
2 k.0 lQ,,l/d(-h ,, dV,,).
E. If d(-h,,, , dV,,,) + 0, then there exist an n 1 2 no and a constar~tko > 0
such that diarn S,,, (I/,,,) > ko IQ,,,l/d(-h,, . dV,,,) for every rn > nl for which
h711 #
PROOF.Also the set Vl has at least two elements since a2,,/(l + v/;,)C V,.
where a ~ , #
, 0. hloreover. -h,, E for all n >
no since s;'(li~,- ) C L17,, arid
. n := no.
-hrl = ~ ; ~ ( - h , l ) Let
A. This follows directly from (7.2).
B. Let h,, # x. Whether m E or not, we must have
238 LISA LORENTZEN
h2,
- WI,
+ i.,.I' diarn + d(-h2,. aKi)
+ -
1
2
as u i x , ( f ~ ~ ~ , # x ) .
Hence the result follows for eve11 tn from the first inequality in (8.1) for any kll < f.
If x @ the result follows similarly for odd nl. b
n ~ ~ , , - and
~ /~~
G2~ ,,
, -,I / G ~ . where
, + ~ G., := ~QII 1
d(-hl,. a v n )
If one (or both) of these products diverges t o m , then GL,,4 0 and/or G>,,+l 4 0.
arid so (V). V,) are uniform twin value sets for K ( n l , / l ) by Lemma 8.2 arid the
nestedness (2.1). If both the products converge, it turns out that wc still get
convergence of K ( n , , / l ) . To help us see this. we shall use the eslimate in the
followirig lemma, which of course is empty if d t K l = 0:
l ~ Let (I/;), V , ) be twin value sets for K ( n r , / l ) with -1 @
L ~ n f n 8.3.
such that C*l and C"2 are sc~tisfied, and assume that d(-h,,, ~ I v , , +
U
) 0 and
r.
linlsup Ih,, 1 < x.Then there exist n X > 0 and u n n o E N such that
(8.2) R,, :=
G,~-L
-- >
1 + Xd(-h,,-y,dV,,) for n > no. where GA :=
lQk l
GTi d ( - h ~. ski ) '
PROOF.We shall first prove (8.2) for the case where n is even. Let r,, :=
s2rl-l o sllr for all n. That is.
and thus
CONVERGENCE CRITERIA FOR CONTINUED FRACTIONS 239
Since limsup ( h n (< ca,we also have that lirnsup )a,,) = lirnsup J(h,,-1)h,,-1) < m.
Since all a,, # m, this means that { a , ) is bounded. Hence. by Proposition 2.2 we
may without loss of generality assume that Ki and V l are bounded, say I/;) U V l C
B ( 0 , M ) for some M > 0.
Let p, q and E be positive constants such that C,t12holds with this M > 0, and
let ro I, := d&) n B(C(1.p) and rl, := dV1 n B ( - 1 - 6 , p ) for an arbitrarily chosen
Co E d i V j . We shall first see that for no E N large enough and p > 0 small enough,
we have:
( 1 ) O @ r o , u ~ l , a n d-1 @ ~opurlp.
( 2 ) h,, E -Wn for all n 1. >
(3) {f~,),"=~,,is bounded and bounded away from 0 and 1.
(4) {a,)~=,,(, is bounded and bounded away from 0.
( 5 ) {a,, + h,,-I),"=,,, is bounded and bounded away from 0.
Evidently, d t ~ ;and
, d t V l are bounded and closed sets in e.
Since a*Vl = - 1 - 3"Vo
and -1 @ d*Vo U d* Vl , we also have 0 @ 8*VnU a*Vl. Hence both -1 and 0 have
( 2 ) follows by Lemma 5.2. (3) is a consequence of the fact that d(-h,,, dtV,,) 0.
( 4 ) follows since Lemma 8.1 implies that (a,, + hk ) -,
-
positive distances to dTQjU d T V l. Since d t V, is closed and bounded, this proves (1).
-+
0. Finally (5) follows from
+
(3) and Lemma 8.1, since a,, h,-1 = h,-l(l - h,,-1) +
(a,, f ~ : ~ ~ ) .
+
Let no be so large and p > 0 be so small that ( 1 ) - ( 5 ) hold, and let 2n > no 2.
Straightforward computation shows that
Since by ( 8 . 4 ) ,
where
240 LISA LORENTZEN
where by (8.4)
+
Since { n , , ) and {a2,,_1 h2n-2)~=rl,,i2
are bounded and bounded away from 0 (see
+
Properties (4) and (5) above), this proves that ( h s r L P 2 z,,) --t 0. From this it
follows that if no is chosen large enough, then z,, has the following properties for
all n > no/2:
(6) (~271-1 - 2 7 1 171=no/2is bounded.
( 7 ) z,, E To ,. (See Condition C ; [ 2 ( a ) . )
( 8 ) z,, is uniquely determined.
(9) { ~ , ; ~ ( z , , ) ) is
~ bounded
= ~ ~ ~ , and
~ ~ bounded away from 0 and -1.
It remains to prove (8.8).
Let no be so large that these properties (6)-(9) hold. Let further no be so large
and q > 0 in C,:,2 be chosen so small that lasn-l - P,,, I >
2q for all n n0/2.>
This is always possible since by Lemma 8.1, (al,,-l - z,) (-h&-2 +
h ~ , ~ - l=)
h 2 1 f - L ( 1- h 2 r l - ~which
) is bounded away from 0.(See Property (3) above.)
Since T o , is smooth. we know that the vector from z,, to P,,, is normal to d f i at
z,, . Sincr 7;' is a conformal mapping, this means that the vector from 7;' (z,) to
PI, is normal to r;'(aK) at ~ ; ' ( z , ) , where F,, is the center of the disk r ; ' ( D Z n ) .
A
Let us write a,, := - c i for all n , where we choose c, to be the square root of -a,
'
which is closest to h , , 1 . Since 7,; is a conformal mapping, we find that
CONVERGENCE CRITERIA FOR CONTINUED FRACTIONS 241
>
for n n,/2. This proves (8.8) with no := nE and ,u = [/2 for even n. The case
where n is odd follows similarly. 0
242 LISA LORENTZEN
PROOF OF THEOREM 2.7. We first observe that since Rl and Vl are bounded,
a*
we must have - 1 # V oU V 1 . This means in particular that - 1 # Kl U d* V L, and
thus also that 0 # d*l.;, U d * V l . Let K ( a , , / l ) be a continued fraction from ( E l , E2).
Assume first that a*% = 8 or that d(-h,, d * V n )does not vanish as n -x. Then
the convergence of K ( a , , / l ) follows from Theorem 2.4.
Assume next that a*M/;I# 0 and that d(-h,,, a"V,,) 4 0. Then limsup Ih,/ <
x,and thus d(-h,, , dl L',) 4 0 by Condition C*1 and Lemma 8.1 with Remark
R8.1.2. hioreover -h,, E WR for all n > 1 since -1 E W,". (See Lemma 5.2.) It
follows therefor? from Lemma 8.2B and D that there exist constants ko > 0 and
no E N such that
2IQ7,1 for all n > no.
(8.14) k"'Qn' <dialnsn(K)5
d(-hill W , )- d(-h,,, aV,,)
Hence d,, := diarn s,, (K) 4 0 if and only if G,, := IQ, ]Id(-h,, , d V n ) 4 0. It
follows from Lernma 8.3 that Rr,, := G L n - ~ / G 2 , isL bounded from below by 1 f
Xd(-hLn-L, 3l.4) for all n >n o / 2 for some X > 0 and no E W. This means that
1 , , = X , and thus G 2 , 4 0 , then diam s,,(E)
if ~ ~ = 1 , 1RL,, 4 0 by (8.14) and the
nestedness (2.1) of s,,(%). It remains to consider the case where nz=nI,,2 RL,,<
oo.
Similarly, ~f n ~ = , l l ,RL,,+1 = oo. then G27,+I + 0. and thus, again,
diam S , (V,)--t 0. Assume therefore that both n,"=,lolL Rln and nF==noll R2n+l
converge to finite values. Then C:=,,=,o d(-h,, dV,,) converges to a finite value. It
follows from Lemma 8.3 and formula (8.14) that 6, := diam S,(K)4 6 > 0. In
fact. since by (8.14) koG,, 5 b,, I 2G,,. we have JQl,J= p,, . d(-h,,, dV,). where
& , / 2 5 p,, 5 6 , l h . Hence C~=r,l,12 l Q 7, 1 < oo.
However, Q,, = h,, (S,, (0)- S,,_ ( 0 ) ) .where {-h21,+J) has all its limit points in
8 4. Hence, {h,,} has no hmit point at 0 , and so x:=,, IS, /
( 0 )-ST,-1 ( 0 ) converges
to a finite value. But this means that C:=,, (s,,(o) -S,- ( 0 ) ) converges, and thus,
so does {S,,( 0 ) ) . 0
Let P k denote the center of BI, and z,,, E dV,,, be such that d(-h,,, ,dV,,,) =
+
1 h,,, z,,, 1. Since
it follows that diam ST,,+,(V,,,) < diarn S,,, (Bk). where S,,, (Bk) is a circular disk
with radius
hfl
5 diam S,,, (K). d(-hrLh , aVnA) + 2hI1'
since x/(2x + A) increases with x when A > 0. This proves the lemma. 0
LEMMA9.2. Let (V), VL) be t w m value sets uizth 0 6 dl(/;,U dV1. and let rn E N
and h 1 > 0. T h e n there erzsts a d,,, > 0 such that d(S,,,(O). dT/;/;,)> d,,, for every
contznued fractzon K(a,,/l) wzth all la,, I 5 A 1 from the correspondzng elements sets
( E l , E2).
A A
PROOF OF THEOREM 2.10. AS in the proof of Lemma 9.1, we find that h,, # x
for all n EW. and thus that S, can be represented by (9.2). This also means that
S, ( m ) # m for all n. Let m := sup rnk. It follows from L e n m a 9.1 that if
C d(-h,, , dVnh) = m , then diam S,,, (V,,,)4 0. and thus diam S,,(V,,)4 0 by
the nestedness of S,,(V,,). In particular S, ( x ) = S,,_,(0) converges.
Assume that d(-h,,, , dVn, ) < m . F'rorn Lemma 8.2B and C it follows
that (8.14) holds with ko := 1. Hencr. if {Q,,, Id(-h,,, , dV,,, )) has a subsrqnence
comerging t o 0, then K(a,, 11) again converges by the arguments above. Assume
that no such subsequence exists. that is, /Q,,, 1 = ph d(-h,,, , W,,,) for some pk > 0.
244 LISA LORENTZEN
where { p k ) is bounded and bounded away from 0 . Then C IQ,, / < m. NO^, for
j>l
- Qn, Qn,
- -
~ l " ~ )+
(mh,, ) s~Y?(o)
+ h,, '
where s F h ) ( z ):= z . Clearly, dm in Lemma 9.2 may be taken to be non-increasing
with m. Using Lemma 9.2, this shows that
IQn, I l IQn, I
(9.5) ISn,+,(m) - sn, (m)l= s~?;'(o)
+ h,, I I d
IQnh
( ( 0 )V )
5 --
dm+~
for 0 < j < m for some dm+1 > 0 depending only on m, M and (Vo,V l ) . Let first
j := m k + l <nz. Then (9.5) implies that C /S,,+,(m) - S,,(m)l < m, and thus
{S,, ( m ) ) converges t o a finite value f .
Choosing j := jk < m k + l in ( 9 . 5 ) , shows that also {S,,+,,(m)) converges to
f . Hence S , ( m )= S,- ( 0 ) converges to f . 17
<
where y, ,O E R with 21PI 5 1 / q , and lg(t)l 5 w ( t ) t 2 , lg(t)l w(t)ltl and R e g ( t ) 2
-1 +6 for all t E I . T h e n there exists a continuous, real-valued function p ( z ) for
z i n ( C \ V )n B ( z O , q / 2 ) such that the following hold:
A. d ( z , dV) = Im { ( z - z o ) e p " } + p ( z ) . [Re { ( z - zo)e-"}]2 for all z E (@ \ V )n
B ( % !q / 2 ) .
B. To every E > 0 there exists an r > 0 , depending only o n E, 6 , q and w , such
that + PI < E for all z E (@ \ V )n B ( z 0 , r ) .
write
Let z E ( C \
PROOF. V )n B ( z o , z) be arbitrarily chosen. For simplicity we
(10.2) X := Re { ( z - z o ) e p q ) , Y := Im { ( z - z ( l ) e - i 7 ) , z := d ( z ,d V ) .
CONVERGENCE CRITERIA FOR CONTINUED FRACTIONS 245
A. We define
p ( z ) :=
{ (x) X 2 if X # 0,
r?T//'
if X = 0.
(Note that 1 - 2 Y P > under our conditions, since IY/ lz - zol < q / 2 and
21Pl 5 119.) This proves the expression for x = d ( z , d V ) since ) z - zo/ < q / 2
implies that x = lz - z01 = Y if X = 0. To see that p ( z ) is continuous, we only
have to check that
lim
x-Y -
----- -
-P
z-Z, X2 1 -2YP'
x 2 = d ( z ,d ~ = min
) lz~ - z ( t ) I 2
1
(10.4) = min / X
t
+iY - t - i p t 2 - g(t)12
where u := Reg and v := Img. This minimum is attained for t = t l . That is.
+
2 ( X - t l - u ( t l ) ) ( - 1 - u ( t 1 ) ) 2 ( Y - Pt: - v ( t l ) ) ( - 2 P t l - v ( t 1 ) ) = 0 , where
< <
J u ( t , ) l d ( t l ) t ? , lu(tl)l w ( t l ) l t l I , Iv(t1)l 5 w(t1)t.fand b ( t l )I i w ( t l ) l t l I. This
means that
where the @term is bounded by some function w1 ( X ) . where UJI E R only depends
on 6, q and w. Inserted into (10.3) this gives
Let q. x and ~ l ( tbe ) the quantities for which Condition C t 2 holds for our h l > 0.
Without loss of generality we assume that q > 0 is chosen such that there exists
a d > 0 such that d(O,I', ,)> d for all <, E d t ~n
, B ( 0 , M); J = 0 , l . Let further
a = -cL E Ej+l with c E VJ n B(<, , p), where p > 0 is some (small) number t o be
chosen. We require that p < 912. Then the condition that r,, is a simple arc in
C i 2 ( a ) implies that d(c, dV,) = Ic - w, 1 for a w, E T, ,. That is, we may write
(10.7) c = w, iRez7l where y,
- = y(dV,, w,) and R := d(c,dV,) > 0.
For simplicity we also write
We also observe that if q > 0 and r > 0 are chosen small enough, then c # 0 and
<, + + < +
1 ~ 21 d/2. Now lz - W , 1 5 iz - I I<, - ci ic - w, I r 2p < q/2 when p and
r are chosen small enough. That is, z E H73" n B(w,, q/2), and since C t 2 ( a ) holds,
it is a consequence of Lemma 10.1 (with V := V, and zo := w,) that
(10.9) x := d(z, dV,) = Y + (-P, + v ) X 2 and thus Y = x + (PJ - v)X2,
where 0, := - ir;(dVJ, w,) and v E IR can be made arbitrarily close t o 0. inde-
<,
pendently of E ,
n B(0, ill) and w, E I', , by having z close enough t o w,;
that is, by choosing p and r small enough. Moreover. s p l ( z ) E s p l ( - 1 - F + l ) &
-1 - T m,+L
C and
CONVERGENCE CRITERIA FOR CONTINUED FRACTIONS 247
where / z - wjI < + r 2p. When p > 0 and r > 0 are small enough, we thus have
1 s - l ( z ) + l + w , / < 912. B y using Lemma 10.1 with V := - 1 - W,", zo := -1-u!,,
and the representation z = -1 - q ( - t , w,) for a ( - 1 - W;) we therefore get
( w , - iRez?j)"
(s-l(z) + 1 + w , ) e p Z ? =~ +
w, ( X + i Y ) e " ~
+ w,) e - ' . ~
=X +i +
{ ~2 R + ( ( Y+ R)' - x 2 ) q 2- 2 X ( Y +R)ql}
e2Y~ ~ ' Y J
vi :=Re---, qz :=Im--- and T : = I X + i Y I + R I r + 3 p ,
W3 W~
<
where 10(?'))G F v o r some %? > 0 with := /XI zT^ + +
R 5 r 4p. From +
Condition C t 2 ( b ) we know that 2p, - 11.2 >
x > 0. Let r > 0 and p > 0 be
chosen so small that we also have Iul 5 x / 8 , la1 I x / 8 and 5 x / 4 when%?T^
248 LISA LORENTZEN
and thus d(-h,, d t V n ) + 0 and d(a,,, - ( d i R , - l ) 2 ) -t 0. Let X > 0 , 7- > 0 and p > 0
be some constants for which (10.5) holds. Let further c,, be the square root of -a,
closest to -h,-l, and no E N be so large that d(c,,, dtV,,-l) < min{p,p, 7-12} and
> >
Ic, +hn- 1 / < 7-12for all n no. Then c, E V,,-1 for n no with d(c,, dt V , ,1 ) < p.
Let <,-l E dtVn-l besuch that d ( c , , d t ~ , - ~ )= /c,-<,-11. Then Ih,-l+<,-lI <
r ; i.e., -h,-1 E (-1 - V,) n B(<,-,, r ) . It follows therefore from (10.5) with
z := -h,-l that
1. L l(a)and ro(a)are parts of the two circles with radius and centers at -1
i.
and 0 respectively. They are connected at z = - The ray L, (a)is tangent to
the circle with center at j , and it takes off where the arc I?, ends, a t an angle
cw with the imaginary axis. Clearly dV(cw) = d * V ( a ) = d i V ( a ) is a Jordan
curve in e.
It has continuous curvature, except at the points where the parts
of dV(cu) are "glued together", where dV(cu) only has one-sided curvature. A
picture of the set C can be found in [ll].
2. For a = 0 the Worpitzky Strip coincides with the strip Co in Thron's Uniform
Parabola Theorem with a = 0. (See (1.15).)
3. We have 0 E V, -1 @ V ,oo E dV, V = @ \ (-1 - V ) and dV = -1 - dV for
V := V(cu). Hence Theorem 11.1 is consistent with Theorem 3.3.
Lange's second class of strip regions is his class of Trancendental Strips. Again
we have changed Lange's notation t o match ours.
THEOREM11.2. (The Trancendental Strips [ll,Thm. 4.1, p. 2291). Let
-12<
- d 5 4
be a given constant, and let Vd be the Jordan domain in @ with 0 E Vd
and boundary
1 2d
(11.2) dVd : z = zd(t) := -- -
2
-
7r
tan-'t + it for - x < t < ca.
Then Vd is a value set for continued fractions K ( a , / l ) with corresponding element
set E = -C2, where C := % n (-vd) \ {m).
REMARKS
R11.2.
1. Clearly, z E dVd with d # 0 iff z = -;+ +
z iy with y = - t a n ( ~ z l 2 d ) .It is
also clear that K d is the complex conjugate of Vd. For d = 0 the Trancendental
Strip coincides with the strip Co in Thron's Uniform Parabola Theorem. As a
consequence of Theorem 4.4 we find that Vd is the unique open value set for E
in Theorem 11.2.
2. In his paper [26] Thron defined conditional convergence regions E (or C ) for
continued fractions K(a, 11) to be regions such that every continued fraction
K ( a , / l ) from E has the property that K(a,,/l) converges if and only if
Strips do not have continous curvature at the three points where the arcs are "glued
together". Moreover, C6rdova's conditions Y 2 and Y 3 do not hold either for this
value set. It seems reasonable to expect that Corollary 2.14 can be made valid for
cases where we only have one-sided curvature; that is, for cases where Condition
C t 2 is slightly weakened. Since V ( a )satisfies the weaker C6rdova Condition C * l ,
this would have made it possible to apply Corollary 2.14 to the Worpitzky Strips.
However, at the present point, we rather apply Corollary 2.9:
THEOREM <
11.3. Let cu E R with 0 a < 7r/3, and let V = V ( a )and E be as
i n Theorem 11 . I . T h e n the following holds.
A. E is a best convergence set locally i n the sense that E is a convergence set
locally, but E U { a ) is not a convergence set locally for any a @ E U {m).
B. V is the unique open value set for E , and V is the best limit set corresponding
t o E . Actually, every w E V \ ( 0 , m) is the value of a convergent continued
fraction K ( a , / l ) with a1 E dE1 and azlL = - ( 1 +
0 2 ,a2",+1 = -(2 for
n > 1 , where < E d V .
C. Parts A and B also hold i f we replace E and V by their complex conjugates.
THEOREM
11.4. Let a , r E R be given constants with 0 <
a < 7r/2 and 0 <
r < 1 such that
max{r, 1 - r ) < cosa, and let & ( a , I-)be the Jordan domain with
252 LISA LORENTZEN
convex. Hence it follows from Proposition 3.4 that the C6rdova Condition Y2 is
sat,isfied. Y1 holds trivially. It remains to prove that Y3 is satisfied.
According to Remark R5.2.2 it suffices to prove that
where zd is given by (11.2). Since 1 + zd(t) = -zd(-t), it suffices to prove the first
inequality in (11.7). We have
Lange's Trancendental Strips can also be extended to twin strips. The sim-
plest extension is possibly the following, where we have just introduced a second
parameter r:
254 LISA LORENTZEN
THEOREM <
11.6. Let d , r E R be given constants with Id1 min{r, 1 - r ) , and
let Vd.r be the Jordan domain with 0 E Vd,r and boundary given by
2d
(11.9) z = ~ d . ~ (:=
t )- r - -
ll
tan-' t + it; t E R.
References
A. Y. Cbrdova, A convergence theorem for continued fractions, Thesis. Bayerischen Julius -
I
CONVERGENCE CRITERIA F O R CONTINUED FRACTIONS 255
12. L. J . Lange and W . J. Thron. A two parameter family of best twzn convergence regions for
continued fractions, Math. Zeitschr. 73 (196O), 295-311.
13. W . Leighton and W . J. Thron, O n value regions of contznued fractzons, Bull. Amer. Math.
Soc. 48 (1942), 917 920.
14. W . Leighton and W . J . Thron, Continued fractions wzth complex elements, Duke Math. J . 9
(1942), 763 -772.
15. L. Lorentzen, Compositions of contractions, J . Comput. Appl. Math. 32 (1990); 169 178.
16. L. Lorentzen, Bestness of the parabola theorem for contznued fractions, J . Coniput. Appl.
Math. 40 (1992), 297-304.
17. L. Lorentzen, Propertzes of lzmzt sets and com~ergenceof contznued fractions, J . hlath. Anal.
and Appl. 185(2) (1994), 229 255.
18. L. Lorentzen, Convergence of composztions of self-mappings, Ann. Univ. Marie Curie Sklo-
dowska (to appear)
19. L. Lorentzen, Convergence of limzt periodzc contznued fractzons K ( a , , / l ) of parabolzc type,
(in preparation).
20. L. Lorentzen and St. Ruscheweyh, Simple convergence sets for contznued fractzon,s K ( a , , / l ) ,
Math. Anal. and Appl., 179(2) (1993), 349 370. -
21. L. Lorentzen and H. Waadeland, Continued fractions with applicatzons, Studies in Compu-
tational Mathematics 3 (Elsevier Science, 1992).
22. J . F. Paydon and H. S. Wall, The contznued fraction as a sequence of h e a r fractzonal
transformations, Duke Math. J . 9 (1942), 360-372.
23. G. Piranian and W . J . Thron, Convergence propertzes of sequences of h e a r fractzonal trans-
formations, Michigan Math. J . 4 (1957), 129-135.
24. W . T . Scott and H. S.Wall, Value regions for contznued fractions, Proc. Amer. Math. Soc.
47 (1941), 580-585.
25. W . J. Thron, T w o famzlzes of twin convergence regions for contznued fractzons, Duke Math.
J . 10 (1943), 677-685.
26. W . J. Thron, A family of simple convergence regzons for continued fractions, Duke Math. J .
11 (1944), 779-791.
+
27. W . J. Thron, Zwzllingskonvergenzgebtete fur Kettenbruche 1 K ( a , , / l ) , d w e n eines dze
<
Krezsscheibe iazrL-l 1 p2 ist, Math. Zeitschr. 70 (1959), 310-344.
28. W . J. Thron, O n parabolic convergence regions for contznued fractzons, Math. Zeitschr. 69
(1958), 173 182.
-
29. H. Waadeland, Boundary versions of Worpitzky's theorem and of parabola theorems, In:
Analytic Theory of Continued Fractions I11 (cd.: L. Jacobsen), Lecture Notes in hlath.,
Springer-Verlag 1406 (1989), 135 142.
DIVISION
OF ~ I A T H E M A . T I ( . A I . S C I E N ~ ENOR\VE(;IAN
S, I ' N I V E R S I TOYF S < . I E N ( ' E AN11 .I'I<('H-
NOLO<:Y.N 7 4 9 1 TRONUHF:IM. NORWAY
E-mail address: lisaomath .ntnu.no
Contemporary M a t h e m a t i c s
Volume 236. 1999
Olav Njgstad
1. Preliminaries
The Stieltjes transform F, of a finite measure y on ( - m , oo) is the function
(1.2) c,, = 1
-z
t n c i p ( t ) , n = 0.*I, * 2 . . . .
1991 Mathematzcs Subject Classification. AMS Subject Classification: 30E05, 42C05, 44A60.
Key words: Strong moment problems, Orthogonal Laurent polynomials.
257
258 OLAV NJASTAD
When the SHMP is solvable, this linear functional gives rise to an inner product
(,) on the space of real Laurent polynomials by the formula
(1.5) (ftd = MIS .9l.
Then for all solutions ,u of the SHMP we have
1f
3i
The focus of this paper is on the study of solutions of the SSMP. W e shall in
all that follows assume that the SSMP zs solvable. Many of the basic results in the
theory are, however, valid as long as the SHMP is solvable.
For the basic theory of orthogonal Laurent polynomials and strong moment
problems we refer to [5]-[6]. [lo], [12], [14]-[18], [23]-[29]. For the analogous theory
of orthogonal polynomials and classical moment problems, we refer to e.g. [I]-(41,
[7]-[9], [13], [19]-[22], [30]-[35], [37].
(See e.g. [16. p. 3651, and [5, p. 191, [6, p. 371 where different. but equivalent
formulations of the recurrence relation are given in the more general setting of a
regular SHILIP. Cf. also [24, p. 3221. (In [24, p. 3221 initial conditions are given
for the general situation. These conditions reduce to (2.12) when co = 1. which we
have assumed in this paper.) The coefficients f,,,g,,, h,, are given in terms of the
leading and trailing coefficients by
260 OLAV NJASTAD
The zeros <;I, k = 1,.. . n , of p, are simple and positive. A discrete measure
p ( n ) with positive masses xP)at the points <PI,
k = 1,.. . , n , is defined through
a Gaussian quadrature formula. These measures satisfy
M [ L ]=
S
0
( t ) L E A-2m.2rrL-1
~ ( t ) d p ( ~ ~ )for
[24, p. 3381.
s}
The approxzmants of the continued fraction defined by the recurrence formulas
(2.10)-(2.12) are the quotients y,/p,,. The sequences { - and { - ,,.+I(.,
converge in the complex plane outside the non-negative real axis to the Stieltje
}
transforms F,ci,, (z) and FA,(,, ( 2 ) (which coincide if and only if the SSMP is deter-
minate). See [18, p. 5151, [24, p. 3381.
T
(3.4) 7J~21,,+,
( 2 , ~=) V h + L( 2 ) - ;v2rn ( 2 ) .
The quas~-approzzmantsof the continued fraction determined by forinulas (2.10)-
(2.12) are the quotients
The zeros <r)(~). k = 1 , .. . , n,of p,,(z, T ) are real and simple. and at least
(
n - 1 of them are positive. See [24. p. 3261. A discrete measure p;) with positive
) the points < l 1 ' ) ( r )k, = 1 , .. . , n , is defined through a certain
masses x ~ ' ) ( Tat
quadrature formnula. Thcse measures satisfy
(3.7) n r [ q= 1
-X
x
forl )L ( tt A
~(t)d~y~~~+ ) -,.,,,,,,.
See [ I s , p. 548-5491. [24, p. 3271.
The quasi-approximant R , , ( z , T )is the negative of the Stieltjes transform of the
measure p:').
For each z in the open upper half-plane UJ, the AIoebius transformation
maps the extented real line @ onto a circle in U. We use the notation A , , ( z ) for
the open disk bourldecl by this circle, d A , , ( z ) for the circle itself, and a,,
( z ) for the
closed disk A,, ( 2 ) U 3A,,( z ) . Thc points in ~ A , , ( are
z ) exactly the values at z of
(
the Stieltjes transforms of the measures p:').
The sequence {A,, ( z ) ) is nested, i.e., A,,+l( z ) C A,, ( z ) . \Ye define
n &,(.I.
X
(3.9) a x ( z =,
72=1
no] when n = 2 m
All the zeros of , , ( z , r ) are positive if and only if r E [h27rr,
and if and only if r E [-no, hLr,,+l]when n = 2 m +
1. See [24, p. 3261. The
(
corresponding measures ,LL$) satisfy
-1
x
(3.13) AqL] = 1
0
~ ( t ) d p ~ " +( tl )) for 1, E A - 2 n i . ~ ~ ~ -
The values at z of the Stieltjes transforms of these measures cover a subarc R,,(z)
of a h , , ( z ) . These arcs tend to a limiting subarc R,(z) (which may reduce to a
single point) of a A x ( z ) as n tends to infinity. This arc consists of the values at z
of the Stieltjes transforms F, of quasi-natural solutions p of the SSMP. (That is:
quasi-natural solutions of the SHMP which are also solutions of the SSMP.) Thus
(3.14) R,(z) = {w = E;,(z) : /I is a quasi-natural solution of the SSMP).
See [24, p. 3301.
The end points of R , ( z ) are the values F,(o, ( z ) and Fp(,, ( 2 ) .
The pseudo-orthogonal L a u r m t polynomzals an( z ,r ) of order n are defined by
(3.18) h l [ L ]=
i
-x
~ ( t ) d v ? " ) ( t ) for L E A - ( L m - ~~) ~ n - 1
maps onto a circle dD,(z) in U.We denote by D n ( z ) the open disk bounded by
this circle and by D,, ( z ) the closed disk D, ( z )u d D , ( 2 ) . The points in dD,, ( 2 ) are
exactly the values at z of the Stieltjes transforms of the measures up).
All the zeros of a,, ( z ,T ) are positive if and only if T E [-co,gznL]when n = 2 m
+
and if and only if r E [g2,,,+l,m] when n = 2 m 1. See [24, p. 3301. In these
cases, and only zn these all the weights K ~ ) ( T ) k, = 1 , . . . , n, are positive. The
corresponding positive measures u p ' satisfy
(3.21) M [ L ]=
7
0
~ ( t ) d v y " ~ for
( t )L E A - ( 2 m - ~ ) . 2 m - i
(3.22) M [ L ]=
i
o
) L E A-(2m-1).2,n+l.
~ ( t ) d u ! 2 " + ' ) ( t for
The values at z of the Stieltjes transforms of these positive measures cover a subarc
r , ( z ) of dD,,(z). Furthermore, the arc r , ( z ) consists of exactly that part of d D n ( z )
which is contained in An-l ( 2 ) . Similarly the arc R,(z) consists of exactly that part
of d A , ( z ) which is contained in D,-I ( 2 ) . I.e.,
This set is a single point or a proper closed disk. It can be shown that an invariance
result holds also for these sets: Either D,(z) is a single point for every z E UJ or
D , ( z ) is a proper disk for every z E UJ. See [28].The arcs r , ( z ) tend t o a limiting
subarc ,?I ( z ) (which may reduce t o a single point) of d D , ( z ) as n tends to infinity.
The points of r,(z) are exactly the values at z of the Stieltjes transforms of all
solutions p of the SSMP obtained as limits of subsequences of sequences {u!:)).
See [24, p. 3401. Such solutions are called pseudo-natural solutzons. Thus:
(Recall that we rlo not obtain solutiorls of the SHhlP from sequences {v!::)} where
T, does not eventually belong to the T-intervals [-m, g ~ ~and , ] [gLrn+l,
m ] . ) The
end points of r, (z) are the values F,,l~i( 2 ) and I.;,(, i (z). \Ire also have
W e agazn stress the fact that pseudo-natural solutzons sn the sense here dzs-
cussed only exsst when the SSMP zs solvable. and these solutions are automatically
solutions of the SShIP (not only of the SHhIP).
A simple convexity argument shows that every point in the lens-shaped region
a, (z) n D, (z) is the value of the Stieltjes transform at z of a solution of the
SSMP. It can be shown that all such values exactly cover thzs regzon, see [28].
m-l
OLAV NJASTAD
and
Repeating this process and taking into account the initial conditions (2.12) we get
(4.5).
In a similar way we get (4.6). By using analogous arguments involving functions
$, as well as cp, and taking into account that (see (2.12))
we obtain (4.7)-(4.8). Finally by using arguments involving only the functions '$',
we obtain (4.9)-(4.10). 0
THEOREM 4.2. Let a,b,c,d be arbitrary complex numbers and let z , be complex <
numbers different from zero. T h e n the following formulas hold for m = 1 , 2 , .. . :
[a$2rn(z) + b(P2m(z)][~$2m-l
( C ) + d ( ~ 2 m - (<)I
- [ a d ~ 2 r n l ( z+
I
) b ~ 2 r n -l ( z ) I [ ~ $ 2 r n( C ) + d ~ 2 (<)I
m 1
PROOF:By multiplying out on both sides of (4.15) and (4.16) and substi-
tuting from (4.5)-(4.6), (4.9)-(4.10) and (4.7)-(4.8) (the last also with z and <
interchanged), the result follows. 0
It follows that the limit point situation for the system { A , ( z ) ) occurs if and only
if the series CT=o lpk( 2 )l 2 diverges. For discussions of these properties of the disk
system { & , ( z ) ) we refer t o [16. p. 3731, [26, p. 3321, [29, p. 151.
Our aim in this section is to study the disk system { D , , ( z ) )in some detail.
Again by standard properties of Moebius transformations we find that the
center s , , ( z )of D , , ( z ) is given by
n1- 1
U2k+l U 2 k + 2
U: E D2ni+l(z) 6 If? C -
uZk+l u2k+2
1 + 2 k + l ( ~ )+i~i02k+l(zV
k=O
(5.8) rn
)@2k(~) + ~ p 2 k ( ~5 )ulz
-. / ~
-
z-z
WZ
k=O
The point U: belongs to dD,, ( z ) if and only if 7 belongs to k. i.e., if and only if the
imaginary part of the expression [ p , ( z ) d + v", +
( z ) ] [p,,- I ( z ) ~& - I ( z ) ] is Zero.
This means
STRONG STIELTJES MOMENT PROBLEMS 269
(Recall that the coefficlcnts in the orthogonal Laurcnt pol! nornials are real.)
Since D,, (2) is contained in UJ and hence J = 0 does not hdong to Dl, (2). we
conclude that DLllr (z)and (z)arc given by the inequalities (5.7)-(5.8).0
PROOF:From (4.1)-(4.2) we find that the numerator (inside the absolute value
'LLn 1
sign) in (5.6) equals ---- . - . The denominator is given by (4.5)-(4.6)with = 2.
un-1 z
<
Substituting these expressions in formula (5.6) for r,,(z) we obtain (5.15)-(5.16).
THEOREM
5.3. T h e limiting disk D,(z) reduces t o a single point zf a n d only
if t h e series
diverges.
PROOF: The limit point case occurs if and only if lim r , = 0. Hence the
n+x
result follows from Theorem 5.2. 0
PROOF: First assume that the SSMP is indeterminate. It follows from (3.28)
and the fact that the arc R,(z) does not reduce to a single point that ~ , ( z )does
not reduce to a single point.
STRONG STIELTJES MOMENT PROBLEMS 271
Here the absolute value of the first term t o the left is l / r , , ( 2 ) . while for even n the
absolute value of the second term to the left is l / p , , ( z ) . (A similar formula with
odd n may also be used.) Since we have assumed that the SHhlP is indeterminate.
it follows that l / p n ( z ) is bounded for each z in UJ. Since we have also assumed
that the SSMP is determinate, it follows that the sequence {Y '1 (2)
--)converges for
%I (')
every z E UJ (see [18, p. 5151, [24, p. 3381). and consequently the right-hand side of
(6.1) is unbounded for every z E U.We conclude that l l r , , ( z ) is unbounded, which
implies that lim r,, ( 2 ) = 0. This means that D,(z) reduces to a single point.
1L"X
which completes the proof.
COROLLARY 6.2 Assume that the SHMP is indeterminate. T h e n the SSMP is
determinate if and onlg if
for all z in U.
PROOF: This follows immediately from Theorem 5.3 and Theorem 6.1. O
COROLLARY 6.3 Assume that the SHMP is indeterminate and the SSMP de-
terminate. T h e n the sequence {-) i s unbounded.
Vll
1L71
PROOF: This follows immediately from Corollary 6.2 and the fact that the
X
SHMP is indeterminate if and only if C
Jp,,( z ) l L< cm. 0
n=1
For the sake of completeness we shall also state a general necessary arid sufficient
condition for the SSMP to be determinate.
We define polynomials P, , Q , by
zV'$ct.)nt(z)
Anl(z)=
Lnr 277-2 ' ' ' $2 7 '
272 OLAV NJASTAD
where
The -polvnomials P,, znd Q,, are thus canonical numerators and denominators
of a positive T-fraction (see [5]-[6]. [17]-[18]). Since "-- - '"(')
- - , and since the
Qn(z) ~n(z)
@ (2) converges outside the non-negative real
SShZP is determinate if and only if {"--)
~n (2)
axis. it follows that the SShlP is determinate if and only if the positive T-fraction
x X
converges. This is the case if and only if at least one of the series C d,,
n=l
, C en
72=1
diverges. (See 117. p. 1461, [18, p. 5151). Consequently, by substituting from (2.13)-
(2.15) in (6.8)-(6.9), we obtain a characterization of determinacy:
THEOREM 6.4. The SSMP is determinate if and only if at least one of the
following series dzverges:
PROOF:This follows from Corollary 6.2, Theorem 6.4 and the fact that the
s
SHMP is determinate if and only if C lvn(z)12 = m. 0
n= 1
References
1. N.I. Akhiezer, The Classical Moment Problem and Some Related Questions
in Analysis, Hafner, New York, 1965.
2. C. Berg, Markov's theorem revisited, J . Approximation Theory 78 (1994)
260-275.
3. C. Berg, In,determinate moment problems and the theory of entire functions,
J . Comp. Appl. Math. 65 (1995) 27-55.
4. C. Berg and G. Valent, The Nevanlinna parametrization for som indeter-
minate Stieltjes moment problems associated with birth and death processes,
Methods Appl. Anal. 1 (1994) 169-209.
5. C.M. Bonan-Hamada, W.B. Jones and 0 . Nj6stad, Natural solutions of inde-
terminate strong Stieltjes moment problems derived from PC-fractions, In:
Orthogonal Fun~t~ions, Moment Theory and Continued Fractions: Theory
and Applications, (W.B. Jones and A. Sri Ranga, editors), Marcel Dekker,
1998, pp. 15-30.
6. C.M. Bonan-Hamada, W.B. Jones, 0 . Njbstad and W.J. Thron, A class
of indeterminate strong Stieltjes moment problems with discrete distribu-
tions, In: Orthogonal Functions, Moment Theory and Continued Fractions:
Theory and Applications, (W.B. Jones and A. Sri Ranga, editors), Marcel
Dekker, 1998, pp. 31-55.
7. T.S. Chihara, On indeterminate Hamburger moment problems, Pacific J .
Math. 27 (1968), 475-484.
8. T.S. Chihara, An Introduction to Orthogonal Polynomials, Gordon and
Breach, New York, 1978.
9. T.S. Chihara, Indeterminate symmetric moment problems, J . Math. Anal.
Appl. 85 (1982) 331-346.
10. L. Cochran and S. Clement Cooper, Orthogonal Laurent polynomials on the
real line, In: Continued Fractions and Orthogonal Functions, (S. Clement
Cooper and W.J. Thron, eds.), Marcel Dekker, New York, 1994. pp. 47-100.
11. V.A. Fil'stinskii, The power moment problem on the entire axis with a fi-
nite number of empty intervals in the spectrum, Zap. hlekh. Mat. Fak.
Khar'kov. Gos. Univ. Khar'kov Mat. Obshch 30 Ser 4 (1964) 186-200 (in
Russian).
12. P. Gonzalez-Vera and 0 . Njbstad, Convergence of two-point Pad4 approxi-
mants to series of Stieltjes, J. Comp. Appl. Math. 32 (1990) 97-105.
13. H. Hamburger, ~ b e eine
r Erweiterung des Stieltjesschen Momentproblems,
Parts I, 11, 111,Math. Annalen 81 (1920), 235-319; 82 (1921), 120-164; 82
(1921), 168-187.
14. E. Hendriksen and H. van Rossum, Orthogonal Laurent polynomials, Indag,
Math. 89 (1986). 17-36.
274 OLAV NJASTAD
35. T.J. Stieltjes. Rech,erches s u r les fractions continues. Aim. Fac. Sci. Toulouse
8 (l894), J . 1-122, 9 (1894), A 1-47; Oeuveres 2, 402-566.
36. K.I. ~ v e c o v ,On Hamburger's m o m e n t problem with the supplementary re-
quirement that masses are absent o n a given interval. Cornnxln. Soc. Alat,h.
Kharkov 16 (1939) 121-128 (in Russian).
37. W. Van Assche, T h e ,impact of Stieltjes' umrk o n c o n t i ' n ~ ~ ef rdn c t ~ o n s and
orthogonal polynornials. In: T.J. Stieltjes. Collected Papers. Springer-V~-lag
1993.
DEP~\RTSIENTO F . \ I A ' I ' H E ~ I ~ \ T I c ~ \ IS, C I E N C E S , ~ O R \ Z ' E G I A NI;XI\'EI<SII.Y
OF S C I E N C E .\XI) ~'I.:('H-
K-7031 T R O X U H E I ~ I KOH\VAY
NOLOGY, ,
E-mad address: n j astadQmath.ntnu. no
Contemporary Mathematics
Vnlump 236. 1999
1. Introduction
Let M be the set of all nonnegative finite Bore1 measures a on [O, 271.). normal-
ized by
It is well known that there exists a uniquely determined sequence of morlic orthog-
onal polynomials {P,,(z,a ) = zrL . . . + )rz0 on the unit circle with respect to a.
that means
Often it is more natural to deal with orthonormal polynomials a,, (z, a ) instead of
the monic ortliogonal polynornials. which satisfy
i.e., the reflection coefficients are the evaluation of the monic orthogonal polynomi-
als at the point z = 0.Furthermore. they satisfy
(1.5) la,,(a)l < 1 for a1ln.E NO.
In this papcr we study two sequences of ortlionormal polynon~ials{@,,(z,a,)))
and {(a,, (z, a ) ) . where we suppose that the second sequence results from the first
one by a compact perturbation of the reflection coefficients, i.e., we assume that
It =O
In [14] the authors studied ratio asymptotics of the perturbed and unperturbed
orthogonal polynomials outside the supports of the orthogonality measures.
To state and prove our results, let us consider some basic facts which hold true
for all orthogonal polynomials on the unit circle: Let p E M be a given measure.
From (1.3) and the relation
for all n E N,compare [7,formula (6)]. Let us consider the infinite matrix (compare
again [7, p.4011)
and as a consequence
) a11 k E N
z k u ( z , , u ) = U ~ ( ~ ) U ( Z , ~ for
Notice that Uk((CLis a triangle matrix with ( k + 1) superdiagonals. Let us denote
U k ( P I = (&(cL)):~=,,, then
and yet
(2.6) lirn
n--'x
(lT ezkpQrz(eZ*',
a)@,,+,(ez*',a ) d a ( p )
holds for all k , g E Z.We start with the case that k E No. By (2.3) we have to
show that
k
lim
I1'X (ufi,,+, (a)- U , n + (00))
~ = 0
for all fixed J E Z. It is not difficult t o see that by the triangle structure of
the matrix U the elements on the diagonals {u: ,,+,( p ) ) ~ x 7 , p, I ,E {go, a ) , are all
generated by the same law in terms of a finite number (depending on j and k) of
reflection coefficients, i.e.. u?,,+,) (71+1)+J(p)
results from uk ,+,(p) by a 1-shift of
the indices of the reflection coefficients. Now recall that a,,(p) = P,,+, (0, p ) and
Then (2.6) follows from (2.1). For negative k one only has to take complex conju-
gation in (2.6).
Hence, we have shown the desired limit relation (2.4) resp. (2.5) for all functions
g of the form g ( p ) = e " ~ ,where k is an integer. By linearity and continuity (recall
also the Banach-Steinhaus Theorem) relation (2.5) is also true for an arbitrary
continuous 2~-periodicfunction. Applying one side approximation arguments (see
e.g. [20, Theorem 1.5.4]), the assertion follows. 0
Let us note that the limit relation (2.4) (resp. (2.5)) holds true if one replaces
) any subinterval [ d l ,&], because one can choose the function g to vanish
[O, 2 ~ by
identically outside the interval [dl,d2].
Especially, if we put j = 0 in Theorem 1, then we get
If the reflection coefficients P,(O, a ) and P,, (0, ao)tend to zero, w e n more is known:
Let us first consider the Szego-class, that is the class of all finite and nonnegative
ORTHOGONAL POLYNOMIALS AND FUNCTIONS O F T H E SECOND KIND 281
Borel-measures p on [ O , ~ Twhich
) satisfy
or equivalently
.x
One of the main results of Szego's theory is the limit relation (see [20, Chap. X
and XI, pp. 274-2951 and [2, Chap.V, pp.187-2721)
Again, if both a:, > 0 and a' > 0 almost everywhere in [0, 27r) then
If the reflection coefficients Pn(O,00)do not tend to zero, then in contrast to the
case just discussed, the limit relation (2.7), and hence the results of Theorem 1. can-
not be improved in general, neither by replacing the weak-* limit by LI-convergence
nor by pointwise convergence almost everywhere. We will demonstrate this fact by
the following
Counterexample. Consider the two sequences of reflection coefficients
where, for simplicity, a E ( - 1 , l ) \ (0). It is known. see e.g. [3, 7, 151. t,hat go
and a are absolutely continuous on [a,27r - a], with cos a / 2 := -./, The
282 FRANZ PEHERSTORFER AND ROBERT STEINBAUER
for cp E [a,27r -
and
a[)(cp) = al(cp) = 0 for cp E [O,27r) \ [a, 27r - a ] .
For the representation of f we have used [15, formula (2.21) and Theorem 4.31.
The value a can be chosen such that a0 is an absolutely continuous measure and
that a has at most two mass-points at * P , where /3 E (0, a); a possible choice for
instance is a = 0.6.
We will show that
- a].
Let us remark that it makes sense to consider convergence only on [a, 27r - a] and
not on [O,27r) because supp(ao) \ [a,27r - a] = 0 and @, (e*@, a ) --- 0 as n + CQ.
It sufficies only to show (2.9) because then, as a consequence,
from which (2.10) follows. For this last implication we made use of the fact that
the a, (e", ao)'s and a,, (ezp,0)'s are uniformly bounded away both from zero and
from infinity on [a + c,27r - a - €1, for all 6 > 0, which can be seen from [16,
Corollary 2.2 resp. Lemma 3.11.
Moreover, the left hand side in (2.11) cannot converge to zero almost every-
where, because otherwise, by the equi-integrable property of the orthonormal poly-
nomials and by standard results from measure theory, this would imply that also
the L1-limit in (2.10) would be zero.
Let us now prove (2.9): It can be derived from [17, Corollary 11 that
for every measurable and Riemann integrable function g. Further, recall the fact
that for every measure ,LL E M by [ 2 , Theorem 2.21
Now, if the left hand side in (2.9) would converge t o 1 in the weak-* sense then
and by (2.14)
.-
For the first integral at the right hand side we can write
sin 5 by (2.13)
(see [12, Corollary 3.11) by P,+, (z, a(]) and if we use the limit relations, given in
[16. Theorem 2.11. we arrive at
But this value is in general different from the right hand side of (2.15), which has
the form
c, :=
2.rr-tu
S,
.
sln - a . sin Q
sin
5
"$0.
5
To see this: we calculate both values cl and c 2 with a = 0.6 (compare above). Using
Mathematica, we get
cl = 4.80319 # 3.78767 = CZ.
Further, let
which is by [6, Theoorem 13.11 of order O(z7"') for J z /< 1. Hence, the functions of
the second kind can be considered as a measure how well the Carathkodory-function
F ( z , p ) can be approximated by the ratio PFL( 2 , p)/@: (z, p ) .
(ii) Associated measures: For a given measure p from M we can define the u-th
associated measure p("), which is related to the shifted reflection coefficients
) ) aTL+"(p), n E NO.
a 7 , ( ~ ( y:=
In view of [12, Theorem 3.11 we have
Thus, if one wants to know the asymptotic behaviour of the u-th associated poly-
nomials @, (z, p") with respect to n as well as with respect to u one should know
about the behaviour of the functions of the second kind (see [14, 161).
(iii) Schur functions: Let
Furthermore, all f,'s are again Schur functions, that is f,, analytic and 1 f,(z)l < 1
on { ( z /< I ) , and by (3.4) and (3.7) there holds
i.e.,
f1,(z, P) = f (z, p i ' ) ) .
(compare also [6, Theorem 18.2, formula (18.8)]).
(iv) Connectzon wzth functzons of the second kznd on the real h e : It is known
and important for asymptotic investigations that polynomials orthogonal on [-I, 11
286 FRANZ PEHERSTORFER AND ROBERT STEINBAUER
can be expressed with the help of polynomials orthogonal on the unit circle. It has
been shown by the first author [ll,Section 31 that there is a corresponding relation
between functions of the second kind and how to get asymptotics with the help of
this relation.
(v) Stieltjes-polynomials and Gauss-Kronrod quadrature: Stieltjes-polynomials
are defined as partial sums of the reciprocal of the function of the second kind.
Roughly speaking, to get asymptotics for Stieltjes-polynomials means t o derive
asymptotics for the functions of the second kind (see [ll,Section 41).
Note that the functions F ( z , p ) , &(z, p ) and E,(z, p ) are elements of the
Hardy spaces H, = H,({lzl < I ) ) , p < 1 (see e.g. [l,Theorem 3.21). But if we
take the integral representation in (3.1) and (3.3), we can consider the functions
F ( z , p ) , 6, ( 2 , p ) and 'Hn( 2 , p ) to be defined on @ \ {e" : p cp supp(p)), where
supp(p) denotes the support of the measure p. Then it is not difficult to obtain
the following relations
(this can be easily derived from [6, formula (12.13), p.201 by setting ~j = F ( z ,p ) )
and
(follows directly from the integral representation or from (3.8)) hold true.
From the well known relation (see [6, formula (5.6), p.71)
= ( ) ( Z ) + 1
1
0
2"
(icn(l')ezTfP+terms of lower order)@,,(eZc,P) ~ P ( P )
Since K has a positive distance to the set {eZq: p E supp(p)) - note that supp(p)
is a closed set - we obtain that
Hence,
>
1 1 on K n {lzl 5 1) (where
and the identity (3.11) is only possible if I@;L(z)Gr,(z)
we have strict > 1 on { J z J< 1)). The analog statement on z@,,'HT,follows from
(3.8).
Finally, the limit relation
lim
z-ix
-
27l
2x e" + z
--- /@n(e'",p)12+(P) =
eZ+- z
1
-I;; 1
0
27I
@,.(ezP.P ) I ' ~ P ( P ) = -1
71-X
lim Q:,(z.p) = D(Z, uniformly on {Izl < r < 1)
and it is shown by the first author in [ll,Theorem 2.11 that
TI -
linl
X
G,, (z. p ) = 2 . D ( z , p') uniformly on {lzl < r < 1).
Herr D ( z , pl) denotes the Szegii-function defined by
+
el*? z
el*?- z
(3.22)
n
11
-
lim
t
x
\,I
za,, (z. o)'H,, (z, o ) = O(Z:rr, A())- 1
and one obtains that O(z) # O on {izl # 1). The limit relations (3.21) again follow
from (3.15). 0
Let us recall the above mentioned relations of Carathkodory-, Schur- and func-
tions of the second kind
The next theorem gives asymptotics of the n-th associated Carathkodory and
Schur functions.
THEOREM 2. Let the measures a(),a E M be related by (2.1). Further let
S c (121 = 1) be the closure of all zeros of G,(z,P) - zXn(z,P), where ,LL E {a(),a)
and n E No. Then
and let 1V be a closed subset of [O,27r)\ (supp(p) U {p : el' E S)). Then it is not
difficult to derive from (3.27) that
( " ), ) = lirn ~e re", p('") = 0
~e ~ ( e 'p ~ for p E N
7 -1
where the constant c,, only depends on v. Thus, for every fixed v the limit relations
ORTHOGONAL POLYNOMIALS AND FUNCTIONS O F T H E SECOND KIND 291
where by [14, formula ((4)) in the proof of Lemma 3.1 and Remark S . l ( h ) ] the
constant can be chosen independent of n and v .
Collecting all the above given informations we derive
lim
71 - X
) -0
f
- F ( Z , ~ ~ ' )3 ~ and
< - I *ill@tob,
(2,
a("')
Then take no = n o ( € )as large such that by (3.29) for all n > no
Let us point out once more that there are no assumptions on the reflection coeffi-
cients {a,(ao)) and {a, (a))apart from (2.1).
From (3.18) and [14, Lemma 3.1(a)] we know that
lim l-In(z,p ) =0 uniformly on K,
n-x
if and only if
(3.33) lim
( a )
-
1 J(z - + 41aI2z + z ( l - 2a) - 1
n-= G,(z, a )
- -
z J(z - +4 +
/ ~ / ~(1z- 2a) - z
uniformly on {lzl < r < 1).
Proof. Let us assume that lim,,, a n ( a ) = a. Take { a n ( a o ) := a)nEWoas a
comparison sequence. Then by [15,Thm.4.4 resp. Thm.5.1 and Example 5.l(a).
p.100]3
F ( z , go) = F ( z , o, )
(n)
=
(z - +
41aI2z - ( a + az) for all n E No.
(1 - a ) - (1 - ti)z
By Theorem 2 we have
<
uniformly on {Izi r < 1); recall that by (3.9) there holds Iz'H,,(z. a)/G,(z, a ) / <
IzI for IzI < 1. NOW(3.33) follows quite easily.
Next suppose that (3.33) holds. We will apply this relation at the point z = 0.
From (3.3) we obtain
and, hence,
Now, 1'Hospital's rule shows that the above right hand side is nothing but -a. This
proves the assertion. 0
Remark. (a) Let us mention that a corresponding result as stated in Corollary 1
also holds true if we consider asymptotically N-periodic reflection coefficients a,,(a),
i.e.,
(3.35) ) a3(a0) for j
linl a 3 + , , ~ ( u= = 0,.. . ,N - 1,
n+?r
where a 3 + 7 L ~ (=
u~a3(uo)
) for all n E N and j E ( 0 , . . . ,N - 1). Indeed, it can be
shown with the help of the relations
Under these assumptions we have shown the following asymptotics of t,he corre-
sponding orthonormal polynomials (see also our survey paper [13]):
THEOREM 3. ([14, Theorem 2.2 and 3.21) Let a 0 and u be given orthogonal
measures from M which satisfy (3.36) and (3.37). Then there hold:
294 FRANZ PEHERSTORFER AND ROBERT STEINBAUER
(a) Let K 1 be a closed subset of {Izl < 1) and assume that the sequence
lim
(b) Let K 2 be a closed subset of {lzl = I), which has a positive distance to
*
supp(p) and to the zeros of 4, (z, p ) z'H, ( 2 , p ) , p E {ao,a ) . Further assume that
instead of (3.38) the sequence
(3.40)
{ C LI @ ; )L ~ ,
I@:(z,
a())12
md2
}ncNo
is uniformly bounded o n K 2
With the help of Theorem 3 and the above results we can derive ratio asymp-
totics for the functions of the second kind.
THEOREM 4. Let 00 and a be given orthogonality measures from M which
satisfy (3.36) and (3.37) and let K be a closed set i n ,: which has a positive
*
distance to supp(p) and to the zeros of 4,,(z, p ) z'H, (z, p ) , p E {no, a ) . Further
suppose that (3.38) holds i n the interior of the unit circle and (3.40) o n a set
K n (1 - t 5 lzl 5 l ) , E > 0 sufJiciently small, if this set is nonempty. T h e n
lim =~ ( z )o n K n {lzl 5 1)
n-x 4n(z, 0)
lim "n(z'""
'H,(z,a)
T ~ + X
= A(z) on K n {lzl 2 I),
where both limit relations hold uniformly o n compact subsets ( i n c).
Proof. From the integral representation (3.3) of G,(z,p), p E {ao,a),it is
easily to see that
and by (3.8)
ORTHOGONAL POLYNOMIALS AND FUNCTIONS O F T H E SECOND KIND 295
lim
where we used (3.11) for the last identity. Taking the lirnit TL +m at both sides,
we obtain by Theorem 4
uniformly on K 1 . By [14, Lemma 2.21 the function A has no zeros on KI" and is
analytic there. Hence, (3.41) reduces to
References
P. L. Duren, T h e o q of
H P spaces, Academic Press, New York and London, 1970
G. Freud, Orthogonal polynomials, Pergamon, Oxford/New York/Toronto, 1971
Ya. L. Geronimus, O n polynomials orthogonal o n the circle, on trigonometric moment-
problem and on allied Caratheodory and Schur Funktzons, C.R. (Doklady), Acad. Sci. URSS
(N.S.) 29(1943), 291 295
-
Ya. L. Geronimus, Polynomzals orthogonal on a circle and znterval, Pergamon Press, New
York, 1960
Ya. L. Geronimus, Orthogonal Polynomials, Consultants Bureau, New York, 1961
Ya. L. Geronimus, Polynomzals orthogonal o n a circle and thezr applzcations, Amer. Math.
Soc. Transl. 3(1962), 1 78 -
ernma ma 2.2 in [14]says that A ( t ) # 0 on 121 < 1. But the analog result also holds true
on K 1 n {It1 = 1) by exact the same proof as given for Lemma 2.2 in [14]by using Theorem 3.2
instead of Theorem 2.2 in [14]
296 FRANZ PEHERSTORFER AND ROBERT STEINBAUER
[7] L. Golinskii, P. Nevai, and W. Van Assche, Perturbatzon of orthogonal polynomzals o n a n arc
of the unzt circle, J . Approx. Theory 83(1995), 392 422
-
[8] A. MBtC, P. Nevai, and V. Totik, Strong and weak convergence of orthogonal polynomials,
Amer. Jour. Math. 109(1987), 239 282
[9] P. Nevai, Weakly convergent sequences of functions and orthogonal polynomials, J . Approx.
Theory 65(1991), 322 340
-
[12] F. Peherstorfer, A special class of polynomzals orthogonal o n the unit czrcle includzng the
associated polynomials, Constr. Approx. 12(1996), 161 186
-
[13] F. Peherstorfer and R. Steinbauer, Perturbation of orthogonal polynomials o n the unit circle.
A survey, Proc. of the Workshop "Orthogonal polynomials on the unit circle: Theory and
Application", Madrid-Leganes, 1994, 97 119
-
[14] F. Peherstorfer and R. Steinbauer, Comparative asymptotics for perturbed orthogonal poly-
nomials, Trans. Amer. Math. Soc. 348(4)(1996), 1459 1486
-
[15] F. Peherstorfer and R . Steinbauer, Orthogonal polynomials on arcs of the unzt circle II.
Orthogonal polynomzals with perzodic reflection coeficients, J . Approx. Theory 87(1996), 60
- 102
[16] F. Peherstorfer and R. Steinbauer, Asymptotic behaviour of orthogonal polynomials o n the
unzt circle wzth asymptotzcally periodic reflectzon coeficzeds, J . Approx. Theory 88(1997),
316 353
-
[20] G. Szego, Orthogonal polynomials, 4th ed., Amer. Math. Soc. Colloq. Publ., Vol. 23, Amer.
Math. Soc., Providence, R.I., 1975
1991 Mathematics Subject Classzfication. Primary 11504, 11506; Secondary 11D09, llH70,
llK55.
Where:
V i = O , . . . , n , a , N*
~
We associate to S the matrix Ms defined by:
And these two numbers give birth to the binary quadratic form:
The discriminant:
So we have:
When the situation is simple, as for example in the case of the classical Markoff
theory, we have only:
In the rest of this article, we suppose that the sequence S*can be written with
two other sequences of integers X I and X z , and an integer b E N* in the following
form:
S*= ( X I , 6, X2)
Such a decomposition is also written, when we want t,o identify the structure
of S * :
With:
b = a,
XI = ( a n , . . . ,%+I)
X2 = (a,-1, . . . , a")
(15) %, = [y: f 2
I where = det(Mxl) = (- 1)"-'
(16) Mx2=
mz
k.21
m2 - k.2
k.21 - 12 I where ~2 = det (Mx,) =
Writting Ms.with those two matrices, we find many relations among which
(-1)"
Those relations show that the numbers m , m l , rn2 are positive terms. We have
also:
(28) +>0
a + r + 1 = a + (l/[S]) (l/[S*])
As for the form Fs, it gives the following relations:
(29) Fe(K1,m ) = -E
(30) Fe(K2 - ( a + l ) m , m ) = -E
And if we introduce now the number A given by:
(31) Fo(h,m1) = A
We easily find with the expression (6) of Fe that we have:
This can be compared to the other expression coming from the development of
the matrix Ms. :
Considering now the new quadratic form cP0 built with z = rnx - K1y
It can be noticed, with remarks made before, that when u > 0, we must have:
(39) El = E2 = 1
So we have then to consider the diophantine equation M + ( a ,6, u):
the same as in the former equation (Ad:). There is no difficulty to understand the
link between those new values and the old ones.
We showed that the solutions are organized in a tree similar to the one existing
for the classical Markoff theory [I]. So the fundamental solution (1,1,1) gives all
(42)
(43)
X : (m, m l , m2)
Y : (m, m1.m2)
--
the other solutions with the three following transformations:
To each Cohn triple, we can associate a sequence S* having the expression (14).
Here, we have a = b = 3. We also showed that we can write:
(48) X = ( 0 2of-1
, 7 . . . ,o o )
We define:
(ax) = ( I , % - ~ , Q ? I , .,no) ..
( a a - l + 1 , a , - 2 , . . . , a ~ )i f o , = 1
such a tree of sequences associated to the tree of Cohn triples can be put in a new
form. Beginning with:
On the right, the things are more complicated, and we had to separate two
cases:
1) We have for two consecutive steps on the right:
2 ) We have for one step on the right consecutive to a step on the left:
It does not seem possible to give such pretty formulas for only one step on the
right. But in fact, it is not necessary, because the formulas given before are sufficient
to build the entire tree of sequences. We could think that a more convenient tree
could be presented for the same triples and corresponding sequences, but the author
of the present article has not yet found it. Anyway, for the case studied here, the
top of the tree of sequences can be described by Table I.
The same general relations (51) t o (60) can be used to present the classical
Markoff theory, which is linked t o the diophantine equation M + ( 2 , 0 , 0 ) . In this
case, we take a = b = 2. All what we have described can be immediately transposed.
At the beginning of the tree, we use also X z = T = 8. And the formulas for going
right or left are the same as those given up.
With such a method, the Cohn triple are again organized in a tree, but this
one is very different from the known ones ( see [I] or 131 ). The construction that
we have made is very peculiar. In fact, in order to build all the sequences of the
T R E E S O F APPROXIMATION CONSTANTS 303
Tree of Tree of
solutions sequences Cohn triple Sequence X2 Sequence T
-
XY 0
XYX D
XY Z G
XYXY DD
XYXZ GD
XY ZX DG
XYZY GG
XYXYX DDD
XYXYZ GDD
XYXZX DGD
XYXZY GGD
XYZXY GDG
XYZXZ DDG
XYZYX DGG
XYZYZ GGG
classical Markoff theory, we need only the following Table I1 of what happens at
the top of our tree:
Tree of Tree of
solutions sequences Cohn triple Sequence X2 Sequence T
We can then compute Mx,and M p . This gives all the numbers m, K1, K2,1,
etc.. .
304 SERGE PERRINE
So:
We now see that when a , 6, and u are given, any solution ( p , A) of this equation
gives At7,. In fact usually, such a solution gives birth to an infinity of solutions,
because the equation is quadratic, and we can use a group of hyperbolic rotations
leaving the equation unchanged. We then find an infinity of solutions of the dio-
phantine equation M + ( a ,6, u ) .
Working on an example, we see now how all this works. Let us choose here
; t
T = (1,2). We find S* = ( 1 , 1 , 1 , 2 , 2 , 0 ), with a = b = 2, and X2 = 0. In this case,
it is easily seen that 6 = 1 and u = 2. So we consider the diophantine equation
M+(2,1,2):
With such a solution, we find a infinity of other solutions using the three trans-
formations:
It is easy to see that it gives an infinity of sequences S*. All of them are
associated to our new equation (76).
With X I we can easily find another quadratic diophantine equation equivalent
to ( E Q ~ ) For
. that, we use the relation (32) under the new form:
We verify that ( m l , k l ) = (50,31) is a solution, and that we can find again the
equation ( E Q ~writting
) Mx,with MT. With the relation (20) we can compute
K 2:
It is now clear why the relation (80) seen before between X1 and ~f gives a
lot of solutions of ( E Q ) and (EQT).
With the Dickson lemma (see [5] vol2, pp. 408-409), we can then compute the
approximation constants of the algebraic numbers associated t o the sequences S*
that we consider. We will not do it here for all the tree derived from the equation
M + ( 2 , 1 , 2 ) , but it is possible to do it. For rapidity, we will do it only for the
sequences built with (80).
306 SERGE PERRINE
The comparison between S1 and 63 is a little more difficult, but not too much.
We see that 61, and 63bare very near, with < Ssb, though 63, and 6lb are quite
far from each other, with 63, < Sib. SO S1 is the greatest of the three considered
numbers, giving the approximation constant:
Those two expressions can be given with only ml and kl, which are linked by
(83). It is easy, with (88), to see that the limit of the constant when n increases is
( 1 / ( 1 + A)).The limit of 6;
' can also be computed and is (113). In fact, we can
write:
This would give a new diophantine equation similar t o (38) with a tree of
solutions (see [6]).
The comparison with a result of Hightower [7] is also interesting, because with
our new example, we can understand the gaps discovered by this author.
T R E E S O F APPROXIMATION CONSTANTS 307
TABLE
111. Analysis of the sequences of the (3,0, -1) theory
We use the computations made before, with Mx,being any matrix built with
X2 sequence of integers, MT given the same way but with T palindromic, and a = b.
We find:
We have then the demonstration of a remark made in [9]that all the diophantine
equations M + ( 3 , 0 ,F2n)have at least a tree of solutions, and we know how to build
one.
The Dickson lemma gives the possibility to compute the corresponding approx-
imation constant. With the same method as used in the former paragraph, we have
to consider the greatest number among the four ones:
We see that d3 = S2 and b4 = 6i. It is also easy t o verify, with the same
notations as before, that we have:
When i is increasing, the value of C converges towards (113). And because the
subsequences of 1 appearing in S* have an odd length, we are not in the classical
Markoff theory. So we have:
And XI and X 2 give the sequence S*with a relation like (14). Our formalism
now creates a tree of sequences applying the rules (52) to (61). Because T is
palindromic, we have always K1 = K2. The corresponding diophantine equation
have the form M + ( 2 , 0 ,u ) . With the methods already known, we can compute u .
Thanks to Fibonacci numbers, we find always the same value:
We can also compute the different values of (m, m l , m2) found before and giving
a triple of solutions of our diophantine equation M + ( 2 , 0 , -2):
So we have established:
THEOREM
5.1. Let us consider the diophantine equation M + ( 2 , 0 ,-2)
Or:
2
F2t+4 - F 2 i + ~ f i z + 3= -1
This equality is an evidence with:
We have so a direct verification of the main part of our theorem. The rest can
also be verified with similar methods to those used in [ l l ] .
310 SERGE PERRINE
References
[I] J.W.S. Cassels, A n introduction to dzophantzne approximation, Cambridge tracts in Math.
Physics 45 (1957).
[2] H. Cohn, Representatzon of Markoff's binary quadratzc forms by geodesics on a perforated
torus, Acta Arithmetica XVIII (1971), 123-136.
[3] T.W. Cusick and M.E. Flahive, The Markoff and Lagrange Spectra, Mathematical surveys
and monographs, 30, A.M.S., 1989.
[4] T . W . Cusick, O n Per-rine's generalized Markoff equation, Aequationes Mathematicae 3 (1993),
203-211.
[5] L.E. Dickson, History of the Theory of Numbers, Chelsea reprints, New York, 1992.
[6] El Khuti El'Huti (M.Kh.), Cubic surfaces of Markoff type, Mat. Sb. 93 (1974), 331-336.
[7] C . J . Hightower, T h e mznzma of indefinzte binary quadratic forms, J . Number Theory 2 (1970),
364-378.
[8] A.A. Markoff, Sur les formes quadratzques inde'finies, Math. Ann. 6 ; 17 (1879; 1880), 381-406;
379 399.
[9] S. Perrine, SUT une ge'ne'ralisation de la the'orze de Markoff, J. Number Theory 37(2) (1991),
211-230.
[lo] S. Perrine, Sur des equations diophantzennes ge'ne'ralisant celle de Markoff, Annales de la
Facult4 des Sciences de Toulouse VI(1) (1997), 127-141.
[Ill S. Perrine, Un arbre de constantes d'approximation analogue a celui de l'e'quation diophanti-
enne de Markoff, submitted.
Igor Rodnianski
To Jerry Lange
In the case of a particle of mass m on the circle T = R/Z in the electric field of
the potential V(x), the Hamiltonian H can be represented as follows:
1 d"
H =
2m dx2
+ V (x), x E T.
1991 Mathematics Subject Classif?cation. Primary 81Q05, llL03, 35D10; Secondary llL07,
46335.
Key words and phrases. Schrodinger evolution, Besov spaces, continued fractions.
Research was partially supported by NSF grant DMS-9623520
311
312 IGOR RODNIANSKI
The kernel of the propagator e P t H " i sthe solution of equation (1.3) with V(x)
and the initial data
- 0
Note that when t is in the upper half-plane, then Eo(t,x , y) is (essentially) Jacobi's
n
The solution operator e P t H 0 is a unitary operator in the scale of Sobolev spaces
Hs := H"(T). Recall that function f = xnEZ
fne2xznxE H S if and only if
Clearly, for each y, E ( t , . , y) E n Hpa-' for all t 2 0 and does not belong t o a
c>O
better Sobolev space . This shows that the series in (1.5) does not define a smooth
function.
The question of regularity for E ( t ,x , y) fits in the general framework of the
study of the smoothness of the fundamental solution for a Schrijdinger evolution.
The Schrodinger equation on Rn has been extensively studied. In that case,
in general, one can expect that the fundamental solution becomes infinitely smooth
for positive times. In fact, this has been proved for a wide class of Hamiltonians
H with potentials V growing slower than quadratically at infinity, see [Z,1983],
[KR,1996], [Y,1996]. Moreover, a similar result has been proved for the more
general class of equations with variable coefficients. The smoothing of the fun-
damental solution occurs provided bicharacteristics of the corresponding classical
Hamiltonian are not trapped, see [CKS,1995], [KS,1996]. The situation of the
Schrodinger equation on more general than Rn non-compact manifolds has been
addressed in [W,1998].
In the case when the underlying space is compact, one can no longer expect the
fundamental solution t o become smoother for positive times. This is supported by
the results of K. Yajima [Y,1996], who showed that in the case of the Schrodinger
equation on an interval with Dirichlet boundary conditions, the fundamental solu-
tion is nowhere smooth, and by the work of S. Doi [Do,1996].
In the paper [KR,1998] we proposed t o address the question of regularity of the
fundamental solution E o ( t , x , y) for the Schrodinger evolution on a circle from the
point of view of Besov spaces B;,, rather than Sobolev spaces. We discovered
that although the fundamental solution is not smooth, its regularity is changing
with time. Regularity at the moment in time t is related t o arithmetic properties
of t. For all rational times E o ( t , ., , y) E B,:,, when for a generic irrational time
t , E o ( t ,., y) is almost " ;-derivative better". It belongs to the space n B;!;'. In
r>O
general, the worse t is approximated by the rationals, the better the regularity of
the fundamental solution at time t.
The Besov spaces' approach requires investigation of the partial sums (incom-
plete theta-functions)
For every N and fixed t and x , (1.6) determines a point on the complex plane.
Connecting the points with consecutive N by straight lines we obtain interest-
ing geometric figures, curlicues in the terminology of M. Berry and J. Goldberg
[BG,1988]. Their pattern depends upon the arithmetic properties of the parame-
ter t, in particular, on the closeness to the set of rational numbers. Similar sums
appear in the diffraction experiment with planar waves incident to a periodic grat-
ing with N narrow slits. The intensity of light on the screen behind the grating
is determined by the square of the absolute value of the sum in (1.6), where the
role of parameter t is played by the distance t o the screen. Therefore, the image
on the screen depends upon arithmetic properties of the distance. M. Berry and J.
Goldberg suggested to call this experiment the arithmetic microscope [BG,1988].
The studies of the geometric patterns generated by theta-sums are also contained
in [DM1981], [D,1985], [BH,1987], [CK,1987].
In this paper we address the question of regularity of the fundamental solution
for the Schrodinger evolution on a circle in the case of a non-zero complex-valued
potential V(x). We show that the picture observed in the case of a free particle
on a circle is completely stable for a wide class of potentials. On the compact
manifold, a smooth bounded potential V(x) constitutes a weak perturbation of
the Hamiltonian Ho. Therefore, one might expect that qualitatively the effects
observed in the case of the free particle will remain the same for this class of
potentials. The surprising element is that the behavior of the fundamental solution
is exactly the same as in the case of the Hamiltonian Ho. This stability suggests
314 IGOR RODNIANSKI
that should any physical experiment relating the smoothness properties of the wave
function at time t and arithmetic properties of t be carried out, one could neglect
the contribution of the electric field.
The effect can be best demonstrated in the special case of a real-valued potential
V. In that case we have the following representation for the fundamental solution:
where An are the eigenvalues of the operator - &&+ V ( z ) , and 4, are its eigen-
functions. Treating the potential V as a perturbation, we can rewrite E ( t , x, y)
+
in the form Eo(t,x, y) error and show that the error term defines a function
smoother than Eo(t, x, y) .
We should mention that K. Oskolkov investigated the pointwise behavior of the
solutions of periodic Schrodinger equations, see [0,1992] and references therein.
In particular, he investigated the series
and
m
We define the Besov spaces B;,, on T for the following values of parameters s,
p, and q: - m < s < + m , 1 5 p, q <
+ m . By definition, the space B;,, is
CONTINUED FRACTIONS AND SCHRODINGER EVOLUTION 315
composed of all distributions f ( x ) = Emf , e2"imx such that the following norm
is finite:
In this paper we use the spaces B& most frequently. Therefore, we make a
simpler notation for them: B&.
We shall say that a distribution f belongs to B& sharp if f E B& and
f ?? ut>o B Z .
The fundamental solution E ( t ,x , y) of the Schrodinger equation on the circle
T with the potential V is a solution of the following initial value problem:
1
-
an
Note, that the expansion is not unique: we also have = [ao,a l , . . . , an- 1 , a,- 1,1],
+
if a, # 1, and 24 = [a0,a l , . . . , an-1 11, if a n = 1.
For an irrational t the continued fraction expansion
a k + ~= [-I,t1
k
1
tlci-1 = - - ak+l,
tk
and the initial conditions
ao = [t], to = t - a0 .
For the basic theory of continued fractions, see [K,1964] and [S,1980].
We state our main result.
. Let E ( t , x) be a solution of equation (2.4) with the initial con-
T H E O R EIM
dition
where F ( t , .) E n B;ct
t>O
for any fixed t .
REMARK 2.1. The definition of Besov spaces (2.3) implies that for all t func-
1
tion E o ( t , .) belongs to the space B2:z sharp. Because of the embedding B G L
1
1
B i z , we can conclude that, for all t , E o ( t , .) does not belong to a space better
-
than B,'. Therefore, Theorem I implies that E o ( t , .) completely determines the
regularity of E ( t , .).
We can simply reiterate the known regularity properties of Eo(t , .). We use the
results obtained in [KR,1998]. Note that E o ( t ,x) is E ( t , x) in the notations of
[KR,1998]. In the case of a rational time we have the following theorem (compare
[KR,1998, Theorem I]).
THEOREM 11. Let t be a rational number, and t = - its simple fraction
representation. Let E, k = 1, . . . , n - 1, be partial convergents to t determined by
a finite continued fraction [ao,a l l .. . , a n ] with a n odd number of quotients (i.e., n
is even). Then,
I ) (formula)
where
E=~.q(mod2), T = P , - I . ~ , - I ( ~ o ~ ~ ) ,
and x o ( t ) is a n eighth root of 1;
2) (regularity)
E ( t , .) E B & ~ sharp.
Denote by J<, the set of all irrational t such that there exists a constant ct such
that for an infinite number of n
THEOREM
111.
(i) For almost all irrational t , E ( t , .) E n B,""'
t>O
(ii) If t is an irrational number with bounded quotients, i.e., there is a constant
C > 0 such that a,, < C , for all n, then E ( t , . ) E B,"~ sharp.
(iii) If t E J?, , then
3. Proof of Theorem I.
Consider the Schrijdinger equation on a circle
It is easy to see that e i A t ~ Ois a solution of equation (3.1) with the constant
potential 7rVo and the initial data e2"inx . Therefore, as in the case of a free
equation, it can be represented by the series
CONTINUED FRACTIONS AND SCHRODINGER EVOLUTION 319
where E o ( t ,x ) is a solution of equation (3.1) with the potential equal to zero. The
right-hand side of (3.6) coincides with the first term of identity (2.8) in Theorem I.
Therefore, we need t o show that for all t , function
Apply operator Pk to both sides of (3.5). Using (3.8) and (3.9) we obtain
(3.10)
t
Obviously,
Therefore.
320 IGOR RODNIANSKI
1 J - + - - e - T L ( " + ~ ~ l ) ( t)P~VP[V(I
-~) .) d r .
- P/)E(T,
7ri(k2 - 12)
0
PhV P , = Vk-lP1 r
Sinlple computation shows that the first two terms can be bound by c
last term on the right-hand side of (3.14) can be estimated by
w. The
1 1 c
Therefore,
for some E > 0. The fundamental solution E ( T , .) belongs to the space n H - $ - ' .
e>u
Also, by the assumption of the theorem, V E U H++'. Therefore, with the help
t>O
of the inequality
Combining estimates (3.15) and (3.19) we conclude that for any E > 0 and t E R
there exists a constant c ( t , E ) such that
The bound for the first term on t,he right-hand side of (3.22) has been found in
(3.15). For the second term. we have similar to (3.16)
Following the estimates leading to (3.18) but taking into account, (3.21) we have
and
Also.
CONTINUED FRACTIONS AND SCHRODINGER EVOLUTION 323
References
J. Bergh and J . Lofstrom, Interpolation spaces, an zntroductzon, Springer, Berlin,
1976.
hl.V. Berry and 3. Goldberg, Renormalzzatzon of curlzcues, Nonlinearity 1 (1988),
1-26.
E. A. Coutsias and N. D. Kazarinoff, Disorder, renorma~zzabi~zty, theta functzons
and Cornu spirals, Physica D 26 (1987), 295-310.
W. Craig, T . Kappeler and W. Strauss, Microlocal dzsperszve smoothing for the
Schrodinger equation, Comm. Pure Appl. Math. 48 (1995); 769-860.
F. M. Dekking and M. Mend& France, Un,zform dzstrzbution modulo one: a geo-
metric view point, J. Reine Angew. Math. 329 (1981), 143-153.
3.-M. Deshouillers, Geometric aspect of Weyl sums, Elementary and analytic theory
of numbers, Banach Center publications, volume 17, PWN-Polish Scientific Publish-
ers, Warsaw, 1985, pp. 75-82.
[Do,1996] S. Doi, Smoothzng effects of Schrodinger evolution groups o n Rzeman,nzan manzfolds,
Duke Math. J . 82 (1996), no. 3, 679-706.
[HB,1980] J . H. Hannay and M. V. Berry, Quantzzation of h e a r maps o n a torus Fresnel
dzffraction by a perzodzc gratzng, Physica D 1 (1980); 267-290.
[KR,1996] L. Kapitanski and I. Rodnianski, Regulated smoothzng for Schrodznger evolutzorz,
International Math. Research Notices 2 (1996), 41-54.
[KR,19981 L. Kapitanski and I. Rodnianski, Does a quantum particle know the tzme?. t o appear
in Proceedings of the workshop on Emerging Applications of Nunlher Theory. ILIA
Volumes in Mathematics and its Applications, 1998.
L. Kapitanski and Yu. Safarov, Disperszue smoothzng for Schriidznger equatzons,
Math. Res. Letters 2 (1996), 41-54.
A. Ya. Khinchin, Contznued fractzons, Translated from the 3rd Russian edition of
1961, The University of Chicago Press, 1964.
P. I. Lizorkin, On bases and multipliers for the spaces BF,o(II),Trudy Mat. Inst.
Steklov 143 (1977), 88-104.
I(.I. Oskolkov. A class of I. M. I/zn,ogradov's series and its applications zn harmonic
analysis, Springer Series in Computational Mathematics, 19, Progress in Approxi-
mation Theory, a n International perspective (A. A. Gonchar and E. B . Staff, eds.),
Springer-Verlag, New York, 1992.
Schmidt, Diophantzne approximatzon, Lecture Notes in Mathematics, 785, Springer-
Vcrlag, Berlin Heidelberg New York, 1980.
H. Triebel, Theory of functzon spaces, Birkhauser, Basel, 1983.
J . Wunsch, Mzcrolocal analysis of the time-dependent Schrodznger operator, Thesis,
Harvard University (1998).
K. Yajima, Smoothness and non-smoothness of the fundamental solutzon of tzme
dependent Schrodinger equations, Comm. Math. Phys. (1996) (to appear).
S.Zelditch, Reconstructzon of singu~aritiesfor solutaons of Schrodinger's equatzon,
Comm. Math. Phys. 90 (1983), 1-26.
1. Orthogonal polynomials
The notion of orthogonal polynomials is an old one. going back to the previ-
ous century (Chebyshev, Stieltjes). A very good source of information is Szego's
book [Sz] and a more recent exposition can be found in [VAl]. hlonic ortilogonal
polynomials for a positive Bore1 measure p on the real line are polynomials P,,
(n = 0 , 1 , . . . ) of degree n and leading coefficient one such that
These n orthogonality conditions given n linear equations for the n unknown co-
efficients ah ,, ( k = 0 , 1 , . . . , n - 1) of the rnonic polynonlial p , , ( ~
=) a k ,,z I, x:=,,
(where a,, ,= 1). This system of n equations for n unknowns always has a unique
solution since the matrix with entries xl'J d p ( x ) (0 5 z. J <
n - 1) (this is known
as the Gram matrix) is positive definite and hence nonsingular. It is well known
that such polynomials satisfy a three-term recurrence relation
1991 Mathematzcs Subject Classzficatzon. Primary 42C05. 33C45; Secondary 11372, 11382.
Research Director of the Belgian National Fund for Scientific Research (FWO). This research
is supported by FWO research project G.0278.97 and INTAS 93-219ext.
@ 1999 A r n r r l r a n h l a t h e m a t i c a l Society
326 WALTER VAN ASSCHE
so t>hatyic = (I
P:(x) d p ( ~ ) ) - " >
~ 0. The recurrence relation for these orthonor-
nlal polynomials is given by
where
from which we conclude that every zero xJ ,, of pn is an eigenvalue of J,, with eigen-
vector (po(x, ,,), pl (J, ,,), . . . ,p , - ~ ( x , , , ) ) I . Since all zeros of orthogonal polynomi-
als are simple, it follows that the eigenvalues of J, are the same as the zeros of p,, .
One can also consider the infinite matrix J , which acts as an operator Jx : t 2 4 t 2
on an appropriate domain. The symmetric matrix J , has selfadjoint extensions
to its maximal domain (sometimes several) and the study of such a selfadjoint op-
erator (especially the spectral theory of such operators) gives useful insight in the
corresponding orthogonal polynomials.
Let us mention at this point the best known examples: the very classical or-
thogonal polynoinials of Jacobi, Laguerre and Hermite These three families of
orthogonal polynomials satisfy a second order linear differential equation, their
derivatives are again orthogonal polynoniials of the same family (but with different
MULTIPLE ORTHOGONAL POLYNOMIALS 327
parameters) and they can all be obtained from a Rodrigues formula. These prop-
erties can be extended to difference operators and q-difference operators and give
the classical orthogonal polynomials in the Askey table [KS].
1.1. Jacobi polynomials. The monic Jacobi polynomials P:~") satisfy the
orthogonality relations
where a , @ > -1. There are two important differential operators, namely the
lowering operator D for which
Both (1.8) and (1.9) can be proved using integration by parts and the orthogonalitmy
(1.7). Combining the raising and lowering operators gives the differential equation
where y(x) = P?")(X), and repeated use of the raising operator DaX3gives the
Rodrigues formula
(1.13)
where cr:
I" x 0,
~ ~ ( x ) x " e -dx~ = ~ k = 0, 1 , 2 , .. . , n - 1,
Both (1.14) and (1.15) can easily be proved using integration by parts and the
orthogonality relation (1.13). Combining the lowering and the raising operator
gives the differential equation
from which one can obtain an explicit expression using Leibnitz' formula
The historical normalization for the Laguerre polynomials is to take LE(O) = )I(:"(
so that L : ( x ) = ( - l ) ' " n ! ~ : ( x ) . See Szeg6 [Sz, Chapter V] for more properties of
Laguerre polynomials.
1.3. Hermite polynomials. The monic Hermite polynonlials H,, satisfy the
orthogonality relations
Combining the lowering and the raising operator gives the differential equation
(1.22) y"(x) - 2 x y 1 ( x )= - 2 n y ( x ) ,
where y ( z ) = ~ , ( r )Repeated
. application of the raising operator gives t,he Ro-
drigues formula
and by using Faa di Bruno's formula for the nth derivative of a composite function
we obtain
Historically the Hermite polynomials are normalized to have leading coefficient 2",
hence H,,(z) = 2'"1, ( x ) . For more properties we refer to Szegij [Sz, Chapter 51.
MULTIPLE ORTHOGONAL POLYNOMIALS 329
For r = 1 we get the ordinary monic orthogonal polynomials, but for r > 1 we get
multiple orthogonal polynonlials (or polyorthogonal polynornials). All together the
orthogonality conditions give 161 linear equations for the 161 unknown coefficients
EL=,,
ah 5 of the polynomial PC(x) = 161 ah ,-xh, where ali; i; = 1. However, the matrix
of coefficients of this system can be singular and we need some extra conditions on
the r measures p l , p2, . . . , pr in order that the multiple orthogonal polynomial is
unique. If the polynomial Pr?(x)is unique, then we say that n' is a norrnal index
and if all indices are normal then we have a complete system. Our interest is of
course in systems of r measures for which all multi-indices are normal.
Again for r = 1 we find the ordinary orthogonal polynomials. but for r > 1 we
find multiple orthogonal polynomials. There are in'l - 1 orthogonality conditions
which give 161 - 1 homogenous linear equations for the 1 61 unknown coefficients
,
of all the polynomials Ar? ( 3 = 1 , 2 , . . . , r ) . Hence we can determine the vector
(A< 1 , Aii 2 , . . . , A< ,) up to a multiplicative factor.
2.3. Operator approach. The t hree-term recurrence relation (1.2) and the
Jacobi rnatrix (1.6) have an interesting extension for multiple orthogonal polyno-
mials, which was subsequently exploited by Kalyagin [KZ] to initiate the spectral
330 WALTER VAN ASSCHE
These indices are nearly diagonal and if all these indices are normal, then we have
a weakly complete system. Now consider the monic type I1 multiple orthogonal
polynomials with these indices, i.e., we put Q,(x) = P,-(,)(x), then one can use
all the orthogonality relations (2.1)-(2.3) to show that they satisfy a recurrence
relation of the form
Again it turns out that the eigenvalues of H , coincide with the zeros of the multiple
orthogonal polynomial Q, = P,-(,,). The matrix H , is not symmetric, thus there is
no a priori reason why the eigenvalues should be real. Nevertheless it turns out that
+
in many cases the zeros are indeed real. Hence the coefficients a, (n i) (n E W),
with 0 5 i < r are not arbitrary and have some hidden structure. An interesting
open problem is to find this hidden structure and to find out under which conditions
on these coefficients the matrix H,, corresponds to n~ultipleorthogonal polynomials.
3. Hermite-Pad6 approximation
3.1. Simultaneous Pad6 approximation. Multiple orthogonal polynomials
are closely related to simultaneous Padk approximation, which is nowadays known
as Hermite-Padk approximation. Suppose we are given r Markov functions (or
Stieltjes functions)
MULTIPLE ORTHOGONAL POLYNOMIALS 331
and
Combining (3.1) with ( 3 . 5 ) gives a nice formula for the remainder in type I Hermite-
Pad6 approximation in terms of the multiple orthogonal polynomials
and combining (3.2)-(3.4) with (3.6) gives a formula for the remainder in type I1
Hermite-Pad6 approxirnation
332 WALTER VAN ASSCHE
Hence for Hermite-Pad4 approximation one only needs to obtain the multiple or-
thogonal polynomials, from which all other quantities can be deduced.
3.2. Useful s y s t e m s of functions. In order to get a system for which most
multi-indices are normal, one needs to pose additional assumptions on the hlarkov
functions ( f l , fi. . . . , f , ) or on the measures ( p l . p2, . . . , y,). Two kinds of systems
turn out to be quite useful.
DEFINITION 3.1 (see [ A n ] ) . An Angelesco s y s t e m consists of r measures on r
pairwise disjoint intervals:
Angelesco systems are always complete systems and are the prototype for study-
ing multiple orthogorlal polynomials on r pairwise disjoint intervals. It turns out
that the zeros of type I1 multiple orthogonal polynomials have a very simple 10-
cation property for Angelesco systems: there will be n, zeros of P,- on the in-
terval A, for z = 1 , 2 . . . . , r . The polyrlornial PITcall therefore be factored as
P,-(x) = P,,, (x)PI,, (x) . . . P,,,(x) and each P,,, is an ordinary orthogonal polyno-
mial on A, for the varying measure n?+
PI,, (x) dp,(x).
For multiple orthogonal polynomials on one interval there is another useful
system of functions.
DEFINITION3.2 (see [ N l ] ) . An AT s y s t e m consists of r measures on the same
interval A such that each measure p, is absolutely continuous with respect to
p = p1 and such that the Radon-Nikodym derivatives u, = dpJ/dp (with u1 = 1)
are such that (1,s... . , x n 1 ' . u 2 2, ~ 2 .... , x7'2-' u2.. . . , u, , x u , , . . . , x ' ~ , - ~ u , is
) a
Chebyshev system 011 A for every multi-index ( n l , n2, . . . , n,).
Recall that a set of n linearly independent functions { 4 1 , 4 ~ ., . ,4,,) is a
Chebyshev system on A if every linear combination C;=, akQr,has at most n - 1
zeros on A. For an AT system it thus follows that P,,,-1(x) PI,,-1 ( x ) u ~ ( z ) + +
. . . + P,,, (z)u, ( x ) has at most (61
- 1 zeros on A whenever Pr, is a polynomial
of systems of hlarkov functions on one interval and allow a rather detailed analysis.
See. e.g.. Driver and Stahl [ D S l ] [DSZ] for normality of Nikishin systems and
MULTIPLE ORTHOGONAL POLYNOMIALS 333
Bustamante and L6pez [BL]and Aptekarev [A31 for asymptotic results. Driver
and Stahl [ D S l ]have shown that the type I1 multiple orthogonal polynomial Pi;
for a Nikishin system of order r on A, has (61simple zeros on this interval A,
whenever n, 2 max(n,+l,. . . , n,) - 1 for every j = 1 , 2 , .. . , r - 1. In particular
this is true for the diagonal indices s'(n)given in (2.5).
Recently Gonchar, Rakhlnanov and Sorokin [ G R S ]have considered a mixture
of Angelesco and Nikishin systems and have analyzed the asymptotic behavior of
the corresponding Hermite-Padit approximants. The asymptotic behavior of the
Hermite-Padit approximants for Angelesco and Nikishin systems was obtained in
[GR] and [ S 5 ] . One of the important results is the asymptotic distribution of the
zeros of the polynomials. For ordinary orthogonal polynomials the distribution
of zeros is usually obtained by investigating an extremal problem for logarithmic
potentials. For multiple orthogonal polynomials one needs to study an extremal
problem in logarithmic potential theory for vector potentials.
where y(z) = pkI3"(2). Observe that the coefficients of y"' and y" are polyno-
mials independent of n but that the coefficient of y' depends on n, so that Sturm-
Liouville theory of linear differential equations can not be used here. If we write
PJir= PAC:, '
') and P2,1+l= pi';:,?, then we have a third order recurrence relation
334 WALTER VAN ASSCHE
for which
a+l 2 a+l 2
lim bZn =
n-cc
-
9
+ p, lim b2n+l =
71-30
-
9
+ -x2,
3
4
lim c,, = -(a2 - a + l),
n-x 81
4 4
lim d z , = - - x l ( x 1 -a ) ( x-
~ l), lim d2n+l = - - x 2 ( x 2 - a )(x2 - 1 )
71-x 27 n-x 27
where x l and 2 2 are the roots of the quadratic equation h l ( x ) = 0 , with h ( x ) =
x ( x - a ) ( x - 1).
where the order in the product of r differential operators can be changed. For
(ao1W.~2)
r = 2 we put PZn= ~ i f f 2 . ~and ~ ~ ) = P7L+l.n and we have the third
~ . P2n+l
order recurrence relation
i m dn
n-x
= ($) :I
MULTIPLE ORTHOGONAL POLYNOMIALS 335
(4.7) dpj(x)=xU~e-xdx, j = 1 , 2 , . . . ,r ,
with a j > -1 for j = 1 , 2 , . . . , r and a , - a, @ Z for i # j . These polynomials are
a limiting case of the Jacobi-Pineiro polynomials:
in a similar way as the Laguerre polynomials are a limiting case of the Jacobi
polynomials. There are again r raising operators
and again the product of the differential operators can be taken in any order.
336 WALTER VAN ASSCHE
where a > -1 and p,(x) = x u / 2 ~ u ( 2 & )(v 2 O), with K u the Macdonald func-
tion (modified Bessel function of the second kind). This AT system has a lot of
properties that very classical polynomials usually have. The differential operator
D is a lowering operator for the type I1 multiple orthogonal polynomials near the
diagonal
and D, = x-"Dx" is a raising operator for the type I multiple orthogonal polyno-
mials: putting QE ,,,, = A ~ , L , l L )+, lA~Tu L . n ) , 2we
TL,, ~ u have
+~
Using the raising operators repeatedly gives a Rodrigues formula for the type I
multiple orthogonal polynomials
Putting PL', = P& and P2rL+l= P:+l gives the recurrence relation
with
b, = ( n + a + 1)(3n+a+2v) - ( a + l ) ( v - I),
c,, = n(n+a)(n+cr+v)(3n+2a+v),
d,, = n(n- l ) ( n + a - l)(n+a)(n+cr+v-l)(n+a+v)
MULTIPLE ORTHOGONAL POLYNOMIALS 337
5. I r r a t i o n a l i t y
Multiple orthogonal polynomials turn out to be quite useful in proving irra-
tionality of certain real numbers. The key to such irrationality proofs is the follow-
ing lemma:
LEMMA5.1. Let x be a real number. Suppose there exists integers p, and q,
(n E N) such that
1. q,x - pn # 0 for all n E N,
2. lim,,,(q,x -p,) = 0.
Then x is irrational.
PROOF.Suppose x is rational so that x = plq with integers p and q. Then
By assumption (1) we know that q,p - pnq is an integer different from zero, hence
Iq,p - p,ql 1. But then
dx.
j!
We can use that information and multiple orthogonal polynomials for the measures
dpj (x) = (-l)Jlogj(x) dx on [O,1] for j = 0 , 1 , 2 to prove the irrationality of C(3).
When Apkry [Ap] first announced the proof of irrationality of ((3), the proof was
considered as rather mysterious. Later Beukers [Bl][B2]showed how Ap6ry's
approximants can be obtained using Hermite-Pad6 approximation.
5.2 (Ap6ry). C(3) is irrational.
THEOREM
PROOF. In order to obtain good rational approximants to <(3) we will use
Hermite-Pad6 approximation to the three Markov functions
Z-x
logx
dx, f3(z) =
1' 2
Z - x dx,
338 WALTER VAN ASSCHE
and use that f 3 ( 1 ) = ( ( 3 ) . We want to find a vector of polynomials (A,, , B,), where
A , and B , are polynomials of degree n , and polynomials C, and D, such that
Observe that (5.2) is a type I approximation problem for the system ( f l , f 2 ) and
(5.3) is a type I approximation problem for ( f 2 , f 3 ) , whereas the combination (5.2)-
(5.3) is a vector type I1 approximation problem with common vector denominator
( A , , B,,). The solution of this problem in Hermite-Pad6 approximation is given by
the generalized polynomial A, ( x )-Bn (x) log x for which the following orthogonality
relations hold:
1
(5.4) (A(x)-B,,(x)logx)xkdx = 0, k = 0 , 1 , . . . ,a - 1
(5.5) l(A,(x)-~,(x)logx)log(x)xkdr = 0, k = 0 , 1 , . . . ,n - 1 ,
0
together with the extra condition that A,(1) = 0. Let P, be the Legendre polyno-
mials of degree n on [ 0 , 1 ] ,
is of the form A,,(x) - B,(x) logx with A,(1) = 0. Taking the Mellin transform
gives
J
/'o Frl(x)x.'logx dx =2 1 1
P n ( x ) x sdx 1 1
P,(x)xs log x d x ,
and for s = 0 , 1 , . . . , n - 1 this gives the orthogonality conditions (5.5). The re-
mainder in (5.3) is
- 1 1
An ( x ) - Bn ( x )log x
2 - 2
log x d x ,
gives
and
1-x
Bn(x) log2 dx
~ f - Hj2)
)
k=O 1 = 0
j-k '
j#k
where ~ f = ) l/j2. Clearly B,(1) is a positive integer, but D,, (1) contains
rational terms. However, if we multiply D n ( l ) by a common multiple of the nurnbers
2 9, 3 ", . . . , n%hen we would get an integer. Taking d, = lcm(1,2,3,.. . , n ) (least
common multiple) then
6. Transcendence
Historically, Hermite introduced the notion of Hermite-Pad6 approximation to
prove the transcendence of e. The basis lemma for proving transcendence is a
generalization of Lemma 5.1, which goes as follows:
N
LEMMA6.1. Let x be a real number. Suppose that for every posztzve znteger m E
.
and for all zntegers ao, a l , . . . , a, E Z we can find zntegers po ,, , p l , . . . ,p,.,
( n E N) such that
1. C;koakpk ,, # 0 for all n E W.
340 WALTER VAN ASSCHE
2. limn,,(po.nxk -~ ~ 0)
k . = for k = 1 , 2 , .. . , m .
Then x is transcendent.
PROOF.Suppose that x is algebraic, then there exists a positive integer m and
integers ao, a l , . . . , a m such that CF=o a k x k = 0. By adding and subtracting we
find
We know by assumption (1) that the integer CTxO akpk., is different from zero,
hence in absolute value it is at least 1. As a consequence
I m I
References
[An] M. A. Angelesco, S u r d e u x e x t e n s i o n s des fractions c o n t i n u e s alge'brzques, C.R. Acad. Sci.
Paris 18 (1919), 262-263.
[Ap] R. Apkry, Irrationalit6 d e C(2) e t C(3), Ast6risque 61 (1979), 11-13.
[All A. I. Aptekarev, A s y m p t o t i c s of simultaneously orthogonal polynomials in t h e Angelesco
case, Mat. Sb. 136 (1988), 56-84 (in Russian); Math. USSR Sb. 64 (1989), 57-84.
[A21 A. I. Aptekarev, Multiple orthogonal polynomials, J . Comput. Appl. Math. 99 (1998),
423-447.
[A31 A. I. Aptekarev, S t r o n g a s y m p t o t i c s of multiple orthogonal polynomials for Nzkishin sys-
t e m s , Mat. Sb. (to appear).
[AMR] A. I. Aptekarev, F. Marcellin, I. A. Rocha, Semiclassical multzple orthogonal polynomials
a n d t h e properties of Jacobi-Bessel polynomials, J . Approx. Theory 90 (1997), 117-146.
[AS] A. I. Aptekarev, H. Stahl, A s y m p t o t i c s of H e r m i t e - P a d 6 polynomials, in 'Progress in
Approximation Theory' (A. A. Gonchar, E.B. Saff, eds.), Springer-Verlag, Berlin, 1992,
pp. 127-167.
[Bl] F. Beukers, A n o t e o n t h e irrationalzty of ((2) a n d C(3), Bull. London Math. Soc. 11
(1979), 268-272.
[BZ] I?. Beukers, P a d 6 a p p r o x i m a n t i o n s in n u m b e r theory, in 'Pad6 Approximation and its
Applications', Lecture Notes in Mathematics 888,Springer-Verlag, Berlin, 1981, pp. 90-
99.
[BE] P. Borwein, T. Erdblyi, P o l y n o m i a l s a n d P o l y n o m i a l Inequalities, Springer-Verlag, Berlin,
1995.
[dB] M. G . de Bruin, S i m u l t a n e o u s P a d 6 a p p r o x i m a n t s a n d orthogonality, in 'Polyn8mes Or-
thogonaux et Applications' (C. Brezinski et al., eds.), Lecture Notes in Mathematics 1171,
Springer-Verlag, 1985, pp. 74-83.
MULTIPLE ORTHOGONAL POLYNOMIALS 341
1. Introduction
We remark on general properties of simple continued fraction expansions of formal
power series (hereafter omitting the qualifier 'simple'). We explain various basic
properties of the expansions. For example, a little experimentation with 'random'
formal power series, say with coefficients 0 , 1, and -1 only, suggests that - unless
there are large gaps of zero coefficients - all, certainly almost all, their partial
quotients are of degree one, and that the coefficients of those partial quotients
increase in complexity at an exponential rate. It follows that,, generically, the
convergents t o formal power series are rational functions with coefficients growing
at a rate compounding that of the partial quotients. A consequence is that those
convergents are not helpful in the study of diophantine properties of values taken by
a series in its domain of convergence. Of course there are atypical cases; we mention
some classes of examples. Indeed, we briefly suggest which cases are 'unexpectedly'
atypical and thus interesting.
An evident consequence of the explosive growth of the numerical coefficients
is that computer programs computing such expansions soon choke - they slow
down severely. One might choose to avoid that by studying the continued fraction
expansions of the series reduced modulo p , say for some collection of primes p , thus
'confining' and taming the coefficients of the partial quotients. We formulate and
prove a 'Reduction Principle' relating the reduction mod p of a continued fraction
Only the partial quotients matter, so such a continued fraction expansion may be
.
conveniently detailed by [ao a, , a2 , a3 , . . . ] = [a(), a1 , a2 , . . . , a,, , F h + l ] .
The truncations [ a " , a l , . . . , ah ] are rational functions ph/qh. Here. the pairs
of relatively prime polynomials p h ( X ) , q h ( X ) are given by the matrix identities
and the remark that the empty matrix product is the identity matrix. This alleged
correspondence, whereby these matrix products provide the convergents p h / q h ,
may be confirmed by induction on the number of matrices on noticing the definition
1.2. Normality. It's actually quite easy t o see why, for formal power series
over an infinite field, all partial quotients are almost always of degree one. We need
only notice that a remainder C h > a h x P h has a reciprocal with polynomial part
of degree greater than one, and thus gives rise to a partial quotient of degree greater
than one, if and only if a1 = 0. Moreover, the partial quotient is a r l x - a2a,' if
a1 # 0 , and the next remainder is (a; -ala3)a,"~-' +terms of lower degree in X .
A little more precisely, one observes that the sequential remainders have lead-
ing coefficient some determinant, and hence some multivariate polynomial, in the
coefficients of the formal power series. But such a polynomial is 'nonzero "almost
always", or "with probability one" ' - [9], p.375. The matter is considered in ex-
tenso by Knuth [9] in the context of his discussion of the Euclidean algorithm for
polynomials over a field. Moreover, that discussion also shows that the 'explosive
growth' of the coefficients of the partial quotients is precisely the better known
extraordinary growth of the coefficients in performing the Euclidean algorithm on
a pair of polynomials over the integers Z.
Over the finite field F, , on the other hand, the nonvanishing of the coefficient
of X-I of all remainders is as unlikely as is the nonappearance of the digit 0 in
the base p expansion of a random real number. Thus we should expect that the
reduction mod p of a formal power series almost always has partial quotients of
346 ALFRED J. VAN DER POORTEN
degree greater than one and that almost always the original sequence of partial
quotients nowhere has good reduction.
Power series with substantial gaps are of course not generic. To assist we
mention
PROPOSITION
1.3 (Folding Lemma [ll]).W e have
>
For k 3 the truncations Gk.+( X ;n) of the infinite product Gk,*( X ) are readily
seen to be convergents of Gk,*. It also turns out to be easy to see that in the cases
k , +, with k even not 2 , they are every second convergent of G k . + .It follows that
every ~ a r t i a quotient
l of the G2k.+ has integer coefficients.
REDUCTION OF CONTINUED FRACTIONS OF FORMAL POWER SERIES 347
All other partial quotients are of degree one, except in the case 2, - .
In summary, the exceptional cases among the Gk.+ are the Gzk,+, in having all
their partial quotients specialisable, that is, with integer coefficients and thus with
good reduction at every prime - compare the G2k,- for k 2 2 ; and the example
G 2 , , in having a partial quotient of degree more than one. All other cases that
have partial quotients of degree other than one can be credited to 'blatant' gaps in
the power series.
G2,- ( X ) = n
h20
(1 - x - 2 " ) =
Contrary to general forecast, there is a partial quotient of degree other than one;
it is generated, in effect, by a numerical accident. All other partial quotients are
indeed of degree one, other than the images of the said partial quotient of degree 2
under the functional equation. Of course the detail reported above has no particular
meaning other than as vivid illustration of the explosive growth of the coefficients
of the partial quotients.
For example, over IF3 we find that
As suggested, to the extent that the original expansion has good reduction the
new expansion is its reduction; whilst the first term with bad reduction 'collapses'
t o a term of higher degree. Beyond that term, the expansion is not immediately
recognisable in terms of the original. Notice that for this example the functional
equation entails that once the 'collapse' yields a partial quotient of high degree
then partial quotients of increasingly high degree are propagated throughout the
expansion.
2.2. Reduction Principle. We will want to use p t o denote a prime; so our
convergents will here be x l y . As above, given a series F , we denote its sequence
of partial quotients by ( a h ) ,and of its complete quotients by ( F h ) .
The idea is to return to a first principles genesis of the sequence of polynomials
(yh) yielding the convergents to F . We know that the convergents xh/yh are
characterised by the property
The last inequality informs us that the corresponding next partial quotient of F,
let's call it bj+l , has degree at least deg yh(,)+, - deg Yh(J). But
where we recall yh(,) = yh(, - 1)+1 , and that by the formalism yh(-l )+1 = 1 SO that
yh(-l)+l is of degree zero.
However, it is plain - say again from the matrix formulaire, that
It follows that the inequalities above all are equalities, that is, deg yh(, = -
d e g y h ( ~ - i ) +and
~ degyh(j)+i - degyh(,) = degbj+l, and the yh(,) must account
for all the convergents of F as claimed. 0
In [2] it is shown that all partial quotients of the product
G3,+( X ) = (1 + X-"I) =
h=O
have degree 1 .
Indeed, the partial quotients do all appear to be linear, and their coefficients
grow in complexity at a furious rate - the 30th partial quotient is
- 1374389534720~
- 13743895347200 .
15737111 456376219 1
the 50th is
Indeed, it is the simplest case of the Principle that if degyh = h for all h then
the same is true for the original function, that is: deg yh = h for all h . But it is
easily confirmed that degyh = h for all h for (1 X-')-'/' + over IF3; specifically
(1 + =
~ - l ) - l ' ~ [ l,X , -X + 1 ,X - 1 1 .
So of course dh = deg yh = h is true for &,,(l+
- X-3"), as Cantor pointed out.
3. Series with a Specialisable Continued Fraction Expansion
As remarked, examples with large gaps - doubling the exponent or more - are
not interesting. In that spirit, the example of [16]:
ts interesting. Here the sequence of exponents (F,,) of the series is defined by the
recurrence relation Fh+2 +
= Fh+l F h and the initial values F2 = 1 , F3 = 2 ; and
+
the point is that the golden ratio (1 &)/2 is less than 2. Notwithstanding that,
in [16]we succeed in proving the validity of the expansion by contorted applications
of the Folding Lemma (Proposition 1.3).
Here the expansion is speczalisable, in the sense that we may replace X by
a positive integer, other than 1, and readily obtain a valid numerical continued
fraction expansion - because it is a simple matter to modify nonpositive partial
quotients in a continued fraction expansion. Of course, this is the same thing as
the expansion having integer coefficients - and thus good reduction at all primes.
One wonders which other recurrence relations produce sequences of exponents
proving a specialisable continued fraction expansion. Experiment suggests that
this class may well be restricted t o certain sequences (Gn.h) satisfying the relation
Gn,h+ntl + Gn.h = 2Gn,h+, . Those are translates by a constant of sequences
satisfying
+
Gn.h+n = Gn.h+n-1 Gn.h+n-2 . . . Gn,h . + +
The case n = 2 of course includes the sequence of Fibonacci numbers.
chosen positive if arid only if > 1 and its co11,jugate 7; satisfies -1 < 7; < 0 .
Nonetheless, various simple transformations. specifically the rules
.
so that y(x) = [ao(x) , a , (x) : . . . a,.-, (T) 1 . Here. a polynomial partial ql~otient
is inadmissible only if it is constant.
Now suppose that y" = D ( r ) is a polynomial, riot a perfect square in K[.r].
Just as in the numerical case, prcsuniing the exist,encc of a unit p ( x ) - q(.r)y shows
t,hat has an expansion
with each of the a , of positive degree. hloreovcr, because the trace T ( x ) vanisl~es.
it follows that the word ( 1 1 (3.).. . . , a , - ( x ) is a palindrorrie. Not ice, incidentally.
2
that the equation (JI(T))' - ( q ( ~ ) y) = &1 entails that D has even degrec. and
leading coefficient a square in K.
352 ALFRED J. VAN DER POORTEN
But these ~ondit~ions do not suffice for the existence of a unit p(x) - q ( r ) y :
unless K is a finite field. In brief, in number fields one proves the existence of
nontrivial units by applications of the box principle. For infinite fields K , though.
such arguments fail for function fields over K because there are i~lfinit~ely many
polynomials of each degree. In all. given a generator y of a quadratic extension it
is unusual for t,here to he a unit of the shape p ( r ) - q ( r ) y . Indeed, with y" D(x)
suppose there exists such a nontrivial unit. It being a unit. its divisor class on t,he
Jacobian of the curve y2 = D ( x ) is supported at the two divisors at infinity: say
x+ and o o . Thus for some positive integer n , n ( m + - m - ) is the divisor of a
function. I11 other words. the Jacobian has a torsion point of order n .
Not much is as yet known about such t,orsion points on Jacobians of arbitrary
genus g - corresponding t,o deg D ( x ) = 2g+2 ([8]is a useful current reference with
helpful examples and extensive guide to the literature). However, if K = Q and
g = 1 we have an elliptic curve over Q and then a theorem of hIazur [lo]details all
possibilities for rational torsion. In particular, the highest torsion possible is 12.
IVe saw above that has a preperiod of length one: it follows that the
case 71. = 12 should yield a period of length 11. For a more leisurely explanation
see [3]. I am not aware of such a periodic continued fraction expansion ever having
been ~nentiorledin the literature arid take this opport,unity to display a computation
of Xuan Chuong Tran detailing up t,o isomorphism the entire class of Q-curves
K X: our criterion does not demand that its norm be h l . Below we display the
first twelve partial quotients of the expansion of y , but just the first half of its
ptlriod - which is of length 22; twice as long as suggested. What has occurred
is that the 11-th convergent p l o / q l o yields a 'unit' plo(x) - qI0(x)y with norm
+
- +
c(t) = 2t(2t - 1)'(3t2 3t l ) ' / ( t 1)" in Q X . Suppose [ a o . w , a l l ] denotes
-
the first 12 partial quotients of y . Then the complete expansion of y - given by the
2 2 i t
unit (plo(.r)-qlo(.r)y) / ( c ( t ) ) i s o n t h e o n e h a n d [ a o , w . a l l , a l l , w , O . a o ] .
by the known symmetry, on the other hand [ u o , w , a l l , ao/c(t) , c ( t ) Z , 0 , a0 1 .
i
h\- our present remark on yuasz-units. Here the word f;; is the string of partial
q;oticGs in reverse order a n d [ c ( t ) ] 1s the continued fraction o f c(t)[ ; one
of course recalls that d [ n , h , c , . . . ] = [ a d , h/d , cd , . . . ] . It's now plain that w
must itself be c-sy~mnetric in the sense that c z = f;;; : as one readily confirms.
-
By the way. these phcnornena are present in the numcrical case. It's just that then
multiplication by \om? constar~tc is rather more difficult to notice; cf [12].
REDUCTION O F CONTINUED FRACTIONS O F FORMAL POWER SERIES 353
a,, ( 2 )
Take t = 2 as an instance. Of course, this example is detailed purely for fun, to show
it can be done. But the principle it exemplifies is important in the context of our
discussiorl of formal power series with 'exceptional' continued fraction expansion:
with good reduction almost everywhere. That phenomenon is rare. It occurs only
when the Jacobian of the corresponding curve has a torsion point at infinity.
354 ALFRED 3. VAN DER POORTEN
The issue is classical and motivatmedhbcl's st,udy of what we now call abclia~i
integrals. In the present,. hyperelliptic. case t,he quest,ion concerns so-c:allrd quasi-
ellipt,ic int,egrals for which there is some polynomial f ( s )so that the differential
-
But, for us, this is an aside. The point is that onl?- in t,he exceptional cases just
described the square root of a polynomial has a continlied fraction expansion with
collvergents growing in height at a reasonable exponent,ial rate. One guesses t>hatin
all but thosc cxceptional ca.ses in which one expects some sort of quasi-periodicity
-
for the continued fraction expansion of all generators of the quadratic cxtmsio~ifield
- t,he cont,inued fraction expansion of an algebraic power series will be 'generic' in
t,he spirit suggested in the ~ a r l i e rsections of our remarks.
L$'it,h a finitc basc field all these expansions will be periodic. However. only in
the except,ional cascs (wlien there is an appropriate t,orsion divisor on the .Jacobian)
should we expect that t,he cont,inued fraction expansions have good reduction allnost
everywhcrc. It may be interest,ing in t,he generic case t,o see just how the length and
complexity of the period varics as one considers rtductions rnodulo p with diffcrcnt
primes p .
That mattcr is not totally unexplored. One reason t o study the coritirlued
fraction expansion of is to lcarn about the expansion of for intcgcrs
rr . This is the theme uriderlj-ing such studies as [7]. hloreover. it is a theorem of
Scliinzel [18] that the lrngt,h of the pcriod of tht, c:ontinucd fraction expansion of
JD(n) is bounded as n varies if and only if y = JD(I) provitles a nont,rivial ~ m i t
p ( s ) - q ( . c ) y which moreover has p arid q with coefficient in $ ~ [ ;sbut
] that last
condit,ion is obtainable in effect by restricting n to certain arithmetic progressioris.
The issue of 'convergentms'with coefficients of reasonable height is relevant in the
theory of diophantine approximation. Recent work of Bornbieri and Paula Cohen
[4] shows that t,he phenomena briefly described above have their a d o g u e in t,he
simultaneous Pad6 approximation of algr,braic: fur~ctionsof dt,grt:c, highrr t ha11 two.
5. Acknowledgements
hly rernarks linashamedly constitute a 'confcwnce paper' to the extent that they
describe work and thinking in progress and that, in their first three sections. the)
have subst ant ial intcrscction with earlier such remarks [14]. That intersection.
however, seemed essential both to allow this work to be read indcpmdcnt 1) and t o
REDUCTION O F CONTINUED FRACTIONS O F FORMAL P O W E R SERIES 355
References
[1] Willianl U'. Adarr~sand hIichacl J . Razar. 'Multiples of points on elliptic curves an(1 continuc(l
fractions', Proc. Sondon Alnth. Soc. 41 (1980). 481 398.
[2] .J. P. Allourhe, XI. hIcn(l&s France and A. J . van der Poorten, 'An infinite product with
tmurldctl partial qnotients', Actn Amth,. 59 (1091), 171 182.
[:3] T. G . Berry. 'On pcrio(licity of conti~iucdfractions in hyperelliptic functiou fields', Arch.
Math. 55 (1990), 259 266.
[4] Enrico Bornbicri am1 Paula B. Cohcn. 'Sicgcl's Lemma, Pad6 Approxinlations and Jacobians'
(with a n appendix t ~ yUni1)crto Zannicr); t o appear in the De Giorgi volurnc. Anrlali Scuola
Norr~lalcSuperiorc. Pisa
[5] David G . Cantor, 'Oil thc corrtiiluetl fractions of quadratic surds', Actn Arzth. 68 (199-1),
2%; ,305.
[ f i ] LIicl~clDckkiug. XIichcl hIen(i&h France arid Alf van dcr Poorten. 'FOLDS!', The Alathe-
4 (1082). 130-138; 11: LSyrrnri~try
rnat7rid I~i,te11z,qer~t:t~~. disturbed'. zbzd. 173-181; 111: 'More
morphisnrs', ah7d. 190-195.
[7] E . Duhois ct R . Paysan-LC Roux. 'Stir la longeur tiu dcvelopperricnt erl fraction continue de
m'. A S ~ ~ ~ Z198S ~200
U C (1991), 107 119
[8] Evcrctt U'. Howe, Frar~ckLeprevost. and Bjorn Poonen. 'Large torsion subgroups of split
Jacobiarls of curvcs of gcrrus two or three'. preprint .57, September 8 , 1998. located at
< http://~x~ww.~riatl~.~~iuc,cdu/Algebraic-Nmber-Theory/ >.
[I)] Donald E. K r ~ u t h ,T ~ LArt C of Computer Programming, 2 Se~nirlurncricalAlgorithms (1969:
211tl printing). pp:HiOff'.
[10] B. hlazur, .hIodular curves and the Eiseristcin ideal', Inst. Hnutes ~ t u d e sScz. Publ. Alnth.
47 (1977). :33 186.
[ I l l hlichel hIetltl?s Fra~icc,'Sur les fractions contirlucs limitdcs'. Acta Arzth. 23 (1973). 207 21.5.
[12] R. A. hlollin. A. J . van tler Poortcn an(l H. C. \.lTillianls. 'Halfway t o a solution of S 2 - IIY' =
. ' ( ' . .I. Thc:or.~cd c Norr~bri'sde B o r d e i ~ u . 6
~ , (10144). 421 -159
[13] A . J . van tlcr Poortcn, 'An iritrotluctiori t o continued frartions'. in J . H. Loxton and A. .J.
van der Poorten cds.. Dzophant7r~eAnnly~7s(Cambridge University Press, 1986). 149 138.
[14] 41f van tlcr Poortcn. 'Formal power series and their cor~tinucdfraction rxpansion', iri Joe
Buhlcr erl.. Algor7thmzc N u r ~ i h ~l'hcory
r (Proc. Third Intcrnatiorial Syrnposium, ANTS-111,
Portlan(1. Ort:gon, .June 1998), Springer. Lecture Notes in Cornputer Science 1423 (1998),
12pp.
[l>]A. .J. van tler Poortcrl am1 ,J. Sliallit, 'Foltied co11tirnlt:d fractions'. J . Number Theory 40
(1992), 237 250 .
[l6] A. .T. van (lcr Poorten a i d J. Shallit. 'A specialiscti contir~uctlfraction', Cnnad. J . Afath. 45
(1!)93), 10fi7-1079
[17] C:. N. Raney. 'On continucd fr;~ctionsand finite automata'. Math. Ann. 206 (1973). 265 283.
[18] A. Schinzcl. 'On some prot~lerrlsof the arithmetical theory of continued fractions II', Acta
Ar~th7 . (1962). 287 298.
Conterrlporary M a t h e m a t i c s
\.'c,lurnr 236, 1999
Haakon Waadeland
A s s ~ ~ a c .Recently
r. there has been some interest in the use of Szego polyno-
mials (or morc generally orthogonal rational functions) as a tool for solving the
frequency analysis problem for trigonometric signals. One step in the method
is to create. from a measure defined by observations of the signal. a family
of positive definite sequences of moments. This is, in different ways, used to
construct a family of Szego polynomials. Asymptotic properties of these lead
t o the unknown frequences. In the limit processes involved in the method the
sequence of asymptotic values of the moments is no longer positive definite.
Some Toeplitz determinants, who else could have been used to compute the
limits of Szego polynornials, are zero, and hence useless for that purpose. This
has, so far, been dealt with in essentially two different ways: by going t o subse-
quences or t o introduce a n additional parameter. In the present paper a quite
different approach is presented in two versions. Both start with the "useless"
determinants. One version is t o perturb one of the moments, the other one
is t o rcplacc one (or more) of the clenlents by zero. Under some mild comli-
tions both versions lead t o polynomials where some of the zeros immediately
produce the frequenccs.
The method was a t first merely a n observation of a strange (surprising)
property, hut was after a while properly established.
1. Background
A signal is received, assumed to be of the form
and let the observed sigiial value at thc time r r ~. At bc denoted s \ ( m ) . The
observatioiis are then givrri by
a i d 0 elsewllcrc. IZ'e dssunie that s\(0) # 0. The ilunibers u/,are sonietlnies called
normalized frequencies. Ile shall here call then1 frequencics, and the question of
cletcr~niliiiigthem shall be referred t o as the frequency analysis problem.
A method for solving the problern (with roots back t o \$'iener and Levinson)
may b r roughly described in the following way: Froin t,he signal values a cert,ain
absoIut,clj. coritiriuous rncasure on the unit circle is constructed. This givt~srise
t,o ail inner product, and in turn t o rnoinerlt,s and nionic orthogonal polynonlials
(Szrgo polynomials) on the unit circle. Asymptotic values of some zeros of the
polynomials then lead t,o the frequencies. Recently ort,hogonal polynorn,ia,Ls have
been replaced by orthogonal rattonal fun,ctions.
It is beyo~ldt,he scope ant1 the purpose of the prcscnt paper t o describe this
method. or rather these methods. since t,here arc variat,ions, essentially obtained by
different choiccs of the measure. Instead we refer to t.he papers [I]. [ 2 ] . [3], [4], [ 5 ]
and. for thc different choiecs of measures. thc paper 171.
Tlw method depends heavily upon tlw possibility of finding the Szego polyno-
mials (or more generally t h t orthogonal rational functions). The best way, from a
practical point of vicw, is t o use thc Levinson algorit,hin. Another way uses contin-
ued fract,ioris. Prrron-Carath4odory fractioris for Szegii-polynomials, Nevanlinna-
Pick fractions for orthogonal rational functions. A t,hird w q , although not a vcry
practical one. is the one of strongest relevance t o the present paper. There (in
the polynomial case) t,he Szegii polynomials arc given as det,errninant formulas.
wherc tlie elements are irlornerits. If, for a fixed and a fixed d e ~ r e en the Szego
polynoiiiial is denoted by p ! , \ ) ( z ) . thc formula is
where D(,' ' is the value of the upper left corner n x n Torplitz- determinant
\Ye bllall not be mterested in the explicit expressions for p i L ) ,but we need t o
kiiow tlie following rclation (see v.g. 121):
111 thc followiiig. p,,, shall always mean thcsc, abyrnptotic values of moments.
+
1%. dcfiiic r i l l := 2 1 L. where L = 1 if A. # 0 and 710 := 21 if Au = 0 (i.e.
r , , , is t 1 1 1111irrber
~ of frrcp(m~ies).A c r u ictl point in the solution of the frcquericy
SOME OBSERVATIONS IN FREQUENCY ANALYSIS 359
im ( z ) = (z - 1 ) ( z - e l " l ) ( z - P'"J 1.
A - x
J=1
Normally the number of frequencies is not known. But for any fixed rL the sequence
{&,") is uniformly bouridcd on the closed unit disk. An evcn more important point
is that for fixed n > no an) convergent subsequence has a limit polynomial with
(1.6) as a factor.
i%'e shall for convenience introduce x, = c o s ~ j , .which leads to the equality
We also have cos mij, = Tm(xJ),where T,, ( z j )is the Tchebycheff polynomial of the
first kind and degree m. Having introduced this, the formula (1.5) for the moment
limits may be written in the form
For any fixed N the sequence { p ! , ~ ) / l ~ ) , ,is positive-definite, and hence the
denominator determinant in (1.4) is always positive, and the forrnula produces a
sequence of Szego polynomials for every fixed N. This changes if we let N tend to
infinity. The sequence {p,,,} is exactly no-definite, implying. in view of (1.4). (1.5)
and (1.6). that for n = no we have
I Z ... Zrf I
whereds for n-values greater than n o , we get a fraction where numerator and de-
nominator both are 0 (see e.g. [ 2 ] ) . There are different ways of deahng with this:
To go back to (1.4) and go to subsequences, or to introduce an additional parameter
etc.. all described in the papers referred to in the present paper. Here we shall look
at the numerator determinant in (1.8) for the case rL > no. but w ~ t hsome changes
in the elements. These changes lead to certain observations, which at first sight
may look surprising. In the next section these. observations will be described, and
some examples will be presented. In the final section a key theorem will be stated
and proved. By using this theorem the observdtions may all be explained (proved).
360 HAAKON WAADELAND
2. The observations
Moment perturbat~on
Take in the numerator determinant in (1.8) n = no 1 and replace for a fixed +
+
rn the moment prr, by p,,, F . In the particular case where nl = 0 this looks like
PO+€ PI ... Pn
PI PO+t ... P11- 1
(2.1)
k - 1 - 2 ... P1
1 z ... zrl
If we expand this in powers of t, the "constant term" (non-epsilon term) will vanish,
and we get
(2.2) 6 .Q b ) + 0(f2),
where Q ( z ) is a polynomial in z, possibly identically 0.
with the zeros e""-" and e*'"~, which are the ones we want (inter~stingzeros). and
an additional (uninteresting) zero.
+
If, instead of n = no 1 we take n = no + k , the expansion in powers of t will
start with .'t i.e. it will be of the form
eh&(z) + 0(th+l),
where Q ( z ) is a polynornlal with zeros in all frequency points. or the determinant
is identically zero. This possibility occurs for instance if the c-modification is made
+
on p,,,, where m 2 no 2. since in that case the highest possible power of r in
the exparisiorl is 5 k . The last remark came out of a mathematical discussiori with
Vigdis Petersen.
SOME OBSERVATIONS IN FREQUENCY ANALYSIS 361
Example 2. Take a signal with two frequencies *w, 0 < w < T , and with
amplitude &
/.I' +
Take n = 4 = no 2 and let the 6- perturbation be on p2. Then
k = 2. By using MACSYMA we find
as a factor.
An alternative form of the observation follows from the expansion of the deter-
minant according- to row number p. Let at, be the element in row number 1 and
column number m, and let Al,, denote the cofactor. The determinant is
(2.4)
The left-hand expression is the determinant (2.4) with the element a,>, replaced
by 0. The right-hand side is, apart from the factor a,, and possibly the sign, the
determinant in the observation as stated above. Hence. in the observation we may
r e p l a c ~the determinant where row number p and colurnn number q are removed,
362 HAAKON WAADELAND
by the original detr~iniliant.but where the element in the pth row and qth col~imn
is replaced by 0 as seen below:
The removed element is p,,, for some rn. If this element happens t o be 0. its
removal will not change the value of the determinant (2.4). and we will still have a
determinant identically equal t o 0. That may (exceptionally) happen also in other
cases.
Example 3
h r a signal with the five frequencies. 0 arid fd L ,k = 1 , 2 and the arnplitudcs
A(,= 1, 2'44 = a. 2A? = b the inoinent limits arc
p,,, = 1 + aT,,, ( s ) + bT,,, (y). a > 0. 6 > 0,
wlieres:=cosdl, ~ : = C O S ~ / ~ .
Replace thc element in row number 2 and colurnn nurnber 4 by 0. This leads
to the following polynomial of degree 6 in z. factorized by I\IACSYhlA:
+
This will for instance be the case for q = no 2, in particular in the trivial case
+ +
(p, q) = (no 1. no 2). For C # 0 we have a polynomial of degree no 1, with the +
frequency points as its zeros and an additional zero at z = - D / C . If in particular
D = 0,the additional zero is z = 0. This happens if q = 1. Observe the trivial case
(P. q) = (131).
Proof, part 1. Since the elements are linear combinations of Tchebycheff polynomi-
als, a natural approach to a simplification of the determinant is to use the recurrence
relations for these. Repeated use will lead to a step-by-step removal of the variables
xJ.This method is closely related to the method used by Vigdis Petersen in [6].
It is convenient in the proof to use also the negative subscripts for the moments
(and rnornent limits). although we have avoided it in the formulas so far, since
p-,, = plrl. The moment limits are then given by
Next we repeat the process. but with x2 instead of xl and find in the same way
and the last step, now with XI,gives 0. With s = 0 this process yields a linear
combination of p l l I ,, C L ~ .,. ~. pr
- l,-nO
~ , equal to zero, and where the coefficients are
independent of m, and where the coefficient of p,, is 1. Lemma 2 is thus proved.
Remarks
For no = 2 the relation ( 3 . 6 ) takes the form
7 -( 1 +x ) ~ ~ +- (2l + ~ 1 ~ 2 ) ~ r 7 -1 - 2(x1
2 + ~2)1*.r!i-j + / 1 r ! ! - 1 = 0.
It is easy to go from one value of no in ( 3 . 6 ) to the next one (difference 2 ) by
the same type of recurrence as in the proof.
Lemma 2 shows that there is a linear combination of the colurrms 2 to no 2 +
+
in ( 3 . 2 ) such that the "new'' ( n o 2)th column has merely zeros except possibly in
place p and in the bottom place. By applying the same linear combination to the
+
columns I to no 1 we accomplish the sarnc in the "new" (no+ 1)th column. It
remains to find out what happens to the bottom place: Since the bottom elements in
the process are subject to the same procedure as the others, we find the polynomials
in the bottom elrnlents of the new two rightmost elements by replacing in (3.6) the
moments plrl-n by the power z""+'-' for the element in row no 2 and by z""-' +
for the element in row no + 1. From the above remarks we see that for no = 2 we
get
z ' - 2 x , z 2 + Z, z L - 2.r12+ 1 ,
366 HAAKON WAADELAND
+
and z times the latter in row no 2 = 6. Rather than using (3.6) we may apply
.
the procedure drscribed above directly to { z l )!=,ll,+, leading to more convenient
forms of the new elements. After one step we get {zl (zL 2 x p l)}~=,l,,p,.
- +
after two steps {zr (z' - +
2x12 1)(z2- 2x22 l))!_,,,,pj + etc., and in the end the
+ + +
(no 2, no 2)th element in the new determinant. The (no 2. no + l ) t h elemtnt is
obtamed in the same way, if we start from {z'))',,,~,. This conrludes part 1 of the
proof of Theorem 1. and hence part 2 is justified. We thus have a complete proof
of Theorem 1.
In the case when A. # 0 the procedure is the same, except for one additional
step at the end, producing the factor ( z - 1) in the two rightinost bottom elements,
and thus in (3.3) in the theorem.
Let it finally be mentioned, that in Theorem 1 the rlurnber 0 is not essential.
as ma) be seen from the proof, the theorem will still be true if it is replaced by any
other fixed number.
Remarks
\Ve need to indicate the bridge between Theorem 1 and the observation on the
t-perturbation. Take first n = no + 1, and replace a fixed p,,, by p,, + t . Then the
determinant may be written as a sum of three types of terms: First the original
determinant without any 6. That determinant has the value 0. Next t times a
sum of determinants obtained from the original one by removing exactly one row
(not the bottom one) and one column. Finally a term 0 ( c 2 ) . From Theorem 1
(keeping in mind the two equivalent forms of the observation) it follows that each
of the determinants where exactly one row and one column are removed either is
identically 0 or divisible by (3.6). Thus the observation on the 6-perturbation is
established as a mathematical fact in the case n = no + 1. That we, for n = no + k .
in the t- perturbation get the factor t' follows as in [6].
If. instead of having n = no + 1 in Theorem 1 we have n = n o + k , where
k 2 2 , we use the same column operation as the one established in Lemma 1. but
this timt not only 2 times, but k +
1 times, leading to a determinantt,where the
k + 1 right~nostcolumns have merely the element 0, except for k distinct rows.
say p l . p2, . . . , pk . where we have constants possibly different from zero, and thc
bottom row. where we have the polynomial (3.6). multiplied by powers of z. in
turn z", z' , . . . z L . Expansion according to the rightmost colurnn (number no + k +
1) leads to a polynomial of degree k multiplied by the factor (3.6). If the first
polynomial is not identically 0 it gives (up to) k uninteresting zeros, whereas the
second polynomial gives the interesting ones.
References
[I] W.B.Jones, 0.Njastati. W.J.Thron and H.Waadcland, Szego polynomaals applzed t o frequency
analyszs J . Comput. Appl. hlath. 46 (1993), 217 228.
[2] W.B.Joncs, O.Nj&stad and H.Waadcland, A n alternatzve w a y of using Szegii polynomials zn
frequency analyszs, in S.C.Cooper and W.J. Thron eds. Continued Fractions and Orthogonal
Functions, Marcel Dekker, New York, 141 152, 1994.
[3] O.Njastad and H.Waadeland General Szego theory i n f r e q u e m y analyszs, J . Math. An. and
Appl. 206 (1997), 280 307
[4] O.Njastarl and H.Waadeland, A s y m p t o t z c propertzes of zeros of orthogonal rational functzons,
J . Comput. Appl. hlath. 77 (1997), 25F -275.
[5] K.Pan and E.B.Saff, A s y m p t o t z c s for Zeros of Szego polynomaals assoczated 111zthtrzgolzomct-
rzc polynomzal szgnals, J . Approx. Theory 71 (1992), 2 3 9 2.51.
SOME OBSERVATIONS IN FREQUENCY ANALYSIS 367
Franck Wielonsky
1. Introduction
Padk approximation may be seen as one of the many ways of performing approx-
imation t o analytic functions in the complex plane. One of the main features of the
Padk approximants comes from the algebraic nature of their definition. Through-
out, P k will denote the set of polynomials with complex coefficients, of degree at
most k.
DEFINITION 1.1. Let f be a function analytic at the origin. The Pad6 approx-
imant of degree (m, n) is defined as the rational function P,,,/Q,,, such that
(Qrn,, f - P m , n ) ( ~=) O ( Z ~ + ~ - ' ) as z -+ 0,
with P,,, E P, and Q,,, E P,.
Thus, given the Taylor's coefficients of the function f at the origin. the Pad6 ap-
proximants can be explicitely computed by solving a set of linear equations. There
exists a whole theory based on algebraic tools such as determinants which leads
t o numerous identities, recursion relations and algorithms. In this connection, one
can also mention the strong links that exist between Pad6 approximants, continued
fractions and orthogonal polynomials.
The other aspect of the theory is the analytic aspect and the main interest,
here, lies in properties such as convergence, asymptotics and distribution of zeros.
In this respect, the Padk (or Baker-Gammel-Wills) conjecture plays a prominent
role which predicts t h a t , given a meromorphic function f . there exists an infinite
subsequence N C N such that the Pad6 approximants of degree (n. n), n E N
converge locally uniformly t o f , away from the poles of f . as n tends t o infinity
(cf. [Stag71 for a recent overview of this conjecture). In general, analyticity is not
Let us just recall that there exists another type of Hermitepad6 approximants.
the german type or type 11, which consists of simultaneous rational approximants.
Concerning the algebraic and analytic aspects of Hermite-Padi. approximants, we
refer to [BGM96, dB85, Coa66, Coa67, Mah681 and [AS92, Nut84, Stat381
respectively.
In the sequel, we shall only consider the approximants in Definition 1.2 when
specializing the choice of t:ie vector of functions ( f o . . . , f 7 n ) to be the vector of
exponentials (1, e Z ,. . . . em '). An important property of such a vector is that it
constitutes an example of a perfect system. It means that for any multi-index
n = ( n o , .. . , n,). any solution Ao, . . . , A m of (1.1) satisfies
Note that upon dividing the previous equation by eZ and changing z into -z, we
get
Pm,n(-z)eZ - Qm.n(-z) = O ( zm+n+l 1.
which implies, by uniqueness, that Q,,,(r) arid P,,,(-r) are equal, up t o a con-
stant. The following theorem of Pad6 shows that for a rational function, interpo-
lating the exponential function at zero with an order as high as possible suffices to
imply uniform convergence in the complex plane to this exponential function.
THEOREM 2.1. W i t h P,,, and Q,., satisfying ( 2 . 1 ) , w e have
+ m ) !/ + a)"'tndt,
1
Pm,n ( z ) = e-'(t
(n
where Co is a circle centered at the origin and of radius less than 1: define poly-
nomials of degree np - 1 satisfying ( 1 . 1 ) . Then, it is natural to ask whether the
previous theorem can be generalized to Hermiteepad6 approximants. The following
result, whose proof can be found in [Wie97]>answers this question in the diagonal
+
case, that is when considering m 1 polynomials A o , . . . , A , such that
and all the polynorriials Ao, . . . A, are of degree less than the same constant integer
n E W.
372 FRANCK WIELONSKY
For n large, let A s 2 and A"2 be the polynomials obtained upon dividing ~ ( 1 and :~
A ( ~ respectively
) by the nonzero derivative at zero and nonzero constant coeficient.
m/2
T h e n , as n + c o ,
where C , is a circle centered at the origin and of radius greater than m, one may
also show that in the diagonal case
n
3&
where PO,, may be seen to equal 1
3 4 . G (7)
3. S o m e differential e q u a t i o n s
Let us now establish the differential equations satisfied by the Hermiteepad6
approximants Ao, . . . , A, such that (2.5) holds. For clarity, as before, we shall limit
ourselves t o the diagonal case, though the general case can be treated in a similar
way. First, in connection with the Pad6 approximants Pn := P,,, and Q, := Q,,,
defined by (2.1), with m = n , we set
~ " ( 2 )= e-z/2z-n Pn(.).
Then, w,(z) satisfies Whittaker's equation (cf. [Olv54, p.2601)
- -
On the other hand, the right hand side of the previous equation equals zL(D)R(z).
Hence we get the differential equation
374 FRANCK WIELONSKY
Since the functions 1, e Z ,e2", . . . , em" are linearly independent over the rational
functions, the differential equation holds for each of the summands A p ( z ) e p z of
R ( z ) . Hence
z L ( D ) ( A p e p Z-) ( n - l ) L 1 ( D ) ( A , e p z )= 0 .
Because of the identity D(epZu)= epz(D + p ) u , this implies that
< <
These observations allows one to compute, for each p, 0 p m. a recurrence re-
lation involving the polynomials A , . . . , A,-,-,, '.
Indeed. from what precedes,
(1)
there are linear relations between Ap,n-J and Ap,,, A p , n ,. . . , A$;). Hence, a linear
,,,-,-
relation between A,,,. . . . , A 1 can be established.
comments and especially for having supplied us with a proof of Theorem 3.1, simpler
than that given in a first version of this paper.
References
[AS921 A.I. Aptekarev and H. Stahl. A s y m p t o t i c s of Hermite-Pad6 polynomials, Progress in
Approximation Theory (A.A. Gonchar and E.B. Saff, eds.), Springer Verlag, 1992.
pp. 127-167.
[Ass981 W . Van Assche, A p p r o x i m a t i o n theory a n d analytic n u m b e r theory, Special Functions
and Differential Equations (New Delhi) (K. Srinivasa Eao et al., ed.). Allied Publishers,
1998, pp. 336-355.
[Beu81] F. Beukers, P a d 6 approxzmatzon i n n u m b e r theory, Pad6 approximation and its appli-
cations (M.G. de Bruin and H. van Rossum, e d ~ . )Springer , Lecture Notes, vol. 888,
Springer. 1981, pp. 90-99.
[BGM96] G.A. Baker and P. Graves-Morris. P a d 6 a p p r o x i m a n t s , Cambridge University Press,
1996.
J . Coates, O n the algebraic approximation of functions I, II, III, Indag. Math. 28
(1966), 421-461.
J. Coates, O n the algebraic approxzmation of functions IV. Indag. Math. 29 ( 1 9 6 7 ) ,
205-212.
M.G. de Bruin, Simultaneous Pad6 approximation. and orthogonalzty, Polynomes or-
thogonaux et applications (C. Brezinski et al.. ed.): Springer Lecture Notes. vol. 1171.
Springer, 1985, pp. 74-83.
A. Van der Poorten, A proof that Euler missed ... Aptry's proof of the irrationality of
< ( 3 ) ,New Mathematical Intelligencer 1 ( 1 9 7 9 ) , 195-203.
D.S. Lubinsky, Spurious poles i n diagonal rational approxzmation, Progress in approx-
imation theory (New York) (A.A. Gonchar et al., ed.), Springer Ser. Comput. Math.,
vol. 19, Springer, 1992, pp. 191-213.
K . Mahler, Perfect systems, Comp. Math. 19 ( 1 9 6 8 ) , 95-166.
J . Nutall, Asymptotzcs of diagonal Hermite-Pad4 approzimants, J. Approx. Theory 42
( l 9 8 4 ) , 299-386.
F . W . J . Olver, T h e asymptotic expansion of Bessel functzons of large order. Phil. Trans.
Roy. Soc. London Ser. A 247 (1954), 328-368.
0 . Perron, Die Lehre von den Kettenbruchen, vol. 2, B. G. Teubner, Stuttgart, 1957.
M. Prkvost, A new proof of the irrationalzty of C(2) and C(3) using Pad6 approximants.
J . Comp. and Appl. Math. 67 (1996), 219-235.
H . Stahl, Asymptotics of Hermite-Padt polynomials and related convergence results
- a summary of results, Non linear Numerical Methods and Rational Approximation
(A. Cuyt, e d . ) , Reidel Publ. Corp., 1988, pp. 23-53.
H . Stahl, T h e convergence of diagonal P a d t approximants and the Pad6 conjecture, J .
Comp. and Appl. Math. 86 (1997), 287-296.
E.B. Saff and R.S. Varga, O n the zeros and poles of Pad6 approximants t o e x , Numer.
Math. 25 ( 1 9 7 5 ) , 1-14.
E.B. Saff and R.S. Varga, Zero-free parabolic regzons for sequences of polynomials, SIAM
J . math. Analysis 7 (1976), 344-357.
E.B. Saff and R.S. Varga, O n the zeros and poles of Pad6 approximants to e x , 11, Pad6
and Rational Approximation, Theory and Appl. (New York) (E.B. Saff and R.S. Varga,
eds.), Academic Press, 1977, pp. 195-213.
E.B. Saff and R.S. Varga, O n the zeros and poles of P a d t approximants t o e x , 111,
Numer. Math. 30 (1978), 241-266.
G. SzegG, ~ b e eine r eigenschaft der exponent~alreihe,Sitzungsber. Berl. Math. Ges. 23
( 1 9 2 4 ) , 50-64.
H. Wallin, T h e convergence of Pad6 approximants and the size of the power series
coeficients, Appl. Anal. 4 (1974), 235-251.
F. Wielonsky, Asymptotics of diagonal Hermite-Pad6 approxzmants to e Z , J . Approx.
Theory 90 (1997), 283-298.